JumpThreading.cpp revision f6832bbda08b6975caa264de183bf3ee1d472aca
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/ConstantFolding.h" 20#include "llvm/Analysis/InstructionSimplify.h" 21#include "llvm/Analysis/LazyValueInfo.h" 22#include "llvm/Analysis/Loads.h" 23#include "llvm/Transforms/Utils/BasicBlockUtils.h" 24#include "llvm/Transforms/Utils/Local.h" 25#include "llvm/Transforms/Utils/SSAUpdater.h" 26#include "llvm/Target/TargetData.h" 27#include "llvm/ADT/DenseMap.h" 28#include "llvm/ADT/DenseSet.h" 29#include "llvm/ADT/Statistic.h" 30#include "llvm/ADT/STLExtras.h" 31#include "llvm/ADT/SmallPtrSet.h" 32#include "llvm/ADT/SmallSet.h" 33#include "llvm/Support/CommandLine.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/Support/ValueHandle.h" 36#include "llvm/Support/raw_ostream.h" 37using namespace llvm; 38 39STATISTIC(NumThreads, "Number of jumps threaded"); 40STATISTIC(NumFolds, "Number of terminators folded"); 41STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 42 43static cl::opt<unsigned> 44Threshold("jump-threading-threshold", 45 cl::desc("Max block size to duplicate for jump threading"), 46 cl::init(6), cl::Hidden); 47 48namespace { 49 // These are at global scope so static functions can use them too. 50 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo; 51 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy; 52 53 // This is used to keep track of what kind of constant we're currently hoping 54 // to find. 55 enum ConstantPreference { 56 WantInteger, 57 WantBlockAddress 58 }; 59 60 /// This pass performs 'jump threading', which looks at blocks that have 61 /// multiple predecessors and multiple successors. If one or more of the 62 /// predecessors of the block can be proven to always jump to one of the 63 /// successors, we forward the edge from the predecessor to the successor by 64 /// duplicating the contents of this block. 65 /// 66 /// An example of when this can occur is code like this: 67 /// 68 /// if () { ... 69 /// X = 4; 70 /// } 71 /// if (X < 3) { 72 /// 73 /// In this case, the unconditional branch at the end of the first if can be 74 /// revectored to the false side of the second if. 75 /// 76 class JumpThreading : public FunctionPass { 77 TargetData *TD; 78 LazyValueInfo *LVI; 79#ifdef NDEBUG 80 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 81#else 82 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 83#endif 84 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet; 85 86 // RAII helper for updating the recursion stack. 87 struct RecursionSetRemover { 88 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet; 89 std::pair<Value*, BasicBlock*> ThePair; 90 91 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S, 92 std::pair<Value*, BasicBlock*> P) 93 : TheSet(S), ThePair(P) { } 94 95 ~RecursionSetRemover() { 96 TheSet.erase(ThePair); 97 } 98 }; 99 public: 100 static char ID; // Pass identification 101 JumpThreading() : FunctionPass(ID) { 102 initializeJumpThreadingPass(*PassRegistry::getPassRegistry()); 103 } 104 105 bool runOnFunction(Function &F); 106 107 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 108 AU.addRequired<LazyValueInfo>(); 109 AU.addPreserved<LazyValueInfo>(); 110 } 111 112 void FindLoopHeaders(Function &F); 113 bool ProcessBlock(BasicBlock *BB); 114 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 115 BasicBlock *SuccBB); 116 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 117 const SmallVectorImpl<BasicBlock *> &PredBBs); 118 119 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 120 PredValueInfo &Result, 121 ConstantPreference Preference); 122 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 123 ConstantPreference Preference); 124 125 bool ProcessBranchOnPHI(PHINode *PN); 126 bool ProcessBranchOnXOR(BinaryOperator *BO); 127 128 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 129 }; 130} 131 132char JumpThreading::ID = 0; 133INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", 134 "Jump Threading", false, false) 135INITIALIZE_PASS_DEPENDENCY(LazyValueInfo) 136INITIALIZE_PASS_END(JumpThreading, "jump-threading", 137 "Jump Threading", false, false) 138 139// Public interface to the Jump Threading pass 140FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 141 142/// runOnFunction - Top level algorithm. 143/// 144bool JumpThreading::runOnFunction(Function &F) { 145 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 146 TD = getAnalysisIfAvailable<TargetData>(); 147 LVI = &getAnalysis<LazyValueInfo>(); 148 149 FindLoopHeaders(F); 150 151 bool Changed, EverChanged = false; 152 do { 153 Changed = false; 154 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 155 BasicBlock *BB = I; 156 // Thread all of the branches we can over this block. 157 while (ProcessBlock(BB)) 158 Changed = true; 159 160 ++I; 161 162 // If the block is trivially dead, zap it. This eliminates the successor 163 // edges which simplifies the CFG. 164 if (pred_begin(BB) == pred_end(BB) && 165 BB != &BB->getParent()->getEntryBlock()) { 166 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 167 << "' with terminator: " << *BB->getTerminator() << '\n'); 168 LoopHeaders.erase(BB); 169 LVI->eraseBlock(BB); 170 DeleteDeadBlock(BB); 171 Changed = true; 172 continue; 173 } 174 175 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 176 177 // Can't thread an unconditional jump, but if the block is "almost 178 // empty", we can replace uses of it with uses of the successor and make 179 // this dead. 180 if (BI && BI->isUnconditional() && 181 BB != &BB->getParent()->getEntryBlock() && 182 // If the terminator is the only non-phi instruction, try to nuke it. 183 BB->getFirstNonPHIOrDbg()->isTerminator()) { 184 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 185 // block, we have to make sure it isn't in the LoopHeaders set. We 186 // reinsert afterward if needed. 187 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 188 BasicBlock *Succ = BI->getSuccessor(0); 189 190 // FIXME: It is always conservatively correct to drop the info 191 // for a block even if it doesn't get erased. This isn't totally 192 // awesome, but it allows us to use AssertingVH to prevent nasty 193 // dangling pointer issues within LazyValueInfo. 194 LVI->eraseBlock(BB); 195 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 196 Changed = true; 197 // If we deleted BB and BB was the header of a loop, then the 198 // successor is now the header of the loop. 199 BB = Succ; 200 } 201 202 if (ErasedFromLoopHeaders) 203 LoopHeaders.insert(BB); 204 } 205 } 206 EverChanged |= Changed; 207 } while (Changed); 208 209 LoopHeaders.clear(); 210 return EverChanged; 211} 212 213/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 214/// thread across it. 215static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 216 /// Ignore PHI nodes, these will be flattened when duplication happens. 217 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 218 219 // FIXME: THREADING will delete values that are just used to compute the 220 // branch, so they shouldn't count against the duplication cost. 221 222 223 // Sum up the cost of each instruction until we get to the terminator. Don't 224 // include the terminator because the copy won't include it. 225 unsigned Size = 0; 226 for (; !isa<TerminatorInst>(I); ++I) { 227 // Debugger intrinsics don't incur code size. 228 if (isa<DbgInfoIntrinsic>(I)) continue; 229 230 // If this is a pointer->pointer bitcast, it is free. 231 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 232 continue; 233 234 // All other instructions count for at least one unit. 235 ++Size; 236 237 // Calls are more expensive. If they are non-intrinsic calls, we model them 238 // as having cost of 4. If they are a non-vector intrinsic, we model them 239 // as having cost of 2 total, and if they are a vector intrinsic, we model 240 // them as having cost 1. 