JumpThreading.cpp revision 6f84a5feee5c7322a5145992bd4704225987909c
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/Transforms/Utils/BasicBlockUtils.h" 21#include "llvm/Transforms/Utils/Local.h" 22#include "llvm/Transforms/Utils/SSAUpdater.h" 23#include "llvm/Target/TargetData.h" 24#include "llvm/ADT/DenseMap.h" 25#include "llvm/ADT/Statistic.h" 26#include "llvm/ADT/STLExtras.h" 27#include "llvm/ADT/SmallPtrSet.h" 28#include "llvm/ADT/SmallSet.h" 29#include "llvm/Support/CommandLine.h" 30#include "llvm/Support/Debug.h" 31#include "llvm/Support/raw_ostream.h" 32using namespace llvm; 33 34STATISTIC(NumThreads, "Number of jumps threaded"); 35STATISTIC(NumFolds, "Number of terminators folded"); 36STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 37 38static cl::opt<unsigned> 39Threshold("jump-threading-threshold", 40 cl::desc("Max block size to duplicate for jump threading"), 41 cl::init(6), cl::Hidden); 42 43namespace { 44 /// This pass performs 'jump threading', which looks at blocks that have 45 /// multiple predecessors and multiple successors. If one or more of the 46 /// predecessors of the block can be proven to always jump to one of the 47 /// successors, we forward the edge from the predecessor to the successor by 48 /// duplicating the contents of this block. 49 /// 50 /// An example of when this can occur is code like this: 51 /// 52 /// if () { ... 53 /// X = 4; 54 /// } 55 /// if (X < 3) { 56 /// 57 /// In this case, the unconditional branch at the end of the first if can be 58 /// revectored to the false side of the second if. 59 /// 60 class JumpThreading : public FunctionPass { 61 TargetData *TD; 62#ifdef NDEBUG 63 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 64#else 65 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 66#endif 67 public: 68 static char ID; // Pass identification 69 JumpThreading() : FunctionPass(&ID) {} 70 71 bool runOnFunction(Function &F); 72 void FindLoopHeaders(Function &F); 73 74 bool ProcessBlock(BasicBlock *BB); 75 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 76 BasicBlock *SuccBB); 77 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 78 BasicBlock *PredBB); 79 80 typedef SmallVectorImpl<std::pair<ConstantInt*, 81 BasicBlock*> > PredValueInfo; 82 83 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 84 PredValueInfo &Result); 85 bool ProcessThreadableEdges(Instruction *CondInst, BasicBlock *BB); 86 87 88 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 89 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 90 91 bool ProcessJumpOnPHI(PHINode *PN); 92 93 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 94 }; 95} 96 97char JumpThreading::ID = 0; 98static RegisterPass<JumpThreading> 99X("jump-threading", "Jump Threading"); 100 101// Public interface to the Jump Threading pass 102FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 103 104/// runOnFunction - Top level algorithm. 105/// 106bool JumpThreading::runOnFunction(Function &F) { 107 DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n"); 108 TD = getAnalysisIfAvailable<TargetData>(); 109 110 FindLoopHeaders(F); 111 112 bool AnotherIteration = true, EverChanged = false; 113 while (AnotherIteration) { 114 AnotherIteration = false; 115 bool Changed = false; 116 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 117 BasicBlock *BB = I; 118 // Thread all of the branches we can over this block. 119 while (ProcessBlock(BB)) 120 Changed = true; 121 122 ++I; 123 124 // If the block is trivially dead, zap it. This eliminates the successor 125 // edges which simplifies the CFG. 126 if (pred_begin(BB) == pred_end(BB) && 127 BB != &BB->getParent()->getEntryBlock()) { 128 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName() 129 << "' with terminator: " << *BB->getTerminator() << '\n'); 130 LoopHeaders.erase(BB); 131 DeleteDeadBlock(BB); 132 Changed = true; 133 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 134 // Can't thread an unconditional jump, but if the block is "almost 135 // empty", we can replace uses of it with uses of the successor and make 136 // this dead. 137 if (BI->isUnconditional() && 138 BB != &BB->getParent()->getEntryBlock()) { 139 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 140 // Ignore dbg intrinsics. 141 while (isa<DbgInfoIntrinsic>(BBI)) 142 ++BBI; 143 // If the terminator is the only non-phi instruction, try to nuke it. 144 if (BBI->isTerminator()) { 145 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 146 // block, we have to make sure it isn't in the LoopHeaders set. We 147 // reinsert afterward in the rare case when the block isn't deleted. 148 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 149 150 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) 151 Changed = true; 152 else if (ErasedFromLoopHeaders) 153 LoopHeaders.insert(BB); 154 } 155 } 156 } 157 } 158 AnotherIteration = Changed; 159 EverChanged |= Changed; 160 } 161 162 LoopHeaders.clear(); 163 return EverChanged; 164} 165 166/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 167/// thread across it. 168static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 169 /// Ignore PHI nodes, these will be flattened when duplication happens. 170 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 171 172 // Sum up the cost of each instruction until we get to the terminator. Don't 173 // include the terminator because the copy won't include it. 174 unsigned Size = 0; 175 for (; !isa<TerminatorInst>(I); ++I) { 176 // Debugger intrinsics don't incur code size. 177 if (isa<DbgInfoIntrinsic>(I)) continue; 178 179 // If this is a pointer->pointer bitcast, it is free. 180 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 181 continue; 182 183 // All other instructions count for at least one unit. 184 ++Size; 185 186 // Calls are more expensive. If they are non-intrinsic calls, we model them 187 // as having cost of 4. If they are a non-vector intrinsic, we model them 188 // as having cost of 2 total, and if they are a vector intrinsic, we model 189 // them as having cost 1. 190 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 191 if (!isa<IntrinsicInst>(CI)) 192 Size += 3; 193 else if (!isa<VectorType>(CI->getType())) 194 Size += 1; 195 } 196 } 197 198 // Threading through a switch statement is particularly profitable. If this 199 // block ends in a switch, decrease its cost to make it more likely to happen. 