JumpThreading.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
1//===- JumpThreading.cpp - Thread control through conditional blocks ------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the Jump Threading pass. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "jump-threading" 15#include "llvm/Transforms/Scalar.h" 16#include "llvm/ADT/DenseMap.h" 17#include "llvm/ADT/DenseSet.h" 18#include "llvm/ADT/STLExtras.h" 19#include "llvm/ADT/SmallPtrSet.h" 20#include "llvm/ADT/SmallSet.h" 21#include "llvm/ADT/Statistic.h" 22#include "llvm/Analysis/CFG.h" 23#include "llvm/Analysis/ConstantFolding.h" 24#include "llvm/Analysis/InstructionSimplify.h" 25#include "llvm/Analysis/LazyValueInfo.h" 26#include "llvm/Analysis/Loads.h" 27#include "llvm/IR/DataLayout.h" 28#include "llvm/IR/IntrinsicInst.h" 29#include "llvm/IR/LLVMContext.h" 30#include "llvm/IR/ValueHandle.h" 31#include "llvm/Pass.h" 32#include "llvm/Support/CommandLine.h" 33#include "llvm/Support/Debug.h" 34#include "llvm/Support/raw_ostream.h" 35#include "llvm/Target/TargetLibraryInfo.h" 36#include "llvm/Transforms/Utils/BasicBlockUtils.h" 37#include "llvm/Transforms/Utils/Local.h" 38#include "llvm/Transforms/Utils/SSAUpdater.h" 39using namespace llvm; 40 41STATISTIC(NumThreads, "Number of jumps threaded"); 42STATISTIC(NumFolds, "Number of terminators folded"); 43STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 44 45static cl::opt<unsigned> 46Threshold("jump-threading-threshold", 47 cl::desc("Max block size to duplicate for jump threading"), 48 cl::init(6), cl::Hidden); 49 50namespace { 51 // These are at global scope so static functions can use them too. 52 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo; 53 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy; 54 55 // This is used to keep track of what kind of constant we're currently hoping 56 // to find. 57 enum ConstantPreference { 58 WantInteger, 59 WantBlockAddress 60 }; 61 62 /// This pass performs 'jump threading', which looks at blocks that have 63 /// multiple predecessors and multiple successors. If one or more of the 64 /// predecessors of the block can be proven to always jump to one of the 65 /// successors, we forward the edge from the predecessor to the successor by 66 /// duplicating the contents of this block. 67 /// 68 /// An example of when this can occur is code like this: 69 /// 70 /// if () { ... 71 /// X = 4; 72 /// } 73 /// if (X < 3) { 74 /// 75 /// In this case, the unconditional branch at the end of the first if can be 76 /// revectored to the false side of the second if. 77 /// 78 class JumpThreading : public FunctionPass { 79 const DataLayout *DL; 80 TargetLibraryInfo *TLI; 81 LazyValueInfo *LVI; 82#ifdef NDEBUG 83 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 84#else 85 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 86#endif 87 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet; 88 89 // RAII helper for updating the recursion stack. 90 struct RecursionSetRemover { 91 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet; 92 std::pair<Value*, BasicBlock*> ThePair; 93 94 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S, 95 std::pair<Value*, BasicBlock*> P) 96 : TheSet(S), ThePair(P) { } 97 98 ~RecursionSetRemover() { 99 TheSet.erase(ThePair); 100 } 101 }; 102 public: 103 static char ID; // Pass identification 104 JumpThreading() : FunctionPass(ID) { 105 initializeJumpThreadingPass(*PassRegistry::getPassRegistry()); 106 } 107 108 bool runOnFunction(Function &F) override; 109 110 void getAnalysisUsage(AnalysisUsage &AU) const override { 111 AU.addRequired<LazyValueInfo>(); 112 AU.addPreserved<LazyValueInfo>(); 113 AU.addRequired<TargetLibraryInfo>(); 114 } 115 116 void FindLoopHeaders(Function &F); 117 bool ProcessBlock(BasicBlock *BB); 118 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 119 BasicBlock *SuccBB); 120 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 121 const SmallVectorImpl<BasicBlock *> &PredBBs); 122 123 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 124 PredValueInfo &Result, 125 ConstantPreference Preference); 126 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 127 ConstantPreference Preference); 128 129 bool ProcessBranchOnPHI(PHINode *PN); 130 bool ProcessBranchOnXOR(BinaryOperator *BO); 131 132 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 133 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB); 134 }; 135} 136 137char JumpThreading::ID = 0; 138INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", 139 "Jump Threading", false, false) 140INITIALIZE_PASS_DEPENDENCY(LazyValueInfo) 141INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 142INITIALIZE_PASS_END(JumpThreading, "jump-threading", 143 "Jump Threading", false, false) 144 145// Public interface to the Jump Threading pass 146FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 147 148/// runOnFunction - Top level algorithm. 149/// 150bool JumpThreading::runOnFunction(Function &F) { 151 if (skipOptnoneFunction(F)) 152 return false; 153 154 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 155 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); 156 DL = DLP ? &DLP->getDataLayout() : 0; 157 TLI = &getAnalysis<TargetLibraryInfo>(); 158 LVI = &getAnalysis<LazyValueInfo>(); 159 160 FindLoopHeaders(F); 161 162 bool Changed, EverChanged = false; 163 do { 164 Changed = false; 165 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 166 BasicBlock *BB = I; 167 // Thread all of the branches we can over this block. 168 while (ProcessBlock(BB)) 169 Changed = true; 170 171 ++I; 172 173 // If the block is trivially dead, zap it. This eliminates the successor 174 // edges which simplifies the CFG. 175 if (pred_begin(BB) == pred_end(BB) && 176 BB != &BB->getParent()->getEntryBlock()) { 177 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 178 << "' with terminator: " << *BB->getTerminator() << '\n'); 179 LoopHeaders.erase(BB); 180 LVI->eraseBlock(BB); 181 DeleteDeadBlock(BB); 182 Changed = true; 183 continue; 184 } 185 186 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 187 188 // Can't thread an unconditional jump, but if the block is "almost 189 // empty", we can replace uses of it with uses of the successor and make 190 // this dead. 191 if (BI && BI->isUnconditional() && 192 BB != &BB->getParent()->getEntryBlock() && 193 // If the terminator is the only non-phi instruction, try to nuke it. 194 BB->getFirstNonPHIOrDbg()->isTerminator()) { 195 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 196 // block, we have to make sure it isn't in the LoopHeaders set. We 197 // reinsert afterward if needed. 198 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 199 BasicBlock *Succ = BI->getSuccessor(0); 200 201 // FIXME: It is always conservatively correct to drop the info 202 // for a block even if it doesn't get erased. This isn't totally 203 // awesome, but it allows us to use AssertingVH to prevent nasty 204 // dangling pointer issues within LazyValueInfo. 205 LVI->eraseBlock(BB); 206 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 207 Changed = true; 208 // If we deleted BB and BB was the header of a loop, then the 209 // successor is now the header of the loop. 210 BB = Succ; 211 } 212 213 if (ErasedFromLoopHeaders) 214 LoopHeaders.insert(BB); 215 } 216 } 217 EverChanged |= Changed; 218 } while (Changed); 219 220 LoopHeaders.clear(); 221 return EverChanged; 222} 223 224/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 225/// thread across it. Stop scanning the block when passing the threshold. 226static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB, 227 unsigned Threshold) { 228 /// Ignore PHI nodes, these will be flattened when duplication happens. 229 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 230 231 // FIXME: THREADING will delete values that are just used to compute the 232 // branch, so they shouldn't count against the duplication cost. 233 234 // Sum up the cost of each instruction until we get to the terminator. Don't 235 // include the terminator because the copy won't include it. 236 unsigned Size = 0; 237 for (; !isa<TerminatorInst>(I); ++I) { 238 239 // Stop scanning the block if we've reached the threshold. 