JumpThreading.cpp revision 2ab36d350293c77fc8941ce1023e4899df7e3a82
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 /// This pass performs 'jump threading', which looks at blocks that have 49 /// multiple predecessors and multiple successors. If one or more of the 50 /// predecessors of the block can be proven to always jump to one of the 51 /// successors, we forward the edge from the predecessor to the successor by 52 /// duplicating the contents of this block. 53 /// 54 /// An example of when this can occur is code like this: 55 /// 56 /// if () { ... 57 /// X = 4; 58 /// } 59 /// if (X < 3) { 60 /// 61 /// In this case, the unconditional branch at the end of the first if can be 62 /// revectored to the false side of the second if. 63 /// 64 class JumpThreading : public FunctionPass { 65 TargetData *TD; 66 LazyValueInfo *LVI; 67#ifdef NDEBUG 68 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 69#else 70 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 71#endif 72 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet; 73 74 // RAII helper for updating the recursion stack. 75 struct RecursionSetRemover { 76 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet; 77 std::pair<Value*, BasicBlock*> ThePair; 78 79 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S, 80 std::pair<Value*, BasicBlock*> P) 81 : TheSet(S), ThePair(P) { } 82 83 ~RecursionSetRemover() { 84 TheSet.erase(ThePair); 85 } 86 }; 87 public: 88 static char ID; // Pass identification 89 JumpThreading() : FunctionPass(ID) {} 90 91 bool runOnFunction(Function &F); 92 93 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 94 AU.addRequired<LazyValueInfo>(); 95 AU.addPreserved<LazyValueInfo>(); 96 } 97 98 void FindLoopHeaders(Function &F); 99 bool ProcessBlock(BasicBlock *BB); 100 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 101 BasicBlock *SuccBB); 102 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 103 const SmallVectorImpl<BasicBlock *> &PredBBs); 104 105 typedef SmallVectorImpl<std::pair<ConstantInt*, 106 BasicBlock*> > PredValueInfo; 107 108 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 109 PredValueInfo &Result); 110 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB); 111 112 113 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 114 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 115 116 bool ProcessBranchOnPHI(PHINode *PN); 117 bool ProcessBranchOnXOR(BinaryOperator *BO); 118 119 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 120 }; 121} 122 123char JumpThreading::ID = 0; 124INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", 125 "Jump Threading", false, false) 126INITIALIZE_PASS_DEPENDENCY(LazyValueInfo) 127INITIALIZE_PASS_END(JumpThreading, "jump-threading", 128 "Jump Threading", false, false) 129 130// Public interface to the Jump Threading pass 131FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 132 133/// runOnFunction - Top level algorithm. 134/// 135bool JumpThreading::runOnFunction(Function &F) { 136 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 137 TD = getAnalysisIfAvailable<TargetData>(); 138 LVI = &getAnalysis<LazyValueInfo>(); 139 140 FindLoopHeaders(F); 141 142 bool Changed, EverChanged = false; 143 do { 144 Changed = false; 145 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 146 BasicBlock *BB = I; 147 // Thread all of the branches we can over this block. 148 while (ProcessBlock(BB)) 149 Changed = true; 150 151 ++I; 152 153 // If the block is trivially dead, zap it. This eliminates the successor 154 // edges which simplifies the CFG. 155 if (pred_begin(BB) == pred_end(BB) && 156 BB != &BB->getParent()->getEntryBlock()) { 157 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 158 << "' with terminator: " << *BB->getTerminator() << '\n'); 159 LoopHeaders.erase(BB); 160 LVI->eraseBlock(BB); 161 DeleteDeadBlock(BB); 162 Changed = true; 163 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 164 // Can't thread an unconditional jump, but if the block is "almost 165 // empty", we can replace uses of it with uses of the successor and make 166 // this dead. 167 if (BI->isUnconditional() && 168 BB != &BB->getParent()->getEntryBlock()) { 169 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 170 // Ignore dbg intrinsics. 171 while (isa<DbgInfoIntrinsic>(BBI)) 172 ++BBI; 173 // If the terminator is the only non-phi instruction, try to nuke it. 174 if (BBI->isTerminator()) { 175 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 176 // block, we have to make sure it isn't in the LoopHeaders set. We 177 // reinsert afterward if needed. 178 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 179 BasicBlock *Succ = BI->getSuccessor(0); 180 181 // FIXME: It is always conservatively correct to drop the info 182 // for a block even if it doesn't get erased. This isn't totally 183 // awesome, but it allows us to use AssertingVH to prevent nasty 184 // dangling pointer issues within LazyValueInfo. 185 LVI->eraseBlock(BB); 186 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 187 Changed = true; 188 // If we deleted BB and BB was the header of a loop, then the 189 // successor is now the header of the loop. 190 BB = Succ; 191 } 192 193 if (ErasedFromLoopHeaders) 194 LoopHeaders.insert(BB); 195 } 196 } 197 } 198 } 199 EverChanged |= Changed; 200 } while (Changed); 201 202 LoopHeaders.clear(); 203 return EverChanged; 204} 205 206/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 207/// thread across it. 208static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 209 /// Ignore PHI nodes, these will be flattened when duplication happens. 210 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 211 212 // FIXME: THREADING will delete values that are just used to compute the 213 // branch, so they shouldn't count against the duplication cost. 214 215 216 // Sum up the cost of each instruction until we get to the terminator. Don't 217 // include the terminator because the copy won't include it. 218 unsigned Size = 0; 219 for (; !isa<TerminatorInst>(I); ++I) { 220 // Debugger intrinsics don't incur code size. 221 if (isa<DbgInfoIntrinsic>(I)) continue; 222 223 // If this is a pointer->pointer bitcast, it is free. 224 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 225 continue; 226 227 // All other instructions count for at least one unit. 228 ++Size; 229 230 // Calls are more expensive. If they are non-intrinsic calls, we model them 231 // as having cost of 4. If they are a non-vector intrinsic, we model them 232 // as having cost of 2 total, and if they are a vector intrinsic, we model 233 // them as having cost 1. 234 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 235 if (!isa<IntrinsicInst>(CI)) 236 Size += 3; 237 else if (!CI->getType()->isVectorTy()) 238 Size += 1; 239 } 240 } 241 242 // Threading through a switch statement is particularly profitable. If this 243 // block ends in a switch, decrease its cost to make it more likely to happen. 244 if (isa<SwitchInst>(I)) 245 Size = Size > 6 ? Size-6 : 0; 246 247 return Size; 248} 249 250/// FindLoopHeaders - We do not want jump threading to turn proper loop 251/// structures into irreducible loops. Doing this breaks up the loop nesting 252/// hierarchy and pessimizes later transformations. To prevent this from 253/// happening, we first have to find the loop headers. Here we approximate this 254/// by finding targets of backedges in the CFG. 255/// 256/// Note that there definitely are cases when we want to allow threading of 257/// edges across a loop header. For example, threading a jump from outside the 258/// loop (the preheader) to an exit block of the loop is definitely profitable. 