241 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 242 if (!isa<IntrinsicInst>(CI)) 243 Size += 3; 244 else if (!CI->getType()->isVectorTy()) 245 Size += 1; 246 } 247 } 248 249 // Threading through a switch statement is particularly profitable. If this 250 // block ends in a switch, decrease its cost to make it more likely to happen. 251 if (isa<SwitchInst>(I)) 252 Size = Size > 6 ? Size-6 : 0; 253 254 // The same holds for indirect branches, but slightly more so. 255 if (isa<IndirectBrInst>(I)) 256 Size = Size > 8 ? Size-8 : 0; 257 258 return Size; 259} 260 261/// FindLoopHeaders - We do not want jump threading to turn proper loop 262/// structures into irreducible loops. Doing this breaks up the loop nesting 263/// hierarchy and pessimizes later transformations. To prevent this from 264/// happening, we first have to find the loop headers. Here we approximate this 265/// by finding targets of backedges in the CFG. 266/// 267/// Note that there definitely are cases when we want to allow threading of 268/// edges across a loop header. For example, threading a jump from outside the 269/// loop (the preheader) to an exit block of the loop is definitely profitable. 270/// It is also almost always profitable to thread backedges from within the loop 271/// to exit blocks, and is often profitable to thread backedges to other blocks 272/// within the loop (forming a nested loop). This simple analysis is not rich 273/// enough to track all of these properties and keep it up-to-date as the CFG 274/// mutates, so we don't allow any of these transformations. 275/// 276void JumpThreading::FindLoopHeaders(Function &F) { 277 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 278 FindFunctionBackedges(F, Edges); 279 280 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 281 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 282} 283 284/// getKnownConstant - Helper method to determine if we can thread over a 285/// terminator with the given value as its condition, and if so what value to 286/// use for that. What kind of value this is depends on whether we want an 287/// integer or a block address, but an undef is always accepted. 288/// Returns null if Val is null or not an appropriate constant. 289static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { 290 if (!Val) 291 return 0; 292 293 // Undef is "known" enough. 294 if (UndefValue *U = dyn_cast<UndefValue>(Val)) 295 return U; 296 297 if (Preference == WantBlockAddress) 298 return dyn_cast<BlockAddress>(Val->stripPointerCasts()); 299 300 return dyn_cast<ConstantInt>(Val); 301} 302 303/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 304/// if we can infer that the value is a known ConstantInt/BlockAddress or undef 305/// in any of our predecessors. If so, return the known list of value and pred 306/// BB in the result vector. 307/// 308/// This returns true if there were any known values. 309/// 310bool JumpThreading:: 311ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result, 312 ConstantPreference Preference) { 313 // This method walks up use-def chains recursively. Because of this, we could 314 // get into an infinite loop going around loops in the use-def chain. To 315 // prevent this, keep track of what (value, block) pairs we've already visited 316 // and terminate the search if we loop back to them 317 if (!RecursionSet.insert(std::make_pair(V, BB)).second) 318 return false; 319 320 // An RAII help to remove this pair from the recursion set once the recursion 321 // stack pops back out again. 322 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB)); 323 324 // If V is a constant, then it is known in all predecessors. 325 if (Constant *KC = getKnownConstant(V, Preference)) { 326 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 327 Result.push_back(std::make_pair(KC, *PI)); 328 329 return true; 330 } 331 332 // If V is a non-instruction value, or an instruction in a different block, 333 // then it can't be derived from a PHI. 334 Instruction *I = dyn_cast<Instruction>(V); 335 if (I == 0 || I->getParent() != BB) { 336 337 // Okay, if this is a live-in value, see if it has a known value at the end 338 // of any of our predecessors. 339 // 340 // FIXME: This should be an edge property, not a block end property. 341 /// TODO: Per PR2563, we could infer value range information about a 342 /// predecessor based on its terminator. 343 // 344 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 345 // "I" is a non-local compare-with-a-constant instruction. This would be 346 // able to handle value inequalities better, for example if the compare is 347 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 348 // Perhaps getConstantOnEdge should be smart enough to do this? 349 350 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 351 BasicBlock *P = *PI; 352 // If the value is known by LazyValueInfo to be a constant in a 353 // predecessor, use that information to try to thread this block. 354 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 355 if (Constant *KC = getKnownConstant(PredCst, Preference)) 356 Result.push_back(std::make_pair(KC, P)); 357 } 358 359 return !Result.empty(); 360 } 361 362 /// If I is a PHI node, then we know the incoming values for any constants. 363 if (PHINode *PN = dyn_cast<PHINode>(I)) { 364 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 365 Value *InVal = PN->getIncomingValue(i); 366 if (Constant *KC = getKnownConstant(InVal, Preference)) { 367 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 368 } else { 369 Constant *CI = LVI->getConstantOnEdge(InVal, 370 PN->getIncomingBlock(i), BB); 371 if (Constant *KC = getKnownConstant(CI, Preference)) 372 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 373 } 374 } 375 376 return !Result.empty(); 377 } 378 379 PredValueInfoTy LHSVals, RHSVals; 380 381 // Handle some boolean conditions. 382 if (I->getType()->getPrimitiveSizeInBits() == 1) { 383 assert(Preference == WantInteger && "One-bit non-integer type?"); 384 // X | true -> true 385 // X & false -> false 386 if (I->getOpcode() == Instruction::Or || 387 I->getOpcode() == Instruction::And) { 388 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 389 WantInteger); 390 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals, 391 WantInteger); 392 393 if (LHSVals.empty() && RHSVals.empty()) 394 return false; 395 396 ConstantInt *InterestingVal; 397 if (I->getOpcode() == Instruction::Or) 398 InterestingVal = ConstantInt::getTrue(I->getContext()); 399 else 400 InterestingVal = ConstantInt::getFalse(I->getContext()); 401 402 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 403 404 // Scan for the sentinel. If we find an undef, force it to the 405 // interesting value: x|undef -> true and x&undef -> false. 