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 #include "llvm/Transforms/Scalar.h" 15 #include "llvm/ADT/DenseMap.h" 16 #include "llvm/ADT/DenseSet.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/SmallPtrSet.h" 19 #include "llvm/ADT/SmallSet.h" 20 #include "llvm/ADT/Statistic.h" 21 #include "llvm/Analysis/CFG.h" 22 #include "llvm/Analysis/ConstantFolding.h" 23 #include "llvm/Analysis/InstructionSimplify.h" 24 #include "llvm/Analysis/LazyValueInfo.h" 25 #include "llvm/Analysis/Loads.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/IntrinsicInst.h" 28 #include "llvm/IR/LLVMContext.h" 29 #include "llvm/IR/ValueHandle.h" 30 #include "llvm/Pass.h" 31 #include "llvm/Support/CommandLine.h" 32 #include "llvm/Support/Debug.h" 33 #include "llvm/Support/raw_ostream.h" 34 #include "llvm/Target/TargetLibraryInfo.h" 35 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 36 #include "llvm/Transforms/Utils/Local.h" 37 #include "llvm/Transforms/Utils/SSAUpdater.h" 38 using namespace llvm; 39 40 #define DEBUG_TYPE "jump-threading" 41 42 STATISTIC(NumThreads, "Number of jumps threaded"); 43 STATISTIC(NumFolds, "Number of terminators folded"); 44 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 45 46 static cl::opt<unsigned> 47 Threshold("jump-threading-threshold", 48 cl::desc("Max block size to duplicate for jump threading"), 49 cl::init(6), cl::Hidden); 50 51 namespace { 52 // These are at global scope so static functions can use them too. 53 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo; 54 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy; 55 56 // This is used to keep track of what kind of constant we're currently hoping 57 // to find. 58 enum ConstantPreference { 59 WantInteger, 60 WantBlockAddress 61 }; 62 63 /// This pass performs 'jump threading', which looks at blocks that have 64 /// multiple predecessors and multiple successors. If one or more of the 65 /// predecessors of the block can be proven to always jump to one of the 66 /// successors, we forward the edge from the predecessor to the successor by 67 /// duplicating the contents of this block. 68 /// 69 /// An example of when this can occur is code like this: 70 /// 71 /// if () { ... 72 /// X = 4; 73 /// } 74 /// if (X < 3) { 75 /// 76 /// In this case, the unconditional branch at the end of the first if can be 77 /// revectored to the false side of the second if. 78 /// 79 class JumpThreading : public FunctionPass { 80 const DataLayout *DL; 81 TargetLibraryInfo *TLI; 82 LazyValueInfo *LVI; 83 #ifdef NDEBUG 84 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 85 #else 86 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 87 #endif 88 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet; 89 90 // RAII helper for updating the recursion stack. 91 struct RecursionSetRemover { 92 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet; 93 std::pair<Value*, BasicBlock*> ThePair; 94 95 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S, 96 std::pair<Value*, BasicBlock*> P) 97 : TheSet(S), ThePair(P) { } 98 99 ~RecursionSetRemover() { 100 TheSet.erase(ThePair); 101 } 102 }; 103 public: 104 static char ID; // Pass identification 105 JumpThreading() : FunctionPass(ID) { 106 initializeJumpThreadingPass(*PassRegistry::getPassRegistry()); 107 } 108 109 bool runOnFunction(Function &F) override; 110 111 void getAnalysisUsage(AnalysisUsage &AU) const override { 112 AU.addRequired<LazyValueInfo>(); 113 AU.addPreserved<LazyValueInfo>(); 114 AU.addRequired<TargetLibraryInfo>(); 115 } 116 117 void FindLoopHeaders(Function &F); 118 bool ProcessBlock(BasicBlock *BB); 119 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 120 BasicBlock *SuccBB); 121 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 122 const SmallVectorImpl<BasicBlock *> &PredBBs); 123 124 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 125 PredValueInfo &Result, 126 ConstantPreference Preference); 127 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 128 ConstantPreference Preference); 129 130 bool ProcessBranchOnPHI(PHINode *PN); 131 bool ProcessBranchOnXOR(BinaryOperator *BO); 132 133 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 134 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB); 135 }; 136 } 137 138 char JumpThreading::ID = 0; 139 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", 140 "Jump Threading", false, false) 141 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo) 142 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 143 INITIALIZE_PASS_END(JumpThreading, "jump-threading", 144 "Jump Threading", false, false) 145 146 // Public interface to the Jump Threading pass 147 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 148 149 /// runOnFunction - Top level algorithm. 150 /// 151 bool JumpThreading::runOnFunction(Function &F) { 152 if (skipOptnoneFunction(F)) 153 return false; 154 155 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 156 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); 157 DL = DLP ? &DLP->getDataLayout() : nullptr; 158 TLI = &getAnalysis<TargetLibraryInfo>(); 159 LVI = &getAnalysis<LazyValueInfo>(); 160 161 // Remove unreachable blocks from function as they may result in infinite 162 // loop. We do threading if we found something profitable. Jump threading a 163 // branch can create other opportunities. If these opportunities form a cycle 164 // i.e. if any jump treading is undoing previous threading in the path, then 165 // we will loop forever. We take care of this issue by not jump threading for 166 // back edges. This works for normal cases but not for unreachable blocks as 167 // they may have cycle with no back edge. 168 removeUnreachableBlocks(F); 169 170 FindLoopHeaders(F); 171 172 bool Changed, EverChanged = false; 173 do { 174 Changed = false; 175 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 176 BasicBlock *BB = I; 177 // Thread all of the branches we can over this block. 178 while (ProcessBlock(BB)) 179 Changed = true; 180 181 ++I; 182 183 // If the block is trivially dead, zap it. This eliminates the successor 184 // edges which simplifies the CFG. 185 if (pred_begin(BB) == pred_end(BB) && 186 BB != &BB->getParent()->getEntryBlock()) { 187 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 188 << "' with terminator: " << *BB->getTerminator() << '\n'); 189 LoopHeaders.erase(BB); 190 LVI->eraseBlock(BB); 191 DeleteDeadBlock(BB); 192 Changed = true; 193 continue; 194 } 195 196 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 197 198 // Can't thread an unconditional jump, but if the block is "almost 199 // empty", we can replace uses of it with uses of the successor and make 200 // this dead. 201 if (BI && BI->isUnconditional() && 202 BB != &BB->getParent()->getEntryBlock() && 203 // If the terminator is the only non-phi instruction, try to nuke it. 204 BB->getFirstNonPHIOrDbg()->isTerminator()) { 205 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 206 // block, we have to make sure it isn't in the LoopHeaders set. We 207 // reinsert afterward if needed. 208 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 209 BasicBlock *Succ = BI->getSuccessor(0); 210 211 // FIXME: It is always conservatively correct to drop the info 212 // for a block even if it doesn't get erased. This isn't totally 213 // awesome, but it allows us to use AssertingVH to prevent nasty 214 // dangling pointer issues within LazyValueInfo. 215 LVI->eraseBlock(BB); 216 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 217 Changed = true; 218 // If we deleted BB and BB was the header of a loop, then the 219 // successor is now the header of the loop. 220 BB = Succ; 221 } 222 223 if (ErasedFromLoopHeaders) 224 LoopHeaders.insert(BB); 225 } 226 } 227 EverChanged |= Changed; 228 } while (Changed); 229 230 LoopHeaders.clear(); 231 return EverChanged; 232 } 233 234 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 235 /// thread across it. Stop scanning the block when passing the threshold. 236 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB, 237 unsigned Threshold) { 238 /// Ignore PHI nodes, these will be flattened when duplication happens. 