JumpThreading.cpp revision 3ecfc861b4365f341c5c969b40e1afccde676e6f
1//===- JumpThreading.cpp - Thread control through conditional blocks ------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the Jump Threading pass. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "jump-threading" 15#include "llvm/Transforms/Scalar.h" 16#include "llvm/IntrinsicInst.h" 17#include "llvm/LLVMContext.h" 18#include "llvm/Pass.h" 19#include "llvm/Analysis/InstructionSimplify.h" 20#include "llvm/Analysis/LazyValueInfo.h" 21#include "llvm/Analysis/Loads.h" 22#include "llvm/Transforms/Utils/BasicBlockUtils.h" 23#include "llvm/Transforms/Utils/Local.h" 24#include "llvm/Transforms/Utils/SSAUpdater.h" 25#include "llvm/Target/TargetData.h" 26#include "llvm/ADT/DenseMap.h" 27#include "llvm/ADT/DenseSet.h" 28#include "llvm/ADT/Statistic.h" 29#include "llvm/ADT/STLExtras.h" 30#include "llvm/ADT/SmallPtrSet.h" 31#include "llvm/ADT/SmallSet.h" 32#include "llvm/Support/CommandLine.h" 33#include "llvm/Support/Debug.h" 34#include "llvm/Support/ValueHandle.h" 35#include "llvm/Support/raw_ostream.h" 36using namespace llvm; 37 38STATISTIC(NumThreads, "Number of jumps threaded"); 39STATISTIC(NumFolds, "Number of terminators folded"); 40STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 41 42static cl::opt<unsigned> 43Threshold("jump-threading-threshold", 44 cl::desc("Max block size to duplicate for jump threading"), 45 cl::init(6), cl::Hidden); 46 47namespace { 48 // These are at global scope so static functions can use them too. 49 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo; 50 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy; 51 52 // This is used to keep track of what kind of constant we're currently hoping 53 // to find. 54 enum ConstantPreference { 55 WantInteger, 56 WantBlockAddress 57 }; 58 59 /// This pass performs 'jump threading', which looks at blocks that have 60 /// multiple predecessors and multiple successors. If one or more of the 61 /// predecessors of the block can be proven to always jump to one of the 62 /// successors, we forward the edge from the predecessor to the successor by 63 /// duplicating the contents of this block. 64 /// 65 /// An example of when this can occur is code like this: 66 /// 67 /// if () { ... 68 /// X = 4; 69 /// } 70 /// if (X < 3) { 71 /// 72 /// In this case, the unconditional branch at the end of the first if can be 73 /// revectored to the false side of the second if. 74 /// 75 class JumpThreading : public FunctionPass { 76 TargetData *TD; 77 LazyValueInfo *LVI; 78#ifdef NDEBUG 79 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 80#else 81 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 82#endif 83 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet; 84 85 // RAII helper for updating the recursion stack. 86 struct RecursionSetRemover { 87 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet; 88 std::pair<Value*, BasicBlock*> ThePair; 89 90 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S, 91 std::pair<Value*, BasicBlock*> P) 92 : TheSet(S), ThePair(P) { } 93 94 ~RecursionSetRemover() { 95 TheSet.erase(ThePair); 96 } 97 }; 98 public: 99 static char ID; // Pass identification 100 JumpThreading() : FunctionPass(ID) { 101 initializeJumpThreadingPass(*PassRegistry::getPassRegistry()); 102 } 103 104 bool runOnFunction(Function &F); 105 106 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 107 AU.addRequired<LazyValueInfo>(); 108 AU.addPreserved<LazyValueInfo>(); 109 } 110 111 void FindLoopHeaders(Function &F); 112 bool ProcessBlock(BasicBlock *BB); 113 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 114 BasicBlock *SuccBB); 115 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 116 const SmallVectorImpl<BasicBlock *> &PredBBs); 117 118 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 119 PredValueInfo &Result, 120 ConstantPreference Preference); 121 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 122 ConstantPreference Preference); 123 124 bool ProcessBranchOnPHI(PHINode *PN); 125 bool ProcessBranchOnXOR(BinaryOperator *BO); 126 127 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 128 }; 129} 130 131char JumpThreading::ID = 0; 132INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", 133 "Jump Threading", false, false) 134INITIALIZE_PASS_DEPENDENCY(LazyValueInfo) 135INITIALIZE_PASS_END(JumpThreading, "jump-threading", 136 "Jump Threading", false, false) 137 138// Public interface to the Jump Threading pass 139FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 140 141/// runOnFunction - Top level algorithm. 142/// 143bool JumpThreading::runOnFunction(Function &F) { 144 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 145 TD = getAnalysisIfAvailable<TargetData>(); 146 LVI = &getAnalysis<LazyValueInfo>(); 147 148 FindLoopHeaders(F); 149 150 bool Changed, EverChanged = false; 151 do { 152 Changed = false; 153 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 154 BasicBlock *BB = I; 155 // Thread all of the branches we can over this block. 156 while (ProcessBlock(BB)) 157 Changed = true; 158 159 ++I; 160 161 // If the block is trivially dead, zap it. This eliminates the successor 162 // edges which simplifies the CFG. 163 if (pred_begin(BB) == pred_end(BB) && 164 BB != &BB->getParent()->getEntryBlock()) { 165 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 166 << "' with terminator: " << *BB->getTerminator() << '\n'); 167 LoopHeaders.erase(BB); 168 LVI->eraseBlock(BB); 169 DeleteDeadBlock(BB); 170 Changed = true; 171 continue; 172 } 173 174 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 175 176 // Can't thread an unconditional jump, but if the block is "almost 177 // empty", we can replace uses of it with uses of the successor and make 178 // this dead. 179 if (BI && BI->isUnconditional() && 180 BB != &BB->getParent()->getEntryBlock() && 181 // If the terminator is the only non-phi instruction, try to nuke it. 182 BB->getFirstNonPHIOrDbg()->isTerminator()) { 183 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 184 // block, we have to make sure it isn't in the LoopHeaders set. We 185 // reinsert afterward if needed. 186 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 187 BasicBlock *Succ = BI->getSuccessor(0); 188 189 // FIXME: It is always conservatively correct to drop the info 190 // for a block even if it doesn't get erased. This isn't totally 191 // awesome, but it allows us to use AssertingVH to prevent nasty 192 // dangling pointer issues within LazyValueInfo. 193 LVI->eraseBlock(BB); 194 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 195 Changed = true; 196 // If we deleted BB and BB was the header of a loop, then the 197 // successor is now the header of the loop. 198 BB = Succ; 199 } 200 201 if (ErasedFromLoopHeaders) 202 LoopHeaders.insert(BB); 203 } 204 } 205 EverChanged |= Changed; 206 } while (Changed); 207 208 LoopHeaders.clear(); 209 return EverChanged; 210} 211 212/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 213/// thread across it. 214static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 215 /// Ignore PHI nodes, these will be flattened when duplication happens. 