200 if (isa<SwitchInst>(I)) 201 Size = Size > 6 ? Size-6 : 0; 202 203 return Size; 204} 205 206 207//===----------------------------------------------------------------------===// 208 209/// ReplaceAndSimplifyAllUses - Perform From->replaceAllUsesWith(To) and then 210/// delete the From instruction. In addition to a basic RAUW, this does a 211/// recursive simplification of the newly formed instructions. This catches 212/// things where one simplification exposes other opportunities. This only 213/// simplifies and deletes scalar operations, it does not change the CFG. 214/// 215static void ReplaceAndSimplifyAllUses(Instruction *From, Value *To, 216 const TargetData *TD) { 217 assert(From != To && "ReplaceAndSimplifyAllUses(X,X) is not valid!"); 218 219 // FromHandle - This keeps a weakvh on the from value so that we can know if 220 // it gets deleted out from under us in a recursive simplification. 221 WeakVH FromHandle(From); 222 223 while (!From->use_empty()) { 224 // Update the instruction to use the new value. 225 Use &U = From->use_begin().getUse(); 226 Instruction *User = cast<Instruction>(U.getUser()); 227 U = To; 228 229 // See if we can simplify it. 230 if (Value *V = SimplifyInstruction(User, TD)) { 231 // Recursively simplify this. 232 ReplaceAndSimplifyAllUses(User, V, TD); 233 234 // If the recursive simplification ended up revisiting and deleting 'From' 235 // then we're done. 236 if (FromHandle == 0) 237 return; 238 } 239 } 240 From->eraseFromParent(); 241} 242 243 244/// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this 245/// method is called when we're about to delete Pred as a predecessor of BB. If 246/// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred. 247/// 248/// Unlike the removePredecessor method, this attempts to simplify uses of PHI 249/// nodes that collapse into identity values. For example, if we have: 250/// x = phi(1, 0, 0, 0) 251/// y = and x, z 252/// 253/// .. and delete the predecessor corresponding to the '1', this will attempt to 254/// recursively fold the and to 0. 255static void RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred, 256 TargetData *TD) { 257 // This only adjusts blocks with PHI nodes. 258 if (!isa<PHINode>(BB->begin())) 259 return; 260 261 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify 262 // them down. This will leave us with single entry phi nodes and other phis 263 // that can be removed. 264 BB->removePredecessor(Pred, true); 265 266 WeakVH PhiIt = &BB->front(); 267 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) { 268 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt)); 269 270 Value *PNV = PN->hasConstantValue(); 271 if (PNV == 0) continue; 272 273 // If we're able to simplify the phi to a single value, substitute the new 274 // value into all of its uses. 275 assert(PNV != PN && "hasConstantValue broken"); 276 277 ReplaceAndSimplifyAllUses(PN, PNV, TD); 278 279 // If recursive simplification ended up deleting the next PHI node we would 280 // iterate to, then our iterator is invalid, restart scanning from the top 281 // of the block. 282 if (PhiIt == 0) PhiIt = &BB->front(); 283 } 284} 285 286//===----------------------------------------------------------------------===// 287 288 289/// FindLoopHeaders - We do not want jump threading to turn proper loop 290/// structures into irreducible loops. Doing this breaks up the loop nesting 291/// hierarchy and pessimizes later transformations. To prevent this from 292/// happening, we first have to find the loop headers. Here we approximate this 293/// by finding targets of backedges in the CFG. 294/// 295/// Note that there definitely are cases when we want to allow threading of 296/// edges across a loop header. For example, threading a jump from outside the 297/// loop (the preheader) to an exit block of the loop is definitely profitable. 298/// It is also almost always profitable to thread backedges from within the loop 299/// to exit blocks, and is often profitable to thread backedges to other blocks 300/// within the loop (forming a nested loop). This simple analysis is not rich 301/// enough to track all of these properties and keep it up-to-date as the CFG 302/// mutates, so we don't allow any of these transformations. 303/// 304void JumpThreading::FindLoopHeaders(Function &F) { 305 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 306 FindFunctionBackedges(F, Edges); 307 308 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 309 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 310} 311 312/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 313/// if we can infer that the value is a known ConstantInt in any of our 314/// predecessors. If so, return the known list of value and pred BB in the 315/// result vector. If a value is known to be undef, it is returned as null. 316/// 317/// The BB basic block is known to start with a PHI node. 318/// 319/// This returns true if there were any known values. 320/// 321/// 322/// TODO: Per PR2563, we could infer value range information about a predecessor 323/// based on its terminator. 324bool JumpThreading:: 325ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 326 PHINode *TheFirstPHI = cast<PHINode>(BB->begin()); 327 328 // If V is a constantint, then it is known in all predecessors. 329 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 330 ConstantInt *CI = dyn_cast<ConstantInt>(V); 331 Result.resize(TheFirstPHI->getNumIncomingValues()); 332 for (unsigned i = 0, e = Result.size(); i != e; ++i) 333 Result[i] = std::make_pair(CI, TheFirstPHI->getIncomingBlock(i)); 334 return true; 335 } 336 337 // If V is a non-instruction value, or an instruction in a different block, 338 // then it can't be derived from a PHI. 339 Instruction *I = dyn_cast<Instruction>(V); 340 if (I == 0 || I->getParent() != BB) 341 return false; 342 343 /// If I is a PHI node, then we know the incoming values for any constants. 344 if (PHINode *PN = dyn_cast<PHINode>(I)) { 345 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 346 Value *InVal = PN->getIncomingValue(i); 347 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 348 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 349 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 350 } 351 } 352 return !Result.empty(); 353 } 354 355 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 356 357 // Handle some boolean conditions. 