239 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 240 241 // FIXME: THREADING will delete values that are just used to compute the 242 // branch, so they shouldn't count against the duplication cost. 243 244 // Sum up the cost of each instruction until we get to the terminator. Don't 245 // include the terminator because the copy won't include it. 246 unsigned Size = 0; 247 for (; !isa<TerminatorInst>(I); ++I) { 248 249 // Stop scanning the block if we've reached the threshold. 250 if (Size > Threshold) 251 return Size; 252 253 // Debugger intrinsics don't incur code size. 254 if (isa<DbgInfoIntrinsic>(I)) continue; 255 256 // If this is a pointer->pointer bitcast, it is free. 257 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 258 continue; 259 260 // All other instructions count for at least one unit. 261 ++Size; 262 263 // Calls are more expensive. If they are non-intrinsic calls, we model them 264 // as having cost of 4. If they are a non-vector intrinsic, we model them 265 // as having cost of 2 total, and if they are a vector intrinsic, we model 266 // them as having cost 1. 267 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 268 if (CI->cannotDuplicate()) 269 // Blocks with NoDuplicate are modelled as having infinite cost, so they 270 // are never duplicated. 271 return ~0U; 272 else if (!isa<IntrinsicInst>(CI)) 273 Size += 3; 274 else if (!CI->getType()->isVectorTy()) 275 Size += 1; 276 } 277 } 278 279 // Threading through a switch statement is particularly profitable. If this 280 // block ends in a switch, decrease its cost to make it more likely to happen. 281 if (isa<SwitchInst>(I)) 282 Size = Size > 6 ? Size-6 : 0; 283 284 // The same holds for indirect branches, but slightly more so. 285 if (isa<IndirectBrInst>(I)) 286 Size = Size > 8 ? Size-8 : 0; 287 288 return Size; 289 } 290 291 /// FindLoopHeaders - We do not want jump threading to turn proper loop 292 /// structures into irreducible loops. Doing this breaks up the loop nesting 293 /// hierarchy and pessimizes later transformations. To prevent this from 294 /// happening, we first have to find the loop headers. Here we approximate this 295 /// by finding targets of backedges in the CFG. 296 /// 297 /// Note that there definitely are cases when we want to allow threading of 298 /// edges across a loop header. For example, threading a jump from outside the 299 /// loop (the preheader) to an exit block of the loop is definitely profitable. 300 /// It is also almost always profitable to thread backedges from within the loop 301 /// to exit blocks, and is often profitable to thread backedges to other blocks 302 /// within the loop (forming a nested loop). This simple analysis is not rich 303 /// enough to track all of these properties and keep it up-to-date as the CFG 304 /// mutates, so we don't allow any of these transformations. 305 /// 306 void JumpThreading::FindLoopHeaders(Function &F) { 307 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 308 FindFunctionBackedges(F, Edges); 309 310 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 311 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 312 } 313 314 /// getKnownConstant - Helper method to determine if we can thread over a 315 /// terminator with the given value as its condition, and if so what value to 316 /// use for that. What kind of value this is depends on whether we want an 317 /// integer or a block address, but an undef is always accepted. 318 /// Returns null if Val is null or not an appropriate constant. 319 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { 320 if (!Val) 321 return nullptr; 322 323 // Undef is "known" enough. 324 if (UndefValue *U = dyn_cast<UndefValue>(Val)) 325 return U; 326 327 if (Preference == WantBlockAddress) 328 return dyn_cast<BlockAddress>(Val->stripPointerCasts()); 329 330 return dyn_cast<ConstantInt>(Val); 331 } 332 333 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 334 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef 335 /// in any of our predecessors. If so, return the known list of value and pred 336 /// BB in the result vector. 337 /// 338 /// This returns true if there were any known values. 339 /// 340 bool JumpThreading:: 341 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result, 342 ConstantPreference Preference) { 343 // This method walks up use-def chains recursively. Because of this, we could 344 // get into an infinite loop going around loops in the use-def chain. To 345 // prevent this, keep track of what (value, block) pairs we've already visited 346 // and terminate the search if we loop back to them 347 if (!RecursionSet.insert(std::make_pair(V, BB)).second) 348 return false; 349 350 // An RAII help to remove this pair from the recursion set once the recursion 351 // stack pops back out again. 352 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB)); 353 354 // If V is a constant, then it is known in all predecessors. 355 if (Constant *KC = getKnownConstant(V, Preference)) { 356 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 357 Result.push_back(std::make_pair(KC, *PI)); 358 359 return true; 360 } 361 362 // If V is a non-instruction value, or an instruction in a different block, 363 // then it can't be derived from a PHI. 364 Instruction *I = dyn_cast<Instruction>(V); 365 if (!I || I->getParent() != BB) { 366 367 // Okay, if this is a live-in value, see if it has a known value at the end 368 // of any of our predecessors. 369 // 370 // FIXME: This should be an edge property, not a block end property. 371 /// TODO: Per PR2563, we could infer value range information about a 372 /// predecessor based on its terminator. 373 // 374 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 375 // "I" is a non-local compare-with-a-constant instruction. This would be 376 // able to handle value inequalities better, for example if the compare is 377 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 378 // Perhaps getConstantOnEdge should be smart enough to do this? 379 380 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 381 BasicBlock *P = *PI; 382 // If the value is known by LazyValueInfo to be a constant in a 383 // predecessor, use that information to try to thread this block. 384 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 385 if (Constant *KC = getKnownConstant(PredCst, Preference)) 386 Result.push_back(std::make_pair(KC, P)); 387 } 388 389 return !Result.empty(); 390 } 391 392 /// If I is a PHI node, then we know the incoming values for any constants. 393 if (PHINode *PN = dyn_cast<PHINode>(I)) { 394 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 395 Value *InVal = PN->getIncomingValue(i); 396 if (Constant *KC = getKnownConstant(InVal, Preference)) { 397 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 398 } else { 399 Constant *CI = LVI->getConstantOnEdge(InVal, 400 PN->getIncomingBlock(i), BB); 401 if (Constant *KC = getKnownConstant(CI, Preference)) 402 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 403 } 404 } 405 406 return !Result.empty(); 407 } 408 409 PredValueInfoTy LHSVals, RHSVals; 410 411 // Handle some boolean conditions. 412 if (I->getType()->getPrimitiveSizeInBits() == 1) { 413 assert(Preference == WantInteger && "One-bit non-integer type?"); 414 // X | true -> true 415 // X & false -> false 416 if (I->getOpcode() == Instruction::Or || 417 I->getOpcode() == Instruction::And) { 418 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 419 WantInteger); 420 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals, 421 WantInteger); 422 423 if (LHSVals.empty() && RHSVals.empty()) 424 return false; 425 426 ConstantInt *InterestingVal; 427 if (I->getOpcode() == Instruction::Or) 428 InterestingVal = ConstantInt::getTrue(I->getContext()); 429 else 430 InterestingVal = ConstantInt::getFalse(I->getContext()); 431 432 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 433 434 // Scan for the sentinel. If we find an undef, force it to the 435 // interesting value: x|undef -> true and x&undef -> false. 