216 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 217 218 // FIXME: THREADING will delete values that are just used to compute the 219 // branch, so they shouldn't count against the duplication cost. 220 221 222 // Sum up the cost of each instruction until we get to the terminator. Don't 223 // include the terminator because the copy won't include it. 224 unsigned Size = 0; 225 for (; !isa<TerminatorInst>(I); ++I) { 226 // Debugger intrinsics don't incur code size. 227 if (isa<DbgInfoIntrinsic>(I)) continue; 228 229 // If this is a pointer->pointer bitcast, it is free. 230 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 231 continue; 232 233 // All other instructions count for at least one unit. 234 ++Size; 235 236 // Calls are more expensive. If they are non-intrinsic calls, we model them 237 // as having cost of 4. If they are a non-vector intrinsic, we model them 238 // as having cost of 2 total, and if they are a vector intrinsic, we model 239 // them as having cost 1. 240 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 241 if (!isa<IntrinsicInst>(CI)) 242 Size += 3; 243 else if (!CI->getType()->isVectorTy()) 244 Size += 1; 245 } 246 } 247 248 // Threading through a switch statement is particularly profitable. If this 249 // block ends in a switch, decrease its cost to make it more likely to happen. 250 if (isa<SwitchInst>(I)) 251 Size = Size > 6 ? Size-6 : 0; 252 253 // The same holds for indirect branches, but slightly more so. 254 if (isa<IndirectBrInst>(I)) 255 Size = Size > 8 ? Size-8 : 0; 256 257 return Size; 258} 259 260/// FindLoopHeaders - We do not want jump threading to turn proper loop 261/// structures into irreducible loops. Doing this breaks up the loop nesting 262/// hierarchy and pessimizes later transformations. To prevent this from 263/// happening, we first have to find the loop headers. Here we approximate this 264/// by finding targets of backedges in the CFG. 265/// 266/// Note that there definitely are cases when we want to allow threading of 267/// edges across a loop header. For example, threading a jump from outside the 268/// loop (the preheader) to an exit block of the loop is definitely profitable. 269/// It is also almost always profitable to thread backedges from within the loop 270/// to exit blocks, and is often profitable to thread backedges to other blocks 271/// within the loop (forming a nested loop). This simple analysis is not rich 272/// enough to track all of these properties and keep it up-to-date as the CFG 273/// mutates, so we don't allow any of these transformations. 274/// 275void JumpThreading::FindLoopHeaders(Function &F) { 276 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 277 FindFunctionBackedges(F, Edges); 278 279 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 280 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 281} 282 283/// getKnownConstant - Helper method to determine if we can thread over a 284/// terminator with the given value as its condition, and if so what value to 285/// use for that. What kind of value this is depends on whether we want an 286/// integer or a block address, but an undef is always accepted. 287/// Returns null if Val is null or not an appropriate constant. 288static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { 289 if (!Val) 290 return 0; 291 292 // Undef is "known" enough. 293 if (UndefValue *U = dyn_cast<UndefValue>(Val)) 294 return U; 295 296 if (Preference == WantBlockAddress) 297 return dyn_cast<BlockAddress>(Val->stripPointerCasts()); 298 299 return dyn_cast<ConstantInt>(Val); 300} 301 302/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 303/// if we can infer that the value is a known ConstantInt/BlockAddress or undef 304/// in any of our predecessors. If so, return the known list of value and pred 305/// BB in the result vector. 306/// 307/// This returns true if there were any known values. 308/// 309bool JumpThreading:: 310ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result, 311 ConstantPreference Preference) { 312 // This method walks up use-def chains recursively. Because of this, we could 313 // get into an infinite loop going around loops in the use-def chain. To 314 // prevent this, keep track of what (value, block) pairs we've already visited 315 // and terminate the search if we loop back to them 316 if (!RecursionSet.insert(std::make_pair(V, BB)).second) 317 return false; 318 319 // An RAII help to remove this pair from the recursion set once the recursion 320 // stack pops back out again. 321 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB)); 322 323 // If V is a constant, then it is known in all predecessors. 324 if (Constant *KC = getKnownConstant(V, Preference)) { 325 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 326 Result.push_back(std::make_pair(KC, *PI)); 327 328 return true; 329 } 330 331 // If V is a non-instruction value, or an instruction in a different block, 332 // then it can't be derived from a PHI. 333 Instruction *I = dyn_cast<Instruction>(V); 334 if (I == 0 || I->getParent() != BB) { 335 336 // Okay, if this is a live-in value, see if it has a known value at the end 337 // of any of our predecessors. 338 // 339 // FIXME: This should be an edge property, not a block end property. 340 /// TODO: Per PR2563, we could infer value range information about a 341 /// predecessor based on its terminator. 342 // 343 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 344 // "I" is a non-local compare-with-a-constant instruction. This would be 345 // able to handle value inequalities better, for example if the compare is 346 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 347 // Perhaps getConstantOnEdge should be smart enough to do this? 348 349 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 350 BasicBlock *P = *PI; 351 // If the value is known by LazyValueInfo to be a constant in a 352 // predecessor, use that information to try to thread this block. 353 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 354 if (Constant *KC = getKnownConstant(PredCst, Preference)) 355 Result.push_back(std::make_pair(KC, P)); 356 } 357 358 return !Result.empty(); 359 } 360 361 /// If I is a PHI node, then we know the incoming values for any constants. 362 if (PHINode *PN = dyn_cast<PHINode>(I)) { 363 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 364 Value *InVal = PN->getIncomingValue(i); 365 if (Constant *KC = getKnownConstant(InVal, Preference)) { 366 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 367 } else { 368 Constant *CI = LVI->getConstantOnEdge(InVal, 369 PN->getIncomingBlock(i), BB); 370 if (Constant *KC = getKnownConstant(CI, Preference)) 371 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 372 } 373 } 374 375 return !Result.empty(); 376 } 377 378 PredValueInfoTy LHSVals, RHSVals; 379 380 // Handle some boolean conditions. 381 if (I->getType()->getPrimitiveSizeInBits() == 1) { 382 assert(Preference == WantInteger && "One-bit non-integer type?"); 383 // X | true -> true 384 // X & false -> false 385 if (I->getOpcode() == Instruction::Or || 386 I->getOpcode() == Instruction::And) { 387 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 388 WantInteger); 389 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals, 390 WantInteger); 391 392 if (LHSVals.empty() && RHSVals.empty()) 393 return false; 394 395 ConstantInt *InterestingVal; 396 if (I->getOpcode() == Instruction::Or) 397 InterestingVal = ConstantInt::getTrue(I->getContext()); 398 else 399 InterestingVal = ConstantInt::getFalse(I->getContext()); 400 401 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 402 403 // Scan for the sentinel. If we find an undef, force it to the 404 // interesting value: x|undef -> true and x&undef -> false. 