JumpThreading.cpp revision 01abcf339f8d42921c680fefb2ff988cfeee1198
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} 608 609/// ProcessBlock - If there are any predecessors whose control can be threaded 610/// through to a successor, transform them now. 611bool JumpThreading::ProcessBlock(BasicBlock *BB) { 612 // If the block is trivially dead, just return and let the caller nuke it. 613 // This simplifies other transformations. 614 if (pred_begin(BB) == pred_end(BB) && 615 BB != &BB->getParent()->getEntryBlock()) 616 return false; 617 618 // If this block has a single predecessor, and if that pred has a single 619 // successor, merge the blocks. This encourages recursive jump threading 620 // because now the condition in this block can be threaded through 621 // predecessors of our predecessor block. 622 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 623 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 624 SinglePred != BB) { 625 // If SinglePred was a loop header, BB becomes one. 626 if (LoopHeaders.erase(SinglePred)) 627 LoopHeaders.insert(BB); 628 629 // Remember if SinglePred was the entry block of the function. If so, we 630 // will need to move BB back to the entry position. 631 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 632 LVI->eraseBlock(SinglePred); 633 MergeBasicBlockIntoOnlyPred(BB); 634 635 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 636 BB->moveBefore(&BB->getParent()->getEntryBlock()); 637 return true; 638 } 639 } 640 641 // What kind of constant we're looking for. 642 ConstantPreference Preference = WantInteger; 643 644 // Look to see if the terminator is a conditional branch, switch or indirect 645 // branch, if not we can't thread it. 646 Value *Condition; 647 Instruction *Terminator = BB->getTerminator(); 648 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { 649 // Can't thread an unconditional jump. 650 if (BI->isUnconditional()) return false; 651 Condition = BI->getCondition(); 652 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { 653 Condition = SI->getCondition(); 654 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { 655 Condition = IB->getAddress()->stripPointerCasts(); 656 Preference = WantBlockAddress; 657 } else { 658 return false; // Must be an invoke. 659 } 660 661 // If the terminator is branching on an undef, we can pick any of the 662 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 663 if (isa<UndefValue>(Condition)) { 664 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 665 666 // Fold the branch/switch. 667 TerminatorInst *BBTerm = BB->getTerminator(); 668 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 669 if (i == BestSucc) continue; 670 BBTerm->getSuccessor(i)->removePredecessor(BB, true); 671 } 672 673 DEBUG(dbgs() << " In block '" << BB->getName() 674 << "' folding undef terminator: " << *BBTerm << '\n'); 675 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 676 BBTerm->eraseFromParent(); 677 return true; 678 } 679 680 // If the terminator of this block is branching on a constant, simplify the 681 // terminator to an unconditional branch. This can occur due to threading in 682 // other blocks. 683 if (getKnownConstant(Condition, Preference)) { 684 DEBUG(dbgs() << " In block '" << BB->getName() 685 << "' folding terminator: " << *BB->getTerminator() << '\n'); 686 ++NumFolds; 687 ConstantFoldTerminator(BB); 688 return true; 689 } 690 691 Instruction *CondInst = dyn_cast<Instruction>(Condition); 692 693 // All the rest of our checks depend on the condition being an instruction. 694 if (CondInst == 0) { 695 // FIXME: Unify this with code below. 696 if (ProcessThreadableEdges(Condition, BB, Preference)) 697 return true; 698 return false; 699 } 700 701 702 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 703 // For a comparison where the LHS is outside this block, it's possible 704 // that we've branched on it before. Used LVI to see if we can simplify 705 // the branch based on that. 706 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 707 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 708 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 709 if (CondBr && CondConst && CondBr->isConditional() && PI != PE && 710 (!isa<Instruction>(CondCmp->getOperand(0)) || 711 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { 712 // For predecessor edge, determine if the comparison is true or false 713 // on that edge. If they're all true or all false, we can simplify the 714 // branch. 