JumpThreading.cpp revision 081c34b725980f995be9080eaec24cd3dfaaf065
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 initializeJumpThreadingPass(*PassRegistry::getPassRegistry()); 91 } 92 93 bool runOnFunction(Function &F); 94 95 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 96 AU.addRequired<LazyValueInfo>(); 97 AU.addPreserved<LazyValueInfo>(); 98 } 99 100 void FindLoopHeaders(Function &F); 101 bool ProcessBlock(BasicBlock *BB); 102 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 103 BasicBlock *SuccBB); 104 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 105 const SmallVectorImpl<BasicBlock *> &PredBBs); 106 107 typedef SmallVectorImpl<std::pair<ConstantInt*, 108 BasicBlock*> > PredValueInfo; 109 110 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 111 PredValueInfo &Result); 112 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB); 113 114 115 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 116 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 117 118 bool ProcessBranchOnPHI(PHINode *PN); 119 bool ProcessBranchOnXOR(BinaryOperator *BO); 120 121 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 122 }; 123} 124 125char JumpThreading::ID = 0; 126INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", 127 "Jump Threading", false, false) 128INITIALIZE_PASS_DEPENDENCY(LazyValueInfo) 129INITIALIZE_PASS_END(JumpThreading, "jump-threading", 130 "Jump Threading", false, false) 131 132// Public interface to the Jump Threading pass 133FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 134 135/// runOnFunction - Top level algorithm. 136/// 137bool JumpThreading::runOnFunction(Function &F) { 138 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 139 TD = getAnalysisIfAvailable<TargetData>(); 140 LVI = &getAnalysis<LazyValueInfo>(); 141 142 FindLoopHeaders(F); 143 144 bool Changed, EverChanged = false; 145 do { 146 Changed = false; 147 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 148 BasicBlock *BB = I; 149 // Thread all of the branches we can over this block. 150 while (ProcessBlock(BB)) 151 Changed = true; 152 153 ++I; 154 155 // If the block is trivially dead, zap it. This eliminates the successor 156 // edges which simplifies the CFG. 157 if (pred_begin(BB) == pred_end(BB) && 158 BB != &BB->getParent()->getEntryBlock()) { 159 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 160 << "' with terminator: " << *BB->getTerminator() << '\n'); 161 LoopHeaders.erase(BB); 162 LVI->eraseBlock(BB); 163 DeleteDeadBlock(BB); 164 Changed = true; 165 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 166 // Can't thread an unconditional jump, but if the block is "almost 167 // empty", we can replace uses of it with uses of the successor and make 168 // this dead. 169 if (BI->isUnconditional() && 170 BB != &BB->getParent()->getEntryBlock()) { 171 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 172 // Ignore dbg intrinsics. 173 while (isa<DbgInfoIntrinsic>(BBI)) 174 ++BBI; 175 // If the terminator is the only non-phi instruction, try to nuke it. 176 if (BBI->isTerminator()) { 177 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 178 // block, we have to make sure it isn't in the LoopHeaders set. We 179 // reinsert afterward if needed. 180 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 181 BasicBlock *Succ = BI->getSuccessor(0); 182 183 // FIXME: It is always conservatively correct to drop the info 184 // for a block even if it doesn't get erased. This isn't totally 185 // awesome, but it allows us to use AssertingVH to prevent nasty 186 // dangling pointer issues within LazyValueInfo. 187 LVI->eraseBlock(BB); 188 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 189 Changed = true; 190 // If we deleted BB and BB was the header of a loop, then the 191 // successor is now the header of the loop. 192 BB = Succ; 193 } 194 195 if (ErasedFromLoopHeaders) 196 LoopHeaders.insert(BB); 197 } 198 } 199 } 200 } 201 EverChanged |= Changed; 202 } while (Changed); 203 204 LoopHeaders.clear(); 205 return EverChanged; 206} 207 208/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 209/// thread across it. 210static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 211 /// Ignore PHI nodes, these will be flattened when duplication happens. 212 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 213 214 // FIXME: THREADING will delete values that are just used to compute the 215 // branch, so they shouldn't count against the duplication cost. 216 217 218 // Sum up the cost of each instruction until we get to the terminator. Don't 219 // include the terminator because the copy won't include it. 220 unsigned Size = 0; 221 for (; !isa<TerminatorInst>(I); ++I) { 222 // Debugger intrinsics don't incur code size. 223 if (isa<DbgInfoIntrinsic>(I)) continue; 224 225 // If this is a pointer->pointer bitcast, it is free. 226 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 227 continue; 228 229 // All other instructions count for at least one unit. 230 ++Size; 231 232 // Calls are more expensive. If they are non-intrinsic calls, we model them 233 // as having cost of 4. If they are a non-vector intrinsic, we model them 234 // as having cost of 2 total, and if they are a vector intrinsic, we model 235 // them as having cost 1. 236 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 237 if (!isa<IntrinsicInst>(CI)) 238 Size += 3; 239 else if (!CI->getType()->isVectorTy()) 240 Size += 1; 241 } 242 } 243 244 // Threading through a switch statement is particularly profitable. If this 245 // block ends in a switch, decrease its cost to make it more likely to happen. 246 if (isa<SwitchInst>(I)) 247 Size = Size > 6 ? Size-6 : 0; 248 249 return Size; 250} 251 252/// FindLoopHeaders - We do not want jump threading to turn proper loop 253/// structures into irreducible loops. Doing this breaks up the loop nesting 254/// hierarchy and pessimizes later transformations. To prevent this from 255/// happening, we first have to find the loop headers. Here we approximate this 256/// by finding targets of backedges in the CFG. 257/// 258/// Note that there definitely are cases when we want to allow threading of 259/// edges across a loop header. For example, threading a jump from outside the 260/// loop (the preheader) to an exit block of the loop is definitely profitable. 261/// It is also almost always profitable to thread backedges from within the loop 262/// to exit blocks, and is often profitable to thread backedges to other blocks 263/// within the loop (forming a nested loop). This simple analysis is not rich 264/// enough to track all of these properties and keep it up-to-date as the CFG 265/// mutates, so we don't allow any of these transformations. 266/// 267void JumpThreading::FindLoopHeaders(Function &F) { 268 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 269 FindFunctionBackedges(F, Edges); 270 271 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 272 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 273} 274 275// Helper method for ComputeValueKnownInPredecessors. If Value is a 276// ConstantInt, push it. If it's an undef, push 0. Otherwise, do nothing. 