1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements the Jump Threading pass. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/Transforms/Scalar/JumpThreading.h" 14 #include "llvm/ADT/DenseMap.h" 15 #include "llvm/ADT/DenseSet.h" 16 #include "llvm/ADT/MapVector.h" 17 #include "llvm/ADT/Optional.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SmallPtrSet.h" 20 #include "llvm/ADT/SmallVector.h" 21 #include "llvm/ADT/Statistic.h" 22 #include "llvm/Analysis/AliasAnalysis.h" 23 #include "llvm/Analysis/BlockFrequencyInfo.h" 24 #include "llvm/Analysis/BranchProbabilityInfo.h" 25 #include "llvm/Analysis/CFG.h" 26 #include "llvm/Analysis/ConstantFolding.h" 27 #include "llvm/Analysis/DomTreeUpdater.h" 28 #include "llvm/Analysis/GlobalsModRef.h" 29 #include "llvm/Analysis/GuardUtils.h" 30 #include "llvm/Analysis/InstructionSimplify.h" 31 #include "llvm/Analysis/LazyValueInfo.h" 32 #include "llvm/Analysis/Loads.h" 33 #include "llvm/Analysis/LoopInfo.h" 34 #include "llvm/Analysis/MemoryLocation.h" 35 #include "llvm/Analysis/TargetLibraryInfo.h" 36 #include "llvm/Analysis/TargetTransformInfo.h" 37 #include "llvm/Analysis/ValueTracking.h" 38 #include "llvm/IR/BasicBlock.h" 39 #include "llvm/IR/CFG.h" 40 #include "llvm/IR/Constant.h" 41 #include "llvm/IR/ConstantRange.h" 42 #include "llvm/IR/Constants.h" 43 #include "llvm/IR/DataLayout.h" 44 #include "llvm/IR/Dominators.h" 45 #include "llvm/IR/Function.h" 46 #include "llvm/IR/InstrTypes.h" 47 #include "llvm/IR/Instruction.h" 48 #include "llvm/IR/Instructions.h" 49 #include "llvm/IR/IntrinsicInst.h" 50 #include "llvm/IR/Intrinsics.h" 51 #include "llvm/IR/LLVMContext.h" 52 #include "llvm/IR/MDBuilder.h" 53 #include "llvm/IR/Metadata.h" 54 #include "llvm/IR/Module.h" 55 #include "llvm/IR/PassManager.h" 56 #include "llvm/IR/PatternMatch.h" 57 #include "llvm/IR/Type.h" 58 #include "llvm/IR/Use.h" 59 #include "llvm/IR/Value.h" 60 #include "llvm/InitializePasses.h" 61 #include "llvm/Pass.h" 62 #include "llvm/Support/BlockFrequency.h" 63 #include "llvm/Support/BranchProbability.h" 64 #include "llvm/Support/Casting.h" 65 #include "llvm/Support/CommandLine.h" 66 #include "llvm/Support/Debug.h" 67 #include "llvm/Support/raw_ostream.h" 68 #include "llvm/Transforms/Scalar.h" 69 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 70 #include "llvm/Transforms/Utils/Cloning.h" 71 #include "llvm/Transforms/Utils/Local.h" 72 #include "llvm/Transforms/Utils/SSAUpdater.h" 73 #include "llvm/Transforms/Utils/ValueMapper.h" 74 #include <algorithm> 75 #include <cassert> 76 #include <cstdint> 77 #include <iterator> 78 #include <memory> 79 #include <utility> 80 81 using namespace llvm; 82 using namespace jumpthreading; 83 84 #define DEBUG_TYPE "jump-threading" 85 86 STATISTIC(NumThreads, "Number of jumps threaded"); 87 STATISTIC(NumFolds, "Number of terminators folded"); 88 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi"); 89 90 static cl::opt<unsigned> 91 BBDuplicateThreshold("jump-threading-threshold", 92 cl::desc("Max block size to duplicate for jump threading"), 93 cl::init(6), cl::Hidden); 94 95 static cl::opt<unsigned> 96 ImplicationSearchThreshold( 97 "jump-threading-implication-search-threshold", 98 cl::desc("The number of predecessors to search for a stronger " 99 "condition to use to thread over a weaker condition"), 100 cl::init(3), cl::Hidden); 101 102 static cl::opt<bool> PrintLVIAfterJumpThreading( 103 "print-lvi-after-jump-threading", 104 cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false), 105 cl::Hidden); 106 107 static cl::opt<bool> ThreadAcrossLoopHeaders( 108 "jump-threading-across-loop-headers", 109 cl::desc("Allow JumpThreading to thread across loop headers, for testing"), 110 cl::init(false), cl::Hidden); 111 112 113 namespace { 114 115 /// This pass performs 'jump threading', which looks at blocks that have 116 /// multiple predecessors and multiple successors. If one or more of the 117 /// predecessors of the block can be proven to always jump to one of the 118 /// successors, we forward the edge from the predecessor to the successor by 119 /// duplicating the contents of this block. 120 /// 121 /// An example of when this can occur is code like this: 122 /// 123 /// if () { ... 124 /// X = 4; 125 /// } 126 /// if (X < 3) { 127 /// 128 /// In this case, the unconditional branch at the end of the first if can be 129 /// revectored to the false side of the second if. 130 class JumpThreading : public FunctionPass { 131 JumpThreadingPass Impl; 132 133 public: 134 static char ID; // Pass identification 135 136 JumpThreading(int T = -1) : FunctionPass(ID), Impl(T) { 137 initializeJumpThreadingPass(*PassRegistry::getPassRegistry()); 138 } 139 140 bool runOnFunction(Function &F) override; 141 142 void getAnalysisUsage(AnalysisUsage &AU) const override { 143 AU.addRequired<DominatorTreeWrapperPass>(); 144 AU.addPreserved<DominatorTreeWrapperPass>(); 145 AU.addRequired<AAResultsWrapperPass>(); 146 AU.addRequired<LazyValueInfoWrapperPass>(); 147 AU.addPreserved<LazyValueInfoWrapperPass>(); 148 AU.addPreserved<GlobalsAAWrapperPass>(); 149 AU.addRequired<TargetLibraryInfoWrapperPass>(); 150 AU.addRequired<TargetTransformInfoWrapperPass>(); 151 } 152 153 void releaseMemory() override { Impl.releaseMemory(); } 154 }; 155 156 } // end anonymous namespace 157 158 char JumpThreading::ID = 0; 159 160 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", 161 "Jump Threading", false, false) 162 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 163 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass) 164 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 165 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 166 INITIALIZE_PASS_END(JumpThreading, "jump-threading", 167 "Jump Threading", false, false) 168 169 // Public interface to the Jump Threading pass 170 FunctionPass *llvm::createJumpThreadingPass(int Threshold) { 171 return new JumpThreading(Threshold); 172 } 173 174 JumpThreadingPass::JumpThreadingPass(int T) { 175 DefaultBBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T); 176 } 177 178 // Update branch probability information according to conditional 179 // branch probability. This is usually made possible for cloned branches 180 // in inline instances by the context specific profile in the caller. 181 // For instance, 182 // 183 // [Block PredBB] 184 // [Branch PredBr] 185 // if (t) { 186 // Block A; 187 // } else { 188 // Block B; 189 // } 190 // 191 // [Block BB] 192 // cond = PN([true, %A], [..., %B]); // PHI node 193 // [Branch CondBr] 194 // if (cond) { 195 // ... // P(cond == true) = 1% 196 // } 197 // 198 // Here we know that when block A is taken, cond must be true, which means 199 // P(cond == true | A) = 1 200 // 201 // Given that P(cond == true) = P(cond == true | A) * P(A) + 202 // P(cond == true | B) * P(B) 203 // we get: 204 // P(cond == true ) = P(A) + P(cond == true | B) * P(B) 205 // 206 // which gives us: 207 // P(A) is less than P(cond == true), i.e. 208 // P(t == true) <= P(cond == true) 209 // 210 // In other words, if we know P(cond == true) is unlikely, we know 211 // that P(t == true) is also unlikely. 212 // 213 static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) { 214 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 215 if (!CondBr) 216 return; 217 218 uint64_t TrueWeight, FalseWeight; 219 if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight)) 220 return; 221 222 if (TrueWeight + FalseWeight == 0) 223 // Zero branch_weights do not give a hint for getting branch probabilities. 224 // Technically it would result in division by zero denominator, which is 225 // TrueWeight + FalseWeight. 