405 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 406 if (LHSVals[i].first == InterestingVal || 407 isa<UndefValue>(LHSVals[i].first)) { 408 Result.push_back(LHSVals[i]); 409 Result.back().first = InterestingVal; 410 LHSKnownBBs.insert(LHSVals[i].second); 411 } 412 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 413 if (RHSVals[i].first == InterestingVal || 414 isa<UndefValue>(RHSVals[i].first)) { 415 // If we already inferred a value for this block on the LHS, don't 416 // re-add it. 417 if (!LHSKnownBBs.count(RHSVals[i].second)) { 418 Result.push_back(RHSVals[i]); 419 Result.back().first = InterestingVal; 420 } 421 } 422 423 return !Result.empty(); 424 } 425 426 // Handle the NOT form of XOR. 427 if (I->getOpcode() == Instruction::Xor && 428 isa<ConstantInt>(I->getOperand(1)) && 429 cast<ConstantInt>(I->getOperand(1))->isOne()) { 430 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result, 431 WantInteger); 432 if (Result.empty()) 433 return false; 434 435 // Invert the known values. 436 for (unsigned i = 0, e = Result.size(); i != e; ++i) 437 Result[i].first = ConstantExpr::getNot(Result[i].first); 438 439 return true; 440 } 441 442 // Try to simplify some other binary operator values. 443 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 444 assert(Preference != WantBlockAddress 445 && "A binary operator creating a block address?"); 446 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 447 PredValueInfoTy LHSVals; 448 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals, 449 WantInteger); 450 451 // Try to use constant folding to simplify the binary operator. 452 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 453 Constant *V = LHSVals[i].first; 454 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); 455 456 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 457 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 458 } 459 } 460 461 return !Result.empty(); 462 } 463 464 // Handle compare with phi operand, where the PHI is defined in this block. 465 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 466 assert(Preference == WantInteger && "Compares only produce integers"); 467 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 468 if (PN && PN->getParent() == BB) { 469 // We can do this simplification if any comparisons fold to true or false. 470 // See if any do. 471 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 472 BasicBlock *PredBB = PN->getIncomingBlock(i); 473 Value *LHS = PN->getIncomingValue(i); 474 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 475 476 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 477 if (Res == 0) { 478 if (!isa<Constant>(RHS)) 479 continue; 480 481 LazyValueInfo::Tristate 482 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 483 cast<Constant>(RHS), PredBB, BB); 484 if (ResT == LazyValueInfo::Unknown) 485 continue; 486 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 487 } 488 489 if (Constant *KC = getKnownConstant(Res, WantInteger)) 490 Result.push_back(std::make_pair(KC, PredBB)); 491 } 492 493 return !Result.empty(); 494 } 495 496 497 // If comparing a live-in value against a constant, see if we know the 498 // live-in value on any predecessors. 499 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) { 500 if (!isa<Instruction>(Cmp->getOperand(0)) || 501 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 502 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 503 504 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ 505 BasicBlock *P = *PI; 506 // If the value is known by LazyValueInfo to be a constant in a 507 // predecessor, use that information to try to thread this block. 508 LazyValueInfo::Tristate Res = 509 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 510 RHSCst, P, BB); 511 if (Res == LazyValueInfo::Unknown) 512 continue; 513 514 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 515 Result.push_back(std::make_pair(ResC, P)); 516 } 517 518 return !Result.empty(); 519 } 520 521 // Try to find a constant value for the LHS of a comparison, 522 // and evaluate it statically if we can. 523 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { 524 PredValueInfoTy LHSVals; 525 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 526 WantInteger); 527 528 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 529 Constant *V = LHSVals[i].first; 530 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(), 531 V, CmpConst); 532 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 533 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 534 } 535 536 return !Result.empty(); 537 } 538 } 539 } 540 541 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 542 // Handle select instructions where at least one operand is a known constant 543 // and we can figure out the condition value for any predecessor block. 544 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference); 545 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference); 546 PredValueInfoTy Conds; 547 if ((TrueVal || FalseVal) && 548 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds, 549 WantInteger)) { 550 for (unsigned i = 0, e = Conds.size(); i != e; ++i) { 551 Constant *Cond = Conds[i].first; 552 553 // Figure out what value to use for the condition. 554 bool KnownCond; 555 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) { 556 // A known boolean. 557 KnownCond = CI->isOne(); 558 } else { 559 assert(isa<UndefValue>(Cond) && "Unexpected condition value"); 560 // Either operand will do, so be sure to pick the one that's a known 561 // constant. 562 // FIXME: Do this more cleverly if both values are known constants? 563 KnownCond = (TrueVal != 0); 564 } 565 566 // See if the select has a known constant value for this predecessor. 567 if (Constant *Val = KnownCond ? TrueVal : FalseVal) 568 Result.push_back(std::make_pair(Val, Conds[i].second)); 569 } 570 571 return !Result.empty(); 572 } 573 } 574 575 // If all else fails, see if LVI can figure out a constant value for us. 576 Constant *CI = LVI->getConstant(V, BB); 577 if (Constant *KC = getKnownConstant(CI, Preference)) { 578 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 579 Result.push_back(std::make_pair(KC, *PI)); 580 } 581 582 return !Result.empty(); 583} 584 585 586 587/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 588/// in an undefined jump, decide which block is best to revector to. 589/// 590/// Since we can pick an arbitrary destination, we pick the successor with the 591/// fewest predecessors. This should reduce the in-degree of the others. 592/// 593static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 594 TerminatorInst *BBTerm = BB->getTerminator(); 595 unsigned MinSucc = 0; 596 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 597 // Compute the successor with the minimum number of predecessors. 598 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 599 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 600 TestBB = BBTerm->getSuccessor(i); 601 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 602 if (NumPreds < MinNumPreds) 603 MinSucc = i; 604 } 605 606 return MinSucc; 607} 608 609static bool hasAddressTakenAndUsed(BasicBlock *BB) { 610 if (!BB->hasAddressTaken()) return false; 611 612 // If the block has its address taken, it may be a tree of dead constants 613 // hanging off of it. These shouldn't keep the block alive. 