JumpThreading.cpp revision dce4a407a24b04eebc6a376f8e62b41aaa7b071f
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" 38using namespace llvm; 39 40#define DEBUG_TYPE "jump-threading" 41 42STATISTIC(NumThreads, "Number of jumps threaded"); 43STATISTIC(NumFolds, "Number of terminators folded"); 44STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 45 46static cl::opt<unsigned> 47Threshold("jump-threading-threshold", 48 cl::desc("Max block size to duplicate for jump threading"), 49 cl::init(6), cl::Hidden); 50 51namespace { 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 138char JumpThreading::ID = 0; 139INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", 140 "Jump Threading", false, false) 141INITIALIZE_PASS_DEPENDENCY(LazyValueInfo) 142INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 143INITIALIZE_PASS_END(JumpThreading, "jump-threading", 144 "Jump Threading", false, false) 145 146// Public interface to the Jump Threading pass 147FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 148 149/// runOnFunction - Top level algorithm. 150/// 151bool 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 FindLoopHeaders(F); 162 163 bool Changed, EverChanged = false; 164 do { 165 Changed = false; 166 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 167 BasicBlock *BB = I; 168 // Thread all of the branches we can over this block. 169 while (ProcessBlock(BB)) 170 Changed = true; 171 172 ++I; 173 174 // If the block is trivially dead, zap it. This eliminates the successor 175 // edges which simplifies the CFG. 176 if (pred_begin(BB) == pred_end(BB) && 177 BB != &BB->getParent()->getEntryBlock()) { 178 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 179 << "' with terminator: " << *BB->getTerminator() << '\n'); 180 LoopHeaders.erase(BB); 181 LVI->eraseBlock(BB); 182 DeleteDeadBlock(BB); 183 Changed = true; 184 continue; 185 } 186 187 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 188 189 // Can't thread an unconditional jump, but if the block is "almost 190 // empty", we can replace uses of it with uses of the successor and make 191 // this dead. 192 if (BI && BI->isUnconditional() && 193 BB != &BB->getParent()->getEntryBlock() && 194 // If the terminator is the only non-phi instruction, try to nuke it. 195 BB->getFirstNonPHIOrDbg()->isTerminator()) { 196 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 197 // block, we have to make sure it isn't in the LoopHeaders set. We 198 // reinsert afterward if needed. 199 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 200 BasicBlock *Succ = BI->getSuccessor(0); 201 202 // FIXME: It is always conservatively correct to drop the info 203 // for a block even if it doesn't get erased. This isn't totally 204 // awesome, but it allows us to use AssertingVH to prevent nasty 205 // dangling pointer issues within LazyValueInfo. 206 LVI->eraseBlock(BB); 207 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 208 Changed = true; 209 // If we deleted BB and BB was the header of a loop, then the 210 // successor is now the header of the loop. 211 BB = Succ; 212 } 213 214 if (ErasedFromLoopHeaders) 215 LoopHeaders.insert(BB); 216 } 217 } 218 EverChanged |= Changed; 219 } while (Changed); 220 221 LoopHeaders.clear(); 222 return EverChanged; 223} 224 225/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 226/// thread across it. Stop scanning the block when passing the threshold. 227static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB, 228 unsigned Threshold) { 229 /// Ignore PHI nodes, these will be flattened when duplication happens. 230 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 231 232 // FIXME: THREADING will delete values that are just used to compute the 233 // branch, so they shouldn't count against the duplication cost. 234 235 // Sum up the cost of each instruction until we get to the terminator. Don't 236 // include the terminator because the copy won't include it. 237 unsigned Size = 0; 238 for (; !isa<TerminatorInst>(I); ++I) { 239 240 // Stop scanning the block if we've reached the threshold. 241 if (Size > Threshold) 242 return Size; 243 244 // Debugger intrinsics don't incur code size. 245 if (isa<DbgInfoIntrinsic>(I)) continue; 246 247 // If this is a pointer->pointer bitcast, it is free. 248 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 249 continue; 250 251 // All other instructions count for at least one unit. 252 ++Size; 253 254 // Calls are more expensive. If they are non-intrinsic calls, we model them 255 // as having cost of 4. If they are a non-vector intrinsic, we model them 256 // as having cost of 2 total, and if they are a vector intrinsic, we model 257 // them as having cost 1. 258 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 259 if (CI->cannotDuplicate()) 260 // Blocks with NoDuplicate are modelled as having infinite cost, so they 261 // are never duplicated. 262 return ~0U; 263 else if (!isa<IntrinsicInst>(CI)) 264 Size += 3; 265 else if (!CI->getType()->isVectorTy()) 266 Size += 1; 267 } 268 } 269 270 // Threading through a switch statement is particularly profitable. If this 271 // block ends in a switch, decrease its cost to make it more likely to happen. 272 if (isa<SwitchInst>(I)) 273 Size = Size > 6 ? Size-6 : 0; 274 275 // The same holds for indirect branches, but slightly more so. 276 if (isa<IndirectBrInst>(I)) 277 Size = Size > 8 ? Size-8 : 0; 278 279 return Size; 280} 281 282/// FindLoopHeaders - We do not want jump threading to turn proper loop 283/// structures into irreducible loops. Doing this breaks up the loop nesting 284/// hierarchy and pessimizes later transformations. To prevent this from 285/// happening, we first have to find the loop headers. Here we approximate this 286/// by finding targets of backedges in the CFG. 287/// 288/// Note that there definitely are cases when we want to allow threading of 289/// edges across a loop header. For example, threading a jump from outside the 290/// loop (the preheader) to an exit block of the loop is definitely profitable. 291/// It is also almost always profitable to thread backedges from within the loop 292/// to exit blocks, and is often profitable to thread backedges to other blocks 293/// within the loop (forming a nested loop). This simple analysis is not rich 294/// enough to track all of these properties and keep it up-to-date as the CFG 295/// mutates, so we don't allow any of these transformations. 296/// 297void JumpThreading::FindLoopHeaders(Function &F) { 298 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 299 FindFunctionBackedges(F, Edges); 300 301 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 302 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 303} 304 305/// getKnownConstant - Helper method to determine if we can thread over a 306/// terminator with the given value as its condition, and if so what value to 307/// use for that. What kind of value this is depends on whether we want an 308/// integer or a block address, but an undef is always accepted. 