226 return; 227 228 // Returns the outgoing edge of the dominating predecessor block 229 // that leads to the PhiNode's incoming block: 230 auto GetPredOutEdge = 231 [](BasicBlock *IncomingBB, 232 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> { 233 auto *PredBB = IncomingBB; 234 auto *SuccBB = PhiBB; 235 SmallPtrSet<BasicBlock *, 16> Visited; 236 while (true) { 237 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()); 238 if (PredBr && PredBr->isConditional()) 239 return {PredBB, SuccBB}; 240 Visited.insert(PredBB); 241 auto *SinglePredBB = PredBB->getSinglePredecessor(); 242 if (!SinglePredBB) 243 return {nullptr, nullptr}; 244 245 // Stop searching when SinglePredBB has been visited. It means we see 246 // an unreachable loop. 247 if (Visited.count(SinglePredBB)) 248 return {nullptr, nullptr}; 249 250 SuccBB = PredBB; 251 PredBB = SinglePredBB; 252 } 253 }; 254 255 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 256 Value *PhiOpnd = PN->getIncomingValue(i); 257 ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd); 258 259 if (!CI || !CI->getType()->isIntegerTy(1)) 260 continue; 261 262 BranchProbability BP = 263 (CI->isOne() ? BranchProbability::getBranchProbability( 264 TrueWeight, TrueWeight + FalseWeight) 265 : BranchProbability::getBranchProbability( 266 FalseWeight, TrueWeight + FalseWeight)); 267 268 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB); 269 if (!PredOutEdge.first) 270 return; 271 272 BasicBlock *PredBB = PredOutEdge.first; 273 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()); 274 if (!PredBr) 275 return; 276 277 uint64_t PredTrueWeight, PredFalseWeight; 278 // FIXME: We currently only set the profile data when it is missing. 279 // With PGO, this can be used to refine even existing profile data with 280 // context information. This needs to be done after more performance 281 // testing. 282 if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight)) 283 continue; 284 285 // We can not infer anything useful when BP >= 50%, because BP is the 286 // upper bound probability value. 287 if (BP >= BranchProbability(50, 100)) 288 continue; 289 290 SmallVector<uint32_t, 2> Weights; 291 if (PredBr->getSuccessor(0) == PredOutEdge.second) { 292 Weights.push_back(BP.getNumerator()); 293 Weights.push_back(BP.getCompl().getNumerator()); 294 } else { 295 Weights.push_back(BP.getCompl().getNumerator()); 296 Weights.push_back(BP.getNumerator()); 297 } 298 PredBr->setMetadata(LLVMContext::MD_prof, 299 MDBuilder(PredBr->getParent()->getContext()) 300 .createBranchWeights(Weights)); 301 } 302 } 303 304 /// runOnFunction - Toplevel algorithm. 305 bool JumpThreading::runOnFunction(Function &F) { 306 if (skipFunction(F)) 307 return false; 308 auto TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 309 // Jump Threading has no sense for the targets with divergent CF 310 if (TTI->hasBranchDivergence()) 311 return false; 312 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 313 auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 314 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI(); 315 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 316 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy); 317 std::unique_ptr<BlockFrequencyInfo> BFI; 318 std::unique_ptr<BranchProbabilityInfo> BPI; 319 if (F.hasProfileData()) { 320 LoopInfo LI{*DT}; 321 BPI.reset(new BranchProbabilityInfo(F, LI, TLI)); 322 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI)); 323 } 324 325 bool Changed = Impl.runImpl(F, TLI, TTI, LVI, AA, &DTU, F.hasProfileData(), 326 std::move(BFI), std::move(BPI)); 327 if (PrintLVIAfterJumpThreading) { 328 dbgs() << "LVI for function '" << F.getName() << "':\n"; 329 LVI->printLVI(F, DTU.getDomTree(), dbgs()); 330 } 331 return Changed; 332 } 333 334 PreservedAnalyses JumpThreadingPass::run(Function &F, 335 FunctionAnalysisManager &AM) { 336 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 337 // Jump Threading has no sense for the targets with divergent CF 338 if (TTI.hasBranchDivergence()) 339 return PreservedAnalyses::all(); 340 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 341 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 342 auto &LVI = AM.getResult<LazyValueAnalysis>(F); 343 auto &AA = AM.getResult<AAManager>(F); 344 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy); 345 346 std::unique_ptr<BlockFrequencyInfo> BFI; 347 std::unique_ptr<BranchProbabilityInfo> BPI; 348 if (F.hasProfileData()) { 349 LoopInfo LI{DominatorTree(F)}; 350 BPI.reset(new BranchProbabilityInfo(F, LI, &TLI)); 351 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI)); 352 } 353 354 bool Changed = runImpl(F, &TLI, &TTI, &LVI, &AA, &DTU, F.hasProfileData(), 355 std::move(BFI), std::move(BPI)); 356 357 if (PrintLVIAfterJumpThreading) { 358 dbgs() << "LVI for function '" << F.getName() << "':\n"; 359 LVI.printLVI(F, DTU.getDomTree(), dbgs()); 360 } 361 362 if (!Changed) 363 return PreservedAnalyses::all(); 364 PreservedAnalyses PA; 365 PA.preserve<DominatorTreeAnalysis>(); 366 PA.preserve<LazyValueAnalysis>(); 367 return PA; 368 } 369 370 bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_, 371 TargetTransformInfo *TTI_, LazyValueInfo *LVI_, 372 AliasAnalysis *AA_, DomTreeUpdater *DTU_, 373 bool HasProfileData_, 374 std::unique_ptr<BlockFrequencyInfo> BFI_, 375 std::unique_ptr<BranchProbabilityInfo> BPI_) { 376 LLVM_DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n"); 377 TLI = TLI_; 378 TTI = TTI_; 379 LVI = LVI_; 380 AA = AA_; 381 DTU = DTU_; 382 BFI.reset(); 383 BPI.reset(); 384 // When profile data is available, we need to update edge weights after 385 // successful jump threading, which requires both BPI and BFI being available. 386 HasProfileData = HasProfileData_; 387 auto *GuardDecl = F.getParent()->getFunction( 388 Intrinsic::getName(Intrinsic::experimental_guard)); 389 HasGuards = GuardDecl && !GuardDecl->use_empty(); 390 if (HasProfileData) { 391 BPI = std::move(BPI_); 392 BFI = std::move(BFI_); 393 } 394 395 // Reduce the number of instructions duplicated when optimizing strictly for 396 // size. 397 if (BBDuplicateThreshold.getNumOccurrences()) 398 BBDupThreshold = BBDuplicateThreshold; 399 else if (F.hasFnAttribute(Attribute::MinSize)) 400 BBDupThreshold = 3; 401 else 402 BBDupThreshold = DefaultBBDupThreshold; 403 404 // JumpThreading must not processes blocks unreachable from entry. It's a 405 // waste of compute time and can potentially lead to hangs. 406 SmallPtrSet<BasicBlock *, 16> Unreachable; 407 assert(DTU && "DTU isn't passed into JumpThreading before using it."); 408 assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed."); 409 DominatorTree &DT = DTU->getDomTree(); 410 for (auto &BB : F) 411 if (!DT.isReachableFromEntry(&BB)) 412 Unreachable.insert(&BB); 413 414 if (!ThreadAcrossLoopHeaders) 415 findLoopHeaders(F); 416 417 bool EverChanged = false; 418 bool Changed; 419 do { 420 Changed = false; 421 for (auto &BB : F) { 422 if (Unreachable.count(&BB)) 423 continue; 424 while (processBlock(&BB)) // Thread all of the branches we can over BB. 425 Changed = true; 426 427 // Jump threading may have introduced redundant debug values into BB 428 // which should be removed. 429 if (Changed) 430 RemoveRedundantDbgInstrs(&BB); 431 432 // Stop processing BB if it's the entry or is now deleted. The following 433 // routines attempt to eliminate BB and locating a suitable replacement 434 // for the entry is non-trivial. 435 if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB)) 436 continue; 437 438 if (pred_empty(&BB)) { 439 // When processBlock makes BB unreachable it doesn't bother to fix up 440 // the instructions in it. We must remove BB to prevent invalid IR. 441 LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName() 442 << "' with terminator: " << *BB.getTerminator() 443 << '\n'); 444 LoopHeaders.erase(&BB); 445 LVI->eraseBlock(&BB); 446 DeleteDeadBlock(&BB, DTU); 447 Changed = true; 448 continue; 449 } 450 451 // processBlock doesn't thread BBs with unconditional TIs. However, if BB 452 // is "almost empty", we attempt to merge BB with its sole successor. 453 auto *BI = dyn_cast<BranchInst>(BB.getTerminator()); 454 if (BI && BI->isUnconditional()) { 455 BasicBlock *Succ = BI->getSuccessor(0); 456 if ( 457 // The terminator must be the only non-phi instruction in BB. 458 BB.getFirstNonPHIOrDbg(true)->isTerminator() && 459 // Don't alter Loop headers and latches to ensure another pass can 460 // detect and transform nested loops later. 461 !LoopHeaders.count(&BB) && !LoopHeaders.count(Succ) && 462 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU)) { 463 RemoveRedundantDbgInstrs(Succ); 464 // BB is valid for cleanup here because we passed in DTU. F remains 465 // BB's parent until a DTU->getDomTree() event. 