614 BlockAddress *BA = BlockAddress::get(BB); 615 BA->removeDeadConstantUsers(); 616 return !BA->use_empty(); 617} 618 619/// ProcessBlock - If there are any predecessors whose control can be threaded 620/// through to a successor, transform them now. 621bool JumpThreading::ProcessBlock(BasicBlock *BB) { 622 // If the block is trivially dead, just return and let the caller nuke it. 623 // This simplifies other transformations. 624 if (pred_begin(BB) == pred_end(BB) && 625 BB != &BB->getParent()->getEntryBlock()) 626 return false; 627 628 // If this block has a single predecessor, and if that pred has a single 629 // successor, merge the blocks. This encourages recursive jump threading 630 // because now the condition in this block can be threaded through 631 // predecessors of our predecessor block. 632 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 633 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 634 SinglePred != BB && !hasAddressTakenAndUsed(BB)) { 635 // If SinglePred was a loop header, BB becomes one. 636 if (LoopHeaders.erase(SinglePred)) 637 LoopHeaders.insert(BB); 638 639 // Remember if SinglePred was the entry block of the function. If so, we 640 // will need to move BB back to the entry position. 641 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 642 LVI->eraseBlock(SinglePred); 643 MergeBasicBlockIntoOnlyPred(BB); 644 645 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 646 BB->moveBefore(&BB->getParent()->getEntryBlock()); 647 return true; 648 } 649 } 650 651 // What kind of constant we're looking for. 652 ConstantPreference Preference = WantInteger; 653 654 // Look to see if the terminator is a conditional branch, switch or indirect 655 // branch, if not we can't thread it. 656 Value *Condition; 657 Instruction *Terminator = BB->getTerminator(); 658 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { 659 // Can't thread an unconditional jump. 660 if (BI->isUnconditional()) return false; 661 Condition = BI->getCondition(); 662 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { 663 Condition = SI->getCondition(); 664 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { 665 Condition = IB->getAddress()->stripPointerCasts(); 666 Preference = WantBlockAddress; 667 } else { 668 return false; // Must be an invoke. 669 } 670 671 // If the terminator is branching on an undef, we can pick any of the 672 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 673 if (isa<UndefValue>(Condition)) { 674 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 675 676 // Fold the branch/switch. 677 TerminatorInst *BBTerm = BB->getTerminator(); 678 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 679 if (i == BestSucc) continue; 680 BBTerm->getSuccessor(i)->removePredecessor(BB, true); 681 } 682 683 DEBUG(dbgs() << " In block '" << BB->getName() 684 << "' folding undef terminator: " << *BBTerm << '\n'); 685 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 686 BBTerm->eraseFromParent(); 687 return true; 688 } 689 690 // If the terminator of this block is branching on a constant, simplify the 691 // terminator to an unconditional branch. This can occur due to threading in 692 // other blocks. 693 if (getKnownConstant(Condition, Preference)) { 694 DEBUG(dbgs() << " In block '" << BB->getName() 695 << "' folding terminator: " << *BB->getTerminator() << '\n'); 696 ++NumFolds; 697 ConstantFoldTerminator(BB); 698 return true; 699 } 700 701 Instruction *CondInst = dyn_cast<Instruction>(Condition); 702 703 // All the rest of our checks depend on the condition being an instruction. 704 if (CondInst == 0) { 705 // FIXME: Unify this with code below. 706 if (ProcessThreadableEdges(Condition, BB, Preference)) 707 return true; 708 return false; 709 } 710 711 712 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 713 // For a comparison where the LHS is outside this block, it's possible 714 // that we've branched on it before. Used LVI to see if we can simplify 715 // the branch based on that. 716 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 717 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 718 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 719 if (CondBr && CondConst && CondBr->isConditional() && PI != PE && 720 (!isa<Instruction>(CondCmp->getOperand(0)) || 721 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { 722 // For predecessor edge, determine if the comparison is true or false 723 // on that edge. If they're all true or all false, we can simplify the 724 // branch. 725 // FIXME: We could handle mixed true/false by duplicating code. 726 LazyValueInfo::Tristate Baseline = 727 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0), 728 CondConst, *PI, BB); 729 if (Baseline != LazyValueInfo::Unknown) { 730 // Check that all remaining incoming values match the first one. 731 while (++PI != PE) { 732 LazyValueInfo::Tristate Ret = 733 LVI->getPredicateOnEdge(CondCmp->getPredicate(), 734 CondCmp->getOperand(0), CondConst, *PI, BB); 735 if (Ret != Baseline) break; 736 } 737 738 // If we terminated early, then one of the values didn't match. 739 if (PI == PE) { 740 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0; 741 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1; 742 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); 743 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); 744 CondBr->eraseFromParent(); 745 return true; 746 } 747 } 748 } 749 } 750 751 // Check for some cases that are worth simplifying. Right now we want to look 752 // for loads that are used by a switch or by the condition for the branch. If 753 // we see one, check to see if it's partially redundant. If so, insert a PHI 754 // which can then be used to thread the values. 755 // 756 Value *SimplifyValue = CondInst; 757 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 758 if (isa<Constant>(CondCmp->getOperand(1))) 759 SimplifyValue = CondCmp->getOperand(0); 760 761 // TODO: There are other places where load PRE would be profitable, such as 762 // more complex comparisons. 763 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 764 if (SimplifyPartiallyRedundantLoad(LI)) 765 return true; 766 767 768 // Handle a variety of cases where we are branching on something derived from 769 // a PHI node in the current block. If we can prove that any predecessors 770 // compute a predictable value based on a PHI node, thread those predecessors. 771 // 772 if (ProcessThreadableEdges(CondInst, BB, Preference)) 773 return true; 774 775 // If this is an otherwise-unfoldable branch on a phi node in the current 776 // block, see if we can simplify. 777 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 778 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 779 return ProcessBranchOnPHI(PN); 780 781 782 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 783 if (CondInst->getOpcode() == Instruction::Xor && 784 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 785 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 786 787 788 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 789 // "(X == 4)", thread through this block. 790 791 return false; 792} 793 794 795/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 796/// load instruction, eliminate it by replacing it with a PHI node. This is an 797/// important optimization that encourages jump threading, and needs to be run 798/// interlaced with other jump threading tasks. 799bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 800 // Don't hack volatile loads. 