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 #include "llvm/Transforms/Scalar.h" 15 #include "llvm/ADT/DenseMap.h" 16 #include "llvm/ADT/DenseSet.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/SmallPtrSet.h" 19 #include "llvm/ADT/SmallSet.h" 20 #include "llvm/ADT/Statistic.h" 21 #include "llvm/Analysis/GlobalsModRef.h" 22 #include "llvm/Analysis/CFG.h" 23 #include "llvm/Analysis/BlockFrequencyInfo.h" 24 #include "llvm/Analysis/BlockFrequencyInfoImpl.h" 25 #include "llvm/Analysis/BranchProbabilityInfo.h" 26 #include "llvm/Analysis/ConstantFolding.h" 27 #include "llvm/Analysis/InstructionSimplify.h" 28 #include "llvm/Analysis/LazyValueInfo.h" 29 #include "llvm/Analysis/Loads.h" 30 #include "llvm/Analysis/LoopInfo.h" 31 #include "llvm/Analysis/TargetLibraryInfo.h" 32 #include "llvm/Analysis/ValueTracking.h" 33 #include "llvm/IR/DataLayout.h" 34 #include "llvm/IR/IntrinsicInst.h" 35 #include "llvm/IR/LLVMContext.h" 36 #include "llvm/IR/MDBuilder.h" 37 #include "llvm/IR/Metadata.h" 38 #include "llvm/IR/ValueHandle.h" 39 #include "llvm/Pass.h" 40 #include "llvm/Support/CommandLine.h" 41 #include "llvm/Support/Debug.h" 42 #include "llvm/Support/raw_ostream.h" 43 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 44 #include "llvm/Transforms/Utils/Local.h" 45 #include "llvm/Transforms/Utils/SSAUpdater.h" 46 #include <algorithm> 47 #include <memory> 48 using namespace llvm; 49 50 #define DEBUG_TYPE "jump-threading" 51 52 STATISTIC(NumThreads, "Number of jumps threaded"); 53 STATISTIC(NumFolds, "Number of terminators folded"); 54 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 55 56 static cl::opt<unsigned> 57 BBDuplicateThreshold("jump-threading-threshold", 58 cl::desc("Max block size to duplicate for jump threading"), 59 cl::init(6), cl::Hidden); 60 61 static cl::opt<unsigned> 62 ImplicationSearchThreshold( 63 "jump-threading-implication-search-threshold", 64 cl::desc("The number of predecessors to search for a stronger " 65 "condition to use to thread over a weaker condition"), 66 cl::init(3), cl::Hidden); 67 68 namespace { 69 // These are at global scope so static functions can use them too. 70 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo; 71 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy; 72 73 // This is used to keep track of what kind of constant we're currently hoping 74 // to find. 75 enum ConstantPreference { 76 WantInteger, 77 WantBlockAddress 78 }; 79 80 /// This pass performs 'jump threading', which looks at blocks that have 81 /// multiple predecessors and multiple successors. If one or more of the 82 /// predecessors of the block can be proven to always jump to one of the 83 /// successors, we forward the edge from the predecessor to the successor by 84 /// duplicating the contents of this block. 85 /// 86 /// An example of when this can occur is code like this: 87 /// 88 /// if () { ... 89 /// X = 4; 90 /// } 91 /// if (X < 3) { 92 /// 93 /// In this case, the unconditional branch at the end of the first if can be 94 /// revectored to the false side of the second if. 95 /// 96 class JumpThreading : public FunctionPass { 97 TargetLibraryInfo *TLI; 98 LazyValueInfo *LVI; 99 std::unique_ptr<BlockFrequencyInfo> BFI; 100 std::unique_ptr<BranchProbabilityInfo> BPI; 101 bool HasProfileData; 102 #ifdef NDEBUG 103 SmallPtrSet<const BasicBlock *, 16> LoopHeaders; 104 #else 105 SmallSet<AssertingVH<const BasicBlock>, 16> LoopHeaders; 106 #endif 107 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet; 108 109 unsigned BBDupThreshold; 110 111 // RAII helper for updating the recursion stack. 112 struct RecursionSetRemover { 113 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet; 114 std::pair<Value*, BasicBlock*> ThePair; 115 116 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S, 117 std::pair<Value*, BasicBlock*> P) 118 : TheSet(S), ThePair(P) { } 119 120 ~RecursionSetRemover() { 121 TheSet.erase(ThePair); 122 } 123 }; 124 public: 125 static char ID; // Pass identification 126 JumpThreading(int T = -1) : FunctionPass(ID) { 127 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T); 128 initializeJumpThreadingPass(*PassRegistry::getPassRegistry()); 129 } 130 131 bool runOnFunction(Function &F) override; 132 133 void getAnalysisUsage(AnalysisUsage &AU) const override { 134 AU.addRequired<LazyValueInfo>(); 135 AU.addPreserved<LazyValueInfo>(); 136 AU.addPreserved<GlobalsAAWrapperPass>(); 137 AU.addRequired<TargetLibraryInfoWrapperPass>(); 138 } 139 140 void releaseMemory() override { 141 BFI.reset(); 142 BPI.reset(); 143 } 144 145 void FindLoopHeaders(Function &F); 146 bool ProcessBlock(BasicBlock *BB); 147 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, 148 BasicBlock *SuccBB); 149 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 150 const SmallVectorImpl<BasicBlock *> &PredBBs); 151 152 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, 153 PredValueInfo &Result, 154 ConstantPreference Preference, 155 Instruction *CxtI = nullptr); 156 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 157 ConstantPreference Preference, 158 Instruction *CxtI = nullptr); 159 160 bool ProcessBranchOnPHI(PHINode *PN); 161 bool ProcessBranchOnXOR(BinaryOperator *BO); 162 bool ProcessImpliedCondition(BasicBlock *BB); 163 164 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 165 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB); 166 bool TryToUnfoldSelectInCurrBB(BasicBlock *BB); 167 168 private: 169 BasicBlock *SplitBlockPreds(BasicBlock *BB, ArrayRef<BasicBlock *> Preds, 170 const char *Suffix); 171 void UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB, BasicBlock *BB, 172 BasicBlock *NewBB, BasicBlock *SuccBB); 173 }; 174 } 175 176 char JumpThreading::ID = 0; 177 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", 178 "Jump Threading", false, false) 179 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo) 180 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 181 INITIALIZE_PASS_END(JumpThreading, "jump-threading", 182 "Jump Threading", false, false) 183 184 // Public interface to the Jump Threading pass 185 FunctionPass *llvm::createJumpThreadingPass(int Threshold) { return new JumpThreading(Threshold); } 186 187 /// runOnFunction - Top level algorithm. 188 /// 189 bool JumpThreading::runOnFunction(Function &F) { 190 if (skipOptnoneFunction(F)) 191 return false; 192 193 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 194 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 195 LVI = &getAnalysis<LazyValueInfo>(); 196 BFI.reset(); 197 BPI.reset(); 198 // When profile data is available, we need to update edge weights after 199 // successful jump threading, which requires both BPI and BFI being available. 200 HasProfileData = F.getEntryCount().hasValue(); 201 if (HasProfileData) { 202 LoopInfo LI{DominatorTree(F)}; 203 BPI.reset(new BranchProbabilityInfo(F, LI)); 204 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI)); 205 } 206 207 // Remove unreachable blocks from function as they may result in infinite 208 // loop. We do threading if we found something profitable. Jump threading a 209 // branch can create other opportunities. If these opportunities form a cycle 210 // i.e. if any jump threading is undoing previous threading in the path, then 211 // we will loop forever. We take care of this issue by not jump threading for 212 // back edges. This works for normal cases but not for unreachable blocks as 213 // they may have cycle with no back edge. 214 bool EverChanged = false; 215 EverChanged |= removeUnreachableBlocks(F, LVI); 216 217 FindLoopHeaders(F); 218 219 bool Changed; 220 do { 221 Changed = false; 222 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 223 BasicBlock *BB = &*I; 224 // Thread all of the branches we can over this block. 225 while (ProcessBlock(BB)) 226 Changed = true; 227 228 ++I; 229 230 // If the block is trivially dead, zap it. This eliminates the successor 231 // edges which simplifies the CFG. 232 if (pred_empty(BB) && 233 BB != &BB->getParent()->getEntryBlock()) { 234 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName() 235 << "' with terminator: " << *BB->getTerminator() << '\n'); 236 LoopHeaders.erase(BB); 237 LVI->eraseBlock(BB); 238 DeleteDeadBlock(BB); 239 Changed = true; 240 continue; 241 } 242 243 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 244 245 // Can't thread an unconditional jump, but if the block is "almost 246 // empty", we can replace uses of it with uses of the successor and make 247 // this dead. 248 // We should not eliminate the loop header either, because eliminating 249 // a loop header might later prevent LoopSimplify from transforming nested 250 // loops into simplified form. 251 if (BI && BI->isUnconditional() && 252 BB != &BB->getParent()->getEntryBlock() && 253 // If the terminator is the only non-phi instruction, try to nuke it. 254 BB->getFirstNonPHIOrDbg()->isTerminator() && !LoopHeaders.count(BB)) { 255 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the 256 // block, we have to make sure it isn't in the LoopHeaders set. We 257 // reinsert afterward if needed. 258 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); 259 BasicBlock *Succ = BI->getSuccessor(0); 260 261 // FIXME: It is always conservatively correct to drop the info 262 // for a block even if it doesn't get erased. This isn't totally 263 // awesome, but it allows us to use AssertingVH to prevent nasty 264 // dangling pointer issues within LazyValueInfo. 