240 if (Size > Threshold) 241 return Size; 242 243 // Debugger intrinsics don't incur code size. 244 if (isa<DbgInfoIntrinsic>(I)) continue; 245 246 // If this is a pointer->pointer bitcast, it is free. 247 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 248 continue; 249 250 // All other instructions count for at least one unit. 251 ++Size; 252 253 // Calls are more expensive. If they are non-intrinsic calls, we model them 254 // as having cost of 4. If they are a non-vector intrinsic, we model them 255 // as having cost of 2 total, and if they are a vector intrinsic, we model 256 // them as having cost 1. 257 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 258 if (CI->cannotDuplicate()) 259 // Blocks with NoDuplicate are modelled as having infinite cost, so they 260 // are never duplicated. 261 return ~0U; 262 else if (!isa<IntrinsicInst>(CI)) 263 Size += 3; 264 else if (!CI->getType()->isVectorTy()) 265 Size += 1; 266 } 267 } 268 269 // Threading through a switch statement is particularly profitable. If this 270 // block ends in a switch, decrease its cost to make it more likely to happen. 271 if (isa<SwitchInst>(I)) 272 Size = Size > 6 ? Size-6 : 0; 273 274 // The same holds for indirect branches, but slightly more so. 275 if (isa<IndirectBrInst>(I)) 276 Size = Size > 8 ? Size-8 : 0; 277 278 return Size; 279} 280 281/// FindLoopHeaders - We do not want jump threading to turn proper loop 282/// structures into irreducible loops. Doing this breaks up the loop nesting 283/// hierarchy and pessimizes later transformations. To prevent this from 284/// happening, we first have to find the loop headers. Here we approximate this 285/// by finding targets of backedges in the CFG. 286/// 287/// Note that there definitely are cases when we want to allow threading of 288/// edges across a loop header. For example, threading a jump from outside the 289/// loop (the preheader) to an exit block of the loop is definitely profitable. 290/// It is also almost always profitable to thread backedges from within the loop 291/// to exit blocks, and is often profitable to thread backedges to other blocks 292/// within the loop (forming a nested loop). This simple analysis is not rich 293/// enough to track all of these properties and keep it up-to-date as the CFG 294/// mutates, so we don't allow any of these transformations. 295/// 296void JumpThreading::FindLoopHeaders(Function &F) { 297 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 298 FindFunctionBackedges(F, Edges); 299 300 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 301 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 302} 303 304/// getKnownConstant - Helper method to determine if we can thread over a 305/// terminator with the given value as its condition, and if so what value to 306/// use for that. What kind of value this is depends on whether we want an 307/// integer or a block address, but an undef is always accepted. 308/// Returns null if Val is null or not an appropriate constant. 309static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { 310 if (!Val) 311 return 0; 312 313 // Undef is "known" enough. 314 if (UndefValue *U = dyn_cast<UndefValue>(Val)) 315 return U; 316 317 if (Preference == WantBlockAddress) 318 return dyn_cast<BlockAddress>(Val->stripPointerCasts()); 319 320 return dyn_cast<ConstantInt>(Val); 321} 322 323/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 324/// if we can infer that the value is a known ConstantInt/BlockAddress or undef 325/// in any of our predecessors. If so, return the known list of value and pred 326/// BB in the result vector. 327/// 328/// This returns true if there were any known values. 329/// 330bool JumpThreading:: 331ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result, 332 ConstantPreference Preference) { 333 // This method walks up use-def chains recursively. Because of this, we could 334 // get into an infinite loop going around loops in the use-def chain. To 335 // prevent this, keep track of what (value, block) pairs we've already visited 336 // and terminate the search if we loop back to them 337 if (!RecursionSet.insert(std::make_pair(V, BB)).second) 338 return false; 339 340 // An RAII help to remove this pair from the recursion set once the recursion 341 // stack pops back out again. 342 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB)); 343 344 // If V is a constant, then it is known in all predecessors. 345 if (Constant *KC = getKnownConstant(V, Preference)) { 346 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 347 Result.push_back(std::make_pair(KC, *PI)); 348 349 return true; 350 } 351 352 // If V is a non-instruction value, or an instruction in a different block, 353 // then it can't be derived from a PHI. 354 Instruction *I = dyn_cast<Instruction>(V); 355 if (I == 0 || I->getParent() != BB) { 356 357 // Okay, if this is a live-in value, see if it has a known value at the end 358 // of any of our predecessors. 359 // 360 // FIXME: This should be an edge property, not a block end property. 361 /// TODO: Per PR2563, we could infer value range information about a 362 /// predecessor based on its terminator. 363 // 364 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 365 // "I" is a non-local compare-with-a-constant instruction. This would be 366 // able to handle value inequalities better, for example if the compare is 367 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 368 // Perhaps getConstantOnEdge should be smart enough to do this? 369 370 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 371 BasicBlock *P = *PI; 372 // If the value is known by LazyValueInfo to be a constant in a 373 // predecessor, use that information to try to thread this block. 374 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 375 if (Constant *KC = getKnownConstant(PredCst, Preference)) 376 Result.push_back(std::make_pair(KC, P)); 377 } 378 379 return !Result.empty(); 380 } 381 382 /// If I is a PHI node, then we know the incoming values for any constants. 383 if (PHINode *PN = dyn_cast<PHINode>(I)) { 384 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 385 Value *InVal = PN->getIncomingValue(i); 386 if (Constant *KC = getKnownConstant(InVal, Preference)) { 387 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 388 } else { 389 Constant *CI = LVI->getConstantOnEdge(InVal, 390 PN->getIncomingBlock(i), BB); 391 if (Constant *KC = getKnownConstant(CI, Preference)) 392 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 393 } 394 } 395 396 return !Result.empty(); 397 } 398 399 PredValueInfoTy LHSVals, RHSVals; 400 401 // Handle some boolean conditions. 402 if (I->getType()->getPrimitiveSizeInBits() == 1) { 403 assert(Preference == WantInteger && "One-bit non-integer type?"); 404 // X | true -> true 405 // X & false -> false 406 if (I->getOpcode() == Instruction::Or || 407 I->getOpcode() == Instruction::And) { 408 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 409 WantInteger); 410 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals, 411 WantInteger); 412 413 if (LHSVals.empty() && RHSVals.empty()) 414 return false; 415 416 ConstantInt *InterestingVal; 417 if (I->getOpcode() == Instruction::Or) 418 InterestingVal = ConstantInt::getTrue(I->getContext()); 419 else 420 InterestingVal = ConstantInt::getFalse(I->getContext()); 421 422 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 423 424 // Scan for the sentinel. If we find an undef, force it to the 425 // interesting value: x|undef -> true and x&undef -> false. 