277static void PushConstantIntOrUndef(SmallVectorImpl<std::pair<ConstantInt*, 278 BasicBlock*> > &Result, 279 Constant *Value, BasicBlock* BB){ 280 if (ConstantInt *FoldedCInt = dyn_cast<ConstantInt>(Value)) 281 Result.push_back(std::make_pair(FoldedCInt, BB)); 282 else if (isa<UndefValue>(Value)) 283 Result.push_back(std::make_pair((ConstantInt*)0, BB)); 284} 285 286/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 287/// if we can infer that the value is a known ConstantInt in any of our 288/// predecessors. If so, return the known list of value and pred BB in the 289/// result vector. If a value is known to be undef, it is returned as null. 290/// 291/// This returns true if there were any known values. 292/// 293bool JumpThreading:: 294ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){ 295 // This method walks up use-def chains recursively. Because of this, we could 296 // get into an infinite loop going around loops in the use-def chain. To 297 // prevent this, keep track of what (value, block) pairs we've already visited 298 // and terminate the search if we loop back to them 299 if (!RecursionSet.insert(std::make_pair(V, BB)).second) 300 return false; 301 302 // An RAII help to remove this pair from the recursion set once the recursion 303 // stack pops back out again. 304 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB)); 305 306 // If V is a constantint, then it is known in all predecessors. 307 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) { 308 ConstantInt *CI = dyn_cast<ConstantInt>(V); 309 310 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 311 Result.push_back(std::make_pair(CI, *PI)); 312 313 return true; 314 } 315 316 // If V is a non-instruction value, or an instruction in a different block, 317 // then it can't be derived from a PHI. 318 Instruction *I = dyn_cast<Instruction>(V); 319 if (I == 0 || I->getParent() != BB) { 320 321 // Okay, if this is a live-in value, see if it has a known value at the end 322 // of any of our predecessors. 323 // 324 // FIXME: This should be an edge property, not a block end property. 325 /// TODO: Per PR2563, we could infer value range information about a 326 /// predecessor based on its terminator. 327 // 328 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 329 // "I" is a non-local compare-with-a-constant instruction. This would be 330 // able to handle value inequalities better, for example if the compare is 331 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 332 // Perhaps getConstantOnEdge should be smart enough to do this? 333 334 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 335 BasicBlock *P = *PI; 336 // If the value is known by LazyValueInfo to be a constant in a 337 // predecessor, use that information to try to thread this block. 338 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB); 339 if (PredCst == 0 || 340 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst))) 341 continue; 342 343 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P)); 344 } 345 346 return !Result.empty(); 347 } 348 349 /// If I is a PHI node, then we know the incoming values for any constants. 350 if (PHINode *PN = dyn_cast<PHINode>(I)) { 351 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 352 Value *InVal = PN->getIncomingValue(i); 353 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) { 354 ConstantInt *CI = dyn_cast<ConstantInt>(InVal); 355 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i))); 356 } else { 357 Constant *CI = LVI->getConstantOnEdge(InVal, 358 PN->getIncomingBlock(i), BB); 359 // LVI returns null is no value could be determined. 360 if (!CI) continue; 361 PushConstantIntOrUndef(Result, CI, PN->getIncomingBlock(i)); 362 } 363 } 364 365 return !Result.empty(); 366 } 367 368 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals; 369 370 // Handle some boolean conditions. 371 if (I->getType()->getPrimitiveSizeInBits() == 1) { 372 // X | true -> true 373 // X & false -> false 374 if (I->getOpcode() == Instruction::Or || 375 I->getOpcode() == Instruction::And) { 376 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 377 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals); 378 379 if (LHSVals.empty() && RHSVals.empty()) 380 return false; 381 382 ConstantInt *InterestingVal; 383 if (I->getOpcode() == Instruction::Or) 384 InterestingVal = ConstantInt::getTrue(I->getContext()); 385 else 386 InterestingVal = ConstantInt::getFalse(I->getContext()); 387 388 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 389 390 // Scan for the sentinel. If we find an undef, force it to the 391 // interesting value: x|undef -> true and x&undef -> false. 392 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) 393 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) { 394 Result.push_back(LHSVals[i]); 395 Result.back().first = InterestingVal; 396 LHSKnownBBs.insert(LHSVals[i].second); 397 } 398 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) 399 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) { 400 // If we already inferred a value for this block on the LHS, don't 401 // re-add it. 402 if (!LHSKnownBBs.count(RHSVals[i].second)) { 403 Result.push_back(RHSVals[i]); 404 Result.back().first = InterestingVal; 405 } 406 } 407 408 return !Result.empty(); 409 } 410 411 // Handle the NOT form of XOR. 412 if (I->getOpcode() == Instruction::Xor && 413 isa<ConstantInt>(I->getOperand(1)) && 414 cast<ConstantInt>(I->getOperand(1))->isOne()) { 415 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result); 416 if (Result.empty()) 417 return false; 418 419 // Invert the known values. 420 for (unsigned i = 0, e = Result.size(); i != e; ++i) 421 if (Result[i].first) 422 Result[i].first = 423 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first)); 424 425 return true; 426 } 427 428 // Try to simplify some other binary operator values. 429 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 430 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 431 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals; 432 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals); 433 434 // Try to use constant folding to simplify the binary operator. 435 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 436 Constant *V = LHSVals[i].first; 437 if (V == 0) V = UndefValue::get(BO->getType()); 438 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); 439 440 PushConstantIntOrUndef(Result, Folded, LHSVals[i].second); 441 } 442 } 443 444 return !Result.empty(); 445 } 446 447 // Handle compare with phi operand, where the PHI is defined in this block. 448 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 449 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 450 if (PN && PN->getParent() == BB) { 451 // We can do this simplification if any comparisons fold to true or false. 452 // See if any do. 453 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 454 BasicBlock *PredBB = PN->getIncomingBlock(i); 455 Value *LHS = PN->getIncomingValue(i); 456 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 457 458 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD); 459 if (Res == 0) { 460 if (!isa<Constant>(RHS)) 461 continue; 462 463 LazyValueInfo::Tristate 464 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 465 cast<Constant>(RHS), PredBB, BB); 466 if (ResT == LazyValueInfo::Unknown) 467 continue; 468 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 469 } 470 471 if (Constant *ConstRes = dyn_cast<Constant>(Res)) 472 PushConstantIntOrUndef(Result, ConstRes, PredBB); 473 } 474 475 return !Result.empty(); 476 } 477 478 479 // If comparing a live-in value against a constant, see if we know the 480 // live-in value on any predecessors. 