259/// It is also almost always profitable to thread backedges from within the loop 260/// to exit blocks, and is often profitable to thread backedges to other blocks 261/// within the loop (forming a nested loop). This simple analysis is not rich 262/// enough to track all of these properties and keep it up-to-date as the CFG 263/// mutates, so we don't allow any of these transformations. 264/// 265void JumpThreading::FindLoopHeaders(Function &F) { 266 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 267 FindFunctionBackedges(F, Edges); 268 269 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 270 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 271} 272 273// Helper method for ComputeValueKnownInPredecessors. If Value is a 274// ConstantInt, push it. If it's an undef, push 0. Otherwise, do nothing. 275static void PushConstantIntOrUndef(SmallVectorImpl<std::pair<ConstantInt*, 276 BasicBlock*> > &Result, 277 Constant *Value, BasicBlock* BB){ 278 if (ConstantInt *FoldedCInt = dyn_cast<ConstantInt>(Value)) 279 Result.push_back(std::make_pair(FoldedCInt, BB)); 280 else if (isa<UndefValue>(Value)) 281 Result.push_back(std::make_pair((ConstantInt*)0, BB)); 282} 283 284/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 285/// if we can infer that the value is a known ConstantInt in any of our 286/// predecessors. If so, return the known list of value and pred BB in the 287/// result vector. If a value is known to be undef, it is returned as null. 288/// 289/// This returns true if there were any known values. 290/// 291bool JumpThreading:: 292ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 293 // This method walks up use-def chains recursively. Because of this, we could 294 // get into an infinite loop going around loops in the use-def chain. To 295 // prevent this, keep track of what (value, block) pairs we've already visited 296 // and terminate the search if we loop back to them 297 if (!RecursionSet.insert(std::make_pair(V, BB)).second) 298 return false; 299 300 // An RAII help to remove this pair from the recursion set once the recursion 301 // stack pops back out again. 302 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB)); 303 304 // If V is a constantint, then it is known in all predecessors. 305 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 306 ConstantInt *CI = dyn_cast<ConstantInt>(V); 307 308 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 309 Result.push_back(std::make_pair(CI, *PI)); 310 311 return true; 312 } 313 314 // If V is a non-instruction value, or an instruction in a different block, 315 // then it can't be derived from a PHI. 316 Instruction *I = dyn_cast<Instruction>(V); 317 if (I == 0 || I->getParent() != BB) { 318 319 // Okay, if this is a live-in value, see if it has a known value at the end 320 // of any of our predecessors. 321 // 322 // FIXME: This should be an edge property, not a block end property. 323 /// TODO: Per PR2563, we could infer value range information about a 324 /// predecessor based on its terminator. 325 // 326 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 327 // "I" is a non-local compare-with-a-constant instruction. This would be 328 // able to handle value inequalities better, for example if the compare is 329 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 330 // Perhaps getConstantOnEdge should be smart enough to do this? 331 332 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 333 BasicBlock *P = *PI; 334 // If the value is known by LazyValueInfo to be a constant in a 335 // predecessor, use that information to try to thread this block. 336 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 337 if (PredCst == 0 || 338 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst))) 339 continue; 340 341 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P)); 342 } 343 344 return !Result.empty(); 345 } 346 347 /// If I is a PHI node, then we know the incoming values for any constants. 348 if (PHINode *PN = dyn_cast<PHINode>(I)) { 349 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 350 Value *InVal = PN->getIncomingValue(i); 351 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 352 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 353 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 354 } else { 355 Constant *CI = LVI->getConstantOnEdge(InVal, 356 PN->getIncomingBlock(i), BB); 357 // LVI returns null is no value could be determined. 358 if (!CI) continue; 359 PushConstantIntOrUndef(Result, CI, PN->getIncomingBlock(i)); 360 } 361 } 362 363 return !Result.empty(); 364 } 365 366 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 367 368 // Handle some boolean conditions. 369 if (I->getType()->getPrimitiveSizeInBits() == 1) { 370 // X | true -> true 371 // X & false -> false 372 if (I->getOpcode() == Instruction::Or || 373 I->getOpcode() == Instruction::And) { 374 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 375 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 376 377 if (LHSVals.empty() && RHSVals.empty()) 378 return false; 379 380 ConstantInt *InterestingVal; 381 if (I->getOpcode() == Instruction::Or) 382 InterestingVal = ConstantInt::getTrue(I->getContext()); 383 else 384 InterestingVal = ConstantInt::getFalse(I->getContext()); 385 386 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 387 388 // Scan for the sentinel. If we find an undef, force it to the 389 // interesting value: x|undef -> true and x&undef -> false. 390 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 391 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) { 392 Result.push_back(LHSVals[i]); 393 Result.back().first = InterestingVal; 394 LHSKnownBBs.insert(LHSVals[i].second); 395 } 396 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 397 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) { 398 // If we already inferred a value for this block on the LHS, don't 399 // re-add it. 400 if (!LHSKnownBBs.count(RHSVals[i].second)) { 401 Result.push_back(RHSVals[i]); 402 Result.back().first = InterestingVal; 403 } 404 } 405 406 return !Result.empty(); 407 } 408 409 // Handle the NOT form of XOR. 410 if (I->getOpcode() == Instruction::Xor && 411 isa<ConstantInt>(I->getOperand(1)) && 412 cast<ConstantInt>(I->getOperand(1))->isOne()) { 413 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result); 414 if (Result.empty()) 415 return false; 416 417 // Invert the known values. 418 for (unsigned i = 0, e = Result.size(); i != e; ++i) 419 if (Result[i].first) 420 Result[i].first = 421 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first)); 422 423 return true; 424 } 425 426 // Try to simplify some other binary operator values. 427 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 428 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 429 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals; 430 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals); 431 432 // Try to use constant folding to simplify the binary operator. 433 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 434 Constant *V = LHSVals[i].first; 435 if (V == 0) V = UndefValue::get(BO->getType()); 436 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); 437 438 PushConstantIntOrUndef(Result, Folded, LHSVals[i].second); 439 } 440 } 441 442 return !Result.empty(); 443 } 444 445 // Handle compare with phi operand, where the PHI is defined in this block. 