406 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 407 if (LHSVals[i].first == InterestingVal || 408 isa<UndefValue>(LHSVals[i].first)) { 409 Result.push_back(LHSVals[i]); 410 Result.back().first = InterestingVal; 411 LHSKnownBBs.insert(LHSVals[i].second); 412 } 413 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 414 if (RHSVals[i].first == InterestingVal || 415 isa<UndefValue>(RHSVals[i].first)) { 416 // If we already inferred a value for this block on the LHS, don't 417 // re-add it. 418 if (!LHSKnownBBs.count(RHSVals[i].second)) { 419 Result.push_back(RHSVals[i]); 420 Result.back().first = InterestingVal; 421 } 422 } 423 424 return !Result.empty(); 425 } 426 427 // Handle the NOT form of XOR. 428 if (I->getOpcode() == Instruction::Xor && 429 isa<ConstantInt>(I->getOperand(1)) && 430 cast<ConstantInt>(I->getOperand(1))->isOne()) { 431 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result, 432 WantInteger); 433 if (Result.empty()) 434 return false; 435 436 // Invert the known values. 437 for (unsigned i = 0, e = Result.size(); i != e; ++i) 438 Result[i].first = ConstantExpr::getNot(Result[i].first); 439 440 return true; 441 } 442 443 // Try to simplify some other binary operator values. 444 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 445 assert(Preference != WantBlockAddress 446 && "A binary operator creating a block address?"); 447 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 448 PredValueInfoTy LHSVals; 449 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals, 450 WantInteger); 451 452 // Try to use constant folding to simplify the binary operator. 453 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 454 Constant *V = LHSVals[i].first; 455 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); 456 457 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 458 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 459 } 460 } 461 462 return !Result.empty(); 463 } 464 465 // Handle compare with phi operand, where the PHI is defined in this block. 466 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 467 assert(Preference == WantInteger && "Compares only produce integers"); 468 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 469 if (PN && PN->getParent() == BB) { 470 // We can do this simplification if any comparisons fold to true or false. 471 // See if any do. 472 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 473 BasicBlock *PredBB = PN->getIncomingBlock(i); 474 Value *LHS = PN->getIncomingValue(i); 475 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 476 477 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 478 if (Res == 0) { 479 if (!isa<Constant>(RHS)) 480 continue; 481 482 LazyValueInfo::Tristate 483 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 484 cast<Constant>(RHS), PredBB, BB); 485 if (ResT == LazyValueInfo::Unknown) 486 continue; 487 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 488 } 489 490 if (Constant *KC = getKnownConstant(Res, WantInteger)) 491 Result.push_back(std::make_pair(KC, PredBB)); 492 } 493 494 return !Result.empty(); 495 } 496 497 498 // If comparing a live-in value against a constant, see if we know the 499 // live-in value on any predecessors. 500 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) { 501 if (!isa<Instruction>(Cmp->getOperand(0)) || 502 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 503 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 504 505 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ 506 BasicBlock *P = *PI; 507 // If the value is known by LazyValueInfo to be a constant in a 508 // predecessor, use that information to try to thread this block. 509 LazyValueInfo::Tristate Res = 510 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 511 RHSCst, P, BB); 512 if (Res == LazyValueInfo::Unknown) 513 continue; 514 515 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 516 Result.push_back(std::make_pair(ResC, P)); 517 } 518 519 return !Result.empty(); 520 } 521 522 // Try to find a constant value for the LHS of a comparison, 523 // and evaluate it statically if we can. 524 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { 525 PredValueInfoTy LHSVals; 526 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 527 WantInteger); 528 529 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 530 Constant *V = LHSVals[i].first; 531 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(), 532 V, CmpConst); 533 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 534 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 535 } 536 537 return !Result.empty(); 538 } 539 } 540 } 541 542 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 543 // Handle select instructions where at least one operand is a known constant 544 // and we can figure out the condition value for any predecessor block. 545 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference); 546 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference); 547 PredValueInfoTy Conds; 548 if ((TrueVal || FalseVal) && 549 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds, 550 WantInteger)) { 551 for (unsigned i = 0, e = Conds.size(); i != e; ++i) { 552 Constant *Cond = Conds[i].first; 553 554 // Figure out what value to use for the condition. 555 bool KnownCond; 556 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) { 557 // A known boolean. 558 KnownCond = CI->isOne(); 559 } else { 560 assert(isa<UndefValue>(Cond) && "Unexpected condition value"); 561 // Either operand will do, so be sure to pick the one that's a known 562 // constant. 563 // FIXME: Do this more cleverly if both values are known constants? 564 KnownCond = (TrueVal != 0); 565 } 566 567 // See if the select has a known constant value for this predecessor. 568 if (Constant *Val = KnownCond ? TrueVal : FalseVal) 569 Result.push_back(std::make_pair(Val, Conds[i].second)); 570 } 571 572 return !Result.empty(); 573 } 574 } 575 576 // If all else fails, see if LVI can figure out a constant value for us. 577 Constant *CI = LVI->getConstant(V, BB); 578 if (Constant *KC = getKnownConstant(CI, Preference)) { 579 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 580 Result.push_back(std::make_pair(KC, *PI)); 581 } 582 583 return !Result.empty(); 584} 585 586 587 588/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 589/// in an undefined jump, decide which block is best to revector to. 590/// 591/// Since we can pick an arbitrary destination, we pick the successor with the 592/// fewest predecessors. This should reduce the in-degree of the others. 593/// 594static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 595 TerminatorInst *BBTerm = BB->getTerminator(); 596 unsigned MinSucc = 0; 597 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 598 // Compute the successor with the minimum number of predecessors. 599 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 600 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 601 TestBB = BBTerm->getSuccessor(i); 602 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 603 if (NumPreds < MinNumPreds) 604 MinSucc = i; 605 } 606 607 return MinSucc; 608} 609 610static bool hasAddressTakenAndUsed(BasicBlock *BB) { 611 if (!BB->hasAddressTaken()) return false; 612 613 // If the block has its address taken, it may be a tree of dead constants 614 // hanging off of it. These shouldn't keep the block alive. 615 BlockAddress *BA = BlockAddress::get(BB); 616 BA->removeDeadConstantUsers(); 617 return !BA->use_empty(); 618} 619 620/// ProcessBlock - If there are any predecessors whose control can be threaded 621/// through to a successor, transform them now. 622bool JumpThreading::ProcessBlock(BasicBlock *BB) { 623 // If the block is trivially dead, just return and let the caller nuke it. 624 // This simplifies other transformations. 