481 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) { 482 if (!isa<Instruction>(Cmp->getOperand(0)) || 483 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 484 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 485 486 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ 487 BasicBlock *P = *PI; 488 // If the value is known by LazyValueInfo to be a constant in a 489 // predecessor, use that information to try to thread this block. 490 LazyValueInfo::Tristate Res = 491 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 492 RHSCst, P, BB); 493 if (Res == LazyValueInfo::Unknown) 494 continue; 495 496 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 497 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P)); 498 } 499 500 return !Result.empty(); 501 } 502 503 // Try to find a constant value for the LHS of a comparison, 504 // and evaluate it statically if we can. 505 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { 506 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals; 507 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals); 508 509 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { 510 Constant *V = LHSVals[i].first; 511 if (V == 0) V = UndefValue::get(CmpConst->getType()); 512 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(), 513 V, CmpConst); 514 PushConstantIntOrUndef(Result, Folded, LHSVals[i].second); 515 } 516 517 return !Result.empty(); 518 } 519 } 520 } 521 522 // If all else fails, see if LVI can figure out a constant value for us. 523 Constant *CI = LVI->getConstant(V, BB); 524 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI); 525 if (CInt) { 526 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 527 Result.push_back(std::make_pair(CInt, *PI)); 528 } 529 530 return !Result.empty(); 531} 532 533 534 535/// GetBestDestForBranchOnUndef - If we determine that the specified block ends 536/// in an undefined jump, decide which block is best to revector to. 537/// 538/// Since we can pick an arbitrary destination, we pick the successor with the 539/// fewest predecessors. This should reduce the in-degree of the others. 540/// 541static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 542 TerminatorInst *BBTerm = BB->getTerminator(); 543 unsigned MinSucc = 0; 544 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 545 // Compute the successor with the minimum number of predecessors. 546 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 547 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 548 TestBB = BBTerm->getSuccessor(i); 549 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 550 if (NumPreds < MinNumPreds) 551 MinSucc = i; 552 } 553 554 return MinSucc; 555} 556 557/// ProcessBlock - If there are any predecessors whose control can be threaded 558/// through to a successor, transform them now. 559bool JumpThreading::ProcessBlock(BasicBlock *BB) { 560 // If the block is trivially dead, just return and let the caller nuke it. 561 // This simplifies other transformations. 562 if (pred_begin(BB) == pred_end(BB) && 563 BB != &BB->getParent()->getEntryBlock()) 564 return false; 565 566 // If this block has a single predecessor, and if that pred has a single 567 // successor, merge the blocks. This encourages recursive jump threading 568 // because now the condition in this block can be threaded through 569 // predecessors of our predecessor block. 570 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 571 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 572 SinglePred != BB) { 573 // If SinglePred was a loop header, BB becomes one. 574 if (LoopHeaders.erase(SinglePred)) 575 LoopHeaders.insert(BB); 576 577 // Remember if SinglePred was the entry block of the function. If so, we 578 // will need to move BB back to the entry position. 579 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 580 LVI->eraseBlock(SinglePred); 581 MergeBasicBlockIntoOnlyPred(BB); 582 583 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 584 BB->moveBefore(&BB->getParent()->getEntryBlock()); 585 return true; 586 } 587 } 588 589 // Look to see if the terminator is a branch of switch, if not we can't thread 590 // it. 591 Value *Condition; 592 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 593 // Can't thread an unconditional jump. 594 if (BI->isUnconditional()) return false; 595 Condition = BI->getCondition(); 596 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 597 Condition = SI->getCondition(); 598 else 599 return false; // Must be an invoke. 600 601 // If the terminator of this block is branching on a constant, simplify the 602 // terminator to an unconditional branch. This can occur due to threading in 603 // other blocks. 604 if (isa<ConstantInt>(Condition)) { 605 DEBUG(dbgs() << " In block '" << BB->getName() 606 << "' folding terminator: " << *BB->getTerminator() << '\n'); 607 ++NumFolds; 608 ConstantFoldTerminator(BB); 609 return true; 610 } 611 612 // If the terminator is branching on an undef, we can pick any of the 613 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 614 if (isa<UndefValue>(Condition)) { 615 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 616 617 // Fold the branch/switch. 618 TerminatorInst *BBTerm = BB->getTerminator(); 619 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 620 if (i == BestSucc) continue; 621 BBTerm->getSuccessor(i)->removePredecessor(BB, true); 622 } 623 624 DEBUG(dbgs() << " In block '" << BB->getName() 625 << "' folding undef terminator: " << *BBTerm << '\n'); 626 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 627 BBTerm->eraseFromParent(); 628 return true; 629 } 630 631 Instruction *CondInst = dyn_cast<Instruction>(Condition); 632 633 // All the rest of our checks depend on the condition being an instruction. 634 if (CondInst == 0) { 635 // FIXME: Unify this with code below. 636 if (ProcessThreadableEdges(Condition, BB)) 637 return true; 638 return false; 639 } 640 641 642 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 643 // For a comparison where the LHS is outside this block, it's possible 644 // that we've branched on it before. Used LVI to see if we can simplify 645 // the branch based on that. 646 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 647 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 648 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 649 if (CondBr && CondConst && CondBr->isConditional() && PI != PE && 650 (!isa<Instruction>(CondCmp->getOperand(0)) || 651 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { 652 // For predecessor edge, determine if the comparison is true or false 653 // on that edge. If they're all true or all false, we can simplify the 654 // branch. 655 // FIXME: We could handle mixed true/false by duplicating code. 656 LazyValueInfo::Tristate Baseline = 657 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0), 658 CondConst, *PI, BB); 659 if (Baseline != LazyValueInfo::Unknown) { 660 // Check that all remaining incoming values match the first one. 661 while (++PI != PE) { 662 LazyValueInfo::Tristate Ret = 663 LVI->getPredicateOnEdge(CondCmp->getPredicate(), 664 CondCmp->getOperand(0), CondConst, *PI, BB); 665 if (Ret != Baseline) break; 666 } 667 668 // If we terminated early, then one of the values didn't match. 