JumpThreading.cpp revision 5660846f15847e540066ae320a4adef7357d597d
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 ProcessJumpOnPHI(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(errs() << "Jump threading on function '" << F.getName() << "'\n"); 122 TD = getAnalysisIfAvailable<TargetData>(); 123 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0; 124 125 FindLoopHeaders(F); 126 127 bool AnotherIteration = true, EverChanged = false; 128 while (AnotherIteration) { 129 AnotherIteration = false; 130 bool Changed = false; 131 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 132 BasicBlock *BB = I; 133 // Thread all of the branches we can over this block. 134 while (ProcessBlock(BB)) 135 Changed = true; 136 137 ++I; 138 139 // If the block is trivially dead, zap it. This eliminates the successor 140 // edges which simplifies the CFG. 141 if (pred_begin(BB) == pred_end(BB) && 142 BB != &BB->getParent()->getEntryBlock()) { 143 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName() 144 << "' with terminator: " << *BB->getTerminator() << '\n'); 145 LoopHeaders.erase(BB); 146 DeleteDeadBlock(BB); 147 Changed = true; 148 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 149 // Can't thread an unconditional jump, but if the block is "almost 150 // empty", we can replace uses of it with uses of the successor and make 151 // this dead. 152 if (BI->isUnconditional() && 153 BB != &BB->getParent()->getEntryBlock()) { 154 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 155 // Ignore dbg intrinsics. 156 while (isa<DbgInfoIntrinsic>(BBI)) 157 ++BBI; 158 // If the terminator is the only non-phi instruction, try to nuke it. 159 if (BBI->isTerminator()) { 160 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 161 // block, we have to make sure it isn't in the LoopHeaders set. We 162 // reinsert afterward if needed. 163 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 164 BasicBlock *Succ = BI->getSuccessor(0); 165 166 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 167 Changed = true; 168 // If we deleted BB and BB was the header of a loop, then the 169 // successor is now the header of the loop. 170 BB = Succ; 171 } 172 173 if (ErasedFromLoopHeaders) 174 LoopHeaders.insert(BB); 175 } 176 } 177 } 178 } 179 AnotherIteration = Changed; 180 EverChanged |= Changed; 181 } 182 183 LoopHeaders.clear(); 184 return EverChanged; 185} 186 187/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 188/// thread across it. 189static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 190 /// Ignore PHI nodes, these will be flattened when duplication happens. 191 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 192 193 // FIXME: THREADING will delete values that are just used to compute the 194 // branch, so they shouldn't count against the duplication cost. 195 196 197 // Sum up the cost of each instruction until we get to the terminator. Don't 198 // include the terminator because the copy won't include it. 199 unsigned Size = 0; 200 for (; !isa<TerminatorInst>(I); ++I) { 201 // Debugger intrinsics don't incur code size. 202 if (isa<DbgInfoIntrinsic>(I)) continue; 203 204 // If this is a pointer->pointer bitcast, it is free. 205 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 206 continue; 207 208 // All other instructions count for at least one unit. 209 ++Size; 210 211 // Calls are more expensive. If they are non-intrinsic calls, we model them 212 // as having cost of 4. If they are a non-vector intrinsic, we model them 213 // as having cost of 2 total, and if they are a vector intrinsic, we model 214 // them as having cost 1. 215 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 216 if (!isa<IntrinsicInst>(CI)) 217 Size += 3; 218 else if (!isa<VectorType>(CI->getType())) 219 Size += 1; 220 } 221 } 222 223 // Threading through a switch statement is particularly profitable. If this 224 // block ends in a switch, decrease its cost to make it more likely to happen. 225 if (isa<SwitchInst>(I)) 226 Size = Size > 6 ? Size-6 : 0; 227 228 return Size; 229} 230 231/// FindLoopHeaders - We do not want jump threading to turn proper loop 232/// structures into irreducible loops. Doing this breaks up the loop nesting 233/// hierarchy and pessimizes later transformations. To prevent this from 234/// happening, we first have to find the loop headers. Here we approximate this 235/// by finding targets of backedges in the CFG. 236/// 237/// Note that there definitely are cases when we want to allow threading of 238/// edges across a loop header. For example, threading a jump from outside the 239/// loop (the preheader) to an exit block of the loop is definitely profitable. 240/// It is also almost always profitable to thread backedges from within the loop 241/// to exit blocks, and is often profitable to thread backedges to other blocks 242/// within the loop (forming a nested loop). This simple analysis is not rich 243/// enough to track all of these properties and keep it up-to-date as the CFG 244/// mutates, so we don't allow any of these transformations. 245/// 246void JumpThreading::FindLoopHeaders(Function &F) { 247 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 248 FindFunctionBackedges(F, Edges); 249 250 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 251 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 252} 253 254/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 255/// if we can infer that the value is a known ConstantInt in any of our 256/// predecessors. If so, return the known list of value and pred BB in the 257/// result vector. If a value is known to be undef, it is returned as null. 258/// 259/// This returns true if there were any known values. 260/// 261bool JumpThreading:: 262ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 263 // If V is a constantint, then it is known in all predecessors. 