625 if (pred_begin(BB) == pred_end(BB) && 626 BB != &BB->getParent()->getEntryBlock()) 627 return false; 628 629 // If this block has a single predecessor, and if that pred has a single 630 // successor, merge the blocks. This encourages recursive jump threading 631 // because now the condition in this block can be threaded through 632 // predecessors of our predecessor block. 633 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 634 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 635 SinglePred != BB && !hasAddressTakenAndUsed(BB)) { 636 // If SinglePred was a loop header, BB becomes one. 637 if (LoopHeaders.erase(SinglePred)) 638 LoopHeaders.insert(BB); 639 640 // Remember if SinglePred was the entry block of the function. If so, we 641 // will need to move BB back to the entry position. 642 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 643 LVI->eraseBlock(SinglePred); 644 MergeBasicBlockIntoOnlyPred(BB); 645 646 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 647 BB->moveBefore(&BB->getParent()->getEntryBlock()); 648 return true; 649 } 650 } 651 652 // What kind of constant we're looking for. 653 ConstantPreference Preference = WantInteger; 654 655 // Look to see if the terminator is a conditional branch, switch or indirect 656 // branch, if not we can't thread it. 657 Value *Condition; 658 Instruction *Terminator = BB->getTerminator(); 659 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { 660 // Can't thread an unconditional jump. 661 if (BI->isUnconditional()) return false; 662 Condition = BI->getCondition(); 663 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { 664 Condition = SI->getCondition(); 665 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { 666 Condition = IB->getAddress()->stripPointerCasts(); 667 Preference = WantBlockAddress; 668 } else { 669 return false; // Must be an invoke. 670 } 671 672 // Run constant folding to see if we can reduce the condition to a simple 673 // constant. 674 if (Instruction *I = dyn_cast<Instruction>(Condition)) { 675 Value *SimpleVal = ConstantFoldInstruction(I, TD); 676 if (SimpleVal) { 677 I->replaceAllUsesWith(SimpleVal); 678 I->eraseFromParent(); 679 Condition = SimpleVal; 680 } 681 } 682 683 // If the terminator is branching on an undef, we can pick any of the 684 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 685 if (isa<UndefValue>(Condition)) { 686 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 687 688 // Fold the branch/switch. 689 TerminatorInst *BBTerm = BB->getTerminator(); 690 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 691 if (i == BestSucc) continue; 692 BBTerm->getSuccessor(i)->removePredecessor(BB, true); 693 } 694 695 DEBUG(dbgs() << " In block '" << BB->getName() 696 << "' folding undef terminator: " << *BBTerm << '\n'); 697 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 698 BBTerm->eraseFromParent(); 699 return true; 700 } 701 702 // If the terminator of this block is branching on a constant, simplify the 703 // terminator to an unconditional branch. This can occur due to threading in 704 // other blocks. 705 if (getKnownConstant(Condition, Preference)) { 706 DEBUG(dbgs() << " In block '" << BB->getName() 707 << "' folding terminator: " << *BB->getTerminator() << '\n'); 708 ++NumFolds; 709 ConstantFoldTerminator(BB); 710 return true; 711 } 712 713 Instruction *CondInst = dyn_cast<Instruction>(Condition); 714 715 // All the rest of our checks depend on the condition being an instruction. 716 if (CondInst == 0) { 717 // FIXME: Unify this with code below. 718 if (ProcessThreadableEdges(Condition, BB, Preference)) 719 return true; 720 return false; 721 } 722 723 724 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 725 // For a comparison where the LHS is outside this block, it's possible 726 // that we've branched on it before. Used LVI to see if we can simplify 727 // the branch based on that. 728 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 729 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 730 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 731 if (CondBr && CondConst && CondBr->isConditional() && PI != PE && 732 (!isa<Instruction>(CondCmp->getOperand(0)) || 733 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { 734 // For predecessor edge, determine if the comparison is true or false 735 // on that edge. If they're all true or all false, we can simplify the 736 // branch. 737 // FIXME: We could handle mixed true/false by duplicating code. 738 LazyValueInfo::Tristate Baseline = 739 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0), 740 CondConst, *PI, BB); 741 if (Baseline != LazyValueInfo::Unknown) { 742 // Check that all remaining incoming values match the first one. 743 while (++PI != PE) { 744 LazyValueInfo::Tristate Ret = 745 LVI->getPredicateOnEdge(CondCmp->getPredicate(), 746 CondCmp->getOperand(0), CondConst, *PI, BB); 747 if (Ret != Baseline) break; 748 } 749 750 // If we terminated early, then one of the values didn't match. 751 if (PI == PE) { 752 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0; 753 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1; 754 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); 755 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); 756 CondBr->eraseFromParent(); 757 return true; 758 } 759 } 760 } 761 } 762 763 // Check for some cases that are worth simplifying. Right now we want to look 764 // for loads that are used by a switch or by the condition for the branch. If 765 // we see one, check to see if it's partially redundant. If so, insert a PHI 766 // which can then be used to thread the values. 767 // 768 Value *SimplifyValue = CondInst; 769 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 770 if (isa<Constant>(CondCmp->getOperand(1))) 771 SimplifyValue = CondCmp->getOperand(0); 772 773 // TODO: There are other places where load PRE would be profitable, such as 774 // more complex comparisons. 775 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 776 if (SimplifyPartiallyRedundantLoad(LI)) 777 return true; 778 779 780 // Handle a variety of cases where we are branching on something derived from 781 // a PHI node in the current block. If we can prove that any predecessors 782 // compute a predictable value based on a PHI node, thread those predecessors. 783 // 784 if (ProcessThreadableEdges(CondInst, BB, Preference)) 785 return true; 786 787 // If this is an otherwise-unfoldable branch on a phi node in the current 788 // block, see if we can simplify. 789 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 790 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 791 return ProcessBranchOnPHI(PN); 792 793 794 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 795 if (CondInst->getOpcode() == Instruction::Xor && 796 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 797 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 798 799 800 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 801 // "(X == 4)", thread through this block. 802 803 return false; 804} 805 806 807/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 808/// load instruction, eliminate it by replacing it with a PHI node. This is an 809/// important optimization that encourages jump threading, and needs to be run 810/// interlaced with other jump threading tasks. 811bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 812 // Don't hack volatile loads. 