358 if (I->getType()->getPrimitiveSizeInBits() == 1) { 359 // X | true -> true 360 // X & false -> false 361 if (I->getOpcode() == Instruction::Or || 362 I->getOpcode() == Instruction::And) { 363 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 364 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 365 366 if (LHSVals.empty() && RHSVals.empty()) 367 return false; 368 369 ConstantInt *InterestingVal; 370 if (I->getOpcode() == Instruction::Or) 371 InterestingVal = ConstantInt::getTrue(I->getContext()); 372 else 373 InterestingVal = ConstantInt::getFalse(I->getContext()); 374 375 // Scan for the sentinel. 376 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 377 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) 378 Result.push_back(LHSVals[i]); 379 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 380 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) 381 Result.push_back(RHSVals[i]); 382 return !Result.empty(); 383 } 384 385 // TODO: Should handle the NOT form of XOR. 386 387 } 388 389 // Handle compare with phi operand, where the PHI is defined in this block. 390 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 391 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 392 if (PN && PN->getParent() == BB) { 393 // We can do this simplification if any comparisons fold to true or false. 394 // See if any do. 395 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 396 BasicBlock *PredBB = PN->getIncomingBlock(i); 397 Value *LHS = PN->getIncomingValue(i); 398 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 399 400 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS); 401 if (Res == 0) continue; 402 403 if (isa<UndefValue>(Res)) 404 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 405 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 406 Result.push_back(std::make_pair(CI, PredBB)); 407 } 408 409 return !Result.empty(); 410 } 411 412 // TODO: We could also recurse to see if we can determine constants another 413 // way. 414 } 415 return false; 416} 417 418 419 420/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 421/// in an undefined jump, decide which block is best to revector to. 422/// 423/// Since we can pick an arbitrary destination, we pick the successor with the 424/// fewest predecessors. This should reduce the in-degree of the others. 425/// 426static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 427 TerminatorInst *BBTerm = BB->getTerminator(); 428 unsigned MinSucc = 0; 429 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 430 // Compute the successor with the minimum number of predecessors. 431 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 432 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 433 TestBB = BBTerm->getSuccessor(i); 434 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 435 if (NumPreds < MinNumPreds) 436 MinSucc = i; 437 } 438 439 return MinSucc; 440} 441 442/// ProcessBlock - If there are any predecessors whose control can be threaded 443/// through to a successor, transform them now. 444bool JumpThreading::ProcessBlock(BasicBlock *BB) { 445 // If this block has a single predecessor, and if that pred has a single 446 // successor, merge the blocks. This encourages recursive jump threading 447 // because now the condition in this block can be threaded through 448 // predecessors of our predecessor block. 449 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 450 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 451 SinglePred != BB) { 452 // If SinglePred was a loop header, BB becomes one. 453 if (LoopHeaders.erase(SinglePred)) 454 LoopHeaders.insert(BB); 455 456 // Remember if SinglePred was the entry block of the function. If so, we 457 // will need to move BB back to the entry position. 458 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 459 MergeBasicBlockIntoOnlyPred(BB); 460 461 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 462 BB->moveBefore(&BB->getParent()->getEntryBlock()); 463 return true; 464 } 465 } 466 467 // Look to see if the terminator is a branch of switch, if not we can't thread 468 // it. 469 Value *Condition; 470 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 471 // Can't thread an unconditional jump. 472 if (BI->isUnconditional()) return false; 473 Condition = BI->getCondition(); 474 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 475 Condition = SI->getCondition(); 476 else 477 return false; // Must be an invoke. 478 479 // If the terminator of this block is branching on a constant, simplify the 480 // terminator to an unconditional branch. This can occur due to threading in 481 // other blocks. 482 if (isa<ConstantInt>(Condition)) { 483 DEBUG(errs() << " In block '" << BB->getName() 484 << "' folding terminator: " << *BB->getTerminator() << '\n'); 485 ++NumFolds; 486 ConstantFoldTerminator(BB); 487 return true; 488 } 489 490 // If the terminator is branching on an undef, we can pick any of the 491 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 492 if (isa<UndefValue>(Condition)) { 493 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 494 495 // Fold the branch/switch. 496 TerminatorInst *BBTerm = BB->getTerminator(); 497 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 498 if (i == BestSucc) continue; 499 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 500 } 501 502 DEBUG(errs() << " In block '" << BB->getName() 503 << "' folding undef terminator: " << *BBTerm << '\n'); 504 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 505 BBTerm->eraseFromParent(); 506 return true; 507 } 508 509 Instruction *CondInst = dyn_cast<Instruction>(Condition); 510 511 // If the condition is an instruction defined in another block, see if a 512 // predecessor has the same condition: 513 // br COND, BBX, BBY 514 // BBX: 515 // br COND, BBZ, BBW 516 if (!Condition->hasOneUse() && // Multiple uses. 517 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 518 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 519 if (isa<BranchInst>(BB->getTerminator())) { 520 for (; PI != E; ++PI) 521 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 522 if (PBI->isConditional() && PBI->getCondition() == Condition && 523 ProcessBranchOnDuplicateCond(*PI, BB)) 524 return true; 525 } else { 526 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 527 for (; PI != E; ++PI) 528 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 529 if (PSI->getCondition() == Condition && 530 ProcessSwitchOnDuplicateCond(*PI, BB)) 531 return true; 532 } 533 } 534 535 // All the rest of our checks depend on the condition being an instruction. 