466 LVI->eraseBlock(&BB); 467 Changed = true; 468 } 469 } 470 } 471 EverChanged |= Changed; 472 } while (Changed); 473 474 LoopHeaders.clear(); 475 return EverChanged; 476 } 477 478 // Replace uses of Cond with ToVal when safe to do so. If all uses are 479 // replaced, we can remove Cond. We cannot blindly replace all uses of Cond 480 // because we may incorrectly replace uses when guards/assumes are uses of 481 // of `Cond` and we used the guards/assume to reason about the `Cond` value 482 // at the end of block. RAUW unconditionally replaces all uses 483 // including the guards/assumes themselves and the uses before the 484 // guard/assume. 485 static bool replaceFoldableUses(Instruction *Cond, Value *ToVal, 486 BasicBlock *KnownAtEndOfBB) { 487 bool Changed = false; 488 assert(Cond->getType() == ToVal->getType()); 489 // We can unconditionally replace all uses in non-local blocks (i.e. uses 490 // strictly dominated by BB), since LVI information is true from the 491 // terminator of BB. 492 if (Cond->getParent() == KnownAtEndOfBB) 493 Changed |= replaceNonLocalUsesWith(Cond, ToVal); 494 for (Instruction &I : reverse(*KnownAtEndOfBB)) { 495 // Reached the Cond whose uses we are trying to replace, so there are no 496 // more uses. 497 if (&I == Cond) 498 break; 499 // We only replace uses in instructions that are guaranteed to reach the end 500 // of BB, where we know Cond is ToVal. 501 if (!isGuaranteedToTransferExecutionToSuccessor(&I)) 502 break; 503 Changed |= I.replaceUsesOfWith(Cond, ToVal); 504 } 505 if (Cond->use_empty() && !Cond->mayHaveSideEffects()) { 506 Cond->eraseFromParent(); 507 Changed = true; 508 } 509 return Changed; 510 } 511 512 /// Return the cost of duplicating a piece of this block from first non-phi 513 /// and before StopAt instruction to thread across it. Stop scanning the block 514 /// when exceeding the threshold. If duplication is impossible, returns ~0U. 515 static unsigned getJumpThreadDuplicationCost(const TargetTransformInfo *TTI, 516 BasicBlock *BB, 517 Instruction *StopAt, 518 unsigned Threshold) { 519 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?"); 520 /// Ignore PHI nodes, these will be flattened when duplication happens. 521 BasicBlock::const_iterator I(BB->getFirstNonPHI()); 522 523 // FIXME: THREADING will delete values that are just used to compute the 524 // branch, so they shouldn't count against the duplication cost. 525 526 unsigned Bonus = 0; 527 if (BB->getTerminator() == StopAt) { 528 // Threading through a switch statement is particularly profitable. If this 529 // block ends in a switch, decrease its cost to make it more likely to 530 // happen. 531 if (isa<SwitchInst>(StopAt)) 532 Bonus = 6; 533 534 // The same holds for indirect branches, but slightly more so. 535 if (isa<IndirectBrInst>(StopAt)) 536 Bonus = 8; 537 } 538 539 // Bump the threshold up so the early exit from the loop doesn't skip the 540 // terminator-based Size adjustment at the end. 541 Threshold += Bonus; 542 543 // Sum up the cost of each instruction until we get to the terminator. Don't 544 // include the terminator because the copy won't include it. 545 unsigned Size = 0; 546 for (; &*I != StopAt; ++I) { 547 548 // Stop scanning the block if we've reached the threshold. 549 if (Size > Threshold) 550 return Size; 551 552 // Bail out if this instruction gives back a token type, it is not possible 553 // to duplicate it if it is used outside this BB. 554 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB)) 555 return ~0U; 556 557 // Blocks with NoDuplicate are modelled as having infinite cost, so they 558 // are never duplicated. 559 if (const CallInst *CI = dyn_cast<CallInst>(I)) 560 if (CI->cannotDuplicate() || CI->isConvergent()) 561 return ~0U; 562 563 if (TTI->getUserCost(&*I, TargetTransformInfo::TCK_SizeAndLatency) 564 == TargetTransformInfo::TCC_Free) 565 continue; 566 567 // All other instructions count for at least one unit. 568 ++Size; 569 570 // Calls are more expensive. If they are non-intrinsic calls, we model them 571 // as having cost of 4. If they are a non-vector intrinsic, we model them 572 // as having cost of 2 total, and if they are a vector intrinsic, we model 573 // them as having cost 1. 574 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 575 if (!isa<IntrinsicInst>(CI)) 576 Size += 3; 577 else if (!CI->getType()->isVectorTy()) 578 Size += 1; 579 } 580 } 581 582 return Size > Bonus ? Size - Bonus : 0; 583 } 584 585 /// findLoopHeaders - We do not want jump threading to turn proper loop 586 /// structures into irreducible loops. Doing this breaks up the loop nesting 587 /// hierarchy and pessimizes later transformations. To prevent this from 588 /// happening, we first have to find the loop headers. Here we approximate this 589 /// by finding targets of backedges in the CFG. 590 /// 591 /// Note that there definitely are cases when we want to allow threading of 592 /// edges across a loop header. For example, threading a jump from outside the 593 /// loop (the preheader) to an exit block of the loop is definitely profitable. 594 /// It is also almost always profitable to thread backedges from within the loop 595 /// to exit blocks, and is often profitable to thread backedges to other blocks 596 /// within the loop (forming a nested loop). This simple analysis is not rich 597 /// enough to track all of these properties and keep it up-to-date as the CFG 598 /// mutates, so we don't allow any of these transformations. 599 void JumpThreadingPass::findLoopHeaders(Function &F) { 600 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 601 FindFunctionBackedges(F, Edges); 602 603 for (const auto &Edge : Edges) 604 LoopHeaders.insert(Edge.second); 605 } 606 607 /// getKnownConstant - Helper method to determine if we can thread over a 608 /// terminator with the given value as its condition, and if so what value to 609 /// use for that. What kind of value this is depends on whether we want an 610 /// integer or a block address, but an undef is always accepted. 611 /// Returns null if Val is null or not an appropriate constant. 612 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { 613 if (!Val) 614 return nullptr; 615 616 // Undef is "known" enough. 617 if (UndefValue *U = dyn_cast<UndefValue>(Val)) 618 return U; 619 620 if (Preference == WantBlockAddress) 621 return dyn_cast<BlockAddress>(Val->stripPointerCasts()); 622 623 return dyn_cast<ConstantInt>(Val); 624 } 625 626 /// computeValueKnownInPredecessors - Given a basic block BB and a value V, see 627 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef 628 /// in any of our predecessors. If so, return the known list of value and pred 629 /// BB in the result vector. 630 /// 631 /// This returns true if there were any known values. 632 bool JumpThreadingPass::computeValueKnownInPredecessorsImpl( 633 Value *V, BasicBlock *BB, PredValueInfo &Result, 634 ConstantPreference Preference, DenseSet<Value *> &RecursionSet, 635 Instruction *CxtI) { 636 // This method walks up use-def chains recursively. Because of this, we could 637 // get into an infinite loop going around loops in the use-def chain. To 638 // prevent this, keep track of what (value, block) pairs we've already visited 639 // and terminate the search if we loop back to them 640 if (!RecursionSet.insert(V).second) 641 return false; 642 643 // If V is a constant, then it is known in all predecessors. 644 if (Constant *KC = getKnownConstant(V, Preference)) { 645 for (BasicBlock *Pred : predecessors(BB)) 646 Result.emplace_back(KC, Pred); 647 648 return !Result.empty(); 649 } 650 651 // If V is a non-instruction value, or an instruction in a different block, 652 // then it can't be derived from a PHI. 653 Instruction *I = dyn_cast<Instruction>(V); 654 if (!I || I->getParent() != BB) { 655 656 // Okay, if this is a live-in value, see if it has a known value at the end 657 // of any of our predecessors. 658 // 659 // FIXME: This should be an edge property, not a block end property. 660 /// TODO: Per PR2563, we could infer value range information about a 661 /// predecessor based on its terminator. 