813 if (LI->isVolatile()) return false; 814 815 // If the load is defined in a block with exactly one predecessor, it can't be 816 // partially redundant. 817 BasicBlock *LoadBB = LI->getParent(); 818 if (LoadBB->getSinglePredecessor()) 819 return false; 820 821 Value *LoadedPtr = LI->getOperand(0); 822 823 // If the loaded operand is defined in the LoadBB, it can't be available. 824 // TODO: Could do simple PHI translation, that would be fun :) 825 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 826 if (PtrOp->getParent() == LoadBB) 827 return false; 828 829 // Scan a few instructions up from the load, to see if it is obviously live at 830 // the entry to its block. 831 BasicBlock::iterator BBIt = LI; 832 833 if (Value *AvailableVal = 834 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 835 // If the value if the load is locally available within the block, just use 836 // it. This frequently occurs for reg2mem'd allocas. 837 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 838 839 // If the returned value is the load itself, replace with an undef. This can 840 // only happen in dead loops. 841 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 842 LI->replaceAllUsesWith(AvailableVal); 843 LI->eraseFromParent(); 844 return true; 845 } 846 847 // Otherwise, if we scanned the whole block and got to the top of the block, 848 // we know the block is locally transparent to the load. If not, something 849 // might clobber its value. 850 if (BBIt != LoadBB->begin()) 851 return false; 852 853 854 SmallPtrSet<BasicBlock*, 8> PredsScanned; 855 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 856 AvailablePredsTy AvailablePreds; 857 BasicBlock *OneUnavailablePred = 0; 858 859 // If we got here, the loaded value is transparent through to the start of the 860 // block. Check to see if it is available in any of the predecessor blocks. 861 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 862 PI != PE; ++PI) { 863 BasicBlock *PredBB = *PI; 864 865 // If we already scanned this predecessor, skip it. 866 if (!PredsScanned.insert(PredBB)) 867 continue; 868 869 // Scan the predecessor to see if the value is available in the pred. 870 BBIt = PredBB->end(); 871 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 872 if (!PredAvailable) { 873 OneUnavailablePred = PredBB; 874 continue; 875 } 876 877 // If so, this load is partially redundant. Remember this info so that we 878 // can create a PHI node. 879 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 880 } 881 882 // If the loaded value isn't available in any predecessor, it isn't partially 883 // redundant. 884 if (AvailablePreds.empty()) return false; 885 886 // Okay, the loaded value is available in at least one (and maybe all!) 887 // predecessors. If the value is unavailable in more than one unique 888 // predecessor, we want to insert a merge block for those common predecessors. 889 // This ensures that we only have to insert one reload, thus not increasing 890 // code size. 891 BasicBlock *UnavailablePred = 0; 892 893 // If there is exactly one predecessor where the value is unavailable, the 894 // already computed 'OneUnavailablePred' block is it. If it ends in an 895 // unconditional branch, we know that it isn't a critical edge. 896 if (PredsScanned.size() == AvailablePreds.size()+1 && 897 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 898 UnavailablePred = OneUnavailablePred; 899 } else if (PredsScanned.size() != AvailablePreds.size()) { 900 // Otherwise, we had multiple unavailable predecessors or we had a critical 901 // edge from the one. 902 SmallVector<BasicBlock*, 8> PredsToSplit; 903 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 904 905 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 906 AvailablePredSet.insert(AvailablePreds[i].first); 907 908 // Add all the unavailable predecessors to the PredsToSplit list. 909 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 910 PI != PE; ++PI) { 911 BasicBlock *P = *PI; 912 // If the predecessor is an indirect goto, we can't split the edge. 913 if (isa<IndirectBrInst>(P->getTerminator())) 914 return false; 915 916 if (!AvailablePredSet.count(P)) 917 PredsToSplit.push_back(P); 918 } 919 920 // Split them out to their own block. 921 UnavailablePred = 922 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 923 "thread-pre-split", this); 924 } 925 926 // If the value isn't available in all predecessors, then there will be 927 // exactly one where it isn't available. Insert a load on that edge and add 928 // it to the AvailablePreds list. 929 if (UnavailablePred) { 930 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 931 "Can't handle critical edge here!"); 932 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 933 LI->getAlignment(), 934 UnavailablePred->getTerminator()); 935 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 936 } 937 938 // Now we know that each predecessor of this block has a value in 939 // AvailablePreds, sort them for efficient access as we're walking the preds. 940 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 941 942 // Create a PHI node at the start of the block for the PRE'd load value. 943 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB); 944 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "", 945 LoadBB->begin()); 946 PN->takeName(LI); 947 948 // Insert new entries into the PHI for each predecessor. A single block may 949 // have multiple entries here. 950 for (pred_iterator PI = PB; PI != PE; ++PI) { 951 BasicBlock *P = *PI; 952 AvailablePredsTy::iterator I = 953 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 954 std::make_pair(P, (Value*)0)); 955 956 assert(I != AvailablePreds.end() && I->first == P && 957 "Didn't find entry for predecessor!"); 958 959 PN->addIncoming(I->second, I->first); 960 } 961 962 //cerr << "PRE: " << *LI << *PN << "\n"; 963 964 LI->replaceAllUsesWith(PN); 965 LI->eraseFromParent(); 966 967 return true; 968} 969 970/// FindMostPopularDest - The specified list contains multiple possible 971/// threadable destinations. Pick the one that occurs the most frequently in 972/// the list. 973static BasicBlock * 974FindMostPopularDest(BasicBlock *BB, 975 const SmallVectorImpl<std::pair<BasicBlock*, 976 BasicBlock*> > &PredToDestList) { 977 assert(!PredToDestList.empty()); 978 979 // Determine popularity. If there are multiple possible destinations, we 980 // explicitly choose to ignore 'undef' destinations. We prefer to thread 981 // blocks with known and real destinations to threading undef. We'll handle 982 // them later if interesting. 983 DenseMap<BasicBlock*, unsigned> DestPopularity; 984 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 985 if (PredToDestList[i].second) 986 DestPopularity[PredToDestList[i].second]++; 987 988 // Find the most popular dest. 989 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 990 BasicBlock *MostPopularDest = DPI->first; 991 unsigned Popularity = DPI->second; 992 SmallVector<BasicBlock*, 4> SamePopularity; 993 994 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 995 // If the popularity of this entry isn't higher than the popularity we've 996 // seen so far, ignore it. 997 if (DPI->second < Popularity) 998 ; // ignore. 