801 if (LI->isVolatile()) return false; 802 803 // If the load is defined in a block with exactly one predecessor, it can't be 804 // partially redundant. 805 BasicBlock *LoadBB = LI->getParent(); 806 if (LoadBB->getSinglePredecessor()) 807 return false; 808 809 Value *LoadedPtr = LI->getOperand(0); 810 811 // If the loaded operand is defined in the LoadBB, it can't be available. 812 // TODO: Could do simple PHI translation, that would be fun :) 813 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 814 if (PtrOp->getParent() == LoadBB) 815 return false; 816 817 // Scan a few instructions up from the load, to see if it is obviously live at 818 // the entry to its block. 819 BasicBlock::iterator BBIt = LI; 820 821 if (Value *AvailableVal = 822 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 823 // If the value if the load is locally available within the block, just use 824 // it. This frequently occurs for reg2mem'd allocas. 825 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 826 827 // If the returned value is the load itself, replace with an undef. This can 828 // only happen in dead loops. 829 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 830 LI->replaceAllUsesWith(AvailableVal); 831 LI->eraseFromParent(); 832 return true; 833 } 834 835 // Otherwise, if we scanned the whole block and got to the top of the block, 836 // we know the block is locally transparent to the load. If not, something 837 // might clobber its value. 838 if (BBIt != LoadBB->begin()) 839 return false; 840 841 842 SmallPtrSet<BasicBlock*, 8> PredsScanned; 843 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 844 AvailablePredsTy AvailablePreds; 845 BasicBlock *OneUnavailablePred = 0; 846 847 // If we got here, the loaded value is transparent through to the start of the 848 // block. Check to see if it is available in any of the predecessor blocks. 849 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 850 PI != PE; ++PI) { 851 BasicBlock *PredBB = *PI; 852 853 // If we already scanned this predecessor, skip it. 854 if (!PredsScanned.insert(PredBB)) 855 continue; 856 857 // Scan the predecessor to see if the value is available in the pred. 858 BBIt = PredBB->end(); 859 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 860 if (!PredAvailable) { 861 OneUnavailablePred = PredBB; 862 continue; 863 } 864 865 // If so, this load is partially redundant. Remember this info so that we 866 // can create a PHI node. 867 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 868 } 869 870 // If the loaded value isn't available in any predecessor, it isn't partially 871 // redundant. 872 if (AvailablePreds.empty()) return false; 873 874 // Okay, the loaded value is available in at least one (and maybe all!) 875 // predecessors. If the value is unavailable in more than one unique 876 // predecessor, we want to insert a merge block for those common predecessors. 877 // This ensures that we only have to insert one reload, thus not increasing 878 // code size. 879 BasicBlock *UnavailablePred = 0; 880 881 // If there is exactly one predecessor where the value is unavailable, the 882 // already computed 'OneUnavailablePred' block is it. If it ends in an 883 // unconditional branch, we know that it isn't a critical edge. 884 if (PredsScanned.size() == AvailablePreds.size()+1 && 885 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 886 UnavailablePred = OneUnavailablePred; 887 } else if (PredsScanned.size() != AvailablePreds.size()) { 888 // Otherwise, we had multiple unavailable predecessors or we had a critical 889 // edge from the one. 890 SmallVector<BasicBlock*, 8> PredsToSplit; 891 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 892 893 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 894 AvailablePredSet.insert(AvailablePreds[i].first); 895 896 // Add all the unavailable predecessors to the PredsToSplit list. 897 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 898 PI != PE; ++PI) { 899 BasicBlock *P = *PI; 900 // If the predecessor is an indirect goto, we can't split the edge. 901 if (isa<IndirectBrInst>(P->getTerminator())) 902 return false; 903 904 if (!AvailablePredSet.count(P)) 905 PredsToSplit.push_back(P); 906 } 907 908 // Split them out to their own block. 909 UnavailablePred = 910 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 911 "thread-pre-split", this); 912 } 913 914 // If the value isn't available in all predecessors, then there will be 915 // exactly one where it isn't available. Insert a load on that edge and add 916 // it to the AvailablePreds list. 917 if (UnavailablePred) { 918 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 919 "Can't handle critical edge here!"); 920 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 921 LI->getAlignment(), 922 UnavailablePred->getTerminator()); 923 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 924 } 925 926 // Now we know that each predecessor of this block has a value in 927 // AvailablePreds, sort them for efficient access as we're walking the preds. 928 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 929 930 // Create a PHI node at the start of the block for the PRE'd load value. 931 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB); 932 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "", 933 LoadBB->begin()); 934 PN->takeName(LI); 935 936 // Insert new entries into the PHI for each predecessor. A single block may 937 // have multiple entries here. 938 for (pred_iterator PI = PB; PI != PE; ++PI) { 939 BasicBlock *P = *PI; 940 AvailablePredsTy::iterator I = 941 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 942 std::make_pair(P, (Value*)0)); 943 944 assert(I != AvailablePreds.end() && I->first == P && 945 "Didn't find entry for predecessor!"); 946 947 PN->addIncoming(I->second, I->first); 948 } 949 950 //cerr << "PRE: " << *LI << *PN << "\n"; 951 952 LI->replaceAllUsesWith(PN); 953 LI->eraseFromParent(); 954 955 return true; 956} 957 958/// FindMostPopularDest - The specified list contains multiple possible 959/// threadable destinations. Pick the one that occurs the most frequently in 960/// the list. 961static BasicBlock * 962FindMostPopularDest(BasicBlock *BB, 963 const SmallVectorImpl<std::pair<BasicBlock*, 964 BasicBlock*> > &PredToDestList) { 965 assert(!PredToDestList.empty()); 966 967 // Determine popularity. If there are multiple possible destinations, we 968 // explicitly choose to ignore 'undef' destinations. We prefer to thread 969 // blocks with known and real destinations to threading undef. We'll handle 970 // them later if interesting. 971 DenseMap<BasicBlock*, unsigned> DestPopularity; 972 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 973 if (PredToDestList[i].second) 974 DestPopularity[PredToDestList[i].second]++; 975 976 // Find the most popular dest. 977 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 978 BasicBlock *MostPopularDest = DPI->first; 979 unsigned Popularity = DPI->second; 980 SmallVector<BasicBlock*, 4> SamePopularity; 981 982 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 983 // If the popularity of this entry isn't higher than the popularity we've 984 // seen so far, ignore it. 985 if (DPI->second < Popularity) 986 ; // ignore. 