265 LVI->eraseBlock(BB); 266 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { 267 Changed = true; 268 // If we deleted BB and BB was the header of a loop, then the 269 // successor is now the header of the loop. 270 BB = Succ; 271 } 272 273 if (ErasedFromLoopHeaders) 274 LoopHeaders.insert(BB); 275 } 276 } 277 EverChanged |= Changed; 278 } while (Changed); 279 280 LoopHeaders.clear(); 281 return EverChanged; 282 } 283 284 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 285 /// thread across it. Stop scanning the block when passing the threshold. 286 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB, 287 unsigned Threshold) { 288 /// Ignore PHI nodes, these will be flattened when duplication happens. 289 BasicBlock::const_iterator I(BB->getFirstNonPHI()); 290 291 // FIXME: THREADING will delete values that are just used to compute the 292 // branch, so they shouldn't count against the duplication cost. 293 294 unsigned Bonus = 0; 295 const TerminatorInst *BBTerm = BB->getTerminator(); 296 // Threading through a switch statement is particularly profitable. If this 297 // block ends in a switch, decrease its cost to make it more likely to happen. 298 if (isa<SwitchInst>(BBTerm)) 299 Bonus = 6; 300 301 // The same holds for indirect branches, but slightly more so. 302 if (isa<IndirectBrInst>(BBTerm)) 303 Bonus = 8; 304 305 // Bump the threshold up so the early exit from the loop doesn't skip the 306 // terminator-based Size adjustment at the end. 307 Threshold += Bonus; 308 309 // Sum up the cost of each instruction until we get to the terminator. Don't 310 // include the terminator because the copy won't include it. 311 unsigned Size = 0; 312 for (; !isa<TerminatorInst>(I); ++I) { 313 314 // Stop scanning the block if we've reached the threshold. 315 if (Size > Threshold) 316 return Size; 317 318 // Debugger intrinsics don't incur code size. 319 if (isa<DbgInfoIntrinsic>(I)) continue; 320 321 // If this is a pointer->pointer bitcast, it is free. 322 if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) 323 continue; 324 325 // Bail out if this instruction gives back a token type, it is not possible 326 // to duplicate it if it is used outside this BB. 327 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB)) 328 return ~0U; 329 330 // All other instructions count for at least one unit. 331 ++Size; 332 333 // Calls are more expensive. If they are non-intrinsic calls, we model them 334 // as having cost of 4. If they are a non-vector intrinsic, we model them 335 // as having cost of 2 total, and if they are a vector intrinsic, we model 336 // them as having cost 1. 337 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 338 if (CI->cannotDuplicate() || CI->isConvergent()) 339 // Blocks with NoDuplicate are modelled as having infinite cost, so they 340 // are never duplicated. 341 return ~0U; 342 else if (!isa<IntrinsicInst>(CI)) 343 Size += 3; 344 else if (!CI->getType()->isVectorTy()) 345 Size += 1; 346 } 347 } 348 349 return Size > Bonus ? Size - Bonus : 0; 350 } 351 352 /// FindLoopHeaders - We do not want jump threading to turn proper loop 353 /// structures into irreducible loops. Doing this breaks up the loop nesting 354 /// hierarchy and pessimizes later transformations. To prevent this from 355 /// happening, we first have to find the loop headers. Here we approximate this 356 /// by finding targets of backedges in the CFG. 357 /// 358 /// Note that there definitely are cases when we want to allow threading of 359 /// edges across a loop header. For example, threading a jump from outside the 360 /// loop (the preheader) to an exit block of the loop is definitely profitable. 361 /// It is also almost always profitable to thread backedges from within the loop 362 /// to exit blocks, and is often profitable to thread backedges to other blocks 363 /// within the loop (forming a nested loop). This simple analysis is not rich 364 /// enough to track all of these properties and keep it up-to-date as the CFG 365 /// mutates, so we don't allow any of these transformations. 366 /// 367 void JumpThreading::FindLoopHeaders(Function &F) { 368 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 369 FindFunctionBackedges(F, Edges); 370 371 for (const auto &Edge : Edges) 372 LoopHeaders.insert(Edge.second); 373 } 374 375 /// getKnownConstant - Helper method to determine if we can thread over a 376 /// terminator with the given value as its condition, and if so what value to 377 /// use for that. What kind of value this is depends on whether we want an 378 /// integer or a block address, but an undef is always accepted. 379 /// Returns null if Val is null or not an appropriate constant. 380 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { 381 if (!Val) 382 return nullptr; 383 384 // Undef is "known" enough. 385 if (UndefValue *U = dyn_cast<UndefValue>(Val)) 386 return U; 387 388 if (Preference == WantBlockAddress) 389 return dyn_cast<BlockAddress>(Val->stripPointerCasts()); 390 391 return dyn_cast<ConstantInt>(Val); 392 } 393 394 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see 395 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef 396 /// in any of our predecessors. If so, return the known list of value and pred 397 /// BB in the result vector. 398 /// 399 /// This returns true if there were any known values. 400 /// 401 bool JumpThreading:: 402 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result, 403 ConstantPreference Preference, 404 Instruction *CxtI) { 405 // This method walks up use-def chains recursively. Because of this, we could 406 // get into an infinite loop going around loops in the use-def chain. To 407 // prevent this, keep track of what (value, block) pairs we've already visited 408 // and terminate the search if we loop back to them 409 if (!RecursionSet.insert(std::make_pair(V, BB)).second) 410 return false; 411 412 // An RAII help to remove this pair from the recursion set once the recursion 413 // stack pops back out again. 414 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB)); 415 416 // If V is a constant, then it is known in all predecessors. 417 if (Constant *KC = getKnownConstant(V, Preference)) { 418 for (BasicBlock *Pred : predecessors(BB)) 419 Result.push_back(std::make_pair(KC, Pred)); 420 421 return !Result.empty(); 422 } 423 424 // If V is a non-instruction value, or an instruction in a different block, 425 // then it can't be derived from a PHI. 426 Instruction *I = dyn_cast<Instruction>(V); 427 if (!I || I->getParent() != BB) { 428 429 // Okay, if this is a live-in value, see if it has a known value at the end 430 // of any of our predecessors. 431 // 432 // FIXME: This should be an edge property, not a block end property. 433 /// TODO: Per PR2563, we could infer value range information about a 434 /// predecessor based on its terminator. 435 // 436 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 437 // "I" is a non-local compare-with-a-constant instruction. This would be 438 // able to handle value inequalities better, for example if the compare is 439 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 440 // Perhaps getConstantOnEdge should be smart enough to do this? 441 442 for (BasicBlock *P : predecessors(BB)) { 443 // If the value is known by LazyValueInfo to be a constant in a 444 // predecessor, use that information to try to thread this block. 445 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI); 446 if (Constant *KC = getKnownConstant(PredCst, Preference)) 447 Result.push_back(std::make_pair(KC, P)); 448 } 449 450 return !Result.empty(); 451 } 452 453 /// If I is a PHI node, then we know the incoming values for any constants. 454 if (PHINode *PN = dyn_cast<PHINode>(I)) { 455 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 456 Value *InVal = PN->getIncomingValue(i); 457 if (Constant *KC = getKnownConstant(InVal, Preference)) { 458 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 459 } else { 460 Constant *CI = LVI->getConstantOnEdge(InVal, 461 PN->getIncomingBlock(i), 462 BB, CxtI); 463 if (Constant *KC = getKnownConstant(CI, Preference)) 464 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); 465 } 466 } 467 468 return !Result.empty(); 469 } 470 471 // Handle Cast instructions. Only see through Cast when the source operand is 472 // PHI or Cmp and the source type is i1 to save the compilation time. 473 if (CastInst *CI = dyn_cast<CastInst>(I)) { 474 Value *Source = CI->getOperand(0); 475 if (!Source->getType()->isIntegerTy(1)) 476 return false; 477 if (!isa<PHINode>(Source) && !isa<CmpInst>(Source)) 478 return false; 479 ComputeValueKnownInPredecessors(Source, BB, Result, Preference, CxtI); 480 if (Result.empty()) 481 return false; 482 483 // Convert the known values. 484 for (auto &R : Result) 485 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType()); 486 487 return true; 488 } 489 490 PredValueInfoTy LHSVals, RHSVals; 491 492 // Handle some boolean conditions. 