669 if (PI == PE) { 670 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0; 671 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1; 672 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); 673 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); 674 CondBr->eraseFromParent(); 675 return true; 676 } 677 } 678 } 679 } 680 681 // Check for some cases that are worth simplifying. Right now we want to look 682 // for loads that are used by a switch or by the condition for the branch. If 683 // we see one, check to see if it's partially redundant. If so, insert a PHI 684 // which can then be used to thread the values. 685 // 686 Value *SimplifyValue = CondInst; 687 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 688 if (isa<Constant>(CondCmp->getOperand(1))) 689 SimplifyValue = CondCmp->getOperand(0); 690 691 // TODO: There are other places where load PRE would be profitable, such as 692 // more complex comparisons. 693 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 694 if (SimplifyPartiallyRedundantLoad(LI)) 695 return true; 696 697 698 // Handle a variety of cases where we are branching on something derived from 699 // a PHI node in the current block. If we can prove that any predecessors 700 // compute a predictable value based on a PHI node, thread those predecessors. 701 // 702 if (ProcessThreadableEdges(CondInst, BB)) 703 return true; 704 705 // If this is an otherwise-unfoldable branch on a phi node in the current 706 // block, see if we can simplify. 707 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 708 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 709 return ProcessBranchOnPHI(PN); 710 711 712 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 713 if (CondInst->getOpcode() == Instruction::Xor && 714 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 715 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 716 717 718 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 719 // "(X == 4)", thread through this block. 720 721 return false; 722} 723 724/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 725/// block that jump on exactly the same condition. This means that we almost 726/// always know the direction of the edge in the DESTBB: 727/// PREDBB: 728/// br COND, DESTBB, BBY 729/// DESTBB: 730/// br COND, BBZ, BBW 731/// 732/// If DESTBB has multiple predecessors, we can't just constant fold the branch 733/// in DESTBB, we have to thread over it. 734bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 735 BasicBlock *BB) { 736 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 737 738 // If both successors of PredBB go to DESTBB, we don't know anything. We can 739 // fold the branch to an unconditional one, which allows other recursive 740 // simplifications. 741 bool BranchDir; 742 if (PredBI->getSuccessor(1) != BB) 743 BranchDir = true; 744 else if (PredBI->getSuccessor(0) != BB) 745 BranchDir = false; 746 else { 747 DEBUG(dbgs() << " In block '" << PredBB->getName() 748 << "' folding terminator: " << *PredBB->getTerminator() << '\n'); 749 ++NumFolds; 750 ConstantFoldTerminator(PredBB); 751 return true; 752 } 753 754 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 755 756 // If the dest block has one predecessor, just fix the branch condition to a 757 // constant and fold it. 758 if (BB->getSinglePredecessor()) { 759 DEBUG(dbgs() << " In block '" << BB->getName() 760 << "' folding condition to '" << BranchDir << "': " 761 << *BB->getTerminator() << '\n'); 762 ++NumFolds; 763 Value *OldCond = DestBI->getCondition(); 764 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()), 765 BranchDir)); 766 // Delete dead instructions before we fold the branch. Folding the branch 767 // can eliminate edges from the CFG which can end up deleting OldCond. 768 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 769 ConstantFoldTerminator(BB); 770 return true; 771 } 772 773 774 // Next, figure out which successor we are threading to. 775 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 776 777 SmallVector<BasicBlock*, 2> Preds; 778 Preds.push_back(PredBB); 779 780 // Ok, try to thread it! 781 return ThreadEdge(BB, Preds, SuccBB); 782} 783 784/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 785/// block that switch on exactly the same condition. This means that we almost 786/// always know the direction of the edge in the DESTBB: 787/// PREDBB: 788/// switch COND [... DESTBB, BBY ... ] 789/// DESTBB: 790/// switch COND [... BBZ, BBW ] 791/// 792/// Optimizing switches like this is very important, because simplifycfg builds 793/// switches out of repeated 'if' conditions. 794bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 795 BasicBlock *DestBB) { 796 // Can't thread edge to self. 797 if (PredBB == DestBB) 798 return false; 799 800 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 801 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 802 803 // There are a variety of optimizations that we can potentially do on these 804 // blocks: we order them from most to least preferable. 805 806 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 807 // directly to their destination. This does not introduce *any* code size 808 // growth. Skip debug info first. 809 BasicBlock::iterator BBI = DestBB->begin(); 810 while (isa<DbgInfoIntrinsic>(BBI)) 811 BBI++; 812 813 // FIXME: Thread if it just contains a PHI. 814 if (isa<SwitchInst>(BBI)) { 815 bool MadeChange = false; 816 // Ignore the default edge for now. 817 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 818 ConstantInt *DestVal = DestSI->getCaseValue(i); 819 BasicBlock *DestSucc = DestSI->getSuccessor(i); 820 821 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 822 // PredSI has an explicit case for it. If so, forward. If it is covered 823 // by the default case, we can't update PredSI. 824 unsigned PredCase = PredSI->findCaseValue(DestVal); 825 if (PredCase == 0) continue; 826 827 // If PredSI doesn't go to DestBB on this value, then it won't reach the 828 // case on this condition. 829 if (PredSI->getSuccessor(PredCase) != DestBB && 830 DestSI->getSuccessor(i) != DestBB) 831 continue; 832 833 // Do not forward this if it already goes to this destination, this would 834 // be an infinite loop. 835 if (PredSI->getSuccessor(PredCase) == DestSucc) 836 continue; 837 838 // Otherwise, we're safe to make the change. Make sure that the edge from 839 // DestSI to DestSucc is not critical and has no PHI nodes. 840 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI); 841 DEBUG(dbgs() << "THROUGH: " << *DestSI); 842 843 // If the destination has PHI nodes, just split the edge for updating 844 // simplicity. 845 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 846 SplitCriticalEdge(DestSI, i, this); 847 DestSucc = DestSI->getSuccessor(i); 848 } 849 FoldSingleEntryPHINodes(DestSucc); 850 PredSI->setSuccessor(PredCase, DestSucc); 851 MadeChange = true; 852 } 853 854 if (MadeChange) 855 return true; 856 } 857 858 return false; 859} 860 861 862/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 863/// load instruction, eliminate it by replacing it with a PHI node. This is an 864/// important optimization that encourages jump threading, and needs to be run 865/// interlaced with other jump threading tasks. 866bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 867 // Don't hack volatile loads. 868 if (LI->isVolatile()) return false; 869 870 // If the load is defined in a block with exactly one predecessor, it can't be 871 // partially redundant. 