309/// Returns null if Val is null or not an appropriate constant. 310static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { 311 if (!Val) 312 return nullptr; 313 314 // Undef is "known" enough. 315 if (UndefValue *U = dyn_cast<UndefValue>(Val)) 316 return U; 317 318 if (Preference == WantBlockAddress) 319 return dyn_cast<BlockAddress>(Val->stripPointerCasts()); 320 321 return dyn_cast<ConstantInt>(Val); 322} 323 324/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 325/// if we can infer that the value is a known ConstantInt/BlockAddress or undef 326/// in any of our predecessors. If so, return the known list of value and pred 327/// BB in the result vector. 328/// 329/// This returns true if there were any known values. 330/// 331bool JumpThreading:: 332ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result, 333 ConstantPreference Preference) { 334 // This method walks up use-def chains recursively. Because of this, we could 335 // get into an infinite loop going around loops in the use-def chain. To 336 // prevent this, keep track of what (value, block) pairs we've already visited 337 // and terminate the search if we loop back to them 338 if (!RecursionSet.insert(std::make_pair(V, BB)).second) 339 return false; 340 341 // An RAII help to remove this pair from the recursion set once the recursion 342 // stack pops back out again. 343 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB)); 344 345 // If V is a constant, then it is known in all predecessors. 346 if (Constant *KC = getKnownConstant(V, Preference)) { 347 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 348 Result.push_back(std::make_pair(KC, *PI)); 349 350 return true; 351 } 352 353 // If V is a non-instruction value, or an instruction in a different block, 354 // then it can't be derived from a PHI. 355 Instruction *I = dyn_cast<Instruction>(V); 356 if (!I || I->getParent() != BB) { 357 358 // Okay, if this is a live-in value, see if it has a known value at the end 359 // of any of our predecessors. 360 // 361 // FIXME: This should be an edge property, not a block end property. 362 /// TODO: Per PR2563, we could infer value range information about a 363 /// predecessor based on its terminator. 364 // 365 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 366 // "I" is a non-local compare-with-a-constant instruction. This would be 367 // able to handle value inequalities better, for example if the compare is 368 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 369 // Perhaps getConstantOnEdge should be smart enough to do this? 370 371 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 372 BasicBlock *P = *PI; 373 // If the value is known by LazyValueInfo to be a constant in a 374 // predecessor, use that information to try to thread this block. 375 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 376 if (Constant *KC = getKnownConstant(PredCst, Preference)) 377 Result.push_back(std::make_pair(KC, P)); 378 } 379 380 return !Result.empty(); 381 } 382 383 /// If I is a PHI node, then we know the incoming values for any constants. 384 if (PHINode *PN = dyn_cast<PHINode>(I)) { 385 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 386 Value *InVal = PN->getIncomingValue(i); 387 if (Constant *KC = getKnownConstant(InVal, Preference)) { 388 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 389 } else { 390 Constant *CI = LVI->getConstantOnEdge(InVal, 391 PN->getIncomingBlock(i), BB); 392 if (Constant *KC = getKnownConstant(CI, Preference)) 393 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 394 } 395 } 396 397 return !Result.empty(); 398 } 399 400 PredValueInfoTy LHSVals, RHSVals; 401 402 // Handle some boolean conditions. 403 if (I->getType()->getPrimitiveSizeInBits() == 1) { 404 assert(Preference == WantInteger && "One-bit non-integer type?"); 405 // X | true -> true 406 // X & false -> false 407 if (I->getOpcode() == Instruction::Or || 408 I->getOpcode() == Instruction::And) { 409 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 410 WantInteger); 411 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals, 412 WantInteger); 413 414 if (LHSVals.empty() && RHSVals.empty()) 415 return false; 416 417 ConstantInt *InterestingVal; 418 if (I->getOpcode() == Instruction::Or) 419 InterestingVal = ConstantInt::getTrue(I->getContext()); 420 else 421 InterestingVal = ConstantInt::getFalse(I->getContext()); 422 423 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 424 425 // Scan for the sentinel. If we find an undef, force it to the 426 // interesting value: x|undef -> true and x&undef -> false. 427 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 428 if (LHSVals[i].first == InterestingVal || 429 isa<UndefValue>(LHSVals[i].first)) { 430 Result.push_back(LHSVals[i]); 431 Result.back().first = InterestingVal; 432 LHSKnownBBs.insert(LHSVals[i].second); 433 } 434 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 435 if (RHSVals[i].first == InterestingVal || 436 isa<UndefValue>(RHSVals[i].first)) { 437 // If we already inferred a value for this block on the LHS, don't 438 // re-add it. 439 if (!LHSKnownBBs.count(RHSVals[i].second)) { 440 Result.push_back(RHSVals[i]); 441 Result.back().first = InterestingVal; 442 } 443 } 444 445 return !Result.empty(); 446 } 447 448 // Handle the NOT form of XOR. 449 if (I->getOpcode() == Instruction::Xor && 450 isa<ConstantInt>(I->getOperand(1)) && 451 cast<ConstantInt>(I->getOperand(1))->isOne()) { 452 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result, 453 WantInteger); 454 if (Result.empty()) 455 return false; 456 457 // Invert the known values. 458 for (unsigned i = 0, e = Result.size(); i != e; ++i) 459 Result[i].first = ConstantExpr::getNot(Result[i].first); 460 461 return true; 462 } 463 464 // Try to simplify some other binary operator values. 465 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 466 assert(Preference != WantBlockAddress 467 && "A binary operator creating a block address?"); 468 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 469 PredValueInfoTy LHSVals; 470 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals, 471 WantInteger); 472 473 // Try to use constant folding to simplify the binary operator. 474 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 475 Constant *V = LHSVals[i].first; 476 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); 477 478 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 479 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 480 } 481 } 482 483 return !Result.empty(); 484 } 485 486 // Handle compare with phi operand, where the PHI is defined in this block. 487 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 488 assert(Preference == WantInteger && "Compares only produce integers"); 489 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 490 if (PN && PN->getParent() == BB) { 491 // We can do this simplification if any comparisons fold to true or false. 