715 // FIXME: We could handle mixed true/false by duplicating code. 716 LazyValueInfo::Tristate Baseline = 717 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0), 718 CondConst, *PI, BB); 719 if (Baseline != LazyValueInfo::Unknown) { 720 // Check that all remaining incoming values match the first one. 721 while (++PI != PE) { 722 LazyValueInfo::Tristate Ret = 723 LVI->getPredicateOnEdge(CondCmp->getPredicate(), 724 CondCmp->getOperand(0), CondConst, *PI, BB); 725 if (Ret != Baseline) break; 726 } 727 728 // If we terminated early, then one of the values didn't match. 729 if (PI == PE) { 730 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0; 731 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1; 732 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); 733 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); 734 CondBr->eraseFromParent(); 735 return true; 736 } 737 } 738 } 739 } 740 741 // Check for some cases that are worth simplifying. Right now we want to look 742 // for loads that are used by a switch or by the condition for the branch. If 743 // we see one, check to see if it's partially redundant. If so, insert a PHI 744 // which can then be used to thread the values. 745 // 746 Value *SimplifyValue = CondInst; 747 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 748 if (isa<Constant>(CondCmp->getOperand(1))) 749 SimplifyValue = CondCmp->getOperand(0); 750 751 // TODO: There are other places where load PRE would be profitable, such as 752 // more complex comparisons. 753 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 754 if (SimplifyPartiallyRedundantLoad(LI)) 755 return true; 756 757 758 // Handle a variety of cases where we are branching on something derived from 759 // a PHI node in the current block. If we can prove that any predecessors 760 // compute a predictable value based on a PHI node, thread those predecessors. 761 // 762 if (ProcessThreadableEdges(CondInst, BB, Preference)) 763 return true; 764 765 // If this is an otherwise-unfoldable branch on a phi node in the current 766 // block, see if we can simplify. 767 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 768 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 769 return ProcessBranchOnPHI(PN); 770 771 772 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 773 if (CondInst->getOpcode() == Instruction::Xor && 774 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 775 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 776 777 778 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 779 // "(X == 4)", thread through this block. 780 781 return false; 782} 783 784 785/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 786/// load instruction, eliminate it by replacing it with a PHI node. This is an 787/// important optimization that encourages jump threading, and needs to be run 788/// interlaced with other jump threading tasks. 789bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 790 // Don't hack volatile loads. 791 if (LI->isVolatile()) return false; 792 793 // If the load is defined in a block with exactly one predecessor, it can't be 794 // partially redundant. 795 BasicBlock *LoadBB = LI->getParent(); 796 if (LoadBB->getSinglePredecessor()) 797 return false; 798 799 Value *LoadedPtr = LI->getOperand(0); 800 801 // If the loaded operand is defined in the LoadBB, it can't be available. 802 // TODO: Could do simple PHI translation, that would be fun :) 803 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 804 if (PtrOp->getParent() == LoadBB) 805 return false; 806 807 // Scan a few instructions up from the load, to see if it is obviously live at 808 // the entry to its block. 809 BasicBlock::iterator BBIt = LI; 810 811 if (Value *AvailableVal = 812 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 813 // If the value if the load is locally available within the block, just use 814 // it. This frequently occurs for reg2mem'd allocas. 815 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 816 817 // If the returned value is the load itself, replace with an undef. This can 818 // only happen in dead loops. 819 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 820 LI->replaceAllUsesWith(AvailableVal); 821 LI->eraseFromParent(); 822 return true; 823 } 824 825 // Otherwise, if we scanned the whole block and got to the top of the block, 826 // we know the block is locally transparent to the load. If not, something 827 // might clobber its value. 828 if (BBIt != LoadBB->begin()) 829 return false; 830 831 832 SmallPtrSet<BasicBlock*, 8> PredsScanned; 833 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 834 AvailablePredsTy AvailablePreds; 835 BasicBlock *OneUnavailablePred = 0; 836 837 // If we got here, the loaded value is transparent through to the start of the 838 // block. Check to see if it is available in any of the predecessor blocks. 