987 else if (DPI->second == Popularity) { 988 // If it is the same as what we've seen so far, keep track of it. 989 SamePopularity.push_back(DPI->first); 990 } else { 991 // If it is more popular, remember it. 992 SamePopularity.clear(); 993 MostPopularDest = DPI->first; 994 Popularity = DPI->second; 995 } 996 } 997 998 // Okay, now we know the most popular destination. If there is more than one 999 // destination, we need to determine one. This is arbitrary, but we need 1000 // to make a deterministic decision. Pick the first one that appears in the 1001 // successor list. 1002 if (!SamePopularity.empty()) { 1003 SamePopularity.push_back(MostPopularDest); 1004 TerminatorInst *TI = BB->getTerminator(); 1005 for (unsigned i = 0; ; ++i) { 1006 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1007 1008 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1009 TI->getSuccessor(i)) == SamePopularity.end()) 1010 continue; 1011 1012 MostPopularDest = TI->getSuccessor(i); 1013 break; 1014 } 1015 } 1016 1017 // Okay, we have finally picked the most popular destination. 1018 return MostPopularDest; 1019} 1020 1021bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 1022 ConstantPreference Preference) { 1023 // If threading this would thread across a loop header, don't even try to 1024 // thread the edge. 1025 if (LoopHeaders.count(BB)) 1026 return false; 1027 1028 PredValueInfoTy PredValues; 1029 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference)) 1030 return false; 1031 1032 assert(!PredValues.empty() && 1033 "ComputeValueKnownInPredecessors returned true with no values"); 1034 1035 DEBUG(dbgs() << "IN BB: " << *BB; 1036 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1037 dbgs() << " BB '" << BB->getName() << "': FOUND condition = " 1038 << *PredValues[i].first 1039 << " for pred '" << PredValues[i].second->getName() << "'.\n"; 1040 }); 1041 1042 // Decide what we want to thread through. Convert our list of known values to 1043 // a list of known destinations for each pred. This also discards duplicate 1044 // predecessors and keeps track of the undefined inputs (which are represented 1045 // as a null dest in the PredToDestList). 1046 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1047 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1048 1049 BasicBlock *OnlyDest = 0; 1050 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1051 1052 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1053 BasicBlock *Pred = PredValues[i].second; 1054 if (!SeenPreds.insert(Pred)) 1055 continue; // Duplicate predecessor entry. 1056 1057 // If the predecessor ends with an indirect goto, we can't change its 1058 // destination. 1059 if (isa<IndirectBrInst>(Pred->getTerminator())) 1060 continue; 1061 1062 Constant *Val = PredValues[i].first; 1063 1064 BasicBlock *DestBB; 1065 if (isa<UndefValue>(Val)) 1066 DestBB = 0; 1067 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1068 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); 1069 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 1070 DestBB = SI->getSuccessor(SI->findCaseValue(cast<ConstantInt>(Val))); 1071 else { 1072 assert(isa<IndirectBrInst>(BB->getTerminator()) 1073 && "Unexpected terminator"); 1074 DestBB = cast<BlockAddress>(Val)->getBasicBlock(); 1075 } 1076 1077 // If we have exactly one destination, remember it for efficiency below. 1078 if (PredToDestList.empty()) 1079 OnlyDest = DestBB; 1080 else if (OnlyDest != DestBB) 1081 OnlyDest = MultipleDestSentinel; 1082 1083 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1084 } 1085 1086 // If all edges were unthreadable, we fail. 1087 if (PredToDestList.empty()) 1088 return false; 1089 1090 // Determine which is the most common successor. If we have many inputs and 1091 // this block is a switch, we want to start by threading the batch that goes 1092 // to the most popular destination first. If we only know about one 1093 // threadable destination (the common case) we can avoid this. 1094 BasicBlock *MostPopularDest = OnlyDest; 1095 1096 if (MostPopularDest == MultipleDestSentinel) 1097 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1098 1099 // Now that we know what the most popular destination is, factor all 1100 // predecessors that will jump to it into a single predecessor. 1101 SmallVector<BasicBlock*, 16> PredsToFactor; 1102 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1103 if (PredToDestList[i].second == MostPopularDest) { 1104 BasicBlock *Pred = PredToDestList[i].first; 1105 1106 // This predecessor may be a switch or something else that has multiple 1107 // edges to the block. Factor each of these edges by listing them 1108 // according to # occurrences in PredsToFactor. 1109 TerminatorInst *PredTI = Pred->getTerminator(); 1110 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1111 if (PredTI->getSuccessor(i) == BB) 1112 PredsToFactor.push_back(Pred); 1113 } 1114 1115 // If the threadable edges are branching on an undefined value, we get to pick 1116 // the destination that these predecessors should get to. 1117 if (MostPopularDest == 0) 1118 MostPopularDest = BB->getTerminator()-> 1119 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1120 1121 // Ok, try to thread it! 1122 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1123} 1124 1125/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1126/// a PHI node in the current block. See if there are any simplifications we 1127/// can do based on inputs to the phi node. 1128/// 1129bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1130 BasicBlock *BB = PN->getParent(); 1131 1132 // TODO: We could make use of this to do it once for blocks with common PHI 1133 // values. 1134 SmallVector<BasicBlock*, 1> PredBBs; 1135 PredBBs.resize(1); 1136 1137 // If any of the predecessor blocks end in an unconditional branch, we can 1138 // *duplicate* the conditional branch into that block in order to further 1139 // encourage jump threading and to eliminate cases where we have branch on a 1140 // phi of an icmp (branch on icmp is much better). 1141 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1142 BasicBlock *PredBB = PN->getIncomingBlock(i); 1143 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1144 if (PredBr->isUnconditional()) { 1145 PredBBs[0] = PredBB; 1146 // Try to duplicate BB into PredBB. 1147 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1148 return true; 1149 } 1150 } 1151 1152 return false; 1153} 1154 1155/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1156/// a xor instruction in the current block. See if there are any 1157/// simplifications we can do based on inputs to the xor. 1158/// 1159bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1160 BasicBlock *BB = BO->getParent(); 1161 1162 // If either the LHS or RHS of the xor is a constant, don't do this 1163 // optimization. 1164 if (isa<ConstantInt>(BO->getOperand(0)) || 1165 isa<ConstantInt>(BO->getOperand(1))) 1166 return false; 1167 1168 // If the first instruction in BB isn't a phi, we won't be able to infer 1169 // anything special about any particular predecessor. 1170 if (!isa<PHINode>(BB->front())) 1171 return false; 1172 1173 // If we have a xor as the branch input to this block, and we know that the 1174 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1175 // the condition into the predecessor and fix that value to true, saving some 1176 // logical ops on that path and encouraging other paths to simplify. 1177 // 1178 // This copies something like this: 1179 // 1180 // BB: 1181 // %X = phi i1 [1], [%X'] 1182 // %Y = icmp eq i32 %A, %B 1183 // %Z = xor i1 %X, %Y 1184 // br i1 %Z, ... 1185 // 1186 // Into: 1187 // BB': 1188 // %Y = icmp ne i32 %A, %B 1189 // br i1 %Z, ... 