493 if (I->getType()->getPrimitiveSizeInBits() == 1) { 494 assert(Preference == WantInteger && "One-bit non-integer type?"); 495 // X | true -> true 496 // X & false -> false 497 if (I->getOpcode() == Instruction::Or || 498 I->getOpcode() == Instruction::And) { 499 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 500 WantInteger, CxtI); 501 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals, 502 WantInteger, CxtI); 503 504 if (LHSVals.empty() && RHSVals.empty()) 505 return false; 506 507 ConstantInt *InterestingVal; 508 if (I->getOpcode() == Instruction::Or) 509 InterestingVal = ConstantInt::getTrue(I->getContext()); 510 else 511 InterestingVal = ConstantInt::getFalse(I->getContext()); 512 513 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 514 515 // Scan for the sentinel. If we find an undef, force it to the 516 // interesting value: x|undef -> true and x&undef -> false. 517 for (const auto &LHSVal : LHSVals) 518 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) { 519 Result.emplace_back(InterestingVal, LHSVal.second); 520 LHSKnownBBs.insert(LHSVal.second); 521 } 522 for (const auto &RHSVal : RHSVals) 523 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) { 524 // If we already inferred a value for this block on the LHS, don't 525 // re-add it. 526 if (!LHSKnownBBs.count(RHSVal.second)) 527 Result.emplace_back(InterestingVal, RHSVal.second); 528 } 529 530 return !Result.empty(); 531 } 532 533 // Handle the NOT form of XOR. 534 if (I->getOpcode() == Instruction::Xor && 535 isa<ConstantInt>(I->getOperand(1)) && 536 cast<ConstantInt>(I->getOperand(1))->isOne()) { 537 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result, 538 WantInteger, CxtI); 539 if (Result.empty()) 540 return false; 541 542 // Invert the known values. 543 for (auto &R : Result) 544 R.first = ConstantExpr::getNot(R.first); 545 546 return true; 547 } 548 549 // Try to simplify some other binary operator values. 550 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 551 assert(Preference != WantBlockAddress 552 && "A binary operator creating a block address?"); 553 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 554 PredValueInfoTy LHSVals; 555 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals, 556 WantInteger, CxtI); 557 558 // Try to use constant folding to simplify the binary operator. 559 for (const auto &LHSVal : LHSVals) { 560 Constant *V = LHSVal.first; 561 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); 562 563 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 564 Result.push_back(std::make_pair(KC, LHSVal.second)); 565 } 566 } 567 568 return !Result.empty(); 569 } 570 571 // Handle compare with phi operand, where the PHI is defined in this block. 572 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 573 assert(Preference == WantInteger && "Compares only produce integers"); 574 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); 575 if (PN && PN->getParent() == BB) { 576 const DataLayout &DL = PN->getModule()->getDataLayout(); 577 // We can do this simplification if any comparisons fold to true or false. 578 // See if any do. 579 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 580 BasicBlock *PredBB = PN->getIncomingBlock(i); 581 Value *LHS = PN->getIncomingValue(i); 582 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); 583 584 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL); 585 if (!Res) { 586 if (!isa<Constant>(RHS)) 587 continue; 588 589 LazyValueInfo::Tristate 590 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, 591 cast<Constant>(RHS), PredBB, BB, 592 CxtI ? CxtI : Cmp); 593 if (ResT == LazyValueInfo::Unknown) 594 continue; 595 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 596 } 597 598 if (Constant *KC = getKnownConstant(Res, WantInteger)) 599 Result.push_back(std::make_pair(KC, PredBB)); 600 } 601 602 return !Result.empty(); 603 } 604 605 // If comparing a live-in value against a constant, see if we know the 606 // live-in value on any predecessors. 607 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) { 608 if (!isa<Instruction>(Cmp->getOperand(0)) || 609 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { 610 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); 611 612 for (BasicBlock *P : predecessors(BB)) { 613 // If the value is known by LazyValueInfo to be a constant in a 614 // predecessor, use that information to try to thread this block. 615 LazyValueInfo::Tristate Res = 616 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), 617 RHSCst, P, BB, CxtI ? CxtI : Cmp); 618 if (Res == LazyValueInfo::Unknown) 619 continue; 620 621 Constant *ResC = ConstantInt::get(Cmp->getType(), Res); 622 Result.push_back(std::make_pair(ResC, P)); 623 } 624 625 return !Result.empty(); 626 } 627 628 // Try to find a constant value for the LHS of a comparison, 629 // and evaluate it statically if we can. 630 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { 631 PredValueInfoTy LHSVals; 632 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, 633 WantInteger, CxtI); 634 635 for (const auto &LHSVal : LHSVals) { 636 Constant *V = LHSVal.first; 637 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(), 638 V, CmpConst); 639 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 640 Result.push_back(std::make_pair(KC, LHSVal.second)); 641 } 642 643 return !Result.empty(); 644 } 645 } 646 } 647 648 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 649 // Handle select instructions where at least one operand is a known constant 650 // and we can figure out the condition value for any predecessor block. 651 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference); 652 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference); 653 PredValueInfoTy Conds; 654 if ((TrueVal || FalseVal) && 655 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds, 656 WantInteger, CxtI)) { 657 for (auto &C : Conds) { 658 Constant *Cond = C.first; 659 660 // Figure out what value to use for the condition. 661 bool KnownCond; 662 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) { 663 // A known boolean. 664 KnownCond = CI->isOne(); 665 } else { 666 assert(isa<UndefValue>(Cond) && "Unexpected condition value"); 667 // Either operand will do, so be sure to pick the one that's a known 668 // constant. 669 // FIXME: Do this more cleverly if both values are known constants? 670 KnownCond = (TrueVal != nullptr); 671 } 672 673 // See if the select has a known constant value for this predecessor. 674 if (Constant *Val = KnownCond ? TrueVal : FalseVal) 675 Result.push_back(std::make_pair(Val, C.second)); 676 } 677 678 return !Result.empty(); 679 } 680 } 681 682 // If all else fails, see if LVI can figure out a constant value for us. 683 Constant *CI = LVI->getConstant(V, BB, CxtI); 684 if (Constant *KC = getKnownConstant(CI, Preference)) { 685 for (BasicBlock *Pred : predecessors(BB)) 686 Result.push_back(std::make_pair(KC, Pred)); 687 } 688 689 return !Result.empty(); 690 } 691 692 693 694 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends 695 /// in an undefined jump, decide which block is best to revector to. 696 /// 697 /// Since we can pick an arbitrary destination, we pick the successor with the 698 /// fewest predecessors. This should reduce the in-degree of the others. 699 /// 700 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { 701 TerminatorInst *BBTerm = BB->getTerminator(); 702 unsigned MinSucc = 0; 703 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 704 // Compute the successor with the minimum number of predecessors. 705 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 706 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 707 TestBB = BBTerm->getSuccessor(i); 708 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 709 if (NumPreds < MinNumPreds) { 710 MinSucc = i; 711 MinNumPreds = NumPreds; 712 } 713 } 714 715 return MinSucc; 716 } 717 718 static bool hasAddressTakenAndUsed(BasicBlock *BB) { 719 if (!BB->hasAddressTaken()) return false; 720 721 // If the block has its address taken, it may be a tree of dead constants 722 // hanging off of it. These shouldn't keep the block alive. 723 BlockAddress *BA = BlockAddress::get(BB); 724 BA->removeDeadConstantUsers(); 725 return !BA->use_empty(); 726 } 727 728 /// ProcessBlock - If there are any predecessors whose control can be threaded 729 /// through to a successor, transform them now. 730 bool JumpThreading::ProcessBlock(BasicBlock *BB) { 731 // If the block is trivially dead, just return and let the caller nuke it. 732 // This simplifies other transformations. 733 if (pred_empty(BB) && 734 BB != &BB->getParent()->getEntryBlock()) 735 return false; 736 737 // If this block has a single predecessor, and if that pred has a single 738 // successor, merge the blocks. This encourages recursive jump threading 739 // because now the condition in this block can be threaded through 740 // predecessors of our predecessor block. 741 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { 742 const TerminatorInst *TI = SinglePred->getTerminator(); 743 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 && 744 SinglePred != BB && !hasAddressTakenAndUsed(BB)) { 745 // If SinglePred was a loop header, BB becomes one. 746 if (LoopHeaders.erase(SinglePred)) 747 LoopHeaders.insert(BB); 748 749 LVI->eraseBlock(SinglePred); 750 MergeBasicBlockIntoOnlyPred(BB); 751 752 return true; 753 } 754 } 755 756 if (TryToUnfoldSelectInCurrBB(BB)) 757 return true; 758 759 // What kind of constant we're looking for. 760 ConstantPreference Preference = WantInteger; 761 762 // Look to see if the terminator is a conditional branch, switch or indirect 763 // branch, if not we can't thread it. 764 Value *Condition; 765 Instruction *Terminator = BB->getTerminator(); 766 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { 767 // Can't thread an unconditional jump. 768 if (BI->isUnconditional()) return false; 769 Condition = BI->getCondition(); 770 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { 771 Condition = SI->getCondition(); 772 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { 773 // Can't thread indirect branch with no successors. 