839 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 840 PI != PE; ++PI) { 841 BasicBlock *PredBB = *PI; 842 843 // If we already scanned this predecessor, skip it. 844 if (!PredsScanned.insert(PredBB)) 845 continue; 846 847 // Scan the predecessor to see if the value is available in the pred. 848 BBIt = PredBB->end(); 849 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 850 if (!PredAvailable) { 851 OneUnavailablePred = PredBB; 852 continue; 853 } 854 855 // If so, this load is partially redundant. Remember this info so that we 856 // can create a PHI node. 857 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 858 } 859 860 // If the loaded value isn't available in any predecessor, it isn't partially 861 // redundant. 862 if (AvailablePreds.empty()) return false; 863 864 // Okay, the loaded value is available in at least one (and maybe all!) 865 // predecessors. If the value is unavailable in more than one unique 866 // predecessor, we want to insert a merge block for those common predecessors. 867 // This ensures that we only have to insert one reload, thus not increasing 868 // code size. 869 BasicBlock *UnavailablePred = 0; 870 871 // If there is exactly one predecessor where the value is unavailable, the 872 // already computed 'OneUnavailablePred' block is it. If it ends in an 873 // unconditional branch, we know that it isn't a critical edge. 874 if (PredsScanned.size() == AvailablePreds.size()+1 && 875 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 876 UnavailablePred = OneUnavailablePred; 877 } else if (PredsScanned.size() != AvailablePreds.size()) { 878 // Otherwise, we had multiple unavailable predecessors or we had a critical 879 // edge from the one. 880 SmallVector<BasicBlock*, 8> PredsToSplit; 881 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 882 883 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 884 AvailablePredSet.insert(AvailablePreds[i].first); 885 886 // Add all the unavailable predecessors to the PredsToSplit list. 887 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 888 PI != PE; ++PI) { 889 BasicBlock *P = *PI; 890 // If the predecessor is an indirect goto, we can't split the edge. 891 if (isa<IndirectBrInst>(P->getTerminator())) 892 return false; 893 894 if (!AvailablePredSet.count(P)) 895 PredsToSplit.push_back(P); 896 } 897 898 // Split them out to their own block. 899 UnavailablePred = 900 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 901 "thread-pre-split", this); 902 } 903 904 // If the value isn't available in all predecessors, then there will be 905 // exactly one where it isn't available. Insert a load on that edge and add 906 // it to the AvailablePreds list. 907 if (UnavailablePred) { 908 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 909 "Can't handle critical edge here!"); 910 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 911 LI->getAlignment(), 912 UnavailablePred->getTerminator()); 913 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 914 } 915 916 // Now we know that each predecessor of this block has a value in 917 // AvailablePreds, sort them for efficient access as we're walking the preds. 918 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 919 920 // Create a PHI node at the start of the block for the PRE'd load value. 921 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 922 PN->takeName(LI); 923 924 // Insert new entries into the PHI for each predecessor. A single block may 925 // have multiple entries here. 926 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 927 ++PI) { 928 BasicBlock *P = *PI; 929 AvailablePredsTy::iterator I = 930 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 931 std::make_pair(P, (Value*)0)); 932 933 assert(I != AvailablePreds.end() && I->first == P && 934 "Didn't find entry for predecessor!"); 935 936 PN->addIncoming(I->second, I->first); 937 } 938 939 //cerr << "PRE: " << *LI << *PN << "\n"; 940 941 LI->replaceAllUsesWith(PN); 942 LI->eraseFromParent(); 943 944 return true; 945} 946 947/// FindMostPopularDest - The specified list contains multiple possible 948/// threadable destinations. Pick the one that occurs the most frequently in 949/// the list. 950static BasicBlock * 951FindMostPopularDest(BasicBlock *BB, 952 const SmallVectorImpl<std::pair<BasicBlock*, 953 BasicBlock*> > &PredToDestList) { 954 assert(!PredToDestList.empty()); 955 956 // Determine popularity. If there are multiple possible destinations, we 957 // explicitly choose to ignore 'undef' destinations. We prefer to thread 958 // blocks with known and real destinations to threading undef. We'll handle 959 // them later if interesting. 