1190 1191 PredValueInfoTy XorOpValues; 1192 bool isLHS = true; 1193 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, 1194 WantInteger)) { 1195 assert(XorOpValues.empty()); 1196 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, 1197 WantInteger)) 1198 return false; 1199 isLHS = false; 1200 } 1201 1202 assert(!XorOpValues.empty() && 1203 "ComputeValueKnownInPredecessors returned true with no values"); 1204 1205 // Scan the information to see which is most popular: true or false. The 1206 // predecessors can be of the set true, false, or undef. 1207 unsigned NumTrue = 0, NumFalse = 0; 1208 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1209 if (isa<UndefValue>(XorOpValues[i].first)) 1210 // Ignore undefs for the count. 1211 continue; 1212 if (cast<ConstantInt>(XorOpValues[i].first)->isZero()) 1213 ++NumFalse; 1214 else 1215 ++NumTrue; 1216 } 1217 1218 // Determine which value to split on, true, false, or undef if neither. 1219 ConstantInt *SplitVal = 0; 1220 if (NumTrue > NumFalse) 1221 SplitVal = ConstantInt::getTrue(BB->getContext()); 1222 else if (NumTrue != 0 || NumFalse != 0) 1223 SplitVal = ConstantInt::getFalse(BB->getContext()); 1224 1225 // Collect all of the blocks that this can be folded into so that we can 1226 // factor this once and clone it once. 1227 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1228 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1229 if (XorOpValues[i].first != SplitVal && 1230 !isa<UndefValue>(XorOpValues[i].first)) 1231 continue; 1232 1233 BlocksToFoldInto.push_back(XorOpValues[i].second); 1234 } 1235 1236 // If we inferred a value for all of the predecessors, then duplication won't 1237 // help us. However, we can just replace the LHS or RHS with the constant. 1238 if (BlocksToFoldInto.size() == 1239 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1240 if (SplitVal == 0) { 1241 // If all preds provide undef, just nuke the xor, because it is undef too. 1242 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1243 BO->eraseFromParent(); 1244 } else if (SplitVal->isZero()) { 1245 // If all preds provide 0, replace the xor with the other input. 1246 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1247 BO->eraseFromParent(); 1248 } else { 1249 // If all preds provide 1, set the computed value to 1. 1250 BO->setOperand(!isLHS, SplitVal); 1251 } 1252 1253 return true; 1254 } 1255 1256 // Try to duplicate BB into PredBB. 1257 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1258} 1259 1260 1261/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1262/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1263/// NewPred using the entries from OldPred (suitably mapped). 1264static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1265 BasicBlock *OldPred, 1266 BasicBlock *NewPred, 1267 DenseMap<Instruction*, Value*> &ValueMap) { 1268 for (BasicBlock::iterator PNI = PHIBB->begin(); 1269 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1270 // Ok, we have a PHI node. Figure out what the incoming value was for the 1271 // DestBlock. 1272 Value *IV = PN->getIncomingValueForBlock(OldPred); 1273 1274 // Remap the value if necessary. 1275 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1276 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1277 if (I != ValueMap.end()) 1278 IV = I->second; 1279 } 1280 1281 PN->addIncoming(IV, NewPred); 1282 } 1283} 1284 1285/// ThreadEdge - We have decided that it is safe and profitable to factor the 1286/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1287/// across BB. Transform the IR to reflect this change. 1288bool JumpThreading::ThreadEdge(BasicBlock *BB, 1289 const SmallVectorImpl<BasicBlock*> &PredBBs, 1290 BasicBlock *SuccBB) { 1291 // If threading to the same block as we come from, we would infinite loop. 1292 if (SuccBB == BB) { 1293 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1294 << "' - would thread to self!\n"); 1295 return false; 1296 } 1297 1298 // If threading this would thread across a loop header, don't thread the edge. 1299 // See the comments above FindLoopHeaders for justifications and caveats. 1300 if (LoopHeaders.count(BB)) { 1301 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1302 << "' to dest BB '" << SuccBB->getName() 1303 << "' - it might create an irreducible loop!\n"); 1304 return false; 1305 } 1306 1307 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1308 if (JumpThreadCost > Threshold) { 1309 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1310 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1311 return false; 1312 } 1313 1314 // And finally, do it! Start by factoring the predecessors is needed. 1315 BasicBlock *PredBB; 1316 if (PredBBs.size() == 1) 1317 PredBB = PredBBs[0]; 1318 else { 1319 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1320 << " common predecessors.\n"); 1321 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1322 ".thr_comm", this); 1323 } 1324 1325 // And finally, do it! 1326 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1327 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1328 << ", across block:\n " 1329 << *BB << "\n"); 1330 1331 LVI->threadEdge(PredBB, BB, SuccBB); 1332 1333 // We are going to have to map operands from the original BB block to the new 1334 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1335 // account for entry from PredBB. 1336 DenseMap<Instruction*, Value*> ValueMapping; 1337 1338 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1339 BB->getName()+".thread", 1340 BB->getParent(), BB); 1341 NewBB->moveAfter(PredBB); 1342 1343 BasicBlock::iterator BI = BB->begin(); 1344 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1345 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1346 1347 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1348 // mapping and using it to remap operands in the cloned instructions. 1349 for (; !isa<TerminatorInst>(BI); ++BI) { 1350 Instruction *New = BI->clone(); 1351 New->setName(BI->getName()); 1352 NewBB->getInstList().push_back(New); 1353 ValueMapping[BI] = New; 1354 1355 // Remap operands to patch up intra-block references. 1356 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1357 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1358 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1359 if (I != ValueMapping.end()) 1360 New->setOperand(i, I->second); 1361 } 1362 } 1363 1364 // We didn't copy the terminator from BB over to NewBB, because there is now 1365 // an unconditional jump to SuccBB. Insert the unconditional jump. 1366 BranchInst::Create(SuccBB, NewBB); 1367 1368 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1369 // PHI nodes for NewBB now. 1370 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1371 1372 // If there were values defined in BB that are used outside the block, then we 1373 // now have to update all uses of the value to use either the original value, 1374 // the cloned value, or some PHI derived value. This can require arbitrary 1375 // PHI insertion, of which we are prepared to do, clean these up now. 1376 SSAUpdater SSAUpdate; 1377 SmallVector<Use*, 16> UsesToRename; 1378 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1379 // Scan all uses of this instruction to see if it is used outside of its 1380 // block, and if so, record them in UsesToRename. 