662 // 663 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if 664 // "I" is a non-local compare-with-a-constant instruction. This would be 665 // able to handle value inequalities better, for example if the compare is 666 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. 667 // Perhaps getConstantOnEdge should be smart enough to do this? 668 for (BasicBlock *P : predecessors(BB)) { 669 // If the value is known by LazyValueInfo to be a constant in a 670 // predecessor, use that information to try to thread this block. 671 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI); 672 if (Constant *KC = getKnownConstant(PredCst, Preference)) 673 Result.emplace_back(KC, P); 674 } 675 676 return !Result.empty(); 677 } 678 679 /// If I is a PHI node, then we know the incoming values for any constants. 680 if (PHINode *PN = dyn_cast<PHINode>(I)) { 681 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 682 Value *InVal = PN->getIncomingValue(i); 683 if (Constant *KC = getKnownConstant(InVal, Preference)) { 684 Result.emplace_back(KC, PN->getIncomingBlock(i)); 685 } else { 686 Constant *CI = LVI->getConstantOnEdge(InVal, 687 PN->getIncomingBlock(i), 688 BB, CxtI); 689 if (Constant *KC = getKnownConstant(CI, Preference)) 690 Result.emplace_back(KC, PN->getIncomingBlock(i)); 691 } 692 } 693 694 return !Result.empty(); 695 } 696 697 // Handle Cast instructions. 698 if (CastInst *CI = dyn_cast<CastInst>(I)) { 699 Value *Source = CI->getOperand(0); 700 computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference, 701 RecursionSet, CxtI); 702 if (Result.empty()) 703 return false; 704 705 // Convert the known values. 706 for (auto &R : Result) 707 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType()); 708 709 return true; 710 } 711 712 if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) { 713 Value *Source = FI->getOperand(0); 714 computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference, 715 RecursionSet, CxtI); 716 717 erase_if(Result, [](auto &Pair) { 718 return !isGuaranteedNotToBeUndefOrPoison(Pair.first); 719 }); 720 721 return !Result.empty(); 722 } 723 724 // Handle some boolean conditions. 725 if (I->getType()->getPrimitiveSizeInBits() == 1) { 726 using namespace PatternMatch; 727 if (Preference != WantInteger) 728 return false; 729 // X | true -> true 730 // X & false -> false 731 Value *Op0, *Op1; 732 if (match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))) || 733 match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) { 734 PredValueInfoTy LHSVals, RHSVals; 735 736 computeValueKnownInPredecessorsImpl(Op0, BB, LHSVals, WantInteger, 737 RecursionSet, CxtI); 738 computeValueKnownInPredecessorsImpl(Op1, BB, RHSVals, WantInteger, 739 RecursionSet, CxtI); 740 741 if (LHSVals.empty() && RHSVals.empty()) 742 return false; 743 744 ConstantInt *InterestingVal; 745 if (match(I, m_LogicalOr())) 746 InterestingVal = ConstantInt::getTrue(I->getContext()); 747 else 748 InterestingVal = ConstantInt::getFalse(I->getContext()); 749 750 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; 751 752 // Scan for the sentinel. If we find an undef, force it to the 753 // interesting value: x|undef -> true and x&undef -> false. 754 for (const auto &LHSVal : LHSVals) 755 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) { 756 Result.emplace_back(InterestingVal, LHSVal.second); 757 LHSKnownBBs.insert(LHSVal.second); 758 } 759 for (const auto &RHSVal : RHSVals) 760 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) { 761 // If we already inferred a value for this block on the LHS, don't 762 // re-add it. 763 if (!LHSKnownBBs.count(RHSVal.second)) 764 Result.emplace_back(InterestingVal, RHSVal.second); 765 } 766 767 return !Result.empty(); 768 } 769 770 // Handle the NOT form of XOR. 771 if (I->getOpcode() == Instruction::Xor && 772 isa<ConstantInt>(I->getOperand(1)) && 773 cast<ConstantInt>(I->getOperand(1))->isOne()) { 774 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result, 775 WantInteger, RecursionSet, CxtI); 776 if (Result.empty()) 777 return false; 778 779 // Invert the known values. 780 for (auto &R : Result) 781 R.first = ConstantExpr::getNot(R.first); 782 783 return true; 784 } 785 786 // Try to simplify some other binary operator values. 787 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 788 if (Preference != WantInteger) 789 return false; 790 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { 791 const DataLayout &DL = BO->getModule()->getDataLayout(); 792 PredValueInfoTy LHSVals; 793 computeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals, 794 WantInteger, RecursionSet, CxtI); 795 796 // Try to use constant folding to simplify the binary operator. 797 for (const auto &LHSVal : LHSVals) { 798 Constant *V = LHSVal.first; 799 Constant *Folded = 800 ConstantFoldBinaryOpOperands(BO->getOpcode(), V, CI, DL); 801 802 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 803 Result.emplace_back(KC, LHSVal.second); 804 } 805 } 806 807 return !Result.empty(); 808 } 809 810 // Handle compare with phi operand, where the PHI is defined in this block. 811 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { 812 if (Preference != WantInteger) 813 return false; 814 Type *CmpType = Cmp->getType(); 815 Value *CmpLHS = Cmp->getOperand(0); 816 Value *CmpRHS = Cmp->getOperand(1); 817 CmpInst::Predicate Pred = Cmp->getPredicate(); 818 819 PHINode *PN = dyn_cast<PHINode>(CmpLHS); 820 if (!PN) 821 PN = dyn_cast<PHINode>(CmpRHS); 822 if (PN && PN->getParent() == BB) { 823 const DataLayout &DL = PN->getModule()->getDataLayout(); 824 // We can do this simplification if any comparisons fold to true or false. 825 // See if any do. 826 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 827 BasicBlock *PredBB = PN->getIncomingBlock(i); 828 Value *LHS, *RHS; 829 if (PN == CmpLHS) { 830 LHS = PN->getIncomingValue(i); 831 RHS = CmpRHS->DoPHITranslation(BB, PredBB); 832 } else { 833 LHS = CmpLHS->DoPHITranslation(BB, PredBB); 834 RHS = PN->getIncomingValue(i); 835 } 836 Value *Res = simplifyCmpInst(Pred, LHS, RHS, {DL}); 837 if (!Res) { 838 if (!isa<Constant>(RHS)) 839 continue; 840 841 // getPredicateOnEdge call will make no sense if LHS is defined in BB. 842 auto LHSInst = dyn_cast<Instruction>(LHS); 843 if (LHSInst && LHSInst->getParent() == BB) 844 continue; 845 846 LazyValueInfo::Tristate 847 ResT = LVI->getPredicateOnEdge(Pred, LHS, 848 cast<Constant>(RHS), PredBB, BB, 849 CxtI ? CxtI : Cmp); 850 if (ResT == LazyValueInfo::Unknown) 851 continue; 852 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); 853 } 854 855 if (Constant *KC = getKnownConstant(Res, WantInteger)) 856 Result.emplace_back(KC, PredBB); 857 } 858 859 return !Result.empty(); 860 } 861 862 // If comparing a live-in value against a constant, see if we know the 863 // live-in value on any predecessors. 864 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) { 865 Constant *CmpConst = cast<Constant>(CmpRHS); 866 867 if (!isa<Instruction>(CmpLHS) || 868 cast<Instruction>(CmpLHS)->getParent() != BB) { 869 for (BasicBlock *P : predecessors(BB)) { 870 // If the value is known by LazyValueInfo to be a constant in a 871 // predecessor, use that information to try to thread this block. 872 LazyValueInfo::Tristate Res = 873 LVI->getPredicateOnEdge(Pred, CmpLHS, 874 CmpConst, P, BB, CxtI ? CxtI : Cmp); 875 if (Res == LazyValueInfo::Unknown) 876 continue; 877 878 Constant *ResC = ConstantInt::get(CmpType, Res); 879 Result.emplace_back(ResC, P); 880 } 881 882 return !Result.empty(); 883 } 884 885 // InstCombine can fold some forms of constant range checks into 886 // (icmp (add (x, C1)), C2). See if we have we have such a thing with 887 // x as a live-in. 888 { 889 using namespace PatternMatch; 890 891 Value *AddLHS; 892 ConstantInt *AddConst; 893 if (isa<ConstantInt>(CmpConst) && 894 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) { 895 if (!isa<Instruction>(AddLHS) || 896 cast<Instruction>(AddLHS)->getParent() != BB) { 897 for (BasicBlock *P : predecessors(BB)) { 898 // If the value is known by LazyValueInfo to be a ConstantRange in 899 // a predecessor, use that information to try to thread this 900 // block. 901 ConstantRange CR = LVI->getConstantRangeOnEdge( 902 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS)); 903 // Propagate the range through the addition. 