960 DenseMap<BasicBlock*, unsigned> DestPopularity; 961 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 962 if (PredToDestList[i].second) 963 DestPopularity[PredToDestList[i].second]++; 964 965 // Find the most popular dest. 966 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 967 BasicBlock *MostPopularDest = DPI->first; 968 unsigned Popularity = DPI->second; 969 SmallVector<BasicBlock*, 4> SamePopularity; 970 971 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 972 // If the popularity of this entry isn't higher than the popularity we've 973 // seen so far, ignore it. 974 if (DPI->second < Popularity) 975 ; // ignore. 976 else if (DPI->second == Popularity) { 977 // If it is the same as what we've seen so far, keep track of it. 978 SamePopularity.push_back(DPI->first); 979 } else { 980 // If it is more popular, remember it. 981 SamePopularity.clear(); 982 MostPopularDest = DPI->first; 983 Popularity = DPI->second; 984 } 985 } 986 987 // Okay, now we know the most popular destination. If there is more than one 988 // destination, we need to determine one. This is arbitrary, but we need 989 // to make a deterministic decision. Pick the first one that appears in the 990 // successor list. 991 if (!SamePopularity.empty()) { 992 SamePopularity.push_back(MostPopularDest); 993 TerminatorInst *TI = BB->getTerminator(); 994 for (unsigned i = 0; ; ++i) { 995 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 996 997 if (std::find(SamePopularity.begin(), SamePopularity.end(), 998 TI->getSuccessor(i)) == SamePopularity.end()) 999 continue; 1000 1001 MostPopularDest = TI->getSuccessor(i); 1002 break; 1003 } 1004 } 1005 1006 // Okay, we have finally picked the most popular destination. 1007 return MostPopularDest; 1008} 1009 1010bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 1011 ConstantPreference Preference) { 1012 // If threading this would thread across a loop header, don't even try to 1013 // thread the edge. 1014 if (LoopHeaders.count(BB)) 1015 return false; 1016 1017 PredValueInfoTy PredValues; 1018 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference)) 1019 return false; 1020 1021 assert(!PredValues.empty() && 1022 "ComputeValueKnownInPredecessors returned true with no values"); 1023 1024 DEBUG(dbgs() << "IN BB: " << *BB; 1025 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1026 dbgs() << " BB '" << BB->getName() << "': FOUND condition = " 1027 << *PredValues[i].first 1028 << " for pred '" << PredValues[i].second->getName() << "'.\n"; 1029 }); 1030 1031 // Decide what we want to thread through. Convert our list of known values to 1032 // a list of known destinations for each pred. This also discards duplicate 1033 // predecessors and keeps track of the undefined inputs (which are represented 1034 // as a null dest in the PredToDestList). 1035 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1036 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1037 1038 BasicBlock *OnlyDest = 0; 1039 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1040 1041 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1042 BasicBlock *Pred = PredValues[i].second; 1043 if (!SeenPreds.insert(Pred)) 1044 continue; // Duplicate predecessor entry. 1045 1046 // If the predecessor ends with an indirect goto, we can't change its 1047 // destination. 1048 if (isa<IndirectBrInst>(Pred->getTerminator())) 1049 continue; 1050 1051 Constant *Val = PredValues[i].first; 1052 1053 BasicBlock *DestBB; 1054 if (isa<UndefValue>(Val)) 1055 DestBB = 0; 1056 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1057 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); 1058 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 1059 DestBB = SI->getSuccessor(SI->findCaseValue(cast<ConstantInt>(Val))); 1060 else { 1061 assert(isa<IndirectBrInst>(BB->getTerminator()) 1062 && "Unexpected terminator"); 1063 DestBB = cast<BlockAddress>(Val)->getBasicBlock(); 1064 } 1065 1066 // If we have exactly one destination, remember it for efficiency below. 1067 if (PredToDestList.empty()) 1068 OnlyDest = DestBB; 1069 else if (OnlyDest != DestBB) 1070 OnlyDest = MultipleDestSentinel; 1071 1072 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1073 } 1074 1075 // If all edges were unthreadable, we fail. 1076 if (PredToDestList.empty()) 1077 return false; 1078 1079 // Determine which is the most common successor. If we have many inputs and 1080 // this block is a switch, we want to start by threading the batch that goes 1081 // to the most popular destination first. If we only know about one 1082 // threadable destination (the common case) we can avoid this. 1083 BasicBlock *MostPopularDest = OnlyDest; 1084 1085 if (MostPopularDest == MultipleDestSentinel) 1086 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1087 1088 // Now that we know what the most popular destination is, factor all 1089 // predecessors that will jump to it into a single predecessor. 