1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the Jump Threading pass. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #define DEBUG_TYPE "jump-threading" 15 #include "llvm/Transforms/Scalar.h" 16 #include "llvm/IntrinsicInst.h" 17 #include "llvm/LLVMContext.h" 18 #include "llvm/Pass.h" 19 #include "llvm/Analysis/InstructionSimplify.h" 20 #include "llvm/Analysis/LazyValueInfo.h" 21 #include "llvm/Analysis/Loads.h" 22 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 23 #include "llvm/Transforms/Utils/Local.h" 24 #include "llvm/Transforms/Utils/SSAUpdater.h" 25 #include "llvm/Target/TargetData.h" 26 #include "llvm/ADT/DenseMap.h" 27 #include "llvm/ADT/Statistic.h" 28 #include "llvm/ADT/STLExtras.h" 29 #include "llvm/ADT/SmallPtrSet.h" 30 #include "llvm/ADT/SmallSet.h" 31 #include "llvm/Support/CommandLine.h" 32 #include "llvm/Support/Debug.h" 33 #include "llvm/Support/ValueHandle.h" 34 #include "llvm/Support/raw_ostream.h" 35 using namespace llvm; 36 37 STATISTIC(NumThreads, "Number of jumps threaded"); 38 STATISTIC(NumFolds, "Number of terminators folded"); 39 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 40 41 static cl::opt<unsigned> 42 Threshold("jump-threading-threshold", 43 cl::desc("Max block size to duplicate for jump threading"), 44 cl::init(6), cl::Hidden); 45 46 // Turn on use of LazyValueInfo. 47 static cl::opt<bool> 48 EnableLVI("enable-jump-threading-lvi", 49 cl::desc("Use LVI for jump threading"), 50 cl::init(true), 51 cl::ReallyHidden); 52 53 54 55 namespace { 56 /// This pass performs 'jump threading', which looks at blocks that have 57 /// multiple predecessors and multiple successors. If one or more of the 58 /// predecessors of the block can be proven to always jump to one of the 59 /// successors, we forward the edge from the predecessor to the successor by 60 /// duplicating the contents of this block. 61 /// 62 /// An example of when this can occur is code like this: 63 /// 64 /// if () { ... 65 /// X = 4; 66 /// } 67 /// if (X < 3) { 68 /// 69 /// In this case, the unconditional branch at the end of the first if can be 70 /// revectored to the false side of the second if. 71 /// 72 class JumpThreading : public FunctionPass { 73 TargetData *TD; 74 LazyValueInfo *LVI; 75 #ifdef NDEBUG 76 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 77 #else 78 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 79 #endif 80 public: 81 static char ID; // Pass identification 82 JumpThreading() : FunctionPass(ID) {} 83 84 bool runOnFunction(Function &F); 85 86 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 87 if (EnableLVI) 88 AU.addRequired<LazyValueInfo>(); 89 } 90 91 void FindLoopHeaders(Function &F); 92 bool ProcessBlock(BasicBlock *BB); 93 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 94 BasicBlock *SuccBB); 95 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 96 const SmallVectorImpl<BasicBlock *> &PredBBs); 97 98 typedef SmallVectorImpl<std::pair<ConstantInt*, 99 BasicBlock*> > PredValueInfo; 100 101 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 102 PredValueInfo &Result); 103 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB); 104 105 106 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 107 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 108 109 bool ProcessBranchOnPHI(PHINode *PN); 110 bool ProcessBranchOnXOR(BinaryOperator *BO); 111 112 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 113 }; 114 } 115 116 char JumpThreading::ID = 0; 117 INITIALIZE_PASS(JumpThreading, "jump-threading", 118 "Jump Threading", false, false); 119 120 // Public interface to the Jump Threading pass 121 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 122 123 /// runOnFunction - Top level algorithm. 124 /// 125 bool JumpThreading::runOnFunction(Function &F) { 126 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 127 TD = getAnalysisIfAvailable<TargetData>(); 128 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0; 129 130 FindLoopHeaders(F); 131 132 bool Changed, EverChanged = false; 133 do { 134 Changed = false; 135 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 136 BasicBlock *BB = I; 137 // Thread all of the branches we can over this block. 138 while (ProcessBlock(BB)) 139 Changed = true; 140 141 ++I; 142 143 // If the block is trivially dead, zap it. This eliminates the successor 144 // edges which simplifies the CFG. 145 if (pred_begin(BB) == pred_end(BB) && 146 BB != &BB->getParent()->getEntryBlock()) { 147 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 148 << "' with terminator: " << *BB->getTerminator() << '\n'); 149 LoopHeaders.erase(BB); 150 if (LVI) LVI->eraseBlock(BB); 151 DeleteDeadBlock(BB); 152 Changed = true; 153 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 154 // Can't thread an unconditional jump, but if the block is "almost 155 // empty", we can replace uses of it with uses of the successor and make 156 // this dead. 157 if (BI->isUnconditional() && 158 BB != &BB->getParent()->getEntryBlock()) { 159 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 160 // Ignore dbg intrinsics. 161 while (isa<DbgInfoIntrinsic>(BBI)) 162 ++BBI; 163 // If the terminator is the only non-phi instruction, try to nuke it. 164 if (BBI->isTerminator()) { 165 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 166 // block, we have to make sure it isn't in the LoopHeaders set. We 167 // reinsert afterward if needed. 168 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 169 BasicBlock *Succ = BI->getSuccessor(0); 170 171 // FIXME: It is always conservatively correct to drop the info 172 // for a block even if it doesn't get erased. This isn't totally 173 // awesome, but it allows us to use AssertingVH to prevent nasty 174 // dangling pointer issues within LazyValueInfo. 175 if (LVI) LVI->eraseBlock(BB); 176 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 177 Changed = true; 178 // If we deleted BB and BB was the header of a loop, then the 179 // successor is now the header of the loop. 180 BB = Succ; 181 } 182 183 if (ErasedFromLoopHeaders) 184 LoopHeaders.insert(BB); 185 } 186 } 187 } 188 } 189 EverChanged |= Changed; 190 } while (Changed); 191 192 LoopHeaders.clear(); 193 return EverChanged; 194 } 195 196 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 197 /// thread across it. 198 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 199 /// Ignore PHI nodes, these will be flattened when duplication happens. 200 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 201 202 // FIXME: THREADING will delete values that are just used to compute the 203 // branch, so they shouldn't count against the duplication cost. 204 205 206 // Sum up the cost of each instruction until we get to the terminator. Don't 207 // include the terminator because the copy won't include it. 208 unsigned Size = 0; 209 for (; !isa<TerminatorInst>(I); ++I) { 210 // Debugger intrinsics don't incur code size. 211 if (isa<DbgInfoIntrinsic>(I)) continue; 212 213 // If this is a pointer->pointer bitcast, it is free. 214 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 215 continue; 216 217 // All other instructions count for at least one unit. 218 ++Size; 219 220 // Calls are more expensive. If they are non-intrinsic calls, we model them 221 // as having cost of 4. If they are a non-vector intrinsic, we model them 222 // as having cost of 2 total, and if they are a vector intrinsic, we model 223 // them as having cost 1. 