999 else if (DPI->second == Popularity) { 1000 // If it is the same as what we've seen so far, keep track of it. 1001 SamePopularity.push_back(DPI->first); 1002 } else { 1003 // If it is more popular, remember it. 1004 SamePopularity.clear(); 1005 MostPopularDest = DPI->first; 1006 Popularity = DPI->second; 1007 } 1008 } 1009 1010 // Okay, now we know the most popular destination. If there is more than one 1011 // destination, we need to determine one. This is arbitrary, but we need 1012 // to make a deterministic decision. Pick the first one that appears in the 1013 // successor list. 1014 if (!SamePopularity.empty()) { 1015 SamePopularity.push_back(MostPopularDest); 1016 TerminatorInst *TI = BB->getTerminator(); 1017 for (unsigned i = 0; ; ++i) { 1018 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1019 1020 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1021 TI->getSuccessor(i)) == SamePopularity.end()) 1022 continue; 1023 1024 MostPopularDest = TI->getSuccessor(i); 1025 break; 1026 } 1027 } 1028 1029 // Okay, we have finally picked the most popular destination. 1030 return MostPopularDest; 1031} 1032 1033bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 1034 ConstantPreference Preference) { 1035 // If threading this would thread across a loop header, don't even try to 1036 // thread the edge. 1037 if (LoopHeaders.count(BB)) 1038 return false; 1039 1040 PredValueInfoTy PredValues; 1041 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference)) 1042 return false; 1043 1044 assert(!PredValues.empty() && 1045 "ComputeValueKnownInPredecessors returned true with no values"); 1046 1047 DEBUG(dbgs() << "IN BB: " << *BB; 1048 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1049 dbgs() << " BB '" << BB->getName() << "': FOUND condition = " 1050 << *PredValues[i].first 1051 << " for pred '" << PredValues[i].second->getName() << "'.\n"; 1052 }); 1053 1054 // Decide what we want to thread through. Convert our list of known values to 1055 // a list of known destinations for each pred. This also discards duplicate 1056 // predecessors and keeps track of the undefined inputs (which are represented 1057 // as a null dest in the PredToDestList). 1058 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1059 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1060 1061 BasicBlock *OnlyDest = 0; 1062 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1063 1064 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1065 BasicBlock *Pred = PredValues[i].second; 1066 if (!SeenPreds.insert(Pred)) 1067 continue; // Duplicate predecessor entry. 1068 1069 // If the predecessor ends with an indirect goto, we can't change its 1070 // destination. 1071 if (isa<IndirectBrInst>(Pred->getTerminator())) 1072 continue; 1073 1074 Constant *Val = PredValues[i].first; 1075 1076 BasicBlock *DestBB; 1077 if (isa<UndefValue>(Val)) 1078 DestBB = 0; 1079 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1080 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); 1081 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 1082 DestBB = SI->getSuccessor(SI->findCaseValue(cast<ConstantInt>(Val))); 1083 else { 1084 assert(isa<IndirectBrInst>(BB->getTerminator()) 1085 && "Unexpected terminator"); 1086 DestBB = cast<BlockAddress>(Val)->getBasicBlock(); 1087 } 1088 1089 // If we have exactly one destination, remember it for efficiency below. 1090 if (PredToDestList.empty()) 1091 OnlyDest = DestBB; 1092 else if (OnlyDest != DestBB) 1093 OnlyDest = MultipleDestSentinel; 1094 1095 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1096 } 1097 1098 // If all edges were unthreadable, we fail. 1099 if (PredToDestList.empty()) 1100 return false; 1101 1102 // Determine which is the most common successor. If we have many inputs and 1103 // this block is a switch, we want to start by threading the batch that goes 1104 // to the most popular destination first. If we only know about one 1105 // threadable destination (the common case) we can avoid this. 1106 BasicBlock *MostPopularDest = OnlyDest; 1107 1108 if (MostPopularDest == MultipleDestSentinel) 1109 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1110 1111 // Now that we know what the most popular destination is, factor all 1112 // predecessors that will jump to it into a single predecessor. 1113 SmallVector<BasicBlock*, 16> PredsToFactor; 1114 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1115 if (PredToDestList[i].second == MostPopularDest) { 1116 BasicBlock *Pred = PredToDestList[i].first; 1117 1118 // This predecessor may be a switch or something else that has multiple 1119 // edges to the block. Factor each of these edges by listing them 1120 // according to # occurrences in PredsToFactor. 1121 TerminatorInst *PredTI = Pred->getTerminator(); 1122 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1123 if (PredTI->getSuccessor(i) == BB) 1124 PredsToFactor.push_back(Pred); 1125 } 1126 1127 // If the threadable edges are branching on an undefined value, we get to pick 1128 // the destination that these predecessors should get to. 1129 if (MostPopularDest == 0) 1130 MostPopularDest = BB->getTerminator()-> 1131 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1132 1133 // Ok, try to thread it! 1134 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1135} 1136 1137/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1138/// a PHI node in the current block. See if there are any simplifications we 1139/// can do based on inputs to the phi node. 1140/// 1141bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1142 BasicBlock *BB = PN->getParent(); 1143 1144 // TODO: We could make use of this to do it once for blocks with common PHI 1145 // values. 1146 SmallVector<BasicBlock*, 1> PredBBs; 1147 PredBBs.resize(1); 1148 1149 // If any of the predecessor blocks end in an unconditional branch, we can 1150 // *duplicate* the conditional branch into that block in order to further 1151 // encourage jump threading and to eliminate cases where we have branch on a 1152 // phi of an icmp (branch on icmp is much better). 1153 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1154 BasicBlock *PredBB = PN->getIncomingBlock(i); 1155 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1156 if (PredBr->isUnconditional()) { 1157 PredBBs[0] = PredBB; 1158 // Try to duplicate BB into PredBB. 1159 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1160 return true; 1161 } 1162 } 1163 1164 return false; 1165} 1166 1167/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1168/// a xor instruction in the current block. See if there are any 1169/// simplifications we can do based on inputs to the xor. 1170/// 1171bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1172 BasicBlock *BB = BO->getParent(); 1173 1174 // If either the LHS or RHS of the xor is a constant, don't do this 1175 // optimization. 1176 if (isa<ConstantInt>(BO->getOperand(0)) || 1177 isa<ConstantInt>(BO->getOperand(1))) 1178 return false; 1179 1180 // If the first instruction in BB isn't a phi, we won't be able to infer 1181 // anything special about any particular predecessor. 1182 if (!isa<PHINode>(BB->front())) 1183 return false; 1184 1185 // If we have a xor as the branch input to this block, and we know that the 1186 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1187 // the condition into the predecessor and fix that value to true, saving some 1188 // logical ops on that path and encouraging other paths to simplify. 1189 // 1190 // This copies something like this: 1191 // 1192 // BB: 1193 // %X = phi i1 [1], [%X'] 1194 // %Y = icmp eq i32 %A, %B 1195 // %Z = xor i1 %X, %Y 1196 // br i1 %Z, ... 1197 // 1198 // Into: 1199 // BB': 1200 // %Y = icmp ne i32 %A, %B 1201 // br i1 %Z, ... 