872 BasicBlock *LoadBB = LI->getParent(); 873 if (LoadBB->getSinglePredecessor()) 874 return false; 875 876 Value *LoadedPtr = LI->getOperand(0); 877 878 // If the loaded operand is defined in the LoadBB, it can't be available. 879 // TODO: Could do simple PHI translation, that would be fun :) 880 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 881 if (PtrOp->getParent() == LoadBB) 882 return false; 883 884 // Scan a few instructions up from the load, to see if it is obviously live at 885 // the entry to its block. 886 BasicBlock::iterator BBIt = LI; 887 888 if (Value *AvailableVal = 889 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { 890 // If the value if the load is locally available within the block, just use 891 // it. This frequently occurs for reg2mem'd allocas. 892 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 893 894 // If the returned value is the load itself, replace with an undef. This can 895 // only happen in dead loops. 896 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 897 LI->replaceAllUsesWith(AvailableVal); 898 LI->eraseFromParent(); 899 return true; 900 } 901 902 // Otherwise, if we scanned the whole block and got to the top of the block, 903 // we know the block is locally transparent to the load. If not, something 904 // might clobber its value. 905 if (BBIt != LoadBB->begin()) 906 return false; 907 908 909 SmallPtrSet<BasicBlock*, 8> PredsScanned; 910 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 911 AvailablePredsTy AvailablePreds; 912 BasicBlock *OneUnavailablePred = 0; 913 914 // If we got here, the loaded value is transparent through to the start of the 915 // block. Check to see if it is available in any of the predecessor blocks. 916 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 917 PI != PE; ++PI) { 918 BasicBlock *PredBB = *PI; 919 920 // If we already scanned this predecessor, skip it. 921 if (!PredsScanned.insert(PredBB)) 922 continue; 923 924 // Scan the predecessor to see if the value is available in the pred. 925 BBIt = PredBB->end(); 926 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 927 if (!PredAvailable) { 928 OneUnavailablePred = PredBB; 929 continue; 930 } 931 932 // If so, this load is partially redundant. Remember this info so that we 933 // can create a PHI node. 934 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 935 } 936 937 // If the loaded value isn't available in any predecessor, it isn't partially 938 // redundant. 939 if (AvailablePreds.empty()) return false; 940 941 // Okay, the loaded value is available in at least one (and maybe all!) 942 // predecessors. If the value is unavailable in more than one unique 943 // predecessor, we want to insert a merge block for those common predecessors. 944 // This ensures that we only have to insert one reload, thus not increasing 945 // code size. 946 BasicBlock *UnavailablePred = 0; 947 948 // If there is exactly one predecessor where the value is unavailable, the 949 // already computed 'OneUnavailablePred' block is it. If it ends in an 950 // unconditional branch, we know that it isn't a critical edge. 951 if (PredsScanned.size() == AvailablePreds.size()+1 && 952 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 953 UnavailablePred = OneUnavailablePred; 954 } else if (PredsScanned.size() != AvailablePreds.size()) { 955 // Otherwise, we had multiple unavailable predecessors or we had a critical 956 // edge from the one. 957 SmallVector<BasicBlock*, 8> PredsToSplit; 958 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 959 960 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 961 AvailablePredSet.insert(AvailablePreds[i].first); 962 963 // Add all the unavailable predecessors to the PredsToSplit list. 964 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 965 PI != PE; ++PI) { 966 BasicBlock *P = *PI; 967 // If the predecessor is an indirect goto, we can't split the edge. 968 if (isa<IndirectBrInst>(P->getTerminator())) 969 return false; 970 971 if (!AvailablePredSet.count(P)) 972 PredsToSplit.push_back(P); 973 } 974 975 // Split them out to their own block. 976 UnavailablePred = 977 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 978 "thread-pre-split", this); 979 } 980 981 // If the value isn't available in all predecessors, then there will be 982 // exactly one where it isn't available. Insert a load on that edge and add 983 // it to the AvailablePreds list. 984 if (UnavailablePred) { 985 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 986 "Can't handle critical edge here!"); 987 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 988 LI->getAlignment(), 989 UnavailablePred->getTerminator()); 990 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 991 } 992 993 // Now we know that each predecessor of this block has a value in 994 // AvailablePreds, sort them for efficient access as we're walking the preds. 995 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 996 997 // Create a PHI node at the start of the block for the PRE'd load value. 998 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 999 PN->takeName(LI); 1000 1001 // Insert new entries into the PHI for each predecessor. A single block may 1002 // have multiple entries here. 1003 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 1004 ++PI) { 1005 BasicBlock *P = *PI; 1006 AvailablePredsTy::iterator I = 1007 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 1008 std::make_pair(P, (Value*)0)); 1009 1010 assert(I != AvailablePreds.end() && I->first == P && 1011 "Didn't find entry for predecessor!"); 1012 1013 PN->addIncoming(I->second, I->first); 1014 } 1015 1016 //cerr << "PRE: " << *LI << *PN << "\n"; 1017 1018 LI->replaceAllUsesWith(PN); 1019 LI->eraseFromParent(); 1020 1021 return true; 1022} 1023 1024/// FindMostPopularDest - The specified list contains multiple possible 1025/// threadable destinations. Pick the one that occurs the most frequently in 1026/// the list. 1027static BasicBlock * 1028FindMostPopularDest(BasicBlock *BB, 1029 const SmallVectorImpl<std::pair<BasicBlock*, 1030 BasicBlock*> > &PredToDestList) { 1031 assert(!PredToDestList.empty()); 1032 1033 // Determine popularity. If there are multiple possible destinations, we 1034 // explicitly choose to ignore 'undef' destinations. We prefer to thread 1035 // blocks with known and real destinations to threading undef. We'll handle 1036 // them later if interesting. 1037 DenseMap<BasicBlock*, unsigned> DestPopularity; 1038 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1039 if (PredToDestList[i].second) 1040 DestPopularity[PredToDestList[i].second]++; 1041 1042 // Find the most popular dest. 1043 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 1044 BasicBlock *MostPopularDest = DPI->first; 1045 unsigned Popularity = DPI->second; 1046 SmallVector<BasicBlock*, 4> SamePopularity; 1047 1048 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 1049 // If the popularity of this entry isn't higher than the popularity we've 1050 // seen so far, ignore it. 1051 if (DPI->second < Popularity) 1052 ; // ignore. 