224 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 225 if (!isa<IntrinsicInst>(CI)) 226 Size += 3; 227 else if (!CI->getType()->isVectorTy()) 228 Size += 1; 229 } 230 } 231 232 // Threading through a switch statement is particularly profitable. If this 233 // block ends in a switch, decrease its cost to make it more likely to happen. 234 if (isa<SwitchInst>(I)) 235 Size = Size > 6 ? Size-6 : 0; 236 237 return Size; 238 } 239 240 /// FindLoopHeaders - We do not want jump threading to turn proper loop 241 /// structures into irreducible loops. Doing this breaks up the loop nesting 242 /// hierarchy and pessimizes later transformations. To prevent this from 243 /// happening, we first have to find the loop headers. Here we approximate this 244 /// by finding targets of backedges in the CFG. 245 /// 246 /// Note that there definitely are cases when we want to allow threading of 247 /// edges across a loop header. For example, threading a jump from outside the 248 /// loop (the preheader) to an exit block of the loop is definitely profitable. 249 /// It is also almost always profitable to thread backedges from within the loop 250 /// to exit blocks, and is often profitable to thread backedges to other blocks 251 /// within the loop (forming a nested loop). This simple analysis is not rich 252 /// enough to track all of these properties and keep it up-to-date as the CFG 253 /// mutates, so we don't allow any of these transformations. 254 /// 255 void JumpThreading::FindLoopHeaders(Function &F) { 256 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 257 FindFunctionBackedges(F, Edges); 258 259 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 260 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 261 } 262 263 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 264 /// if we can infer that the value is a known ConstantInt in any of our 265 /// predecessors. If so, return the known list of value and pred BB in the 266 /// result vector. If a value is known to be undef, it is returned as null. 267 /// 268 /// This returns true if there were any known values. 269 /// 270 bool JumpThreading:: 271 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 272 // If V is a constantint, then it is known in all predecessors. 273 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 274 ConstantInt *CI = dyn_cast<ConstantInt>(V); 275 276 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 277 Result.push_back(std::make_pair(CI, *PI)); 278 return true; 279 } 280 281 // If V is a non-instruction value, or an instruction in a different block, 282 // then it can't be derived from a PHI. 283 Instruction *I = dyn_cast<Instruction>(V); 284 if (I == 0 || I->getParent() != BB) { 285 286 // Okay, if this is a live-in value, see if it has a known value at the end 287 // of any of our predecessors. 288 // 289 // FIXME: This should be an edge property, not a block end property. 290 /// TODO: Per PR2563, we could infer value range information about a 291 /// predecessor based on its terminator. 292 // 293 if (LVI) { 294 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 295 // "I" is a non-local compare-with-a-constant instruction. This would be 296 // able to handle value inequalities better, for example if the compare is 297 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 298 // Perhaps getConstantOnEdge should be smart enough to do this? 299 300 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 301 BasicBlock *P = *PI; 302 // If the value is known by LazyValueInfo to be a constant in a 303 // predecessor, use that information to try to thread this block. 304 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 305 if (PredCst == 0 || 306 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst))) 307 continue; 308 309 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P)); 310 } 311 312 return !Result.empty(); 313 } 314 315 return false; 316 } 317 318 /// If I is a PHI node, then we know the incoming values for any constants. 319 if (PHINode *PN = dyn_cast<PHINode>(I)) { 320 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 321 Value *InVal = PN->getIncomingValue(i); 322 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 323 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 324 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 325 } else if (LVI) { 326 Constant *CI = LVI->getConstantOnEdge(InVal, 327 PN->getIncomingBlock(i), BB); 328 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI); 329 if (CInt) 330 Result.push_back(std::make_pair(CInt, PN->getIncomingBlock(i))); 331 } 332 } 333 return !Result.empty(); 334 } 335 336 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 337 338 // Handle some boolean conditions. 339 if (I->getType()->getPrimitiveSizeInBits() == 1) { 340 // X | true -> true 341 // X & false -> false 342 if (I->getOpcode() == Instruction::Or || 343 I->getOpcode() == Instruction::And) { 344 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 345 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 346 347 if (LHSVals.empty() && RHSVals.empty()) 348 return false; 349 350 ConstantInt *InterestingVal; 351 if (I->getOpcode() == Instruction::Or) 352 InterestingVal = ConstantInt::getTrue(I->getContext()); 353 else 354 InterestingVal = ConstantInt::getFalse(I->getContext()); 355 356 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 357 358 // Scan for the sentinel. If we find an undef, force it to the 359 // interesting value: x|undef -> true and x&undef -> false. 360 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 361 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) { 362 Result.push_back(LHSVals[i]); 363 Result.back().first = InterestingVal; 364 LHSKnownBBs.insert(LHSVals[i].second); 365 } 366 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 367 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) { 368 // If we already inferred a value for this block on the LHS, don't 369 // re-add it. 370 if (!LHSKnownBBs.count(RHSVals[i].second)) { 371 Result.push_back(RHSVals[i]); 372 Result.back().first = InterestingVal; 373 } 374 } 375 return !Result.empty(); 376 377 // Try to process a few other binary operator patterns. 378 } else if (isa<BinaryOperator>(I)) { 379 380 } 381 382 // Handle the NOT form of XOR. 383 if (I->getOpcode() == Instruction::Xor && 384 isa<ConstantInt>(I->getOperand(1)) && 385 cast<ConstantInt>(I->getOperand(1))->isOne()) { 386 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result); 387 if (Result.empty()) 388 return false; 389 390 // Invert the known values. 391 for (unsigned i = 0, e = Result.size(); i != e; ++i) 392 if (Result[i].first) 393 Result[i].first = 394 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first)); 395 return true; 396 } 397 398 // Try to simplify some other binary operator values. 399 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 400 // AND or OR of a value with itself is that value. 401 ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1)); 402 if (CI && (BO->getOpcode() == Instruction::And || 403 BO->getOpcode() == Instruction::Or)) { 404 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals; 405 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals); 406 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 407 if (LHSVals[i].first == CI) 408 Result.push_back(std::make_pair(CI, LHSVals[i].second)); 409 410 return !Result.empty(); 411 } 412 } 413 414 // Handle compare with phi operand, where the PHI is defined in this block. 415 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 416 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 417 if (PN && PN->getParent() == BB) { 418 // We can do this simplification if any comparisons fold to true or false. 