536 if (CondInst == 0) 537 return false; 538 539 // See if this is a phi node in the current block. 540 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 541 if (PN->getParent() == BB) 542 return ProcessJumpOnPHI(PN); 543 544 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 545 if (!isa<PHINode>(CondCmp->getOperand(0)) || 546 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) { 547 // If we have a comparison, loop over the predecessors to see if there is 548 // a condition with a lexically identical value. 549 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 550 for (; PI != E; ++PI) 551 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 552 if (PBI->isConditional() && *PI != BB) { 553 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 554 if (CI->getOperand(0) == CondCmp->getOperand(0) && 555 CI->getOperand(1) == CondCmp->getOperand(1) && 556 CI->getPredicate() == CondCmp->getPredicate()) { 557 // TODO: Could handle things like (x != 4) --> (x == 17) 558 if (ProcessBranchOnDuplicateCond(*PI, BB)) 559 return true; 560 } 561 } 562 } 563 } 564 } 565 566 // Check for some cases that are worth simplifying. Right now we want to look 567 // for loads that are used by a switch or by the condition for the branch. If 568 // we see one, check to see if it's partially redundant. If so, insert a PHI 569 // which can then be used to thread the values. 570 // 571 // This is particularly important because reg2mem inserts loads and stores all 572 // over the place, and this blocks jump threading if we don't zap them. 573 Value *SimplifyValue = CondInst; 574 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 575 if (isa<Constant>(CondCmp->getOperand(1))) 576 SimplifyValue = CondCmp->getOperand(0); 577 578 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 579 if (SimplifyPartiallyRedundantLoad(LI)) 580 return true; 581 582 583 // Handle a variety of cases where we are branching on something derived from 584 // a PHI node in the current block. If we can prove that any predecessors 585 // compute a predictable value based on a PHI node, thread those predecessors. 586 // 587 // We only bother doing this if the current block has a PHI node and if the 588 // conditional instruction lives in the current block. If either condition 589 // fails, this won't be a computable value anyway. 590 if (CondInst->getParent() == BB && isa<PHINode>(BB->front())) 591 if (ProcessThreadableEdges(CondInst, BB)) 592 return true; 593 594 595 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 596 // "(X == 4)" thread through this block. 597 598 return false; 599} 600 601/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 602/// block that jump on exactly the same condition. This means that we almost 603/// always know the direction of the edge in the DESTBB: 604/// PREDBB: 605/// br COND, DESTBB, BBY 606/// DESTBB: 607/// br COND, BBZ, BBW 608/// 609/// If DESTBB has multiple predecessors, we can't just constant fold the branch 610/// in DESTBB, we have to thread over it. 611bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 612 BasicBlock *BB) { 613 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 614 615 // If both successors of PredBB go to DESTBB, we don't know anything. We can 616 // fold the branch to an unconditional one, which allows other recursive 617 // simplifications. 618 bool BranchDir; 619 if (PredBI->getSuccessor(1) != BB) 620 BranchDir = true; 621 else if (PredBI->getSuccessor(0) != BB) 622 BranchDir = false; 623 else { 624 DEBUG(errs() << " In block '" << PredBB->getName() 625 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 626 ++NumFolds; 627 ConstantFoldTerminator(PredBB); 628 return true; 629 } 630 631 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 632 633 // If the dest block has one predecessor, just fix the branch condition to a 634 // constant and fold it. 635 if (BB->getSinglePredecessor()) { 636 DEBUG(errs() << " In block '" << BB->getName() 637 << "' folding condition to '" << BranchDir << "': " 638 << *BB->getTerminator() << '\n'); 639 ++NumFolds; 640 Value *OldCond = DestBI->getCondition(); 641 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 642 BranchDir)); 643 ConstantFoldTerminator(BB); 644 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 645 return true; 646 } 647 648 649 // Next, figure out which successor we are threading to. 650 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 651 652 SmallVector<BasicBlock*, 2> Preds; 653 Preds.push_back(PredBB); 654 655 // Ok, try to thread it! 656 return ThreadEdge(BB, Preds, SuccBB); 657} 658 659/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 660/// block that switch on exactly the same condition. This means that we almost 661/// always know the direction of the edge in the DESTBB: 662/// PREDBB: 663/// switch COND [... DESTBB, BBY ... ] 664/// DESTBB: 665/// switch COND [... BBZ, BBW ] 666/// 667/// Optimizing switches like this is very important, because simplifycfg builds 668/// switches out of repeated 'if' conditions. 669bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 670 BasicBlock *DestBB) { 671 // Can't thread edge to self. 672 if (PredBB == DestBB) 673 return false; 674 675 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 676 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 677 678 // There are a variety of optimizations that we can potentially do on these 679 // blocks: we order them from most to least preferable. 680 681 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 682 // directly to their destination. This does not introduce *any* code size 683 // growth. Skip debug info first. 684 BasicBlock::iterator BBI = DestBB->begin(); 685 while (isa<DbgInfoIntrinsic>(BBI)) 686 BBI++; 687 688 // FIXME: Thread if it just contains a PHI. 689 if (isa<SwitchInst>(BBI)) { 690 bool MadeChange = false; 691 // Ignore the default edge for now. 692 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 693 ConstantInt *DestVal = DestSI->getCaseValue(i); 694 BasicBlock *DestSucc = DestSI->getSuccessor(i); 695 696 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 697 // PredSI has an explicit case for it. If so, forward. If it is covered 698 // by the default case, we can't update PredSI. 699 unsigned PredCase = PredSI->findCaseValue(DestVal); 700 if (PredCase == 0) continue; 701 702 // If PredSI doesn't go to DestBB on this value, then it won't reach the 703 // case on this condition. 