JumpThreading.cpp revision c69908695af9cf509bf498a492854db5f0556e0f
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/ConstantFolding.h" 23#include "llvm/Analysis/InstructionSimplify.h" 24#include "llvm/Analysis/LazyValueInfo.h" 25#include "llvm/Analysis/Loads.h" 26#include "llvm/DataLayout.h" 27#include "llvm/IntrinsicInst.h" 28#include "llvm/LLVMContext.h" 29#include "llvm/Pass.h" 30#include "llvm/Support/CommandLine.h" 31#include "llvm/Support/Debug.h" 32#include "llvm/Support/ValueHandle.h" 33#include "llvm/Support/raw_ostream.h" 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 40STATISTIC(NumThreads, "Number of jumps threaded"); 41STATISTIC(NumFolds, "Number of terminators folded"); 42STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 43 44static cl::opt<unsigned> 45Threshold("jump-threading-threshold", 46 cl::desc("Max block size to duplicate for jump threading"), 47 cl::init(6), cl::Hidden); 48 49namespace { 50 // These are at global scope so static functions can use them too. 51 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo; 52 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy; 53 54 // This is used to keep track of what kind of constant we're currently hoping 55 // to find. 56 enum ConstantPreference { 57 WantInteger, 58 WantBlockAddress 59 }; 60 61 /// This pass performs 'jump threading', which looks at blocks that have 62 /// multiple predecessors and multiple successors. If one or more of the 63 /// predecessors of the block can be proven to always jump to one of the 64 /// successors, we forward the edge from the predecessor to the successor by 65 /// duplicating the contents of this block. 66 /// 67 /// An example of when this can occur is code like this: 68 /// 69 /// if () { ... 70 /// X = 4; 71 /// } 72 /// if (X < 3) { 73 /// 74 /// In this case, the unconditional branch at the end of the first if can be 75 /// revectored to the false side of the second if. 76 /// 77 class JumpThreading : public FunctionPass { 78 DataLayout *TD; 79 TargetLibraryInfo *TLI; 80 LazyValueInfo *LVI; 81#ifdef NDEBUG 82 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 83#else 84 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 85#endif 86 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet; 87 88 // RAII helper for updating the recursion stack. 89 struct RecursionSetRemover { 90 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet; 91 std::pair<Value*, BasicBlock*> ThePair; 92 93 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S, 94 std::pair<Value*, BasicBlock*> P) 95 : TheSet(S), ThePair(P) { } 96 97 ~RecursionSetRemover() { 98 TheSet.erase(ThePair); 99 } 100 }; 101 public: 102 static char ID; // Pass identification 103 JumpThreading() : FunctionPass(ID) { 104 initializeJumpThreadingPass(*PassRegistry::getPassRegistry()); 105 } 106 107 bool runOnFunction(Function &F); 108 109 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 110 AU.addRequired<LazyValueInfo>(); 111 AU.addPreserved<LazyValueInfo>(); 112 AU.addRequired<TargetLibraryInfo>(); 113 } 114 115 void FindLoopHeaders(Function &F); 116 bool ProcessBlock(BasicBlock *BB); 117 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 118 BasicBlock *SuccBB); 119 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 120 const SmallVectorImpl<BasicBlock *> &PredBBs); 121 122 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 123 PredValueInfo &Result, 124 ConstantPreference Preference); 125 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 126 ConstantPreference Preference); 127 128 bool ProcessBranchOnPHI(PHINode *PN); 129 bool ProcessBranchOnXOR(BinaryOperator *BO); 130 131 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 132 }; 133} 134 135char JumpThreading::ID = 0; 136INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", 137 "Jump Threading", false, false) 138INITIALIZE_PASS_DEPENDENCY(LazyValueInfo) 139INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 140INITIALIZE_PASS_END(JumpThreading, "jump-threading", 141 "Jump Threading", false, false) 142 143// Public interface to the Jump Threading pass 144FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 145 146/// runOnFunction - Top level algorithm. 147/// 148bool JumpThreading::runOnFunction(Function &F) { 149 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 150 TD = getAnalysisIfAvailable<DataLayout>(); 151 TLI = &getAnalysis<TargetLibraryInfo>(); 152 LVI = &getAnalysis<LazyValueInfo>(); 153 154 FindLoopHeaders(F); 155 156 bool Changed, EverChanged = false; 157 do { 158 Changed = false; 159 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 160 BasicBlock *BB = I; 161 // Thread all of the branches we can over this block. 162 while (ProcessBlock(BB)) 163 Changed = true; 164 165 ++I; 166 167 // If the block is trivially dead, zap it. This eliminates the successor 168 // edges which simplifies the CFG. 169 if (pred_begin(BB) == pred_end(BB) && 170 BB != &BB->getParent()->getEntryBlock()) { 171 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 172 << "' with terminator: " << *BB->getTerminator() << '\n'); 173 LoopHeaders.erase(BB); 174 LVI->eraseBlock(BB); 175 DeleteDeadBlock(BB); 176 Changed = true; 177 continue; 178 } 179 180 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 181 182 // Can't thread an unconditional jump, but if the block is "almost 183 // empty", we can replace uses of it with uses of the successor and make 184 // this dead. 185 if (BI && BI->isUnconditional() && 186 BB != &BB->getParent()->getEntryBlock() && 187 // If the terminator is the only non-phi instruction, try to nuke it. 188 BB->getFirstNonPHIOrDbg()->isTerminator()) { 189 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 190 // block, we have to make sure it isn't in the LoopHeaders set. We 191 // reinsert afterward if needed. 192 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 193 BasicBlock *Succ = BI->getSuccessor(0); 194 195 // FIXME: It is always conservatively correct to drop the info 196 // for a block even if it doesn't get erased. This isn't totally 197 // awesome, but it allows us to use AssertingVH to prevent nasty 198 // dangling pointer issues within LazyValueInfo. 199 LVI->eraseBlock(BB); 200 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 201 Changed = true; 202 // If we deleted BB and BB was the header of a loop, then the 203 // successor is now the header of the loop. 204 BB = Succ; 205 } 206 207 if (ErasedFromLoopHeaders) 208 LoopHeaders.insert(BB); 209 } 210 } 211 EverChanged |= Changed; 212 } while (Changed); 213 214 LoopHeaders.clear(); 215 return EverChanged; 216} 217 218/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 219/// thread across it. Stop scanning the block when passing the threshold. 220static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB, 221 unsigned Threshold) { 222 /// Ignore PHI nodes, these will be flattened when duplication happens. 223 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 224 225 // FIXME: THREADING will delete values that are just used to compute the 226 // branch, so they shouldn't count against the duplication cost. 227 228 // Sum up the cost of each instruction until we get to the terminator. Don't 229 // include the terminator because the copy won't include it. 230 unsigned Size = 0; 231 for (; !isa<TerminatorInst>(I); ++I) { 232 233 // Stop scanning the block if we've reached the threshold. 