446 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 447 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 448 if (PN && PN->getParent() == BB) { 449 // We can do this simplification if any comparisons fold to true or false. 450 // See if any do. 451 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 452 BasicBlock *PredBB = PN->getIncomingBlock(i); 453 Value *LHS = PN->getIncomingValue(i); 454 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 455 456 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 457 if (Res == 0) { 458 if (!isa<Constant>(RHS)) 459 continue; 460 461 LazyValueInfo::Tristate 462 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 463 cast<Constant>(RHS), PredBB, BB); 464 if (ResT == LazyValueInfo::Unknown) 465 continue; 466 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 467 } 468 469 if (Constant *ConstRes = dyn_cast<Constant>(Res)) 470 PushConstantIntOrUndef(Result, ConstRes, PredBB); 471 } 472 473 return !Result.empty(); 474 } 475 476 477 // If comparing a live-in value against a constant, see if we know the 478 // live-in value on any predecessors. 479 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) { 480 if (!isa<Instruction>(Cmp->getOperand(0)) || 481 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 482 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 483 484 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ 485 BasicBlock *P = *PI; 486 // If the value is known by LazyValueInfo to be a constant in a 487 // predecessor, use that information to try to thread this block. 488 LazyValueInfo::Tristate Res = 489 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 490 RHSCst, P, BB); 491 if (Res == LazyValueInfo::Unknown) 492 continue; 493 494 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 495 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P)); 496 } 497 498 return !Result.empty(); 499 } 500 501 // Try to find a constant value for the LHS of a comparison, 502 // and evaluate it statically if we can. 503 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { 504 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals; 505 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 506 507 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 508 Constant *V = LHSVals[i].first; 509 if (V == 0) V = UndefValue::get(CmpConst->getType()); 510 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(), 511 V, CmpConst); 512 PushConstantIntOrUndef(Result, Folded, LHSVals[i].second); 513 } 514 515 return !Result.empty(); 516 } 517 } 518 } 519 520 // If all else fails, see if LVI can figure out a constant value for us. 521 Constant *CI = LVI->getConstant(V, BB); 522 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI); 523 if (CInt) { 524 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 525 Result.push_back(std::make_pair(CInt, *PI)); 526 } 527 528 return !Result.empty(); 529} 530 531 532 533/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 534/// in an undefined jump, decide which block is best to revector to. 535/// 536/// Since we can pick an arbitrary destination, we pick the successor with the 537/// fewest predecessors. This should reduce the in-degree of the others. 538/// 539static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 540 TerminatorInst *BBTerm = BB->getTerminator(); 541 unsigned MinSucc = 0; 542 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 543 // Compute the successor with the minimum number of predecessors. 544 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 545 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 546 TestBB = BBTerm->getSuccessor(i); 547 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 548 if (NumPreds < MinNumPreds) 549 MinSucc = i; 550 } 551 552 return MinSucc; 553} 554 555/// ProcessBlock - If there are any predecessors whose control can be threaded 556/// through to a successor, transform them now. 557bool JumpThreading::ProcessBlock(BasicBlock *BB) { 558 // If the block is trivially dead, just return and let the caller nuke it. 559 // This simplifies other transformations. 560 if (pred_begin(BB) == pred_end(BB) && 561 BB != &BB->getParent()->getEntryBlock()) 562 return false; 563 564 // If this block has a single predecessor, and if that pred has a single 565 // successor, merge the blocks. This encourages recursive jump threading 566 // because now the condition in this block can be threaded through 567 // predecessors of our predecessor block. 568 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 569 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 570 SinglePred != BB) { 571 // If SinglePred was a loop header, BB becomes one. 572 if (LoopHeaders.erase(SinglePred)) 573 LoopHeaders.insert(BB); 574 575 // Remember if SinglePred was the entry block of the function. If so, we 576 // will need to move BB back to the entry position. 577 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 578 LVI->eraseBlock(SinglePred); 579 MergeBasicBlockIntoOnlyPred(BB); 580 581 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 582 BB->moveBefore(&BB->getParent()->getEntryBlock()); 583 return true; 584 } 585 } 586 587 // Look to see if the terminator is a branch of switch, if not we can't thread 588 // it. 589 Value *Condition; 590 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 591 // Can't thread an unconditional jump. 592 if (BI->isUnconditional()) return false; 593 Condition = BI->getCondition(); 594 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 595 Condition = SI->getCondition(); 596 else 597 return false; // Must be an invoke. 598 599 // If the terminator of this block is branching on a constant, simplify the 600 // terminator to an unconditional branch. This can occur due to threading in 601 // other blocks. 602 if (isa<ConstantInt>(Condition)) { 603 DEBUG(dbgs() << " In block '" << BB->getName() 604 << "' folding terminator: " << *BB->getTerminator() << '\n'); 605 ++NumFolds; 606 ConstantFoldTerminator(BB); 607 return true; 608 } 609 610 // If the terminator is branching on an undef, we can pick any of the 611 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 612 if (isa<UndefValue>(Condition)) { 613 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 614 615 // Fold the branch/switch. 616 TerminatorInst *BBTerm = BB->getTerminator(); 617 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 618 if (i == BestSucc) continue; 619 BBTerm->getSuccessor(i)->removePredecessor(BB, true); 620 } 621 622 DEBUG(dbgs() << " In block '" << BB->getName() 623 << "' folding undef terminator: " << *BBTerm << '\n'); 624 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 625 BBTerm->eraseFromParent(); 626 return true; 627 } 628 629 Instruction *CondInst = dyn_cast<Instruction>(Condition); 630 631 // All the rest of our checks depend on the condition being an instruction. 632 if (CondInst == 0) { 633 // FIXME: Unify this with code below. 634 if (ProcessThreadableEdges(Condition, BB)) 635 return true; 636 return false; 637 } 638 639 640 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 641 // For a comparison where the LHS is outside this block, it's possible 642 // that we've branched on it before. Used LVI to see if we can simplify 643 // the branch based on that. 644 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 645 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 646 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 647 if (CondBr && CondConst && CondBr->isConditional() && PI != PE && 648 (!isa<Instruction>(CondCmp->getOperand(0)) || 649 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { 650 // For predecessor edge, determine if the comparison is true or false 651 // on that edge. If they're all true or all false, we can simplify the 652 // branch. 