419 // See if any do. 420 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 421 BasicBlock *PredBB = PN->getIncomingBlock(i); 422 Value *LHS = PN->getIncomingValue(i); 423 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 424 425 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 426 if (Res == 0) { 427 if (!LVI || !isa<Constant>(RHS)) 428 continue; 429 430 LazyValueInfo::Tristate 431 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 432 cast<Constant>(RHS), PredBB, BB); 433 if (ResT == LazyValueInfo::Unknown) 434 continue; 435 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 436 } 437 438 if (isa<UndefValue>(Res)) 439 Result.push_back(std::make_pair((ConstantInt*)0, PredBB)); 440 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res)) 441 Result.push_back(std::make_pair(CI, PredBB)); 442 } 443 444 return !Result.empty(); 445 } 446 447 448 // If comparing a live-in value against a constant, see if we know the 449 // live-in value on any predecessors. 450 if (LVI && isa<Constant>(Cmp->getOperand(1)) && 451 Cmp->getType()->isIntegerTy()) { 452 if (!isa<Instruction>(Cmp->getOperand(0)) || 453 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 454 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 455 456 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ 457 BasicBlock *P = *PI; 458 // If the value is known by LazyValueInfo to be a constant in a 459 // predecessor, use that information to try to thread this block. 460 LazyValueInfo::Tristate Res = 461 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 462 RHSCst, P, BB); 463 if (Res == LazyValueInfo::Unknown) 464 continue; 465 466 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 467 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P)); 468 } 469 470 return !Result.empty(); 471 } 472 473 // Try to find a constant value for the LHS of an equality comparison, 474 // and evaluate it statically if we can. 475 if (Cmp->getPredicate() == CmpInst::ICMP_EQ || 476 Cmp->getPredicate() == CmpInst::ICMP_NE) { 477 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals; 478 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 479 480 ConstantInt *True = ConstantInt::getTrue(I->getContext()); 481 ConstantInt *False = ConstantInt::getFalse(I->getContext()); 482 if (Cmp->getPredicate() == CmpInst::ICMP_NE) std::swap(True, False); 483 484 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 485 if (LHSVals[i].first == Cmp->getOperand(1)) 486 Result.push_back(std::make_pair(True, LHSVals[i].second)); 487 else 488 Result.push_back(std::make_pair(False, LHSVals[i].second)); 489 } 490 491 return !Result.empty(); 492 } 493 } 494 } 495 496 if (LVI) { 497 // If all else fails, see if LVI can figure out a constant value for us. 498 Constant *CI = LVI->getConstant(V, BB); 499 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI); 500 if (CInt) { 501 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 502 Result.push_back(std::make_pair(CInt, *PI)); 503 } 504 505 return !Result.empty(); 506 } 507 508 return false; 509 } 510 511 512 513 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends 514 /// in an undefined jump, decide which block is best to revector to. 515 /// 516 /// Since we can pick an arbitrary destination, we pick the successor with the 517 /// fewest predecessors. This should reduce the in-degree of the others. 518 /// 519 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 520 TerminatorInst *BBTerm = BB->getTerminator(); 521 unsigned MinSucc = 0; 522 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 523 // Compute the successor with the minimum number of predecessors. 524 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 525 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 526 TestBB = BBTerm->getSuccessor(i); 527 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 528 if (NumPreds < MinNumPreds) 529 MinSucc = i; 530 } 531 532 return MinSucc; 533 } 534 535 /// ProcessBlock - If there are any predecessors whose control can be threaded 536 /// through to a successor, transform them now. 537 bool JumpThreading::ProcessBlock(BasicBlock *BB) { 538 // If the block is trivially dead, just return and let the caller nuke it. 539 // This simplifies other transformations. 540 if (pred_begin(BB) == pred_end(BB) && 541 BB != &BB->getParent()->getEntryBlock()) 542 return false; 543 544 // If this block has a single predecessor, and if that pred has a single 545 // successor, merge the blocks. This encourages recursive jump threading 546 // because now the condition in this block can be threaded through 547 // predecessors of our predecessor block. 548 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 549 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 550 SinglePred != BB) { 551 // If SinglePred was a loop header, BB becomes one. 552 if (LoopHeaders.erase(SinglePred)) 553 LoopHeaders.insert(BB); 554 555 // Remember if SinglePred was the entry block of the function. If so, we 556 // will need to move BB back to the entry position. 557 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 558 if (LVI) LVI->eraseBlock(SinglePred); 559 MergeBasicBlockIntoOnlyPred(BB); 560 561 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 562 BB->moveBefore(&BB->getParent()->getEntryBlock()); 563 return true; 564 } 565 } 566 567 // Look to see if the terminator is a branch of switch, if not we can't thread 568 // it. 569 Value *Condition; 570 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 571 // Can't thread an unconditional jump. 572 if (BI->isUnconditional()) return false; 573 Condition = BI->getCondition(); 574 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 575 Condition = SI->getCondition(); 576 else 577 return false; // Must be an invoke. 578 579 // If the terminator of this block is branching on a constant, simplify the 580 // terminator to an unconditional branch. This can occur due to threading in 581 // other blocks. 582 if (isa<ConstantInt>(Condition)) { 583 DEBUG(dbgs() << " In block '" << BB->getName() 584 << "' folding terminator: " << *BB->getTerminator() << '\n'); 585 ++NumFolds; 586 ConstantFoldTerminator(BB); 587 return true; 588 } 589 590 // If the terminator is branching on an undef, we can pick any of the 591 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 592 if (isa<UndefValue>(Condition)) { 593 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 594 595 // Fold the branch/switch. 596 TerminatorInst *BBTerm = BB->getTerminator(); 597 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 598 if (i == BestSucc) continue; 599 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD); 600 } 601 602 DEBUG(dbgs() << " In block '" << BB->getName() 603 << "' folding undef terminator: " << *BBTerm << '\n'); 604 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 605 BBTerm->eraseFromParent(); 606 return true; 607 } 608 609 Instruction *CondInst = dyn_cast<Instruction>(Condition); 610 611 // If the condition is an instruction defined in another block, see if a 612 // predecessor has the same condition: 613 // br COND, BBX, BBY 614 // BBX: 615 // br COND, BBZ, BBW 616 if (!LVI && 617 !Condition->hasOneUse() && // Multiple uses. 618 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 619 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 620 if (isa<BranchInst>(BB->getTerminator())) { 621 for (; PI != E; ++PI) { 622 BasicBlock *P = *PI; 623 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator())) 624 if (PBI->isConditional() && PBI->getCondition() == Condition && 625 ProcessBranchOnDuplicateCond(P, BB)) 626 return true; 627 } 628 } else { 629 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 630 for (; PI != E; ++PI) { 631 BasicBlock *P = *PI; 632 if (SwitchInst *PSI = dyn_cast<SwitchInst>(P->getTerminator())) 633 if (PSI->getCondition() == Condition && 634 ProcessSwitchOnDuplicateCond(P, BB)) 635 return true; 636 } 637 } 638 } 639 640 // All the rest of our checks depend on the condition being an instruction. 