1090 SmallVector<BasicBlock*, 16> PredsToFactor; 1091 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1092 if (PredToDestList[i].second == MostPopularDest) { 1093 BasicBlock *Pred = PredToDestList[i].first; 1094 1095 // This predecessor may be a switch or something else that has multiple 1096 // edges to the block. Factor each of these edges by listing them 1097 // according to # occurrences in PredsToFactor. 1098 TerminatorInst *PredTI = Pred->getTerminator(); 1099 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1100 if (PredTI->getSuccessor(i) == BB) 1101 PredsToFactor.push_back(Pred); 1102 } 1103 1104 // If the threadable edges are branching on an undefined value, we get to pick 1105 // the destination that these predecessors should get to. 1106 if (MostPopularDest == 0) 1107 MostPopularDest = BB->getTerminator()-> 1108 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1109 1110 // Ok, try to thread it! 1111 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1112} 1113 1114/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1115/// a PHI node in the current block. See if there are any simplifications we 1116/// can do based on inputs to the phi node. 1117/// 1118bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1119 BasicBlock *BB = PN->getParent(); 1120 1121 // TODO: We could make use of this to do it once for blocks with common PHI 1122 // values. 1123 SmallVector<BasicBlock*, 1> PredBBs; 1124 PredBBs.resize(1); 1125 1126 // If any of the predecessor blocks end in an unconditional branch, we can 1127 // *duplicate* the conditional branch into that block in order to further 1128 // encourage jump threading and to eliminate cases where we have branch on a 1129 // phi of an icmp (branch on icmp is much better). 1130 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1131 BasicBlock *PredBB = PN->getIncomingBlock(i); 1132 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1133 if (PredBr->isUnconditional()) { 1134 PredBBs[0] = PredBB; 1135 // Try to duplicate BB into PredBB. 1136 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1137 return true; 1138 } 1139 } 1140 1141 return false; 1142} 1143 1144/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1145/// a xor instruction in the current block. See if there are any 1146/// simplifications we can do based on inputs to the xor. 1147/// 1148bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1149 BasicBlock *BB = BO->getParent(); 1150 1151 // If either the LHS or RHS of the xor is a constant, don't do this 1152 // optimization. 1153 if (isa<ConstantInt>(BO->getOperand(0)) || 1154 isa<ConstantInt>(BO->getOperand(1))) 1155 return false; 1156 1157 // If the first instruction in BB isn't a phi, we won't be able to infer 1158 // anything special about any particular predecessor. 1159 if (!isa<PHINode>(BB->front())) 1160 return false; 1161 1162 // If we have a xor as the branch input to this block, and we know that the 1163 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1164 // the condition into the predecessor and fix that value to true, saving some 1165 // logical ops on that path and encouraging other paths to simplify. 1166 // 1167 // This copies something like this: 1168 // 1169 // BB: 1170 // %X = phi i1 [1], [%X'] 1171 // %Y = icmp eq i32 %A, %B 1172 // %Z = xor i1 %X, %Y 1173 // br i1 %Z, ... 1174 // 1175 // Into: 1176 // BB': 1177 // %Y = icmp ne i32 %A, %B 1178 // br i1 %Z, ... 1179 1180 PredValueInfoTy XorOpValues; 1181 bool isLHS = true; 1182 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, 1183 WantInteger)) { 1184 assert(XorOpValues.empty()); 1185 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, 1186 WantInteger)) 1187 return false; 1188 isLHS = false; 1189 } 1190 1191 assert(!XorOpValues.empty() && 1192 "ComputeValueKnownInPredecessors returned true with no values"); 1193 1194 // Scan the information to see which is most popular: true or false. The 1195 // predecessors can be of the set true, false, or undef. 1196 unsigned NumTrue = 0, NumFalse = 0; 1197 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1198 if (isa<UndefValue>(XorOpValues[i].first)) 1199 // Ignore undefs for the count. 1200 continue; 1201 if (cast<ConstantInt>(XorOpValues[i].first)->isZero()) 1202 ++NumFalse; 1203 else 1204 ++NumTrue; 1205 } 1206 1207 // Determine which value to split on, true, false, or undef if neither. 