704 if (PredSI->getSuccessor(PredCase) != DestBB && 705 DestSI->getSuccessor(i) != DestBB) 706 continue; 707 708 // Otherwise, we're safe to make the change. Make sure that the edge from 709 // DestSI to DestSucc is not critical and has no PHI nodes. 710 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 711 DEBUG(errs() << "THROUGH: " << *DestSI); 712 713 // If the destination has PHI nodes, just split the edge for updating 714 // simplicity. 715 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 716 SplitCriticalEdge(DestSI, i, this); 717 DestSucc = DestSI->getSuccessor(i); 718 } 719 FoldSingleEntryPHINodes(DestSucc); 720 PredSI->setSuccessor(PredCase, DestSucc); 721 MadeChange = true; 722 } 723 724 if (MadeChange) 725 return true; 726 } 727 728 return false; 729} 730 731 732/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 733/// load instruction, eliminate it by replacing it with a PHI node. This is an 734/// important optimization that encourages jump threading, and needs to be run 735/// interlaced with other jump threading tasks. 736bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 737 // Don't hack volatile loads. 738 if (LI->isVolatile()) return false; 739 740 // If the load is defined in a block with exactly one predecessor, it can't be 741 // partially redundant. 742 BasicBlock *LoadBB = LI->getParent(); 743 if (LoadBB->getSinglePredecessor()) 744 return false; 745 746 Value *LoadedPtr = LI->getOperand(0); 747 748 // If the loaded operand is defined in the LoadBB, it can't be available. 749 // FIXME: Could do PHI translation, that would be fun :) 750 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 751 if (PtrOp->getParent() == LoadBB) 752 return false; 753 754 // Scan a few instructions up from the load, to see if it is obviously live at 755 // the entry to its block. 756 BasicBlock::iterator BBIt = LI; 757 758 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB, 759 BBIt, 6)) { 760 // If the value if the load is locally available within the block, just use 761 // it. This frequently occurs for reg2mem'd allocas. 762 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 763 764 // If the returned value is the load itself, replace with an undef. This can 765 // only happen in dead loops. 766 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 767 LI->replaceAllUsesWith(AvailableVal); 768 LI->eraseFromParent(); 769 return true; 770 } 771 772 // Otherwise, if we scanned the whole block and got to the top of the block, 773 // we know the block is locally transparent to the load. If not, something 774 // might clobber its value. 775 if (BBIt != LoadBB->begin()) 776 return false; 777 778 779 SmallPtrSet<BasicBlock*, 8> PredsScanned; 780 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 781 AvailablePredsTy AvailablePreds; 782 BasicBlock *OneUnavailablePred = 0; 783 784 // If we got here, the loaded value is transparent through to the start of the 785 // block. Check to see if it is available in any of the predecessor blocks. 786 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 787 PI != PE; ++PI) { 788 BasicBlock *PredBB = *PI; 789 790 // If we already scanned this predecessor, skip it. 791 if (!PredsScanned.insert(PredBB)) 792 continue; 793 794 // Scan the predecessor to see if the value is available in the pred. 795 BBIt = PredBB->end(); 796 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 797 if (!PredAvailable) { 798 OneUnavailablePred = PredBB; 799 continue; 800 } 801 802 // If so, this load is partially redundant. Remember this info so that we 803 // can create a PHI node. 804 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 805 } 806 807 // If the loaded value isn't available in any predecessor, it isn't partially 808 // redundant. 809 if (AvailablePreds.empty()) return false; 810 811 // Okay, the loaded value is available in at least one (and maybe all!) 812 // predecessors. If the value is unavailable in more than one unique 813 // predecessor, we want to insert a merge block for those common predecessors. 814 // This ensures that we only have to insert one reload, thus not increasing 815 // code size. 816 BasicBlock *UnavailablePred = 0; 817 818 // If there is exactly one predecessor where the value is unavailable, the 819 // already computed 'OneUnavailablePred' block is it. If it ends in an 820 // unconditional branch, we know that it isn't a critical edge. 821 if (PredsScanned.size() == AvailablePreds.size()+1 && 822 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 823 UnavailablePred = OneUnavailablePred; 824 } else if (PredsScanned.size() != AvailablePreds.size()) { 825 // Otherwise, we had multiple unavailable predecessors or we had a critical 826 // edge from the one. 827 SmallVector<BasicBlock*, 8> PredsToSplit; 828 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 829 830 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 831 AvailablePredSet.insert(AvailablePreds[i].first); 832 833 // Add all the unavailable predecessors to the PredsToSplit list. 834 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 835 PI != PE; ++PI) 836 if (!AvailablePredSet.count(*PI)) 837 PredsToSplit.push_back(*PI); 838 839 // Split them out to their own block. 840 UnavailablePred = 841 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 842 "thread-split", this); 843 } 844 845 // If the value isn't available in all predecessors, then there will be 846 // exactly one where it isn't available. Insert a load on that edge and add 847 // it to the AvailablePreds list. 848 if (UnavailablePred) { 849 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 850 "Can't handle critical edge here!"); 851 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", 852 UnavailablePred->getTerminator()); 853 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 854 } 855 856 // Now we know that each predecessor of this block has a value in 857 // AvailablePreds, sort them for efficient access as we're walking the preds. 858 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 859 860 // Create a PHI node at the start of the block for the PRE'd load value. 