426 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 427 if (LHSVals[i].first == InterestingVal || 428 isa<UndefValue>(LHSVals[i].first)) { 429 Result.push_back(LHSVals[i]); 430 Result.back().first = InterestingVal; 431 LHSKnownBBs.insert(LHSVals[i].second); 432 } 433 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 434 if (RHSVals[i].first == InterestingVal || 435 isa<UndefValue>(RHSVals[i].first)) { 436 // If we already inferred a value for this block on the LHS, don't 437 // re-add it. 438 if (!LHSKnownBBs.count(RHSVals[i].second)) { 439 Result.push_back(RHSVals[i]); 440 Result.back().first = InterestingVal; 441 } 442 } 443 444 return !Result.empty(); 445 } 446 447 // Handle the NOT form of XOR. 448 if (I->getOpcode() == Instruction::Xor && 449 isa<ConstantInt>(I->getOperand(1)) && 450 cast<ConstantInt>(I->getOperand(1))->isOne()) { 451 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result, 452 WantInteger); 453 if (Result.empty()) 454 return false; 455 456 // Invert the known values. 457 for (unsigned i = 0, e = Result.size(); i != e; ++i) 458 Result[i].first = ConstantExpr::getNot(Result[i].first); 459 460 return true; 461 } 462 463 // Try to simplify some other binary operator values. 464 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 465 assert(Preference != WantBlockAddress 466 && "A binary operator creating a block address?"); 467 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 468 PredValueInfoTy LHSVals; 469 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals, 470 WantInteger); 471 472 // Try to use constant folding to simplify the binary operator. 473 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 474 Constant *V = LHSVals[i].first; 475 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); 476 477 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 478 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 479 } 480 } 481 482 return !Result.empty(); 483 } 484 485 // Handle compare with phi operand, where the PHI is defined in this block. 486 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 487 assert(Preference == WantInteger && "Compares only produce integers"); 488 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 489 if (PN && PN->getParent() == BB) { 490 // We can do this simplification if any comparisons fold to true or false. 491 // See if any do. 492 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 493 BasicBlock *PredBB = PN->getIncomingBlock(i); 494 Value *LHS = PN->getIncomingValue(i); 495 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 496 497 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL); 498 if (Res == 0) { 499 if (!isa<Constant>(RHS)) 500 continue; 501 502 LazyValueInfo::Tristate 503 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 504 cast<Constant>(RHS), PredBB, BB); 505 if (ResT == LazyValueInfo::Unknown) 506 continue; 507 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 508 } 509 510 if (Constant *KC = getKnownConstant(Res, WantInteger)) 511 Result.push_back(std::make_pair(KC, PredBB)); 512 } 513 514 return !Result.empty(); 515 } 516 517 518 // If comparing a live-in value against a constant, see if we know the 519 // live-in value on any predecessors. 520 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) { 521 if (!isa<Instruction>(Cmp->getOperand(0)) || 522 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 523 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 524 525 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ 526 BasicBlock *P = *PI; 527 // If the value is known by LazyValueInfo to be a constant in a 528 // predecessor, use that information to try to thread this block. 529 LazyValueInfo::Tristate Res = 530 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 531 RHSCst, P, BB); 532 if (Res == LazyValueInfo::Unknown) 533 continue; 534 535 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 536 Result.push_back(std::make_pair(ResC, P)); 537 } 538 539 return !Result.empty(); 540 } 541 542 // Try to find a constant value for the LHS of a comparison, 543 // and evaluate it statically if we can. 544 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { 545 PredValueInfoTy LHSVals; 546 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 547 WantInteger); 548 549 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 550 Constant *V = LHSVals[i].first; 551 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(), 552 V, CmpConst); 553 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 554 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 555 } 556 557 return !Result.empty(); 558 } 559 } 560 } 561 562 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 563 // Handle select instructions where at least one operand is a known constant 564 // and we can figure out the condition value for any predecessor block. 565 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference); 566 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference); 567 PredValueInfoTy Conds; 568 if ((TrueVal || FalseVal) && 569 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds, 570 WantInteger)) { 571 for (unsigned i = 0, e = Conds.size(); i != e; ++i) { 572 Constant *Cond = Conds[i].first; 573 574 // Figure out what value to use for the condition. 575 bool KnownCond; 576 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) { 577 // A known boolean. 578 KnownCond = CI->isOne(); 579 } else { 580 assert(isa<UndefValue>(Cond) && "Unexpected condition value"); 581 // Either operand will do, so be sure to pick the one that's a known 582 // constant. 583 // FIXME: Do this more cleverly if both values are known constants? 584 KnownCond = (TrueVal != 0); 585 } 586 587 // See if the select has a known constant value for this predecessor. 588 if (Constant *Val = KnownCond ? TrueVal : FalseVal) 589 Result.push_back(std::make_pair(Val, Conds[i].second)); 590 } 591 592 return !Result.empty(); 593 } 594 } 595 596 // If all else fails, see if LVI can figure out a constant value for us. 597 Constant *CI = LVI->getConstant(V, BB); 598 if (Constant *KC = getKnownConstant(CI, Preference)) { 599 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 600 Result.push_back(std::make_pair(KC, *PI)); 601 } 602 603 return !Result.empty(); 604} 605 606 607 608/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 609/// in an undefined jump, decide which block is best to revector to. 610/// 611/// Since we can pick an arbitrary destination, we pick the successor with the 612/// fewest predecessors. This should reduce the in-degree of the others. 613/// 614static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 615 TerminatorInst *BBTerm = BB->getTerminator(); 616 unsigned MinSucc = 0; 617 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 618 // Compute the successor with the minimum number of predecessors. 619 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 620 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 621 TestBB = BBTerm->getSuccessor(i); 622 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 623 if (NumPreds < MinNumPreds) { 624 MinSucc = i; 625 MinNumPreds = NumPreds; 626 } 627 } 628 629 return MinSucc; 630} 631 632static bool hasAddressTakenAndUsed(BasicBlock *BB) { 633 if (!BB->hasAddressTaken()) return false; 634 635 // If the block has its address taken, it may be a tree of dead constants 636 // hanging off of it. These shouldn't keep the block alive. 637 BlockAddress *BA = BlockAddress::get(BB); 638 BA->removeDeadConstantUsers(); 639 return !BA->use_empty(); 640} 641 642/// ProcessBlock - If there are any predecessors whose control can be threaded 643/// through to a successor, transform them now. 644bool JumpThreading::ProcessBlock(BasicBlock *BB) { 645 // If the block is trivially dead, just return and let the caller nuke it. 646 // This simplifies other transformations. 647 if (pred_begin(BB) == pred_end(BB) && 648 BB != &BB->getParent()->getEntryBlock()) 649 return false; 650 651 // If this block has a single predecessor, and if that pred has a single 652 // successor, merge the blocks. This encourages recursive jump threading 653 // because now the condition in this block can be threaded through 654 // predecessors of our predecessor block. 655 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 656 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 657 SinglePred != BB && !hasAddressTakenAndUsed(BB)) { 658 // If SinglePred was a loop header, BB becomes one. 659 if (LoopHeaders.erase(SinglePred)) 660 LoopHeaders.insert(BB); 661 662 // Remember if SinglePred was the entry block of the function. If so, we 663 // will need to move BB back to the entry position. 