653 // FIXME: We could handle mixed true/false by duplicating code. 654 LazyValueInfo::Tristate Baseline = 655 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0), 656 CondConst, *PI, BB); 657 if (Baseline != LazyValueInfo::Unknown) { 658 // Check that all remaining incoming values match the first one. 659 while (++PI != PE) { 660 LazyValueInfo::Tristate Ret = 661 LVI->getPredicateOnEdge(CondCmp->getPredicate(), 662 CondCmp->getOperand(0), CondConst, *PI, BB); 663 if (Ret != Baseline) break; 664 } 665 666 // If we terminated early, then one of the values didn't match. 667 if (PI == PE) { 668 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0; 669 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1; 670 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); 671 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); 672 CondBr->eraseFromParent(); 673 return true; 674 } 675 } 676 } 677 } 678 679 // Check for some cases that are worth simplifying. Right now we want to look 680 // for loads that are used by a switch or by the condition for the branch. If 681 // we see one, check to see if it's partially redundant. If so, insert a PHI 682 // which can then be used to thread the values. 683 // 684 Value *SimplifyValue = CondInst; 685 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 686 if (isa<Constant>(CondCmp->getOperand(1))) 687 SimplifyValue = CondCmp->getOperand(0); 688 689 // TODO: There are other places where load PRE would be profitable, such as 690 // more complex comparisons. 691 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 692 if (SimplifyPartiallyRedundantLoad(LI)) 693 return true; 694 695 696 // Handle a variety of cases where we are branching on something derived from 697 // a PHI node in the current block. If we can prove that any predecessors 698 // compute a predictable value based on a PHI node, thread those predecessors. 699 // 700 if (ProcessThreadableEdges(CondInst, BB)) 701 return true; 702 703 // If this is an otherwise-unfoldable branch on a phi node in the current 704 // block, see if we can simplify. 705 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 706 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 707 return ProcessBranchOnPHI(PN); 708 709 710 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 711 if (CondInst->getOpcode() == Instruction::Xor && 712 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 713 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 714 715 716 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 717 // "(X == 4)", thread through this block. 718 719 return false; 720} 721 722/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 723/// block that jump on exactly the same condition. This means that we almost 724/// always know the direction of the edge in the DESTBB: 725/// PREDBB: 726/// br COND, DESTBB, BBY 727/// DESTBB: 728/// br COND, BBZ, BBW 729/// 730/// If DESTBB has multiple predecessors, we can't just constant fold the branch 731/// in DESTBB, we have to thread over it. 732bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 733 BasicBlock *BB) { 734 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 735 736 // If both successors of PredBB go to DESTBB, we don't know anything. We can 737 // fold the branch to an unconditional one, which allows other recursive 738 // simplifications. 739 bool BranchDir; 740 if (PredBI->getSuccessor(1) != BB) 741 BranchDir = true; 742 else if (PredBI->getSuccessor(0) != BB) 743 BranchDir = false; 744 else { 745 DEBUG(dbgs() << " In block '" << PredBB->getName() 746 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 747 ++NumFolds; 748 ConstantFoldTerminator(PredBB); 749 return true; 750 } 751 752 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 753 754 // If the dest block has one predecessor, just fix the branch condition to a 755 // constant and fold it. 756 if (BB->getSinglePredecessor()) { 757 DEBUG(dbgs() << " In block '" << BB->getName() 758 << "' folding condition to '" << BranchDir << "': " 759 << *BB->getTerminator() << '\n'); 760 ++NumFolds; 761 Value *OldCond = DestBI->getCondition(); 762 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 763 BranchDir)); 764 // Delete dead instructions before we fold the branch. Folding the branch 765 // can eliminate edges from the CFG which can end up deleting OldCond. 766 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 767 ConstantFoldTerminator(BB); 768 return true; 769 } 770 771 772 // Next, figure out which successor we are threading to. 773 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 774 775 SmallVector<BasicBlock*, 2> Preds; 776 Preds.push_back(PredBB); 777 778 // Ok, try to thread it! 779 return ThreadEdge(BB, Preds, SuccBB); 780} 781 782/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 783/// block that switch on exactly the same condition. This means that we almost 784/// always know the direction of the edge in the DESTBB: 785/// PREDBB: 786/// switch COND [... DESTBB, BBY ... ] 787/// DESTBB: 788/// switch COND [... BBZ, BBW ] 789/// 790/// Optimizing switches like this is very important, because simplifycfg builds 791/// switches out of repeated 'if' conditions. 792bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 793 BasicBlock *DestBB) { 794 // Can't thread edge to self. 795 if (PredBB == DestBB) 796 return false; 797 798 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 799 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 800 801 // There are a variety of optimizations that we can potentially do on these 802 // blocks: we order them from most to least preferable. 803 804 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 805 // directly to their destination. This does not introduce *any* code size 806 // growth. Skip debug info first. 807 BasicBlock::iterator BBI = DestBB->begin(); 808 while (isa<DbgInfoIntrinsic>(BBI)) 809 BBI++; 810 811 // FIXME: Thread if it just contains a PHI. 812 if (isa<SwitchInst>(BBI)) { 813 bool MadeChange = false; 814 // Ignore the default edge for now. 815 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 816 ConstantInt *DestVal = DestSI->getCaseValue(i); 817 BasicBlock *DestSucc = DestSI->getSuccessor(i); 818 819 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 820 // PredSI has an explicit case for it. If so, forward. If it is covered 821 // by the default case, we can't update PredSI. 822 unsigned PredCase = PredSI->findCaseValue(DestVal); 823 if (PredCase == 0) continue; 824 825 // If PredSI doesn't go to DestBB on this value, then it won't reach the 826 // case on this condition. 827 if (PredSI->getSuccessor(PredCase) != DestBB && 828 DestSI->getSuccessor(i) != DestBB) 829 continue; 830 831 // Do not forward this if it already goes to this destination, this would 832 // be an infinite loop. 833 if (PredSI->getSuccessor(PredCase) == DestSucc) 834 continue; 835 836 // Otherwise, we're safe to make the change. Make sure that the edge from 837 // DestSI to DestSucc is not critical and has no PHI nodes. 838 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 839 DEBUG(dbgs() << "THROUGH: " << *DestSI); 840 841 // If the destination has PHI nodes, just split the edge for updating 842 // simplicity. 