1202 1203 PredValueInfoTy XorOpValues; 1204 bool isLHS = true; 1205 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, 1206 WantInteger)) { 1207 assert(XorOpValues.empty()); 1208 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, 1209 WantInteger)) 1210 return false; 1211 isLHS = false; 1212 } 1213 1214 assert(!XorOpValues.empty() && 1215 "ComputeValueKnownInPredecessors returned true with no values"); 1216 1217 // Scan the information to see which is most popular: true or false. The 1218 // predecessors can be of the set true, false, or undef. 1219 unsigned NumTrue = 0, NumFalse = 0; 1220 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1221 if (isa<UndefValue>(XorOpValues[i].first)) 1222 // Ignore undefs for the count. 1223 continue; 1224 if (cast<ConstantInt>(XorOpValues[i].first)->isZero()) 1225 ++NumFalse; 1226 else 1227 ++NumTrue; 1228 } 1229 1230 // Determine which value to split on, true, false, or undef if neither. 1231 ConstantInt *SplitVal = 0; 1232 if (NumTrue > NumFalse) 1233 SplitVal = ConstantInt::getTrue(BB->getContext()); 1234 else if (NumTrue != 0 || NumFalse != 0) 1235 SplitVal = ConstantInt::getFalse(BB->getContext()); 1236 1237 // Collect all of the blocks that this can be folded into so that we can 1238 // factor this once and clone it once. 1239 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1240 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1241 if (XorOpValues[i].first != SplitVal && 1242 !isa<UndefValue>(XorOpValues[i].first)) 1243 continue; 1244 1245 BlocksToFoldInto.push_back(XorOpValues[i].second); 1246 } 1247 1248 // If we inferred a value for all of the predecessors, then duplication won't 1249 // help us. However, we can just replace the LHS or RHS with the constant. 1250 if (BlocksToFoldInto.size() == 1251 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1252 if (SplitVal == 0) { 1253 // If all preds provide undef, just nuke the xor, because it is undef too. 1254 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1255 BO->eraseFromParent(); 1256 } else if (SplitVal->isZero()) { 1257 // If all preds provide 0, replace the xor with the other input. 1258 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1259 BO->eraseFromParent(); 1260 } else { 1261 // If all preds provide 1, set the computed value to 1. 1262 BO->setOperand(!isLHS, SplitVal); 1263 } 1264 1265 return true; 1266 } 1267 1268 // Try to duplicate BB into PredBB. 1269 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1270} 1271 1272 1273/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1274/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1275/// NewPred using the entries from OldPred (suitably mapped). 1276static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1277 BasicBlock *OldPred, 1278 BasicBlock *NewPred, 1279 DenseMap<Instruction*, Value*> &ValueMap) { 1280 for (BasicBlock::iterator PNI = PHIBB->begin(); 1281 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1282 // Ok, we have a PHI node. Figure out what the incoming value was for the 1283 // DestBlock. 1284 Value *IV = PN->getIncomingValueForBlock(OldPred); 1285 1286 // Remap the value if necessary. 1287 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1288 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1289 if (I != ValueMap.end()) 1290 IV = I->second; 1291 } 1292 1293 PN->addIncoming(IV, NewPred); 1294 } 1295} 1296 1297/// ThreadEdge - We have decided that it is safe and profitable to factor the 1298/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1299/// across BB. Transform the IR to reflect this change. 1300bool JumpThreading::ThreadEdge(BasicBlock *BB, 1301 const SmallVectorImpl<BasicBlock*> &PredBBs, 1302 BasicBlock *SuccBB) { 1303 // If threading to the same block as we come from, we would infinite loop. 1304 if (SuccBB == BB) { 1305 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1306 << "' - would thread to self!\n"); 1307 return false; 1308 } 1309 1310 // If threading this would thread across a loop header, don't thread the edge. 1311 // See the comments above FindLoopHeaders for justifications and caveats. 1312 if (LoopHeaders.count(BB)) { 1313 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1314 << "' to dest BB '" << SuccBB->getName() 1315 << "' - it might create an irreducible loop!\n"); 1316 return false; 1317 } 1318 1319 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1320 if (JumpThreadCost > Threshold) { 1321 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1322 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1323 return false; 1324 } 1325 1326 // And finally, do it! Start by factoring the predecessors is needed. 1327 BasicBlock *PredBB; 1328 if (PredBBs.size() == 1) 1329 PredBB = PredBBs[0]; 1330 else { 1331 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1332 << " common predecessors.\n"); 1333 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1334 ".thr_comm", this); 1335 } 1336 1337 // And finally, do it! 1338 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1339 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1340 << ", across block:\n " 1341 << *BB << "\n"); 1342 1343 LVI->threadEdge(PredBB, BB, SuccBB); 1344 1345 // We are going to have to map operands from the original BB block to the new 1346 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1347 // account for entry from PredBB. 1348 DenseMap<Instruction*, Value*> ValueMapping; 1349 1350 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1351 BB->getName()+".thread", 1352 BB->getParent(), BB); 1353 NewBB->moveAfter(PredBB); 1354 1355 BasicBlock::iterator BI = BB->begin(); 1356 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1357 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1358 1359 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1360 // mapping and using it to remap operands in the cloned instructions. 1361 for (; !isa<TerminatorInst>(BI); ++BI) { 1362 Instruction *New = BI->clone(); 1363 New->setName(BI->getName()); 1364 NewBB->getInstList().push_back(New); 1365 ValueMapping[BI] = New; 1366 1367 // Remap operands to patch up intra-block references. 1368 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1369 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1370 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1371 if (I != ValueMapping.end()) 1372 New->setOperand(i, I->second); 1373 } 1374 } 1375 1376 // We didn't copy the terminator from BB over to NewBB, because there is now 1377 // an unconditional jump to SuccBB. Insert the unconditional jump. 1378 BranchInst::Create(SuccBB, NewBB); 1379 1380 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1381 // PHI nodes for NewBB now. 1382 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1383 1384 // If there were values defined in BB that are used outside the block, then we 1385 // now have to update all uses of the value to use either the original value, 1386 // the cloned value, or some PHI derived value. This can require arbitrary 1387 // PHI insertion, of which we are prepared to do, clean these up now. 1388 SSAUpdater SSAUpdate; 1389 SmallVector<Use*, 16> UsesToRename; 1390 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1391 // Scan all uses of this instruction to see if it is used outside of its 1392 // block, and if so, record them in UsesToRename. 