774 if (IB->getNumSuccessors() == 0) return false; 775 Condition = IB->getAddress()->stripPointerCasts(); 776 Preference = WantBlockAddress; 777 } else { 778 return false; // Must be an invoke. 779 } 780 781 // Run constant folding to see if we can reduce the condition to a simple 782 // constant. 783 if (Instruction *I = dyn_cast<Instruction>(Condition)) { 784 Value *SimpleVal = 785 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI); 786 if (SimpleVal) { 787 I->replaceAllUsesWith(SimpleVal); 788 I->eraseFromParent(); 789 Condition = SimpleVal; 790 } 791 } 792 793 // If the terminator is branching on an undef, we can pick any of the 794 // successors to branch to. Let GetBestDestForJumpOnUndef decide. 795 if (isa<UndefValue>(Condition)) { 796 unsigned BestSucc = GetBestDestForJumpOnUndef(BB); 797 798 // Fold the branch/switch. 799 TerminatorInst *BBTerm = BB->getTerminator(); 800 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 801 if (i == BestSucc) continue; 802 BBTerm->getSuccessor(i)->removePredecessor(BB, true); 803 } 804 805 DEBUG(dbgs() << " In block '" << BB->getName() 806 << "' folding undef terminator: " << *BBTerm << '\n'); 807 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 808 BBTerm->eraseFromParent(); 809 return true; 810 } 811 812 // If the terminator of this block is branching on a constant, simplify the 813 // terminator to an unconditional branch. This can occur due to threading in 814 // other blocks. 815 if (getKnownConstant(Condition, Preference)) { 816 DEBUG(dbgs() << " In block '" << BB->getName() 817 << "' folding terminator: " << *BB->getTerminator() << '\n'); 818 ++NumFolds; 819 ConstantFoldTerminator(BB, true); 820 return true; 821 } 822 823 Instruction *CondInst = dyn_cast<Instruction>(Condition); 824 825 // All the rest of our checks depend on the condition being an instruction. 826 if (!CondInst) { 827 // FIXME: Unify this with code below. 828 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator)) 829 return true; 830 return false; 831 } 832 833 834 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 835 // If we're branching on a conditional, LVI might be able to determine 836 // it's value at the branch instruction. We only handle comparisons 837 // against a constant at this time. 838 // TODO: This should be extended to handle switches as well. 839 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 840 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); 841 if (CondBr && CondConst && CondBr->isConditional()) { 842 LazyValueInfo::Tristate Ret = 843 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0), 844 CondConst, CondBr); 845 if (Ret != LazyValueInfo::Unknown) { 846 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0; 847 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1; 848 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); 849 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); 850 CondBr->eraseFromParent(); 851 if (CondCmp->use_empty()) 852 CondCmp->eraseFromParent(); 853 else if (CondCmp->getParent() == BB) { 854 // If the fact we just learned is true for all uses of the 855 // condition, replace it with a constant value 856 auto *CI = Ret == LazyValueInfo::True ? 857 ConstantInt::getTrue(CondCmp->getType()) : 858 ConstantInt::getFalse(CondCmp->getType()); 859 CondCmp->replaceAllUsesWith(CI); 860 CondCmp->eraseFromParent(); 861 } 862 return true; 863 } 864 } 865 866 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB)) 867 return true; 868 } 869 870 // Check for some cases that are worth simplifying. Right now we want to look 871 // for loads that are used by a switch or by the condition for the branch. If 872 // we see one, check to see if it's partially redundant. If so, insert a PHI 873 // which can then be used to thread the values. 874 // 875 Value *SimplifyValue = CondInst; 876 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 877 if (isa<Constant>(CondCmp->getOperand(1))) 878 SimplifyValue = CondCmp->getOperand(0); 879 880 // TODO: There are other places where load PRE would be profitable, such as 881 // more complex comparisons. 882 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 883 if (SimplifyPartiallyRedundantLoad(LI)) 884 return true; 885 886 887 // Handle a variety of cases where we are branching on something derived from 888 // a PHI node in the current block. If we can prove that any predecessors 889 // compute a predictable value based on a PHI node, thread those predecessors. 890 // 891 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator)) 892 return true; 893 894 // If this is an otherwise-unfoldable branch on a phi node in the current 895 // block, see if we can simplify. 896 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 897 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 898 return ProcessBranchOnPHI(PN); 899 900 901 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 902 if (CondInst->getOpcode() == Instruction::Xor && 903 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 904 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); 905 906 // Search for a stronger dominating condition that can be used to simplify a 907 // conditional branch leaving BB. 908 if (ProcessImpliedCondition(BB)) 909 return true; 910 911 return false; 912 } 913 914 bool JumpThreading::ProcessImpliedCondition(BasicBlock *BB) { 915 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 916 if (!BI || !BI->isConditional()) 917 return false; 918 919 Value *Cond = BI->getCondition(); 920 BasicBlock *CurrentBB = BB; 921 BasicBlock *CurrentPred = BB->getSinglePredecessor(); 922 unsigned Iter = 0; 923 924 auto &DL = BB->getModule()->getDataLayout(); 925 926 while (CurrentPred && Iter++ < ImplicationSearchThreshold) { 927 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator()); 928 if (!PBI || !PBI->isConditional() || PBI->getSuccessor(0) != CurrentBB) 929 return false; 930 931 if (isImpliedCondition(PBI->getCondition(), Cond, DL)) { 932 BI->getSuccessor(1)->removePredecessor(BB); 933 BranchInst::Create(BI->getSuccessor(0), BI); 934 BI->eraseFromParent(); 935 return true; 936 } 937 CurrentBB = CurrentPred; 938 CurrentPred = CurrentBB->getSinglePredecessor(); 939 } 940 941 return false; 942 } 943 944 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 945 /// load instruction, eliminate it by replacing it with a PHI node. This is an 946 /// important optimization that encourages jump threading, and needs to be run 947 /// interlaced with other jump threading tasks. 948 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 949 // Don't hack volatile/atomic loads. 950 if (!LI->isSimple()) return false; 951 952 // If the load is defined in a block with exactly one predecessor, it can't be 953 // partially redundant. 954 BasicBlock *LoadBB = LI->getParent(); 955 if (LoadBB->getSinglePredecessor()) 956 return false; 957 958 // If the load is defined in an EH pad, it can't be partially redundant, 959 // because the edges between the invoke and the EH pad cannot have other 960 // instructions between them. 961 if (LoadBB->isEHPad()) 962 return false; 963 964 Value *LoadedPtr = LI->getOperand(0); 965 966 // If the loaded operand is defined in the LoadBB, it can't be available. 967 // TODO: Could do simple PHI translation, that would be fun :) 968 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 969 if (PtrOp->getParent() == LoadBB) 970 return false; 971 972 // Scan a few instructions up from the load, to see if it is obviously live at 973 // the entry to its block. 974 BasicBlock::iterator BBIt(LI); 975 976 if (Value *AvailableVal = 977 FindAvailableLoadedValue(LI, LoadBB, BBIt, DefMaxInstsToScan)) { 978 // If the value of the load is locally available within the block, just use 979 // it. This frequently occurs for reg2mem'd allocas. 980 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 981 982 // If the returned value is the load itself, replace with an undef. This can 983 // only happen in dead loops. 984 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); 985 if (AvailableVal->getType() != LI->getType()) 986 AvailableVal = 987 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI); 988 LI->replaceAllUsesWith(AvailableVal); 989 LI->eraseFromParent(); 990 return true; 991 } 992 993 // Otherwise, if we scanned the whole block and got to the top of the block, 994 // we know the block is locally transparent to the load. If not, something 995 // might clobber its value. 996 if (BBIt != LoadBB->begin()) 997 return false; 998 999 // If all of the loads and stores that feed the value have the same AA tags, 1000 // then we can propagate them onto any newly inserted loads. 1001 AAMDNodes AATags; 1002 LI->getAAMetadata(AATags); 1003 1004 SmallPtrSet<BasicBlock*, 8> PredsScanned; 1005 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 1006 AvailablePredsTy AvailablePreds; 1007 BasicBlock *OneUnavailablePred = nullptr; 1008 1009 // If we got here, the loaded value is transparent through to the start of the 1010 // block. Check to see if it is available in any of the predecessor blocks. 