436 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 437 if (LHSVals[i].first == InterestingVal || 438 isa<UndefValue>(LHSVals[i].first)) { 439 Result.push_back(LHSVals[i]); 440 Result.back().first = InterestingVal; 441 LHSKnownBBs.insert(LHSVals[i].second); 442 } 443 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 444 if (RHSVals[i].first == InterestingVal || 445 isa<UndefValue>(RHSVals[i].first)) { 446 // If we already inferred a value for this block on the LHS, don't 447 // re-add it. 448 if (!LHSKnownBBs.count(RHSVals[i].second)) { 449 Result.push_back(RHSVals[i]); 450 Result.back().first = InterestingVal; 451 } 452 } 453 454 return !Result.empty(); 455 } 456 457 // Handle the NOT form of XOR. 458 if (I->getOpcode() == Instruction::Xor && 459 isa<ConstantInt>(I->getOperand(1)) && 460 cast<ConstantInt>(I->getOperand(1))->isOne()) { 461 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result, 462 WantInteger); 463 if (Result.empty()) 464 return false; 465 466 // Invert the known values. 467 for (unsigned i = 0, e = Result.size(); i != e; ++i) 468 Result[i].first = ConstantExpr::getNot(Result[i].first); 469 470 return true; 471 } 472 473 // Try to simplify some other binary operator values. 474 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 475 assert(Preference != WantBlockAddress 476 && "A binary operator creating a block address?"); 477 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 478 PredValueInfoTy LHSVals; 479 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals, 480 WantInteger); 481 482 // Try to use constant folding to simplify the binary operator. 483 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 484 Constant *V = LHSVals[i].first; 485 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); 486 487 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 488 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 489 } 490 } 491 492 return !Result.empty(); 493 } 494 495 // Handle compare with phi operand, where the PHI is defined in this block. 496 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 497 assert(Preference == WantInteger && "Compares only produce integers"); 498 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 499 if (PN && PN->getParent() == BB) { 500 // We can do this simplification if any comparisons fold to true or false. 501 // See if any do. 502 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 503 BasicBlock *PredBB = PN->getIncomingBlock(i); 504 Value *LHS = PN->getIncomingValue(i); 505 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 506 507 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL); 508 if (!Res) { 509 if (!isa<Constant>(RHS)) 510 continue; 511 512 LazyValueInfo::Tristate 513 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 514 cast<Constant>(RHS), PredBB, BB); 515 if (ResT == LazyValueInfo::Unknown) 516 continue; 517 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 518 } 519 520 if (Constant *KC = getKnownConstant(Res, WantInteger)) 521 Result.push_back(std::make_pair(KC, PredBB)); 522 } 523 524 return !Result.empty(); 525 } 526 527 528 // If comparing a live-in value against a constant, see if we know the 529 // live-in value on any predecessors. 530 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) { 531 if (!isa<Instruction>(Cmp->getOperand(0)) || 532 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 533 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 534 535 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ 536 BasicBlock *P = *PI; 537 // If the value is known by LazyValueInfo to be a constant in a 538 // predecessor, use that information to try to thread this block. 539 LazyValueInfo::Tristate Res = 540 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 541 RHSCst, P, BB); 542 if (Res == LazyValueInfo::Unknown) 543 continue; 544 545 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 546 Result.push_back(std::make_pair(ResC, P)); 547 } 548 549 return !Result.empty(); 550 } 551 552 // Try to find a constant value for the LHS of a comparison, 553 // and evaluate it statically if we can. 554 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { 555 PredValueInfoTy LHSVals; 556 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 557 WantInteger); 558 559 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 560 Constant *V = LHSVals[i].first; 561 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(), 562 V, CmpConst); 563 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 564 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 565 } 566 567 return !Result.empty(); 568 } 569 } 570 } 571 572 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 573 // Handle select instructions where at least one operand is a known constant 574 // and we can figure out the condition value for any predecessor block. 575 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference); 576 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference); 577 PredValueInfoTy Conds; 578 if ((TrueVal || FalseVal) && 579 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds, 580 WantInteger)) { 581 for (unsigned i = 0, e = Conds.size(); i != e; ++i) { 582 Constant *Cond = Conds[i].first; 583 584 // Figure out what value to use for the condition. 585 bool KnownCond; 586 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) { 587 // A known boolean. 588 KnownCond = CI->isOne(); 589 } else { 590 assert(isa<UndefValue>(Cond) && "Unexpected condition value"); 591 // Either operand will do, so be sure to pick the one that's a known 592 // constant. 593 // FIXME: Do this more cleverly if both values are known constants? 594 KnownCond = (TrueVal != nullptr); 595 } 596 597 // See if the select has a known constant value for this predecessor. 598 if (Constant *Val = KnownCond ? TrueVal : FalseVal) 599 Result.push_back(std::make_pair(Val, Conds[i].second)); 600 } 601 602 return !Result.empty(); 603 } 604 } 605 606 // If all else fails, see if LVI can figure out a constant value for us. 607 Constant *CI = LVI->getConstant(V, BB); 608 if (Constant *KC = getKnownConstant(CI, Preference)) { 609 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 610 Result.push_back(std::make_pair(KC, *PI)); 611 } 612 613 return !Result.empty(); 614 } 615 616 617 618 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends 619 /// in an undefined jump, decide which block is best to revector to. 620 /// 621 /// Since we can pick an arbitrary destination, we pick the successor with the 622 /// fewest predecessors. This should reduce the in-degree of the others. 623 /// 624 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 625 TerminatorInst *BBTerm = BB->getTerminator(); 626 unsigned MinSucc = 0; 627 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 628 // Compute the successor with the minimum number of predecessors. 629 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 630 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 631 TestBB = BBTerm->getSuccessor(i); 632 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 633 if (NumPreds < MinNumPreds) { 634 MinSucc = i; 635 MinNumPreds = NumPreds; 636 } 637 } 638 639 return MinSucc; 640 } 641 642 static bool hasAddressTakenAndUsed(BasicBlock *BB) { 643 if (!BB->hasAddressTaken()) return false; 644 645 // If the block has its address taken, it may be a tree of dead constants 646 // hanging off of it. These shouldn't keep the block alive. 647 BlockAddress *BA = BlockAddress::get(BB); 648 BA->removeDeadConstantUsers(); 649 return !BA->use_empty(); 650 } 651 652 /// ProcessBlock - If there are any predecessors whose control can be threaded 653 /// through to a successor, transform them now. 654 bool JumpThreading::ProcessBlock(BasicBlock *BB) { 655 // If the block is trivially dead, just return and let the caller nuke it. 656 // This simplifies other transformations. 657 if (pred_begin(BB) == pred_end(BB) && 658 BB != &BB->getParent()->getEntryBlock()) 659 return false; 660 661 // If this block has a single predecessor, and if that pred has a single 662 // successor, merge the blocks. This encourages recursive jump threading 663 // because now the condition in this block can be threaded through 664 // predecessors of our predecessor block. 