664 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 665 LVI->eraseBlock(SinglePred); 666 MergeBasicBlockIntoOnlyPred(BB); 667 668 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 669 BB->moveBefore(&BB->getParent()->getEntryBlock()); 670 return true; 671 } 672 } 673 674 // What kind of constant we're looking for. 675 ConstantPreference Preference = WantInteger; 676 677 // Look to see if the terminator is a conditional branch, switch or indirect 678 // branch, if not we can't thread it. 679 Value *Condition; 680 Instruction *Terminator = BB->getTerminator(); 681 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { 682 // Can't thread an unconditional jump. 683 if (BI->isUnconditional()) return false; 684 Condition = BI->getCondition(); 685 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { 686 Condition = SI->getCondition(); 687 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { 688 // Can't thread indirect branch with no successors. 689 if (IB->getNumSuccessors() == 0) return false; 690 Condition = IB->getAddress()->stripPointerCasts(); 691 Preference = WantBlockAddress; 692 } else { 693 return false; // Must be an invoke. 694 } 695 696 // Run constant folding to see if we can reduce the condition to a simple 697 // constant. 698 if (Instruction *I = dyn_cast<Instruction>(Condition)) { 699 Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI); 700 if (SimpleVal) { 701 I->replaceAllUsesWith(SimpleVal); 702 I->eraseFromParent(); 703 Condition = SimpleVal; 704 } 705 } 706 707 // If the terminator is branching on an undef, we can pick any of the 708 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 709 if (isa<UndefValue>(Condition)) { 710 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 711 712 // Fold the branch/switch. 713 TerminatorInst *BBTerm = BB->getTerminator(); 714 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 715 if (i == BestSucc) continue; 716 BBTerm->getSuccessor(i)->removePredecessor(BB, true); 717 } 718 719 DEBUG(dbgs() << " In block '" << BB->getName() 720 << "' folding undef terminator: " << *BBTerm << '\n'); 721 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 722 BBTerm->eraseFromParent(); 723 return true; 724 } 725 726 // If the terminator of this block is branching on a constant, simplify the 727 // terminator to an unconditional branch. This can occur due to threading in 728 // other blocks. 729 if (getKnownConstant(Condition, Preference)) { 730 DEBUG(dbgs() << " In block '" << BB->getName() 731 << "' folding terminator: " << *BB->getTerminator() << '\n'); 732 ++NumFolds; 733 ConstantFoldTerminator(BB, true); 734 return true; 735 } 736 737 Instruction *CondInst = dyn_cast<Instruction>(Condition); 738 739 // All the rest of our checks depend on the condition being an instruction. 740 if (CondInst == 0) { 741 // FIXME: Unify this with code below. 742 if (ProcessThreadableEdges(Condition, BB, Preference)) 743 return true; 744 return false; 745 } 746 747 748 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 749 // For a comparison where the LHS is outside this block, it's possible 750 // that we've branched on it before. Used LVI to see if we can simplify 751 // the branch based on that. 752 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 753 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 754 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 755 if (CondBr && CondConst && CondBr->isConditional() && PI != PE && 756 (!isa<Instruction>(CondCmp->getOperand(0)) || 757 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { 758 // For predecessor edge, determine if the comparison is true or false 759 // on that edge. If they're all true or all false, we can simplify the 760 // branch. 761 // FIXME: We could handle mixed true/false by duplicating code. 762 LazyValueInfo::Tristate Baseline = 763 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0), 764 CondConst, *PI, BB); 765 if (Baseline != LazyValueInfo::Unknown) { 766 // Check that all remaining incoming values match the first one. 767 while (++PI != PE) { 768 LazyValueInfo::Tristate Ret = 769 LVI->getPredicateOnEdge(CondCmp->getPredicate(), 770 CondCmp->getOperand(0), CondConst, *PI, BB); 771 if (Ret != Baseline) break; 772 } 773 774 // If we terminated early, then one of the values didn't match. 775 if (PI == PE) { 776 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0; 777 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1; 778 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); 779 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); 780 CondBr->eraseFromParent(); 781 return true; 782 } 783 } 784 785 } 786 787 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB)) 788 return true; 789 } 790 791 // Check for some cases that are worth simplifying. Right now we want to look 792 // for loads that are used by a switch or by the condition for the branch. If 793 // we see one, check to see if it's partially redundant. If so, insert a PHI 794 // which can then be used to thread the values. 795 // 796 Value *SimplifyValue = CondInst; 797 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 798 if (isa<Constant>(CondCmp->getOperand(1))) 799 SimplifyValue = CondCmp->getOperand(0); 800 801 // TODO: There are other places where load PRE would be profitable, such as 802 // more complex comparisons. 803 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 804 if (SimplifyPartiallyRedundantLoad(LI)) 805 return true; 806 807 808 // Handle a variety of cases where we are branching on something derived from 809 // a PHI node in the current block. If we can prove that any predecessors 810 // compute a predictable value based on a PHI node, thread those predecessors. 811 // 812 if (ProcessThreadableEdges(CondInst, BB, Preference)) 813 return true; 814 815 // If this is an otherwise-unfoldable branch on a phi node in the current 816 // block, see if we can simplify. 817 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 818 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 819 return ProcessBranchOnPHI(PN); 820 821 822 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 823 if (CondInst->getOpcode() == Instruction::Xor && 824 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 825 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 826 827 828 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 829 // "(X == 4)", thread through this block. 830 831 return false; 832} 833 834/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 835/// load instruction, eliminate it by replacing it with a PHI node. This is an 836/// important optimization that encourages jump threading, and needs to be run 837/// interlaced with other jump threading tasks. 838bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 839 // Don't hack volatile/atomic loads. 840 if (!LI->isSimple()) return false; 841 842 // If the load is defined in a block with exactly one predecessor, it can't be 843 // partially redundant. 844 BasicBlock *LoadBB = LI->getParent(); 845 if (LoadBB->getSinglePredecessor()) 846 return false; 847 848 // If the load is defined in a landing pad, it can't be partially redundant, 849 // because the edges between the invoke and the landing pad cannot have other 850 // instructions between them. 851 if (LoadBB->isLandingPad()) 852 return false; 853 854 Value *LoadedPtr = LI->getOperand(0); 855 856 // If the loaded operand is defined in the LoadBB, it can't be available. 857 // TODO: Could do simple PHI translation, that would be fun :) 858 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 859 if (PtrOp->getParent() == LoadBB) 860 return false; 861 862 // Scan a few instructions up from the load, to see if it is obviously live at 863 // the entry to its block. 864 BasicBlock::iterator BBIt = LI; 865 866 if (Value *AvailableVal = 867 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 868 // If the value if the load is locally available within the block, just use 869 // it. This frequently occurs for reg2mem'd allocas. 