1011 for (BasicBlock *PredBB : predecessors(LoadBB)) { 1012 // If we already scanned this predecessor, skip it. 1013 if (!PredsScanned.insert(PredBB).second) 1014 continue; 1015 1016 // Scan the predecessor to see if the value is available in the pred. 1017 BBIt = PredBB->end(); 1018 AAMDNodes ThisAATags; 1019 Value *PredAvailable = FindAvailableLoadedValue(LI, PredBB, BBIt, 1020 DefMaxInstsToScan, 1021 nullptr, &ThisAATags); 1022 if (!PredAvailable) { 1023 OneUnavailablePred = PredBB; 1024 continue; 1025 } 1026 1027 // If AA tags disagree or are not present, forget about them. 1028 if (AATags != ThisAATags) AATags = AAMDNodes(); 1029 1030 // If so, this load is partially redundant. Remember this info so that we 1031 // can create a PHI node. 1032 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 1033 } 1034 1035 // If the loaded value isn't available in any predecessor, it isn't partially 1036 // redundant. 1037 if (AvailablePreds.empty()) return false; 1038 1039 // Okay, the loaded value is available in at least one (and maybe all!) 1040 // predecessors. If the value is unavailable in more than one unique 1041 // predecessor, we want to insert a merge block for those common predecessors. 1042 // This ensures that we only have to insert one reload, thus not increasing 1043 // code size. 1044 BasicBlock *UnavailablePred = nullptr; 1045 1046 // If there is exactly one predecessor where the value is unavailable, the 1047 // already computed 'OneUnavailablePred' block is it. If it ends in an 1048 // unconditional branch, we know that it isn't a critical edge. 1049 if (PredsScanned.size() == AvailablePreds.size()+1 && 1050 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 1051 UnavailablePred = OneUnavailablePred; 1052 } else if (PredsScanned.size() != AvailablePreds.size()) { 1053 // Otherwise, we had multiple unavailable predecessors or we had a critical 1054 // edge from the one. 1055 SmallVector<BasicBlock*, 8> PredsToSplit; 1056 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 1057 1058 for (const auto &AvailablePred : AvailablePreds) 1059 AvailablePredSet.insert(AvailablePred.first); 1060 1061 // Add all the unavailable predecessors to the PredsToSplit list. 1062 for (BasicBlock *P : predecessors(LoadBB)) { 1063 // If the predecessor is an indirect goto, we can't split the edge. 1064 if (isa<IndirectBrInst>(P->getTerminator())) 1065 return false; 1066 1067 if (!AvailablePredSet.count(P)) 1068 PredsToSplit.push_back(P); 1069 } 1070 1071 // Split them out to their own block. 1072 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split"); 1073 } 1074 1075 // If the value isn't available in all predecessors, then there will be 1076 // exactly one where it isn't available. Insert a load on that edge and add 1077 // it to the AvailablePreds list. 1078 if (UnavailablePred) { 1079 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 1080 "Can't handle critical edge here!"); 1081 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, 1082 LI->getAlignment(), 1083 UnavailablePred->getTerminator()); 1084 NewVal->setDebugLoc(LI->getDebugLoc()); 1085 if (AATags) 1086 NewVal->setAAMetadata(AATags); 1087 1088 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 1089 } 1090 1091 // Now we know that each predecessor of this block has a value in 1092 // AvailablePreds, sort them for efficient access as we're walking the preds. 1093 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 1094 1095 // Create a PHI node at the start of the block for the PRE'd load value. 1096 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB); 1097 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "", 1098 &LoadBB->front()); 1099 PN->takeName(LI); 1100 PN->setDebugLoc(LI->getDebugLoc()); 1101 1102 // Insert new entries into the PHI for each predecessor. A single block may 1103 // have multiple entries here. 1104 for (pred_iterator PI = PB; PI != PE; ++PI) { 1105 BasicBlock *P = *PI; 1106 AvailablePredsTy::iterator I = 1107 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 1108 std::make_pair(P, (Value*)nullptr)); 1109 1110 assert(I != AvailablePreds.end() && I->first == P && 1111 "Didn't find entry for predecessor!"); 1112 1113 // If we have an available predecessor but it requires casting, insert the 1114 // cast in the predecessor and use the cast. Note that we have to update the 1115 // AvailablePreds vector as we go so that all of the PHI entries for this 1116 // predecessor use the same bitcast. 1117 Value *&PredV = I->second; 1118 if (PredV->getType() != LI->getType()) 1119 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "", 1120 P->getTerminator()); 1121 1122 PN->addIncoming(PredV, I->first); 1123 } 1124 1125 //cerr << "PRE: " << *LI << *PN << "\n"; 1126 1127 LI->replaceAllUsesWith(PN); 1128 LI->eraseFromParent(); 1129 1130 return true; 1131 } 1132 1133 /// FindMostPopularDest - The specified list contains multiple possible 1134 /// threadable destinations. Pick the one that occurs the most frequently in 1135 /// the list. 1136 static BasicBlock * 1137 FindMostPopularDest(BasicBlock *BB, 1138 const SmallVectorImpl<std::pair<BasicBlock*, 1139 BasicBlock*> > &PredToDestList) { 1140 assert(!PredToDestList.empty()); 1141 1142 // Determine popularity. If there are multiple possible destinations, we 1143 // explicitly choose to ignore 'undef' destinations. We prefer to thread 1144 // blocks with known and real destinations to threading undef. We'll handle 1145 // them later if interesting. 1146 DenseMap<BasicBlock*, unsigned> DestPopularity; 1147 for (const auto &PredToDest : PredToDestList) 1148 if (PredToDest.second) 1149 DestPopularity[PredToDest.second]++; 1150 1151 // Find the most popular dest. 1152 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); 1153 BasicBlock *MostPopularDest = DPI->first; 1154 unsigned Popularity = DPI->second; 1155 SmallVector<BasicBlock*, 4> SamePopularity; 1156 1157 for (++DPI; DPI != DestPopularity.end(); ++DPI) { 1158 // If the popularity of this entry isn't higher than the popularity we've 1159 // seen so far, ignore it. 1160 if (DPI->second < Popularity) 1161 ; // ignore. 1162 else if (DPI->second == Popularity) { 1163 // If it is the same as what we've seen so far, keep track of it. 1164 SamePopularity.push_back(DPI->first); 1165 } else { 1166 // If it is more popular, remember it. 1167 SamePopularity.clear(); 1168 MostPopularDest = DPI->first; 1169 Popularity = DPI->second; 1170 } 1171 } 1172 1173 // Okay, now we know the most popular destination. If there is more than one 1174 // destination, we need to determine one. This is arbitrary, but we need 1175 // to make a deterministic decision. Pick the first one that appears in the 1176 // successor list. 1177 if (!SamePopularity.empty()) { 1178 SamePopularity.push_back(MostPopularDest); 1179 TerminatorInst *TI = BB->getTerminator(); 1180 for (unsigned i = 0; ; ++i) { 1181 assert(i != TI->getNumSuccessors() && "Didn't find any successor!"); 1182 1183 if (std::find(SamePopularity.begin(), SamePopularity.end(), 1184 TI->getSuccessor(i)) == SamePopularity.end()) 1185 continue; 1186 1187 MostPopularDest = TI->getSuccessor(i); 1188 break; 1189 } 1190 } 1191 1192 // Okay, we have finally picked the most popular destination. 1193 return MostPopularDest; 1194 } 1195 1196 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB, 1197 ConstantPreference Preference, 1198 Instruction *CxtI) { 1199 // If threading this would thread across a loop header, don't even try to 1200 // thread the edge. 1201 if (LoopHeaders.count(BB)) 1202 return false; 1203 1204 PredValueInfoTy PredValues; 1205 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI)) 1206 return false; 1207 1208 assert(!PredValues.empty() && 1209 "ComputeValueKnownInPredecessors returned true with no values"); 1210 1211 DEBUG(dbgs() << "IN BB: " << *BB; 1212 for (const auto &PredValue : PredValues) { 1213 dbgs() << " BB '" << BB->getName() << "': FOUND condition = " 1214 << *PredValue.first 1215 << " for pred '" << PredValue.second->getName() << "'.\n"; 1216 }); 1217 1218 // Decide what we want to thread through. Convert our list of known values to 1219 // a list of known destinations for each pred. This also discards duplicate 1220 // predecessors and keeps track of the undefined inputs (which are represented 1221 // as a null dest in the PredToDestList). 1222 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1223 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1224 1225 BasicBlock *OnlyDest = nullptr; 1226 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1227 1228 for (const auto &PredValue : PredValues) { 1229 BasicBlock *Pred = PredValue.second; 1230 if (!SeenPreds.insert(Pred).second) 1231 continue; // Duplicate predecessor entry. 1232 1233 // If the predecessor ends with an indirect goto, we can't change its 1234 // destination. 1235 if (isa<IndirectBrInst>(Pred->getTerminator())) 1236 continue; 1237 1238 Constant *Val = PredValue.first; 1239 1240 BasicBlock *DestBB; 1241 if (isa<UndefValue>(Val)) 1242 DestBB = nullptr; 1243 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 1244 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); 1245 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 1246 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor(); 1247 } else { 1248 assert(isa<IndirectBrInst>(BB->getTerminator()) 1249 && "Unexpected terminator"); 1250 DestBB = cast<BlockAddress>(Val)->getBasicBlock(); 1251 } 1252 1253 // If we have exactly one destination, remember it for efficiency below. 