1053 else if (DPI->second == Popularity) { 1054 // If it is the same as what we've seen so far, keep track of it. 1055 SamePopularity.push_back(DPI->first); 1056 } else { 1057 // If it is more popular, remember it. 1058 SamePopularity.clear(); 1059 MostPopularDest = DPI->first; 1060 Popularity = DPI->second; 1061 } 1062 } 1063 1064 // Okay, now we know the most popular destination. If there is more than 1065 // destination, we need to determine one. This is arbitrary, but we need 1066 // to make a deterministic decision. Pick the first one that appears in the 1067 // successor list. 1068 if (!SamePopularity.empty()) { 1069 SamePopularity.push_back(MostPopularDest); 1070 TerminatorInst *TI = BB->getTerminator(); 1071 for (unsigned i = 0; ; ++i) { 1072 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1073 1074 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1075 TI->getSuccessor(i)) == SamePopularity.end()) 1076 continue; 1077 1078 MostPopularDest = TI->getSuccessor(i); 1079 break; 1080 } 1081 } 1082 1083 // Okay, we have finally picked the most popular destination. 1084 return MostPopularDest; 1085} 1086 1087bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) { 1088 // If threading this would thread across a loop header, don't even try to 1089 // thread the edge. 1090 if (LoopHeaders.count(BB)) 1091 return false; 1092 1093 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues; 1094 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) 1095 return false; 1096 1097 assert(!PredValues.empty() && 1098 "ComputeValueKnownInPredecessors returned true with no values"); 1099 1100 DEBUG(dbgs() << "IN BB: " << *BB; 1101 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1102 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "; 1103 if (PredValues[i].first) 1104 dbgs() << *PredValues[i].first; 1105 else 1106 dbgs() << "UNDEF"; 1107 dbgs() << " for pred '" << PredValues[i].second->getName() 1108 << "'.\n"; 1109 }); 1110 1111 // Decide what we want to thread through. Convert our list of known values to 1112 // a list of known destinations for each pred. This also discards duplicate 1113 // predecessors and keeps track of the undefined inputs (which are represented 1114 // as a null dest in the PredToDestList). 1115 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1116 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1117 1118 BasicBlock *OnlyDest = 0; 1119 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1120 1121 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { 1122 BasicBlock *Pred = PredValues[i].second; 1123 if (!SeenPreds.insert(Pred)) 1124 continue; // Duplicate predecessor entry. 1125 1126 // If the predecessor ends with an indirect goto, we can't change its 1127 // destination. 1128 if (isa<IndirectBrInst>(Pred->getTerminator())) 1129 continue; 1130 1131 ConstantInt *Val = PredValues[i].first; 1132 1133 BasicBlock *DestBB; 1134 if (Val == 0) // Undef. 1135 DestBB = 0; 1136 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1137 DestBB = BI->getSuccessor(Val->isZero()); 1138 else { 1139 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 1140 DestBB = SI->getSuccessor(SI->findCaseValue(Val)); 1141 } 1142 1143 // If we have exactly one destination, remember it for efficiency below. 1144 if (i == 0) 1145 OnlyDest = DestBB; 1146 else if (OnlyDest != DestBB) 1147 OnlyDest = MultipleDestSentinel; 1148 1149 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1150 } 1151 1152 // If all edges were unthreadable, we fail. 1153 if (PredToDestList.empty()) 1154 return false; 1155 1156 // Determine which is the most common successor. If we have many inputs and 1157 // this block is a switch, we want to start by threading the batch that goes 1158 // to the most popular destination first. If we only know about one 1159 // threadable destination (the common case) we can avoid this. 1160 BasicBlock *MostPopularDest = OnlyDest; 1161 1162 if (MostPopularDest == MultipleDestSentinel) 1163 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1164 1165 // Now that we know what the most popular destination is, factor all 1166 // predecessors that will jump to it into a single predecessor. 1167 SmallVector<BasicBlock*, 16> PredsToFactor; 1168 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) 1169 if (PredToDestList[i].second == MostPopularDest) { 1170 BasicBlock *Pred = PredToDestList[i].first; 1171 1172 // This predecessor may be a switch or something else that has multiple 1173 // edges to the block. Factor each of these edges by listing them 1174 // according to # occurrences in PredsToFactor. 1175 TerminatorInst *PredTI = Pred->getTerminator(); 1176 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) 1177 if (PredTI->getSuccessor(i) == BB) 1178 PredsToFactor.push_back(Pred); 1179 } 1180 1181 // If the threadable edges are branching on an undefined value, we get to pick 1182 // the destination that these predecessors should get to. 1183 if (MostPopularDest == 0) 1184 MostPopularDest = BB->getTerminator()-> 1185 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1186 1187 // Ok, try to thread it! 1188 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1189} 1190 1191/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1192/// a PHI node in the current block. See if there are any simplifications we 1193/// can do based on inputs to the phi node. 1194/// 1195bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1196 BasicBlock *BB = PN->getParent(); 1197 1198 // TODO: We could make use of this to do it once for blocks with common PHI 1199 // values. 1200 SmallVector<BasicBlock*, 1> PredBBs; 1201 PredBBs.resize(1); 1202 1203 // If any of the predecessor blocks end in an unconditional branch, we can 1204 // *duplicate* the conditional branch into that block in order to further 1205 // encourage jump threading and to eliminate cases where we have branch on a 1206 // phi of an icmp (branch on icmp is much better). 1207 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1208 BasicBlock *PredBB = PN->getIncomingBlock(i); 1209 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1210 if (PredBr->isUnconditional()) { 1211 PredBBs[0] = PredBB; 1212 // Try to duplicate BB into PredBB. 1213 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1214 return true; 1215 } 1216 } 1217 1218 return false; 1219} 1220 1221/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1222/// a xor instruction in the current block. See if there are any 1223/// simplifications we can do based on inputs to the xor. 1224/// 1225bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1226 BasicBlock *BB = BO->getParent(); 1227 1228 // If either the LHS or RHS of the xor is a constant, don't do this 1229 // optimization. 1230 if (isa<ConstantInt>(BO->getOperand(0)) || 1231 isa<ConstantInt>(BO->getOperand(1))) 1232 return false; 1233 1234 // If the first instruction in BB isn't a phi, we won't be able to infer 1235 // anything special about any particular predecessor. 1236 if (!isa<PHINode>(BB->front())) 1237 return false; 1238 1239 // If we have a xor as the branch input to this block, and we know that the 1240 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1241 // the condition into the predecessor and fix that value to true, saving some 1242 // logical ops on that path and encouraging other paths to simplify. 1243 // 1244 // This copies something like this: 1245 // 1246 // BB: 1247 // %X = phi i1 [1], [%X'] 1248 // %Y = icmp eq i32 %A, %B 1249 // %Z = xor i1 %X, %Y 1250 // br i1 %Z, ... 1251 // 1252 // Into: 1253 // BB': 1254 // %Y = icmp ne i32 %A, %B 1255 // br i1 %Z, ... 