264 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 265 ConstantInt *CI = dyn_cast<ConstantInt>(V); 266 267 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 268 Result.push_back(std::make_pair(CI, *PI)); 269 return true; 270 } 271 272 // If V is a non-instruction value, or an instruction in a different block, 273 // then it can't be derived from a PHI. 274 Instruction *I = dyn_cast<Instruction>(V); 275 if (I == 0 || I->getParent() != BB) { 276 277 // Okay, if this is a live-in value, see if it has a known value at the end 278 // of any of our predecessors. 279 // 280 // FIXME: This should be an edge property, not a block end property. 281 /// TODO: Per PR2563, we could infer value range information about a 282 /// predecessor based on its terminator. 283 // 284 if (LVI) { 285 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 286 // "I" is a non-local compare-with-a-constant instruction. This would be 287 // able to handle value inequalities better, for example if the compare is 288 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 289 // Perhaps getConstantOnEdge should be smart enough to do this? 290 291 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 292 // If the value is known by LazyValueInfo to be a constant in a 293 // predecessor, use that information to try to thread this block. 294 Constant *PredCst = LVI->getConstantOnEdge(V, *PI, BB); 295 if (PredCst == 0 || 296 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst))) 297 continue; 298 299 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), *PI)); 300 } 301 302 return !Result.empty(); 303 } 304 305 return false; 306 } 307 308 /// If I is a PHI node, then we know the incoming values for any constants. 309 if (PHINode *PN = dyn_cast<PHINode>(I)) { 310 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 311 Value *InVal = PN->getIncomingValue(i); 312 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 313 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 314 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 315 } 316 } 317 return !Result.empty(); 318 } 319 320 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 321 322 // Handle some boolean conditions. 323 if (I->getType()->getPrimitiveSizeInBits() == 1) { 324 // X | true -> true 325 // X & false -> false 326 if (I->getOpcode() == Instruction::Or || 327 I->getOpcode() == Instruction::And) { 328 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 329 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 330 331 if (LHSVals.empty() && RHSVals.empty()) 332 return false; 333 334 ConstantInt *InterestingVal; 335 if (I->getOpcode() == Instruction::Or) 336 InterestingVal = ConstantInt::getTrue(I->getContext()); 337 else 338 InterestingVal = ConstantInt::getFalse(I->getContext()); 339 340 // Scan for the sentinel. 341 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 342 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) 343 Result.push_back(LHSVals[i]); 344 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 345 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) 346 Result.push_back(RHSVals[i]); 347 return !Result.empty(); 348 } 349 350 // Handle the NOT form of XOR. 351 if (I->getOpcode() == Instruction::Xor && 352 isa<ConstantInt>(I->getOperand(1)) && 353 cast<ConstantInt>(I->getOperand(1))->isOne()) { 354 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result); 355 if (Result.empty()) 356 return false; 357 358 // Invert the known values. 359 for (unsigned i = 0, e = Result.size(); i != e; ++i) 360 if (Result[i].first) 361 Result[i].first = 362 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first)); 363 return true; 364 } 365 } 366 367 // Handle compare with phi operand, where the PHI is defined in this block. 368 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 369 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 370 if (PN && PN->getParent() == BB) { 371 // We can do this simplification if any comparisons fold to true or false. 372 // See if any do. 373 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 374 BasicBlock *PredBB = PN->getIncomingBlock(i); 375 Value *LHS = PN->getIncomingValue(i); 376 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 377 378 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 379 if (Res == 0) { 380 if (!LVI || !isa<Constant>(RHS)) 381 continue; 382 383 LazyValueInfo::Tristate 384 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 385 cast<Constant>(RHS), PredBB, BB); 386 if (ResT == LazyValueInfo::Unknown) 387 continue; 388 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 389 } 390 391 if (isa<UndefValue>(Res)) 392 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 393 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 394 Result.push_back(std::make_pair(CI, PredBB)); 395 } 396 397 return !Result.empty(); 398 } 399 400 401 // If comparing a live-in value against a constant, see if we know the 402 // live-in value on any predecessors. 403 if (LVI && isa<Constant>(Cmp->getOperand(1)) && 404 Cmp->getType()->isInteger() && // Not vector compare. 405 (!isa<Instruction>(Cmp->getOperand(0)) || 406 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) { 407 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 408 409 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 410 // If the value is known by LazyValueInfo to be a constant in a 411 // predecessor, use that information to try to thread this block. 412 LazyValueInfo::Tristate 413 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 414 RHSCst, *PI, BB); 415 if (Res == LazyValueInfo::Unknown) 416 continue; 417 418 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 419 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), *PI)); 420 } 421 422 return !Result.empty(); 423 } 424 } 425 return false; 426} 427 428 429 430/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 431/// in an undefined jump, decide which block is best to revector to. 432/// 433/// Since we can pick an arbitrary destination, we pick the successor with the 434/// fewest predecessors. This should reduce the in-degree of the others. 435/// 436static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 437 TerminatorInst *BBTerm = BB->getTerminator(); 438 unsigned MinSucc = 0; 439 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 440 // Compute the successor with the minimum number of predecessors. 