1254 if (PredToDestList.empty()) 1255 OnlyDest = DestBB; 1256 else if (OnlyDest != DestBB) 1257 OnlyDest = MultipleDestSentinel; 1258 1259 PredToDestList.push_back(std::make_pair(Pred, DestBB)); 1260 } 1261 1262 // If all edges were unthreadable, we fail. 1263 if (PredToDestList.empty()) 1264 return false; 1265 1266 // Determine which is the most common successor. If we have many inputs and 1267 // this block is a switch, we want to start by threading the batch that goes 1268 // to the most popular destination first. If we only know about one 1269 // threadable destination (the common case) we can avoid this. 1270 BasicBlock *MostPopularDest = OnlyDest; 1271 1272 if (MostPopularDest == MultipleDestSentinel) 1273 MostPopularDest = FindMostPopularDest(BB, PredToDestList); 1274 1275 // Now that we know what the most popular destination is, factor all 1276 // predecessors that will jump to it into a single predecessor. 1277 SmallVector<BasicBlock*, 16> PredsToFactor; 1278 for (const auto &PredToDest : PredToDestList) 1279 if (PredToDest.second == MostPopularDest) { 1280 BasicBlock *Pred = PredToDest.first; 1281 1282 // This predecessor may be a switch or something else that has multiple 1283 // edges to the block. Factor each of these edges by listing them 1284 // according to # occurrences in PredsToFactor. 1285 for (BasicBlock *Succ : successors(Pred)) 1286 if (Succ == BB) 1287 PredsToFactor.push_back(Pred); 1288 } 1289 1290 // If the threadable edges are branching on an undefined value, we get to pick 1291 // the destination that these predecessors should get to. 1292 if (!MostPopularDest) 1293 MostPopularDest = BB->getTerminator()-> 1294 getSuccessor(GetBestDestForJumpOnUndef(BB)); 1295 1296 // Ok, try to thread it! 1297 return ThreadEdge(BB, PredsToFactor, MostPopularDest); 1298 } 1299 1300 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on 1301 /// a PHI node in the current block. See if there are any simplifications we 1302 /// can do based on inputs to the phi node. 1303 /// 1304 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { 1305 BasicBlock *BB = PN->getParent(); 1306 1307 // TODO: We could make use of this to do it once for blocks with common PHI 1308 // values. 1309 SmallVector<BasicBlock*, 1> PredBBs; 1310 PredBBs.resize(1); 1311 1312 // If any of the predecessor blocks end in an unconditional branch, we can 1313 // *duplicate* the conditional branch into that block in order to further 1314 // encourage jump threading and to eliminate cases where we have branch on a 1315 // phi of an icmp (branch on icmp is much better). 1316 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1317 BasicBlock *PredBB = PN->getIncomingBlock(i); 1318 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1319 if (PredBr->isUnconditional()) { 1320 PredBBs[0] = PredBB; 1321 // Try to duplicate BB into PredBB. 1322 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1323 return true; 1324 } 1325 } 1326 1327 return false; 1328 } 1329 1330 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on 1331 /// a xor instruction in the current block. See if there are any 1332 /// simplifications we can do based on inputs to the xor. 1333 /// 1334 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { 1335 BasicBlock *BB = BO->getParent(); 1336 1337 // If either the LHS or RHS of the xor is a constant, don't do this 1338 // optimization. 1339 if (isa<ConstantInt>(BO->getOperand(0)) || 1340 isa<ConstantInt>(BO->getOperand(1))) 1341 return false; 1342 1343 // If the first instruction in BB isn't a phi, we won't be able to infer 1344 // anything special about any particular predecessor. 1345 if (!isa<PHINode>(BB->front())) 1346 return false; 1347 1348 // If we have a xor as the branch input to this block, and we know that the 1349 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1350 // the condition into the predecessor and fix that value to true, saving some 1351 // logical ops on that path and encouraging other paths to simplify. 1352 // 1353 // This copies something like this: 1354 // 1355 // BB: 1356 // %X = phi i1 [1], [%X'] 1357 // %Y = icmp eq i32 %A, %B 1358 // %Z = xor i1 %X, %Y 1359 // br i1 %Z, ... 1360 // 1361 // Into: 1362 // BB': 1363 // %Y = icmp ne i32 %A, %B 1364 // br i1 %Y, ... 1365 1366 PredValueInfoTy XorOpValues; 1367 bool isLHS = true; 1368 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, 1369 WantInteger, BO)) { 1370 assert(XorOpValues.empty()); 1371 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, 1372 WantInteger, BO)) 1373 return false; 1374 isLHS = false; 1375 } 1376 1377 assert(!XorOpValues.empty() && 1378 "ComputeValueKnownInPredecessors returned true with no values"); 1379 1380 // Scan the information to see which is most popular: true or false. The 1381 // predecessors can be of the set true, false, or undef. 1382 unsigned NumTrue = 0, NumFalse = 0; 1383 for (const auto &XorOpValue : XorOpValues) { 1384 if (isa<UndefValue>(XorOpValue.first)) 1385 // Ignore undefs for the count. 1386 continue; 1387 if (cast<ConstantInt>(XorOpValue.first)->isZero()) 1388 ++NumFalse; 1389 else 1390 ++NumTrue; 1391 } 1392 1393 // Determine which value to split on, true, false, or undef if neither. 1394 ConstantInt *SplitVal = nullptr; 1395 if (NumTrue > NumFalse) 1396 SplitVal = ConstantInt::getTrue(BB->getContext()); 1397 else if (NumTrue != 0 || NumFalse != 0) 1398 SplitVal = ConstantInt::getFalse(BB->getContext()); 1399 1400 // Collect all of the blocks that this can be folded into so that we can 1401 // factor this once and clone it once. 1402 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1403 for (const auto &XorOpValue : XorOpValues) { 1404 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first)) 1405 continue; 1406 1407 BlocksToFoldInto.push_back(XorOpValue.second); 1408 } 1409 1410 // If we inferred a value for all of the predecessors, then duplication won't 1411 // help us. However, we can just replace the LHS or RHS with the constant. 1412 if (BlocksToFoldInto.size() == 1413 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1414 if (!SplitVal) { 1415 // If all preds provide undef, just nuke the xor, because it is undef too. 1416 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1417 BO->eraseFromParent(); 1418 } else if (SplitVal->isZero()) { 1419 // If all preds provide 0, replace the xor with the other input. 1420 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1421 BO->eraseFromParent(); 1422 } else { 1423 // If all preds provide 1, set the computed value to 1. 1424 BO->setOperand(!isLHS, SplitVal); 1425 } 1426 1427 return true; 1428 } 1429 1430 // Try to duplicate BB into PredBB. 1431 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1432 } 1433 1434 1435 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1436 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1437 /// NewPred using the entries from OldPred (suitably mapped). 1438 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1439 BasicBlock *OldPred, 1440 BasicBlock *NewPred, 1441 DenseMap<Instruction*, Value*> &ValueMap) { 1442 for (BasicBlock::iterator PNI = PHIBB->begin(); 1443 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { 1444 // Ok, we have a PHI node. Figure out what the incoming value was for the 1445 // DestBlock. 1446 Value *IV = PN->getIncomingValueForBlock(OldPred); 1447 1448 // Remap the value if necessary. 1449 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1450 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1451 if (I != ValueMap.end()) 1452 IV = I->second; 1453 } 1454 1455 PN->addIncoming(IV, NewPred); 1456 } 1457 } 1458 1459 /// ThreadEdge - We have decided that it is safe and profitable to factor the 1460 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 1461 /// across BB. Transform the IR to reflect this change. 1462 bool JumpThreading::ThreadEdge(BasicBlock *BB, 1463 const SmallVectorImpl<BasicBlock*> &PredBBs, 1464 BasicBlock *SuccBB) { 1465 // If threading to the same block as we come from, we would infinite loop. 1466 if (SuccBB == BB) { 1467 DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 1468 << "' - would thread to self!\n"); 1469 return false; 1470 } 1471 1472 // If threading this would thread across a loop header, don't thread the edge. 1473 // See the comments above FindLoopHeaders for justifications and caveats. 1474 if (LoopHeaders.count(BB)) { 1475 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName() 1476 << "' to dest BB '" << SuccBB->getName() 1477 << "' - it might create an irreducible loop!\n"); 1478 return false; 1479 } 1480 1481 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, BBDupThreshold); 1482 if (JumpThreadCost > BBDupThreshold) { 1483 DEBUG(dbgs() << " Not threading BB '" << BB->getName() 1484 << "' - Cost is too high: " << JumpThreadCost << "\n"); 1485 return false; 1486 } 1487 1488 // And finally, do it! Start by factoring the predecessors if needed. 1489 BasicBlock *PredBB; 1490 if (PredBBs.size() == 1) 1491 PredBB = PredBBs[0]; 1492 else { 1493 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1494 << " common predecessors.