234 if (Size > Threshold) 235 return Size; 236 237 // Debugger intrinsics don't incur code size. 238 if (isa<DbgInfoIntrinsic>(I)) continue; 239 240 // If this is a pointer->pointer bitcast, it is free. 241 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 242 continue; 243 244 // All other instructions count for at least one unit. 245 ++Size; 246 247 // Calls are more expensive. If they are non-intrinsic calls, we model them 248 // as having cost of 4. If they are a non-vector intrinsic, we model them 249 // as having cost of 2 total, and if they are a vector intrinsic, we model 250 // them as having cost 1. 251 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 252 if (!isa<IntrinsicInst>(CI)) 253 Size += 3; 254 else if (!CI->getType()->isVectorTy()) 255 Size += 1; 256 } 257 } 258 259 // Threading through a switch statement is particularly profitable. If this 260 // block ends in a switch, decrease its cost to make it more likely to happen. 261 if (isa<SwitchInst>(I)) 262 Size = Size > 6 ? Size-6 : 0; 263 264 // The same holds for indirect branches, but slightly more so. 265 if (isa<IndirectBrInst>(I)) 266 Size = Size > 8 ? Size-8 : 0; 267 268 return Size; 269} 270 271/// FindLoopHeaders - We do not want jump threading to turn proper loop 272/// structures into irreducible loops. Doing this breaks up the loop nesting 273/// hierarchy and pessimizes later transformations. To prevent this from 274/// happening, we first have to find the loop headers. Here we approximate this 275/// by finding targets of backedges in the CFG. 276/// 277/// Note that there definitely are cases when we want to allow threading of 278/// edges across a loop header. For example, threading a jump from outside the 279/// loop (the preheader) to an exit block of the loop is definitely profitable. 280/// It is also almost always profitable to thread backedges from within the loop 281/// to exit blocks, and is often profitable to thread backedges to other blocks 282/// within the loop (forming a nested loop). This simple analysis is not rich 283/// enough to track all of these properties and keep it up-to-date as the CFG 284/// mutates, so we don't allow any of these transformations. 285/// 286void JumpThreading::FindLoopHeaders(Function &F) { 287 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 288 FindFunctionBackedges(F, Edges); 289 290 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 291 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 292} 293 294/// getKnownConstant - Helper method to determine if we can thread over a 295/// terminator with the given value as its condition, and if so what value to 296/// use for that. What kind of value this is depends on whether we want an 297/// integer or a block address, but an undef is always accepted. 298/// Returns null if Val is null or not an appropriate constant. 299static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { 300 if (!Val) 301 return 0; 302 303 // Undef is "known" enough. 304 if (UndefValue *U = dyn_cast<UndefValue>(Val)) 305 return U; 306 307 if (Preference == WantBlockAddress) 308 return dyn_cast<BlockAddress>(Val->stripPointerCasts()); 309 310 return dyn_cast<ConstantInt>(Val); 311} 312 313/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 314/// if we can infer that the value is a known ConstantInt/BlockAddress or undef 315/// in any of our predecessors. If so, return the known list of value and pred 316/// BB in the result vector. 317/// 318/// This returns true if there were any known values. 319/// 320bool JumpThreading:: 321ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result, 322 ConstantPreference Preference) { 323 // This method walks up use-def chains recursively. Because of this, we could 324 // get into an infinite loop going around loops in the use-def chain. To 325 // prevent this, keep track of what (value, block) pairs we've already visited 326 // and terminate the search if we loop back to them 327 if (!RecursionSet.insert(std::make_pair(V, BB)).second) 328 return false; 329 330 // An RAII help to remove this pair from the recursion set once the recursion 331 // stack pops back out again. 332 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB)); 333 334 // If V is a constant, then it is known in all predecessors. 335 if (Constant *KC = getKnownConstant(V, Preference)) { 336 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 337 Result.push_back(std::make_pair(KC, *PI)); 338 339 return true; 340 } 341 342 // If V is a non-instruction value, or an instruction in a different block, 343 // then it can't be derived from a PHI. 344 Instruction *I = dyn_cast<Instruction>(V); 345 if (I == 0 || I->getParent() != BB) { 346 347 // Okay, if this is a live-in value, see if it has a known value at the end 348 // of any of our predecessors. 349 // 350 // FIXME: This should be an edge property, not a block end property. 351 /// TODO: Per PR2563, we could infer value range information about a 352 /// predecessor based on its terminator. 353 // 354 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 355 // "I" is a non-local compare-with-a-constant instruction. This would be 356 // able to handle value inequalities better, for example if the compare is 357 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 358 // Perhaps getConstantOnEdge should be smart enough to do this? 359 360 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 361 BasicBlock *P = *PI; 362 // If the value is known by LazyValueInfo to be a constant in a 363 // predecessor, use that information to try to thread this block. 364 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 365 if (Constant *KC = getKnownConstant(PredCst, Preference)) 366 Result.push_back(std::make_pair(KC, P)); 367 } 368 369 return !Result.empty(); 370 } 371 372 /// If I is a PHI node, then we know the incoming values for any constants. 373 if (PHINode *PN = dyn_cast<PHINode>(I)) { 374 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 375 Value *InVal = PN->getIncomingValue(i); 376 if (Constant *KC = getKnownConstant(InVal, Preference)) { 377 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 378 } else { 379 Constant *CI = LVI->getConstantOnEdge(InVal, 380 PN->getIncomingBlock(i), BB); 381 if (Constant *KC = getKnownConstant(CI, Preference)) 382 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 383 } 384 } 385 386 return !Result.empty(); 387 } 388 389 PredValueInfoTy LHSVals, RHSVals; 390 391 // Handle some boolean conditions. 392 if (I->getType()->getPrimitiveSizeInBits() == 1) { 393 assert(Preference == WantInteger && "One-bit non-integer type?"); 394 // X | true -> true 395 // X & false -> false 396 if (I->getOpcode() == Instruction::Or || 397 I->getOpcode() == Instruction::And) { 398 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 399 WantInteger); 400 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals, 401 WantInteger); 402 403 if (LHSVals.empty() && RHSVals.empty()) 404 return false; 405 406 ConstantInt *InterestingVal; 407 if (I->getOpcode() == Instruction::Or) 408 InterestingVal = ConstantInt::getTrue(I->getContext()); 409 else 410 InterestingVal = ConstantInt::getFalse(I->getContext()); 411 412 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 413 414 // Scan for the sentinel. If we find an undef, force it to the 415 // interesting value: x|undef -> true and x&undef -> false. 