441 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 442 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 443 TestBB = BBTerm->getSuccessor(i); 444 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 445 if (NumPreds < MinNumPreds) 446 MinSucc = i; 447 } 448 449 return MinSucc; 450} 451 452/// ProcessBlock - If there are any predecessors whose control can be threaded 453/// through to a successor, transform them now. 454bool JumpThreading::ProcessBlock(BasicBlock *BB) { 455 // If this block has a single predecessor, and if that pred has a single 456 // successor, merge the blocks. This encourages recursive jump threading 457 // because now the condition in this block can be threaded through 458 // predecessors of our predecessor block. 459 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 460 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 461 SinglePred != BB) { 462 // If SinglePred was a loop header, BB becomes one. 463 if (LoopHeaders.erase(SinglePred)) 464 LoopHeaders.insert(BB); 465 466 // Remember if SinglePred was the entry block of the function. If so, we 467 // will need to move BB back to the entry position. 468 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 469 MergeBasicBlockIntoOnlyPred(BB); 470 471 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 472 BB->moveBefore(&BB->getParent()->getEntryBlock()); 473 return true; 474 } 475 } 476 477 // Look to see if the terminator is a branch of switch, if not we can't thread 478 // it. 479 Value *Condition; 480 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 481 // Can't thread an unconditional jump. 482 if (BI->isUnconditional()) return false; 483 Condition = BI->getCondition(); 484 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 485 Condition = SI->getCondition(); 486 else 487 return false; // Must be an invoke. 488 489 // If the terminator of this block is branching on a constant, simplify the 490 // terminator to an unconditional branch. This can occur due to threading in 491 // other blocks. 492 if (isa<ConstantInt>(Condition)) { 493 DEBUG(errs() << " In block '" << BB->getName() 494 << "' folding terminator: " << *BB->getTerminator() << '\n'); 495 ++NumFolds; 496 ConstantFoldTerminator(BB); 497 return true; 498 } 499 500 // If the terminator is branching on an undef, we can pick any of the 501 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 502 if (isa<UndefValue>(Condition)) { 503 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 504 505 // Fold the branch/switch. 506 TerminatorInst *BBTerm = BB->getTerminator(); 507 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 508 if (i == BestSucc) continue; 509 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 510 } 511 512 DEBUG(errs() << " In block '" << BB->getName() 513 << "' folding undef terminator: " << *BBTerm << '\n'); 514 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 515 BBTerm->eraseFromParent(); 516 return true; 517 } 518 519 Instruction *CondInst = dyn_cast<Instruction>(Condition); 520 521 // If the condition is an instruction defined in another block, see if a 522 // predecessor has the same condition: 523 // br COND, BBX, BBY 524 // BBX: 525 // br COND, BBZ, BBW 526 if (!LVI && 527 !Condition->hasOneUse() && // Multiple uses. 528 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 529 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 530 if (isa<BranchInst>(BB->getTerminator())) { 531 for (; PI != E; ++PI) 532 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 533 if (PBI->isConditional() && PBI->getCondition() == Condition && 534 ProcessBranchOnDuplicateCond(*PI, BB)) 535 return true; 536 } else { 537 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 538 for (; PI != E; ++PI) 539 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 540 if (PSI->getCondition() == Condition && 541 ProcessSwitchOnDuplicateCond(*PI, BB)) 542 return true; 543 } 544 } 545 546 // All the rest of our checks depend on the condition being an instruction. 547 if (CondInst == 0) { 548 // FIXME: Unify this with code below. 549 if (LVI && ProcessThreadableEdges(Condition, BB)) 550 return true; 551 return false; 552 } 553 554 555 // See if this is a phi node in the current block. 556 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 557 if (PN->getParent() == BB) 558 return ProcessJumpOnPHI(PN); 559 560 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 561 if (!LVI && 562 (!isa<PHINode>(CondCmp->getOperand(0)) || 563 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) { 564 // If we have a comparison, loop over the predecessors to see if there is 565 // a condition with a lexically identical value. 566 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 567 for (; PI != E; ++PI) 568 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 569 if (PBI->isConditional() && *PI != BB) { 570 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 571 if (CI->getOperand(0) == CondCmp->getOperand(0) && 572 CI->getOperand(1) == CondCmp->getOperand(1) && 573 CI->getPredicate() == CondCmp->getPredicate()) { 574 // TODO: Could handle things like (x != 4) --> (x == 17) 575 if (ProcessBranchOnDuplicateCond(*PI, BB)) 576 return true; 577 } 578 } 579 } 580 } 581 } 582 583 // Check for some cases that are worth simplifying. Right now we want to look 584 // for loads that are used by a switch or by the condition for the branch. If 585 // we see one, check to see if it's partially redundant. If so, insert a PHI 586 // which can then be used to thread the values. 587 // 588 // This is particularly important because reg2mem inserts loads and stores all 589 // over the place, and this blocks jump threading if we don't zap them. 590 Value *SimplifyValue = CondInst; 591 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 592 if (isa<Constant>(CondCmp->getOperand(1))) 593 SimplifyValue = CondCmp->getOperand(0); 594 595 // TODO: There are other places where load PRE would be profitable, such as 596 // more complex comparisons. 