861 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 862 PN->takeName(LI); 863 864 // Insert new entries into the PHI for each predecessor. A single block may 865 // have multiple entries here. 866 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 867 ++PI) { 868 AvailablePredsTy::iterator I = 869 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 870 std::make_pair(*PI, (Value*)0)); 871 872 assert(I != AvailablePreds.end() && I->first == *PI && 873 "Didn't find entry for predecessor!"); 874 875 PN->addIncoming(I->second, I->first); 876 } 877 878 //cerr << "PRE: " << *LI << *PN << "\n"; 879 880 LI->replaceAllUsesWith(PN); 881 LI->eraseFromParent(); 882 883 return true; 884} 885 886/// FindMostPopularDest - The specified list contains multiple possible 887/// threadable destinations. Pick the one that occurs the most frequently in 888/// the list. 889static BasicBlock * 890FindMostPopularDest(BasicBlock *BB, 891 const SmallVectorImpl<std::pair<BasicBlock*, 892 BasicBlock*> > &PredToDestList) { 893 assert(!PredToDestList.empty()); 894 895 // Determine popularity. If there are multiple possible destinations, we 896 // explicitly choose to ignore 'undef' destinations. We prefer to thread 897 // blocks with known and real destinations to threading undef. We'll handle 898 // them later if interesting. 899 DenseMap<BasicBlock*, unsigned> DestPopularity; 900 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 901 if (PredToDestList[i].second) 902 DestPopularity[PredToDestList[i].second]++; 903 904 // Find the most popular dest. 905 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 906 BasicBlock *MostPopularDest = DPI->first; 907 unsigned Popularity = DPI->second; 908 SmallVector<BasicBlock*, 4> SamePopularity; 909 910 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 911 // If the popularity of this entry isn't higher than the popularity we've 912 // seen so far, ignore it. 913 if (DPI->second < Popularity) 914 ; // ignore. 915 else if (DPI->second == Popularity) { 916 // If it is the same as what we've seen so far, keep track of it. 917 SamePopularity.push_back(DPI->first); 918 } else { 919 // If it is more popular, remember it. 920 SamePopularity.clear(); 921 MostPopularDest = DPI->first; 922 Popularity = DPI->second; 923 } 924 } 925 926 // Okay, now we know the most popular destination. If there is more than 927 // destination, we need to determine one. This is arbitrary, but we need 928 // to make a deterministic decision. Pick the first one that appears in the 929 // successor list. 930 if (!SamePopularity.empty()) { 931 SamePopularity.push_back(MostPopularDest); 932 TerminatorInst *TI = BB->getTerminator(); 933 for (unsigned i = 0; ; ++i) { 934 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 935 936 if (std::find(SamePopularity.begin(), SamePopularity.end(), 937 TI->getSuccessor(i)) == SamePopularity.end()) 938 continue; 939 940 MostPopularDest = TI->getSuccessor(i); 941 break; 942 } 943 } 944 945 // Okay, we have finally picked the most popular destination. 946 return MostPopularDest; 947} 948 949bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst, 950 BasicBlock *BB) { 951 // If threading this would thread across a loop header, don't even try to 952 // thread the edge. 953 if (LoopHeaders.count(BB)) 954 return false; 955 956 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 957 if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues)) 958 return false; 959 assert(!PredValues.empty() && 960 "ComputeValueKnownInPredecessors returned true with no values"); 961 962 DEBUG(errs() << "IN BB: " << *BB; 963 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 964 errs() << " BB '" << BB->getName() << "': FOUND condition = "; 965 if (PredValues[i].first) 966 errs() << *PredValues[i].first; 967 else 968 errs() << "UNDEF"; 969 errs() << " for pred '" << PredValues[i].second->getName() 970 << "'.\n"; 971 }); 972 973 // Decide what we want to thread through. Convert our list of known values to 974 // a list of known destinations for each pred. This also discards duplicate 975 // predecessors and keeps track of the undefined inputs (which are represented 976 // as a null dest in the PredToDestList). 977 SmallPtrSet<BasicBlock*, 16> SeenPreds; 978 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 979 980 BasicBlock *OnlyDest = 0; 981 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 982 983 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 984 BasicBlock *Pred = PredValues[i].second; 985 if (!SeenPreds.insert(Pred)) 986 continue; // Duplicate predecessor entry. 987 988 // If the predecessor ends with an indirect goto, we can't change its 989 // destination. 990 if (isa<IndirectBrInst>(Pred->getTerminator())) 991 continue; 992 993 ConstantInt *Val = PredValues[i].first; 994 995 BasicBlock *DestBB; 996 if (Val == 0) // Undef. 997 DestBB = 0; 998 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 999 DestBB = BI->getSuccessor(Val->isZero()); 1000 else { 1001 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1002 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1003 } 1004 1005 // If we have exactly one destination, remember it for efficiency below. 1006 if (i == 0) 1007 OnlyDest = DestBB; 1008 else if (OnlyDest != DestBB) 1009 OnlyDest = MultipleDestSentinel; 1010 1011 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1012 } 1013 1014 // If all edges were unthreadable, we fail. 1015 if (PredToDestList.empty()) 1016 return false; 1017 1018 // Determine which is the most common successor. If we have many inputs and 1019 // this block is a switch, we want to start by threading the batch that goes 1020 // to the most popular destination first. If we only know about one 1021 // threadable destination (the common case) we can avoid this. 1022 BasicBlock *MostPopularDest = OnlyDest; 1023 1024 if (MostPopularDest == MultipleDestSentinel) 1025 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1026 1027 // Now that we know what the most popular destination is, factor all 1028 // predecessors that will jump to it into a single predecessor. 1029 SmallVector<BasicBlock*, 16> PredsToFactor; 1030 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1031 if (PredToDestList[i].second == MostPopularDest) { 1032 BasicBlock *Pred = PredToDestList[i].first; 1033 1034 // This predecessor may be a switch or something else that has multiple 1035 // edges to the block. Factor each of these edges by listing them 1036 // according to # occurrences in PredsToFactor. 