870 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 871 872 // If the returned value is the load itself, replace with an undef. This can 873 // only happen in dead loops. 874 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 875 LI->replaceAllUsesWith(AvailableVal); 876 LI->eraseFromParent(); 877 return true; 878 } 879 880 // Otherwise, if we scanned the whole block and got to the top of the block, 881 // we know the block is locally transparent to the load. If not, something 882 // might clobber its value. 883 if (BBIt != LoadBB->begin()) 884 return false; 885 886 // If all of the loads and stores that feed the value have the same TBAA tag, 887 // then we can propagate it onto any newly inserted loads. 888 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa); 889 890 SmallPtrSet<BasicBlock*, 8> PredsScanned; 891 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 892 AvailablePredsTy AvailablePreds; 893 BasicBlock *OneUnavailablePred = 0; 894 895 // If we got here, the loaded value is transparent through to the start of the 896 // block. Check to see if it is available in any of the predecessor blocks. 897 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 898 PI != PE; ++PI) { 899 BasicBlock *PredBB = *PI; 900 901 // If we already scanned this predecessor, skip it. 902 if (!PredsScanned.insert(PredBB)) 903 continue; 904 905 // Scan the predecessor to see if the value is available in the pred. 906 BBIt = PredBB->end(); 907 MDNode *ThisTBAATag = 0; 908 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6, 909 0, &ThisTBAATag); 910 if (!PredAvailable) { 911 OneUnavailablePred = PredBB; 912 continue; 913 } 914 915 // If tbaa tags disagree or are not present, forget about them. 916 if (TBAATag != ThisTBAATag) TBAATag = 0; 917 918 // If so, this load is partially redundant. Remember this info so that we 919 // can create a PHI node. 920 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 921 } 922 923 // If the loaded value isn't available in any predecessor, it isn't partially 924 // redundant. 925 if (AvailablePreds.empty()) return false; 926 927 // Okay, the loaded value is available in at least one (and maybe all!) 928 // predecessors. If the value is unavailable in more than one unique 929 // predecessor, we want to insert a merge block for those common predecessors. 930 // This ensures that we only have to insert one reload, thus not increasing 931 // code size. 932 BasicBlock *UnavailablePred = 0; 933 934 // If there is exactly one predecessor where the value is unavailable, the 935 // already computed 'OneUnavailablePred' block is it. If it ends in an 936 // unconditional branch, we know that it isn't a critical edge. 937 if (PredsScanned.size() == AvailablePreds.size()+1 && 938 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 939 UnavailablePred = OneUnavailablePred; 940 } else if (PredsScanned.size() != AvailablePreds.size()) { 941 // Otherwise, we had multiple unavailable predecessors or we had a critical 942 // edge from the one. 943 SmallVector<BasicBlock*, 8> PredsToSplit; 944 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 945 946 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 947 AvailablePredSet.insert(AvailablePreds[i].first); 948 949 // Add all the unavailable predecessors to the PredsToSplit list. 950 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 951 PI != PE; ++PI) { 952 BasicBlock *P = *PI; 953 // If the predecessor is an indirect goto, we can't split the edge. 954 if (isa<IndirectBrInst>(P->getTerminator())) 955 return false; 956 957 if (!AvailablePredSet.count(P)) 958 PredsToSplit.push_back(P); 959 } 960 961 // Split them out to their own block. 962 UnavailablePred = 963 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this); 964 } 965 966 // If the value isn't available in all predecessors, then there will be 967 // exactly one where it isn't available. Insert a load on that edge and add 968 // it to the AvailablePreds list. 969 if (UnavailablePred) { 970 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 971 "Can't handle critical edge here!"); 972 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 973 LI->getAlignment(), 974 UnavailablePred->getTerminator()); 975 NewVal->setDebugLoc(LI->getDebugLoc()); 976 if (TBAATag) 977 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag); 978 979 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 980 } 981 982 // Now we know that each predecessor of this block has a value in 983 // AvailablePreds, sort them for efficient access as we're walking the preds. 984 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 985 986 // Create a PHI node at the start of the block for the PRE'd load value. 987 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB); 988 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "", 989 LoadBB->begin()); 990 PN->takeName(LI); 991 PN->setDebugLoc(LI->getDebugLoc()); 992 993 // Insert new entries into the PHI for each predecessor. A single block may 994 // have multiple entries here. 995 for (pred_iterator PI = PB; PI != PE; ++PI) { 996 BasicBlock *P = *PI; 997 AvailablePredsTy::iterator I = 998 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 999 std::make_pair(P, (Value*)0)); 1000 1001 assert(I != AvailablePreds.end() && I->first == P && 1002 "Didn't find entry for predecessor!"); 1003 1004 PN->addIncoming(I->second, I->first); 1005 } 1006 1007 //cerr << "PRE: " << *LI << *PN << "\n"; 1008 1009 LI->replaceAllUsesWith(PN); 1010 LI->eraseFromParent(); 1011 1012 return true; 1013} 1014 1015/// FindMostPopularDest - The specified list contains multiple possible 1016/// threadable destinations. Pick the one that occurs the most frequently in 1017/// the list. 1018static BasicBlock * 1019FindMostPopularDest(BasicBlock *BB, 1020 const SmallVectorImpl<std::pair<BasicBlock*, 1021 BasicBlock*> > &PredToDestList) { 1022 assert(!PredToDestList.empty()); 1023 1024 // Determine popularity. If there are multiple possible destinations, we 1025 // explicitly choose to ignore 'undef' destinations. We prefer to thread 1026 // blocks with known and real destinations to threading undef. We'll handle 1027 // them later if interesting. 1028 DenseMap<BasicBlock*, unsigned> DestPopularity; 1029 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1030 if (PredToDestList[i].second) 1031 DestPopularity[PredToDestList[i].second]++; 1032 1033 // Find the most popular dest. 1034 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 1035 BasicBlock *MostPopularDest = DPI->first; 1036 unsigned Popularity = DPI->second; 1037 SmallVector<BasicBlock*, 4> SamePopularity; 1038 1039 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 1040 // If the popularity of this entry isn't higher than the popularity we've 1041 // seen so far, ignore it. 1042 if (DPI->second < Popularity) 1043 ; // ignore. 1044 else if (DPI->second == Popularity) { 1045 // If it is the same as what we've seen so far, keep track of it. 1046 SamePopularity.push_back(DPI->first); 1047 } else { 1048 // If it is more popular, remember it. 1049 SamePopularity.clear(); 1050 MostPopularDest = DPI->first; 1051 Popularity = DPI->second; 1052 } 1053 } 1054 1055 // Okay, now we know the most popular destination. If there is more than one 1056 // destination, we need to determine one. This is arbitrary, but we need 1057 // to make a deterministic decision. Pick the first one that appears in the 1058 // successor list. 1059 if (!SamePopularity.empty()) { 1060 SamePopularity.push_back(MostPopularDest); 1061 TerminatorInst *TI = BB->getTerminator(); 1062 for (unsigned i = 0; ; ++i) { 1063 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1064 1065 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1066 TI->getSuccessor(i)) == SamePopularity.end()) 1067 continue; 1068 1069 MostPopularDest = TI->getSuccessor(i); 1070 break; 1071 } 1072 } 1073 1074 // Okay, we have finally picked the most popular destination. 