904 CR = CR.add(AddConst->getValue()); 905 906 // Get the range where the compare returns true. 907 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion( 908 Pred, cast<ConstantInt>(CmpConst)->getValue()); 909 910 Constant *ResC; 911 if (CmpRange.contains(CR)) 912 ResC = ConstantInt::getTrue(CmpType); 913 else if (CmpRange.inverse().contains(CR)) 914 ResC = ConstantInt::getFalse(CmpType); 915 else 916 continue; 917 918 Result.emplace_back(ResC, P); 919 } 920 921 return !Result.empty(); 922 } 923 } 924 } 925 926 // Try to find a constant value for the LHS of a comparison, 927 // and evaluate it statically if we can. 928 PredValueInfoTy LHSVals; 929 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals, 930 WantInteger, RecursionSet, CxtI); 931 932 for (const auto &LHSVal : LHSVals) { 933 Constant *V = LHSVal.first; 934 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst); 935 if (Constant *KC = getKnownConstant(Folded, WantInteger)) 936 Result.emplace_back(KC, LHSVal.second); 937 } 938 939 return !Result.empty(); 940 } 941 } 942 943 if (SelectInst *SI = dyn_cast<SelectInst>(I)) { 944 // Handle select instructions where at least one operand is a known constant 945 // and we can figure out the condition value for any predecessor block. 946 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference); 947 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference); 948 PredValueInfoTy Conds; 949 if ((TrueVal || FalseVal) && 950 computeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds, 951 WantInteger, RecursionSet, CxtI)) { 952 for (auto &C : Conds) { 953 Constant *Cond = C.first; 954 955 // Figure out what value to use for the condition. 956 bool KnownCond; 957 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) { 958 // A known boolean. 959 KnownCond = CI->isOne(); 960 } else { 961 assert(isa<UndefValue>(Cond) && "Unexpected condition value"); 962 // Either operand will do, so be sure to pick the one that's a known 963 // constant. 964 // FIXME: Do this more cleverly if both values are known constants? 965 KnownCond = (TrueVal != nullptr); 966 } 967 968 // See if the select has a known constant value for this predecessor. 969 if (Constant *Val = KnownCond ? TrueVal : FalseVal) 970 Result.emplace_back(Val, C.second); 971 } 972 973 return !Result.empty(); 974 } 975 } 976 977 // If all else fails, see if LVI can figure out a constant value for us. 978 assert(CxtI->getParent() == BB && "CxtI should be in BB"); 979 Constant *CI = LVI->getConstant(V, CxtI); 980 if (Constant *KC = getKnownConstant(CI, Preference)) { 981 for (BasicBlock *Pred : predecessors(BB)) 982 Result.emplace_back(KC, Pred); 983 } 984 985 return !Result.empty(); 986 } 987 988 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends 989 /// in an undefined jump, decide which block is best to revector to. 990 /// 991 /// Since we can pick an arbitrary destination, we pick the successor with the 992 /// fewest predecessors. This should reduce the in-degree of the others. 993 static unsigned getBestDestForJumpOnUndef(BasicBlock *BB) { 994 Instruction *BBTerm = BB->getTerminator(); 995 unsigned MinSucc = 0; 996 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 997 // Compute the successor with the minimum number of predecessors. 998 unsigned MinNumPreds = pred_size(TestBB); 999 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 1000 TestBB = BBTerm->getSuccessor(i); 1001 unsigned NumPreds = pred_size(TestBB); 1002 if (NumPreds < MinNumPreds) { 1003 MinSucc = i; 1004 MinNumPreds = NumPreds; 1005 } 1006 } 1007 1008 return MinSucc; 1009 } 1010 1011 static bool hasAddressTakenAndUsed(BasicBlock *BB) { 1012 if (!BB->hasAddressTaken()) return false; 1013 1014 // If the block has its address taken, it may be a tree of dead constants 1015 // hanging off of it. These shouldn't keep the block alive. 1016 BlockAddress *BA = BlockAddress::get(BB); 1017 BA->removeDeadConstantUsers(); 1018 return !BA->use_empty(); 1019 } 1020 1021 /// processBlock - If there are any predecessors whose control can be threaded 1022 /// through to a successor, transform them now. 1023 bool JumpThreadingPass::processBlock(BasicBlock *BB) { 1024 // If the block is trivially dead, just return and let the caller nuke it. 1025 // This simplifies other transformations. 1026 if (DTU->isBBPendingDeletion(BB) || 1027 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock())) 1028 return false; 1029 1030 // If this block has a single predecessor, and if that pred has a single 1031 // successor, merge the blocks. This encourages recursive jump threading 1032 // because now the condition in this block can be threaded through 1033 // predecessors of our predecessor block. 1034 if (maybeMergeBasicBlockIntoOnlyPred(BB)) 1035 return true; 1036 1037 if (tryToUnfoldSelectInCurrBB(BB)) 1038 return true; 1039 1040 // Look if we can propagate guards to predecessors. 1041 if (HasGuards && processGuards(BB)) 1042 return true; 1043 1044 // What kind of constant we're looking for. 1045 ConstantPreference Preference = WantInteger; 1046 1047 // Look to see if the terminator is a conditional branch, switch or indirect 1048 // branch, if not we can't thread it. 1049 Value *Condition; 1050 Instruction *Terminator = BB->getTerminator(); 1051 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { 1052 // Can't thread an unconditional jump. 1053 if (BI->isUnconditional()) return false; 1054 Condition = BI->getCondition(); 1055 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { 1056 Condition = SI->getCondition(); 1057 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { 1058 // Can't thread indirect branch with no successors. 1059 if (IB->getNumSuccessors() == 0) return false; 1060 Condition = IB->getAddress()->stripPointerCasts(); 1061 Preference = WantBlockAddress; 1062 } else { 1063 return false; // Must be an invoke or callbr. 1064 } 1065 1066 // Keep track if we constant folded the condition in this invocation. 1067 bool ConstantFolded = false; 1068 1069 // Run constant folding to see if we can reduce the condition to a simple 1070 // constant. 1071 if (Instruction *I = dyn_cast<Instruction>(Condition)) { 1072 Value *SimpleVal = 1073 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI); 1074 if (SimpleVal) { 1075 I->replaceAllUsesWith(SimpleVal); 1076 if (isInstructionTriviallyDead(I, TLI)) 1077 I->eraseFromParent(); 1078 Condition = SimpleVal; 1079 ConstantFolded = true; 1080 } 1081 } 1082 1083 // If the terminator is branching on an undef or freeze undef, we can pick any 1084 // of the successors to branch to. Let getBestDestForJumpOnUndef decide. 1085 auto *FI = dyn_cast<FreezeInst>(Condition); 1086 if (isa<UndefValue>(Condition) || 1087 (FI && isa<UndefValue>(FI->getOperand(0)) && FI->hasOneUse())) { 1088 unsigned BestSucc = getBestDestForJumpOnUndef(BB); 1089 std::vector<DominatorTree::UpdateType> Updates; 1090 1091 // Fold the branch/switch. 1092 Instruction *BBTerm = BB->getTerminator(); 1093 Updates.reserve(BBTerm->getNumSuccessors()); 1094 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 1095 if (i == BestSucc) continue; 1096 BasicBlock *Succ = BBTerm->getSuccessor(i); 1097 Succ->removePredecessor(BB, true); 1098 Updates.push_back({DominatorTree::Delete, BB, Succ}); 1099 } 1100 1101 LLVM_DEBUG(dbgs() << " In block '" << BB->getName() 1102 << "' folding undef terminator: " << *BBTerm << '\n'); 1103 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); 1104 ++NumFolds; 1105 BBTerm->eraseFromParent(); 1106 DTU->applyUpdatesPermissive(Updates); 1107 if (FI) 1108 FI->eraseFromParent(); 1109 return true; 1110 } 1111 1112 // If the terminator of this block is branching on a constant, simplify the 1113 // terminator to an unconditional branch. This can occur due to threading in 1114 // other blocks. 1115 if (getKnownConstant(Condition, Preference)) { 1116 LLVM_DEBUG(dbgs() << " In block '" << BB->getName() 1117 << "' folding terminator: " << *BB->getTerminator() 1118 << '\n'); 1119 ++NumFolds; 1120 ConstantFoldTerminator(BB, true, nullptr, DTU); 1121 if (HasProfileData) 1122 BPI->eraseBlock(BB); 1123 return true; 1124 } 1125 1126 Instruction *CondInst = dyn_cast<Instruction>(Condition); 1127 1128 // All the rest of our checks depend on the condition being an instruction. 