\n"); 1495 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm"); 1496 } 1497 1498 // And finally, do it! 1499 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" 1500 << SuccBB->getName() << "' with cost: " << JumpThreadCost 1501 << ", across block:\n " 1502 << *BB << "\n"); 1503 1504 LVI->threadEdge(PredBB, BB, SuccBB); 1505 1506 // We are going to have to map operands from the original BB block to the new 1507 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 1508 // account for entry from PredBB. 1509 DenseMap<Instruction*, Value*> ValueMapping; 1510 1511 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 1512 BB->getName()+".thread", 1513 BB->getParent(), BB); 1514 NewBB->moveAfter(PredBB); 1515 1516 // Set the block frequency of NewBB. 1517 if (HasProfileData) { 1518 auto NewBBFreq = 1519 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB); 1520 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency()); 1521 } 1522 1523 BasicBlock::iterator BI = BB->begin(); 1524 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1525 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1526 1527 // Clone the non-phi instructions of BB into NewBB, keeping track of the 1528 // mapping and using it to remap operands in the cloned instructions. 1529 for (; !isa<TerminatorInst>(BI); ++BI) { 1530 Instruction *New = BI->clone(); 1531 New->setName(BI->getName()); 1532 NewBB->getInstList().push_back(New); 1533 ValueMapping[&*BI] = New; 1534 1535 // Remap operands to patch up intra-block references. 1536 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1537 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1538 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1539 if (I != ValueMapping.end()) 1540 New->setOperand(i, I->second); 1541 } 1542 } 1543 1544 // We didn't copy the terminator from BB over to NewBB, because there is now 1545 // an unconditional jump to SuccBB. Insert the unconditional jump. 1546 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB); 1547 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc()); 1548 1549 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 1550 // PHI nodes for NewBB now. 1551 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 1552 1553 // If there were values defined in BB that are used outside the block, then we 1554 // now have to update all uses of the value to use either the original value, 1555 // the cloned value, or some PHI derived value. This can require arbitrary 1556 // PHI insertion, of which we are prepared to do, clean these up now. 1557 SSAUpdater SSAUpdate; 1558 SmallVector<Use*, 16> UsesToRename; 1559 for (Instruction &I : *BB) { 1560 // Scan all uses of this instruction to see if it is used outside of its 1561 // block, and if so, record them in UsesToRename. 1562 for (Use &U : I.uses()) { 1563 Instruction *User = cast<Instruction>(U.getUser()); 1564 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1565 if (UserPN->getIncomingBlock(U) == BB) 1566 continue; 1567 } else if (User->getParent() == BB) 1568 continue; 1569 1570 UsesToRename.push_back(&U); 1571 } 1572 1573 // If there are no uses outside the block, we're done with this instruction. 1574 if (UsesToRename.empty()) 1575 continue; 1576 1577 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n"); 1578 1579 // We found a use of I outside of BB. Rename all uses of I that are outside 1580 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1581 // with the two values we know. 1582 SSAUpdate.Initialize(I.getType(), I.getName()); 1583 SSAUpdate.AddAvailableValue(BB, &I); 1584 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]); 1585 1586 while (!UsesToRename.empty()) 1587 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1588 DEBUG(dbgs() << "\n"); 1589 } 1590 1591 1592 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 1593 // NewBB instead of BB. This eliminates predecessors from BB, which requires 1594 // us to simplify any PHI nodes in BB. 1595 TerminatorInst *PredTerm = PredBB->getTerminator(); 1596 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 1597 if (PredTerm->getSuccessor(i) == BB) { 1598 BB->removePredecessor(PredBB, true); 1599 PredTerm->setSuccessor(i, NewBB); 1600 } 1601 1602 // At this point, the IR is fully up to date and consistent. Do a quick scan 1603 // over the new instructions and zap any that are constants or dead. This 1604 // frequently happens because of phi translation. 1605 SimplifyInstructionsInBlock(NewBB, TLI); 1606 1607 // Update the edge weight from BB to SuccBB, which should be less than before. 1608 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB); 1609 1610 // Threaded an edge! 1611 ++NumThreads; 1612 return true; 1613 } 1614 1615 /// Create a new basic block that will be the predecessor of BB and successor of 1616 /// all blocks in Preds. When profile data is availble, update the frequency of 1617 /// this new block. 1618 BasicBlock *JumpThreading::SplitBlockPreds(BasicBlock *BB, 1619 ArrayRef<BasicBlock *> Preds, 1620 const char *Suffix) { 1621 // Collect the frequencies of all predecessors of BB, which will be used to 1622 // update the edge weight on BB->SuccBB. 1623 BlockFrequency PredBBFreq(0); 1624 if (HasProfileData) 1625 for (auto Pred : Preds) 1626 PredBBFreq += BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB); 1627 1628 BasicBlock *PredBB = SplitBlockPredecessors(BB, Preds, Suffix); 1629 1630 // Set the block frequency of the newly created PredBB, which is the sum of 1631 // frequencies of Preds. 1632 if (HasProfileData) 1633 BFI->setBlockFreq(PredBB, PredBBFreq.getFrequency()); 1634 return PredBB; 1635 } 1636 1637 /// Update the block frequency of BB and branch weight and the metadata on the 1638 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 - 1639 /// Freq(PredBB->BB) / Freq(BB->SuccBB). 1640 void JumpThreading::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB, 1641 BasicBlock *BB, 1642 BasicBlock *NewBB, 1643 BasicBlock *SuccBB) { 1644 if (!HasProfileData) 1645 return; 1646 1647 assert(BFI && BPI && "BFI & BPI should have been created here"); 1648 1649 // As the edge from PredBB to BB is deleted, we have to update the block 1650 // frequency of BB. 1651 auto BBOrigFreq = BFI->getBlockFreq(BB); 1652 auto NewBBFreq = BFI->getBlockFreq(NewBB); 1653 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB); 1654 auto BBNewFreq = BBOrigFreq - NewBBFreq; 1655 BFI->setBlockFreq(BB, BBNewFreq.getFrequency()); 1656 1657 // Collect updated outgoing edges' frequencies from BB and use them to update 1658 // edge probabilities. 1659 SmallVector<uint64_t, 4> BBSuccFreq; 1660 for (BasicBlock *Succ : successors(BB)) { 1661 auto SuccFreq = (Succ == SuccBB) 1662 ? BB2SuccBBFreq - NewBBFreq 1663 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ); 1664 BBSuccFreq.push_back(SuccFreq.getFrequency()); 1665 } 1666 1667 uint64_t MaxBBSuccFreq = 1668 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end()); 1669 1670 SmallVector<BranchProbability, 4> BBSuccProbs; 1671 if (MaxBBSuccFreq == 0) 1672 BBSuccProbs.assign(BBSuccFreq.size(), 1673 {1, static_cast<unsigned>(BBSuccFreq.size())}); 1674 else { 1675 for (uint64_t Freq : BBSuccFreq) 1676 BBSuccProbs.push_back( 1677 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq)); 1678 // Normalize edge probabilities so that they sum up to one. 1679 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(), 1680 BBSuccProbs.end()); 1681 } 1682 1683 // Update edge probabilities in BPI. 1684 for (int I = 0, E = BBSuccProbs.size(); I < E; I++) 1685 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]); 1686 1687 if (BBSuccProbs.size() >= 2) { 1688 SmallVector<uint32_t, 4> Weights; 1689 for (auto Prob : BBSuccProbs) 1690 Weights.push_back(Prob.getNumerator()); 1691 1692 auto TI = BB->getTerminator(); 1693 TI->setMetadata( 1694 LLVMContext::MD_prof, 1695 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights)); 1696 } 1697 } 1698 1699 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 1700 /// to BB which contains an i1 PHI node and a conditional branch on that PHI. 1701 /// If we can duplicate the contents of BB up into PredBB do so now, this 1702 /// improves the odds that the branch will be on an analyzable instruction like 1703 /// a compare. 1704 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, 1705 const SmallVectorImpl<BasicBlock *> &PredBBs) { 1706 assert(!PredBBs.empty() && "Can't handle an empty set"); 1707 1708 // If BB is a loop header, then duplicating this block outside the loop would 1709 // cause us to transform this into an irreducible loop, don't do this. 1710 // See the comments above FindLoopHeaders for justifications and caveats. 1711 if (LoopHeaders.count(BB)) { 1712 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 1713 << "' into predecessor block '" << PredBBs[0]->getName() 1714 << "' - it might create an irreducible loop!\n"); 1715 return false; 1716 } 1717 1718 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, BBDupThreshold); 1719 if (DuplicationCost > BBDupThreshold) { 1720 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 1721 << "' - Cost is too high: " << DuplicationCost << "\n"); 1722 return false; 1723 } 1724 1725 // And finally, do it! Start by factoring the predecessors if needed. 1726 BasicBlock *PredBB; 1727 if (PredBBs.size() == 1) 1728 PredBB = PredBBs[0]; 1729 else { 1730 DEBUG(dbgs() << " Factoring out " << PredBBs.size() 1731 << " common predecessors.