1381 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1382 ++UI) { 1383 Instruction *User = cast<Instruction>(*UI); 1384 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1385 if (UserPN->getIncomingBlock(UI) == BB) 1386 continue; 1387 } else if (User->getParent() == BB) 1388 continue; 1389 1390 UsesToRename.push_back(&UI.getUse()); 1391 } 1392 1393 // If there are no uses outside the block, we're done with this instruction. 1394 if (UsesToRename.empty()) 1395 continue; 1396 1397 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1398 1399 // We found a use of I outside of BB. Rename all uses of I that are outside 1400 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1401 // with the two values we know. 1402 SSAUpdate.Initialize(I->getType(), I->getName()); 1403 SSAUpdate.AddAvailableValue(BB, I); 1404 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1405 1406 while (!UsesToRename.empty()) 1407 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1408 DEBUG(dbgs() << "\n"); 1409 } 1410 1411 1412 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1413 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1414 // us to simplify any PHI nodes in BB. 1415 TerminatorInst *PredTerm = PredBB->getTerminator(); 1416 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1417 if (PredTerm->getSuccessor(i) == BB) { 1418 BB->removePredecessor(PredBB, true); 1419 PredTerm->setSuccessor(i, NewBB); 1420 } 1421 1422 // At this point, the IR is fully up to date and consistent. Do a quick scan 1423 // over the new instructions and zap any that are constants or dead. This 1424 // frequently happens because of phi translation. 1425 SimplifyInstructionsInBlock(NewBB, TD); 1426 1427 // Threaded an edge! 1428 ++NumThreads; 1429 return true; 1430} 1431 1432/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1433/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1434/// If we can duplicate the contents of BB up into PredBB do so now, this 1435/// improves the odds that the branch will be on an analyzable instruction like 1436/// a compare. 1437bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1438 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1439 assert(!PredBBs.empty() && "Can't handle an empty set"); 1440 1441 // If BB is a loop header, then duplicating this block outside the loop would 1442 // cause us to transform this into an irreducible loop, don't do this. 1443 // See the comments above FindLoopHeaders for justifications and caveats. 1444 if (LoopHeaders.count(BB)) { 1445 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1446 << "' into predecessor block '" << PredBBs[0]->getName() 1447 << "' - it might create an irreducible loop!\n"); 1448 return false; 1449 } 1450 1451 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1452 if (DuplicationCost > Threshold) { 1453 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1454 << "' - Cost is too high: " << DuplicationCost << "\n"); 1455 return false; 1456 } 1457 1458 // And finally, do it! Start by factoring the predecessors is needed. 1459 BasicBlock *PredBB; 1460 if (PredBBs.size() == 1) 1461 PredBB = PredBBs[0]; 1462 else { 1463 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1464 << " common predecessors.\n"); 1465 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1466 ".thr_comm", this); 1467 } 1468 1469 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1470 // of PredBB. 1471 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1472 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1473 << DuplicationCost << " block is:" << *BB << "\n"); 1474 1475 // Unless PredBB ends with an unconditional branch, split the edge so that we 1476 // can just clone the bits from BB into the end of the new PredBB. 1477 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1478 1479 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1480 PredBB = SplitEdge(PredBB, BB, this); 1481 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1482 } 1483 1484 // We are going to have to map operands from the original BB block into the 1485 // PredBB block. Evaluate PHI nodes in BB. 1486 DenseMap<Instruction*, Value*> ValueMapping; 1487 1488 BasicBlock::iterator BI = BB->begin(); 1489 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1490 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1491 1492 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1493 // mapping and using it to remap operands in the cloned instructions. 1494 for (; BI != BB->end(); ++BI) { 1495 Instruction *New = BI->clone(); 1496 1497 // Remap operands to patch up intra-block references. 1498 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1499 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1500 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1501 if (I != ValueMapping.end()) 1502 New->setOperand(i, I->second); 1503 } 1504 1505 // If this instruction can be simplified after the operands are updated, 1506 // just use the simplified value instead. This frequently happens due to 1507 // phi translation. 1508 if (Value *IV = SimplifyInstruction(New, TD)) { 1509 delete New; 1510 ValueMapping[BI] = IV; 1511 } else { 1512 // Otherwise, insert the new instruction into the block. 1513 New->setName(BI->getName()); 1514 PredBB->getInstList().insert(OldPredBranch, New); 1515 ValueMapping[BI] = New; 1516 } 1517 } 1518 1519 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1520 // add entries to the PHI nodes for branch from PredBB now. 1521 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1522 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1523 ValueMapping); 1524 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1525 ValueMapping); 1526 1527 // If there were values defined in BB that are used outside the block, then we 1528 // now have to update all uses of the value to use either the original value, 1529 // the cloned value, or some PHI derived value. This can require arbitrary 1530 // PHI insertion, of which we are prepared to do, clean these up now. 1531 SSAUpdater SSAUpdate; 1532 SmallVector<Use*, 16> UsesToRename; 1533 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1534 // Scan all uses of this instruction to see if it is used outside of its 1535 // block, and if so, record them in UsesToRename. 1536 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1537 ++UI) { 1538 Instruction *User = cast<Instruction>(*UI); 1539 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1540 if (UserPN->getIncomingBlock(UI) == BB) 1541 continue; 1542 } else if (User->getParent() == BB) 1543 continue; 1544 1545 UsesToRename.push_back(&UI.getUse()); 1546 } 1547 1548 // If there are no uses outside the block, we're done with this instruction. 1549 if (UsesToRename.empty()) 1550 continue; 1551 1552 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1553 1554 // We found a use of I outside of BB. Rename all uses of I that are outside 1555 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1556 // with the two values we know. 1557 SSAUpdate.Initialize(I->getType(), I->getName()); 1558 SSAUpdate.AddAvailableValue(BB, I); 1559 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1560 1561 while (!UsesToRename.empty()) 1562 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1563 DEBUG(dbgs() << "\n"); 1564 } 1565 1566 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1567 // that we nuked. 1568 BB->removePredecessor(PredBB, true); 1569 1570 // Remove the unconditional branch at the end of the PredBB block. 1571 OldPredBranch->eraseFromParent(); 1572 1573 ++NumDupes; 1574 return true; 1575} 1576 1577 1578