1208 ConstantInt *SplitVal = 0; 1209 if (NumTrue > NumFalse) 1210 SplitVal = ConstantInt::getTrue(BB->getContext()); 1211 else if (NumTrue != 0 || NumFalse != 0) 1212 SplitVal = ConstantInt::getFalse(BB->getContext()); 1213 1214 // Collect all of the blocks that this can be folded into so that we can 1215 // factor this once and clone it once. 1216 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1217 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1218 if (XorOpValues[i].first != SplitVal && 1219 !isa<UndefValue>(XorOpValues[i].first)) 1220 continue; 1221 1222 BlocksToFoldInto.push_back(XorOpValues[i].second); 1223 } 1224 1225 // If we inferred a value for all of the predecessors, then duplication won't 1226 // help us. However, we can just replace the LHS or RHS with the constant. 1227 if (BlocksToFoldInto.size() == 1228 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1229 if (SplitVal == 0) { 1230 // If all preds provide undef, just nuke the xor, because it is undef too. 1231 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1232 BO->eraseFromParent(); 1233 } else if (SplitVal->isZero()) { 1234 // If all preds provide 0, replace the xor with the other input. 1235 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1236 BO->eraseFromParent(); 1237 } else { 1238 // If all preds provide 1, set the computed value to 1. 1239 BO->setOperand(!isLHS, SplitVal); 1240 } 1241 1242 return true; 1243 } 1244 1245 // Try to duplicate BB into PredBB. 1246 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1247} 1248 1249 1250/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1251/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1252/// NewPred using the entries from OldPred (suitably mapped). 1253static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1254 BasicBlock *OldPred, 1255 BasicBlock *NewPred, 1256 DenseMap<Instruction*, Value*> &ValueMap) { 1257 for (BasicBlock::iterator PNI = PHIBB->begin(); 1258 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1259 // Ok, we have a PHI node. Figure out what the incoming value was for the 1260 // DestBlock. 1261 Value *IV = PN->getIncomingValueForBlock(OldPred); 1262 1263 // Remap the value if necessary. 1264 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1265 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1266 if (I != ValueMap.end()) 1267 IV = I->second; 1268 } 1269 1270 PN->addIncoming(IV, NewPred); 1271 } 1272} 1273 1274/// ThreadEdge - We have decided that it is safe and profitable to factor the 1275/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1276/// across BB. Transform the IR to reflect this change. 1277bool JumpThreading::ThreadEdge(BasicBlock *BB, 1278 const SmallVectorImpl<BasicBlock*> &PredBBs, 1279 BasicBlock *SuccBB) { 1280 // If threading to the same block as we come from, we would infinite loop. 1281 if (SuccBB == BB) { 1282 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1283 << "' - would thread to self!\n"); 1284 return false; 1285 } 1286 1287 // If threading this would thread across a loop header, don't thread the edge. 1288 // See the comments above FindLoopHeaders for justifications and caveats. 1289 if (LoopHeaders.count(BB)) { 1290 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1291 << "' to dest BB '" << SuccBB->getName() 1292 << "' - it might create an irreducible loop!\n"); 1293 return false; 1294 } 1295 1296 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1297 if (JumpThreadCost > Threshold) { 1298 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1299 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1300 return false; 1301 } 1302 1303 // And finally, do it! Start by factoring the predecessors is needed. 1304 BasicBlock *PredBB; 1305 if (PredBBs.size() == 1) 1306 PredBB = PredBBs[0]; 1307 else { 1308 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1309 << " common predecessors.\n"); 1310 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1311 ".thr_comm", this); 1312 } 1313 1314 // And finally, do it! 1315 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1316 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1317 << ", across block:\n " 1318 << *BB << "\n"); 1319 1320 LVI->threadEdge(PredBB, BB, SuccBB); 1321 1322 // We are going to have to map operands from the original BB block to the new 1323 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1324 // account for entry from PredBB. 