492 // See if any do. 493 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 494 BasicBlock *PredBB = PN->getIncomingBlock(i); 495 Value *LHS = PN->getIncomingValue(i); 496 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 497 498 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL); 499 if (!Res) { 500 if (!isa<Constant>(RHS)) 501 continue; 502 503 LazyValueInfo::Tristate 504 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 505 cast<Constant>(RHS), PredBB, BB); 506 if (ResT == LazyValueInfo::Unknown) 507 continue; 508 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 509 } 510 511 if (Constant *KC = getKnownConstant(Res, WantInteger)) 512 Result.push_back(std::make_pair(KC, PredBB)); 513 } 514 515 return !Result.empty(); 516 } 517 518 519 // If comparing a live-in value against a constant, see if we know the 520 // live-in value on any predecessors. 521 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) { 522 if (!isa<Instruction>(Cmp->getOperand(0)) || 523 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 524 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 525 526 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ 527 BasicBlock *P = *PI; 528 // If the value is known by LazyValueInfo to be a constant in a 529 // predecessor, use that information to try to thread this block. 530 LazyValueInfo::Tristate Res = 531 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 532 RHSCst, P, BB); 533 if (Res == LazyValueInfo::Unknown) 534 continue; 535 536 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 537 Result.push_back(std::make_pair(ResC, P)); 538 } 539 540 return !Result.empty(); 541 } 542 543 // Try to find a constant value for the LHS of a comparison, 544 // and evaluate it statically if we can. 545 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { 546 PredValueInfoTy LHSVals; 547 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 548 WantInteger); 549 550 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 551 Constant *V = LHSVals[i].first; 552 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(), 553 V, CmpConst); 554 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 555 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 556 } 557 558 return !Result.empty(); 559 } 560 } 561 } 562 563 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 564 // Handle select instructions where at least one operand is a known constant 565 // and we can figure out the condition value for any predecessor block. 566 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference); 567 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference); 568 PredValueInfoTy Conds; 569 if ((TrueVal || FalseVal) && 570 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds, 571 WantInteger)) { 572 for (unsigned i = 0, e = Conds.size(); i != e; ++i) { 573 Constant *Cond = Conds[i].first; 574 575 // Figure out what value to use for the condition. 576 bool KnownCond; 577 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) { 578 // A known boolean. 579 KnownCond = CI->isOne(); 580 } else { 581 assert(isa<UndefValue>(Cond) && "Unexpected condition value"); 582 // Either operand will do, so be sure to pick the one that's a known 583 // constant. 584 // FIXME: Do this more cleverly if both values are known constants? 585 KnownCond = (TrueVal != nullptr); 586 } 587 588 // See if the select has a known constant value for this predecessor. 589 if (Constant *Val = KnownCond ? TrueVal : FalseVal) 590 Result.push_back(std::make_pair(Val, Conds[i].second)); 591 } 592 593 return !Result.empty(); 594 } 595 } 596 597 // If all else fails, see if LVI can figure out a constant value for us. 598 Constant *CI = LVI->getConstant(V, BB); 599 if (Constant *KC = getKnownConstant(CI, Preference)) { 600 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 601 Result.push_back(std::make_pair(KC, *PI)); 602 } 603 604 return !Result.empty(); 605} 606 607 608 609/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 610/// in an undefined jump, decide which block is best to revector to. 611/// 612/// Since we can pick an arbitrary destination, we pick the successor with the 613/// fewest predecessors. This should reduce the in-degree of the others. 614/// 615static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 616 TerminatorInst *BBTerm = BB->getTerminator(); 617 unsigned MinSucc = 0; 618 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 619 // Compute the successor with the minimum number of predecessors. 620 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 621 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 622 TestBB = BBTerm->getSuccessor(i); 623 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 624 if (NumPreds < MinNumPreds) { 625 MinSucc = i; 626 MinNumPreds = NumPreds; 627 } 628 } 629 630 return MinSucc; 631} 632 633static bool hasAddressTakenAndUsed(BasicBlock *BB) { 634 if (!BB->hasAddressTaken()) return false; 635 636 // If the block has its address taken, it may be a tree of dead constants 637 // hanging off of it. These shouldn't keep the block alive. 638 BlockAddress *BA = BlockAddress::get(BB); 639 BA->removeDeadConstantUsers(); 640 return !BA->use_empty(); 641} 642 643/// ProcessBlock - If there are any predecessors whose control can be threaded 644/// through to a successor, transform them now. 645bool JumpThreading::ProcessBlock(BasicBlock *BB) { 646 // If the block is trivially dead, just return and let the caller nuke it. 647 // This simplifies other transformations. 648 if (pred_begin(BB) == pred_end(BB) && 649 BB != &BB->getParent()->getEntryBlock()) 650 return false; 651 652 // If this block has a single predecessor, and if that pred has a single 653 // successor, merge the blocks. This encourages recursive jump threading 654 // because now the condition in this block can be threaded through 655 // predecessors of our predecessor block. 656 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 657 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 658 SinglePred != BB && !hasAddressTakenAndUsed(BB)) { 659 // If SinglePred was a loop header, BB becomes one. 660 if (LoopHeaders.erase(SinglePred)) 661 LoopHeaders.insert(BB); 662 663 // Remember if SinglePred was the entry block of the function. If so, we 664 // will need to move BB back to the entry position. 665 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 666 LVI->eraseBlock(SinglePred); 667 MergeBasicBlockIntoOnlyPred(BB); 668 669 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 670 BB->moveBefore(&BB->getParent()->getEntryBlock()); 671 return true; 672 } 673 } 674 675 // What kind of constant we're looking for. 676 ConstantPreference Preference = WantInteger; 677 678 // Look to see if the terminator is a conditional branch, switch or indirect 679 // branch, if not we can't thread it. 680 Value *Condition; 681 Instruction *Terminator = BB->getTerminator(); 682 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { 683 // Can't thread an unconditional jump. 684 if (BI->isUnconditional()) return false; 685 Condition = BI->getCondition(); 686 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { 687 Condition = SI->getCondition(); 688 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { 689 // Can't thread indirect branch with no successors. 