843 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 844 SplitCriticalEdge(DestSI, i, this); 845 DestSucc = DestSI->getSuccessor(i); 846 } 847 FoldSingleEntryPHINodes(DestSucc); 848 PredSI->setSuccessor(PredCase, DestSucc); 849 MadeChange = true; 850 } 851 852 if (MadeChange) 853 return true; 854 } 855 856 return false; 857} 858 859 860/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 861/// load instruction, eliminate it by replacing it with a PHI node. This is an 862/// important optimization that encourages jump threading, and needs to be run 863/// interlaced with other jump threading tasks. 864bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 865 // Don't hack volatile loads. 866 if (LI->isVolatile()) return false; 867 868 // If the load is defined in a block with exactly one predecessor, it can't be 869 // partially redundant. 870 BasicBlock *LoadBB = LI->getParent(); 871 if (LoadBB->getSinglePredecessor()) 872 return false; 873 874 Value *LoadedPtr = LI->getOperand(0); 875 876 // If the loaded operand is defined in the LoadBB, it can't be available. 877 // TODO: Could do simple PHI translation, that would be fun :) 878 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 879 if (PtrOp->getParent() == LoadBB) 880 return false; 881 882 // Scan a few instructions up from the load, to see if it is obviously live at 883 // the entry to its block. 884 BasicBlock::iterator BBIt = LI; 885 886 if (Value *AvailableVal = 887 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 888 // If the value if the load is locally available within the block, just use 889 // it. This frequently occurs for reg2mem'd allocas. 890 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 891 892 // If the returned value is the load itself, replace with an undef. This can 893 // only happen in dead loops. 894 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 895 LI->replaceAllUsesWith(AvailableVal); 896 LI->eraseFromParent(); 897 return true; 898 } 899 900 // Otherwise, if we scanned the whole block and got to the top of the block, 901 // we know the block is locally transparent to the load. If not, something 902 // might clobber its value. 903 if (BBIt != LoadBB->begin()) 904 return false; 905 906 907 SmallPtrSet<BasicBlock*, 8> PredsScanned; 908 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 909 AvailablePredsTy AvailablePreds; 910 BasicBlock *OneUnavailablePred = 0; 911 912 // If we got here, the loaded value is transparent through to the start of the 913 // block. Check to see if it is available in any of the predecessor blocks. 914 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 915 PI != PE; ++PI) { 916 BasicBlock *PredBB = *PI; 917 918 // If we already scanned this predecessor, skip it. 919 if (!PredsScanned.insert(PredBB)) 920 continue; 921 922 // Scan the predecessor to see if the value is available in the pred. 923 BBIt = PredBB->end(); 924 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 925 if (!PredAvailable) { 926 OneUnavailablePred = PredBB; 927 continue; 928 } 929 930 // If so, this load is partially redundant. Remember this info so that we 931 // can create a PHI node. 932 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 933 } 934 935 // If the loaded value isn't available in any predecessor, it isn't partially 936 // redundant. 937 if (AvailablePreds.empty()) return false; 938 939 // Okay, the loaded value is available in at least one (and maybe all!) 940 // predecessors. If the value is unavailable in more than one unique 941 // predecessor, we want to insert a merge block for those common predecessors. 942 // This ensures that we only have to insert one reload, thus not increasing 943 // code size. 944 BasicBlock *UnavailablePred = 0; 945 946 // If there is exactly one predecessor where the value is unavailable, the 947 // already computed 'OneUnavailablePred' block is it. If it ends in an 948 // unconditional branch, we know that it isn't a critical edge. 949 if (PredsScanned.size() == AvailablePreds.size()+1 && 950 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 951 UnavailablePred = OneUnavailablePred; 952 } else if (PredsScanned.size() != AvailablePreds.size()) { 953 // Otherwise, we had multiple unavailable predecessors or we had a critical 954 // edge from the one. 955 SmallVector<BasicBlock*, 8> PredsToSplit; 956 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 957 958 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 959 AvailablePredSet.insert(AvailablePreds[i].first); 960 961 // Add all the unavailable predecessors to the PredsToSplit list. 962 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 963 PI != PE; ++PI) { 964 BasicBlock *P = *PI; 965 // If the predecessor is an indirect goto, we can't split the edge. 966 if (isa<IndirectBrInst>(P->getTerminator())) 967 return false; 968 969 if (!AvailablePredSet.count(P)) 970 PredsToSplit.push_back(P); 971 } 972 973 // Split them out to their own block. 974 UnavailablePred = 975 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 976 "thread-pre-split", this); 977 } 978 979 // If the value isn't available in all predecessors, then there will be 980 // exactly one where it isn't available. Insert a load on that edge and add 981 // it to the AvailablePreds list. 982 if (UnavailablePred) { 983 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 984 "Can't handle critical edge here!"); 985 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 986 LI->getAlignment(), 987 UnavailablePred->getTerminator()); 988 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 989 } 990 991 // Now we know that each predecessor of this block has a value in 992 // AvailablePreds, sort them for efficient access as we're walking the preds. 993 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 994 995 // Create a PHI node at the start of the block for the PRE'd load value. 996 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 997 PN->takeName(LI); 998 999 // Insert new entries into the PHI for each predecessor. A single block may 1000 // have multiple entries here. 1001 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 1002 ++PI) { 1003 BasicBlock *P = *PI; 1004 AvailablePredsTy::iterator I = 1005 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 1006 std::make_pair(P, (Value*)0)); 1007 1008 assert(I != AvailablePreds.end() && I->first == P && 1009 "Didn't find entry for predecessor!"); 1010 1011 PN->addIncoming(I->second, I->first); 1012 } 1013 1014 //cerr << "PRE: " << *LI << *PN << "\n"; 1015 1016 LI->replaceAllUsesWith(PN); 1017 LI->eraseFromParent(); 1018 1019 return true; 1020} 1021 1022/// FindMostPopularDest - The specified list contains multiple possible 1023/// threadable destinations. Pick the one that occurs the most frequently in 1024/// the list. 1025static BasicBlock * 1026FindMostPopularDest(BasicBlock *BB, 1027 const SmallVectorImpl<std::pair<BasicBlock*, 1028 BasicBlock*> > &PredToDestList) { 1029 assert(!PredToDestList.empty()); 1030 1031 // Determine popularity. If there are multiple possible destinations, we 1032 // explicitly choose to ignore 'undef' destinations. We prefer to thread 1033 // blocks with known and real destinations to threading undef. We'll handle 1034 // them later if interesting. 1035 DenseMap<BasicBlock*, unsigned> DestPopularity; 1036 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1037 if (PredToDestList[i].second) 1038 DestPopularity[PredToDestList[i].second]++; 1039 1040 // Find the most popular dest. 1041 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 1042 BasicBlock *MostPopularDest = DPI->first; 1043 unsigned Popularity = DPI->second; 1044 SmallVector<BasicBlock*, 4> SamePopularity; 1045 1046 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 1047 // If the popularity of this entry isn't higher than the popularity we've 1048 // seen so far, ignore it. 1049 if (DPI->second < Popularity) 1050 ; // ignore. 