1325 DenseMap<Instruction*, Value*> ValueMapping; 1326 1327 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1328 BB->getName()+".thread", 1329 BB->getParent(), BB); 1330 NewBB->moveAfter(PredBB); 1331 1332 BasicBlock::iterator BI = BB->begin(); 1333 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1334 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1335 1336 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1337 // mapping and using it to remap operands in the cloned instructions. 1338 for (; !isa<TerminatorInst>(BI); ++BI) { 1339 Instruction *New = BI->clone(); 1340 New->setName(BI->getName()); 1341 NewBB->getInstList().push_back(New); 1342 ValueMapping[BI] = New; 1343 1344 // Remap operands to patch up intra-block references. 1345 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1346 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1347 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1348 if (I != ValueMapping.end()) 1349 New->setOperand(i, I->second); 1350 } 1351 } 1352 1353 // We didn't copy the terminator from BB over to NewBB, because there is now 1354 // an unconditional jump to SuccBB. Insert the unconditional jump. 1355 BranchInst::Create(SuccBB, NewBB); 1356 1357 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1358 // PHI nodes for NewBB now. 1359 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1360 1361 // If there were values defined in BB that are used outside the block, then we 1362 // now have to update all uses of the value to use either the original value, 1363 // the cloned value, or some PHI derived value. This can require arbitrary 1364 // PHI insertion, of which we are prepared to do, clean these up now. 1365 SSAUpdater SSAUpdate; 1366 SmallVector<Use*, 16> UsesToRename; 1367 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1368 // Scan all uses of this instruction to see if it is used outside of its 1369 // block, and if so, record them in UsesToRename. 1370 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1371 ++UI) { 1372 Instruction *User = cast<Instruction>(*UI); 1373 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1374 if (UserPN->getIncomingBlock(UI) == BB) 1375 continue; 1376 } else if (User->getParent() == BB) 1377 continue; 1378 1379 UsesToRename.push_back(&UI.getUse()); 1380 } 1381 1382 // If there are no uses outside the block, we're done with this instruction. 1383 if (UsesToRename.empty()) 1384 continue; 1385 1386 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1387 1388 // We found a use of I outside of BB. Rename all uses of I that are outside 1389 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1390 // with the two values we know. 1391 SSAUpdate.Initialize(I->getType(), I->getName()); 1392 SSAUpdate.AddAvailableValue(BB, I); 1393 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1394 1395 while (!UsesToRename.empty()) 1396 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1397 DEBUG(dbgs() << "\n"); 1398 } 1399 1400 1401 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1402 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1403 // us to simplify any PHI nodes in BB. 1404 TerminatorInst *PredTerm = PredBB->getTerminator(); 1405 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1406 if (PredTerm->getSuccessor(i) == BB) { 1407 BB->removePredecessor(PredBB, true); 1408 PredTerm->setSuccessor(i, NewBB); 1409 } 1410 1411 // At this point, the IR is fully up to date and consistent. Do a quick scan 1412 // over the new instructions and zap any that are constants or dead. This 1413 // frequently happens because of phi translation. 1414 SimplifyInstructionsInBlock(NewBB, TD); 1415 1416 // Threaded an edge! 1417 ++NumThreads; 1418 return true; 1419} 1420 1421/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1422/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1423/// If we can duplicate the contents of BB up into PredBB do so now, this 1424/// improves the odds that the branch will be on an analyzable instruction like 1425/// a compare. 1426bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1427 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1428 assert(!PredBBs.empty() && "Can't handle an empty set"); 1429 1430 // If BB is a loop header, then duplicating this block outside the loop would 1431 // cause us to transform this into an irreducible loop, don't do this. 1432 // See the comments above FindLoopHeaders for justifications and caveats. 1433 if (LoopHeaders.count(BB)) { 1434 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1435 << "' into predecessor block '" << PredBBs[0]->getName() 1436 << "' - it might create an irreducible loop!\n"); 1437 return false; 1438 } 1439 1440 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1441 if (DuplicationCost > Threshold) { 1442 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1443 << "' - Cost is too high: " << DuplicationCost << "\n"); 1444 return false; 1445 } 1446 1447 // And finally, do it! Start by factoring the predecessors is needed. 