1256 1257 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues; 1258 bool isLHS = true; 1259 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) { 1260 assert(XorOpValues.empty()); 1261 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues)) 1262 return false; 1263 isLHS = false; 1264 } 1265 1266 assert(!XorOpValues.empty() && 1267 "ComputeValueKnownInPredecessors returned true with no values"); 1268 1269 // Scan the information to see which is most popular: true or false. The 1270 // predecessors can be of the set true, false, or undef. 1271 unsigned NumTrue = 0, NumFalse = 0; 1272 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1273 if (!XorOpValues[i].first) continue; // Ignore undefs for the count. 1274 if (XorOpValues[i].first->isZero()) 1275 ++NumFalse; 1276 else 1277 ++NumTrue; 1278 } 1279 1280 // Determine which value to split on, true, false, or undef if neither. 1281 ConstantInt *SplitVal = 0; 1282 if (NumTrue > NumFalse) 1283 SplitVal = ConstantInt::getTrue(BB->getContext()); 1284 else if (NumTrue != 0 || NumFalse != 0) 1285 SplitVal = ConstantInt::getFalse(BB->getContext()); 1286 1287 // Collect all of the blocks that this can be folded into so that we can 1288 // factor this once and clone it once. 1289 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1290 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { 1291 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue; 1292 1293 BlocksToFoldInto.push_back(XorOpValues[i].second); 1294 } 1295 1296 // If we inferred a value for all of the predecessors, then duplication won't 1297 // help us. However, we can just replace the LHS or RHS with the constant. 1298 if (BlocksToFoldInto.size() == 1299 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1300 if (SplitVal == 0) { 1301 // If all preds provide undef, just nuke the xor, because it is undef too. 1302 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1303 BO->eraseFromParent(); 1304 } else if (SplitVal->isZero()) { 1305 // If all preds provide 0, replace the xor with the other input. 1306 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1307 BO->eraseFromParent(); 1308 } else { 1309 // If all preds provide 1, set the computed value to 1. 1310 BO->setOperand(!isLHS, SplitVal); 1311 } 1312 1313 return true; 1314 } 1315 1316 // Try to duplicate BB into PredBB. 1317 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1318} 1319 1320 1321/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1322/// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1323/// NewPred using the entries from OldPred (suitably mapped). 1324static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1325 BasicBlock *OldPred, 1326 BasicBlock *NewPred, 1327 DenseMap<Instruction*, Value*> &ValueMap) { 1328 for (BasicBlock::iterator PNI = PHIBB->begin(); 1329 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1330 // Ok, we have a PHI node. Figure out what the incoming value was for the 1331 // DestBlock. 1332 Value *IV = PN->getIncomingValueForBlock(OldPred); 1333 1334 // Remap the value if necessary. 1335 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1336 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1337 if (I != ValueMap.end()) 1338 IV = I->second; 1339 } 1340 1341 PN->addIncoming(IV, NewPred); 1342 } 1343} 1344 1345/// ThreadEdge - We have decided that it is safe and profitable to factor the 1346/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1347/// across BB. Transform the IR to reflect this change. 1348bool JumpThreading::ThreadEdge(BasicBlock *BB, 1349 const SmallVectorImpl<BasicBlock*> &PredBBs, 1350 BasicBlock *SuccBB) { 1351 // If threading to the same block as we come from, we would infinite loop. 1352 if (SuccBB == BB) { 1353 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1354 << "' - would thread to self!\n"); 1355 return false; 1356 } 1357 1358 // If threading this would thread across a loop header, don't thread the edge. 1359 // See the comments above FindLoopHeaders for justifications and caveats. 1360 if (LoopHeaders.count(BB)) { 1361 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1362 << "' to dest BB '" << SuccBB->getName() 1363 << "' - it might create an irreducible loop!\n"); 1364 return false; 1365 } 1366 1367 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 1368 if (JumpThreadCost > Threshold) { 1369 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1370 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1371 return false; 1372 } 1373 1374 // And finally, do it! Start by factoring the predecessors is needed. 1375 BasicBlock *PredBB; 1376 if (PredBBs.size() == 1) 1377 PredBB = PredBBs[0]; 1378 else { 1379 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1380 << " common predecessors.\n"); 1381 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1382 ".thr_comm", this); 1383 } 1384 1385 // And finally, do it! 1386 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1387 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1388 << ", across block:\n " 1389 << *BB << "\n"); 1390 1391 LVI->threadEdge(PredBB, BB, SuccBB); 1392 1393 // We are going to have to map operands from the original BB block to the new 1394 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1395 // account for entry from PredBB. 1396 DenseMap<Instruction*, Value*> ValueMapping; 1397 1398 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1399 BB->getName()+".thread", 1400 BB->getParent(), BB); 1401 NewBB->moveAfter(PredBB); 1402 1403 BasicBlock::iterator BI = BB->begin(); 1404 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1405 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1406 1407 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1408 // mapping and using it to remap operands in the cloned instructions. 1409 for (; !isa<TerminatorInst>(BI); ++BI) { 1410 Instruction *New = BI->clone(); 1411 New->setName(BI->getName()); 1412 NewBB->getInstList().push_back(New); 1413 ValueMapping[BI] = New; 1414 1415 // Remap operands to patch up intra-block references. 1416 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1417 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1418 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1419 if (I != ValueMapping.end()) 1420 New->setOperand(i, I->second); 1421 } 1422 } 1423 1424 // We didn't copy the terminator from BB over to NewBB, because there is now 1425 // an unconditional jump to SuccBB. Insert the unconditional jump. 1426 BranchInst::Create(SuccBB, NewBB); 1427 1428 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1429 // PHI nodes for NewBB now. 1430 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1431 1432 // If there were values defined in BB that are used outside the block, then we 1433 // now have to update all uses of the value to use either the original value, 1434 // the cloned value, or some PHI derived value. This can require arbitrary 1435 // PHI insertion, of which we are prepared to do, clean these up now. 1436 SSAUpdater SSAUpdate; 1437 SmallVector<Use*, 16> UsesToRename; 1438 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1439 // Scan all uses of this instruction to see if it is used outside of its 1440 // block, and if so, record them in UsesToRename. 1441 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1442 ++UI) { 1443 Instruction *User = cast<Instruction>(*UI); 1444 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1445 if (UserPN->getIncomingBlock(UI) == BB) 1446 continue; 1447 } else if (User->getParent() == BB) 1448 continue; 1449 1450 UsesToRename.push_back(&UI.getUse()); 1451 } 1452 1453 // If there are no uses outside the block, we're done with this instruction. 