1051 else if (DPI->second == Popularity) { 1052 // If it is the same as what we've seen so far, keep track of it. 1053 SamePopularity.push_back(DPI->first); 1054 } else { 1055 // If it is more popular, remember it. 1056 SamePopularity.clear(); 1057 MostPopularDest = DPI->first; 1058 Popularity = DPI->second; 1059 } 1060 } 1061 1062 // Okay, now we know the most popular destination. If there is more than 1063 // destination, we need to determine one. This is arbitrary, but we need 1064 // to make a deterministic decision. Pick the first one that appears in the 1065 // successor list. 1066 if (!SamePopularity.empty()) { 1067 SamePopularity.push_back(MostPopularDest); 1068 TerminatorInst *TI = BB->getTerminator(); 1069 for (unsigned i = 0; ; ++i) { 1070 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1071 1072 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1073 TI->getSuccessor(i)) == SamePopularity.end()) 1074 continue; 1075 1076 MostPopularDest = TI->getSuccessor(i); 1077 break; 1078 } 1079 } 1080 1081 // Okay, we have finally picked the most popular destination. 1082 return MostPopularDest; 1083} 1084 1085bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) { 1086 // If threading this would thread across a loop header, don't even try to 1087 // thread the edge. 1088 if (LoopHeaders.count(BB)) 1089 return false; 1090 1091 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 1092 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) 1093 return false; 1094 1095 assert(!PredValues.empty() && 1096 "ComputeValueKnownInPredecessors returned true with no values"); 1097 1098 DEBUG(dbgs() << "IN BB: " << *BB; 1099 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1100 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "; 1101 if (PredValues[i].first) 1102 dbgs() << *PredValues[i].first; 1103 else 1104 dbgs() << "UNDEF"; 1105 dbgs() << " for pred '" << PredValues[i].second->getName() 1106 << "'.\n"; 1107 }); 1108 1109 // Decide what we want to thread through. Convert our list of known values to 1110 // a list of known destinations for each pred. This also discards duplicate 1111 // predecessors and keeps track of the undefined inputs (which are represented 1112 // as a null dest in the PredToDestList). 1113 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1114 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1115 1116 BasicBlock *OnlyDest = 0; 1117 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1118 1119 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1120 BasicBlock *Pred = PredValues[i].second; 1121 if (!SeenPreds.insert(Pred)) 1122 continue; // Duplicate predecessor entry. 1123 1124 // If the predecessor ends with an indirect goto, we can't change its 1125 // destination. 1126 if (isa<IndirectBrInst>(Pred->getTerminator())) 1127 continue; 1128 1129 ConstantInt *Val = PredValues[i].first; 1130 1131 BasicBlock *DestBB; 1132 if (Val == 0) // Undef. 1133 DestBB = 0; 1134 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1135 DestBB = BI->getSuccessor(Val->isZero()); 1136 else { 1137 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1138 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1139 } 1140 1141 // If we have exactly one destination, remember it for efficiency below. 1142 if (i == 0) 1143 OnlyDest = DestBB; 1144 else if (OnlyDest != DestBB) 1145 OnlyDest = MultipleDestSentinel; 1146 1147 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1148 } 1149 1150 // If all edges were unthreadable, we fail. 1151 if (PredToDestList.empty()) 1152 return false; 1153 1154 // Determine which is the most common successor. If we have many inputs and 1155 // this block is a switch, we want to start by threading the batch that goes 1156 // to the most popular destination first. If we only know about one 1157 // threadable destination (the common case) we can avoid this. 1158 BasicBlock *MostPopularDest = OnlyDest; 1159 1160 if (MostPopularDest == MultipleDestSentinel) 1161 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1162 1163 // Now that we know what the most popular destination is, factor all 1164 // predecessors that will jump to it into a single predecessor. 1165 SmallVector<BasicBlock*, 16> PredsToFactor; 1166 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1167 if (PredToDestList[i].second == MostPopularDest) { 1168 BasicBlock *Pred = PredToDestList[i].first; 1169 1170 // This predecessor may be a switch or something else that has multiple 1171 // edges to the block. Factor each of these edges by listing them 1172 // according to # occurrences in PredsToFactor. 1173 TerminatorInst *PredTI = Pred->getTerminator(); 1174 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1175 if (PredTI->getSuccessor(i) == BB) 1176 PredsToFactor.push_back(Pred); 1177 } 1178 1179 // If the threadable edges are branching on an undefined value, we get to pick 1180 // the destination that these predecessors should get to. 1181 if (MostPopularDest == 0) 1182 MostPopularDest = BB->getTerminator()-> 1183 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1184 1185 // Ok, try to thread it! 1186 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1187} 1188 1189/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1190/// a PHI node in the current block. See if there are any simplifications we 1191/// can do based on inputs to the phi node. 1192/// 1193bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1194 BasicBlock *BB = PN->getParent(); 1195 1196 // TODO: We could make use of this to do it once for blocks with common PHI 1197 // values. 1198 SmallVector<BasicBlock*, 1> PredBBs; 1199 PredBBs.resize(1); 1200 1201 // If any of the predecessor blocks end in an unconditional branch, we can 1202 // *duplicate* the conditional branch into that block in order to further 1203 // encourage jump threading and to eliminate cases where we have branch on a 1204 // phi of an icmp (branch on icmp is much better). 1205 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1206 BasicBlock *PredBB = PN->getIncomingBlock(i); 1207 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1208 if (PredBr->isUnconditional()) { 1209 PredBBs[0] = PredBB; 1210 // Try to duplicate BB into PredBB. 1211 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1212 return true; 1213 } 1214 } 1215 1216 return false; 1217} 1218 1219/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1220/// a xor instruction in the current block. See if there are any 1221/// simplifications we can do based on inputs to the xor. 1222/// 1223bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1224 BasicBlock *BB = BO->getParent(); 1225 1226 // If either the LHS or RHS of the xor is a constant, don't do this 1227 // optimization. 1228 if (isa<ConstantInt>(BO->getOperand(0)) || 1229 isa<ConstantInt>(BO->getOperand(1))) 1230 return false; 1231 1232 // If the first instruction in BB isn't a phi, we won't be able to infer 1233 // anything special about any particular predecessor. 1234 if (!isa<PHINode>(BB->front())) 1235 return false; 1236 1237 // If we have a xor as the branch input to this block, and we know that the 1238 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1239 // the condition into the predecessor and fix that value to true, saving some 1240 // logical ops on that path and encouraging other paths to simplify. 