1448 BasicBlock *PredBB; 1449 if (PredBBs.size() == 1) 1450 PredBB = PredBBs[0]; 1451 else { 1452 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1453 << " common predecessors.\n"); 1454 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1455 ".thr_comm", this); 1456 } 1457 1458 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1459 // of PredBB. 1460 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1461 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1462 << DuplicationCost << " block is:" << *BB << "\n"); 1463 1464 // Unless PredBB ends with an unconditional branch, split the edge so that we 1465 // can just clone the bits from BB into the end of the new PredBB. 1466 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1467 1468 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1469 PredBB = SplitEdge(PredBB, BB, this); 1470 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1471 } 1472 1473 // We are going to have to map operands from the original BB block into the 1474 // PredBB block. Evaluate PHI nodes in BB. 1475 DenseMap<Instruction*, Value*> ValueMapping; 1476 1477 BasicBlock::iterator BI = BB->begin(); 1478 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1479 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1480 1481 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1482 // mapping and using it to remap operands in the cloned instructions. 1483 for (; BI != BB->end(); ++BI) { 1484 Instruction *New = BI->clone(); 1485 1486 // Remap operands to patch up intra-block references. 1487 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1488 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1489 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1490 if (I != ValueMapping.end()) 1491 New->setOperand(i, I->second); 1492 } 1493 1494 // If this instruction can be simplified after the operands are updated, 1495 // just use the simplified value instead. This frequently happens due to 1496 // phi translation. 1497 if (Value *IV = SimplifyInstruction(New, TD)) { 1498 delete New; 1499 ValueMapping[BI] = IV; 1500 } else { 1501 // Otherwise, insert the new instruction into the block. 1502 New->setName(BI->getName()); 1503 PredBB->getInstList().insert(OldPredBranch, New); 1504 ValueMapping[BI] = New; 1505 } 1506 } 1507 1508 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1509 // add entries to the PHI nodes for branch from PredBB now. 1510 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1511 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1512 ValueMapping); 1513 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1514 ValueMapping); 1515 1516 // If there were values defined in BB that are used outside the block, then we 1517 // now have to update all uses of the value to use either the original value, 1518 // the cloned value, or some PHI derived value. This can require arbitrary 1519 // PHI insertion, of which we are prepared to do, clean these up now. 1520 SSAUpdater SSAUpdate; 1521 SmallVector<Use*, 16> UsesToRename; 1522 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1523 // Scan all uses of this instruction to see if it is used outside of its 1524 // block, and if so, record them in UsesToRename. 1525 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1526 ++UI) { 1527 Instruction *User = cast<Instruction>(*UI); 1528 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1529 if (UserPN->getIncomingBlock(UI) == BB) 1530 continue; 1531 } else if (User->getParent() == BB) 1532 continue; 1533 1534 UsesToRename.push_back(&UI.getUse()); 1535 } 1536 1537 // If there are no uses outside the block, we're done with this instruction. 1538 if (UsesToRename.empty()) 1539 continue; 1540 1541 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1542 1543 // We found a use of I outside of BB. Rename all uses of I that are outside 1544 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1545 // with the two values we know. 1546 SSAUpdate.Initialize(I->getType(), I->getName()); 1547 SSAUpdate.AddAvailableValue(BB, I); 1548 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1549 1550 while (!UsesToRename.empty()) 1551 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1552 DEBUG(dbgs() << "\n"); 1553 } 1554 1555 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1556 // that we nuked. 1557 BB->removePredecessor(PredBB, true); 1558 1559 // Remove the unconditional branch at the end of the PredBB block. 1560 OldPredBranch->eraseFromParent(); 1561 1562 ++NumDupes; 1563 return true; 1564} 1565 1566 1567