1075 return MostPopularDest; 1076} 1077 1078bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 1079 ConstantPreference Preference) { 1080 // If threading this would thread across a loop header, don't even try to 1081 // thread the edge. 1082 if (LoopHeaders.count(BB)) 1083 return false; 1084 1085 PredValueInfoTy PredValues; 1086 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference)) 1087 return false; 1088 1089 assert(!PredValues.empty() && 1090 "ComputeValueKnownInPredecessors returned true with no values"); 1091 1092 DEBUG(dbgs() << "IN BB: " << *BB; 1093 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1094 dbgs() << " BB '" << BB->getName() << "': FOUND condition = " 1095 << *PredValues[i].first 1096 << " for pred '" << PredValues[i].second->getName() << "'.\n"; 1097 }); 1098 1099 // Decide what we want to thread through. Convert our list of known values to 1100 // a list of known destinations for each pred. This also discards duplicate 1101 // predecessors and keeps track of the undefined inputs (which are represented 1102 // as a null dest in the PredToDestList). 1103 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1104 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1105 1106 BasicBlock *OnlyDest = 0; 1107 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1108 1109 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1110 BasicBlock *Pred = PredValues[i].second; 1111 if (!SeenPreds.insert(Pred)) 1112 continue; // Duplicate predecessor entry. 1113 1114 // If the predecessor ends with an indirect goto, we can't change its 1115 // destination. 1116 if (isa<IndirectBrInst>(Pred->getTerminator())) 1117 continue; 1118 1119 Constant *Val = PredValues[i].first; 1120 1121 BasicBlock *DestBB; 1122 if (isa<UndefValue>(Val)) 1123 DestBB = 0; 1124 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1125 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); 1126 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 1127 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor(); 1128 } else { 1129 assert(isa<IndirectBrInst>(BB->getTerminator()) 1130 && "Unexpected terminator"); 1131 DestBB = cast<BlockAddress>(Val)->getBasicBlock(); 1132 } 1133 1134 // If we have exactly one destination, remember it for efficiency below. 1135 if (PredToDestList.empty()) 1136 OnlyDest = DestBB; 1137 else if (OnlyDest != DestBB) 1138 OnlyDest = MultipleDestSentinel; 1139 1140 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1141 } 1142 1143 // If all edges were unthreadable, we fail. 1144 if (PredToDestList.empty()) 1145 return false; 1146 1147 // Determine which is the most common successor. If we have many inputs and 1148 // this block is a switch, we want to start by threading the batch that goes 1149 // to the most popular destination first. If we only know about one 1150 // threadable destination (the common case) we can avoid this. 1151 BasicBlock *MostPopularDest = OnlyDest; 1152 1153 if (MostPopularDest == MultipleDestSentinel) 1154 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1155 1156 // Now that we know what the most popular destination is, factor all 1157 // predecessors that will jump to it into a single predecessor. 1158 SmallVector<BasicBlock*, 16> PredsToFactor; 1159 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1160 if (PredToDestList[i].second == MostPopularDest) { 1161 BasicBlock *Pred = PredToDestList[i].first; 1162 1163 // This predecessor may be a switch or something else that has multiple 1164 // edges to the block. Factor each of these edges by listing them 1165 // according to # occurrences in PredsToFactor. 1166 TerminatorInst *PredTI = Pred->getTerminator(); 1167 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1168 if (PredTI->getSuccessor(i) == BB) 1169 PredsToFactor.push_back(Pred); 1170 } 1171 1172 // If the threadable edges are branching on an undefined value, we get to pick 1173 // the destination that these predecessors should get to. 1174 if (MostPopularDest == 0) 1175 MostPopularDest = BB->getTerminator()-> 1176 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1177 1178 // Ok, try to thread it! 1179 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1180} 1181 1182/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1183/// a PHI node in the current block. See if there are any simplifications we 1184/// can do based on inputs to the phi node. 1185/// 1186bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1187 BasicBlock *BB = PN->getParent(); 1188 1189 // TODO: We could make use of this to do it once for blocks with common PHI 1190 // values. 1191 SmallVector<BasicBlock*, 1> PredBBs; 1192 PredBBs.resize(1); 1193 1194 // If any of the predecessor blocks end in an unconditional branch, we can 1195 // *duplicate* the conditional branch into that block in order to further 1196 // encourage jump threading and to eliminate cases where we have branch on a 1197 // phi of an icmp (branch on icmp is much better). 1198 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1199 BasicBlock *PredBB = PN->getIncomingBlock(i); 1200 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1201 if (PredBr->isUnconditional()) { 1202 PredBBs[0] = PredBB; 1203 // Try to duplicate BB into PredBB. 1204 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1205 return true; 1206 } 1207 } 1208 1209 return false; 1210} 1211 1212/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1213/// a xor instruction in the current block. See if there are any 1214/// simplifications we can do based on inputs to the xor. 1215/// 1216bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1217 BasicBlock *BB = BO->getParent(); 1218 1219 // If either the LHS or RHS of the xor is a constant, don't do this 1220 // optimization. 1221 if (isa<ConstantInt>(BO->getOperand(0)) || 1222 isa<ConstantInt>(BO->getOperand(1))) 1223 return false; 1224 1225 // If the first instruction in BB isn't a phi, we won't be able to infer 1226 // anything special about any particular predecessor. 1227 if (!isa<PHINode>(BB->front())) 1228 return false; 1229 1230 // If we have a xor as the branch input to this block, and we know that the 1231 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1232 // the condition into the predecessor and fix that value to true, saving some 1233 // logical ops on that path and encouraging other paths to simplify. 1234 // 1235 // This copies something like this: 1236 // 1237 // BB: 1238 // %X = phi i1 [1], [%X'] 1239 // %Y = icmp eq i32 %A, %B 1240 // %Z = xor i1 %X, %Y 1241 // br i1 %Z, ... 1242 // 1243 // Into: 1244 // BB': 1245 // %Y = icmp ne i32 %A, %B 1246 // br i1 %Z, ... 1247 1248 PredValueInfoTy XorOpValues; 1249 bool isLHS = true; 1250 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, 1251 WantInteger)) { 1252 assert(XorOpValues.empty()); 1253 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, 1254 WantInteger)) 1255 return false; 1256 isLHS = false; 1257 } 1258 1259 assert(!XorOpValues.empty() && 1260 "ComputeValueKnownInPredecessors returned true with no values"); 1261 1262 // Scan the information to see which is most popular: true or false. The 1263 // predecessors can be of the set true, false, or undef. 1264 unsigned NumTrue = 0, NumFalse = 0; 1265 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1266 if (isa<UndefValue>(XorOpValues[i].first)) 1267 // Ignore undefs for the count. 1268 continue; 1269 if (cast<ConstantInt>(XorOpValues[i].first)->isZero()) 1270 ++NumFalse; 1271 else 1272 ++NumTrue; 1273 } 1274 1275 // Determine which value to split on, true, false, or undef if neither. 1276 ConstantInt *SplitVal = 0; 1277 if (NumTrue > NumFalse) 1278 SplitVal = ConstantInt::getTrue(BB->getContext()); 1279 else if (NumTrue != 0 || NumFalse != 0) 1280 SplitVal = ConstantInt::getFalse(BB->getContext()); 1281 1282 // Collect all of the blocks that this can be folded into so that we can 1283 // factor this once and clone it once. 1284 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1285 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1286 if (XorOpValues[i].first != SplitVal && 1287 !isa<UndefValue>(XorOpValues[i].first)) 1288 continue; 1289 1290 BlocksToFoldInto.push_back(XorOpValues[i].second); 1291 } 1292 1293 // If we inferred a value for all of the predecessors, then duplication won't 1294 // help us. However, we can just replace the LHS or RHS with the constant. 