1129 if (!CondInst) { 1130 // FIXME: Unify this with code below. 1131 if (processThreadableEdges(Condition, BB, Preference, Terminator)) 1132 return true; 1133 return ConstantFolded; 1134 } 1135 1136 // Some of the following optimization can safely work on the unfrozen cond. 1137 Value *CondWithoutFreeze = CondInst; 1138 if (auto *FI = dyn_cast<FreezeInst>(CondInst)) 1139 CondWithoutFreeze = FI->getOperand(0); 1140 1141 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondWithoutFreeze)) { 1142 // If we're branching on a conditional, LVI might be able to determine 1143 // it's value at the branch instruction. We only handle comparisons 1144 // against a constant at this time. 1145 if (Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1))) { 1146 LazyValueInfo::Tristate Ret = 1147 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0), 1148 CondConst, BB->getTerminator(), 1149 /*UseBlockValue=*/false); 1150 if (Ret != LazyValueInfo::Unknown) { 1151 // We can safely replace *some* uses of the CondInst if it has 1152 // exactly one value as returned by LVI. RAUW is incorrect in the 1153 // presence of guards and assumes, that have the `Cond` as the use. This 1154 // is because we use the guards/assume to reason about the `Cond` value 1155 // at the end of block, but RAUW unconditionally replaces all uses 1156 // including the guards/assumes themselves and the uses before the 1157 // guard/assume. 1158 auto *CI = Ret == LazyValueInfo::True ? 1159 ConstantInt::getTrue(CondCmp->getType()) : 1160 ConstantInt::getFalse(CondCmp->getType()); 1161 if (replaceFoldableUses(CondCmp, CI, BB)) 1162 return true; 1163 } 1164 1165 // We did not manage to simplify this branch, try to see whether 1166 // CondCmp depends on a known phi-select pattern. 1167 if (tryToUnfoldSelect(CondCmp, BB)) 1168 return true; 1169 } 1170 } 1171 1172 if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 1173 if (tryToUnfoldSelect(SI, BB)) 1174 return true; 1175 1176 // Check for some cases that are worth simplifying. Right now we want to look 1177 // for loads that are used by a switch or by the condition for the branch. If 1178 // we see one, check to see if it's partially redundant. If so, insert a PHI 1179 // which can then be used to thread the values. 1180 Value *SimplifyValue = CondWithoutFreeze; 1181 1182 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 1183 if (isa<Constant>(CondCmp->getOperand(1))) 1184 SimplifyValue = CondCmp->getOperand(0); 1185 1186 // TODO: There are other places where load PRE would be profitable, such as 1187 // more complex comparisons. 1188 if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue)) 1189 if (simplifyPartiallyRedundantLoad(LoadI)) 1190 return true; 1191 1192 // Before threading, try to propagate profile data backwards: 1193 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 1194 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 1195 updatePredecessorProfileMetadata(PN, BB); 1196 1197 // Handle a variety of cases where we are branching on something derived from 1198 // a PHI node in the current block. If we can prove that any predecessors 1199 // compute a predictable value based on a PHI node, thread those predecessors. 1200 if (processThreadableEdges(CondInst, BB, Preference, Terminator)) 1201 return true; 1202 1203 // If this is an otherwise-unfoldable branch on a phi node or freeze(phi) in 1204 // the current block, see if we can simplify. 1205 PHINode *PN = dyn_cast<PHINode>(CondWithoutFreeze); 1206 if (PN && PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 1207 return processBranchOnPHI(PN); 1208 1209 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. 1210 if (CondInst->getOpcode() == Instruction::Xor && 1211 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) 1212 return processBranchOnXOR(cast<BinaryOperator>(CondInst)); 1213 1214 // Search for a stronger dominating condition that can be used to simplify a 1215 // conditional branch leaving BB. 1216 if (processImpliedCondition(BB)) 1217 return true; 1218 1219 return false; 1220 } 1221 1222 bool JumpThreadingPass::processImpliedCondition(BasicBlock *BB) { 1223 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 1224 if (!BI || !BI->isConditional()) 1225 return false; 1226 1227 Value *Cond = BI->getCondition(); 1228 // Assuming that predecessor's branch was taken, if pred's branch condition 1229 // (V) implies Cond, Cond can be either true, undef, or poison. In this case, 1230 // freeze(Cond) is either true or a nondeterministic value. 1231 // If freeze(Cond) has only one use, we can freely fold freeze(Cond) to true 1232 // without affecting other instructions. 1233 auto *FICond = dyn_cast<FreezeInst>(Cond); 1234 if (FICond && FICond->hasOneUse()) 1235 Cond = FICond->getOperand(0); 1236 else 1237 FICond = nullptr; 1238 1239 BasicBlock *CurrentBB = BB; 1240 BasicBlock *CurrentPred = BB->getSinglePredecessor(); 1241 unsigned Iter = 0; 1242 1243 auto &DL = BB->getModule()->getDataLayout(); 1244 1245 while (CurrentPred && Iter++ < ImplicationSearchThreshold) { 1246 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator()); 1247 if (!PBI || !PBI->isConditional()) 1248 return false; 1249 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB) 1250 return false; 1251 1252 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB; 1253 Optional<bool> Implication = 1254 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue); 1255 1256 // If the branch condition of BB (which is Cond) and CurrentPred are 1257 // exactly the same freeze instruction, Cond can be folded into CondIsTrue. 1258 if (!Implication && FICond && isa<FreezeInst>(PBI->getCondition())) { 1259 if (cast<FreezeInst>(PBI->getCondition())->getOperand(0) == 1260 FICond->getOperand(0)) 1261 Implication = CondIsTrue; 1262 } 1263 1264 if (Implication) { 1265 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1); 1266 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0); 1267 RemoveSucc->removePredecessor(BB); 1268 BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI); 1269 UncondBI->setDebugLoc(BI->getDebugLoc()); 1270 ++NumFolds; 1271 BI->eraseFromParent(); 1272 if (FICond) 1273 FICond->eraseFromParent(); 1274 1275 DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}}); 1276 if (HasProfileData) 1277 BPI->eraseBlock(BB); 1278 return true; 1279 } 1280 CurrentBB = CurrentPred; 1281 CurrentPred = CurrentBB->getSinglePredecessor(); 1282 } 1283 1284 return false; 1285 } 1286 1287 /// Return true if Op is an instruction defined in the given block. 1288 static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) { 1289 if (Instruction *OpInst = dyn_cast<Instruction>(Op)) 1290 if (OpInst->getParent() == BB) 1291 return true; 1292 return false; 1293 } 1294 1295 /// simplifyPartiallyRedundantLoad - If LoadI is an obviously partially 1296 /// redundant load instruction, eliminate it by replacing it with a PHI node. 1297 /// This is an important optimization that encourages jump threading, and needs 1298 /// to be run interlaced with other jump threading tasks. 1299 bool JumpThreadingPass::simplifyPartiallyRedundantLoad(LoadInst *LoadI) { 1300 // Don't hack volatile and ordered loads. 1301 if (!LoadI->isUnordered()) return false; 1302 1303 // If the load is defined in a block with exactly one predecessor, it can't be 1304 // partially redundant. 1305 BasicBlock *LoadBB = LoadI->getParent(); 1306 if (LoadBB->getSinglePredecessor()) 1307 return false; 1308 1309 // If the load is defined in an EH pad, it can't be partially redundant, 1310 // because the edges between the invoke and the EH pad cannot have other 1311 // instructions between them. 1312 if (LoadBB->isEHPad()) 1313 return false; 1314 1315 Value *LoadedPtr = LoadI->getOperand(0); 1316 1317 // If the loaded operand is defined in the LoadBB and its not a phi, 1318 // it can't be available in predecessors. 1319 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr)) 1320 return false; 1321 1322 // Scan a few instructions up from the load, to see if it is obviously live at 1323 // the entry to its block. 