JumpThreading.cpp revision 24473120a253a05f3601cd3373403b47e6d03d41
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/Target/TargetLibraryInfo.h" 28#include "llvm/ADT/DenseMap.h" 29#include "llvm/ADT/DenseSet.h" 30#include "llvm/ADT/Statistic.h" 31#include "llvm/ADT/STLExtras.h" 32#include "llvm/ADT/SmallPtrSet.h" 33#include "llvm/ADT/SmallSet.h" 34#include "llvm/Support/CommandLine.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/ValueHandle.h" 37#include "llvm/Support/raw_ostream.h" 38using namespace llvm; 39 40STATISTIC(NumThreads, "Number of jumps threaded"); 41STATISTIC(NumFolds, "Number of terminators folded"); 42STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 43 44static cl::opt<unsigned> 45Threshold("jump-threading-threshold", 46 cl::desc("Max block size to duplicate for jump threading"), 47 cl::init(6), cl::Hidden); 48 49namespace { 50 // These are at global scope so static functions can use them too. 51 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo; 52 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy; 53 54 // This is used to keep track of what kind of constant we're currently hoping 55 // to find. 56 enum ConstantPreference { 57 WantInteger, 58 WantBlockAddress 59 }; 60 61 /// This pass performs 'jump threading', which looks at blocks that have 62 /// multiple predecessors and multiple successors. If one or more of the 63 /// predecessors of the block can be proven to always jump to one of the 64 /// successors, we forward the edge from the predecessor to the successor by 65 /// duplicating the contents of this block. 66 /// 67 /// An example of when this can occur is code like this: 68 /// 69 /// if () { ... 70 /// X = 4; 71 /// } 72 /// if (X < 3) { 73 /// 74 /// In this case, the unconditional branch at the end of the first if can be 75 /// revectored to the false side of the second if. 76 /// 77 class JumpThreading : public FunctionPass { 78 TargetData *TD; 79 TargetLibraryInfo *TLI; 80 LazyValueInfo *LVI; 81#ifdef NDEBUG 82 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 83#else 84 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 85#endif 86 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet; 87 88 // RAII helper for updating the recursion stack. 89 struct RecursionSetRemover { 90 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet; 91 std::pair<Value*, BasicBlock*> ThePair; 92 93 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S, 94 std::pair<Value*, BasicBlock*> P) 95 : TheSet(S), ThePair(P) { } 96 97 ~RecursionSetRemover() { 98 TheSet.erase(ThePair); 99 } 100 }; 101 public: 102 static char ID; // Pass identification 103 JumpThreading() : FunctionPass(ID) { 104 initializeJumpThreadingPass(*PassRegistry::getPassRegistry()); 105 } 106 107 bool runOnFunction(Function &F); 108 109 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 110 AU.addRequired<LazyValueInfo>(); 111 AU.addPreserved<LazyValueInfo>(); 112 AU.addRequired<TargetLibraryInfo>(); 113 } 114 115 void FindLoopHeaders(Function &F); 116 bool ProcessBlock(BasicBlock *BB); 117 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 118 BasicBlock *SuccBB); 119 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 120 const SmallVectorImpl<BasicBlock *> &PredBBs); 121 122 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 123 PredValueInfo &Result, 124 ConstantPreference Preference); 125 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 126 ConstantPreference Preference); 127 128 bool ProcessBranchOnPHI(PHINode *PN); 129 bool ProcessBranchOnXOR(BinaryOperator *BO); 130 131 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 132 }; 133} 134 135char JumpThreading::ID = 0; 136INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", 137 "Jump Threading", false, false) 138INITIALIZE_PASS_DEPENDENCY(LazyValueInfo) 139INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 140INITIALIZE_PASS_END(JumpThreading, "jump-threading", 141 "Jump Threading", false, false) 142 143// Public interface to the Jump Threading pass 144FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 145 146/// runOnFunction - Top level algorithm. 147/// 148bool JumpThreading::runOnFunction(Function &F) { 149 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 150 TD = getAnalysisIfAvailable<TargetData>(); 151 TLI = &getAnalysis<TargetLibraryInfo>(); 152 LVI = &getAnalysis<LazyValueInfo>(); 153 154 FindLoopHeaders(F); 155 156 bool Changed, EverChanged = false; 157 do { 158 Changed = false; 159 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 160 BasicBlock *BB = I; 161 // Thread all of the branches we can over this block. 162 while (ProcessBlock(BB)) 163 Changed = true; 164 165 ++I; 166 167 // If the block is trivially dead, zap it. This eliminates the successor 168 // edges which simplifies the CFG. 169 if (pred_begin(BB) == pred_end(BB) && 170 BB != &BB->getParent()->getEntryBlock()) { 171 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 172 << "' with terminator: " << *BB->getTerminator() << '\n'); 173 LoopHeaders.erase(BB); 174 LVI->eraseBlock(BB); 175 DeleteDeadBlock(BB); 176 Changed = true; 177 continue; 178 } 179 180 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 181 182 // Can't thread an unconditional jump, but if the block is "almost 183 // empty", we can replace uses of it with uses of the successor and make 184 // this dead. 185 if (BI && BI->isUnconditional() && 186 BB != &BB->getParent()->getEntryBlock() && 187 // If the terminator is the only non-phi instruction, try to nuke it. 188 BB->getFirstNonPHIOrDbg()->isTerminator()) { 189 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 190 // block, we have to make sure it isn't in the LoopHeaders set. We 191 // reinsert afterward if needed. 192 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 193 BasicBlock *Succ = BI->getSuccessor(0); 194 195 // FIXME: It is always conservatively correct to drop the info 196 // for a block even if it doesn't get erased. This isn't totally 197 // awesome, but it allows us to use AssertingVH to prevent nasty 198 // dangling pointer issues within LazyValueInfo. 199 LVI->eraseBlock(BB); 200 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 201 Changed = true; 202 // If we deleted BB and BB was the header of a loop, then the 203 // successor is now the header of the loop. 204 BB = Succ; 205 } 206 207 if (ErasedFromLoopHeaders) 208 LoopHeaders.insert(BB); 209 } 210 } 211 EverChanged |= Changed; 212 } while (Changed); 213 214 LoopHeaders.clear(); 215 return EverChanged; 216} 217 218/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 219/// thread across it. 220static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 221 /// Ignore PHI nodes, these will be flattened when duplication happens. 222 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 223 224 // FIXME: THREADING will delete values that are just used to compute the 225 // branch, so they shouldn't count against the duplication cost. 226 227 228 // Sum up the cost of each instruction until we get to the terminator. Don't 229 // include the terminator because the copy won't include it. 230 unsigned Size = 0; 231 for (; !isa<TerminatorInst>(I); ++I) { 232 // Debugger intrinsics don't incur code size. 233 if (isa<DbgInfoIntrinsic>(I)) continue; 234 235 // If this is a pointer->pointer bitcast, it is free. 236 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 237 continue; 238 239 // All other instructions count for at least one unit. 240 ++Size; 241 242 // Calls are more expensive. If they are non-intrinsic calls, we model them 243 // as having cost of 4. If they are a non-vector intrinsic, we model them 244 // as having cost of 2 total, and if they are a vector intrinsic, we model 245 // them as having cost 1. 