665 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 666 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 667 SinglePred != BB && !hasAddressTakenAndUsed(BB)) { 668 // If SinglePred was a loop header, BB becomes one. 669 if (LoopHeaders.erase(SinglePred)) 670 LoopHeaders.insert(BB); 671 672 // Remember if SinglePred was the entry block of the function. If so, we 673 // will need to move BB back to the entry position. 674 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 675 LVI->eraseBlock(SinglePred); 676 MergeBasicBlockIntoOnlyPred(BB); 677 678 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 679 BB->moveBefore(&BB->getParent()->getEntryBlock()); 680 return true; 681 } 682 } 683 684 // What kind of constant we're looking for. 685 ConstantPreference Preference = WantInteger; 686 687 // Look to see if the terminator is a conditional branch, switch or indirect 688 // branch, if not we can't thread it. 689 Value *Condition; 690 Instruction *Terminator = BB->getTerminator(); 691 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { 692 // Can't thread an unconditional jump. 693 if (BI->isUnconditional()) return false; 694 Condition = BI->getCondition(); 695 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { 696 Condition = SI->getCondition(); 697 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { 698 // Can't thread indirect branch with no successors. 699 if (IB->getNumSuccessors() == 0) return false; 700 Condition = IB->getAddress()->stripPointerCasts(); 701 Preference = WantBlockAddress; 702 } else { 703 return false; // Must be an invoke. 704 } 705 706 // Run constant folding to see if we can reduce the condition to a simple 707 // constant. 708 if (Instruction *I = dyn_cast<Instruction>(Condition)) { 709 Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI); 710 if (SimpleVal) { 711 I->replaceAllUsesWith(SimpleVal); 712 I->eraseFromParent(); 713 Condition = SimpleVal; 714 } 715 } 716 717 // If the terminator is branching on an undef, we can pick any of the 718 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 719 if (isa<UndefValue>(Condition)) { 720 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 721 722 // Fold the branch/switch. 723 TerminatorInst *BBTerm = BB->getTerminator(); 724 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 725 if (i == BestSucc) continue; 726 BBTerm->getSuccessor(i)->removePredecessor(BB, true); 727 } 728 729 DEBUG(dbgs() << " In block '" << BB->getName() 730 << "' folding undef terminator: " << *BBTerm << '\n'); 731 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 732 BBTerm->eraseFromParent(); 733 return true; 734 } 735 736 // If the terminator of this block is branching on a constant, simplify the 737 // terminator to an unconditional branch. This can occur due to threading in 738 // other blocks. 739 if (getKnownConstant(Condition, Preference)) { 740 DEBUG(dbgs() << " In block '" << BB->getName() 741 << "' folding terminator: " << *BB->getTerminator() << '\n'); 742 ++NumFolds; 743 ConstantFoldTerminator(BB, true); 744 return true; 745 } 746 747 Instruction *CondInst = dyn_cast<Instruction>(Condition); 748 749 // All the rest of our checks depend on the condition being an instruction. 750 if (!CondInst) { 751 // FIXME: Unify this with code below. 752 if (ProcessThreadableEdges(Condition, BB, Preference)) 753 return true; 754 return false; 755 } 756 757 758 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 759 // For a comparison where the LHS is outside this block, it's possible 760 // that we've branched on it before. Used LVI to see if we can simplify 761 // the branch based on that. 762 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 763 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 764 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 765 if (CondBr && CondConst && CondBr->isConditional() && PI != PE && 766 (!isa<Instruction>(CondCmp->getOperand(0)) || 767 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { 768 // For predecessor edge, determine if the comparison is true or false 769 // on that edge. If they're all true or all false, we can simplify the 770 // branch. 771 // FIXME: We could handle mixed true/false by duplicating code. 772 LazyValueInfo::Tristate Baseline = 773 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0), 774 CondConst, *PI, BB); 775 if (Baseline != LazyValueInfo::Unknown) { 776 // Check that all remaining incoming values match the first one. 777 while (++PI != PE) { 778 LazyValueInfo::Tristate Ret = 779 LVI->getPredicateOnEdge(CondCmp->getPredicate(), 780 CondCmp->getOperand(0), CondConst, *PI, BB); 781 if (Ret != Baseline) break; 782 } 783 784 // If we terminated early, then one of the values didn't match. 785 if (PI == PE) { 786 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0; 787 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1; 788 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); 789 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); 790 CondBr->eraseFromParent(); 791 return true; 792 } 793 } 794 795 } 796 797 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB)) 798 return true; 799 } 800 801 // Check for some cases that are worth simplifying. Right now we want to look 802 // for loads that are used by a switch or by the condition for the branch. If 803 // we see one, check to see if it's partially redundant. If so, insert a PHI 804 // which can then be used to thread the values. 805 // 806 Value *SimplifyValue = CondInst; 807 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 808 if (isa<Constant>(CondCmp->getOperand(1))) 809 SimplifyValue = CondCmp->getOperand(0); 810 811 // TODO: There are other places where load PRE would be profitable, such as 812 // more complex comparisons. 813 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 814 if (SimplifyPartiallyRedundantLoad(LI)) 815 return true; 816 817 818 // Handle a variety of cases where we are branching on something derived from 819 // a PHI node in the current block. If we can prove that any predecessors 820 // compute a predictable value based on a PHI node, thread those predecessors. 821 // 822 if (ProcessThreadableEdges(CondInst, BB, Preference)) 823 return true; 824 825 // If this is an otherwise-unfoldable branch on a phi node in the current 826 // block, see if we can simplify. 827 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 828 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 829 return ProcessBranchOnPHI(PN); 830 831 832 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 833 if (CondInst->getOpcode() == Instruction::Xor && 834 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 835 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 836 837 838 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 839 // "(X == 4)", thread through this block. 840 841 return false; 842 } 843 844 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 845 /// load instruction, eliminate it by replacing it with a PHI node. This is an 846 /// important optimization that encourages jump threading, and needs to be run 847 /// interlaced with other jump threading tasks. 848 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 849 // Don't hack volatile/atomic loads. 850 if (!LI->isSimple()) return false; 851 852 // If the load is defined in a block with exactly one predecessor, it can't be 853 // partially redundant. 854 BasicBlock *LoadBB = LI->getParent(); 855 if (LoadBB->getSinglePredecessor()) 856 return false; 857 858 // If the load is defined in a landing pad, it can't be partially redundant, 859 // because the edges between the invoke and the landing pad cannot have other 860 // instructions between them. 861 if (LoadBB->isLandingPad()) 862 return false; 863 864 Value *LoadedPtr = LI->getOperand(0); 865 866 // If the loaded operand is defined in the LoadBB, it can't be available. 867 // TODO: Could do simple PHI translation, that would be fun :) 868 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 869 if (PtrOp->getParent() == LoadBB) 870 return false; 871 872 // Scan a few instructions up from the load, to see if it is obviously live at 873 // the entry to its block. 