641 if (CondInst == 0) { 642 // FIXME: Unify this with code below. 643 if (LVI && ProcessThreadableEdges(Condition, BB)) 644 return true; 645 return false; 646 } 647 648 649 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 650 if (!LVI && 651 (!isa<PHINode>(CondCmp->getOperand(0)) || 652 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) { 653 // If we have a comparison, loop over the predecessors to see if there is 654 // a condition with a lexically identical value. 655 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 656 for (; PI != E; ++PI) { 657 BasicBlock *P = *PI; 658 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator())) 659 if (PBI->isConditional() && P != BB) { 660 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 661 if (CI->getOperand(0) == CondCmp->getOperand(0) && 662 CI->getOperand(1) == CondCmp->getOperand(1) && 663 CI->getPredicate() == CondCmp->getPredicate()) { 664 // TODO: Could handle things like (x != 4) --> (x == 17) 665 if (ProcessBranchOnDuplicateCond(P, BB)) 666 return true; 667 } 668 } 669 } 670 } 671 } 672 } 673 674 // Check for some cases that are worth simplifying. Right now we want to look 675 // for loads that are used by a switch or by the condition for the branch. If 676 // we see one, check to see if it's partially redundant. If so, insert a PHI 677 // which can then be used to thread the values. 678 // 679 Value *SimplifyValue = CondInst; 680 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 681 if (isa<Constant>(CondCmp->getOperand(1))) 682 SimplifyValue = CondCmp->getOperand(0); 683 684 // TODO: There are other places where load PRE would be profitable, such as 685 // more complex comparisons. 686 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 687 if (SimplifyPartiallyRedundantLoad(LI)) 688 return true; 689 690 691 // Handle a variety of cases where we are branching on something derived from 692 // a PHI node in the current block. If we can prove that any predecessors 693 // compute a predictable value based on a PHI node, thread those predecessors. 694 // 695 if (ProcessThreadableEdges(CondInst, BB)) 696 return true; 697 698 // If this is an otherwise-unfoldable branch on a phi node in the current 699 // block, see if we can simplify. 700 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 701 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 702 return ProcessBranchOnPHI(PN); 703 704 705 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 706 if (CondInst->getOpcode() == Instruction::Xor && 707 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 708 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 709 710 711 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 712 // "(X == 4)", thread through this block. 713 714 return false; 715 } 716 717 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 718 /// block that jump on exactly the same condition. This means that we almost 719 /// always know the direction of the edge in the DESTBB: 720 /// PREDBB: 721 /// br COND, DESTBB, BBY 722 /// DESTBB: 723 /// br COND, BBZ, BBW 724 /// 725 /// If DESTBB has multiple predecessors, we can't just constant fold the branch 726 /// in DESTBB, we have to thread over it. 727 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 728 BasicBlock *BB) { 729 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 730 731 // If both successors of PredBB go to DESTBB, we don't know anything. We can 732 // fold the branch to an unconditional one, which allows other recursive 733 // simplifications. 734 bool BranchDir; 735 if (PredBI->getSuccessor(1) != BB) 736 BranchDir = true; 737 else if (PredBI->getSuccessor(0) != BB) 738 BranchDir = false; 739 else { 740 DEBUG(dbgs() << " In block '" << PredBB->getName() 741 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 742 ++NumFolds; 743 ConstantFoldTerminator(PredBB); 744 return true; 745 } 746 747 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 748 749 // If the dest block has one predecessor, just fix the branch condition to a 750 // constant and fold it. 751 if (BB->getSinglePredecessor()) { 752 DEBUG(dbgs() << " In block '" << BB->getName() 753 << "' folding condition to '" << BranchDir << "': " 754 << *BB->getTerminator() << '\n'); 755 ++NumFolds; 756 Value *OldCond = DestBI->getCondition(); 757 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 758 BranchDir)); 759 // Delete dead instructions before we fold the branch. Folding the branch 760 // can eliminate edges from the CFG which can end up deleting OldCond. 761 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 762 ConstantFoldTerminator(BB); 763 return true; 764 } 765 766 767 // Next, figure out which successor we are threading to. 768 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 769 770 SmallVector<BasicBlock*, 2> Preds; 771 Preds.push_back(PredBB); 772 773 // Ok, try to thread it! 774 return ThreadEdge(BB, Preds, SuccBB); 775 } 776 777 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 778 /// block that switch on exactly the same condition. This means that we almost 779 /// always know the direction of the edge in the DESTBB: 780 /// PREDBB: 781 /// switch COND [... DESTBB, BBY ... ] 782 /// DESTBB: 783 /// switch COND [... BBZ, BBW ] 784 /// 785 /// Optimizing switches like this is very important, because simplifycfg builds 786 /// switches out of repeated 'if' conditions. 787 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 788 BasicBlock *DestBB) { 789 // Can't thread edge to self. 790 if (PredBB == DestBB) 791 return false; 792 793 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 794 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 795 796 // There are a variety of optimizations that we can potentially do on these 797 // blocks: we order them from most to least preferable. 798 799 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 800 // directly to their destination. This does not introduce *any* code size 801 // growth. Skip debug info first. 802 BasicBlock::iterator BBI = DestBB->begin(); 803 while (isa<DbgInfoIntrinsic>(BBI)) 804 BBI++; 805 806 // FIXME: Thread if it just contains a PHI. 807 if (isa<SwitchInst>(BBI)) { 808 bool MadeChange = false; 809 // Ignore the default edge for now. 810 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 811 ConstantInt *DestVal = DestSI->getCaseValue(i); 812 BasicBlock *DestSucc = DestSI->getSuccessor(i); 813 814 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 815 // PredSI has an explicit case for it. If so, forward. If it is covered 816 // by the default case, we can't update PredSI. 817 unsigned PredCase = PredSI->findCaseValue(DestVal); 818 if (PredCase == 0) continue; 819 820 // If PredSI doesn't go to DestBB on this value, then it won't reach the 821 // case on this condition. 822 if (PredSI->getSuccessor(PredCase) != DestBB && 823 DestSI->getSuccessor(i) != DestBB) 824 continue; 825 826 // Do not forward this if it already goes to this destination, this would 827 // be an infinite loop. 828 if (PredSI->getSuccessor(PredCase) == DestSucc) 829 continue; 830 831 // Otherwise, we're safe to make the change. Make sure that the edge from 832 // DestSI to DestSucc is not critical and has no PHI nodes. 