246 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 247 if (!isa<IntrinsicInst>(CI)) 248 Size += 3; 249 else if (!CI->getType()->isVectorTy()) 250 Size += 1; 251 } 252 } 253 254 // Threading through a switch statement is particularly profitable. If this 255 // block ends in a switch, decrease its cost to make it more likely to happen. 256 if (isa<SwitchInst>(I)) 257 Size = Size > 6 ? Size-6 : 0; 258 259 // The same holds for indirect branches, but slightly more so. 260 if (isa<IndirectBrInst>(I)) 261 Size = Size > 8 ? Size-8 : 0; 262 263 return Size; 264} 265 266/// FindLoopHeaders - We do not want jump threading to turn proper loop 267/// structures into irreducible loops. Doing this breaks up the loop nesting 268/// hierarchy and pessimizes later transformations. To prevent this from 269/// happening, we first have to find the loop headers. Here we approximate this 270/// by finding targets of backedges in the CFG. 271/// 272/// Note that there definitely are cases when we want to allow threading of 273/// edges across a loop header. For example, threading a jump from outside the 274/// loop (the preheader) to an exit block of the loop is definitely profitable. 275/// It is also almost always profitable to thread backedges from within the loop 276/// to exit blocks, and is often profitable to thread backedges to other blocks 277/// within the loop (forming a nested loop). This simple analysis is not rich 278/// enough to track all of these properties and keep it up-to-date as the CFG 279/// mutates, so we don't allow any of these transformations. 280/// 281void JumpThreading::FindLoopHeaders(Function &F) { 282 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 283 FindFunctionBackedges(F, Edges); 284 285 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 286 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 287} 288 289/// getKnownConstant - Helper method to determine if we can thread over a 290/// terminator with the given value as its condition, and if so what value to 291/// use for that. What kind of value this is depends on whether we want an 292/// integer or a block address, but an undef is always accepted. 293/// Returns null if Val is null or not an appropriate constant. 294static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { 295 if (!Val) 296 return 0; 297 298 // Undef is "known" enough. 299 if (UndefValue *U = dyn_cast<UndefValue>(Val)) 300 return U; 301 302 if (Preference == WantBlockAddress) 303 return dyn_cast<BlockAddress>(Val->stripPointerCasts()); 304 305 return dyn_cast<ConstantInt>(Val); 306} 307 308/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 309/// if we can infer that the value is a known ConstantInt/BlockAddress or undef 310/// in any of our predecessors. If so, return the known list of value and pred 311/// BB in the result vector. 312/// 313/// This returns true if there were any known values. 314/// 315bool JumpThreading:: 316ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result, 317 ConstantPreference Preference) { 318 // This method walks up use-def chains recursively. Because of this, we could 319 // get into an infinite loop going around loops in the use-def chain. To 320 // prevent this, keep track of what (value, block) pairs we've already visited 321 // and terminate the search if we loop back to them 322 if (!RecursionSet.insert(std::make_pair(V, BB)).second) 323 return false; 324 325 // An RAII help to remove this pair from the recursion set once the recursion 326 // stack pops back out again. 327 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB)); 328 329 // If V is a constant, then it is known in all predecessors. 330 if (Constant *KC = getKnownConstant(V, Preference)) { 331 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 332 Result.push_back(std::make_pair(KC, *PI)); 333 334 return true; 335 } 336 337 // If V is a non-instruction value, or an instruction in a different block, 338 // then it can't be derived from a PHI. 339 Instruction *I = dyn_cast<Instruction>(V); 340 if (I == 0 || I->getParent() != BB) { 341 342 // Okay, if this is a live-in value, see if it has a known value at the end 343 // of any of our predecessors. 344 // 345 // FIXME: This should be an edge property, not a block end property. 346 /// TODO: Per PR2563, we could infer value range information about a 347 /// predecessor based on its terminator. 348 // 349 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 350 // "I" is a non-local compare-with-a-constant instruction. This would be 351 // able to handle value inequalities better, for example if the compare is 352 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 353 // Perhaps getConstantOnEdge should be smart enough to do this? 354 355 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 356 BasicBlock *P = *PI; 357 // If the value is known by LazyValueInfo to be a constant in a 358 // predecessor, use that information to try to thread this block. 359 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 360 if (Constant *KC = getKnownConstant(PredCst, Preference)) 361 Result.push_back(std::make_pair(KC, P)); 362 } 363 364 return !Result.empty(); 365 } 366 367 /// If I is a PHI node, then we know the incoming values for any constants. 368 if (PHINode *PN = dyn_cast<PHINode>(I)) { 369 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 370 Value *InVal = PN->getIncomingValue(i); 371 if (Constant *KC = getKnownConstant(InVal, Preference)) { 372 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 373 } else { 374 Constant *CI = LVI->getConstantOnEdge(InVal, 375 PN->getIncomingBlock(i), BB); 376 if (Constant *KC = getKnownConstant(CI, Preference)) 377 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 378 } 379 } 380 381 return !Result.empty(); 382 } 383 384 PredValueInfoTy LHSVals, RHSVals; 385 386 // Handle some boolean conditions. 387 if (I->getType()->getPrimitiveSizeInBits() == 1) { 388 assert(Preference == WantInteger && "One-bit non-integer type?"); 389 // X | true -> true 390 // X & false -> false 391 if (I->getOpcode() == Instruction::Or || 392 I->getOpcode() == Instruction::And) { 393 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 394 WantInteger); 395 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals, 396 WantInteger); 397 398 if (LHSVals.empty() && RHSVals.empty()) 399 return false; 400 401 ConstantInt *InterestingVal; 402 if (I->getOpcode() == Instruction::Or) 403 InterestingVal = ConstantInt::getTrue(I->getContext()); 404 else 405 InterestingVal = ConstantInt::getFalse(I->getContext()); 406 407 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 408 409 // Scan for the sentinel. If we find an undef, force it to the 410 // interesting value: x|undef -> true and x&undef -> false. 