874 BasicBlock::iterator BBIt = LI; 875 876 if (Value *AvailableVal = 877 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 878 // If the value if the load is locally available within the block, just use 879 // it. This frequently occurs for reg2mem'd allocas. 880 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 881 882 // If the returned value is the load itself, replace with an undef. This can 883 // only happen in dead loops. 884 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 885 LI->replaceAllUsesWith(AvailableVal); 886 LI->eraseFromParent(); 887 return true; 888 } 889 890 // Otherwise, if we scanned the whole block and got to the top of the block, 891 // we know the block is locally transparent to the load. If not, something 892 // might clobber its value. 893 if (BBIt != LoadBB->begin()) 894 return false; 895 896 // If all of the loads and stores that feed the value have the same TBAA tag, 897 // then we can propagate it onto any newly inserted loads. 898 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa); 899 900 SmallPtrSet<BasicBlock*, 8> PredsScanned; 901 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 902 AvailablePredsTy AvailablePreds; 903 BasicBlock *OneUnavailablePred = nullptr; 904 905 // If we got here, the loaded value is transparent through to the start of the 906 // block. Check to see if it is available in any of the predecessor blocks. 907 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 908 PI != PE; ++PI) { 909 BasicBlock *PredBB = *PI; 910 911 // If we already scanned this predecessor, skip it. 912 if (!PredsScanned.insert(PredBB)) 913 continue; 914 915 // Scan the predecessor to see if the value is available in the pred. 916 BBIt = PredBB->end(); 917 MDNode *ThisTBAATag = nullptr; 918 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6, 919 nullptr, &ThisTBAATag); 920 if (!PredAvailable) { 921 OneUnavailablePred = PredBB; 922 continue; 923 } 924 925 // If tbaa tags disagree or are not present, forget about them. 926 if (TBAATag != ThisTBAATag) TBAATag = nullptr; 927 928 // If so, this load is partially redundant. Remember this info so that we 929 // can create a PHI node. 930 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 931 } 932 933 // If the loaded value isn't available in any predecessor, it isn't partially 934 // redundant. 935 if (AvailablePreds.empty()) return false; 936 937 // Okay, the loaded value is available in at least one (and maybe all!) 938 // predecessors. If the value is unavailable in more than one unique 939 // predecessor, we want to insert a merge block for those common predecessors. 940 // This ensures that we only have to insert one reload, thus not increasing 941 // code size. 942 BasicBlock *UnavailablePred = nullptr; 943 944 // If there is exactly one predecessor where the value is unavailable, the 945 // already computed 'OneUnavailablePred' block is it. If it ends in an 946 // unconditional branch, we know that it isn't a critical edge. 947 if (PredsScanned.size() == AvailablePreds.size()+1 && 948 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 949 UnavailablePred = OneUnavailablePred; 950 } else if (PredsScanned.size() != AvailablePreds.size()) { 951 // Otherwise, we had multiple unavailable predecessors or we had a critical 952 // edge from the one. 953 SmallVector<BasicBlock*, 8> PredsToSplit; 954 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 955 956 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 957 AvailablePredSet.insert(AvailablePreds[i].first); 958 959 // Add all the unavailable predecessors to the PredsToSplit list. 960 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 961 PI != PE; ++PI) { 962 BasicBlock *P = *PI; 963 // If the predecessor is an indirect goto, we can't split the edge. 964 if (isa<IndirectBrInst>(P->getTerminator())) 965 return false; 966 967 if (!AvailablePredSet.count(P)) 968 PredsToSplit.push_back(P); 969 } 970 971 // Split them out to their own block. 972 UnavailablePred = 973 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this); 974 } 975 976 // If the value isn't available in all predecessors, then there will be 977 // exactly one where it isn't available. Insert a load on that edge and add 978 // it to the AvailablePreds list. 979 if (UnavailablePred) { 980 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 981 "Can't handle critical edge here!"); 982 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 983 LI->getAlignment(), 984 UnavailablePred->getTerminator()); 985 NewVal->setDebugLoc(LI->getDebugLoc()); 986 if (TBAATag) 987 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag); 988 989 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 990 } 991 992 // Now we know that each predecessor of this block has a value in 993 // AvailablePreds, sort them for efficient access as we're walking the preds. 994 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 995 996 // Create a PHI node at the start of the block for the PRE'd load value. 997 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB); 998 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "", 999 LoadBB->begin()); 1000 PN->takeName(LI); 1001 PN->setDebugLoc(LI->getDebugLoc()); 1002 1003 // Insert new entries into the PHI for each predecessor. A single block may 1004 // have multiple entries here. 1005 for (pred_iterator PI = PB; PI != PE; ++PI) { 1006 BasicBlock *P = *PI; 1007 AvailablePredsTy::iterator I = 1008 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 1009 std::make_pair(P, (Value*)nullptr)); 1010 1011 assert(I != AvailablePreds.end() && I->first == P && 1012 "Didn't find entry for predecessor!"); 1013 1014 PN->addIncoming(I->second, I->first); 1015 } 1016 1017 //cerr << "PRE: " << *LI << *PN << "\n"; 1018 1019 LI->replaceAllUsesWith(PN); 1020 LI->eraseFromParent(); 1021 1022 return true; 1023 } 1024 1025 /// FindMostPopularDest - The specified list contains multiple possible 1026 /// threadable destinations. Pick the one that occurs the most frequently in 1027 /// the list. 1028 static BasicBlock * 1029 FindMostPopularDest(BasicBlock *BB, 1030 const SmallVectorImpl<std::pair<BasicBlock*, 1031 BasicBlock*> > &PredToDestList) { 1032 assert(!PredToDestList.empty()); 1033 1034 // Determine popularity. If there are multiple possible destinations, we 1035 // explicitly choose to ignore 'undef' destinations. We prefer to thread 1036 // blocks with known and real destinations to threading undef. We'll handle 1037 // them later if interesting. 1038 DenseMap<BasicBlock*, unsigned> DestPopularity; 1039 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1040 if (PredToDestList[i].second) 1041 DestPopularity[PredToDestList[i].second]++; 1042 1043 // Find the most popular dest. 1044 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 1045 BasicBlock *MostPopularDest = DPI->first; 1046 unsigned Popularity = DPI->second; 1047 SmallVector<BasicBlock*, 4> SamePopularity; 1048 1049 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 1050 // If the popularity of this entry isn't higher than the popularity we've 1051 // seen so far, ignore it. 1052 if (DPI->second < Popularity) 1053 ; // ignore. 1054 else if (DPI->second == Popularity) { 1055 // If it is the same as what we've seen so far, keep track of it. 1056 SamePopularity.push_back(DPI->first); 1057 } else { 1058 // If it is more popular, remember it. 1059 SamePopularity.clear(); 1060 MostPopularDest = DPI->first; 1061 Popularity = DPI->second; 1062 } 1063 } 1064 1065 // Okay, now we know the most popular destination. If there is more than one 1066 // destination, we need to determine one. This is arbitrary, but we need 1067 // to make a deterministic decision. Pick the first one that appears in the 1068 // successor list. 1069 if (!SamePopularity.empty()) { 1070 SamePopularity.push_back(MostPopularDest); 1071 TerminatorInst *TI = BB->getTerminator(); 1072 for (unsigned i = 0; ; ++i) { 1073 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1074 1075 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1076 TI->getSuccessor(i)) == SamePopularity.end()) 1077 continue; 1078 1079 MostPopularDest = TI->getSuccessor(i); 1080 break; 1081 } 1082 } 1083 1084 // Okay, we have finally picked the most popular destination. 