833 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 834 DEBUG(dbgs() << "THROUGH: " << *DestSI); 835 836 // If the destination has PHI nodes, just split the edge for updating 837 // simplicity. 838 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 839 SplitCriticalEdge(DestSI, i, this); 840 DestSucc = DestSI->getSuccessor(i); 841 } 842 FoldSingleEntryPHINodes(DestSucc); 843 PredSI->setSuccessor(PredCase, DestSucc); 844 MadeChange = true; 845 } 846 847 if (MadeChange) 848 return true; 849 } 850 851 return false; 852 } 853 854 855 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 856 /// load instruction, eliminate it by replacing it with a PHI node. This is an 857 /// important optimization that encourages jump threading, and needs to be run 858 /// interlaced with other jump threading tasks. 859 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 860 // Don't hack volatile loads. 861 if (LI->isVolatile()) return false; 862 863 // If the load is defined in a block with exactly one predecessor, it can't be 864 // partially redundant. 865 BasicBlock *LoadBB = LI->getParent(); 866 if (LoadBB->getSinglePredecessor()) 867 return false; 868 869 Value *LoadedPtr = LI->getOperand(0); 870 871 // If the loaded operand is defined in the LoadBB, it can't be available. 872 // TODO: Could do simple PHI translation, that would be fun :) 873 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 874 if (PtrOp->getParent() == LoadBB) 875 return false; 876 877 // Scan a few instructions up from the load, to see if it is obviously live at 878 // the entry to its block. 879 BasicBlock::iterator BBIt = LI; 880 881 if (Value *AvailableVal = 882 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 883 // If the value if the load is locally available within the block, just use 884 // it. This frequently occurs for reg2mem'd allocas. 885 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 886 887 // If the returned value is the load itself, replace with an undef. This can 888 // only happen in dead loops. 889 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 890 LI->replaceAllUsesWith(AvailableVal); 891 LI->eraseFromParent(); 892 return true; 893 } 894 895 // Otherwise, if we scanned the whole block and got to the top of the block, 896 // we know the block is locally transparent to the load. If not, something 897 // might clobber its value. 898 if (BBIt != LoadBB->begin()) 899 return false; 900 901 902 SmallPtrSet<BasicBlock*, 8> PredsScanned; 903 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 904 AvailablePredsTy AvailablePreds; 905 BasicBlock *OneUnavailablePred = 0; 906 907 // If we got here, the loaded value is transparent through to the start of the 908 // block. Check to see if it is available in any of the predecessor blocks. 909 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 910 PI != PE; ++PI) { 911 BasicBlock *PredBB = *PI; 912 913 // If we already scanned this predecessor, skip it. 914 if (!PredsScanned.insert(PredBB)) 915 continue; 916 917 // Scan the predecessor to see if the value is available in the pred. 918 BBIt = PredBB->end(); 919 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 920 if (!PredAvailable) { 921 OneUnavailablePred = PredBB; 922 continue; 923 } 924 925 // If so, this load is partially redundant. Remember this info so that we 926 // can create a PHI node. 927 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 928 } 929 930 // If the loaded value isn't available in any predecessor, it isn't partially 931 // redundant. 932 if (AvailablePreds.empty()) return false; 933 934 // Okay, the loaded value is available in at least one (and maybe all!) 935 // predecessors. If the value is unavailable in more than one unique 936 // predecessor, we want to insert a merge block for those common predecessors. 937 // This ensures that we only have to insert one reload, thus not increasing 938 // code size. 939 BasicBlock *UnavailablePred = 0; 940 941 // If there is exactly one predecessor where the value is unavailable, the 942 // already computed 'OneUnavailablePred' block is it. If it ends in an 943 // unconditional branch, we know that it isn't a critical edge. 944 if (PredsScanned.size() == AvailablePreds.size()+1 && 945 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 946 UnavailablePred = OneUnavailablePred; 947 } else if (PredsScanned.size() != AvailablePreds.size()) { 948 // Otherwise, we had multiple unavailable predecessors or we had a critical 949 // edge from the one. 950 SmallVector<BasicBlock*, 8> PredsToSplit; 951 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 952 953 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 954 AvailablePredSet.insert(AvailablePreds[i].first); 955 956 // Add all the unavailable predecessors to the PredsToSplit list. 957 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 958 PI != PE; ++PI) { 959 BasicBlock *P = *PI; 960 // If the predecessor is an indirect goto, we can't split the edge. 961 if (isa<IndirectBrInst>(P->getTerminator())) 962 return false; 963 964 if (!AvailablePredSet.count(P)) 965 PredsToSplit.push_back(P); 966 } 967 968 // Split them out to their own block. 969 UnavailablePred = 970 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 971 "thread-pre-split", this); 972 } 973 974 // If the value isn't available in all predecessors, then there will be 975 // exactly one where it isn't available. Insert a load on that edge and add 976 // it to the AvailablePreds list. 977 if (UnavailablePred) { 978 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 979 "Can't handle critical edge here!"); 980 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 981 LI->getAlignment(), 982 UnavailablePred->getTerminator()); 983 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 984 } 985 986 // Now we know that each predecessor of this block has a value in 987 // AvailablePreds, sort them for efficient access as we're walking the preds. 988 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 989 990 // Create a PHI node at the start of the block for the PRE'd load value. 991 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 992 PN->takeName(LI); 993 994 // Insert new entries into the PHI for each predecessor. A single block may 995 // have multiple entries here. 996 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 997 ++PI) { 998 BasicBlock *P = *PI; 999 AvailablePredsTy::iterator I = 1000 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 1001 std::make_pair(P, (Value*)0)); 1002 1003 assert(I != AvailablePreds.end() && I->first == P && 1004 "Didn't find entry for predecessor!"); 1005 1006 PN->addIncoming(I->second, I->first); 1007 } 1008 1009 //cerr << "PRE: " << *LI << *PN << "\n"; 1010 1011 LI->replaceAllUsesWith(PN); 1012 LI->eraseFromParent(); 1013 1014 return true; 1015 } 1016 1017 /// FindMostPopularDest - The specified list contains multiple possible 1018 /// threadable destinations. Pick the one that occurs the most frequently in 1019 /// the list. 1020 static BasicBlock * 1021 FindMostPopularDest(BasicBlock *BB, 1022 const SmallVectorImpl<std::pair<BasicBlock*, 1023 BasicBlock*> > &PredToDestList) { 1024 assert(!PredToDestList.empty()); 1025 1026 // Determine popularity. If there are multiple possible destinations, we 1027 // explicitly choose to ignore 'undef' destinations. We prefer to thread 1028 // blocks with known and real destinations to threading undef. We'll handle 1029 // them later if interesting. 