1324 BasicBlock::iterator BBIt(LoadI); 1325 bool IsLoadCSE; 1326 if (Value *AvailableVal = FindAvailableLoadedValue( 1327 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) { 1328 // If the value of the load is locally available within the block, just use 1329 // it. This frequently occurs for reg2mem'd allocas. 1330 1331 if (IsLoadCSE) { 1332 LoadInst *NLoadI = cast<LoadInst>(AvailableVal); 1333 combineMetadataForCSE(NLoadI, LoadI, false); 1334 }; 1335 1336 // If the returned value is the load itself, replace with poison. This can 1337 // only happen in dead loops. 1338 if (AvailableVal == LoadI) 1339 AvailableVal = PoisonValue::get(LoadI->getType()); 1340 if (AvailableVal->getType() != LoadI->getType()) 1341 AvailableVal = CastInst::CreateBitOrPointerCast( 1342 AvailableVal, LoadI->getType(), "", LoadI); 1343 LoadI->replaceAllUsesWith(AvailableVal); 1344 LoadI->eraseFromParent(); 1345 return true; 1346 } 1347 1348 // Otherwise, if we scanned the whole block and got to the top of the block, 1349 // we know the block is locally transparent to the load. If not, something 1350 // might clobber its value. 1351 if (BBIt != LoadBB->begin()) 1352 return false; 1353 1354 // If all of the loads and stores that feed the value have the same AA tags, 1355 // then we can propagate them onto any newly inserted loads. 1356 AAMDNodes AATags = LoadI->getAAMetadata(); 1357 1358 SmallPtrSet<BasicBlock*, 8> PredsScanned; 1359 1360 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>; 1361 1362 AvailablePredsTy AvailablePreds; 1363 BasicBlock *OneUnavailablePred = nullptr; 1364 SmallVector<LoadInst*, 8> CSELoads; 1365 1366 // If we got here, the loaded value is transparent through to the start of the 1367 // block. Check to see if it is available in any of the predecessor blocks. 1368 for (BasicBlock *PredBB : predecessors(LoadBB)) { 1369 // If we already scanned this predecessor, skip it. 1370 if (!PredsScanned.insert(PredBB).second) 1371 continue; 1372 1373 BBIt = PredBB->end(); 1374 unsigned NumScanedInst = 0; 1375 Value *PredAvailable = nullptr; 1376 // NOTE: We don't CSE load that is volatile or anything stronger than 1377 // unordered, that should have been checked when we entered the function. 1378 assert(LoadI->isUnordered() && 1379 "Attempting to CSE volatile or atomic loads"); 1380 // If this is a load on a phi pointer, phi-translate it and search 1381 // for available load/store to the pointer in predecessors. 1382 Type *AccessTy = LoadI->getType(); 1383 const auto &DL = LoadI->getModule()->getDataLayout(); 1384 MemoryLocation Loc(LoadedPtr->DoPHITranslation(LoadBB, PredBB), 1385 LocationSize::precise(DL.getTypeStoreSize(AccessTy)), 1386 AATags); 1387 PredAvailable = findAvailablePtrLoadStore(Loc, AccessTy, LoadI->isAtomic(), 1388 PredBB, BBIt, DefMaxInstsToScan, 1389 AA, &IsLoadCSE, &NumScanedInst); 1390 1391 // If PredBB has a single predecessor, continue scanning through the 1392 // single predecessor. 1393 BasicBlock *SinglePredBB = PredBB; 1394 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() && 1395 NumScanedInst < DefMaxInstsToScan) { 1396 SinglePredBB = SinglePredBB->getSinglePredecessor(); 1397 if (SinglePredBB) { 1398 BBIt = SinglePredBB->end(); 1399 PredAvailable = findAvailablePtrLoadStore( 1400 Loc, AccessTy, LoadI->isAtomic(), SinglePredBB, BBIt, 1401 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE, 1402 &NumScanedInst); 1403 } 1404 } 1405 1406 if (!PredAvailable) { 1407 OneUnavailablePred = PredBB; 1408 continue; 1409 } 1410 1411 if (IsLoadCSE) 1412 CSELoads.push_back(cast<LoadInst>(PredAvailable)); 1413 1414 // If so, this load is partially redundant. Remember this info so that we 1415 // can create a PHI node. 1416 AvailablePreds.emplace_back(PredBB, PredAvailable); 1417 } 1418 1419 // If the loaded value isn't available in any predecessor, it isn't partially 1420 // redundant. 1421 if (AvailablePreds.empty()) return false; 1422 1423 // Okay, the loaded value is available in at least one (and maybe all!) 1424 // predecessors. If the value is unavailable in more than one unique 1425 // predecessor, we want to insert a merge block for those common predecessors. 1426 // This ensures that we only have to insert one reload, thus not increasing 1427 // code size. 1428 BasicBlock *UnavailablePred = nullptr; 1429 1430 // If the value is unavailable in one of predecessors, we will end up 1431 // inserting a new instruction into them. It is only valid if all the 1432 // instructions before LoadI are guaranteed to pass execution to its 1433 // successor, or if LoadI is safe to speculate. 1434 // TODO: If this logic becomes more complex, and we will perform PRE insertion 1435 // farther than to a predecessor, we need to reuse the code from GVN's PRE. 1436 // It requires domination tree analysis, so for this simple case it is an 1437 // overkill. 1438 if (PredsScanned.size() != AvailablePreds.size() && 1439 !isSafeToSpeculativelyExecute(LoadI)) 1440 for (auto I = LoadBB->begin(); &*I != LoadI; ++I) 1441 if (!isGuaranteedToTransferExecutionToSuccessor(&*I)) 1442 return false; 1443 1444 // If there is exactly one predecessor where the value is unavailable, the 1445 // already computed 'OneUnavailablePred' block is it. If it ends in an 1446 // unconditional branch, we know that it isn't a critical edge. 1447 if (PredsScanned.size() == AvailablePreds.size()+1 && 1448 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 1449 UnavailablePred = OneUnavailablePred; 1450 } else if (PredsScanned.size() != AvailablePreds.size()) { 1451 // Otherwise, we had multiple unavailable predecessors or we had a critical 1452 // edge from the one. 1453 SmallVector<BasicBlock*, 8> PredsToSplit; 1454 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 1455 1456 for (const auto &AvailablePred : AvailablePreds) 1457 AvailablePredSet.insert(AvailablePred.first); 1458 1459 // Add all the unavailable predecessors to the PredsToSplit list. 1460 for (BasicBlock *P : predecessors(LoadBB)) { 1461 // If the predecessor is an indirect goto, we can't split the edge. 1462 if (isa<IndirectBrInst>(P->getTerminator())) 1463 return false; 1464 1465 if (!AvailablePredSet.count(P)) 1466 PredsToSplit.push_back(P); 1467 } 1468 1469 // Split them out to their own block. 1470 UnavailablePred = splitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split"); 1471 } 1472 1473 // If the value isn't available in all predecessors, then there will be 1474 // exactly one where it isn't available. Insert a load on that edge and add 1475 // it to the AvailablePreds list. 1476 if (UnavailablePred) { 1477 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 1478 "Can't handle critical edge here!"); 1479 LoadInst *NewVal = new LoadInst( 1480 LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred), 1481 LoadI->getName() + ".pr", false, LoadI->getAlign(), 1482 LoadI->getOrdering(), LoadI->getSyncScopeID(), 1483 UnavailablePred->getTerminator()); 1484 NewVal->setDebugLoc(LoadI->getDebugLoc()); 1485 if (AATags) 1486 NewVal->setAAMetadata(AATags); 1487 1488 AvailablePreds.emplace_back(UnavailablePred, NewVal); 1489 } 1490 1491 // Now we know that each predecessor of this block has a value in 1492 // AvailablePreds, sort them for efficient access as we're walking the preds. 1493 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 1494 1495 // Create a PHI node at the start of the block for the PRE'd load value. 1496 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB); 1497 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "", 1498 &LoadBB->front()); 1499 PN->takeName(LoadI); 1500 PN->setDebugLoc(LoadI->getDebugLoc()); 1501 1502 // Insert new entries into the PHI for each predecessor. A single block may 1503 // have multiple entries here. 1504 for (pred_iterator PI = PB; PI != PE; ++PI) { 1505 BasicBlock *P = *PI; 1506 AvailablePredsTy::iterator I = 1507 llvm::lower_bound(AvailablePreds, std::make_pair(P, (Value *)nullptr)); 1508 1509 assert(I != AvailablePreds.end() && I->first == P && 1510 "Didn't find entry for predecessor!"); 1511 1512 // If we have an available predecessor but it requires casting, insert the 1513 // cast in the predecessor and use the cast. Note that we have to update the 1514 // AvailablePreds vector as we go so that all of the PHI entries for this 1515 // predecessor use the same bitcast. 