1393 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1394 ++UI) { 1395 Instruction *User = cast<Instruction>(*UI); 1396 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1397 if (UserPN->getIncomingBlock(UI) == BB) 1398 continue; 1399 } else if (User->getParent() == BB) 1400 continue; 1401 1402 UsesToRename.push_back(&UI.getUse()); 1403 } 1404 1405 // If there are no uses outside the block, we're done with this instruction. 1406 if (UsesToRename.empty()) 1407 continue; 1408 1409 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1410 1411 // We found a use of I outside of BB. Rename all uses of I that are outside 1412 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1413 // with the two values we know. 1414 SSAUpdate.Initialize(I->getType(), I->getName()); 1415 SSAUpdate.AddAvailableValue(BB, I); 1416 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1417 1418 while (!UsesToRename.empty()) 1419 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1420 DEBUG(dbgs() << "\n"); 1421 } 1422 1423 1424 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1425 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1426 // us to simplify any PHI nodes in BB. 1427 TerminatorInst *PredTerm = PredBB->getTerminator(); 1428 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1429 if (PredTerm->getSuccessor(i) == BB) { 1430 BB->removePredecessor(PredBB, true); 1431 PredTerm->setSuccessor(i, NewBB); 1432 } 1433 1434 // At this point, the IR is fully up to date and consistent. Do a quick scan 1435 // over the new instructions and zap any that are constants or dead. This 1436 // frequently happens because of phi translation. 1437 SimplifyInstructionsInBlock(NewBB, TD); 1438 1439 // Threaded an edge! 1440 ++NumThreads; 1441 return true; 1442} 1443 1444/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1445/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1446/// If we can duplicate the contents of BB up into PredBB do so now, this 1447/// improves the odds that the branch will be on an analyzable instruction like 1448/// a compare. 1449bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1450 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1451 assert(!PredBBs.empty() && "Can't handle an empty set"); 1452 1453 // If BB is a loop header, then duplicating this block outside the loop would 1454 // cause us to transform this into an irreducible loop, don't do this. 1455 // See the comments above FindLoopHeaders for justifications and caveats. 1456 if (LoopHeaders.count(BB)) { 1457 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1458 << "' into predecessor block '" << PredBBs[0]->getName() 1459 << "' - it might create an irreducible loop!\n"); 1460 return false; 1461 } 1462 1463 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1464 if (DuplicationCost > Threshold) { 1465 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1466 << "' - Cost is too high: " << DuplicationCost << "\n"); 1467 return false; 1468 } 1469 1470 // And finally, do it! Start by factoring the predecessors is needed. 1471 BasicBlock *PredBB; 1472 if (PredBBs.size() == 1) 1473 PredBB = PredBBs[0]; 1474 else { 1475 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1476 << " common predecessors.\n"); 1477 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1478 ".thr_comm", this); 1479 } 1480 1481 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1482 // of PredBB. 1483 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1484 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1485 << DuplicationCost << " block is:" << *BB << "\n"); 1486 1487 // Unless PredBB ends with an unconditional branch, split the edge so that we 1488 // can just clone the bits from BB into the end of the new PredBB. 1489 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1490 1491 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1492 PredBB = SplitEdge(PredBB, BB, this); 1493 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1494 } 1495 1496 // We are going to have to map operands from the original BB block into the 1497 // PredBB block. Evaluate PHI nodes in BB. 1498 DenseMap<Instruction*, Value*> ValueMapping; 1499 1500 BasicBlock::iterator BI = BB->begin(); 1501 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1502 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1503 1504 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1505 // mapping and using it to remap operands in the cloned instructions. 1506 for (; BI != BB->end(); ++BI) { 1507 Instruction *New = BI->clone(); 1508 1509 // Remap operands to patch up intra-block references. 1510 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1511 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1512 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1513 if (I != ValueMapping.end()) 1514 New->setOperand(i, I->second); 1515 } 1516 1517 // If this instruction can be simplified after the operands are updated, 1518 // just use the simplified value instead. This frequently happens due to 1519 // phi translation. 1520 if (Value *IV = SimplifyInstruction(New, TD)) { 1521 delete New; 1522 ValueMapping[BI] = IV; 1523 } else { 1524 // Otherwise, insert the new instruction into the block. 1525 New->setName(BI->getName()); 1526 PredBB->getInstList().insert(OldPredBranch, New); 1527 ValueMapping[BI] = New; 1528 } 1529 } 1530 1531 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1532 // add entries to the PHI nodes for branch from PredBB now. 1533 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1534 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1535 ValueMapping); 1536 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1537 ValueMapping); 1538 1539 // If there were values defined in BB that are used outside the block, then we 1540 // now have to update all uses of the value to use either the original value, 1541 // the cloned value, or some PHI derived value. This can require arbitrary 1542 // PHI insertion, of which we are prepared to do, clean these up now. 1543 SSAUpdater SSAUpdate; 1544 SmallVector<Use*, 16> UsesToRename; 1545 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1546 // Scan all uses of this instruction to see if it is used outside of its 1547 // block, and if so, record them in UsesToRename. 1548 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1549 ++UI) { 1550 Instruction *User = cast<Instruction>(*UI); 1551 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1552 if (UserPN->getIncomingBlock(UI) == BB) 1553 continue; 1554 } else if (User->getParent() == BB) 1555 continue; 1556 1557 UsesToRename.push_back(&UI.getUse()); 1558 } 1559 1560 // If there are no uses outside the block, we're done with this instruction. 1561 if (UsesToRename.empty()) 1562 continue; 1563 1564 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1565 1566 // We found a use of I outside of BB. Rename all uses of I that are outside 1567 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1568 // with the two values we know. 1569 SSAUpdate.Initialize(I->getType(), I->getName()); 1570 SSAUpdate.AddAvailableValue(BB, I); 1571 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1572 1573 while (!UsesToRename.empty()) 1574 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1575 DEBUG(dbgs() << "\n"); 1576 } 1577 1578 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1579 // that we nuked. 1580 BB->removePredecessor(PredBB, true); 1581 1582 // Remove the unconditional branch at the end of the PredBB block. 1583 OldPredBranch->eraseFromParent(); 1584 1585 ++NumDupes; 1586 return true; 1587} 1588 1589 1590