1030 DenseMap<BasicBlock*, unsigned> DestPopularity; 1031 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1032 if (PredToDestList[i].second) 1033 DestPopularity[PredToDestList[i].second]++; 1034 1035 // Find the most popular dest. 1036 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 1037 BasicBlock *MostPopularDest = DPI->first; 1038 unsigned Popularity = DPI->second; 1039 SmallVector<BasicBlock*, 4> SamePopularity; 1040 1041 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 1042 // If the popularity of this entry isn't higher than the popularity we've 1043 // seen so far, ignore it. 1044 if (DPI->second < Popularity) 1045 ; // ignore. 1046 else if (DPI->second == Popularity) { 1047 // If it is the same as what we've seen so far, keep track of it. 1048 SamePopularity.push_back(DPI->first); 1049 } else { 1050 // If it is more popular, remember it. 1051 SamePopularity.clear(); 1052 MostPopularDest = DPI->first; 1053 Popularity = DPI->second; 1054 } 1055 } 1056 1057 // Okay, now we know the most popular destination. If there is more than 1058 // destination, we need to determine one. This is arbitrary, but we need 1059 // to make a deterministic decision. Pick the first one that appears in the 1060 // successor list. 1061 if (!SamePopularity.empty()) { 1062 SamePopularity.push_back(MostPopularDest); 1063 TerminatorInst *TI = BB->getTerminator(); 1064 for (unsigned i = 0; ; ++i) { 1065 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1066 1067 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1068 TI->getSuccessor(i)) == SamePopularity.end()) 1069 continue; 1070 1071 MostPopularDest = TI->getSuccessor(i); 1072 break; 1073 } 1074 } 1075 1076 // Okay, we have finally picked the most popular destination. 1077 return MostPopularDest; 1078 } 1079 1080 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) { 1081 // If threading this would thread across a loop header, don't even try to 1082 // thread the edge. 1083 if (LoopHeaders.count(BB)) 1084 return false; 1085 1086 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 1087 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) 1088 return false; 1089 assert(!PredValues.empty() && 1090 "ComputeValueKnownInPredecessors returned true with no values"); 1091 1092 DEBUG(dbgs() << "IN BB: " << *BB; 1093 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1094 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "; 1095 if (PredValues[i].first) 1096 dbgs() << *PredValues[i].first; 1097 else 1098 dbgs() << "UNDEF"; 1099 dbgs() << " for pred '" << PredValues[i].second->getName() 1100 << "'.\n"; 1101 }); 1102 1103 // Decide what we want to thread through. Convert our list of known values to 1104 // a list of known destinations for each pred. This also discards duplicate 1105 // predecessors and keeps track of the undefined inputs (which are represented 1106 // as a null dest in the PredToDestList). 1107 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1108 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1109 1110 BasicBlock *OnlyDest = 0; 1111 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1112 1113 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1114 BasicBlock *Pred = PredValues[i].second; 1115 if (!SeenPreds.insert(Pred)) 1116 continue; // Duplicate predecessor entry. 1117 1118 // If the predecessor ends with an indirect goto, we can't change its 1119 // destination. 1120 if (isa<IndirectBrInst>(Pred->getTerminator())) 1121 continue; 1122 1123 ConstantInt *Val = PredValues[i].first; 1124 1125 BasicBlock *DestBB; 1126 if (Val == 0) // Undef. 1127 DestBB = 0; 1128 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1129 DestBB = BI->getSuccessor(Val->isZero()); 1130 else { 1131 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1132 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1133 } 1134 1135 // If we have exactly one destination, remember it for efficiency below. 1136 if (i == 0) 1137 OnlyDest = DestBB; 1138 else if (OnlyDest != DestBB) 1139 OnlyDest = MultipleDestSentinel; 1140 1141 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1142 } 1143 1144 // If all edges were unthreadable, we fail. 1145 if (PredToDestList.empty()) 1146 return false; 1147 1148 // Determine which is the most common successor. If we have many inputs and 1149 // this block is a switch, we want to start by threading the batch that goes 1150 // to the most popular destination first. If we only know about one 1151 // threadable destination (the common case) we can avoid this. 1152 BasicBlock *MostPopularDest = OnlyDest; 1153 1154 if (MostPopularDest == MultipleDestSentinel) 1155 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1156 1157 // Now that we know what the most popular destination is, factor all 1158 // predecessors that will jump to it into a single predecessor. 1159 SmallVector<BasicBlock*, 16> PredsToFactor; 1160 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1161 if (PredToDestList[i].second == MostPopularDest) { 1162 BasicBlock *Pred = PredToDestList[i].first; 1163 1164 // This predecessor may be a switch or something else that has multiple 1165 // edges to the block. Factor each of these edges by listing them 1166 // according to # occurrences in PredsToFactor. 1167 TerminatorInst *PredTI = Pred->getTerminator(); 1168 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1169 if (PredTI->getSuccessor(i) == BB) 1170 PredsToFactor.push_back(Pred); 1171 } 1172 1173 // If the threadable edges are branching on an undefined value, we get to pick 1174 // the destination that these predecessors should get to. 1175 if (MostPopularDest == 0) 1176 MostPopularDest = BB->getTerminator()-> 1177 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1178 1179 // Ok, try to thread it! 1180 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1181 } 1182 1183 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1184 /// a PHI node in the current block. See if there are any simplifications we 1185 /// can do based on inputs to the phi node. 1186 /// 1187 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1188 BasicBlock *BB = PN->getParent(); 1189 1190 // TODO: We could make use of this to do it once for blocks with common PHI 1191 // values. 1192 SmallVector<BasicBlock*, 1> PredBBs; 1193 PredBBs.resize(1); 1194 1195 // If any of the predecessor blocks end in an unconditional branch, we can 1196 // *duplicate* the conditional branch into that block in order to further 1197 // encourage jump threading and to eliminate cases where we have branch on a 1198 // phi of an icmp (branch on icmp is much better). 1199 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1200 BasicBlock *PredBB = PN->getIncomingBlock(i); 1201 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1202 if (PredBr->isUnconditional()) { 1203 PredBBs[0] = PredBB; 1204 // Try to duplicate BB into PredBB. 1205 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1206 return true; 1207 } 1208 } 1209 1210 return false; 1211 } 1212 1213 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1214 /// a xor instruction in the current block. See if there are any 1215 /// simplifications we can do based on inputs to the xor. 1216 /// 1217 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1218 BasicBlock *BB = BO->getParent(); 1219 1220 // If either the LHS or RHS of the xor is a constant, don't do this 1221 // optimization. 1222 if (isa<ConstantInt>(BO->getOperand(0)) || 1223 isa<ConstantInt>(BO->getOperand(1))) 1224 return false; 1225 1226 // If the first instruction in BB isn't a phi, we won't be able to infer 1227 // anything special about any particular predecessor. 1228 if (!isa<PHINode>(BB->front())) 1229 return false; 1230 1231 // If we have a xor as the branch input to this block, and we know that the 1232 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1233 // the condition into the predecessor and fix that value to true, saving some 1234 // logical ops on that path and encouraging other paths to simplify. 