690 if (IB->getNumSuccessors() == 0) return false; 691 Condition = IB->getAddress()->stripPointerCasts(); 692 Preference = WantBlockAddress; 693 } else { 694 return false; // Must be an invoke. 695 } 696 697 // Run constant folding to see if we can reduce the condition to a simple 698 // constant. 699 if (Instruction *I = dyn_cast<Instruction>(Condition)) { 700 Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI); 701 if (SimpleVal) { 702 I->replaceAllUsesWith(SimpleVal); 703 I->eraseFromParent(); 704 Condition = SimpleVal; 705 } 706 } 707 708 // If the terminator is branching on an undef, we can pick any of the 709 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 710 if (isa<UndefValue>(Condition)) { 711 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 712 713 // Fold the branch/switch. 714 TerminatorInst *BBTerm = BB->getTerminator(); 715 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 716 if (i == BestSucc) continue; 717 BBTerm->getSuccessor(i)->removePredecessor(BB, true); 718 } 719 720 DEBUG(dbgs() << " In block '" << BB->getName() 721 << "' folding undef terminator: " << *BBTerm << '\n'); 722 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 723 BBTerm->eraseFromParent(); 724 return true; 725 } 726 727 // If the terminator of this block is branching on a constant, simplify the 728 // terminator to an unconditional branch. This can occur due to threading in 729 // other blocks. 730 if (getKnownConstant(Condition, Preference)) { 731 DEBUG(dbgs() << " In block '" << BB->getName() 732 << "' folding terminator: " << *BB->getTerminator() << '\n'); 733 ++NumFolds; 734 ConstantFoldTerminator(BB, true); 735 return true; 736 } 737 738 Instruction *CondInst = dyn_cast<Instruction>(Condition); 739 740 // All the rest of our checks depend on the condition being an instruction. 741 if (!CondInst) { 742 // FIXME: Unify this with code below. 743 if (ProcessThreadableEdges(Condition, BB, Preference)) 744 return true; 745 return false; 746 } 747 748 749 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 750 // For a comparison where the LHS is outside this block, it's possible 751 // that we've branched on it before. Used LVI to see if we can simplify 752 // the branch based on that. 753 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 754 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 755 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 756 if (CondBr && CondConst && CondBr->isConditional() && PI != PE && 757 (!isa<Instruction>(CondCmp->getOperand(0)) || 758 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { 759 // For predecessor edge, determine if the comparison is true or false 760 // on that edge. If they're all true or all false, we can simplify the 761 // branch. 762 // FIXME: We could handle mixed true/false by duplicating code. 763 LazyValueInfo::Tristate Baseline = 764 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0), 765 CondConst, *PI, BB); 766 if (Baseline != LazyValueInfo::Unknown) { 767 // Check that all remaining incoming values match the first one. 768 while (++PI != PE) { 769 LazyValueInfo::Tristate Ret = 770 LVI->getPredicateOnEdge(CondCmp->getPredicate(), 771 CondCmp->getOperand(0), CondConst, *PI, BB); 772 if (Ret != Baseline) break; 773 } 774 775 // If we terminated early, then one of the values didn't match. 776 if (PI == PE) { 777 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0; 778 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1; 779 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); 780 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); 781 CondBr->eraseFromParent(); 782 return true; 783 } 784 } 785 786 } 787 788 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB)) 789 return true; 790 } 791 792 // Check for some cases that are worth simplifying. Right now we want to look 793 // for loads that are used by a switch or by the condition for the branch. If 794 // we see one, check to see if it's partially redundant. If so, insert a PHI 795 // which can then be used to thread the values. 796 // 797 Value *SimplifyValue = CondInst; 798 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 799 if (isa<Constant>(CondCmp->getOperand(1))) 800 SimplifyValue = CondCmp->getOperand(0); 801 802 // TODO: There are other places where load PRE would be profitable, such as 803 // more complex comparisons. 804 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 805 if (SimplifyPartiallyRedundantLoad(LI)) 806 return true; 807 808 809 // Handle a variety of cases where we are branching on something derived from 810 // a PHI node in the current block. If we can prove that any predecessors 811 // compute a predictable value based on a PHI node, thread those predecessors. 812 // 813 if (ProcessThreadableEdges(CondInst, BB, Preference)) 814 return true; 815 816 // If this is an otherwise-unfoldable branch on a phi node in the current 817 // block, see if we can simplify. 818 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 819 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 820 return ProcessBranchOnPHI(PN); 821 822 823 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 824 if (CondInst->getOpcode() == Instruction::Xor && 825 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 826 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 827 828 829 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 830 // "(X == 4)", thread through this block. 831 832 return false; 833} 834 835/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 836/// load instruction, eliminate it by replacing it with a PHI node. This is an 837/// important optimization that encourages jump threading, and needs to be run 838/// interlaced with other jump threading tasks. 839bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 840 // Don't hack volatile/atomic loads. 841 if (!LI->isSimple()) return false; 842 843 // If the load is defined in a block with exactly one predecessor, it can't be 844 // partially redundant. 845 BasicBlock *LoadBB = LI->getParent(); 846 if (LoadBB->getSinglePredecessor()) 847 return false; 848 849 // If the load is defined in a landing pad, it can't be partially redundant, 850 // because the edges between the invoke and the landing pad cannot have other 851 // instructions between them. 852 if (LoadBB->isLandingPad()) 853 return false; 854 855 Value *LoadedPtr = LI->getOperand(0); 856 857 // If the loaded operand is defined in the LoadBB, it can't be available. 858 // TODO: Could do simple PHI translation, that would be fun :) 859 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 860 if (PtrOp->getParent() == LoadBB) 861 return false; 862 863 // Scan a few instructions up from the load, to see if it is obviously live at 864 // the entry to its block. 