411 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 412 if (LHSVals[i].first == InterestingVal || 413 isa<UndefValue>(LHSVals[i].first)) { 414 Result.push_back(LHSVals[i]); 415 Result.back().first = InterestingVal; 416 LHSKnownBBs.insert(LHSVals[i].second); 417 } 418 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 419 if (RHSVals[i].first == InterestingVal || 420 isa<UndefValue>(RHSVals[i].first)) { 421 // If we already inferred a value for this block on the LHS, don't 422 // re-add it. 423 if (!LHSKnownBBs.count(RHSVals[i].second)) { 424 Result.push_back(RHSVals[i]); 425 Result.back().first = InterestingVal; 426 } 427 } 428 429 return !Result.empty(); 430 } 431 432 // Handle the NOT form of XOR. 433 if (I->getOpcode() == Instruction::Xor && 434 isa<ConstantInt>(I->getOperand(1)) && 435 cast<ConstantInt>(I->getOperand(1))->isOne()) { 436 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result, 437 WantInteger); 438 if (Result.empty()) 439 return false; 440 441 // Invert the known values. 442 for (unsigned i = 0, e = Result.size(); i != e; ++i) 443 Result[i].first = ConstantExpr::getNot(Result[i].first); 444 445 return true; 446 } 447 448 // Try to simplify some other binary operator values. 449 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 450 assert(Preference != WantBlockAddress 451 && "A binary operator creating a block address?"); 452 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 453 PredValueInfoTy LHSVals; 454 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals, 455 WantInteger); 456 457 // Try to use constant folding to simplify the binary operator. 458 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 459 Constant *V = LHSVals[i].first; 460 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); 461 462 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 463 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 464 } 465 } 466 467 return !Result.empty(); 468 } 469 470 // Handle compare with phi operand, where the PHI is defined in this block. 471 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 472 assert(Preference == WantInteger && "Compares only produce integers"); 473 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 474 if (PN && PN->getParent() == BB) { 475 // We can do this simplification if any comparisons fold to true or false. 476 // See if any do. 477 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 478 BasicBlock *PredBB = PN->getIncomingBlock(i); 479 Value *LHS = PN->getIncomingValue(i); 480 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 481 482 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 483 if (Res == 0) { 484 if (!isa<Constant>(RHS)) 485 continue; 486 487 LazyValueInfo::Tristate 488 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 489 cast<Constant>(RHS), PredBB, BB); 490 if (ResT == LazyValueInfo::Unknown) 491 continue; 492 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 493 } 494 495 if (Constant *KC = getKnownConstant(Res, WantInteger)) 496 Result.push_back(std::make_pair(KC, PredBB)); 497 } 498 499 return !Result.empty(); 500 } 501 502 503 // If comparing a live-in value against a constant, see if we know the 504 // live-in value on any predecessors. 505 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) { 506 if (!isa<Instruction>(Cmp->getOperand(0)) || 507 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 508 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 509 510 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ 511 BasicBlock *P = *PI; 512 // If the value is known by LazyValueInfo to be a constant in a 513 // predecessor, use that information to try to thread this block. 514 LazyValueInfo::Tristate Res = 515 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 516 RHSCst, P, BB); 517 if (Res == LazyValueInfo::Unknown) 518 continue; 519 520 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 521 Result.push_back(std::make_pair(ResC, P)); 522 } 523 524 return !Result.empty(); 525 } 526 527 // Try to find a constant value for the LHS of a comparison, 528 // and evaluate it statically if we can. 529 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { 530 PredValueInfoTy LHSVals; 531 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 532 WantInteger); 533 534 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 535 Constant *V = LHSVals[i].first; 536 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(), 537 V, CmpConst); 538 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 539 Result.push_back(std::make_pair(KC, LHSVals[i].second)); 540 } 541 542 return !Result.empty(); 543 } 544 } 545 } 546 547 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 548 // Handle select instructions where at least one operand is a known constant 549 // and we can figure out the condition value for any predecessor block. 550 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference); 551 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference); 552 PredValueInfoTy Conds; 553 if ((TrueVal || FalseVal) && 554 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds, 555 WantInteger)) { 556 for (unsigned i = 0, e = Conds.size(); i != e; ++i) { 557 Constant *Cond = Conds[i].first; 558 559 // Figure out what value to use for the condition. 560 bool KnownCond; 561 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) { 562 // A known boolean. 563 KnownCond = CI->isOne(); 564 } else { 565 assert(isa<UndefValue>(Cond) && "Unexpected condition value"); 566 // Either operand will do, so be sure to pick the one that's a known 567 // constant. 568 // FIXME: Do this more cleverly if both values are known constants? 569 KnownCond = (TrueVal != 0); 570 } 571 572 // See if the select has a known constant value for this predecessor. 573 if (Constant *Val = KnownCond ? TrueVal : FalseVal) 574 Result.push_back(std::make_pair(Val, Conds[i].second)); 575 } 576 577 return !Result.empty(); 578 } 579 } 580 581 // If all else fails, see if LVI can figure out a constant value for us. 582 Constant *CI = LVI->getConstant(V, BB); 583 if (Constant *KC = getKnownConstant(CI, Preference)) { 584 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 585 Result.push_back(std::make_pair(KC, *PI)); 586 } 587 588 return !Result.empty(); 589} 590 591 592 593/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 594/// in an undefined jump, decide which block is best to revector to. 595/// 596/// Since we can pick an arbitrary destination, we pick the successor with the 597/// fewest predecessors. This should reduce the in-degree of the others. 598/// 599static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 600 TerminatorInst *BBTerm = BB->getTerminator(); 601 unsigned MinSucc = 0; 602 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 603 // Compute the successor with the minimum number of predecessors. 604 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 605 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 606 TestBB = BBTerm->getSuccessor(i); 607 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 608 if (NumPreds < MinNumPreds) { 609 MinSucc = i; 610 MinNumPreds = NumPreds; 611 } 612 } 613 614 return MinSucc; 615} 616 617static bool hasAddressTakenAndUsed(BasicBlock *BB) { 618 if (!BB->hasAddressTaken()) return false; 619 620 // If the block has its address taken, it may be a tree of dead constants 621 // hanging off of it. These shouldn't keep the block alive. 622 BlockAddress *BA = BlockAddress::get(BB); 623 BA->removeDeadConstantUsers(); 624 return !BA->use_empty(); 625} 626 627/// ProcessBlock - If there are any predecessors whose control can be threaded 628/// through to a successor, transform them now. 629bool JumpThreading::ProcessBlock(BasicBlock *BB) { 630 // If the block is trivially dead, just return and let the caller nuke it. 631 // This simplifies other transformations. 