597 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 598 if (SimplifyPartiallyRedundantLoad(LI)) 599 return true; 600 601 602 // Handle a variety of cases where we are branching on something derived from 603 // a PHI node in the current block. If we can prove that any predecessors 604 // compute a predictable value based on a PHI node, thread those predecessors. 605 // 606 if (ProcessThreadableEdges(CondInst, BB)) 607 return true; 608 609 610 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 611 // "(X == 4)" thread through this block. 612 613 return false; 614} 615 616/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 617/// block that jump on exactly the same condition. This means that we almost 618/// always know the direction of the edge in the DESTBB: 619/// PREDBB: 620/// br COND, DESTBB, BBY 621/// DESTBB: 622/// br COND, BBZ, BBW 623/// 624/// If DESTBB has multiple predecessors, we can't just constant fold the branch 625/// in DESTBB, we have to thread over it. 626bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 627 BasicBlock *BB) { 628 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 629 630 // If both successors of PredBB go to DESTBB, we don't know anything. We can 631 // fold the branch to an unconditional one, which allows other recursive 632 // simplifications. 633 bool BranchDir; 634 if (PredBI->getSuccessor(1) != BB) 635 BranchDir = true; 636 else if (PredBI->getSuccessor(0) != BB) 637 BranchDir = false; 638 else { 639 DEBUG(errs() << " In block '" << PredBB->getName() 640 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 641 ++NumFolds; 642 ConstantFoldTerminator(PredBB); 643 return true; 644 } 645 646 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 647 648 // If the dest block has one predecessor, just fix the branch condition to a 649 // constant and fold it. 650 if (BB->getSinglePredecessor()) { 651 DEBUG(errs() << " In block '" << BB->getName() 652 << "' folding condition to '" << BranchDir << "': " 653 << *BB->getTerminator() << '\n'); 654 ++NumFolds; 655 Value *OldCond = DestBI->getCondition(); 656 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 657 BranchDir)); 658 ConstantFoldTerminator(BB); 659 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 660 return true; 661 } 662 663 664 // Next, figure out which successor we are threading to. 665 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 666 667 SmallVector<BasicBlock*, 2> Preds; 668 Preds.push_back(PredBB); 669 670 // Ok, try to thread it! 671 return ThreadEdge(BB, Preds, SuccBB); 672} 673 674/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 675/// block that switch on exactly the same condition. This means that we almost 676/// always know the direction of the edge in the DESTBB: 677/// PREDBB: 678/// switch COND [... DESTBB, BBY ... ] 679/// DESTBB: 680/// switch COND [... BBZ, BBW ] 681/// 682/// Optimizing switches like this is very important, because simplifycfg builds 683/// switches out of repeated 'if' conditions. 684bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 685 BasicBlock *DestBB) { 686 // Can't thread edge to self. 687 if (PredBB == DestBB) 688 return false; 689 690 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 691 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 692 693 // There are a variety of optimizations that we can potentially do on these 694 // blocks: we order them from most to least preferable. 695 696 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 697 // directly to their destination. This does not introduce *any* code size 698 // growth. Skip debug info first. 699 BasicBlock::iterator BBI = DestBB->begin(); 700 while (isa<DbgInfoIntrinsic>(BBI)) 701 BBI++; 702 703 // FIXME: Thread if it just contains a PHI. 704 if (isa<SwitchInst>(BBI)) { 705 bool MadeChange = false; 706 // Ignore the default edge for now. 707 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 708 ConstantInt *DestVal = DestSI->getCaseValue(i); 709 BasicBlock *DestSucc = DestSI->getSuccessor(i); 710 711 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 712 // PredSI has an explicit case for it. If so, forward. If it is covered 713 // by the default case, we can't update PredSI. 714 unsigned PredCase = PredSI->findCaseValue(DestVal); 715 if (PredCase == 0) continue; 716 717 // If PredSI doesn't go to DestBB on this value, then it won't reach the 718 // case on this condition. 719 if (PredSI->getSuccessor(PredCase) != DestBB && 720 DestSI->getSuccessor(i) != DestBB) 721 continue; 722 723 // Do not forward this if it already goes to this destination, this would 724 // be an infinite loop. 725 if (PredSI->getSuccessor(PredCase) == DestSucc) 726 continue; 727 728 // Otherwise, we're safe to make the change. Make sure that the edge from 729 // DestSI to DestSucc is not critical and has no PHI nodes. 730 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 731 DEBUG(errs() << "THROUGH: " << *DestSI); 732 733 // If the destination has PHI nodes, just split the edge for updating 734 // simplicity. 735 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 736 SplitCriticalEdge(DestSI, i, this); 737 DestSucc = DestSI->getSuccessor(i); 738 } 739 FoldSingleEntryPHINodes(DestSucc); 740 PredSI->setSuccessor(PredCase, DestSucc); 741 MadeChange = true; 742 } 743 744 if (MadeChange) 745 return true; 746 } 747 748 return false; 749} 750 751 752/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 753/// load instruction, eliminate it by replacing it with a PHI node. This is an 754/// important optimization that encourages jump threading, and needs to be run 755/// interlaced with other jump threading tasks. 756bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 757 // Don't hack volatile loads. 758 if (LI->isVolatile()) return false; 759 760 // If the load is defined in a block with exactly one predecessor, it can't be 761 // partially redundant. 762 BasicBlock *LoadBB = LI->getParent(); 763 if (LoadBB->getSinglePredecessor()) 764 return false; 765 766 Value *LoadedPtr = LI->getOperand(0); 767 768 // If the loaded operand is defined in the LoadBB, it can't be available. 