865 BasicBlock::iterator BBIt = LI; 866 867 if (Value *AvailableVal = 868 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 869 // If the value if the load is locally available within the block, just use 870 // it. This frequently occurs for reg2mem'd allocas. 871 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 872 873 // If the returned value is the load itself, replace with an undef. This can 874 // only happen in dead loops. 875 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 876 LI->replaceAllUsesWith(AvailableVal); 877 LI->eraseFromParent(); 878 return true; 879 } 880 881 // Otherwise, if we scanned the whole block and got to the top of the block, 882 // we know the block is locally transparent to the load. If not, something 883 // might clobber its value. 884 if (BBIt != LoadBB->begin()) 885 return false; 886 887 // If all of the loads and stores that feed the value have the same TBAA tag, 888 // then we can propagate it onto any newly inserted loads. 889 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa); 890 891 SmallPtrSet<BasicBlock*, 8> PredsScanned; 892 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 893 AvailablePredsTy AvailablePreds; 894 BasicBlock *OneUnavailablePred = nullptr; 895 896 // If we got here, the loaded value is transparent through to the start of the 897 // block. Check to see if it is available in any of the predecessor blocks. 898 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 899 PI != PE; ++PI) { 900 BasicBlock *PredBB = *PI; 901 902 // If we already scanned this predecessor, skip it. 903 if (!PredsScanned.insert(PredBB)) 904 continue; 905 906 // Scan the predecessor to see if the value is available in the pred. 907 BBIt = PredBB->end(); 908 MDNode *ThisTBAATag = nullptr; 909 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6, 910 nullptr, &ThisTBAATag); 911 if (!PredAvailable) { 912 OneUnavailablePred = PredBB; 913 continue; 914 } 915 916 // If tbaa tags disagree or are not present, forget about them. 917 if (TBAATag != ThisTBAATag) TBAATag = nullptr; 918 919 // If so, this load is partially redundant. Remember this info so that we 920 // can create a PHI node. 921 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 922 } 923 924 // If the loaded value isn't available in any predecessor, it isn't partially 925 // redundant. 926 if (AvailablePreds.empty()) return false; 927 928 // Okay, the loaded value is available in at least one (and maybe all!) 929 // predecessors. If the value is unavailable in more than one unique 930 // predecessor, we want to insert a merge block for those common predecessors. 931 // This ensures that we only have to insert one reload, thus not increasing 932 // code size. 933 BasicBlock *UnavailablePred = nullptr; 934 935 // If there is exactly one predecessor where the value is unavailable, the 936 // already computed 'OneUnavailablePred' block is it. If it ends in an 937 // unconditional branch, we know that it isn't a critical edge. 938 if (PredsScanned.size() == AvailablePreds.size()+1 && 939 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 940 UnavailablePred = OneUnavailablePred; 941 } else if (PredsScanned.size() != AvailablePreds.size()) { 942 // Otherwise, we had multiple unavailable predecessors or we had a critical 943 // edge from the one. 944 SmallVector<BasicBlock*, 8> PredsToSplit; 945 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 946 947 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 948 AvailablePredSet.insert(AvailablePreds[i].first); 949 950 // Add all the unavailable predecessors to the PredsToSplit list. 951 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 952 PI != PE; ++PI) { 953 BasicBlock *P = *PI; 954 // If the predecessor is an indirect goto, we can't split the edge. 955 if (isa<IndirectBrInst>(P->getTerminator())) 956 return false; 957 958 if (!AvailablePredSet.count(P)) 959 PredsToSplit.push_back(P); 960 } 961 962 // Split them out to their own block. 963 UnavailablePred = 964 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this); 965 } 966 967 // If the value isn't available in all predecessors, then there will be 968 // exactly one where it isn't available. Insert a load on that edge and add 969 // it to the AvailablePreds list. 970 if (UnavailablePred) { 971 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 972 "Can't handle critical edge here!"); 973 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 974 LI->getAlignment(), 975 UnavailablePred->getTerminator()); 976 NewVal->setDebugLoc(LI->getDebugLoc()); 977 if (TBAATag) 978 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag); 979 980 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 981 } 982 983 // Now we know that each predecessor of this block has a value in 984 // AvailablePreds, sort them for efficient access as we're walking the preds. 985 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 986 987 // Create a PHI node at the start of the block for the PRE'd load value. 988 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB); 989 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "", 990 LoadBB->begin()); 991 PN->takeName(LI); 992 PN->setDebugLoc(LI->getDebugLoc()); 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 = PB; PI != PE; ++PI) { 997 BasicBlock *P = *PI; 998 AvailablePredsTy::iterator I = 999 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 1000 std::make_pair(P, (Value*)nullptr)); 1001 1002 assert(I != AvailablePreds.end() && I->first == P && 1003 "Didn't find entry for predecessor!"); 1004 1005 PN->addIncoming(I->second, I->first); 1006 } 1007 1008 //cerr << "PRE: " << *LI << *PN << "\n"; 1009 1010 LI->replaceAllUsesWith(PN); 1011 LI->eraseFromParent(); 1012 1013 return true; 1014} 1015 1016/// FindMostPopularDest - The specified list contains multiple possible 1017/// threadable destinations. Pick the one that occurs the most frequently in 1018/// the list. 1019static BasicBlock * 1020FindMostPopularDest(BasicBlock *BB, 1021 const SmallVectorImpl<std::pair<BasicBlock*, 1022 BasicBlock*> > &PredToDestList) { 1023 assert(!PredToDestList.empty()); 1024 1025 // Determine popularity. If there are multiple possible destinations, we 1026 // explicitly choose to ignore 'undef' destinations. We prefer to thread 1027 // blocks with known and real destinations to threading undef. We'll handle 1028 // them later if interesting. 1029 DenseMap<BasicBlock*, unsigned> DestPopularity; 1030 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1031 if (PredToDestList[i].second) 1032 DestPopularity[PredToDestList[i].second]++; 1033 1034 // Find the most popular dest. 1035 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 1036 BasicBlock *MostPopularDest = DPI->first; 1037 unsigned Popularity = DPI->second; 1038 SmallVector<BasicBlock*, 4> SamePopularity; 1039 1040 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 1041 // If the popularity of this entry isn't higher than the popularity we've 1042 // seen so far, ignore it. 1043 if (DPI->second < Popularity) 1044 ; // ignore. 