632 if (pred_begin(BB) == pred_end(BB) && 633 BB != &BB->getParent()->getEntryBlock()) 634 return false; 635 636 // If this block has a single predecessor, and if that pred has a single 637 // successor, merge the blocks. This encourages recursive jump threading 638 // because now the condition in this block can be threaded through 639 // predecessors of our predecessor block. 640 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 641 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 642 SinglePred != BB && !hasAddressTakenAndUsed(BB)) { 643 // If SinglePred was a loop header, BB becomes one. 644 if (LoopHeaders.erase(SinglePred)) 645 LoopHeaders.insert(BB); 646 647 // Remember if SinglePred was the entry block of the function. If so, we 648 // will need to move BB back to the entry position. 649 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 650 LVI->eraseBlock(SinglePred); 651 MergeBasicBlockIntoOnlyPred(BB); 652 653 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 654 BB->moveBefore(&BB->getParent()->getEntryBlock()); 655 return true; 656 } 657 } 658 659 // What kind of constant we're looking for. 660 ConstantPreference Preference = WantInteger; 661 662 // Look to see if the terminator is a conditional branch, switch or indirect 663 // branch, if not we can't thread it. 664 Value *Condition; 665 Instruction *Terminator = BB->getTerminator(); 666 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { 667 // Can't thread an unconditional jump. 668 if (BI->isUnconditional()) return false; 669 Condition = BI->getCondition(); 670 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { 671 Condition = SI->getCondition(); 672 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { 673 Condition = IB->getAddress()->stripPointerCasts(); 674 Preference = WantBlockAddress; 675 } else { 676 return false; // Must be an invoke. 677 } 678 679 // Run constant folding to see if we can reduce the condition to a simple 680 // constant. 681 if (Instruction *I = dyn_cast<Instruction>(Condition)) { 682 Value *SimpleVal = ConstantFoldInstruction(I, TD, TLI); 683 if (SimpleVal) { 684 I->replaceAllUsesWith(SimpleVal); 685 I->eraseFromParent(); 686 Condition = SimpleVal; 687 } 688 } 689 690 // If the terminator is branching on an undef, we can pick any of the 691 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 692 if (isa<UndefValue>(Condition)) { 693 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 694 695 // Fold the branch/switch. 696 TerminatorInst *BBTerm = BB->getTerminator(); 697 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 698 if (i == BestSucc) continue; 699 BBTerm->getSuccessor(i)->removePredecessor(BB, true); 700 } 701 702 DEBUG(dbgs() << " In block '" << BB->getName() 703 << "' folding undef terminator: " << *BBTerm << '\n'); 704 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 705 BBTerm->eraseFromParent(); 706 return true; 707 } 708 709 // If the terminator of this block is branching on a constant, simplify the 710 // terminator to an unconditional branch. This can occur due to threading in 711 // other blocks. 712 if (getKnownConstant(Condition, Preference)) { 713 DEBUG(dbgs() << " In block '" << BB->getName() 714 << "' folding terminator: " << *BB->getTerminator() << '\n'); 715 ++NumFolds; 716 ConstantFoldTerminator(BB, true); 717 return true; 718 } 719 720 Instruction *CondInst = dyn_cast<Instruction>(Condition); 721 722 // All the rest of our checks depend on the condition being an instruction. 723 if (CondInst == 0) { 724 // FIXME: Unify this with code below. 725 if (ProcessThreadableEdges(Condition, BB, Preference)) 726 return true; 727 return false; 728 } 729 730 731 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 732 // For a comparison where the LHS is outside this block, it's possible 733 // that we've branched on it before. Used LVI to see if we can simplify 734 // the branch based on that. 735 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 736 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 737 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 738 if (CondBr && CondConst && CondBr->isConditional() && PI != PE && 739 (!isa<Instruction>(CondCmp->getOperand(0)) || 740 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { 741 // For predecessor edge, determine if the comparison is true or false 742 // on that edge. If they're all true or all false, we can simplify the 743 // branch. 744 // FIXME: We could handle mixed true/false by duplicating code. 745 LazyValueInfo::Tristate Baseline = 746 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0), 747 CondConst, *PI, BB); 748 if (Baseline != LazyValueInfo::Unknown) { 749 // Check that all remaining incoming values match the first one. 750 while (++PI != PE) { 751 LazyValueInfo::Tristate Ret = 752 LVI->getPredicateOnEdge(CondCmp->getPredicate(), 753 CondCmp->getOperand(0), CondConst, *PI, BB); 754 if (Ret != Baseline) break; 755 } 756 757 // If we terminated early, then one of the values didn't match. 758 if (PI == PE) { 759 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0; 760 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1; 761 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); 762 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); 763 CondBr->eraseFromParent(); 764 return true; 765 } 766 } 767 } 768 } 769 770 // Check for some cases that are worth simplifying. Right now we want to look 771 // for loads that are used by a switch or by the condition for the branch. If 772 // we see one, check to see if it's partially redundant. If so, insert a PHI 773 // which can then be used to thread the values. 774 // 775 Value *SimplifyValue = CondInst; 776 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 777 if (isa<Constant>(CondCmp->getOperand(1))) 778 SimplifyValue = CondCmp->getOperand(0); 779 780 // TODO: There are other places where load PRE would be profitable, such as 781 // more complex comparisons. 782 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 783 if (SimplifyPartiallyRedundantLoad(LI)) 784 return true; 785 786 787 // Handle a variety of cases where we are branching on something derived from 788 // a PHI node in the current block. If we can prove that any predecessors 789 // compute a predictable value based on a PHI node, thread those predecessors. 790 // 791 if (ProcessThreadableEdges(CondInst, BB, Preference)) 792 return true; 793 794 // If this is an otherwise-unfoldable branch on a phi node in the current 795 // block, see if we can simplify. 796 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 797 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 798 return ProcessBranchOnPHI(PN); 799 800 801 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 802 if (CondInst->getOpcode() == Instruction::Xor && 803 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 804 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 805 806 807 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 808 // "(X == 4)", thread through this block. 809 810 return false; 811} 812 813 814/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 815/// load instruction, eliminate it by replacing it with a PHI node. This is an 816/// important optimization that encourages jump threading, and needs to be run 817/// interlaced with other jump threading tasks. 818bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 819 // Don't hack volatile/atomic loads. 820 if (!LI->isSimple()) return false; 821 822 // If the load is defined in a block with exactly one predecessor, it can't be 823 // partially redundant. 