1085 return MostPopularDest; 1086 } 1087 1088 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 1089 ConstantPreference Preference) { 1090 // If threading this would thread across a loop header, don't even try to 1091 // thread the edge. 1092 if (LoopHeaders.count(BB)) 1093 return false; 1094 1095 PredValueInfoTy PredValues; 1096 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference)) 1097 return false; 1098 1099 assert(!PredValues.empty() && 1100 "ComputeValueKnownInPredecessors returned true with no values"); 1101 1102 DEBUG(dbgs() << "IN BB: " << *BB; 1103 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1104 dbgs() << " BB '" << BB->getName() << "': FOUND condition = " 1105 << *PredValues[i].first 1106 << " for pred '" << PredValues[i].second->getName() << "'.\n"; 1107 }); 1108 1109 // Decide what we want to thread through. Convert our list of known values to 1110 // a list of known destinations for each pred. This also discards duplicate 1111 // predecessors and keeps track of the undefined inputs (which are represented 1112 // as a null dest in the PredToDestList). 1113 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1114 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1115 1116 BasicBlock *OnlyDest = nullptr; 1117 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1118 1119 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1120 BasicBlock *Pred = PredValues[i].second; 1121 if (!SeenPreds.insert(Pred)) 1122 continue; // Duplicate predecessor entry. 1123 1124 // If the predecessor ends with an indirect goto, we can't change its 1125 // destination. 1126 if (isa<IndirectBrInst>(Pred->getTerminator())) 1127 continue; 1128 1129 Constant *Val = PredValues[i].first; 1130 1131 BasicBlock *DestBB; 1132 if (isa<UndefValue>(Val)) 1133 DestBB = nullptr; 1134 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1135 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); 1136 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 1137 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor(); 1138 } else { 1139 assert(isa<IndirectBrInst>(BB->getTerminator()) 1140 && "Unexpected terminator"); 1141 DestBB = cast<BlockAddress>(Val)->getBasicBlock(); 1142 } 1143 1144 // If we have exactly one destination, remember it for efficiency below. 1145 if (PredToDestList.empty()) 1146 OnlyDest = DestBB; 1147 else if (OnlyDest != DestBB) 1148 OnlyDest = MultipleDestSentinel; 1149 1150 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1151 } 1152 1153 // If all edges were unthreadable, we fail. 1154 if (PredToDestList.empty()) 1155 return false; 1156 1157 // Determine which is the most common successor. If we have many inputs and 1158 // this block is a switch, we want to start by threading the batch that goes 1159 // to the most popular destination first. If we only know about one 1160 // threadable destination (the common case) we can avoid this. 1161 BasicBlock *MostPopularDest = OnlyDest; 1162 1163 if (MostPopularDest == MultipleDestSentinel) 1164 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1165 1166 // Now that we know what the most popular destination is, factor all 1167 // predecessors that will jump to it into a single predecessor. 1168 SmallVector<BasicBlock*, 16> PredsToFactor; 1169 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1170 if (PredToDestList[i].second == MostPopularDest) { 1171 BasicBlock *Pred = PredToDestList[i].first; 1172 1173 // This predecessor may be a switch or something else that has multiple 1174 // edges to the block. Factor each of these edges by listing them 1175 // according to # occurrences in PredsToFactor. 1176 TerminatorInst *PredTI = Pred->getTerminator(); 1177 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1178 if (PredTI->getSuccessor(i) == BB) 1179 PredsToFactor.push_back(Pred); 1180 } 1181 1182 // If the threadable edges are branching on an undefined value, we get to pick 1183 // the destination that these predecessors should get to. 1184 if (!MostPopularDest) 1185 MostPopularDest = BB->getTerminator()-> 1186 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1187 1188 // Ok, try to thread it! 1189 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1190 } 1191 1192 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1193 /// a PHI node in the current block. See if there are any simplifications we 1194 /// can do based on inputs to the phi node. 1195 /// 1196 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1197 BasicBlock *BB = PN->getParent(); 1198 1199 // TODO: We could make use of this to do it once for blocks with common PHI 1200 // values. 1201 SmallVector<BasicBlock*, 1> PredBBs; 1202 PredBBs.resize(1); 1203 1204 // If any of the predecessor blocks end in an unconditional branch, we can 1205 // *duplicate* the conditional branch into that block in order to further 1206 // encourage jump threading and to eliminate cases where we have branch on a 1207 // phi of an icmp (branch on icmp is much better). 1208 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1209 BasicBlock *PredBB = PN->getIncomingBlock(i); 1210 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1211 if (PredBr->isUnconditional()) { 1212 PredBBs[0] = PredBB; 1213 // Try to duplicate BB into PredBB. 1214 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1215 return true; 1216 } 1217 } 1218 1219 return false; 1220 } 1221 1222 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1223 /// a xor instruction in the current block. See if there are any 1224 /// simplifications we can do based on inputs to the xor. 1225 /// 1226 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1227 BasicBlock *BB = BO->getParent(); 1228 1229 // If either the LHS or RHS of the xor is a constant, don't do this 1230 // optimization. 1231 if (isa<ConstantInt>(BO->getOperand(0)) || 1232 isa<ConstantInt>(BO->getOperand(1))) 1233 return false; 1234 1235 // If the first instruction in BB isn't a phi, we won't be able to infer 1236 // anything special about any particular predecessor. 1237 if (!isa<PHINode>(BB->front())) 1238 return false; 1239 1240 // If we have a xor as the branch input to this block, and we know that the 1241 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1242 // the condition into the predecessor and fix that value to true, saving some 1243 // logical ops on that path and encouraging other paths to simplify. 1244 // 1245 // This copies something like this: 1246 // 1247 // BB: 1248 // %X = phi i1 [1], [%X'] 1249 // %Y = icmp eq i32 %A, %B 1250 // %Z = xor i1 %X, %Y 1251 // br i1 %Z, ... 1252 // 1253 // Into: 1254 // BB': 1255 // %Y = icmp ne i32 %A, %B 1256 // br i1 %Z, ... 1257 1258 PredValueInfoTy XorOpValues; 1259 bool isLHS = true; 1260 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, 1261 WantInteger)) { 1262 assert(XorOpValues.empty()); 1263 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, 1264 WantInteger)) 1265 return false; 1266 isLHS = false; 1267 } 1268 1269 assert(!XorOpValues.empty() && 1270 "ComputeValueKnownInPredecessors returned true with no values"); 1271 1272 // Scan the information to see which is most popular: true or false. The 1273 // predecessors can be of the set true, false, or undef. 1274 unsigned NumTrue = 0, NumFalse = 0; 1275 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1276 if (isa<UndefValue>(XorOpValues[i].first)) 1277 // Ignore undefs for the count. 1278 continue; 1279 if (cast<ConstantInt>(XorOpValues[i].first)->isZero()) 1280 ++NumFalse; 1281 else 1282 ++NumTrue; 1283 } 1284 1285 // Determine which value to split on, true, false, or undef if neither. 1286 ConstantInt *SplitVal = nullptr; 1287 if (NumTrue > NumFalse) 1288 SplitVal = ConstantInt::getTrue(BB->getContext()); 1289 else if (NumTrue != 0 || NumFalse != 0) 1290 SplitVal = ConstantInt::getFalse(BB->getContext()); 1291 1292 // Collect all of the blocks that this can be folded into so that we can 1293 // factor this once and clone it once. 1294 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1295 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1296 if (XorOpValues[i].first != SplitVal && 1297 !isa<UndefValue>(XorOpValues[i].first)) 1298 continue; 1299 1300 BlocksToFoldInto.push_back(XorOpValues[i].second); 1301 } 1302 1303 // If we inferred a value for all of the predecessors, then duplication won't 1304 // help us. However, we can just replace the LHS or RHS with the constant. 