416 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 417 if (LHSVals[i].first == InterestingVal || 418 isa<UndefValue>(LHSVals[i].first)) { 419 Result.push_back(LHSVals[i]); 420 Result.back().first = InterestingVal; 421 LHSKnownBBs.insert(LHSVals[i].second); 422 } 423 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 424 if (RHSVals[i].first == InterestingVal || 425 isa<UndefValue>(RHSVals[i].first)) { 426 // If we already inferred a value for this block on the LHS, don't 427 // re-add it. 428 if (!LHSKnownBBs.count(RHSVals[i].second)) { 429 Result.push_back(RHSVals[i]); 430 Result.back().first = InterestingVal; 431 } 432 } 433 434 return !Result.empty(); 435 } 436 437 // Handle the NOT form of XOR. 438 if (I->getOpcode() == Instruction::Xor && 439 isa<ConstantInt>(I->getOperand(1)) && 440 cast<ConstantInt>(I->getOperand(1))->isOne()) { 441 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result, 442 WantInteger); 443 if (Result.empty()) 444 return false; 445 446 // Invert the known values. 447 for (unsigned i = 0, e = Result.size(); i != e; ++i) 448 Result[i].first = ConstantExpr::getNot(Result[i].first); 449 450 return true; 451 } 452 453 // Try to simplify some other binary operator values. 454 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 455 assert(Preference != WantBlockAddress 456 && "A binary operator creating a block address?"); 457 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 458 PredValueInfoTy LHSVals; 459 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals, 460 WantInteger); 461 462 // Try to use constant folding to simplify the binary operator. 463 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 464 Constant *V = LHSVals[i].first; 465 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); 466 467 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 468 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 469 } 470 } 471 472 return !Result.empty(); 473 } 474 475 // Handle compare with phi operand, where the PHI is defined in this block. 476 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 477 assert(Preference == WantInteger && "Compares only produce integers"); 478 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 479 if (PN && PN->getParent() == BB) { 480 // We can do this simplification if any comparisons fold to true or false. 481 // See if any do. 482 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 483 BasicBlock *PredBB = PN->getIncomingBlock(i); 484 Value *LHS = PN->getIncomingValue(i); 485 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 486 487 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 488 if (Res == 0) { 489 if (!isa<Constant>(RHS)) 490 continue; 491 492 LazyValueInfo::Tristate 493 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 494 cast<Constant>(RHS), PredBB, BB); 495 if (ResT == LazyValueInfo::Unknown) 496 continue; 497 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 498 } 499 500 if (Constant *KC = getKnownConstant(Res, WantInteger)) 501 Result.push_back(std::make_pair(KC, PredBB)); 502 } 503 504 return !Result.empty(); 505 } 506 507 508 // If comparing a live-in value against a constant, see if we know the 509 // live-in value on any predecessors. 510 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) { 511 if (!isa<Instruction>(Cmp->getOperand(0)) || 512 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 513 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 514 515 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ 516 BasicBlock *P = *PI; 517 // If the value is known by LazyValueInfo to be a constant in a 518 // predecessor, use that information to try to thread this block. 519 LazyValueInfo::Tristate Res = 520 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 521 RHSCst, P, BB); 522 if (Res == LazyValueInfo::Unknown) 523 continue; 524 525 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 526 Result.push_back(std::make_pair(ResC, P)); 527 } 528 529 return !Result.empty(); 530 } 531 532 // Try to find a constant value for the LHS of a comparison, 533 // and evaluate it statically if we can. 534 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { 535 PredValueInfoTy LHSVals; 536 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 537 WantInteger); 538 539 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 540 Constant *V = LHSVals[i].first; 541 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(), 542 V, CmpConst); 543 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 544 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 545 } 546 547 return !Result.empty(); 548 } 549 } 550 } 551 552 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 553 // Handle select instructions where at least one operand is a known constant 554 // and we can figure out the condition value for any predecessor block. 555 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference); 556 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference); 557 PredValueInfoTy Conds; 558 if ((TrueVal || FalseVal) && 559 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds, 560 WantInteger)) { 561 for (unsigned i = 0, e = Conds.size(); i != e; ++i) { 562 Constant *Cond = Conds[i].first; 563 564 // Figure out what value to use for the condition. 565 bool KnownCond; 566 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) { 567 // A known boolean. 568 KnownCond = CI->isOne(); 569 } else { 570 assert(isa<UndefValue>(Cond) && "Unexpected condition value"); 571 // Either operand will do, so be sure to pick the one that's a known 572 // constant. 573 // FIXME: Do this more cleverly if both values are known constants? 574 KnownCond = (TrueVal != 0); 575 } 576 577 // See if the select has a known constant value for this predecessor. 578 if (Constant *Val = KnownCond ? TrueVal : FalseVal) 579 Result.push_back(std::make_pair(Val, Conds[i].second)); 580 } 581 582 return !Result.empty(); 583 } 584 } 585 586 // If all else fails, see if LVI can figure out a constant value for us. 587 Constant *CI = LVI->getConstant(V, BB); 588 if (Constant *KC = getKnownConstant(CI, Preference)) { 589 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 590 Result.push_back(std::make_pair(KC, *PI)); 591 } 592 593 return !Result.empty(); 594} 595 596 597 598/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 599/// in an undefined jump, decide which block is best to revector to. 600/// 601/// Since we can pick an arbitrary destination, we pick the successor with the 602/// fewest predecessors. This should reduce the in-degree of the others. 603/// 604static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 605 TerminatorInst *BBTerm = BB->getTerminator(); 606 unsigned MinSucc = 0; 607 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 608 // Compute the successor with the minimum number of predecessors. 609 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 610 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 611 TestBB = BBTerm->getSuccessor(i); 612 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 613 if (NumPreds < MinNumPreds) { 614 MinSucc = i; 615 MinNumPreds = NumPreds; 616 } 617 } 618 619 return MinSucc; 620} 621 622static bool hasAddressTakenAndUsed(BasicBlock *BB) { 623 if (!BB->hasAddressTaken()) return false; 624 625 // If the block has its address taken, it may be a tree of dead constants 626 // hanging off of it. These shouldn't keep the block alive. 627 BlockAddress *BA = BlockAddress::get(BB); 628 BA->removeDeadConstantUsers(); 629 return !BA->use_empty(); 630} 631 632/// ProcessBlock - If there are any predecessors whose control can be threaded 633/// through to a successor, transform them now. 634bool JumpThreading::ProcessBlock(BasicBlock *BB) { 635 // If the block is trivially dead, just return and let the caller nuke it. 636 // This simplifies other transformations. 