769 // TODO: Could do simple PHI translation, that would be fun :) 770 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 771 if (PtrOp->getParent() == LoadBB) 772 return false; 773 774 // Scan a few instructions up from the load, to see if it is obviously live at 775 // the entry to its block. 776 BasicBlock::iterator BBIt = LI; 777 778 if (Value *AvailableVal = 779 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 780 // If the value if the load is locally available within the block, just use 781 // it. This frequently occurs for reg2mem'd allocas. 782 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 783 784 // If the returned value is the load itself, replace with an undef. This can 785 // only happen in dead loops. 786 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 787 LI->replaceAllUsesWith(AvailableVal); 788 LI->eraseFromParent(); 789 return true; 790 } 791 792 // Otherwise, if we scanned the whole block and got to the top of the block, 793 // we know the block is locally transparent to the load. If not, something 794 // might clobber its value. 795 if (BBIt != LoadBB->begin()) 796 return false; 797 798 799 SmallPtrSet<BasicBlock*, 8> PredsScanned; 800 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 801 AvailablePredsTy AvailablePreds; 802 BasicBlock *OneUnavailablePred = 0; 803 804 // If we got here, the loaded value is transparent through to the start of the 805 // block. Check to see if it is available in any of the predecessor blocks. 806 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 807 PI != PE; ++PI) { 808 BasicBlock *PredBB = *PI; 809 810 // If we already scanned this predecessor, skip it. 811 if (!PredsScanned.insert(PredBB)) 812 continue; 813 814 // Scan the predecessor to see if the value is available in the pred. 815 BBIt = PredBB->end(); 816 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 817 if (!PredAvailable) { 818 OneUnavailablePred = PredBB; 819 continue; 820 } 821 822 // If so, this load is partially redundant. Remember this info so that we 823 // can create a PHI node. 824 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 825 } 826 827 // If the loaded value isn't available in any predecessor, it isn't partially 828 // redundant. 829 if (AvailablePreds.empty()) return false; 830 831 // Okay, the loaded value is available in at least one (and maybe all!) 832 // predecessors. If the value is unavailable in more than one unique 833 // predecessor, we want to insert a merge block for those common predecessors. 834 // This ensures that we only have to insert one reload, thus not increasing 835 // code size. 836 BasicBlock *UnavailablePred = 0; 837 838 // If there is exactly one predecessor where the value is unavailable, the 839 // already computed 'OneUnavailablePred' block is it. If it ends in an 840 // unconditional branch, we know that it isn't a critical edge. 841 if (PredsScanned.size() == AvailablePreds.size()+1 && 842 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 843 UnavailablePred = OneUnavailablePred; 844 } else if (PredsScanned.size() != AvailablePreds.size()) { 845 // Otherwise, we had multiple unavailable predecessors or we had a critical 846 // edge from the one. 847 SmallVector<BasicBlock*, 8> PredsToSplit; 848 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 849 850 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 851 AvailablePredSet.insert(AvailablePreds[i].first); 852 853 // Add all the unavailable predecessors to the PredsToSplit list. 854 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 855 PI != PE; ++PI) 856 if (!AvailablePredSet.count(*PI)) 857 PredsToSplit.push_back(*PI); 858 859 // Split them out to their own block. 860 UnavailablePred = 861 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 862 "thread-pre-split", this); 863 } 864 865 // If the value isn't available in all predecessors, then there will be 866 // exactly one where it isn't available. Insert a load on that edge and add 867 // it to the AvailablePreds list. 868 if (UnavailablePred) { 869 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 870 "Can't handle critical edge here!"); 871 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 872 LI->getAlignment(), 873 UnavailablePred->getTerminator()); 874 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 875 } 876 877 // Now we know that each predecessor of this block has a value in 878 // AvailablePreds, sort them for efficient access as we're walking the preds. 879 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 880 881 // Create a PHI node at the start of the block for the PRE'd load value. 882 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 883 PN->takeName(LI); 884 885 // Insert new entries into the PHI for each predecessor. A single block may 886 // have multiple entries here. 887 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 888 ++PI) { 889 AvailablePredsTy::iterator I = 890 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 891 std::make_pair(*PI, (Value*)0)); 892 893 assert(I != AvailablePreds.end() && I->first == *PI && 894 "Didn't find entry for predecessor!"); 895 896 PN->addIncoming(I->second, I->first); 897 } 898 899 //cerr << "PRE: " << *LI << *PN << "\n"; 900 901 LI->replaceAllUsesWith(PN); 902 LI->eraseFromParent(); 903 904 return true; 905} 906 907/// FindMostPopularDest - The specified list contains multiple possible 908/// threadable destinations. Pick the one that occurs the most frequently in 909/// the list. 910static BasicBlock * 911FindMostPopularDest(BasicBlock *BB, 912 const SmallVectorImpl<std::pair<BasicBlock*, 913 BasicBlock*> > &PredToDestList) { 914 assert(!PredToDestList.empty()); 915 916 // Determine popularity. If there are multiple possible destinations, we 917 // explicitly choose to ignore 'undef' destinations. We prefer to thread 918 // blocks with known and real destinations to threading undef. We'll handle 919 // them later if interesting. 