1241 // 1242 // This copies something like this: 1243 // 1244 // BB: 1245 // %X = phi i1 [1], [%X'] 1246 // %Y = icmp eq i32 %A, %B 1247 // %Z = xor i1 %X, %Y 1248 // br i1 %Z, ... 1249 // 1250 // Into: 1251 // BB': 1252 // %Y = icmp ne i32 %A, %B 1253 // br i1 %Z, ... 1254 1255 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues; 1256 bool isLHS = true; 1257 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) { 1258 assert(XorOpValues.empty()); 1259 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues)) 1260 return false; 1261 isLHS = false; 1262 } 1263 1264 assert(!XorOpValues.empty() && 1265 "ComputeValueKnownInPredecessors returned true with no values"); 1266 1267 // Scan the information to see which is most popular: true or false. The 1268 // predecessors can be of the set true, false, or undef. 1269 unsigned NumTrue = 0, NumFalse = 0; 1270 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1271 if (!XorOpValues[i].first) continue; // Ignore undefs for the count. 1272 if (XorOpValues[i].first->isZero()) 1273 ++NumFalse; 1274 else 1275 ++NumTrue; 1276 } 1277 1278 // Determine which value to split on, true, false, or undef if neither. 1279 ConstantInt *SplitVal = 0; 1280 if (NumTrue > NumFalse) 1281 SplitVal = ConstantInt::getTrue(BB->getContext()); 1282 else if (NumTrue != 0 || NumFalse != 0) 1283 SplitVal = ConstantInt::getFalse(BB->getContext()); 1284 1285 // Collect all of the blocks that this can be folded into so that we can 1286 // factor this once and clone it once. 1287 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1288 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1289 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue; 1290 1291 BlocksToFoldInto.push_back(XorOpValues[i].second); 1292 } 1293 1294 // If we inferred a value for all of the predecessors, then duplication won't 1295 // help us. However, we can just replace the LHS or RHS with the constant. 1296 if (BlocksToFoldInto.size() == 1297 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1298 if (SplitVal == 0) { 1299 // If all preds provide undef, just nuke the xor, because it is undef too. 1300 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1301 BO->eraseFromParent(); 1302 } else if (SplitVal->isZero()) { 1303 // If all preds provide 0, replace the xor with the other input. 1304 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1305 BO->eraseFromParent(); 1306 } else { 1307 // If all preds provide 1, set the computed value to 1. 1308 BO->setOperand(!isLHS, SplitVal); 1309 } 1310 1311 return true; 1312 } 1313 1314 // Try to duplicate BB into PredBB. 1315 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1316} 1317 1318 1319/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1320/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1321/// NewPred using the entries from OldPred (suitably mapped). 1322static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1323 BasicBlock *OldPred, 1324 BasicBlock *NewPred, 1325 DenseMap<Instruction*, Value*> &ValueMap) { 1326 for (BasicBlock::iterator PNI = PHIBB->begin(); 1327 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1328 // Ok, we have a PHI node. Figure out what the incoming value was for the 1329 // DestBlock. 1330 Value *IV = PN->getIncomingValueForBlock(OldPred); 1331 1332 // Remap the value if necessary. 1333 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1334 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1335 if (I != ValueMap.end()) 1336 IV = I->second; 1337 } 1338 1339 PN->addIncoming(IV, NewPred); 1340 } 1341} 1342 1343/// ThreadEdge - We have decided that it is safe and profitable to factor the 1344/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1345/// across BB. Transform the IR to reflect this change. 1346bool JumpThreading::ThreadEdge(BasicBlock *BB, 1347 const SmallVectorImpl<BasicBlock*> &PredBBs, 1348 BasicBlock *SuccBB) { 1349 // If threading to the same block as we come from, we would infinite loop. 1350 if (SuccBB == BB) { 1351 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1352 << "' - would thread to self!\n"); 1353 return false; 1354 } 1355 1356 // If threading this would thread across a loop header, don't thread the edge. 1357 // See the comments above FindLoopHeaders for justifications and caveats. 1358 if (LoopHeaders.count(BB)) { 1359 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1360 << "' to dest BB '" << SuccBB->getName() 1361 << "' - it might create an irreducible loop!\n"); 1362 return false; 1363 } 1364 1365 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1366 if (JumpThreadCost > Threshold) { 1367 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1368 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1369 return false; 1370 } 1371 1372 // And finally, do it! Start by factoring the predecessors is needed. 1373 BasicBlock *PredBB; 1374 if (PredBBs.size() == 1) 1375 PredBB = PredBBs[0]; 1376 else { 1377 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1378 << " common predecessors.\n"); 1379 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1380 ".thr_comm", this); 1381 } 1382 1383 // And finally, do it! 1384 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1385 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1386 << ", across block:\n " 1387 << *BB << "\n"); 1388 1389 LVI->threadEdge(PredBB, BB, SuccBB); 1390 1391 // We are going to have to map operands from the original BB block to the new 1392 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1393 // account for entry from PredBB. 1394 DenseMap<Instruction*, Value*> ValueMapping; 1395 1396 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1397 BB->getName()+".thread", 1398 BB->getParent(), BB); 1399 NewBB->moveAfter(PredBB); 1400 1401 BasicBlock::iterator BI = BB->begin(); 1402 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1403 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1404 1405 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1406 // mapping and using it to remap operands in the cloned instructions. 1407 for (; !isa<TerminatorInst>(BI); ++BI) { 1408 Instruction *New = BI->clone(); 1409 New->setName(BI->getName()); 1410 NewBB->getInstList().push_back(New); 1411 ValueMapping[BI] = New; 1412 1413 // Remap operands to patch up intra-block references. 1414 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1415 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1416 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1417 if (I != ValueMapping.end()) 1418 New->setOperand(i, I->second); 1419 } 1420 } 1421 1422 // We didn't copy the terminator from BB over to NewBB, because there is now 1423 // an unconditional jump to SuccBB. Insert the unconditional jump. 1424 BranchInst::Create(SuccBB, NewBB); 1425 1426 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1427 // PHI nodes for NewBB now. 1428 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1429 1430 // If there were values defined in BB that are used outside the block, then we 1431 // now have to update all uses of the value to use either the original value, 1432 // the cloned value, or some PHI derived value. This can require arbitrary 1433 // PHI insertion, of which we are prepared to do, clean these up now. 1434 SSAUpdater SSAUpdate; 1435 SmallVector<Use*, 16> UsesToRename; 1436 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1437 // Scan all uses of this instruction to see if it is used outside of its 1438 // block, and if so, record them in UsesToRename. 