637 if (pred_begin(BB) == pred_end(BB) && 638 BB != &BB->getParent()->getEntryBlock()) 639 return false; 640 641 // If this block has a single predecessor, and if that pred has a single 642 // successor, merge the blocks. This encourages recursive jump threading 643 // because now the condition in this block can be threaded through 644 // predecessors of our predecessor block. 645 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 646 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 647 SinglePred != BB && !hasAddressTakenAndUsed(BB)) { 648 // If SinglePred was a loop header, BB becomes one. 649 if (LoopHeaders.erase(SinglePred)) 650 LoopHeaders.insert(BB); 651 652 // Remember if SinglePred was the entry block of the function. If so, we 653 // will need to move BB back to the entry position. 654 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 655 LVI->eraseBlock(SinglePred); 656 MergeBasicBlockIntoOnlyPred(BB); 657 658 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 659 BB->moveBefore(&BB->getParent()->getEntryBlock()); 660 return true; 661 } 662 } 663 664 // What kind of constant we're looking for. 665 ConstantPreference Preference = WantInteger; 666 667 // Look to see if the terminator is a conditional branch, switch or indirect 668 // branch, if not we can't thread it. 669 Value *Condition; 670 Instruction *Terminator = BB->getTerminator(); 671 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { 672 // Can't thread an unconditional jump. 673 if (BI->isUnconditional()) return false; 674 Condition = BI->getCondition(); 675 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { 676 Condition = SI->getCondition(); 677 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { 678 // Can't thread indirect branch with no successors. 679 if (IB->getNumSuccessors() == 0) return false; 680 Condition = IB->getAddress()->stripPointerCasts(); 681 Preference = WantBlockAddress; 682 } else { 683 return false; // Must be an invoke. 684 } 685 686 // Run constant folding to see if we can reduce the condition to a simple 687 // constant. 688 if (Instruction *I = dyn_cast<Instruction>(Condition)) { 689 Value *SimpleVal = ConstantFoldInstruction(I, TD, TLI); 690 if (SimpleVal) { 691 I->replaceAllUsesWith(SimpleVal); 692 I->eraseFromParent(); 693 Condition = SimpleVal; 694 } 695 } 696 697 // If the terminator is branching on an undef, we can pick any of the 698 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 699 if (isa<UndefValue>(Condition)) { 700 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 701 702 // Fold the branch/switch. 703 TerminatorInst *BBTerm = BB->getTerminator(); 704 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 705 if (i == BestSucc) continue; 706 BBTerm->getSuccessor(i)->removePredecessor(BB, true); 707 } 708 709 DEBUG(dbgs() << " In block '" << BB->getName() 710 << "' folding undef terminator: " << *BBTerm << '\n'); 711 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 712 BBTerm->eraseFromParent(); 713 return true; 714 } 715 716 // If the terminator of this block is branching on a constant, simplify the 717 // terminator to an unconditional branch. This can occur due to threading in 718 // other blocks. 719 if (getKnownConstant(Condition, Preference)) { 720 DEBUG(dbgs() << " In block '" << BB->getName() 721 << "' folding terminator: " << *BB->getTerminator() << '\n'); 722 ++NumFolds; 723 ConstantFoldTerminator(BB, true); 724 return true; 725 } 726 727 Instruction *CondInst = dyn_cast<Instruction>(Condition); 728 729 // All the rest of our checks depend on the condition being an instruction. 730 if (CondInst == 0) { 731 // FIXME: Unify this with code below. 732 if (ProcessThreadableEdges(Condition, BB, Preference)) 733 return true; 734 return false; 735 } 736 737 738 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 739 // For a comparison where the LHS is outside this block, it's possible 740 // that we've branched on it before. Used LVI to see if we can simplify 741 // the branch based on that. 742 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 743 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 744 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 745 if (CondBr && CondConst && CondBr->isConditional() && PI != PE && 746 (!isa<Instruction>(CondCmp->getOperand(0)) || 747 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { 748 // For predecessor edge, determine if the comparison is true or false 749 // on that edge. If they're all true or all false, we can simplify the 750 // branch. 751 // FIXME: We could handle mixed true/false by duplicating code. 752 LazyValueInfo::Tristate Baseline = 753 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0), 754 CondConst, *PI, BB); 755 if (Baseline != LazyValueInfo::Unknown) { 756 // Check that all remaining incoming values match the first one. 757 while (++PI != PE) { 758 LazyValueInfo::Tristate Ret = 759 LVI->getPredicateOnEdge(CondCmp->getPredicate(), 760 CondCmp->getOperand(0), CondConst, *PI, BB); 761 if (Ret != Baseline) break; 762 } 763 764 // If we terminated early, then one of the values didn't match. 765 if (PI == PE) { 766 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0; 767 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1; 768 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); 769 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); 770 CondBr->eraseFromParent(); 771 return true; 772 } 773 } 774 } 775 } 776 777 // Check for some cases that are worth simplifying. Right now we want to look 778 // for loads that are used by a switch or by the condition for the branch. If 779 // we see one, check to see if it's partially redundant. If so, insert a PHI 780 // which can then be used to thread the values. 781 // 782 Value *SimplifyValue = CondInst; 783 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 784 if (isa<Constant>(CondCmp->getOperand(1))) 785 SimplifyValue = CondCmp->getOperand(0); 786 787 // TODO: There are other places where load PRE would be profitable, such as 788 // more complex comparisons. 789 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 790 if (SimplifyPartiallyRedundantLoad(LI)) 791 return true; 792 793 794 // Handle a variety of cases where we are branching on something derived from 795 // a PHI node in the current block. If we can prove that any predecessors 796 // compute a predictable value based on a PHI node, thread those predecessors. 797 // 798 if (ProcessThreadableEdges(CondInst, BB, Preference)) 799 return true; 800 801 // If this is an otherwise-unfoldable branch on a phi node in the current 802 // block, see if we can simplify. 803 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 804 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 805 return ProcessBranchOnPHI(PN); 806 807 808 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 809 if (CondInst->getOpcode() == Instruction::Xor && 810 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 811 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 812 813 814 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 815 // "(X == 4)", thread through this block. 816 817 return false; 818} 819 820 821/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 822/// load instruction, eliminate it by replacing it with a PHI node. This is an 823/// important optimization that encourages jump threading, and needs to be run 824/// interlaced with other jump threading tasks. 825bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 826 // Don't hack volatile/atomic loads. 827 if (!LI->isSimple()) return false; 828 829 // If the load is defined in a block with exactly one predecessor, it can't be 830 // partially redundant. 