1045 else if (DPI->second == Popularity) { 1046 // If it is the same as what we've seen so far, keep track of it. 1047 SamePopularity.push_back(DPI->first); 1048 } else { 1049 // If it is more popular, remember it. 1050 SamePopularity.clear(); 1051 MostPopularDest = DPI->first; 1052 Popularity = DPI->second; 1053 } 1054 } 1055 1056 // Okay, now we know the most popular destination. If there is more than one 1057 // destination, we need to determine one. This is arbitrary, but we need 1058 // to make a deterministic decision. Pick the first one that appears in the 1059 // successor list. 1060 if (!SamePopularity.empty()) { 1061 SamePopularity.push_back(MostPopularDest); 1062 TerminatorInst *TI = BB->getTerminator(); 1063 for (unsigned i = 0; ; ++i) { 1064 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1065 1066 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1067 TI->getSuccessor(i)) == SamePopularity.end()) 1068 continue; 1069 1070 MostPopularDest = TI->getSuccessor(i); 1071 break; 1072 } 1073 } 1074 1075 // Okay, we have finally picked the most popular destination. 1076 return MostPopularDest; 1077} 1078 1079bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 1080 ConstantPreference Preference) { 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 PredValueInfoTy PredValues; 1087 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference)) 1088 return false; 1089 1090 assert(!PredValues.empty() && 1091 "ComputeValueKnownInPredecessors returned true with no values"); 1092 1093 DEBUG(dbgs() << "IN BB: " << *BB; 1094 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1095 dbgs() << " BB '" << BB->getName() << "': FOUND condition = " 1096 << *PredValues[i].first 1097 << " for pred '" << PredValues[i].second->getName() << "'.\n"; 1098 }); 1099 1100 // Decide what we want to thread through. Convert our list of known values to 1101 // a list of known destinations for each pred. This also discards duplicate 1102 // predecessors and keeps track of the undefined inputs (which are represented 1103 // as a null dest in the PredToDestList). 1104 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1105 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1106 1107 BasicBlock *OnlyDest = nullptr; 1108 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1109 1110 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1111 BasicBlock *Pred = PredValues[i].second; 1112 if (!SeenPreds.insert(Pred)) 1113 continue; // Duplicate predecessor entry. 1114 1115 // If the predecessor ends with an indirect goto, we can't change its 1116 // destination. 1117 if (isa<IndirectBrInst>(Pred->getTerminator())) 1118 continue; 1119 1120 Constant *Val = PredValues[i].first; 1121 1122 BasicBlock *DestBB; 1123 if (isa<UndefValue>(Val)) 1124 DestBB = nullptr; 1125 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1126 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); 1127 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 1128 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor(); 1129 } else { 1130 assert(isa<IndirectBrInst>(BB->getTerminator()) 1131 && "Unexpected terminator"); 1132 DestBB = cast<BlockAddress>(Val)->getBasicBlock(); 1133 } 1134 1135 // If we have exactly one destination, remember it for efficiency below. 1136 if (PredToDestList.empty()) 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) 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/// 1187bool 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/// 1217bool 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 PredValueInfoTy XorOpValues; 1250 bool isLHS = true; 1251 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, 1252 WantInteger)) { 1253 assert(XorOpValues.empty()); 1254 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, 1255 WantInteger)) 1256 return false; 1257 isLHS = false; 1258 } 1259 1260 assert(!XorOpValues.empty() && 1261 "ComputeValueKnownInPredecessors returned true with no values"); 1262 1263 // Scan the information to see which is most popular: true or false. The 1264 // predecessors can be of the set true, false, or undef. 1265 unsigned NumTrue = 0, NumFalse = 0; 1266 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1267 if (isa<UndefValue>(XorOpValues[i].first)) 1268 // Ignore undefs for the count. 1269 continue; 1270 if (cast<ConstantInt>(XorOpValues[i].first)->isZero()) 1271 ++NumFalse; 1272 else 1273 ++NumTrue; 1274 } 1275 1276 // Determine which value to split on, true, false, or undef if neither. 1277 ConstantInt *SplitVal = nullptr; 1278 if (NumTrue > NumFalse) 1279 SplitVal = ConstantInt::getTrue(BB->getContext()); 1280 else if (NumTrue != 0 || NumFalse != 0) 1281 SplitVal = ConstantInt::getFalse(BB->getContext()); 1282 1283 // Collect all of the blocks that this can be folded into so that we can 1284 // factor this once and clone it once. 1285 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1286 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1287 if (XorOpValues[i].first != SplitVal && 1288 !isa<UndefValue>(XorOpValues[i].first)) 1289 continue; 1290 1291 BlocksToFoldInto.push_back(XorOpValues[i].second); 1292 } 1293 1294 // If we inferred a value for all of the predecessors, then duplication won't 1295 // help us. However, we can just replace the LHS or RHS with the constant. 1296 if (BlocksToFoldInto.size() == 1297 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1298 if (!SplitVal) { 1299 // If all preds provide undef, just nuke the xor, because it is undef too. 1300 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1301 BO->eraseFromParent(); 1302 } else if (SplitVal->isZero()) { 1303 // If all preds provide 0, replace the xor with the other input. 1304 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1305 BO->eraseFromParent(); 1306 } else { 1307 // If all preds provide 1, set the computed value to 1. 1308 BO->setOperand(!isLHS, SplitVal); 1309 } 1310 1311 return true; 1312 } 1313 1314 // Try to duplicate BB into PredBB. 1315 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1316} 1317 1318 1319/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1320/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1321/// NewPred using the entries from OldPred (suitably mapped). 1322static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1323 BasicBlock *OldPred, 1324 BasicBlock *NewPred, 1325 DenseMap<Instruction*, Value*> &ValueMap) { 1326 for (BasicBlock::iterator PNI = PHIBB->begin(); 1327 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1328 // Ok, we have a PHI node. Figure out what the incoming value was for the 1329 // DestBlock. 1330 Value *IV = PN->getIncomingValueForBlock(OldPred); 1331 1332 // Remap the value if necessary. 1333 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1334 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1335 if (I != ValueMap.end()) 1336 IV = I->second; 1337 } 1338 1339 PN->addIncoming(IV, NewPred); 1340 } 1341} 1342 1343/// ThreadEdge - We have decided that it is safe and profitable to factor the 1344/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1345/// across BB. Transform the IR to reflect this change. 