1295 if (BlocksToFoldInto.size() == 1296 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1297 if (SplitVal == 0) { 1298 // If all preds provide undef, just nuke the xor, because it is undef too. 1299 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1300 BO->eraseFromParent(); 1301 } else if (SplitVal->isZero()) { 1302 // If all preds provide 0, replace the xor with the other input. 1303 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1304 BO->eraseFromParent(); 1305 } else { 1306 // If all preds provide 1, set the computed value to 1. 1307 BO->setOperand(!isLHS, SplitVal); 1308 } 1309 1310 return true; 1311 } 1312 1313 // Try to duplicate BB into PredBB. 1314 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1315} 1316 1317 1318/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1319/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1320/// NewPred using the entries from OldPred (suitably mapped). 1321static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1322 BasicBlock *OldPred, 1323 BasicBlock *NewPred, 1324 DenseMap<Instruction*, Value*> &ValueMap) { 1325 for (BasicBlock::iterator PNI = PHIBB->begin(); 1326 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1327 // Ok, we have a PHI node. Figure out what the incoming value was for the 1328 // DestBlock. 1329 Value *IV = PN->getIncomingValueForBlock(OldPred); 1330 1331 // Remap the value if necessary. 1332 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1333 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1334 if (I != ValueMap.end()) 1335 IV = I->second; 1336 } 1337 1338 PN->addIncoming(IV, NewPred); 1339 } 1340} 1341 1342/// ThreadEdge - We have decided that it is safe and profitable to factor the 1343/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1344/// across BB. Transform the IR to reflect this change. 1345bool JumpThreading::ThreadEdge(BasicBlock *BB, 1346 const SmallVectorImpl<BasicBlock*> &PredBBs, 1347 BasicBlock *SuccBB) { 1348 // If threading to the same block as we come from, we would infinite loop. 1349 if (SuccBB == BB) { 1350 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1351 << "' - would thread to self!\n"); 1352 return false; 1353 } 1354 1355 // If threading this would thread across a loop header, don't thread the edge. 1356 // See the comments above FindLoopHeaders for justifications and caveats. 1357 if (LoopHeaders.count(BB)) { 1358 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1359 << "' to dest BB '" << SuccBB->getName() 1360 << "' - it might create an irreducible loop!\n"); 1361 return false; 1362 } 1363 1364 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold); 1365 if (JumpThreadCost > Threshold) { 1366 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1367 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1368 return false; 1369 } 1370 1371 // And finally, do it! Start by factoring the predecessors is needed. 1372 BasicBlock *PredBB; 1373 if (PredBBs.size() == 1) 1374 PredBB = PredBBs[0]; 1375 else { 1376 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1377 << " common predecessors.\n"); 1378 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); 1379 } 1380 1381 // And finally, do it! 1382 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1383 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1384 << ", across block:\n " 1385 << *BB << "\n"); 1386 1387 LVI->threadEdge(PredBB, BB, SuccBB); 1388 1389 // We are going to have to map operands from the original BB block to the new 1390 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1391 // account for entry from PredBB. 1392 DenseMap<Instruction*, Value*> ValueMapping; 1393 1394 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1395 BB->getName()+".thread", 1396 BB->getParent(), BB); 1397 NewBB->moveAfter(PredBB); 1398 1399 BasicBlock::iterator BI = BB->begin(); 1400 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1401 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1402 1403 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1404 // mapping and using it to remap operands in the cloned instructions. 1405 for (; !isa<TerminatorInst>(BI); ++BI) { 1406 Instruction *New = BI->clone(); 1407 New->setName(BI->getName()); 1408 NewBB->getInstList().push_back(New); 1409 ValueMapping[BI] = New; 1410 1411 // Remap operands to patch up intra-block references. 1412 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1413 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1414 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1415 if (I != ValueMapping.end()) 1416 New->setOperand(i, I->second); 1417 } 1418 } 1419 1420 // We didn't copy the terminator from BB over to NewBB, because there is now 1421 // an unconditional jump to SuccBB. Insert the unconditional jump. 1422 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB); 1423 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc()); 1424 1425 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1426 // PHI nodes for NewBB now. 1427 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1428 1429 // If there were values defined in BB that are used outside the block, then we 1430 // now have to update all uses of the value to use either the original value, 1431 // the cloned value, or some PHI derived value. This can require arbitrary 1432 // PHI insertion, of which we are prepared to do, clean these up now. 1433 SSAUpdater SSAUpdate; 1434 SmallVector<Use*, 16> UsesToRename; 1435 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1436 // Scan all uses of this instruction to see if it is used outside of its 1437 // block, and if so, record them in UsesToRename. 1438 for (Use &U : I->uses()) { 1439 Instruction *User = cast<Instruction>(U.getUser()); 1440 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1441 if (UserPN->getIncomingBlock(U) == BB) 1442 continue; 1443 } else if (User->getParent() == BB) 1444 continue; 1445 1446 UsesToRename.push_back(&U); 1447 } 1448 1449 // If there are no uses outside the block, we're done with this instruction. 1450 if (UsesToRename.empty()) 1451 continue; 1452 1453 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1454 1455 // We found a use of I outside of BB. Rename all uses of I that are outside 1456 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1457 // with the two values we know. 1458 SSAUpdate.Initialize(I->getType(), I->getName()); 1459 SSAUpdate.AddAvailableValue(BB, I); 1460 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1461 1462 while (!UsesToRename.empty()) 1463 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1464 DEBUG(dbgs() << "\n"); 1465 } 1466 1467 1468 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1469 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1470 // us to simplify any PHI nodes in BB. 1471 TerminatorInst *PredTerm = PredBB->getTerminator(); 1472 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1473 if (PredTerm->getSuccessor(i) == BB) { 1474 BB->removePredecessor(PredBB, true); 1475 PredTerm->setSuccessor(i, NewBB); 1476 } 1477 1478 // At this point, the IR is fully up to date and consistent. Do a quick scan 1479 // over the new instructions and zap any that are constants or dead. This 1480 // frequently happens because of phi translation. 1481 SimplifyInstructionsInBlock(NewBB, DL, TLI); 1482 1483 // Threaded an edge! 1484 ++NumThreads; 1485 return true; 1486} 1487 1488/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1489/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1490/// If we can duplicate the contents of BB up into PredBB do so now, this 1491/// improves the odds that the branch will be on an analyzable instruction like 1492/// a compare. 