1439 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1440 ++UI) { 1441 Instruction *User = cast<Instruction>(*UI); 1442 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1443 if (UserPN->getIncomingBlock(UI) == BB) 1444 continue; 1445 } else if (User->getParent() == BB) 1446 continue; 1447 1448 UsesToRename.push_back(&UI.getUse()); 1449 } 1450 1451 // If there are no uses outside the block, we're done with this instruction. 1452 if (UsesToRename.empty()) 1453 continue; 1454 1455 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1456 1457 // We found a use of I outside of BB. Rename all uses of I that are outside 1458 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1459 // with the two values we know. 1460 SSAUpdate.Initialize(I->getType(), I->getName()); 1461 SSAUpdate.AddAvailableValue(BB, I); 1462 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1463 1464 while (!UsesToRename.empty()) 1465 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1466 DEBUG(dbgs() << "\n"); 1467 } 1468 1469 1470 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1471 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1472 // us to simplify any PHI nodes in BB. 1473 TerminatorInst *PredTerm = PredBB->getTerminator(); 1474 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1475 if (PredTerm->getSuccessor(i) == BB) { 1476 BB->removePredecessor(PredBB, true); 1477 PredTerm->setSuccessor(i, NewBB); 1478 } 1479 1480 // At this point, the IR is fully up to date and consistent. Do a quick scan 1481 // over the new instructions and zap any that are constants or dead. This 1482 // frequently happens because of phi translation. 1483 SimplifyInstructionsInBlock(NewBB, TD); 1484 1485 // Threaded an edge! 1486 ++NumThreads; 1487 return true; 1488} 1489 1490/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1491/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1492/// If we can duplicate the contents of BB up into PredBB do so now, this 1493/// improves the odds that the branch will be on an analyzable instruction like 1494/// a compare. 1495bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1496 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1497 assert(!PredBBs.empty() && "Can't handle an empty set"); 1498 1499 // If BB is a loop header, then duplicating this block outside the loop would 1500 // cause us to transform this into an irreducible loop, don't do this. 1501 // See the comments above FindLoopHeaders for justifications and caveats. 1502 if (LoopHeaders.count(BB)) { 1503 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1504 << "' into predecessor block '" << PredBBs[0]->getName() 1505 << "' - it might create an irreducible loop!\n"); 1506 return false; 1507 } 1508 1509 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1510 if (DuplicationCost > Threshold) { 1511 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1512 << "' - Cost is too high: " << DuplicationCost << "\n"); 1513 return false; 1514 } 1515 1516 // And finally, do it! Start by factoring the predecessors is needed. 1517 BasicBlock *PredBB; 1518 if (PredBBs.size() == 1) 1519 PredBB = PredBBs[0]; 1520 else { 1521 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1522 << " common predecessors.\n"); 1523 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1524 ".thr_comm", this); 1525 } 1526 1527 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1528 // of PredBB. 1529 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1530 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1531 << DuplicationCost << " block is:" << *BB << "\n"); 1532 1533 // Unless PredBB ends with an unconditional branch, split the edge so that we 1534 // can just clone the bits from BB into the end of the new PredBB. 1535 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1536 1537 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1538 PredBB = SplitEdge(PredBB, BB, this); 1539 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1540 } 1541 1542 // We are going to have to map operands from the original BB block into the 1543 // PredBB block. Evaluate PHI nodes in BB. 1544 DenseMap<Instruction*, Value*> ValueMapping; 1545 1546 BasicBlock::iterator BI = BB->begin(); 1547 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1548 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1549 1550 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1551 // mapping and using it to remap operands in the cloned instructions. 1552 for (; BI != BB->end(); ++BI) { 1553 Instruction *New = BI->clone(); 1554 1555 // Remap operands to patch up intra-block references. 1556 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1557 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1558 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1559 if (I != ValueMapping.end()) 1560 New->setOperand(i, I->second); 1561 } 1562 1563 // If this instruction can be simplified after the operands are updated, 1564 // just use the simplified value instead. This frequently happens due to 1565 // phi translation. 1566 if (Value *IV = SimplifyInstruction(New, TD)) { 1567 delete New; 1568 ValueMapping[BI] = IV; 1569 } else { 1570 // Otherwise, insert the new instruction into the block. 1571 New->setName(BI->getName()); 1572 PredBB->getInstList().insert(OldPredBranch, New); 1573 ValueMapping[BI] = New; 1574 } 1575 } 1576 1577 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1578 // add entries to the PHI nodes for branch from PredBB now. 1579 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1580 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1581 ValueMapping); 1582 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1583 ValueMapping); 1584 1585 // If there were values defined in BB that are used outside the block, then we 1586 // now have to update all uses of the value to use either the original value, 1587 // the cloned value, or some PHI derived value. This can require arbitrary 1588 // PHI insertion, of which we are prepared to do, clean these up now. 1589 SSAUpdater SSAUpdate; 1590 SmallVector<Use*, 16> UsesToRename; 1591 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1592 // Scan all uses of this instruction to see if it is used outside of its 1593 // block, and if so, record them in UsesToRename. 1594 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1595 ++UI) { 1596 Instruction *User = cast<Instruction>(*UI); 1597 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1598 if (UserPN->getIncomingBlock(UI) == BB) 1599 continue; 1600 } else if (User->getParent() == BB) 1601 continue; 1602 1603 UsesToRename.push_back(&UI.getUse()); 1604 } 1605 1606 // If there are no uses outside the block, we're done with this instruction. 1607 if (UsesToRename.empty()) 1608 continue; 1609 1610 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1611 1612 // We found a use of I outside of BB. Rename all uses of I that are outside 1613 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1614 // with the two values we know. 1615 SSAUpdate.Initialize(I->getType(), I->getName()); 1616 SSAUpdate.AddAvailableValue(BB, I); 1617 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1618 1619 while (!UsesToRename.empty()) 1620 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1621 DEBUG(dbgs() << "\n"); 1622 } 1623 1624 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1625 // that we nuked. 1626 BB->removePredecessor(PredBB, true); 1627 1628 // Remove the unconditional branch at the end of the PredBB block. 1629 OldPredBranch->eraseFromParent(); 1630 1631 ++NumDupes; 1632 return true; 1633} 1634 1635 1636