831 BasicBlock *LoadBB = LI->getParent(); 832 if (LoadBB->getSinglePredecessor()) 833 return false; 834 835 Value *LoadedPtr = LI->getOperand(0); 836 837 // If the loaded operand is defined in the LoadBB, it can't be available. 838 // TODO: Could do simple PHI translation, that would be fun :) 839 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 840 if (PtrOp->getParent() == LoadBB) 841 return false; 842 843 // Scan a few instructions up from the load, to see if it is obviously live at 844 // the entry to its block. 845 BasicBlock::iterator BBIt = LI; 846 847 if (Value *AvailableVal = 848 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 849 // If the value if the load is locally available within the block, just use 850 // it. This frequently occurs for reg2mem'd allocas. 851 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 852 853 // If the returned value is the load itself, replace with an undef. This can 854 // only happen in dead loops. 855 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 856 LI->replaceAllUsesWith(AvailableVal); 857 LI->eraseFromParent(); 858 return true; 859 } 860 861 // Otherwise, if we scanned the whole block and got to the top of the block, 862 // we know the block is locally transparent to the load. If not, something 863 // might clobber its value. 864 if (BBIt != LoadBB->begin()) 865 return false; 866 867 // If all of the loads and stores that feed the value have the same TBAA tag, 868 // then we can propagate it onto any newly inserted loads. 869 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa); 870 871 SmallPtrSet<BasicBlock*, 8> PredsScanned; 872 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 873 AvailablePredsTy AvailablePreds; 874 BasicBlock *OneUnavailablePred = 0; 875 876 // If we got here, the loaded value is transparent through to the start of the 877 // block. Check to see if it is available in any of the predecessor blocks. 878 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 879 PI != PE; ++PI) { 880 BasicBlock *PredBB = *PI; 881 882 // If we already scanned this predecessor, skip it. 883 if (!PredsScanned.insert(PredBB)) 884 continue; 885 886 // Scan the predecessor to see if the value is available in the pred. 887 BBIt = PredBB->end(); 888 MDNode *ThisTBAATag = 0; 889 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6, 890 0, &ThisTBAATag); 891 if (!PredAvailable) { 892 OneUnavailablePred = PredBB; 893 continue; 894 } 895 896 // If tbaa tags disagree or are not present, forget about them. 897 if (TBAATag != ThisTBAATag) TBAATag = 0; 898 899 // If so, this load is partially redundant. Remember this info so that we 900 // can create a PHI node. 901 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 902 } 903 904 // If the loaded value isn't available in any predecessor, it isn't partially 905 // redundant. 906 if (AvailablePreds.empty()) return false; 907 908 // Okay, the loaded value is available in at least one (and maybe all!) 909 // predecessors. If the value is unavailable in more than one unique 910 // predecessor, we want to insert a merge block for those common predecessors. 911 // This ensures that we only have to insert one reload, thus not increasing 912 // code size. 913 BasicBlock *UnavailablePred = 0; 914 915 // If there is exactly one predecessor where the value is unavailable, the 916 // already computed 'OneUnavailablePred' block is it. If it ends in an 917 // unconditional branch, we know that it isn't a critical edge. 918 if (PredsScanned.size() == AvailablePreds.size()+1 && 919 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 920 UnavailablePred = OneUnavailablePred; 921 } else if (PredsScanned.size() != AvailablePreds.size()) { 922 // Otherwise, we had multiple unavailable predecessors or we had a critical 923 // edge from the one. 924 SmallVector<BasicBlock*, 8> PredsToSplit; 925 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 926 927 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 928 AvailablePredSet.insert(AvailablePreds[i].first); 929 930 // Add all the unavailable predecessors to the PredsToSplit list. 931 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 932 PI != PE; ++PI) { 933 BasicBlock *P = *PI; 934 // If the predecessor is an indirect goto, we can't split the edge. 935 if (isa<IndirectBrInst>(P->getTerminator())) 936 return false; 937 938 if (!AvailablePredSet.count(P)) 939 PredsToSplit.push_back(P); 940 } 941 942 // Split them out to their own block. 943 UnavailablePred = 944 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this); 945 } 946 947 // If the value isn't available in all predecessors, then there will be 948 // exactly one where it isn't available. Insert a load on that edge and add 949 // it to the AvailablePreds list. 950 if (UnavailablePred) { 951 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 952 "Can't handle critical edge here!"); 953 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 954 LI->getAlignment(), 955 UnavailablePred->getTerminator()); 956 NewVal->setDebugLoc(LI->getDebugLoc()); 957 if (TBAATag) 958 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag); 959 960 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 961 } 962 963 // Now we know that each predecessor of this block has a value in 964 // AvailablePreds, sort them for efficient access as we're walking the preds. 965 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 966 967 // Create a PHI node at the start of the block for the PRE'd load value. 968 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB); 969 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "", 970 LoadBB->begin()); 971 PN->takeName(LI); 972 PN->setDebugLoc(LI->getDebugLoc()); 973 974 // Insert new entries into the PHI for each predecessor. A single block may 975 // have multiple entries here. 976 for (pred_iterator PI = PB; PI != PE; ++PI) { 977 BasicBlock *P = *PI; 978 AvailablePredsTy::iterator I = 979 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 980 std::make_pair(P, (Value*)0)); 981 982 assert(I != AvailablePreds.end() && I->first == P && 983 "Didn't find entry for predecessor!"); 984 985 PN->addIncoming(I->second, I->first); 986 } 987 988 //cerr << "PRE: " << *LI << *PN << "\n"; 989 990 LI->replaceAllUsesWith(PN); 991 LI->eraseFromParent(); 992 993 return true; 994} 995 996/// FindMostPopularDest - The specified list contains multiple possible 997/// threadable destinations. Pick the one that occurs the most frequently in 998/// the list. 999static BasicBlock * 1000FindMostPopularDest(BasicBlock *BB, 1001 const SmallVectorImpl<std::pair<BasicBlock*, 1002 BasicBlock*> > &PredToDestList) { 1003 assert(!PredToDestList.empty()); 1004 1005 // Determine popularity. If there are multiple possible destinations, we 1006 // explicitly choose to ignore 'undef' destinations. We prefer to thread 1007 // blocks with known and real destinations to threading undef. We'll handle 1008 // them later if interesting. 1009 DenseMap<BasicBlock*, unsigned> DestPopularity; 1010 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1011 if (PredToDestList[i].second) 1012 DestPopularity[PredToDestList[i].second]++; 1013 1014 // Find the most popular dest. 1015 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 1016 BasicBlock *MostPopularDest = DPI->first; 1017 unsigned Popularity = DPI->second; 1018 SmallVector<BasicBlock*, 4> SamePopularity; 1019 1020 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 1021 // If the popularity of this entry isn't higher than the popularity we've 1022 // seen so far, ignore it. 1023 if (DPI->second < Popularity) 1024 ; // ignore. 