920 DenseMap<BasicBlock*, unsigned> DestPopularity; 921 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 922 if (PredToDestList[i].second) 923 DestPopularity[PredToDestList[i].second]++; 924 925 // Find the most popular dest. 926 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 927 BasicBlock *MostPopularDest = DPI->first; 928 unsigned Popularity = DPI->second; 929 SmallVector<BasicBlock*, 4> SamePopularity; 930 931 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 932 // If the popularity of this entry isn't higher than the popularity we've 933 // seen so far, ignore it. 934 if (DPI->second < Popularity) 935 ; // ignore. 936 else if (DPI->second == Popularity) { 937 // If it is the same as what we've seen so far, keep track of it. 938 SamePopularity.push_back(DPI->first); 939 } else { 940 // If it is more popular, remember it. 941 SamePopularity.clear(); 942 MostPopularDest = DPI->first; 943 Popularity = DPI->second; 944 } 945 } 946 947 // Okay, now we know the most popular destination. If there is more than 948 // destination, we need to determine one. This is arbitrary, but we need 949 // to make a deterministic decision. Pick the first one that appears in the 950 // successor list. 951 if (!SamePopularity.empty()) { 952 SamePopularity.push_back(MostPopularDest); 953 TerminatorInst *TI = BB->getTerminator(); 954 for (unsigned i = 0; ; ++i) { 955 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 956 957 if (std::find(SamePopularity.begin(), SamePopularity.end(), 958 TI->getSuccessor(i)) == SamePopularity.end()) 959 continue; 960 961 MostPopularDest = TI->getSuccessor(i); 962 break; 963 } 964 } 965 966 // Okay, we have finally picked the most popular destination. 967 return MostPopularDest; 968} 969 970bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) { 971 // If threading this would thread across a loop header, don't even try to 972 // thread the edge. 973 if (LoopHeaders.count(BB)) 974 return false; 975 976 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 977 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) 978 return false; 979 assert(!PredValues.empty() && 980 "ComputeValueKnownInPredecessors returned true with no values"); 981 982 DEBUG(errs() << "IN BB: " << *BB; 983 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 984 errs() << " BB '" << BB->getName() << "': FOUND condition = "; 985 if (PredValues[i].first) 986 errs() << *PredValues[i].first; 987 else 988 errs() << "UNDEF"; 989 errs() << " for pred '" << PredValues[i].second->getName() 990 << "'.\n"; 991 }); 992 993 // Decide what we want to thread through. Convert our list of known values to 994 // a list of known destinations for each pred. This also discards duplicate 995 // predecessors and keeps track of the undefined inputs (which are represented 996 // as a null dest in the PredToDestList). 997 SmallPtrSet<BasicBlock*, 16> SeenPreds; 998 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 999 1000 BasicBlock *OnlyDest = 0; 1001 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1002 1003 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1004 BasicBlock *Pred = PredValues[i].second; 1005 if (!SeenPreds.insert(Pred)) 1006 continue; // Duplicate predecessor entry. 1007 1008 // If the predecessor ends with an indirect goto, we can't change its 1009 // destination. 1010 if (isa<IndirectBrInst>(Pred->getTerminator())) 1011 continue; 1012 1013 ConstantInt *Val = PredValues[i].first; 1014 1015 BasicBlock *DestBB; 1016 if (Val == 0) // Undef. 1017 DestBB = 0; 1018 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1019 DestBB = BI->getSuccessor(Val->isZero()); 1020 else { 1021 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1022 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1023 } 1024 1025 // If we have exactly one destination, remember it for efficiency below. 1026 if (i == 0) 1027 OnlyDest = DestBB; 1028 else if (OnlyDest != DestBB) 1029 OnlyDest = MultipleDestSentinel; 1030 1031 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1032 } 1033 1034 // If all edges were unthreadable, we fail. 1035 if (PredToDestList.empty()) 1036 return false; 1037 1038 // Determine which is the most common successor. If we have many inputs and 1039 // this block is a switch, we want to start by threading the batch that goes 1040 // to the most popular destination first. If we only know about one 1041 // threadable destination (the common case) we can avoid this. 1042 BasicBlock *MostPopularDest = OnlyDest; 1043 1044 if (MostPopularDest == MultipleDestSentinel) 1045 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1046 1047 // Now that we know what the most popular destination is, factor all 1048 // predecessors that will jump to it into a single predecessor. 1049 SmallVector<BasicBlock*, 16> PredsToFactor; 1050 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1051 if (PredToDestList[i].second == MostPopularDest) { 1052 BasicBlock *Pred = PredToDestList[i].first; 1053 1054 // This predecessor may be a switch or something else that has multiple 1055 // edges to the block. Factor each of these edges by listing them 1056 // according to # occurrences in PredsToFactor. 1057 TerminatorInst *PredTI = Pred->getTerminator(); 1058 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1059 if (PredTI->getSuccessor(i) == BB) 1060 PredsToFactor.push_back(Pred); 1061 } 1062 1063 // If the threadable edges are branching on an undefined value, we get to pick 1064 // the destination that these predecessors should get to. 1065 if (MostPopularDest == 0) 1066 MostPopularDest = BB->getTerminator()-> 1067 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1068 1069 // Ok, try to thread it! 1070 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1071} 1072 1073/// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in 1074/// the current block. See if there are any simplifications we can do based on 1075/// inputs to the phi node. 1076/// 1077bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 1078 BasicBlock *BB = PN->getParent(); 1079 1080 // If any of the predecessor blocks end in an unconditional branch, we can 1081 // *duplicate* the jump into that block in order to further encourage jump 1082 // threading and to eliminate cases where we have branch on a phi of an icmp 1083 // (branch on icmp is much better). 