1346bool JumpThreading::ThreadEdge(BasicBlock *BB, 1347 const SmallVectorImpl<BasicBlock*> &PredBBs, 1348 BasicBlock *SuccBB) { 1349 // If threading to the same block as we come from, we would infinite loop. 1350 if (SuccBB == BB) { 1351 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1352 << "' - would thread to self!\n"); 1353 return false; 1354 } 1355 1356 // If threading this would thread across a loop header, don't thread the edge. 1357 // See the comments above FindLoopHeaders for justifications and caveats. 1358 if (LoopHeaders.count(BB)) { 1359 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1360 << "' to dest BB '" << SuccBB->getName() 1361 << "' - it might create an irreducible loop!\n"); 1362 return false; 1363 } 1364 1365 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold); 1366 if (JumpThreadCost > Threshold) { 1367 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1368 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1369 return false; 1370 } 1371 1372 // And finally, do it! Start by factoring the predecessors is needed. 1373 BasicBlock *PredBB; 1374 if (PredBBs.size() == 1) 1375 PredBB = PredBBs[0]; 1376 else { 1377 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1378 << " common predecessors.\n"); 1379 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); 1380 } 1381 1382 // And finally, do it! 1383 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1384 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1385 << ", across block:\n " 1386 << *BB << "\n"); 1387 1388 LVI->threadEdge(PredBB, BB, SuccBB); 1389 1390 // We are going to have to map operands from the original BB block to the new 1391 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1392 // account for entry from PredBB. 1393 DenseMap<Instruction*, Value*> ValueMapping; 1394 1395 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1396 BB->getName()+".thread", 1397 BB->getParent(), BB); 1398 NewBB->moveAfter(PredBB); 1399 1400 BasicBlock::iterator BI = BB->begin(); 1401 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1402 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1403 1404 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1405 // mapping and using it to remap operands in the cloned instructions. 1406 for (; !isa<TerminatorInst>(BI); ++BI) { 1407 Instruction *New = BI->clone(); 1408 New->setName(BI->getName()); 1409 NewBB->getInstList().push_back(New); 1410 ValueMapping[BI] = New; 1411 1412 // Remap operands to patch up intra-block references. 1413 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1414 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1415 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1416 if (I != ValueMapping.end()) 1417 New->setOperand(i, I->second); 1418 } 1419 } 1420 1421 // We didn't copy the terminator from BB over to NewBB, because there is now 1422 // an unconditional jump to SuccBB. Insert the unconditional jump. 1423 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB); 1424 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc()); 1425 1426 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1427 // PHI nodes for NewBB now. 1428 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1429 1430 // If there were values defined in BB that are used outside the block, then we 1431 // now have to update all uses of the value to use either the original value, 1432 // the cloned value, or some PHI derived value. This can require arbitrary 1433 // PHI insertion, of which we are prepared to do, clean these up now. 1434 SSAUpdater SSAUpdate; 1435 SmallVector<Use*, 16> UsesToRename; 1436 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1437 // Scan all uses of this instruction to see if it is used outside of its 1438 // block, and if so, record them in UsesToRename. 1439 for (Use &U : I->uses()) { 1440 Instruction *User = cast<Instruction>(U.getUser()); 1441 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1442 if (UserPN->getIncomingBlock(U) == BB) 1443 continue; 1444 } else if (User->getParent() == BB) 1445 continue; 1446 1447 UsesToRename.push_back(&U); 1448 } 1449 1450 // If there are no uses outside the block, we're done with this instruction. 1451 if (UsesToRename.empty()) 1452 continue; 1453 1454 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1455 1456 // We found a use of I outside of BB. Rename all uses of I that are outside 1457 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1458 // with the two values we know. 1459 SSAUpdate.Initialize(I->getType(), I->getName()); 1460 SSAUpdate.AddAvailableValue(BB, I); 1461 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1462 1463 while (!UsesToRename.empty()) 1464 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1465 DEBUG(dbgs() << "\n"); 1466 } 1467 1468 1469 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1470 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1471 // us to simplify any PHI nodes in BB. 1472 TerminatorInst *PredTerm = PredBB->getTerminator(); 1473 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1474 if (PredTerm->getSuccessor(i) == BB) { 1475 BB->removePredecessor(PredBB, true); 1476 PredTerm->setSuccessor(i, NewBB); 1477 } 1478 1479 // At this point, the IR is fully up to date and consistent. Do a quick scan 1480 // over the new instructions and zap any that are constants or dead. This 1481 // frequently happens because of phi translation. 1482 SimplifyInstructionsInBlock(NewBB, DL, TLI); 1483 1484 // Threaded an edge! 1485 ++NumThreads; 1486 return true; 1487} 1488 1489/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1490/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1491/// If we can duplicate the contents of BB up into PredBB do so now, this 1492/// improves the odds that the branch will be on an analyzable instruction like 1493/// a compare. 1494bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1495 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1496 assert(!PredBBs.empty() && "Can't handle an empty set"); 1497 1498 // If BB is a loop header, then duplicating this block outside the loop would 1499 // cause us to transform this into an irreducible loop, don't do this. 1500 // See the comments above FindLoopHeaders for justifications and caveats. 1501 if (LoopHeaders.count(BB)) { 1502 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1503 << "' into predecessor block '" << PredBBs[0]->getName() 1504 << "' - it might create an irreducible loop!\n"); 1505 return false; 1506 } 1507 1508 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold); 1509 if (DuplicationCost > Threshold) { 1510 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1511 << "' - Cost is too high: " << DuplicationCost << "\n"); 1512 return false; 1513 } 1514 1515 // And finally, do it! Start by factoring the predecessors is needed. 1516 BasicBlock *PredBB; 1517 if (PredBBs.size() == 1) 1518 PredBB = PredBBs[0]; 1519 else { 1520 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1521 << " common predecessors.