1037 TerminatorInst *PredTI = Pred->getTerminator(); 1038 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1039 if (PredTI->getSuccessor(i) == BB) 1040 PredsToFactor.push_back(Pred); 1041 } 1042 1043 // If the threadable edges are branching on an undefined value, we get to pick 1044 // the destination that these predecessors should get to. 1045 if (MostPopularDest == 0) 1046 MostPopularDest = BB->getTerminator()-> 1047 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1048 1049 // Ok, try to thread it! 1050 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1051} 1052 1053/// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in 1054/// the current block. See if there are any simplifications we can do based on 1055/// inputs to the phi node. 1056/// 1057bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 1058 BasicBlock *BB = PN->getParent(); 1059 1060 // If any of the predecessor blocks end in an unconditional branch, we can 1061 // *duplicate* the jump into that block in order to further encourage jump 1062 // threading and to eliminate cases where we have branch on a phi of an icmp 1063 // (branch on icmp is much better). 1064 1065 // We don't want to do this tranformation for switches, because we don't 1066 // really want to duplicate a switch. 1067 if (isa<SwitchInst>(BB->getTerminator())) 1068 return false; 1069 1070 // Look for unconditional branch predecessors. 1071 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1072 BasicBlock *PredBB = PN->getIncomingBlock(i); 1073 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1074 if (PredBr->isUnconditional() && 1075 // Try to duplicate BB into PredBB. 1076 DuplicateCondBranchOnPHIIntoPred(BB, PredBB)) 1077 return true; 1078 } 1079 1080 return false; 1081} 1082 1083 1084/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1085/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1086/// NewPred using the entries from OldPred (suitably mapped). 1087static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1088 BasicBlock *OldPred, 1089 BasicBlock *NewPred, 1090 DenseMap<Instruction*, Value*> &ValueMap) { 1091 for (BasicBlock::iterator PNI = PHIBB->begin(); 1092 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1093 // Ok, we have a PHI node. Figure out what the incoming value was for the 1094 // DestBlock. 1095 Value *IV = PN->getIncomingValueForBlock(OldPred); 1096 1097 // Remap the value if necessary. 1098 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1099 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1100 if (I != ValueMap.end()) 1101 IV = I->second; 1102 } 1103 1104 PN->addIncoming(IV, NewPred); 1105 } 1106} 1107 1108/// ThreadEdge - We have decided that it is safe and profitable to factor the 1109/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1110/// across BB. Transform the IR to reflect this change. 1111bool JumpThreading::ThreadEdge(BasicBlock *BB, 1112 const SmallVectorImpl<BasicBlock*> &PredBBs, 1113 BasicBlock *SuccBB) { 1114 // If threading to the same block as we come from, we would infinite loop. 1115 if (SuccBB == BB) { 1116 DEBUG(errs() << " Not threading across BB '" << BB->getName() 1117 << "' - would thread to self!\n"); 1118 return false; 1119 } 1120 1121 // If threading this would thread across a loop header, don't thread the edge. 1122 // See the comments above FindLoopHeaders for justifications and caveats. 1123 if (LoopHeaders.count(BB)) { 1124 DEBUG(errs() << " Not threading across loop header BB '" << BB->getName() 1125 << "' to dest BB '" << SuccBB->getName() 1126 << "' - it might create an irreducible loop!\n"); 1127 return false; 1128 } 1129 1130 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1131 if (JumpThreadCost > Threshold) { 1132 DEBUG(errs() << " Not threading BB '" << BB->getName() 1133 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1134 return false; 1135 } 1136 1137 // And finally, do it! Start by factoring the predecessors is needed. 1138 BasicBlock *PredBB; 1139 if (PredBBs.size() == 1) 1140 PredBB = PredBBs[0]; 1141 else { 1142 DEBUG(errs() << " Factoring out " << PredBBs.size() 1143 << " common predecessors.\n"); 1144 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1145 ".thr_comm", this); 1146 } 1147 1148 // And finally, do it! 1149 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '" 1150 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1151 << ", across block:\n " 1152 << *BB << "\n"); 1153 1154 // We are going to have to map operands from the original BB block to the new 1155 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1156 // account for entry from PredBB. 1157 DenseMap<Instruction*, Value*> ValueMapping; 1158 1159 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1160 BB->getName()+".thread", 1161 BB->getParent(), BB); 1162 NewBB->moveAfter(PredBB); 1163 1164 BasicBlock::iterator BI = BB->begin(); 1165 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1166 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1167 1168 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1169 // mapping and using it to remap operands in the cloned instructions. 1170 for (; !isa<TerminatorInst>(BI); ++BI) { 1171 Instruction *New = BI->clone(); 1172 New->setName(BI->getName()); 1173 NewBB->getInstList().push_back(New); 1174 ValueMapping[BI] = New; 1175 1176 // Remap operands to patch up intra-block references. 1177 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1178 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1179 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1180 if (I != ValueMapping.end()) 1181 New->setOperand(i, I->second); 1182 } 1183 } 1184 1185 // We didn't copy the terminator from BB over to NewBB, because there is now 1186 // an unconditional jump to SuccBB. Insert the unconditional jump. 1187 BranchInst::Create(SuccBB, NewBB); 1188 1189 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1190 // PHI nodes for NewBB now. 1191 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1192 1193 // If there were values defined in BB that are used outside the block, then we 1194 // now have to update all uses of the value to use either the original value, 1195 // the cloned value, or some PHI derived value. This can require arbitrary 1196 // PHI insertion, of which we are prepared to do, clean these up now. 1197 SSAUpdater SSAUpdate; 1198 SmallVector<Use*, 16> UsesToRename; 1199 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1200 // Scan all uses of this instruction to see if it is used outside of its 1201 // block, and if so, record them in UsesToRename. 1202 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1203 ++UI) { 1204 Instruction *User = cast<Instruction>(*UI); 1205 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1206 if (UserPN->getIncomingBlock(UI) == BB) 1207 continue; 1208 } else if (User->getParent() == BB) 1209 continue; 1210 1211 UsesToRename.push_back(&UI.getUse()); 1212 } 1213 1214 // If there are no uses outside the block, we're done with this instruction. 