1025 else if (DPI->second == Popularity) { 1026 // If it is the same as what we've seen so far, keep track of it. 1027 SamePopularity.push_back(DPI->first); 1028 } else { 1029 // If it is more popular, remember it. 1030 SamePopularity.clear(); 1031 MostPopularDest = DPI->first; 1032 Popularity = DPI->second; 1033 } 1034 } 1035 1036 // Okay, now we know the most popular destination. If there is more than one 1037 // destination, we need to determine one. This is arbitrary, but we need 1038 // to make a deterministic decision. Pick the first one that appears in the 1039 // successor list. 1040 if (!SamePopularity.empty()) { 1041 SamePopularity.push_back(MostPopularDest); 1042 TerminatorInst *TI = BB->getTerminator(); 1043 for (unsigned i = 0; ; ++i) { 1044 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1045 1046 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1047 TI->getSuccessor(i)) == SamePopularity.end()) 1048 continue; 1049 1050 MostPopularDest = TI->getSuccessor(i); 1051 break; 1052 } 1053 } 1054 1055 // Okay, we have finally picked the most popular destination. 1056 return MostPopularDest; 1057} 1058 1059bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 1060 ConstantPreference Preference) { 1061 // If threading this would thread across a loop header, don't even try to 1062 // thread the edge. 1063 if (LoopHeaders.count(BB)) 1064 return false; 1065 1066 PredValueInfoTy PredValues; 1067 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference)) 1068 return false; 1069 1070 assert(!PredValues.empty() && 1071 "ComputeValueKnownInPredecessors returned true with no values"); 1072 1073 DEBUG(dbgs() << "IN BB: " << *BB; 1074 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1075 dbgs() << " BB '" << BB->getName() << "': FOUND condition = " 1076 << *PredValues[i].first 1077 << " for pred '" << PredValues[i].second->getName() << "'.\n"; 1078 }); 1079 1080 // Decide what we want to thread through. Convert our list of known values to 1081 // a list of known destinations for each pred. This also discards duplicate 1082 // predecessors and keeps track of the undefined inputs (which are represented 1083 // as a null dest in the PredToDestList). 1084 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1085 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1086 1087 BasicBlock *OnlyDest = 0; 1088 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1089 1090 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1091 BasicBlock *Pred = PredValues[i].second; 1092 if (!SeenPreds.insert(Pred)) 1093 continue; // Duplicate predecessor entry. 1094 1095 // If the predecessor ends with an indirect goto, we can't change its 1096 // destination. 1097 if (isa<IndirectBrInst>(Pred->getTerminator())) 1098 continue; 1099 1100 Constant *Val = PredValues[i].first; 1101 1102 BasicBlock *DestBB; 1103 if (isa<UndefValue>(Val)) 1104 DestBB = 0; 1105 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1106 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); 1107 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 1108 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor(); 1109 } else { 1110 assert(isa<IndirectBrInst>(BB->getTerminator()) 1111 && "Unexpected terminator"); 1112 DestBB = cast<BlockAddress>(Val)->getBasicBlock(); 1113 } 1114 1115 // If we have exactly one destination, remember it for efficiency below. 1116 if (PredToDestList.empty()) 1117 OnlyDest = DestBB; 1118 else if (OnlyDest != DestBB) 1119 OnlyDest = MultipleDestSentinel; 1120 1121 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1122 } 1123 1124 // If all edges were unthreadable, we fail. 1125 if (PredToDestList.empty()) 1126 return false; 1127 1128 // Determine which is the most common successor. If we have many inputs and 1129 // this block is a switch, we want to start by threading the batch that goes 1130 // to the most popular destination first. If we only know about one 1131 // threadable destination (the common case) we can avoid this. 1132 BasicBlock *MostPopularDest = OnlyDest; 1133 1134 if (MostPopularDest == MultipleDestSentinel) 1135 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1136 1137 // Now that we know what the most popular destination is, factor all 1138 // predecessors that will jump to it into a single predecessor. 1139 SmallVector<BasicBlock*, 16> PredsToFactor; 1140 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1141 if (PredToDestList[i].second == MostPopularDest) { 1142 BasicBlock *Pred = PredToDestList[i].first; 1143 1144 // This predecessor may be a switch or something else that has multiple 1145 // edges to the block. Factor each of these edges by listing them 1146 // according to # occurrences in PredsToFactor. 1147 TerminatorInst *PredTI = Pred->getTerminator(); 1148 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1149 if (PredTI->getSuccessor(i) == BB) 1150 PredsToFactor.push_back(Pred); 1151 } 1152 1153 // If the threadable edges are branching on an undefined value, we get to pick 1154 // the destination that these predecessors should get to. 1155 if (MostPopularDest == 0) 1156 MostPopularDest = BB->getTerminator()-> 1157 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1158 1159 // Ok, try to thread it! 1160 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1161} 1162 1163/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1164/// a PHI node in the current block. See if there are any simplifications we 1165/// can do based on inputs to the phi node. 1166/// 1167bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1168 BasicBlock *BB = PN->getParent(); 1169 1170 // TODO: We could make use of this to do it once for blocks with common PHI 1171 // values. 1172 SmallVector<BasicBlock*, 1> PredBBs; 1173 PredBBs.resize(1); 1174 1175 // If any of the predecessor blocks end in an unconditional branch, we can 1176 // *duplicate* the conditional branch into that block in order to further 1177 // encourage jump threading and to eliminate cases where we have branch on a 1178 // phi of an icmp (branch on icmp is much better). 1179 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1180 BasicBlock *PredBB = PN->getIncomingBlock(i); 1181 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1182 if (PredBr->isUnconditional()) { 1183 PredBBs[0] = PredBB; 1184 // Try to duplicate BB into PredBB. 1185 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1186 return true; 1187 } 1188 } 1189 1190 return false; 1191} 1192 1193/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1194/// a xor instruction in the current block. See if there are any 1195/// simplifications we can do based on inputs to the xor. 1196/// 1197bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1198 BasicBlock *BB = BO->getParent(); 1199 1200 // If either the LHS or RHS of the xor is a constant, don't do this 1201 // optimization. 1202 if (isa<ConstantInt>(BO->getOperand(0)) || 1203 isa<ConstantInt>(BO->getOperand(1))) 1204 return false; 1205 1206 // If the first instruction in BB isn't a phi, we won't be able to infer 1207 // anything special about any particular predecessor. 1208 if (!isa<PHINode>(BB->front())) 1209 return false; 1210 1211 // If we have a xor as the branch input to this block, and we know that the 1212 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1213 // the condition into the predecessor and fix that value to true, saving some 1214 // logical ops on that path and encouraging other paths to simplify. 