824 BasicBlock *LoadBB = LI->getParent(); 825 if (LoadBB->getSinglePredecessor()) 826 return false; 827 828 Value *LoadedPtr = LI->getOperand(0); 829 830 // If the loaded operand is defined in the LoadBB, it can't be available. 831 // TODO: Could do simple PHI translation, that would be fun :) 832 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 833 if (PtrOp->getParent() == LoadBB) 834 return false; 835 836 // Scan a few instructions up from the load, to see if it is obviously live at 837 // the entry to its block. 838 BasicBlock::iterator BBIt = LI; 839 840 if (Value *AvailableVal = 841 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 842 // If the value if the load is locally available within the block, just use 843 // it. This frequently occurs for reg2mem'd allocas. 844 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 845 846 // If the returned value is the load itself, replace with an undef. This can 847 // only happen in dead loops. 848 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 849 LI->replaceAllUsesWith(AvailableVal); 850 LI->eraseFromParent(); 851 return true; 852 } 853 854 // Otherwise, if we scanned the whole block and got to the top of the block, 855 // we know the block is locally transparent to the load. If not, something 856 // might clobber its value. 857 if (BBIt != LoadBB->begin()) 858 return false; 859 860 861 SmallPtrSet<BasicBlock*, 8> PredsScanned; 862 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 863 AvailablePredsTy AvailablePreds; 864 BasicBlock *OneUnavailablePred = 0; 865 866 // If we got here, the loaded value is transparent through to the start of the 867 // block. Check to see if it is available in any of the predecessor blocks. 868 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 869 PI != PE; ++PI) { 870 BasicBlock *PredBB = *PI; 871 872 // If we already scanned this predecessor, skip it. 873 if (!PredsScanned.insert(PredBB)) 874 continue; 875 876 // Scan the predecessor to see if the value is available in the pred. 877 BBIt = PredBB->end(); 878 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 879 if (!PredAvailable) { 880 OneUnavailablePred = PredBB; 881 continue; 882 } 883 884 // If so, this load is partially redundant. Remember this info so that we 885 // can create a PHI node. 886 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 887 } 888 889 // If the loaded value isn't available in any predecessor, it isn't partially 890 // redundant. 891 if (AvailablePreds.empty()) return false; 892 893 // Okay, the loaded value is available in at least one (and maybe all!) 894 // predecessors. If the value is unavailable in more than one unique 895 // predecessor, we want to insert a merge block for those common predecessors. 896 // This ensures that we only have to insert one reload, thus not increasing 897 // code size. 898 BasicBlock *UnavailablePred = 0; 899 900 // If there is exactly one predecessor where the value is unavailable, the 901 // already computed 'OneUnavailablePred' block is it. If it ends in an 902 // unconditional branch, we know that it isn't a critical edge. 903 if (PredsScanned.size() == AvailablePreds.size()+1 && 904 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 905 UnavailablePred = OneUnavailablePred; 906 } else if (PredsScanned.size() != AvailablePreds.size()) { 907 // Otherwise, we had multiple unavailable predecessors or we had a critical 908 // edge from the one. 909 SmallVector<BasicBlock*, 8> PredsToSplit; 910 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 911 912 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 913 AvailablePredSet.insert(AvailablePreds[i].first); 914 915 // Add all the unavailable predecessors to the PredsToSplit list. 916 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 917 PI != PE; ++PI) { 918 BasicBlock *P = *PI; 919 // If the predecessor is an indirect goto, we can't split the edge. 920 if (isa<IndirectBrInst>(P->getTerminator())) 921 return false; 922 923 if (!AvailablePredSet.count(P)) 924 PredsToSplit.push_back(P); 925 } 926 927 // Split them out to their own block. 928 UnavailablePred = 929 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this); 930 } 931 932 // If the value isn't available in all predecessors, then there will be 933 // exactly one where it isn't available. Insert a load on that edge and add 934 // it to the AvailablePreds list. 935 if (UnavailablePred) { 936 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 937 "Can't handle critical edge here!"); 938 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 939 LI->getAlignment(), 940 UnavailablePred->getTerminator()); 941 NewVal->setDebugLoc(LI->getDebugLoc()); 942 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 943 } 944 945 // Now we know that each predecessor of this block has a value in 946 // AvailablePreds, sort them for efficient access as we're walking the preds. 947 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 948 949 // Create a PHI node at the start of the block for the PRE'd load value. 950 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB); 951 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "", 952 LoadBB->begin()); 953 PN->takeName(LI); 954 PN->setDebugLoc(LI->getDebugLoc()); 955 956 // Insert new entries into the PHI for each predecessor. A single block may 957 // have multiple entries here. 958 for (pred_iterator PI = PB; PI != PE; ++PI) { 959 BasicBlock *P = *PI; 960 AvailablePredsTy::iterator I = 961 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 962 std::make_pair(P, (Value*)0)); 963 964 assert(I != AvailablePreds.end() && I->first == P && 965 "Didn't find entry for predecessor!"); 966 967 PN->addIncoming(I->second, I->first); 968 } 969 970 //cerr << "PRE: " << *LI << *PN << "\n"; 971 972 LI->replaceAllUsesWith(PN); 973 LI->eraseFromParent(); 974 975 return true; 976} 977 978/// FindMostPopularDest - The specified list contains multiple possible 979/// threadable destinations. Pick the one that occurs the most frequently in 980/// the list. 981static BasicBlock * 982FindMostPopularDest(BasicBlock *BB, 983 const SmallVectorImpl<std::pair<BasicBlock*, 984 BasicBlock*> > &PredToDestList) { 985 assert(!PredToDestList.empty()); 986 987 // Determine popularity. If there are multiple possible destinations, we 988 // explicitly choose to ignore 'undef' destinations. We prefer to thread 989 // blocks with known and real destinations to threading undef. We'll handle 990 // them later if interesting. 991 DenseMap<BasicBlock*, unsigned> DestPopularity; 992 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 993 if (PredToDestList[i].second) 994 DestPopularity[PredToDestList[i].second]++; 995 996 // Find the most popular dest. 997 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 998 BasicBlock *MostPopularDest = DPI->first; 999 unsigned Popularity = DPI->second; 1000 SmallVector<BasicBlock*, 4> SamePopularity; 1001 1002 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 1003 // If the popularity of this entry isn't higher than the popularity we've 1004 // seen so far, ignore it. 1005 if (DPI->second < Popularity) 1006 ; // ignore. 