1493bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1494 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1495 assert(!PredBBs.empty() && "Can't handle an empty set"); 1496 1497 // If BB is a loop header, then duplicating this block outside the loop would 1498 // cause us to transform this into an irreducible loop, don't do this. 1499 // See the comments above FindLoopHeaders for justifications and caveats. 1500 if (LoopHeaders.count(BB)) { 1501 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1502 << "' into predecessor block '" << PredBBs[0]->getName() 1503 << "' - it might create an irreducible loop!\n"); 1504 return false; 1505 } 1506 1507 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold); 1508 if (DuplicationCost > Threshold) { 1509 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1510 << "' - Cost is too high: " << DuplicationCost << "\n"); 1511 return false; 1512 } 1513 1514 // And finally, do it! Start by factoring the predecessors is needed. 1515 BasicBlock *PredBB; 1516 if (PredBBs.size() == 1) 1517 PredBB = PredBBs[0]; 1518 else { 1519 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1520 << " common predecessors.\n"); 1521 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); 1522 } 1523 1524 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1525 // of PredBB. 1526 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1527 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1528 << DuplicationCost << " block is:" << *BB << "\n"); 1529 1530 // Unless PredBB ends with an unconditional branch, split the edge so that we 1531 // can just clone the bits from BB into the end of the new PredBB. 1532 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1533 1534 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1535 PredBB = SplitEdge(PredBB, BB, this); 1536 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1537 } 1538 1539 // We are going to have to map operands from the original BB block into the 1540 // PredBB block. Evaluate PHI nodes in BB. 1541 DenseMap<Instruction*, Value*> ValueMapping; 1542 1543 BasicBlock::iterator BI = BB->begin(); 1544 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1545 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1546 1547 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1548 // mapping and using it to remap operands in the cloned instructions. 1549 for (; BI != BB->end(); ++BI) { 1550 Instruction *New = BI->clone(); 1551 1552 // Remap operands to patch up intra-block references. 1553 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1554 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1555 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1556 if (I != ValueMapping.end()) 1557 New->setOperand(i, I->second); 1558 } 1559 1560 // If this instruction can be simplified after the operands are updated, 1561 // just use the simplified value instead. This frequently happens due to 1562 // phi translation. 1563 if (Value *IV = SimplifyInstruction(New, DL)) { 1564 delete New; 1565 ValueMapping[BI] = IV; 1566 } else { 1567 // Otherwise, insert the new instruction into the block. 1568 New->setName(BI->getName()); 1569 PredBB->getInstList().insert(OldPredBranch, New); 1570 ValueMapping[BI] = New; 1571 } 1572 } 1573 1574 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1575 // add entries to the PHI nodes for branch from PredBB now. 1576 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1577 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1578 ValueMapping); 1579 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1580 ValueMapping); 1581 1582 // If there were values defined in BB that are used outside the block, then we 1583 // now have to update all uses of the value to use either the original value, 1584 // the cloned value, or some PHI derived value. This can require arbitrary 1585 // PHI insertion, of which we are prepared to do, clean these up now. 1586 SSAUpdater SSAUpdate; 1587 SmallVector<Use*, 16> UsesToRename; 1588 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1589 // Scan all uses of this instruction to see if it is used outside of its 1590 // block, and if so, record them in UsesToRename. 1591 for (Use &U : I->uses()) { 1592 Instruction *User = cast<Instruction>(U.getUser()); 1593 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1594 if (UserPN->getIncomingBlock(U) == BB) 1595 continue; 1596 } else if (User->getParent() == BB) 1597 continue; 1598 1599 UsesToRename.push_back(&U); 1600 } 1601 1602 // If there are no uses outside the block, we're done with this instruction. 1603 if (UsesToRename.empty()) 1604 continue; 1605 1606 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1607 1608 // We found a use of I outside of BB. Rename all uses of I that are outside 1609 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1610 // with the two values we know. 1611 SSAUpdate.Initialize(I->getType(), I->getName()); 1612 SSAUpdate.AddAvailableValue(BB, I); 1613 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1614 1615 while (!UsesToRename.empty()) 1616 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1617 DEBUG(dbgs() << "\n"); 1618 } 1619 1620 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1621 // that we nuked. 1622 BB->removePredecessor(PredBB, true); 1623 1624 // Remove the unconditional branch at the end of the PredBB block. 1625 OldPredBranch->eraseFromParent(); 1626 1627 ++NumDupes; 1628 return true; 1629} 1630 1631/// TryToUnfoldSelect - Look for blocks of the form 1632/// bb1: 1633/// %a = select 1634/// br bb 1635/// 1636/// bb2: 1637/// %p = phi [%a, %bb] ... 1638/// %c = icmp %p 1639/// br i1 %c 1640/// 1641/// And expand the select into a branch structure if one of its arms allows %c 1642/// to be folded. This later enables threading from bb1 over bb2. 1643bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) { 1644 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 1645 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0)); 1646 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1)); 1647 1648 if (!CondBr || !CondBr->isConditional() || !CondLHS || 1649 CondLHS->getParent() != BB) 1650 return false; 1651 1652 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) { 1653 BasicBlock *Pred = CondLHS->getIncomingBlock(I); 1654 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I)); 1655 1656 // Look if one of the incoming values is a select in the corresponding 1657 // predecessor. 1658 if (!SI || SI->getParent() != Pred || !SI->hasOneUse()) 1659 continue; 1660 1661 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator()); 1662 if (!PredTerm || !PredTerm->isUnconditional()) 1663 continue; 1664 1665 // Now check if one of the select values would allow us to constant fold the 1666 // terminator in BB. We don't do the transform if both sides fold, those 1667 // cases will be threaded in any case. 1668 LazyValueInfo::Tristate LHSFolds = 1669 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1), 1670 CondRHS, Pred, BB); 1671 LazyValueInfo::Tristate RHSFolds = 1672 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2), 1673 CondRHS, Pred, BB); 1674 if ((LHSFolds != LazyValueInfo::Unknown || 1675 RHSFolds != LazyValueInfo::Unknown) && 1676 LHSFolds != RHSFolds) { 1677 // Expand the select. 1678 // 1679 // Pred -- 1680 // | v 1681 // | NewBB 1682 // | | 1683 // |----- 1684 // v 1685 // BB 1686 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold", 1687 BB->getParent(), BB); 1688 // Move the unconditional branch to NewBB. 1689 PredTerm->removeFromParent(); 1690 NewBB->getInstList().insert(NewBB->end(), PredTerm); 1691 // Create a conditional branch and update PHI nodes. 1692 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred); 1693 CondLHS->setIncomingValue(I, SI->getFalseValue()); 1694 CondLHS->addIncoming(SI->getTrueValue(), NewBB); 1695 // The select is now dead. 1696 SI->eraseFromParent(); 1697 1698 // Update any other PHI nodes in BB. 1699 for (BasicBlock::iterator BI = BB->begin(); 1700 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI) 1701 if (Phi != CondLHS) 1702 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB); 1703 return true; 1704 } 1705 } 1706 return false; 1707} 1708