1215 // 1216 // This copies something like this: 1217 // 1218 // BB: 1219 // %X = phi i1 [1], [%X'] 1220 // %Y = icmp eq i32 %A, %B 1221 // %Z = xor i1 %X, %Y 1222 // br i1 %Z, ... 1223 // 1224 // Into: 1225 // BB': 1226 // %Y = icmp ne i32 %A, %B 1227 // br i1 %Z, ... 1228 1229 PredValueInfoTy XorOpValues; 1230 bool isLHS = true; 1231 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, 1232 WantInteger)) { 1233 assert(XorOpValues.empty()); 1234 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, 1235 WantInteger)) 1236 return false; 1237 isLHS = false; 1238 } 1239 1240 assert(!XorOpValues.empty() && 1241 "ComputeValueKnownInPredecessors returned true with no values"); 1242 1243 // Scan the information to see which is most popular: true or false. The 1244 // predecessors can be of the set true, false, or undef. 1245 unsigned NumTrue = 0, NumFalse = 0; 1246 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1247 if (isa<UndefValue>(XorOpValues[i].first)) 1248 // Ignore undefs for the count. 1249 continue; 1250 if (cast<ConstantInt>(XorOpValues[i].first)->isZero()) 1251 ++NumFalse; 1252 else 1253 ++NumTrue; 1254 } 1255 1256 // Determine which value to split on, true, false, or undef if neither. 1257 ConstantInt *SplitVal = 0; 1258 if (NumTrue > NumFalse) 1259 SplitVal = ConstantInt::getTrue(BB->getContext()); 1260 else if (NumTrue != 0 || NumFalse != 0) 1261 SplitVal = ConstantInt::getFalse(BB->getContext()); 1262 1263 // Collect all of the blocks that this can be folded into so that we can 1264 // factor this once and clone it once. 1265 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1266 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1267 if (XorOpValues[i].first != SplitVal && 1268 !isa<UndefValue>(XorOpValues[i].first)) 1269 continue; 1270 1271 BlocksToFoldInto.push_back(XorOpValues[i].second); 1272 } 1273 1274 // If we inferred a value for all of the predecessors, then duplication won't 1275 // help us. However, we can just replace the LHS or RHS with the constant. 1276 if (BlocksToFoldInto.size() == 1277 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1278 if (SplitVal == 0) { 1279 // If all preds provide undef, just nuke the xor, because it is undef too. 1280 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1281 BO->eraseFromParent(); 1282 } else if (SplitVal->isZero()) { 1283 // If all preds provide 0, replace the xor with the other input. 1284 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1285 BO->eraseFromParent(); 1286 } else { 1287 // If all preds provide 1, set the computed value to 1. 1288 BO->setOperand(!isLHS, SplitVal); 1289 } 1290 1291 return true; 1292 } 1293 1294 // Try to duplicate BB into PredBB. 1295 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1296} 1297 1298 1299/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1300/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1301/// NewPred using the entries from OldPred (suitably mapped). 1302static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1303 BasicBlock *OldPred, 1304 BasicBlock *NewPred, 1305 DenseMap<Instruction*, Value*> &ValueMap) { 1306 for (BasicBlock::iterator PNI = PHIBB->begin(); 1307 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1308 // Ok, we have a PHI node. Figure out what the incoming value was for the 1309 // DestBlock. 1310 Value *IV = PN->getIncomingValueForBlock(OldPred); 1311 1312 // Remap the value if necessary. 1313 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1314 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1315 if (I != ValueMap.end()) 1316 IV = I->second; 1317 } 1318 1319 PN->addIncoming(IV, NewPred); 1320 } 1321} 1322 1323/// ThreadEdge - We have decided that it is safe and profitable to factor the 1324/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1325/// across BB. Transform the IR to reflect this change. 1326bool JumpThreading::ThreadEdge(BasicBlock *BB, 1327 const SmallVectorImpl<BasicBlock*> &PredBBs, 1328 BasicBlock *SuccBB) { 1329 // If threading to the same block as we come from, we would infinite loop. 1330 if (SuccBB == BB) { 1331 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1332 << "' - would thread to self!\n"); 1333 return false; 1334 } 1335 1336 // If threading this would thread across a loop header, don't thread the edge. 1337 // See the comments above FindLoopHeaders for justifications and caveats. 1338 if (LoopHeaders.count(BB)) { 1339 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1340 << "' to dest BB '" << SuccBB->getName() 1341 << "' - it might create an irreducible loop!\n"); 1342 return false; 1343 } 1344 1345 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold); 1346 if (JumpThreadCost > Threshold) { 1347 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1348 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1349 return false; 1350 } 1351 1352 // And finally, do it! Start by factoring the predecessors is needed. 1353 BasicBlock *PredBB; 1354 if (PredBBs.size() == 1) 1355 PredBB = PredBBs[0]; 1356 else { 1357 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1358 << " common predecessors.\n"); 1359 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); 1360 } 1361 1362 // And finally, do it! 1363 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1364 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1365 << ", across block:\n " 1366 << *BB << "\n"); 1367 1368 LVI->threadEdge(PredBB, BB, SuccBB); 1369 1370 // We are going to have to map operands from the original BB block to the new 1371 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1372 // account for entry from PredBB. 1373 DenseMap<Instruction*, Value*> ValueMapping; 1374 1375 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1376 BB->getName()+".thread", 1377 BB->getParent(), BB); 1378 NewBB->moveAfter(PredBB); 1379 1380 BasicBlock::iterator BI = BB->begin(); 1381 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1382 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1383 1384 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1385 // mapping and using it to remap operands in the cloned instructions. 1386 for (; !isa<TerminatorInst>(BI); ++BI) { 1387 Instruction *New = BI->clone(); 1388 New->setName(BI->getName()); 1389 NewBB->getInstList().push_back(New); 1390 ValueMapping[BI] = New; 1391 1392 // Remap operands to patch up intra-block references. 1393 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1394 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1395 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1396 if (I != ValueMapping.end()) 1397 New->setOperand(i, I->second); 1398 } 1399 } 1400 1401 // We didn't copy the terminator from BB over to NewBB, because there is now 1402 // an unconditional jump to SuccBB. Insert the unconditional jump. 1403 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB); 1404 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc()); 1405 1406 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1407 // PHI nodes for NewBB now. 1408 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1409 1410 // If there were values defined in BB that are used outside the block, then we 1411 // now have to update all uses of the value to use either the original value, 1412 // the cloned value, or some PHI derived value. This can require arbitrary 1413 // PHI insertion, of which we are prepared to do, clean these up now. 1414 SSAUpdater SSAUpdate; 1415 SmallVector<Use*, 16> UsesToRename; 1416 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1417 // Scan all uses of this instruction to see if it is used outside of its 1418 // block, and if so, record them in UsesToRename. 