1007 else if (DPI->second == Popularity) { 1008 // If it is the same as what we've seen so far, keep track of it. 1009 SamePopularity.push_back(DPI->first); 1010 } else { 1011 // If it is more popular, remember it. 1012 SamePopularity.clear(); 1013 MostPopularDest = DPI->first; 1014 Popularity = DPI->second; 1015 } 1016 } 1017 1018 // Okay, now we know the most popular destination. If there is more than one 1019 // destination, we need to determine one. This is arbitrary, but we need 1020 // to make a deterministic decision. Pick the first one that appears in the 1021 // successor list. 1022 if (!SamePopularity.empty()) { 1023 SamePopularity.push_back(MostPopularDest); 1024 TerminatorInst *TI = BB->getTerminator(); 1025 for (unsigned i = 0; ; ++i) { 1026 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1027 1028 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1029 TI->getSuccessor(i)) == SamePopularity.end()) 1030 continue; 1031 1032 MostPopularDest = TI->getSuccessor(i); 1033 break; 1034 } 1035 } 1036 1037 // Okay, we have finally picked the most popular destination. 1038 return MostPopularDest; 1039} 1040 1041bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 1042 ConstantPreference Preference) { 1043 // If threading this would thread across a loop header, don't even try to 1044 // thread the edge. 1045 if (LoopHeaders.count(BB)) 1046 return false; 1047 1048 PredValueInfoTy PredValues; 1049 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference)) 1050 return false; 1051 1052 assert(!PredValues.empty() && 1053 "ComputeValueKnownInPredecessors returned true with no values"); 1054 1055 DEBUG(dbgs() << "IN BB: " << *BB; 1056 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1057 dbgs() << " BB '" << BB->getName() << "': FOUND condition = " 1058 << *PredValues[i].first 1059 << " for pred '" << PredValues[i].second->getName() << "'.\n"; 1060 }); 1061 1062 // Decide what we want to thread through. Convert our list of known values to 1063 // a list of known destinations for each pred. This also discards duplicate 1064 // predecessors and keeps track of the undefined inputs (which are represented 1065 // as a null dest in the PredToDestList). 1066 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1067 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1068 1069 BasicBlock *OnlyDest = 0; 1070 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1071 1072 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1073 BasicBlock *Pred = PredValues[i].second; 1074 if (!SeenPreds.insert(Pred)) 1075 continue; // Duplicate predecessor entry. 1076 1077 // If the predecessor ends with an indirect goto, we can't change its 1078 // destination. 1079 if (isa<IndirectBrInst>(Pred->getTerminator())) 1080 continue; 1081 1082 Constant *Val = PredValues[i].first; 1083 1084 BasicBlock *DestBB; 1085 if (isa<UndefValue>(Val)) 1086 DestBB = 0; 1087 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1088 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); 1089 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 1090 unsigned ValCase = SI->findCaseValue(cast<ConstantInt>(Val)); 1091 DestBB = SI->getSuccessor(SI->resolveSuccessorIndex(ValCase)); 1092 } else { 1093 assert(isa<IndirectBrInst>(BB->getTerminator()) 1094 && "Unexpected terminator"); 1095 DestBB = cast<BlockAddress>(Val)->getBasicBlock(); 1096 } 1097 1098 // If we have exactly one destination, remember it for efficiency below. 1099 if (PredToDestList.empty()) 1100 OnlyDest = DestBB; 1101 else if (OnlyDest != DestBB) 1102 OnlyDest = MultipleDestSentinel; 1103 1104 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1105 } 1106 1107 // If all edges were unthreadable, we fail. 1108 if (PredToDestList.empty()) 1109 return false; 1110 1111 // Determine which is the most common successor. If we have many inputs and 1112 // this block is a switch, we want to start by threading the batch that goes 1113 // to the most popular destination first. If we only know about one 1114 // threadable destination (the common case) we can avoid this. 1115 BasicBlock *MostPopularDest = OnlyDest; 1116 1117 if (MostPopularDest == MultipleDestSentinel) 1118 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1119 1120 // Now that we know what the most popular destination is, factor all 1121 // predecessors that will jump to it into a single predecessor. 1122 SmallVector<BasicBlock*, 16> PredsToFactor; 1123 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1124 if (PredToDestList[i].second == MostPopularDest) { 1125 BasicBlock *Pred = PredToDestList[i].first; 1126 1127 // This predecessor may be a switch or something else that has multiple 1128 // edges to the block. Factor each of these edges by listing them 1129 // according to # occurrences in PredsToFactor. 1130 TerminatorInst *PredTI = Pred->getTerminator(); 1131 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1132 if (PredTI->getSuccessor(i) == BB) 1133 PredsToFactor.push_back(Pred); 1134 } 1135 1136 // If the threadable edges are branching on an undefined value, we get to pick 1137 // the destination that these predecessors should get to. 1138 if (MostPopularDest == 0) 1139 MostPopularDest = BB->getTerminator()-> 1140 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1141 1142 // Ok, try to thread it! 1143 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1144} 1145 1146/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1147/// a PHI node in the current block. See if there are any simplifications we 1148/// can do based on inputs to the phi node. 1149/// 1150bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1151 BasicBlock *BB = PN->getParent(); 1152 1153 // TODO: We could make use of this to do it once for blocks with common PHI 1154 // values. 1155 SmallVector<BasicBlock*, 1> PredBBs; 1156 PredBBs.resize(1); 1157 1158 // If any of the predecessor blocks end in an unconditional branch, we can 1159 // *duplicate* the conditional branch into that block in order to further 1160 // encourage jump threading and to eliminate cases where we have branch on a 1161 // phi of an icmp (branch on icmp is much better). 1162 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1163 BasicBlock *PredBB = PN->getIncomingBlock(i); 1164 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1165 if (PredBr->isUnconditional()) { 1166 PredBBs[0] = PredBB; 1167 // Try to duplicate BB into PredBB. 1168 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1169 return true; 1170 } 1171 } 1172 1173 return false; 1174} 1175 1176/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1177/// a xor instruction in the current block. See if there are any 1178/// simplifications we can do based on inputs to the xor. 1179/// 1180bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1181 BasicBlock *BB = BO->getParent(); 1182 1183 // If either the LHS or RHS of the xor is a constant, don't do this 1184 // optimization. 1185 if (isa<ConstantInt>(BO->getOperand(0)) || 1186 isa<ConstantInt>(BO->getOperand(1))) 1187 return false; 1188 1189 // If the first instruction in BB isn't a phi, we won't be able to infer 1190 // anything special about any particular predecessor. 1191 if (!isa<PHINode>(BB->front())) 1192 return false; 1193 1194 // If we have a xor as the branch input to this block, and we know that the 1195 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1196 // the condition into the predecessor and fix that value to true, saving some 1197 // logical ops on that path and encouraging other paths to simplify. 1198 // 1199 // This copies something like this: 1200 // 1201 // BB: 1202 // %X = phi i1 [1], [%X'] 1203 // %Y = icmp eq i32 %A, %B 1204 // %Z = xor i1 %X, %Y 1205 // br i1 %Z, ... 1206 // 1207 // Into: 1208 // BB': 1209 // %Y = icmp ne i32 %A, %B 1210 // br i1 %Z, ... 