\n"); 1522 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); 1523 } 1524 1525 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1526 // of PredBB. 1527 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1528 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1529 << DuplicationCost << " block is:" << *BB << "\n"); 1530 1531 // Unless PredBB ends with an unconditional branch, split the edge so that we 1532 // can just clone the bits from BB into the end of the new PredBB. 1533 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1534 1535 if (!OldPredBranch || !OldPredBranch->isUnconditional()) { 1536 PredBB = SplitEdge(PredBB, BB, this); 1537 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1538 } 1539 1540 // We are going to have to map operands from the original BB block into the 1541 // PredBB block. Evaluate PHI nodes in BB. 1542 DenseMap<Instruction*, Value*> ValueMapping; 1543 1544 BasicBlock::iterator BI = BB->begin(); 1545 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1546 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1547 1548 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1549 // mapping and using it to remap operands in the cloned instructions. 1550 for (; BI != BB->end(); ++BI) { 1551 Instruction *New = BI->clone(); 1552 1553 // Remap operands to patch up intra-block references. 1554 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1555 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1556 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1557 if (I != ValueMapping.end()) 1558 New->setOperand(i, I->second); 1559 } 1560 1561 // If this instruction can be simplified after the operands are updated, 1562 // just use the simplified value instead. This frequently happens due to 1563 // phi translation. 1564 if (Value *IV = SimplifyInstruction(New, DL)) { 1565 delete New; 1566 ValueMapping[BI] = IV; 1567 } else { 1568 // Otherwise, insert the new instruction into the block. 1569 New->setName(BI->getName()); 1570 PredBB->getInstList().insert(OldPredBranch, New); 1571 ValueMapping[BI] = New; 1572 } 1573 } 1574 1575 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1576 // add entries to the PHI nodes for branch from PredBB now. 1577 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1578 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1579 ValueMapping); 1580 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1581 ValueMapping); 1582 1583 // If there were values defined in BB that are used outside the block, then we 1584 // now have to update all uses of the value to use either the original value, 1585 // the cloned value, or some PHI derived value. This can require arbitrary 1586 // PHI insertion, of which we are prepared to do, clean these up now. 1587 SSAUpdater SSAUpdate; 1588 SmallVector<Use*, 16> UsesToRename; 1589 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1590 // Scan all uses of this instruction to see if it is used outside of its 1591 // block, and if so, record them in UsesToRename. 1592 for (Use &U : I->uses()) { 1593 Instruction *User = cast<Instruction>(U.getUser()); 1594 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1595 if (UserPN->getIncomingBlock(U) == BB) 1596 continue; 1597 } else if (User->getParent() == BB) 1598 continue; 1599 1600 UsesToRename.push_back(&U); 1601 } 1602 1603 // If there are no uses outside the block, we're done with this instruction. 1604 if (UsesToRename.empty()) 1605 continue; 1606 1607 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1608 1609 // We found a use of I outside of BB. Rename all uses of I that are outside 1610 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1611 // with the two values we know. 1612 SSAUpdate.Initialize(I->getType(), I->getName()); 1613 SSAUpdate.AddAvailableValue(BB, I); 1614 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1615 1616 while (!UsesToRename.empty()) 1617 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1618 DEBUG(dbgs() << "\n"); 1619 } 1620 1621 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1622 // that we nuked. 1623 BB->removePredecessor(PredBB, true); 1624 1625 // Remove the unconditional branch at the end of the PredBB block. 1626 OldPredBranch->eraseFromParent(); 1627 1628 ++NumDupes; 1629 return true; 1630} 1631 1632/// TryToUnfoldSelect - Look for blocks of the form 1633/// bb1: 1634/// %a = select 1635/// br bb 1636/// 1637/// bb2: 1638/// %p = phi [%a, %bb] ... 1639/// %c = icmp %p 1640/// br i1 %c 1641/// 1642/// And expand the select into a branch structure if one of its arms allows %c 1643/// to be folded. This later enables threading from bb1 over bb2. 1644bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) { 1645 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 1646 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0)); 1647 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1)); 1648 1649 if (!CondBr || !CondBr->isConditional() || !CondLHS || 1650 CondLHS->getParent() != BB) 1651 return false; 1652 1653 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) { 1654 BasicBlock *Pred = CondLHS->getIncomingBlock(I); 1655 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I)); 1656 1657 // Look if one of the incoming values is a select in the corresponding 1658 // predecessor. 1659 if (!SI || SI->getParent() != Pred || !SI->hasOneUse()) 1660 continue; 1661 1662 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator()); 1663 if (!PredTerm || !PredTerm->isUnconditional()) 1664 continue; 1665 1666 // Now check if one of the select values would allow us to constant fold the 1667 // terminator in BB. We don't do the transform if both sides fold, those 1668 // cases will be threaded in any case. 1669 LazyValueInfo::Tristate LHSFolds = 1670 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1), 1671 CondRHS, Pred, BB); 1672 LazyValueInfo::Tristate RHSFolds = 1673 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2), 1674 CondRHS, Pred, BB); 1675 if ((LHSFolds != LazyValueInfo::Unknown || 1676 RHSFolds != LazyValueInfo::Unknown) && 1677 LHSFolds != RHSFolds) { 1678 // Expand the select. 1679 // 1680 // Pred -- 1681 // | v 1682 // | NewBB 1683 // | | 1684 // |----- 1685 // v 1686 // BB 1687 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold", 1688 BB->getParent(), BB); 1689 // Move the unconditional branch to NewBB. 1690 PredTerm->removeFromParent(); 1691 NewBB->getInstList().insert(NewBB->end(), PredTerm); 1692 // Create a conditional branch and update PHI nodes. 1693 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred); 1694 CondLHS->setIncomingValue(I, SI->getFalseValue()); 1695 CondLHS->addIncoming(SI->getTrueValue(), NewBB); 1696 // The select is now dead. 1697 SI->eraseFromParent(); 1698 1699 // Update any other PHI nodes in BB. 1700 for (BasicBlock::iterator BI = BB->begin(); 1701 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI) 1702 if (Phi != CondLHS) 1703 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB); 1704 return true; 1705 } 1706 } 1707 return false; 1708} 1709