1235 // 1236 // This copies something like this: 1237 // 1238 // BB: 1239 // %X = phi i1 [1], [%X'] 1240 // %Y = icmp eq i32 %A, %B 1241 // %Z = xor i1 %X, %Y 1242 // br i1 %Z, ... 1243 // 1244 // Into: 1245 // BB': 1246 // %Y = icmp ne i32 %A, %B 1247 // br i1 %Z, ... 1248 1249 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues; 1250 bool isLHS = true; 1251 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) { 1252 assert(XorOpValues.empty()); 1253 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues)) 1254 return false; 1255 isLHS = false; 1256 } 1257 1258 assert(!XorOpValues.empty() && 1259 "ComputeValueKnownInPredecessors returned true with no values"); 1260 1261 // Scan the information to see which is most popular: true or false. The 1262 // predecessors can be of the set true, false, or undef. 1263 unsigned NumTrue = 0, NumFalse = 0; 1264 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1265 if (!XorOpValues[i].first) continue; // Ignore undefs for the count. 1266 if (XorOpValues[i].first->isZero()) 1267 ++NumFalse; 1268 else 1269 ++NumTrue; 1270 } 1271 1272 // Determine which value to split on, true, false, or undef if neither. 1273 ConstantInt *SplitVal = 0; 1274 if (NumTrue > NumFalse) 1275 SplitVal = ConstantInt::getTrue(BB->getContext()); 1276 else if (NumTrue != 0 || NumFalse != 0) 1277 SplitVal = ConstantInt::getFalse(BB->getContext()); 1278 1279 // Collect all of the blocks that this can be folded into so that we can 1280 // factor this once and clone it once. 1281 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1282 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1283 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue; 1284 1285 BlocksToFoldInto.push_back(XorOpValues[i].second); 1286 } 1287 1288 // If we inferred a value for all of the predecessors, then duplication won't 1289 // help us. However, we can just replace the LHS or RHS with the constant. 1290 if (BlocksToFoldInto.size() == 1291 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1292 if (SplitVal == 0) { 1293 // If all preds provide undef, just nuke the xor, because it is undef too. 1294 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1295 BO->eraseFromParent(); 1296 } else if (SplitVal->isZero()) { 1297 // If all preds provide 0, replace the xor with the other input. 1298 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1299 BO->eraseFromParent(); 1300 } else { 1301 // If all preds provide 1, set the computed value to 1. 1302 BO->setOperand(!isLHS, SplitVal); 1303 } 1304 1305 return true; 1306 } 1307 1308 // Try to duplicate BB into PredBB. 1309 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1310 } 1311 1312 1313 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1314 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1315 /// NewPred using the entries from OldPred (suitably mapped). 1316 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1317 BasicBlock *OldPred, 1318 BasicBlock *NewPred, 1319 DenseMap<Instruction*, Value*> &ValueMap) { 1320 for (BasicBlock::iterator PNI = PHIBB->begin(); 1321 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1322 // Ok, we have a PHI node. Figure out what the incoming value was for the 1323 // DestBlock. 1324 Value *IV = PN->getIncomingValueForBlock(OldPred); 1325 1326 // Remap the value if necessary. 1327 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1328 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1329 if (I != ValueMap.end()) 1330 IV = I->second; 1331 } 1332 1333 PN->addIncoming(IV, NewPred); 1334 } 1335 } 1336 1337 /// ThreadEdge - We have decided that it is safe and profitable to factor the 1338 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1339 /// across BB. Transform the IR to reflect this change. 1340 bool JumpThreading::ThreadEdge(BasicBlock *BB, 1341 const SmallVectorImpl<BasicBlock*> &PredBBs, 1342 BasicBlock *SuccBB) { 1343 // If threading to the same block as we come from, we would infinite loop. 1344 if (SuccBB == BB) { 1345 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1346 << "' - would thread to self!\n"); 1347 return false; 1348 } 1349 1350 // If threading this would thread across a loop header, don't thread the edge. 1351 // See the comments above FindLoopHeaders for justifications and caveats. 1352 if (LoopHeaders.count(BB)) { 1353 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1354 << "' to dest BB '" << SuccBB->getName() 1355 << "' - it might create an irreducible loop!\n"); 1356 return false; 1357 } 1358 1359 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1360 if (JumpThreadCost > Threshold) { 1361 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1362 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1363 return false; 1364 } 1365 1366 // And finally, do it! Start by factoring the predecessors is needed. 1367 BasicBlock *PredBB; 1368 if (PredBBs.size() == 1) 1369 PredBB = PredBBs[0]; 1370 else { 1371 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1372 << " common predecessors.\n"); 1373 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1374 ".thr_comm", this); 1375 } 1376 1377 // And finally, do it! 1378 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1379 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1380 << ", across block:\n " 1381 << *BB << "\n"); 1382 1383 if (LVI) 1384 LVI->threadEdge(PredBB, BB, SuccBB); 1385 1386 // We are going to have to map operands from the original BB block to the new 1387 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1388 // account for entry from PredBB. 1389 DenseMap<Instruction*, Value*> ValueMapping; 1390 1391 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1392 BB->getName()+".thread", 1393 BB->getParent(), BB); 1394 NewBB->moveAfter(PredBB); 1395 1396 BasicBlock::iterator BI = BB->begin(); 1397 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1398 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1399 1400 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1401 // mapping and using it to remap operands in the cloned instructions. 1402 for (; !isa<TerminatorInst>(BI); ++BI) { 1403 Instruction *New = BI->clone(); 1404 New->setName(BI->getName()); 1405 NewBB->getInstList().push_back(New); 1406 ValueMapping[BI] = New; 1407 1408 // Remap operands to patch up intra-block references. 1409 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1410 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1411 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1412 if (I != ValueMapping.end()) 1413 New->setOperand(i, I->second); 1414 } 1415 } 1416 1417 // We didn't copy the terminator from BB over to NewBB, because there is now 1418 // an unconditional jump to SuccBB. Insert the unconditional jump. 1419 BranchInst::Create(SuccBB, NewBB); 1420 1421 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1422 // PHI nodes for NewBB now. 1423 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1424 1425 // If there were values defined in BB that are used outside the block, then we 1426 // now have to update all uses of the value to use either the original value, 1427 // the cloned value, or some PHI derived value. This can require arbitrary 1428 // PHI insertion, of which we are prepared to do, clean these up now. 1429 SSAUpdater SSAUpdate; 1430 SmallVector<Use*, 16> UsesToRename; 1431 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1432 // Scan all uses of this instruction to see if it is used outside of its 1433 // block, and if so, record them in UsesToRename. 