1211 1212 PredValueInfoTy XorOpValues; 1213 bool isLHS = true; 1214 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, 1215 WantInteger)) { 1216 assert(XorOpValues.empty()); 1217 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, 1218 WantInteger)) 1219 return false; 1220 isLHS = false; 1221 } 1222 1223 assert(!XorOpValues.empty() && 1224 "ComputeValueKnownInPredecessors returned true with no values"); 1225 1226 // Scan the information to see which is most popular: true or false. The 1227 // predecessors can be of the set true, false, or undef. 1228 unsigned NumTrue = 0, NumFalse = 0; 1229 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1230 if (isa<UndefValue>(XorOpValues[i].first)) 1231 // Ignore undefs for the count. 1232 continue; 1233 if (cast<ConstantInt>(XorOpValues[i].first)->isZero()) 1234 ++NumFalse; 1235 else 1236 ++NumTrue; 1237 } 1238 1239 // Determine which value to split on, true, false, or undef if neither. 1240 ConstantInt *SplitVal = 0; 1241 if (NumTrue > NumFalse) 1242 SplitVal = ConstantInt::getTrue(BB->getContext()); 1243 else if (NumTrue != 0 || NumFalse != 0) 1244 SplitVal = ConstantInt::getFalse(BB->getContext()); 1245 1246 // Collect all of the blocks that this can be folded into so that we can 1247 // factor this once and clone it once. 1248 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1249 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1250 if (XorOpValues[i].first != SplitVal && 1251 !isa<UndefValue>(XorOpValues[i].first)) 1252 continue; 1253 1254 BlocksToFoldInto.push_back(XorOpValues[i].second); 1255 } 1256 1257 // If we inferred a value for all of the predecessors, then duplication won't 1258 // help us. However, we can just replace the LHS or RHS with the constant. 1259 if (BlocksToFoldInto.size() == 1260 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1261 if (SplitVal == 0) { 1262 // If all preds provide undef, just nuke the xor, because it is undef too. 1263 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1264 BO->eraseFromParent(); 1265 } else if (SplitVal->isZero()) { 1266 // If all preds provide 0, replace the xor with the other input. 1267 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1268 BO->eraseFromParent(); 1269 } else { 1270 // If all preds provide 1, set the computed value to 1. 1271 BO->setOperand(!isLHS, SplitVal); 1272 } 1273 1274 return true; 1275 } 1276 1277 // Try to duplicate BB into PredBB. 1278 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1279} 1280 1281 1282/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1283/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1284/// NewPred using the entries from OldPred (suitably mapped). 1285static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1286 BasicBlock *OldPred, 1287 BasicBlock *NewPred, 1288 DenseMap<Instruction*, Value*> &ValueMap) { 1289 for (BasicBlock::iterator PNI = PHIBB->begin(); 1290 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1291 // Ok, we have a PHI node. Figure out what the incoming value was for the 1292 // DestBlock. 1293 Value *IV = PN->getIncomingValueForBlock(OldPred); 1294 1295 // Remap the value if necessary. 1296 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1297 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1298 if (I != ValueMap.end()) 1299 IV = I->second; 1300 } 1301 1302 PN->addIncoming(IV, NewPred); 1303 } 1304} 1305 1306/// ThreadEdge - We have decided that it is safe and profitable to factor the 1307/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1308/// across BB. Transform the IR to reflect this change. 1309bool JumpThreading::ThreadEdge(BasicBlock *BB, 1310 const SmallVectorImpl<BasicBlock*> &PredBBs, 1311 BasicBlock *SuccBB) { 1312 // If threading to the same block as we come from, we would infinite loop. 1313 if (SuccBB == BB) { 1314 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1315 << "' - would thread to self!\n"); 1316 return false; 1317 } 1318 1319 // If threading this would thread across a loop header, don't thread the edge. 1320 // See the comments above FindLoopHeaders for justifications and caveats. 1321 if (LoopHeaders.count(BB)) { 1322 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1323 << "' to dest BB '" << SuccBB->getName() 1324 << "' - it might create an irreducible loop!\n"); 1325 return false; 1326 } 1327 1328 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1329 if (JumpThreadCost > Threshold) { 1330 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1331 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1332 return false; 1333 } 1334 1335 // And finally, do it! Start by factoring the predecessors is needed. 1336 BasicBlock *PredBB; 1337 if (PredBBs.size() == 1) 1338 PredBB = PredBBs[0]; 1339 else { 1340 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1341 << " common predecessors.\n"); 1342 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); 1343 } 1344 1345 // And finally, do it! 1346 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1347 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1348 << ", across block:\n " 1349 << *BB << "\n"); 1350 1351 LVI->threadEdge(PredBB, BB, SuccBB); 1352 1353 // We are going to have to map operands from the original BB block to the new 1354 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1355 // account for entry from PredBB. 1356 DenseMap<Instruction*, Value*> ValueMapping; 1357 1358 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1359 BB->getName()+".thread", 1360 BB->getParent(), BB); 1361 NewBB->moveAfter(PredBB); 1362 1363 BasicBlock::iterator BI = BB->begin(); 1364 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1365 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1366 1367 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1368 // mapping and using it to remap operands in the cloned instructions. 1369 for (; !isa<TerminatorInst>(BI); ++BI) { 1370 Instruction *New = BI->clone(); 1371 New->setName(BI->getName()); 1372 NewBB->getInstList().push_back(New); 1373 ValueMapping[BI] = New; 1374 1375 // Remap operands to patch up intra-block references. 1376 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1377 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1378 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1379 if (I != ValueMapping.end()) 1380 New->setOperand(i, I->second); 1381 } 1382 } 1383 1384 // We didn't copy the terminator from BB over to NewBB, because there is now 1385 // an unconditional jump to SuccBB. Insert the unconditional jump. 1386 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB); 1387 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc()); 1388 1389 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1390 // PHI nodes for NewBB now. 1391 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1392 1393 // If there were values defined in BB that are used outside the block, then we 1394 // now have to update all uses of the value to use either the original value, 1395 // the cloned value, or some PHI derived value. This can require arbitrary 1396 // PHI insertion, of which we are prepared to do, clean these up now. 1397 SSAUpdater SSAUpdate; 1398 SmallVector<Use*, 16> UsesToRename; 1399 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1400 // Scan all uses of this instruction to see if it is used outside of its 1401 // block, and if so, record them in UsesToRename. 