1419 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1420 ++UI) { 1421 Instruction *User = cast<Instruction>(*UI); 1422 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1423 if (UserPN->getIncomingBlock(UI) == BB) 1424 continue; 1425 } else if (User->getParent() == BB) 1426 continue; 1427 1428 UsesToRename.push_back(&UI.getUse()); 1429 } 1430 1431 // If there are no uses outside the block, we're done with this instruction. 1432 if (UsesToRename.empty()) 1433 continue; 1434 1435 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1436 1437 // We found a use of I outside of BB. Rename all uses of I that are outside 1438 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1439 // with the two values we know. 1440 SSAUpdate.Initialize(I->getType(), I->getName()); 1441 SSAUpdate.AddAvailableValue(BB, I); 1442 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1443 1444 while (!UsesToRename.empty()) 1445 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1446 DEBUG(dbgs() << "\n"); 1447 } 1448 1449 1450 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1451 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1452 // us to simplify any PHI nodes in BB. 1453 TerminatorInst *PredTerm = PredBB->getTerminator(); 1454 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1455 if (PredTerm->getSuccessor(i) == BB) { 1456 BB->removePredecessor(PredBB, true); 1457 PredTerm->setSuccessor(i, NewBB); 1458 } 1459 1460 // At this point, the IR is fully up to date and consistent. Do a quick scan 1461 // over the new instructions and zap any that are constants or dead. This 1462 // frequently happens because of phi translation. 1463 SimplifyInstructionsInBlock(NewBB, TD, TLI); 1464 1465 // Threaded an edge! 1466 ++NumThreads; 1467 return true; 1468} 1469 1470/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1471/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1472/// If we can duplicate the contents of BB up into PredBB do so now, this 1473/// improves the odds that the branch will be on an analyzable instruction like 1474/// a compare. 1475bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1476 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1477 assert(!PredBBs.empty() && "Can't handle an empty set"); 1478 1479 // If BB is a loop header, then duplicating this block outside the loop would 1480 // cause us to transform this into an irreducible loop, don't do this. 1481 // See the comments above FindLoopHeaders for justifications and caveats. 1482 if (LoopHeaders.count(BB)) { 1483 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1484 << "' into predecessor block '" << PredBBs[0]->getName() 1485 << "' - it might create an irreducible loop!\n"); 1486 return false; 1487 } 1488 1489 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold); 1490 if (DuplicationCost > Threshold) { 1491 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1492 << "' - Cost is too high: " << DuplicationCost << "\n"); 1493 return false; 1494 } 1495 1496 // And finally, do it! Start by factoring the predecessors is needed. 1497 BasicBlock *PredBB; 1498 if (PredBBs.size() == 1) 1499 PredBB = PredBBs[0]; 1500 else { 1501 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1502 << " common predecessors.\n"); 1503 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); 1504 } 1505 1506 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1507 // of PredBB. 1508 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1509 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1510 << DuplicationCost << " block is:" << *BB << "\n"); 1511 1512 // Unless PredBB ends with an unconditional branch, split the edge so that we 1513 // can just clone the bits from BB into the end of the new PredBB. 1514 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1515 1516 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1517 PredBB = SplitEdge(PredBB, BB, this); 1518 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1519 } 1520 1521 // We are going to have to map operands from the original BB block into the 1522 // PredBB block. Evaluate PHI nodes in BB. 1523 DenseMap<Instruction*, Value*> ValueMapping; 1524 1525 BasicBlock::iterator BI = BB->begin(); 1526 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1527 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1528 1529 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1530 // mapping and using it to remap operands in the cloned instructions. 1531 for (; BI != BB->end(); ++BI) { 1532 Instruction *New = BI->clone(); 1533 1534 // Remap operands to patch up intra-block references. 1535 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1536 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1537 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1538 if (I != ValueMapping.end()) 1539 New->setOperand(i, I->second); 1540 } 1541 1542 // If this instruction can be simplified after the operands are updated, 1543 // just use the simplified value instead. This frequently happens due to 1544 // phi translation. 1545 if (Value *IV = SimplifyInstruction(New, TD)) { 1546 delete New; 1547 ValueMapping[BI] = IV; 1548 } else { 1549 // Otherwise, insert the new instruction into the block. 1550 New->setName(BI->getName()); 1551 PredBB->getInstList().insert(OldPredBranch, New); 1552 ValueMapping[BI] = New; 1553 } 1554 } 1555 1556 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1557 // add entries to the PHI nodes for branch from PredBB now. 1558 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1559 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1560 ValueMapping); 1561 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1562 ValueMapping); 1563 1564 // If there were values defined in BB that are used outside the block, then we 1565 // now have to update all uses of the value to use either the original value, 1566 // the cloned value, or some PHI derived value. This can require arbitrary 1567 // PHI insertion, of which we are prepared to do, clean these up now. 1568 SSAUpdater SSAUpdate; 1569 SmallVector<Use*, 16> UsesToRename; 1570 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1571 // Scan all uses of this instruction to see if it is used outside of its 1572 // block, and if so, record them in UsesToRename. 1573 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1574 ++UI) { 1575 Instruction *User = cast<Instruction>(*UI); 1576 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1577 if (UserPN->getIncomingBlock(UI) == BB) 1578 continue; 1579 } else if (User->getParent() == BB) 1580 continue; 1581 1582 UsesToRename.push_back(&UI.getUse()); 1583 } 1584 1585 // If there are no uses outside the block, we're done with this instruction. 1586 if (UsesToRename.empty()) 1587 continue; 1588 1589 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1590 1591 // We found a use of I outside of BB. Rename all uses of I that are outside 1592 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1593 // with the two values we know. 1594 SSAUpdate.Initialize(I->getType(), I->getName()); 1595 SSAUpdate.AddAvailableValue(BB, I); 1596 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1597 1598 while (!UsesToRename.empty()) 1599 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1600 DEBUG(dbgs() << "\n"); 1601 } 1602 1603 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1604 // that we nuked. 1605 BB->removePredecessor(PredBB, true); 1606 1607 // Remove the unconditional branch at the end of the PredBB block. 1608 OldPredBranch->eraseFromParent(); 1609 1610 ++NumDupes; 1611 return true; 1612} 1613 1614 1615