\n"); 1732 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm"); 1733 } 1734 1735 // Okay, we decided to do this! Clone all the instructions in BB onto the end 1736 // of PredBB. 1737 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" 1738 << PredBB->getName() << "' to eliminate branch on phi. Cost: " 1739 << DuplicationCost << " block is:" << *BB << "\n"); 1740 1741 // Unless PredBB ends with an unconditional branch, split the edge so that we 1742 // can just clone the bits from BB into the end of the new PredBB. 1743 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 1744 1745 if (!OldPredBranch || !OldPredBranch->isUnconditional()) { 1746 PredBB = SplitEdge(PredBB, BB); 1747 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 1748 } 1749 1750 // We are going to have to map operands from the original BB block into the 1751 // PredBB block. Evaluate PHI nodes in BB. 1752 DenseMap<Instruction*, Value*> ValueMapping; 1753 1754 BasicBlock::iterator BI = BB->begin(); 1755 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 1756 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 1757 // Clone the non-phi instructions of BB into PredBB, keeping track of the 1758 // mapping and using it to remap operands in the cloned instructions. 1759 for (; BI != BB->end(); ++BI) { 1760 Instruction *New = BI->clone(); 1761 1762 // Remap operands to patch up intra-block references. 1763 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 1764 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 1765 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 1766 if (I != ValueMapping.end()) 1767 New->setOperand(i, I->second); 1768 } 1769 1770 // If this instruction can be simplified after the operands are updated, 1771 // just use the simplified value instead. This frequently happens due to 1772 // phi translation. 1773 if (Value *IV = 1774 SimplifyInstruction(New, BB->getModule()->getDataLayout())) { 1775 delete New; 1776 ValueMapping[&*BI] = IV; 1777 } else { 1778 // Otherwise, insert the new instruction into the block. 1779 New->setName(BI->getName()); 1780 PredBB->getInstList().insert(OldPredBranch->getIterator(), New); 1781 ValueMapping[&*BI] = New; 1782 } 1783 } 1784 1785 // Check to see if the targets of the branch had PHI nodes. If so, we need to 1786 // add entries to the PHI nodes for branch from PredBB now. 1787 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 1788 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 1789 ValueMapping); 1790 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 1791 ValueMapping); 1792 1793 // If there were values defined in BB that are used outside the block, then we 1794 // now have to update all uses of the value to use either the original value, 1795 // the cloned value, or some PHI derived value. This can require arbitrary 1796 // PHI insertion, of which we are prepared to do, clean these up now. 1797 SSAUpdater SSAUpdate; 1798 SmallVector<Use*, 16> UsesToRename; 1799 for (Instruction &I : *BB) { 1800 // Scan all uses of this instruction to see if it is used outside of its 1801 // block, and if so, record them in UsesToRename. 1802 for (Use &U : I.uses()) { 1803 Instruction *User = cast<Instruction>(U.getUser()); 1804 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 1805 if (UserPN->getIncomingBlock(U) == BB) 1806 continue; 1807 } else if (User->getParent() == BB) 1808 continue; 1809 1810 UsesToRename.push_back(&U); 1811 } 1812 1813 // If there are no uses outside the block, we're done with this instruction. 1814 if (UsesToRename.empty()) 1815 continue; 1816 1817 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n"); 1818 1819 // We found a use of I outside of BB. Rename all uses of I that are outside 1820 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 1821 // with the two values we know. 1822 SSAUpdate.Initialize(I.getType(), I.getName()); 1823 SSAUpdate.AddAvailableValue(BB, &I); 1824 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]); 1825 1826 while (!UsesToRename.empty()) 1827 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 1828 DEBUG(dbgs() << "\n"); 1829 } 1830 1831 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 1832 // that we nuked. 1833 BB->removePredecessor(PredBB, true); 1834 1835 // Remove the unconditional branch at the end of the PredBB block. 1836 OldPredBranch->eraseFromParent(); 1837 1838 ++NumDupes; 1839 return true; 1840 } 1841 1842 /// TryToUnfoldSelect - Look for blocks of the form 1843 /// bb1: 1844 /// %a = select 1845 /// br bb 1846 /// 1847 /// bb2: 1848 /// %p = phi [%a, %bb] ... 1849 /// %c = icmp %p 1850 /// br i1 %c 1851 /// 1852 /// And expand the select into a branch structure if one of its arms allows %c 1853 /// to be folded. This later enables threading from bb1 over bb2. 1854 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) { 1855 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 1856 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0)); 1857 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1)); 1858 1859 if (!CondBr || !CondBr->isConditional() || !CondLHS || 1860 CondLHS->getParent() != BB) 1861 return false; 1862 1863 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) { 1864 BasicBlock *Pred = CondLHS->getIncomingBlock(I); 1865 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I)); 1866 1867 // Look if one of the incoming values is a select in the corresponding 1868 // predecessor. 1869 if (!SI || SI->getParent() != Pred || !SI->hasOneUse()) 1870 continue; 1871 1872 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator()); 1873 if (!PredTerm || !PredTerm->isUnconditional()) 1874 continue; 1875 1876 // Now check if one of the select values would allow us to constant fold the 1877 // terminator in BB. We don't do the transform if both sides fold, those 1878 // cases will be threaded in any case. 1879 LazyValueInfo::Tristate LHSFolds = 1880 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1), 1881 CondRHS, Pred, BB, CondCmp); 1882 LazyValueInfo::Tristate RHSFolds = 1883 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2), 1884 CondRHS, Pred, BB, CondCmp); 1885 if ((LHSFolds != LazyValueInfo::Unknown || 1886 RHSFolds != LazyValueInfo::Unknown) && 1887 LHSFolds != RHSFolds) { 1888 // Expand the select. 1889 // 1890 // Pred -- 1891 // | v 1892 // | NewBB 1893 // | | 1894 // |----- 1895 // v 1896 // BB 1897 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold", 1898 BB->getParent(), BB); 1899 // Move the unconditional branch to NewBB. 1900 PredTerm->removeFromParent(); 1901 NewBB->getInstList().insert(NewBB->end(), PredTerm); 1902 // Create a conditional branch and update PHI nodes. 1903 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred); 1904 CondLHS->setIncomingValue(I, SI->getFalseValue()); 1905 CondLHS->addIncoming(SI->getTrueValue(), NewBB); 1906 // The select is now dead. 1907 SI->eraseFromParent(); 1908 1909 // Update any other PHI nodes in BB. 1910 for (BasicBlock::iterator BI = BB->begin(); 1911 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI) 1912 if (Phi != CondLHS) 1913 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB); 1914 return true; 1915 } 1916 } 1917 return false; 1918 } 1919 1920 /// TryToUnfoldSelectInCurrBB - Look for PHI/Select in the same BB of the form 1921 /// bb: 1922 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ... 1923 /// %s = select p, trueval, falseval 1924 /// 1925 /// And expand the select into a branch structure. This later enables 1926 /// jump-threading over bb in this pass. 1927 /// 1928 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold 1929 /// select if the associated PHI has at least one constant. If the unfolded 1930 /// select is not jump-threaded, it will be folded again in the later 1931 /// optimizations. 1932 bool JumpThreading::TryToUnfoldSelectInCurrBB(BasicBlock *BB) { 1933 // If threading this would thread across a loop header, don't thread the edge. 1934 // See the comments above FindLoopHeaders for justifications and caveats. 1935 if (LoopHeaders.count(BB)) 1936 return false; 1937 1938 // Look for a Phi/Select pair in the same basic block. The Phi feeds the 1939 // condition of the Select and at least one of the incoming values is a 1940 // constant. 1941 for (BasicBlock::iterator BI = BB->begin(); 1942 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) { 1943 unsigned NumPHIValues = PN->getNumIncomingValues(); 1944 if (NumPHIValues == 0 || !PN->hasOneUse()) 1945 continue; 1946 1947 SelectInst *SI = dyn_cast<SelectInst>(PN->user_back()); 1948 if (!SI || SI->getParent() != BB) 1949 continue; 1950 1951 Value *Cond = SI->getCondition(); 1952 if (!Cond || Cond != PN || !Cond->getType()->isIntegerTy(1)) 1953 continue; 1954 1955 bool HasConst = false; 1956 for (unsigned i = 0; i != NumPHIValues; ++i) { 1957 if (PN->getIncomingBlock(i) == BB) 1958 return false; 1959 if (isa<ConstantInt>(PN->getIncomingValue(i))) 1960 HasConst = true; 1961 } 1962 1963 if (HasConst) { 1964 // Expand the select. 1965 TerminatorInst *Term = 1966 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false); 1967 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI); 1968 NewPN->addIncoming(SI->getTrueValue(), Term->getParent()); 1969 NewPN->addIncoming(SI->getFalseValue(), BB); 1970 SI->replaceAllUsesWith(NewPN); 1971 SI->eraseFromParent(); 1972 return true; 1973 } 1974 } 1975 1976 return false; 1977 } 1978