1402 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1403 ++UI) { 1404 Instruction *User = cast<Instruction>(*UI); 1405 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1406 if (UserPN->getIncomingBlock(UI) == BB) 1407 continue; 1408 } else if (User->getParent() == BB) 1409 continue; 1410 1411 UsesToRename.push_back(&UI.getUse()); 1412 } 1413 1414 // If there are no uses outside the block, we're done with this instruction. 1415 if (UsesToRename.empty()) 1416 continue; 1417 1418 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1419 1420 // We found a use of I outside of BB. Rename all uses of I that are outside 1421 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1422 // with the two values we know. 1423 SSAUpdate.Initialize(I->getType(), I->getName()); 1424 SSAUpdate.AddAvailableValue(BB, I); 1425 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1426 1427 while (!UsesToRename.empty()) 1428 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1429 DEBUG(dbgs() << "\n"); 1430 } 1431 1432 1433 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1434 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1435 // us to simplify any PHI nodes in BB. 1436 TerminatorInst *PredTerm = PredBB->getTerminator(); 1437 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1438 if (PredTerm->getSuccessor(i) == BB) { 1439 BB->removePredecessor(PredBB, true); 1440 PredTerm->setSuccessor(i, NewBB); 1441 } 1442 1443 // At this point, the IR is fully up to date and consistent. Do a quick scan 1444 // over the new instructions and zap any that are constants or dead. This 1445 // frequently happens because of phi translation. 1446 SimplifyInstructionsInBlock(NewBB, TD); 1447 1448 // Threaded an edge! 1449 ++NumThreads; 1450 return true; 1451} 1452 1453/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1454/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1455/// If we can duplicate the contents of BB up into PredBB do so now, this 1456/// improves the odds that the branch will be on an analyzable instruction like 1457/// a compare. 1458bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1459 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1460 assert(!PredBBs.empty() && "Can't handle an empty set"); 1461 1462 // If BB is a loop header, then duplicating this block outside the loop would 1463 // cause us to transform this into an irreducible loop, don't do this. 1464 // See the comments above FindLoopHeaders for justifications and caveats. 1465 if (LoopHeaders.count(BB)) { 1466 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1467 << "' into predecessor block '" << PredBBs[0]->getName() 1468 << "' - it might create an irreducible loop!\n"); 1469 return false; 1470 } 1471 1472 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1473 if (DuplicationCost > Threshold) { 1474 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1475 << "' - Cost is too high: " << DuplicationCost << "\n"); 1476 return false; 1477 } 1478 1479 // And finally, do it! Start by factoring the predecessors is needed. 1480 BasicBlock *PredBB; 1481 if (PredBBs.size() == 1) 1482 PredBB = PredBBs[0]; 1483 else { 1484 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1485 << " common predecessors.\n"); 1486 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); 1487 } 1488 1489 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1490 // of PredBB. 1491 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1492 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1493 << DuplicationCost << " block is:" << *BB << "\n"); 1494 1495 // Unless PredBB ends with an unconditional branch, split the edge so that we 1496 // can just clone the bits from BB into the end of the new PredBB. 1497 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1498 1499 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1500 PredBB = SplitEdge(PredBB, BB, this); 1501 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1502 } 1503 1504 // We are going to have to map operands from the original BB block into the 1505 // PredBB block. Evaluate PHI nodes in BB. 1506 DenseMap<Instruction*, Value*> ValueMapping; 1507 1508 BasicBlock::iterator BI = BB->begin(); 1509 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1510 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1511 1512 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1513 // mapping and using it to remap operands in the cloned instructions. 1514 for (; BI != BB->end(); ++BI) { 1515 Instruction *New = BI->clone(); 1516 1517 // Remap operands to patch up intra-block references. 1518 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1519 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1520 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1521 if (I != ValueMapping.end()) 1522 New->setOperand(i, I->second); 1523 } 1524 1525 // If this instruction can be simplified after the operands are updated, 1526 // just use the simplified value instead. This frequently happens due to 1527 // phi translation. 1528 if (Value *IV = SimplifyInstruction(New, TD)) { 1529 delete New; 1530 ValueMapping[BI] = IV; 1531 } else { 1532 // Otherwise, insert the new instruction into the block. 1533 New->setName(BI->getName()); 1534 PredBB->getInstList().insert(OldPredBranch, New); 1535 ValueMapping[BI] = New; 1536 } 1537 } 1538 1539 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1540 // add entries to the PHI nodes for branch from PredBB now. 1541 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1542 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1543 ValueMapping); 1544 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1545 ValueMapping); 1546 1547 // If there were values defined in BB that are used outside the block, then we 1548 // now have to update all uses of the value to use either the original value, 1549 // the cloned value, or some PHI derived value. This can require arbitrary 1550 // PHI insertion, of which we are prepared to do, clean these up now. 1551 SSAUpdater SSAUpdate; 1552 SmallVector<Use*, 16> UsesToRename; 1553 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1554 // Scan all uses of this instruction to see if it is used outside of its 1555 // block, and if so, record them in UsesToRename. 1556 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1557 ++UI) { 1558 Instruction *User = cast<Instruction>(*UI); 1559 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1560 if (UserPN->getIncomingBlock(UI) == BB) 1561 continue; 1562 } else if (User->getParent() == BB) 1563 continue; 1564 1565 UsesToRename.push_back(&UI.getUse()); 1566 } 1567 1568 // If there are no uses outside the block, we're done with this instruction. 1569 if (UsesToRename.empty()) 1570 continue; 1571 1572 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1573 1574 // We found a use of I outside of BB. Rename all uses of I that are outside 1575 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1576 // with the two values we know. 1577 SSAUpdate.Initialize(I->getType(), I->getName()); 1578 SSAUpdate.AddAvailableValue(BB, I); 1579 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1580 1581 while (!UsesToRename.empty()) 1582 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1583 DEBUG(dbgs() << "\n"); 1584 } 1585 1586 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1587 // that we nuked. 1588 BB->removePredecessor(PredBB, true); 1589 1590 // Remove the unconditional branch at the end of the PredBB block. 1591 OldPredBranch->eraseFromParent(); 1592 1593 ++NumDupes; 1594 return true; 1595} 1596 1597 1598