1454 if (UsesToRename.empty()) 1455 continue; 1456 1457 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1458 1459 // We found a use of I outside of BB. Rename all uses of I that are outside 1460 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1461 // with the two values we know. 1462 SSAUpdate.Initialize(I->getType(), I->getName()); 1463 SSAUpdate.AddAvailableValue(BB, I); 1464 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); 1465 1466 while (!UsesToRename.empty()) 1467 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1468 DEBUG(dbgs() << "\n"); 1469 } 1470 1471 1472 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1473 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1474 // us to simplify any PHI nodes in BB. 1475 TerminatorInst *PredTerm = PredBB->getTerminator(); 1476 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1477 if (PredTerm->getSuccessor(i) == BB) { 1478 BB->removePredecessor(PredBB, true); 1479 PredTerm->setSuccessor(i, NewBB); 1480 } 1481 1482 // At this point, the IR is fully up to date and consistent. Do a quick scan 1483 // over the new instructions and zap any that are constants or dead. This 1484 // frequently happens because of phi translation. 1485 SimplifyInstructionsInBlock(NewBB, TD); 1486 1487 // Threaded an edge! 1488 ++NumThreads; 1489 return true; 1490} 1491 1492/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1493/// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1494/// If we can duplicate the contents of BB up into PredBB do so now, this 1495/// improves the odds that the branch will be on an analyzable instruction like 1496/// a compare. 1497bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1498 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1499 assert(!PredBBs.empty() && "Can't handle an empty set"); 1500 1501 // If BB is a loop header, then duplicating this block outside the loop would 1502 // cause us to transform this into an irreducible loop, don't do this. 1503 // See the comments above FindLoopHeaders for justifications and caveats. 1504 if (LoopHeaders.count(BB)) { 1505 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1506 << "' into predecessor block '" << PredBBs[0]->getName() 1507 << "' - it might create an irreducible loop!\n"); 1508 return false; 1509 } 1510 1511 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB); 1512 if (DuplicationCost > Threshold) { 1513 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1514 << "' - Cost is too high: " << DuplicationCost << "\n"); 1515 return false; 1516 } 1517 1518 // And finally, do it! Start by factoring the predecessors is needed. 1519 BasicBlock *PredBB; 1520 if (PredBBs.size() == 1) 1521 PredBB = PredBBs[0]; 1522 else { 1523 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1524 << " common predecessors.\n"); 1525 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(), 1526 ".thr_comm", this); 1527 } 1528 1529 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1530 // of PredBB. 1531 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1532 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1533 << DuplicationCost << " block is:" << *BB << "\n"); 1534 1535 // Unless PredBB ends with an unconditional branch, split the edge so that we 1536 // can just clone the bits from BB into the end of the new PredBB. 1537 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1538 1539 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) { 1540 PredBB = SplitEdge(PredBB, BB, this); 1541 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1542 } 1543 1544 // We are going to have to map operands from the original BB block into the 1545 // PredBB block. Evaluate PHI nodes in BB. 1546 DenseMap<Instruction*, Value*> ValueMapping; 1547 1548 BasicBlock::iterator BI = BB->begin(); 1549 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1550 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1551 1552 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1553 // mapping and using it to remap operands in the cloned instructions. 1554 for (; BI != BB->end(); ++BI) { 1555 Instruction *New = BI->clone(); 1556 1557 // Remap operands to patch up intra-block references. 1558 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1559 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1560 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1561 if (I != ValueMapping.end()) 1562 New->setOperand(i, I->second); 1563 } 1564 1565 // If this instruction can be simplified after the operands are updated, 1566 // just use the simplified value instead. This frequently happens due to 1567 // phi translation. 1568 if (Value *IV = SimplifyInstruction(New, TD)) { 1569 delete New; 1570 ValueMapping[BI] = IV; 1571 } else { 1572 // Otherwise, insert the new instruction into the block. 1573 New->setName(BI->getName()); 1574 PredBB->getInstList().insert(OldPredBranch, New); 1575 ValueMapping[BI] = New; 1576 } 1577 } 1578 1579 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1580 // add entries to the PHI nodes for branch from PredBB now. 1581 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1582 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1583 ValueMapping); 1584 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1585 ValueMapping); 1586 1587 // If there were values defined in BB that are used outside the block, then we 1588 // now have to update all uses of the value to use either the original value, 1589 // the cloned value, or some PHI derived value. This can require arbitrary 1590 // PHI insertion, of which we are prepared to do, clean these up now. 1591 SSAUpdater SSAUpdate; 1592 SmallVector<Use*, 16> UsesToRename; 1593 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 1594 // Scan all uses of this instruction to see if it is used outside of its 1595 // block, and if so, record them in UsesToRename. 1596 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 1597 ++UI) { 1598 Instruction *User = cast<Instruction>(*UI); 1599 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1600 if (UserPN->getIncomingBlock(UI) == BB) 1601 continue; 1602 } else if (User->getParent() == BB) 1603 continue; 1604 1605 UsesToRename.push_back(&UI.getUse()); 1606 } 1607 1608 // If there are no uses outside the block, we're done with this instruction. 1609 if (UsesToRename.empty()) 1610 continue; 1611 1612 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"); 1613 1614 // We found a use of I outside of BB. Rename all uses of I that are outside 1615 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1616 // with the two values we know. 1617 SSAUpdate.Initialize(I->getType(), I->getName()); 1618 SSAUpdate.AddAvailableValue(BB, I); 1619 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); 1620 1621 while (!UsesToRename.empty()) 1622 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1623 DEBUG(dbgs() << "\n"); 1624 } 1625 1626 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1627 // that we nuked. 1628 BB->removePredecessor(PredBB, true); 1629 1630 // Remove the unconditional branch at the end of the PredBB block. 1631 OldPredBranch->eraseFromParent(); 1632 1633 ++NumDupes; 1634 return true; 1635} 1636 1637 1638