1305 if (BlocksToFoldInto.size() == 1306 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1307 if (!SplitVal) { 1308 // If all preds provide undef, just nuke the xor, because it is undef too. 1309 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1310 BO->eraseFromParent(); 1311 } else if (SplitVal->isZero()) { 1312 // If all preds provide 0, replace the xor with the other input. 1313 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1314 BO->eraseFromParent(); 1315 } else { 1316 // If all preds provide 1, set the computed value to 1. 1317 BO->setOperand(!isLHS, SplitVal); 1318 } 1319 1320 return true; 1321 } 1322 1323 // Try to duplicate BB into PredBB. 1324 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1325 } 1326 1327 1328 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1329 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1330 /// NewPred using the entries from OldPred (suitably mapped). 1331 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1332 BasicBlock *OldPred, 1333 BasicBlock *NewPred, 1334 DenseMap<Instruction*, Value*> &ValueMap) { 1335 for (BasicBlock::iterator PNI = PHIBB->begin(); 1336 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1337 // Ok, we have a PHI node. Figure out what the incoming value was for the 1338 // DestBlock. 1339 Value *IV = PN->getIncomingValueForBlock(OldPred); 1340 1341 // Remap the value if necessary. 1342 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1343 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1344 if (I != ValueMap.end()) 1345 IV = I->second; 1346 } 1347 1348 PN->addIncoming(IV, NewPred); 1349 } 1350 } 1351 1352 /// ThreadEdge - We have decided that it is safe and profitable to factor the 1353 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1354 /// across BB. Transform the IR to reflect this change. 1355 bool JumpThreading::ThreadEdge(BasicBlock *BB, 1356 const SmallVectorImpl<BasicBlock*> &PredBBs, 1357 BasicBlock *SuccBB) { 1358 // If threading to the same block as we come from, we would infinite loop. 1359 if (SuccBB == BB) { 1360 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1361 << "' - would thread to self!\n"); 1362 return false; 1363 } 1364 1365 // If threading this would thread across a loop header, don't thread the edge. 1366 // See the comments above FindLoopHeaders for justifications and caveats. 1367 if (LoopHeaders.count(BB)) { 1368 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1369 << "' to dest BB '" << SuccBB->getName() 1370 << "' - it might create an irreducible loop!\n"); 1371 return false; 1372 } 1373 1374 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold); 1375 if (JumpThreadCost > Threshold) { 1376 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1377 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1378 return false; 1379 } 1380 1381 // And finally, do it! Start by factoring the predecessors is needed. 1382 BasicBlock *PredBB; 1383 if (PredBBs.size() == 1) 1384 PredBB = PredBBs[0]; 1385 else { 1386 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1387 << " common predecessors.\n"); 1388 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); 1389 } 1390 1391 // And finally, do it! 1392 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1393 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1394 << ", across block:\n " 1395 << *BB << "\n"); 1396 1397 LVI->threadEdge(PredBB, BB, SuccBB); 1398 1399 // We are going to have to map operands from the original BB block to the new 1400 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1401 // account for entry from PredBB. 1402 DenseMap<Instruction*, Value*> ValueMapping; 1403 1404 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1405 BB->getName()+".thread", 1406 BB->getParent(), BB); 1407 NewBB->moveAfter(PredBB); 1408 1409 BasicBlock::iterator BI = BB->begin(); 1410 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1411 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1412 1413 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1414 // mapping and using it to remap operands in the cloned instructions. 1415 for (; !isa<TerminatorInst>(BI); ++BI) { 1416 Instruction *New = BI->clone(); 1417 New->setName(BI->getName()); 1418 NewBB->getInstList().push_back(New); 1419 ValueMapping[BI] = New; 1420 1421 // Remap operands to patch up intra-block references. 1422 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1423 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1424 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1425 if (I != ValueMapping.end()) 1426 New->setOperand(i, I->second); 1427 } 1428 } 1429 1430 // We didn't copy the terminator from BB over to NewBB, because there is now 1431 // an unconditional jump to SuccBB. Insert the unconditional jump. 1432 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB); 1433 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc()); 1434 1435 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1436 // PHI nodes for NewBB now. 1437 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1438 1439 // If there were values defined in BB that are used outside the block, then we 1440 // now have to update all uses of the value to use either the original value, 1441 // the cloned value, or some PHI derived value. This can require arbitrary 1442 // PHI insertion, of which we are prepared to do, clean these up now. 1443 SSAUpdater SSAUpdate; 1444 SmallVector<Use*, 16> UsesToRename; 1445 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1446 // Scan all uses of this instruction to see if it is used outside of its 1447 // block, and if so, record them in UsesToRename. 1448 for (Use &U : I->uses()) { 1449 Instruction *User = cast<Instruction>(U.getUser()); 1450 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1451 if (UserPN->getIncomingBlock(U) == BB) 1452 continue; 1453 } else if (User->getParent() == BB) 1454 continue; 1455 1456 UsesToRename.push_back(&U); 1457 } 1458 1459 // If there are no uses outside the block, we're done with this instruction. 1460 if (UsesToRename.empty()) 1461 continue; 1462 1463 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1464 1465 // We found a use of I outside of BB. Rename all uses of I that are outside 1466 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1467 // with the two values we know. 1468 SSAUpdate.Initialize(I->getType(), I->getName()); 1469 SSAUpdate.AddAvailableValue(BB, I); 1470 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1471 1472 while (!UsesToRename.empty()) 1473 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1474 DEBUG(dbgs() << "\n"); 1475 } 1476 1477 1478 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1479 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1480 // us to simplify any PHI nodes in BB. 1481 TerminatorInst *PredTerm = PredBB->getTerminator(); 1482 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1483 if (PredTerm->getSuccessor(i) == BB) { 1484 BB->removePredecessor(PredBB, true); 1485 PredTerm->setSuccessor(i, NewBB); 1486 } 1487 1488 // At this point, the IR is fully up to date and consistent. Do a quick scan 1489 // over the new instructions and zap any that are constants or dead. This 1490 // frequently happens because of phi translation. 1491 SimplifyInstructionsInBlock(NewBB, DL, TLI); 1492 1493 // Threaded an edge! 1494 ++NumThreads; 1495 return true; 1496 } 1497 1498 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1499 /// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1500 /// If we can duplicate the contents of BB up into PredBB do so now, this 1501 /// improves the odds that the branch will be on an analyzable instruction like 1502 /// a compare. 