1215 if (UsesToRename.empty()) 1216 continue; 1217 1218 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1219 1220 // We found a use of I outside of BB. Rename all uses of I that are outside 1221 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1222 // with the two values we know. 1223 SSAUpdate.Initialize(I); 1224 SSAUpdate.AddAvailableValue(BB, I); 1225 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1226 1227 while (!UsesToRename.empty()) 1228 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1229 DEBUG(errs() << "\n"); 1230 } 1231 1232 1233 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1234 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1235 // us to simplify any PHI nodes in BB. 1236 TerminatorInst *PredTerm = PredBB->getTerminator(); 1237 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1238 if (PredTerm->getSuccessor(i) == BB) { 1239 RemovePredecessorAndSimplify(BB, PredBB, TD); 1240 PredTerm->setSuccessor(i, NewBB); 1241 } 1242 1243 // At this point, the IR is fully up to date and consistent. Do a quick scan 1244 // over the new instructions and zap any that are constants or dead. This 1245 // frequently happens because of phi translation. 1246 BI = NewBB->begin(); 1247 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 1248 Instruction *Inst = BI++; 1249 1250 if (Value *V = SimplifyInstruction(Inst, TD)) { 1251 WeakVH BIHandle(BI); 1252 ReplaceAndSimplifyAllUses(Inst, V, TD); 1253 if (BIHandle == 0) 1254 BI = NewBB->begin(); 1255 continue; 1256 } 1257 1258 RecursivelyDeleteTriviallyDeadInstructions(Inst); 1259 } 1260 1261 // Threaded an edge! 1262 ++NumThreads; 1263 return true; 1264} 1265 1266/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1267/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1268/// If we can duplicate the contents of BB up into PredBB do so now, this 1269/// improves the odds that the branch will be on an analyzable instruction like 1270/// a compare. 1271bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1272 BasicBlock *PredBB) { 1273 // If BB is a loop header, then duplicating this block outside the loop would 1274 // cause us to transform this into an irreducible loop, don't do this. 1275 // See the comments above FindLoopHeaders for justifications and caveats. 1276 if (LoopHeaders.count(BB)) { 1277 DEBUG(errs() << " Not duplicating loop header '" << BB->getName() 1278 << "' into predecessor block '" << PredBB->getName() 1279 << "' - it might create an irreducible loop!\n"); 1280 return false; 1281 } 1282 1283 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1284 if (DuplicationCost > Threshold) { 1285 DEBUG(errs() << " Not duplicating BB '" << BB->getName() 1286 << "' - Cost is too high: " << DuplicationCost << "\n"); 1287 return false; 1288 } 1289 1290 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1291 // of PredBB. 1292 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '" 1293 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1294 << DuplicationCost << " block is:" << *BB << "\n"); 1295 1296 // We are going to have to map operands from the original BB block into the 1297 // PredBB block. Evaluate PHI nodes in BB. 1298 DenseMap<Instruction*, Value*> ValueMapping; 1299 1300 BasicBlock::iterator BI = BB->begin(); 1301 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1302 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1303 1304 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1305 1306 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1307 // mapping and using it to remap operands in the cloned instructions. 1308 for (; BI != BB->end(); ++BI) { 1309 Instruction *New = BI->clone(); 1310 New->setName(BI->getName()); 1311 PredBB->getInstList().insert(OldPredBranch, New); 1312 ValueMapping[BI] = New; 1313 1314 // Remap operands to patch up intra-block references. 1315 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1316 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1317 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1318 if (I != ValueMapping.end()) 1319 New->setOperand(i, I->second); 1320 } 1321 } 1322 1323 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1324 // add entries to the PHI nodes for branch from PredBB now. 1325 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1326 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1327 ValueMapping); 1328 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1329 ValueMapping); 1330 1331 // If there were values defined in BB that are used outside the block, then we 1332 // now have to update all uses of the value to use either the original value, 1333 // the cloned value, or some PHI derived value. This can require arbitrary 1334 // PHI insertion, of which we are prepared to do, clean these up now. 1335 SSAUpdater SSAUpdate; 1336 SmallVector<Use*, 16> UsesToRename; 1337 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1338 // Scan all uses of this instruction to see if it is used outside of its 1339 // block, and if so, record them in UsesToRename. 1340 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1341 ++UI) { 1342 Instruction *User = cast<Instruction>(*UI); 1343 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1344 if (UserPN->getIncomingBlock(UI) == BB) 1345 continue; 1346 } else if (User->getParent() == BB) 1347 continue; 1348 1349 UsesToRename.push_back(&UI.getUse()); 1350 } 1351 1352 // If there are no uses outside the block, we're done with this instruction. 1353 if (UsesToRename.empty()) 1354 continue; 1355 1356 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1357 1358 // We found a use of I outside of BB. Rename all uses of I that are outside 1359 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1360 // with the two values we know. 1361 SSAUpdate.Initialize(I); 1362 SSAUpdate.AddAvailableValue(BB, I); 1363 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1364 1365 while (!UsesToRename.empty()) 1366 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1367 DEBUG(errs() << "\n"); 1368 } 1369 1370 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1371 // that we nuked. 1372 RemovePredecessorAndSimplify(BB, PredBB, TD); 1373 1374 // Remove the unconditional branch at the end of the PredBB block. 1375 OldPredBranch->eraseFromParent(); 1376 1377 ++NumDupes; 1378 return true; 1379} 1380 1381 1382