1084 1085 // We don't want to do this tranformation for switches, because we don't 1086 // really want to duplicate a switch. 1087 if (isa<SwitchInst>(BB->getTerminator())) 1088 return false; 1089 1090 // Look for unconditional branch predecessors. 1091 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1092 BasicBlock *PredBB = PN->getIncomingBlock(i); 1093 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1094 if (PredBr->isUnconditional() && 1095 // Try to duplicate BB into PredBB. 1096 DuplicateCondBranchOnPHIIntoPred(BB, PredBB)) 1097 return true; 1098 } 1099 1100 return false; 1101} 1102 1103 1104/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1105/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1106/// NewPred using the entries from OldPred (suitably mapped). 1107static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1108 BasicBlock *OldPred, 1109 BasicBlock *NewPred, 1110 DenseMap<Instruction*, Value*> &ValueMap) { 1111 for (BasicBlock::iterator PNI = PHIBB->begin(); 1112 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1113 // Ok, we have a PHI node. Figure out what the incoming value was for the 1114 // DestBlock. 1115 Value *IV = PN->getIncomingValueForBlock(OldPred); 1116 1117 // Remap the value if necessary. 1118 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1119 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1120 if (I != ValueMap.end()) 1121 IV = I->second; 1122 } 1123 1124 PN->addIncoming(IV, NewPred); 1125 } 1126} 1127 1128/// ThreadEdge - We have decided that it is safe and profitable to factor the 1129/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1130/// across BB. Transform the IR to reflect this change. 1131bool JumpThreading::ThreadEdge(BasicBlock *BB, 1132 const SmallVectorImpl<BasicBlock*> &PredBBs, 1133 BasicBlock *SuccBB) { 1134 // If threading to the same block as we come from, we would infinite loop. 1135 if (SuccBB == BB) { 1136 DEBUG(errs() << " Not threading across BB '" << BB->getName() 1137 << "' - would thread to self!\n"); 1138 return false; 1139 } 1140 1141 // If threading this would thread across a loop header, don't thread the edge. 1142 // See the comments above FindLoopHeaders for justifications and caveats. 1143 if (LoopHeaders.count(BB)) { 1144 DEBUG(errs() << " Not threading across loop header BB '" << BB->getName() 1145 << "' to dest BB '" << SuccBB->getName() 1146 << "' - it might create an irreducible loop!\n"); 1147 return false; 1148 } 1149 1150 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1151 if (JumpThreadCost > Threshold) { 1152 DEBUG(errs() << " Not threading BB '" << BB->getName() 1153 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1154 return false; 1155 } 1156 1157 // And finally, do it! Start by factoring the predecessors is needed. 1158 BasicBlock *PredBB; 1159 if (PredBBs.size() == 1) 1160 PredBB = PredBBs[0]; 1161 else { 1162 DEBUG(errs() << " Factoring out " << PredBBs.size() 1163 << " common predecessors.\n"); 1164 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1165 ".thr_comm", this); 1166 } 1167 1168 // And finally, do it! 1169 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '" 1170 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1171 << ", across block:\n " 1172 << *BB << "\n"); 1173 1174 // We are going to have to map operands from the original BB block to the new 1175 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1176 // account for entry from PredBB. 1177 DenseMap<Instruction*, Value*> ValueMapping; 1178 1179 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1180 BB->getName()+".thread", 1181 BB->getParent(), BB); 1182 NewBB->moveAfter(PredBB); 1183 1184 BasicBlock::iterator BI = BB->begin(); 1185 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1186 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1187 1188 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1189 // mapping and using it to remap operands in the cloned instructions. 1190 for (; !isa<TerminatorInst>(BI); ++BI) { 1191 Instruction *New = BI->clone(); 1192 New->setName(BI->getName()); 1193 NewBB->getInstList().push_back(New); 1194 ValueMapping[BI] = New; 1195 1196 // Remap operands to patch up intra-block references. 1197 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1198 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1199 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1200 if (I != ValueMapping.end()) 1201 New->setOperand(i, I->second); 1202 } 1203 } 1204 1205 // We didn't copy the terminator from BB over to NewBB, because there is now 1206 // an unconditional jump to SuccBB. Insert the unconditional jump. 1207 BranchInst::Create(SuccBB, NewBB); 1208 1209 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1210 // PHI nodes for NewBB now. 1211 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1212 1213 // If there were values defined in BB that are used outside the block, then we 1214 // now have to update all uses of the value to use either the original value, 1215 // the cloned value, or some PHI derived value. This can require arbitrary 1216 // PHI insertion, of which we are prepared to do, clean these up now. 1217 SSAUpdater SSAUpdate; 1218 SmallVector<Use*, 16> UsesToRename; 1219 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1220 // Scan all uses of this instruction to see if it is used outside of its 1221 // block, and if so, record them in UsesToRename. 1222 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1223 ++UI) { 1224 Instruction *User = cast<Instruction>(*UI); 1225 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1226 if (UserPN->getIncomingBlock(UI) == BB) 1227 continue; 1228 } else if (User->getParent() == BB) 1229 continue; 1230 1231 UsesToRename.push_back(&UI.getUse()); 1232 } 1233 1234 // If there are no uses outside the block, we're done with this instruction. 