1503 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1504 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1505 assert(!PredBBs.empty() && "Can't handle an empty set"); 1506 1507 // If BB is a loop header, then duplicating this block outside the loop would 1508 // cause us to transform this into an irreducible loop, don't do this. 1509 // See the comments above FindLoopHeaders for justifications and caveats. 1510 if (LoopHeaders.count(BB)) { 1511 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1512 << "' into predecessor block '" << PredBBs[0]->getName() 1513 << "' - it might create an irreducible loop!\n"); 1514 return false; 1515 } 1516 1517 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold); 1518 if (DuplicationCost > Threshold) { 1519 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1520 << "' - Cost is too high: " << DuplicationCost << "\n"); 1521 return false; 1522 } 1523 1524 // And finally, do it! Start by factoring the predecessors is needed. 1525 BasicBlock *PredBB; 1526 if (PredBBs.size() == 1) 1527 PredBB = PredBBs[0]; 1528 else { 1529 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1530 << " common predecessors.\n"); 1531 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); 1532 } 1533 1534 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1535 // of PredBB. 1536 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1537 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1538 << DuplicationCost << " block is:" << *BB << "\n"); 1539 1540 // Unless PredBB ends with an unconditional branch, split the edge so that we 1541 // can just clone the bits from BB into the end of the new PredBB. 1542 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1543 1544 if (!OldPredBranch || !OldPredBranch->isUnconditional()) { 1545 PredBB = SplitEdge(PredBB, BB, this); 1546 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1547 } 1548 1549 // We are going to have to map operands from the original BB block into the 1550 // PredBB block. Evaluate PHI nodes in BB. 1551 DenseMap<Instruction*, Value*> ValueMapping; 1552 1553 BasicBlock::iterator BI = BB->begin(); 1554 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1555 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1556 1557 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1558 // mapping and using it to remap operands in the cloned instructions. 1559 for (; BI != BB->end(); ++BI) { 1560 Instruction *New = BI->clone(); 1561 1562 // Remap operands to patch up intra-block references. 1563 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1564 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1565 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1566 if (I != ValueMapping.end()) 1567 New->setOperand(i, I->second); 1568 } 1569 1570 // If this instruction can be simplified after the operands are updated, 1571 // just use the simplified value instead. This frequently happens due to 1572 // phi translation. 1573 if (Value *IV = SimplifyInstruction(New, DL)) { 1574 delete New; 1575 ValueMapping[BI] = IV; 1576 } else { 1577 // Otherwise, insert the new instruction into the block. 1578 New->setName(BI->getName()); 1579 PredBB->getInstList().insert(OldPredBranch, New); 1580 ValueMapping[BI] = New; 1581 } 1582 } 1583 1584 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1585 // add entries to the PHI nodes for branch from PredBB now. 1586 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1587 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1588 ValueMapping); 1589 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1590 ValueMapping); 1591 1592 // If there were values defined in BB that are used outside the block, then we 1593 // now have to update all uses of the value to use either the original value, 1594 // the cloned value, or some PHI derived value. This can require arbitrary 1595 // PHI insertion, of which we are prepared to do, clean these up now. 1596 SSAUpdater SSAUpdate; 1597 SmallVector<Use*, 16> UsesToRename; 1598 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1599 // Scan all uses of this instruction to see if it is used outside of its 1600 // block, and if so, record them in UsesToRename. 1601 for (Use &U : I->uses()) { 1602 Instruction *User = cast<Instruction>(U.getUser()); 1603 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1604 if (UserPN->getIncomingBlock(U) == BB) 1605 continue; 1606 } else if (User->getParent() == BB) 1607 continue; 1608 1609 UsesToRename.push_back(&U); 1610 } 1611 1612 // If there are no uses outside the block, we're done with this instruction. 1613 if (UsesToRename.empty()) 1614 continue; 1615 1616 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1617 1618 // We found a use of I outside of BB. Rename all uses of I that are outside 1619 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1620 // with the two values we know. 1621 SSAUpdate.Initialize(I->getType(), I->getName()); 1622 SSAUpdate.AddAvailableValue(BB, I); 1623 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1624 1625 while (!UsesToRename.empty()) 1626 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1627 DEBUG(dbgs() << "\n"); 1628 } 1629 1630 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1631 // that we nuked. 1632 BB->removePredecessor(PredBB, true); 1633 1634 // Remove the unconditional branch at the end of the PredBB block. 1635 OldPredBranch->eraseFromParent(); 1636 1637 ++NumDupes; 1638 return true; 1639 } 1640 1641 /// TryToUnfoldSelect - Look for blocks of the form 1642 /// bb1: 1643 /// %a = select 1644 /// br bb 1645 /// 1646 /// bb2: 1647 /// %p = phi [%a, %bb] ... 1648 /// %c = icmp %p 1649 /// br i1 %c 1650 /// 1651 /// And expand the select into a branch structure if one of its arms allows %c 1652 /// to be folded. This later enables threading from bb1 over bb2. 1653 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) { 1654 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 1655 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0)); 1656 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1)); 1657 1658 if (!CondBr || !CondBr->isConditional() || !CondLHS || 1659 CondLHS->getParent() != BB) 1660 return false; 1661 1662 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) { 1663 BasicBlock *Pred = CondLHS->getIncomingBlock(I); 1664 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I)); 1665 1666 // Look if one of the incoming values is a select in the corresponding 1667 // predecessor. 1668 if (!SI || SI->getParent() != Pred || !SI->hasOneUse()) 1669 continue; 1670 1671 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator()); 1672 if (!PredTerm || !PredTerm->isUnconditional()) 1673 continue; 1674 1675 // Now check if one of the select values would allow us to constant fold the 1676 // terminator in BB. We don't do the transform if both sides fold, those 1677 // cases will be threaded in any case. 1678 LazyValueInfo::Tristate LHSFolds = 1679 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1), 1680 CondRHS, Pred, BB); 1681 LazyValueInfo::Tristate RHSFolds = 1682 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2), 1683 CondRHS, Pred, BB); 1684 if ((LHSFolds != LazyValueInfo::Unknown || 1685 RHSFolds != LazyValueInfo::Unknown) && 1686 LHSFolds != RHSFolds) { 1687 // Expand the select. 1688 // 1689 // Pred -- 1690 // | v 1691 // | NewBB 1692 // | | 1693 // |----- 1694 // v 1695 // BB 1696 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold", 1697 BB->getParent(), BB); 1698 // Move the unconditional branch to NewBB. 1699 PredTerm->removeFromParent(); 1700 NewBB->getInstList().insert(NewBB->end(), PredTerm); 1701 // Create a conditional branch and update PHI nodes. 1702 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred); 1703 CondLHS->setIncomingValue(I, SI->getFalseValue()); 1704 CondLHS->addIncoming(SI->getTrueValue(), NewBB); 1705 // The select is now dead. 1706 SI->eraseFromParent(); 1707 1708 // Update any other PHI nodes in BB. 1709 for (BasicBlock::iterator BI = BB->begin(); 1710 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI) 1711 if (Phi != CondLHS) 1712 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB); 1713 return true; 1714 } 1715 } 1716 return false; 1717 } 1718