1434 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1435 ++UI) { 1436 Instruction *User = cast<Instruction>(*UI); 1437 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1438 if (UserPN->getIncomingBlock(UI) == BB) 1439 continue; 1440 } else if (User->getParent() == BB) 1441 continue; 1442 1443 UsesToRename.push_back(&UI.getUse()); 1444 } 1445 1446 // If there are no uses outside the block, we're done with this instruction. 1447 if (UsesToRename.empty()) 1448 continue; 1449 1450 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1451 1452 // We found a use of I outside of BB. Rename all uses of I that are outside 1453 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1454 // with the two values we know. 1455 SSAUpdate.Initialize(I); 1456 SSAUpdate.AddAvailableValue(BB, I); 1457 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1458 1459 while (!UsesToRename.empty()) 1460 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1461 DEBUG(dbgs() << "\n"); 1462 } 1463 1464 1465 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1466 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1467 // us to simplify any PHI nodes in BB. 1468 TerminatorInst *PredTerm = PredBB->getTerminator(); 1469 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1470 if (PredTerm->getSuccessor(i) == BB) { 1471 RemovePredecessorAndSimplify(BB, PredBB, TD); 1472 PredTerm->setSuccessor(i, NewBB); 1473 } 1474 1475 // At this point, the IR is fully up to date and consistent. Do a quick scan 1476 // over the new instructions and zap any that are constants or dead. This 1477 // frequently happens because of phi translation. 1478 SimplifyInstructionsInBlock(NewBB, TD); 1479 1480 // Threaded an edge! 1481 ++NumThreads; 1482 return true; 1483 } 1484 1485 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1486 /// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1487 /// If we can duplicate the contents of BB up into PredBB do so now, this 1488 /// improves the odds that the branch will be on an analyzable instruction like 1489 /// a compare. 1490 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1491 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1492 assert(!PredBBs.empty() && "Can't handle an empty set"); 1493 1494 // If BB is a loop header, then duplicating this block outside the loop would 1495 // cause us to transform this into an irreducible loop, don't do this. 1496 // See the comments above FindLoopHeaders for justifications and caveats. 1497 if (LoopHeaders.count(BB)) { 1498 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1499 << "' into predecessor block '" << PredBBs[0]->getName() 1500 << "' - it might create an irreducible loop!\n"); 1501 return false; 1502 } 1503 1504 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1505 if (DuplicationCost > Threshold) { 1506 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1507 << "' - Cost is too high: " << DuplicationCost << "\n"); 1508 return false; 1509 } 1510 1511 // And finally, do it! Start by factoring the predecessors is needed. 1512 BasicBlock *PredBB; 1513 if (PredBBs.size() == 1) 1514 PredBB = PredBBs[0]; 1515 else { 1516 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1517 << " common predecessors.\n"); 1518 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1519 ".thr_comm", this); 1520 } 1521 1522 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1523 // of PredBB. 1524 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1525 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1526 << DuplicationCost << " block is:" << *BB << "\n"); 1527 1528 // Unless PredBB ends with an unconditional branch, split the edge so that we 1529 // can just clone the bits from BB into the end of the new PredBB. 1530 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1531 1532 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1533 PredBB = SplitEdge(PredBB, BB, this); 1534 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1535 } 1536 1537 // We are going to have to map operands from the original BB block into the 1538 // PredBB block. Evaluate PHI nodes in BB. 1539 DenseMap<Instruction*, Value*> ValueMapping; 1540 1541 BasicBlock::iterator BI = BB->begin(); 1542 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1543 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1544 1545 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1546 // mapping and using it to remap operands in the cloned instructions. 1547 for (; BI != BB->end(); ++BI) { 1548 Instruction *New = BI->clone(); 1549 1550 // Remap operands to patch up intra-block references. 1551 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1552 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1553 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1554 if (I != ValueMapping.end()) 1555 New->setOperand(i, I->second); 1556 } 1557 1558 // If this instruction can be simplified after the operands are updated, 1559 // just use the simplified value instead. This frequently happens due to 1560 // phi translation. 1561 if (Value *IV = SimplifyInstruction(New, TD)) { 1562 delete New; 1563 ValueMapping[BI] = IV; 1564 } else { 1565 // Otherwise, insert the new instruction into the block. 1566 New->setName(BI->getName()); 1567 PredBB->getInstList().insert(OldPredBranch, New); 1568 ValueMapping[BI] = New; 1569 } 1570 } 1571 1572 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1573 // add entries to the PHI nodes for branch from PredBB now. 1574 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1575 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1576 ValueMapping); 1577 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1578 ValueMapping); 1579 1580 // If there were values defined in BB that are used outside the block, then we 1581 // now have to update all uses of the value to use either the original value, 1582 // the cloned value, or some PHI derived value. This can require arbitrary 1583 // PHI insertion, of which we are prepared to do, clean these up now. 1584 SSAUpdater SSAUpdate; 1585 SmallVector<Use*, 16> UsesToRename; 1586 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1587 // Scan all uses of this instruction to see if it is used outside of its 1588 // block, and if so, record them in UsesToRename. 1589 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1590 ++UI) { 1591 Instruction *User = cast<Instruction>(*UI); 1592 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1593 if (UserPN->getIncomingBlock(UI) == BB) 1594 continue; 1595 } else if (User->getParent() == BB) 1596 continue; 1597 1598 UsesToRename.push_back(&UI.getUse()); 1599 } 1600 1601 // If there are no uses outside the block, we're done with this instruction. 1602 if (UsesToRename.empty()) 1603 continue; 1604 1605 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1606 1607 // We found a use of I outside of BB. Rename all uses of I that are outside 1608 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1609 // with the two values we know. 1610 SSAUpdate.Initialize(I); 1611 SSAUpdate.AddAvailableValue(BB, I); 1612 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1613 1614 while (!UsesToRename.empty()) 1615 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1616 DEBUG(dbgs() << "\n"); 1617 } 1618 1619 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1620 // that we nuked. 1621 RemovePredecessorAndSimplify(BB, PredBB, TD); 1622 1623 // Remove the unconditional branch at the end of the PredBB block. 1624 OldPredBranch->eraseFromParent(); 1625 1626 ++NumDupes; 1627 return true; 1628 } 1629 1630 1631