1235 if (UsesToRename.empty()) 1236 continue; 1237 1238 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1239 1240 // We found a use of I outside of BB. Rename all uses of I that are outside 1241 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1242 // with the two values we know. 1243 SSAUpdate.Initialize(I); 1244 SSAUpdate.AddAvailableValue(BB, I); 1245 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1246 1247 while (!UsesToRename.empty()) 1248 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1249 DEBUG(errs() << "\n"); 1250 } 1251 1252 1253 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1254 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1255 // us to simplify any PHI nodes in BB. 1256 TerminatorInst *PredTerm = PredBB->getTerminator(); 1257 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1258 if (PredTerm->getSuccessor(i) == BB) { 1259 RemovePredecessorAndSimplify(BB, PredBB, TD); 1260 PredTerm->setSuccessor(i, NewBB); 1261 } 1262 1263 // At this point, the IR is fully up to date and consistent. Do a quick scan 1264 // over the new instructions and zap any that are constants or dead. This 1265 // frequently happens because of phi translation. 1266 BI = NewBB->begin(); 1267 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 1268 Instruction *Inst = BI++; 1269 1270 if (Value *V = SimplifyInstruction(Inst, TD)) { 1271 WeakVH BIHandle(BI); 1272 ReplaceAndSimplifyAllUses(Inst, V, TD); 1273 if (BIHandle == 0) 1274 BI = NewBB->begin(); 1275 continue; 1276 } 1277 1278 RecursivelyDeleteTriviallyDeadInstructions(Inst); 1279 } 1280 1281 // Threaded an edge! 1282 ++NumThreads; 1283 return true; 1284} 1285 1286/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1287/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1288/// If we can duplicate the contents of BB up into PredBB do so now, this 1289/// improves the odds that the branch will be on an analyzable instruction like 1290/// a compare. 1291bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1292 BasicBlock *PredBB) { 1293 // If BB is a loop header, then duplicating this block outside the loop would 1294 // cause us to transform this into an irreducible loop, don't do this. 1295 // See the comments above FindLoopHeaders for justifications and caveats. 1296 if (LoopHeaders.count(BB)) { 1297 DEBUG(errs() << " Not duplicating loop header '" << BB->getName() 1298 << "' into predecessor block '" << PredBB->getName() 1299 << "' - it might create an irreducible loop!\n"); 1300 return false; 1301 } 1302 1303 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1304 if (DuplicationCost > Threshold) { 1305 DEBUG(errs() << " Not duplicating BB '" << BB->getName() 1306 << "' - Cost is too high: " << DuplicationCost << "\n"); 1307 return false; 1308 } 1309 1310 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1311 // of PredBB. 1312 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '" 1313 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1314 << DuplicationCost << " block is:" << *BB << "\n"); 1315 1316 // We are going to have to map operands from the original BB block into the 1317 // PredBB block. Evaluate PHI nodes in BB. 1318 DenseMap<Instruction*, Value*> ValueMapping; 1319 1320 BasicBlock::iterator BI = BB->begin(); 1321 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1322 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1323 1324 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1325 1326 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1327 // mapping and using it to remap operands in the cloned instructions. 1328 for (; BI != BB->end(); ++BI) { 1329 Instruction *New = BI->clone(); 1330 New->setName(BI->getName()); 1331 PredBB->getInstList().insert(OldPredBranch, New); 1332 ValueMapping[BI] = New; 1333 1334 // Remap operands to patch up intra-block references. 1335 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1336 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1337 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1338 if (I != ValueMapping.end()) 1339 New->setOperand(i, I->second); 1340 } 1341 } 1342 1343 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1344 // add entries to the PHI nodes for branch from PredBB now. 1345 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1346 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1347 ValueMapping); 1348 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1349 ValueMapping); 1350 1351 // If there were values defined in BB that are used outside the block, then we 1352 // now have to update all uses of the value to use either the original value, 1353 // the cloned value, or some PHI derived value. This can require arbitrary 1354 // PHI insertion, of which we are prepared to do, clean these up now. 1355 SSAUpdater SSAUpdate; 1356 SmallVector<Use*, 16> UsesToRename; 1357 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1358 // Scan all uses of this instruction to see if it is used outside of its 1359 // block, and if so, record them in UsesToRename. 1360 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1361 ++UI) { 1362 Instruction *User = cast<Instruction>(*UI); 1363 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1364 if (UserPN->getIncomingBlock(UI) == BB) 1365 continue; 1366 } else if (User->getParent() == BB) 1367 continue; 1368 1369 UsesToRename.push_back(&UI.getUse()); 1370 } 1371 1372 // If there are no uses outside the block, we're done with this instruction. 1373 if (UsesToRename.empty()) 1374 continue; 1375 1376 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1377 1378 // We found a use of I outside of BB. Rename all uses of I that are outside 1379 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1380 // with the two values we know. 1381 SSAUpdate.Initialize(I); 1382 SSAUpdate.AddAvailableValue(BB, I); 1383 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1384 1385 while (!UsesToRename.empty()) 1386 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1387 DEBUG(errs() << "\n"); 1388 } 1389 1390 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1391 // that we nuked. 1392 RemovePredecessorAndSimplify(BB, PredBB, TD); 1393 1394 // Remove the unconditional branch at the end of the PredBB block. 1395 OldPredBranch->eraseFromParent(); 1396 1397 ++NumDupes; 1398 return true; 1399} 1400 1401 1402