1516 Value *&PredV = I->second; 1517 if (PredV->getType() != LoadI->getType()) 1518 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "", 1519 P->getTerminator()); 1520 1521 PN->addIncoming(PredV, I->first); 1522 } 1523 1524 for (LoadInst *PredLoadI : CSELoads) { 1525 combineMetadataForCSE(PredLoadI, LoadI, true); 1526 } 1527 1528 LoadI->replaceAllUsesWith(PN); 1529 LoadI->eraseFromParent(); 1530 1531 return true; 1532 } 1533 1534 /// findMostPopularDest - The specified list contains multiple possible 1535 /// threadable destinations. Pick the one that occurs the most frequently in 1536 /// the list. 1537 static BasicBlock * 1538 findMostPopularDest(BasicBlock *BB, 1539 const SmallVectorImpl<std::pair<BasicBlock *, 1540 BasicBlock *>> &PredToDestList) { 1541 assert(!PredToDestList.empty()); 1542 1543 // Determine popularity. If there are multiple possible destinations, we 1544 // explicitly choose to ignore 'undef' destinations. We prefer to thread 1545 // blocks with known and real destinations to threading undef. We'll handle 1546 // them later if interesting. 1547 MapVector<BasicBlock *, unsigned> DestPopularity; 1548 1549 // Populate DestPopularity with the successors in the order they appear in the 1550 // successor list. This way, we ensure determinism by iterating it in the 1551 // same order in std::max_element below. We map nullptr to 0 so that we can 1552 // return nullptr when PredToDestList contains nullptr only. 1553 DestPopularity[nullptr] = 0; 1554 for (auto *SuccBB : successors(BB)) 1555 DestPopularity[SuccBB] = 0; 1556 1557 for (const auto &PredToDest : PredToDestList) 1558 if (PredToDest.second) 1559 DestPopularity[PredToDest.second]++; 1560 1561 // Find the most popular dest. 1562 auto MostPopular = std::max_element( 1563 DestPopularity.begin(), DestPopularity.end(), llvm::less_second()); 1564 1565 // Okay, we have finally picked the most popular destination. 1566 return MostPopular->first; 1567 } 1568 1569 // Try to evaluate the value of V when the control flows from PredPredBB to 1570 // BB->getSinglePredecessor() and then on to BB. 1571 Constant *JumpThreadingPass::evaluateOnPredecessorEdge(BasicBlock *BB, 1572 BasicBlock *PredPredBB, 1573 Value *V) { 1574 BasicBlock *PredBB = BB->getSinglePredecessor(); 1575 assert(PredBB && "Expected a single predecessor"); 1576 1577 if (Constant *Cst = dyn_cast<Constant>(V)) { 1578 return Cst; 1579 } 1580 1581 // Consult LVI if V is not an instruction in BB or PredBB. 1582 Instruction *I = dyn_cast<Instruction>(V); 1583 if (!I || (I->getParent() != BB && I->getParent() != PredBB)) { 1584 return LVI->getConstantOnEdge(V, PredPredBB, PredBB, nullptr); 1585 } 1586 1587 // Look into a PHI argument. 1588 if (PHINode *PHI = dyn_cast<PHINode>(V)) { 1589 if (PHI->getParent() == PredBB) 1590 return dyn_cast<Constant>(PHI->getIncomingValueForBlock(PredPredBB)); 1591 return nullptr; 1592 } 1593 1594 // If we have a CmpInst, try to fold it for each incoming edge into PredBB. 1595 if (CmpInst *CondCmp = dyn_cast<CmpInst>(V)) { 1596 if (CondCmp->getParent() == BB) { 1597 Constant *Op0 = 1598 evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(0)); 1599 Constant *Op1 = 1600 evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(1)); 1601 if (Op0 && Op1) { 1602 return ConstantExpr::getCompare(CondCmp->getPredicate(), Op0, Op1); 1603 } 1604 } 1605 return nullptr; 1606 } 1607 1608 return nullptr; 1609 } 1610 1611 bool JumpThreadingPass::processThreadableEdges(Value *Cond, BasicBlock *BB, 1612 ConstantPreference Preference, 1613 Instruction *CxtI) { 1614 // If threading this would thread across a loop header, don't even try to 1615 // thread the edge. 1616 if (LoopHeaders.count(BB)) 1617 return false; 1618 1619 PredValueInfoTy PredValues; 1620 if (!computeValueKnownInPredecessors(Cond, BB, PredValues, Preference, 1621 CxtI)) { 1622 // We don't have known values in predecessors. See if we can thread through 1623 // BB and its sole predecessor. 1624 return maybethreadThroughTwoBasicBlocks(BB, Cond); 1625 } 1626 1627 assert(!PredValues.empty() && 1628 "computeValueKnownInPredecessors returned true with no values"); 1629 1630 LLVM_DEBUG(dbgs() << "IN BB: " << *BB; 1631 for (const auto &PredValue : PredValues) { 1632 dbgs() << " BB '" << BB->getName() 1633 << "': FOUND condition = " << *PredValue.first 1634 << " for pred '" << PredValue.second->getName() << "'.\n"; 1635 }); 1636 1637 // Decide what we want to thread through. Convert our list of known values to 1638 // a list of known destinations for each pred. This also discards duplicate 1639 // predecessors and keeps track of the undefined inputs (which are represented 1640 // as a null dest in the PredToDestList). 1641 SmallPtrSet<BasicBlock*, 16> SeenPreds; 1642 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; 1643 1644 BasicBlock *OnlyDest = nullptr; 1645 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; 1646 Constant *OnlyVal = nullptr; 1647 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL; 1648 1649 for (const auto &PredValue : PredValues) { 1650 BasicBlock *Pred = PredValue.second; 1651 if (!SeenPreds.insert(Pred).second) 1652 continue; // Duplicate predecessor entry. 1653 1654 Constant *Val = PredValue.first; 1655 1656 BasicBlock *DestBB; 1657 if (isa<UndefValue>(Val)) 1658 DestBB = nullptr; 1659 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 1660 assert(isa<ConstantInt>(Val) && "Expecting a constant integer"); 1661 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); 1662 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 1663 assert(isa<ConstantInt>(Val) && "Expecting a constant integer"); 1664 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor(); 1665 } else { 1666 assert(isa<IndirectBrInst>(BB->getTerminator()) 1667 && "Unexpected terminator"); 1668 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress"); 1669 DestBB = cast<BlockAddress>(Val)->getBasicBlock(); 1670 } 1671 1672 // If we have exactly one destination, remember it for efficiency below. 1673 if (PredToDestList.empty()) { 1674 OnlyDest = DestBB; 1675 OnlyVal = Val; 1676 } else { 1677 if (OnlyDest != DestBB) 1678 OnlyDest = MultipleDestSentinel; 1679 // It possible we have same destination, but different value, e.g. default 1680 // case in switchinst. 1681 if (Val != OnlyVal) 1682 OnlyVal = MultipleVal; 1683 } 1684 1685 // If the predecessor ends with an indirect goto, we can't change its 1686 // destination. 1687 if (isa<IndirectBrInst>(Pred->getTerminator())) 1688 continue; 1689 1690 PredToDestList.emplace_back(Pred, DestBB); 1691 } 1692 1693 // If all edges were unthreadable, we fail. 1694 if (PredToDestList.empty()) 1695 return false; 1696 1697 // If all the predecessors go to a single known successor, we want to fold, 1698 // not thread. By doing so, we do not need to duplicate the current block and 1699 // also miss potential opportunities in case we dont/cant duplicate. 1700 if (OnlyDest && OnlyDest != MultipleDestSentinel) { 1701 if (BB->hasNPredecessors(PredToDestList.size())) { 1702 bool SeenFirstBranchToOnlyDest = false; 1703 std::vector <DominatorTree::UpdateType> Updates; 1704 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1); 1705 for (BasicBlock *SuccBB : successors(BB)) { 1706 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) { 1707 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch. 1708 } else { 1709 SuccBB->removePredecessor(BB, true); // This is unreachable successor. 1710 Updates.push_back({DominatorTree::Delete, BB, SuccBB}); 1711 } 1712 } 1713 1714 // Finally update the terminator. 1715 Instruction *Term = BB->getTerminator(); 1716 BranchInst::Create(OnlyDest, Term); 1717 ++NumFolds; 1718 Term->eraseFromParent(); 1719 DTU->applyUpdatesPermissive(Updates); 1720 if (HasProfileData) 1721 BPI->eraseBlock(BB); 1722 1723 // If the condition is now dead due to the removal of the old terminator, 1724 // erase it. 1725 if (auto *CondInst = dyn_cast<Instruction>(Cond)) { 1726 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects()) 1727 CondInst->eraseFromParent(); 1728 // We can safely replace *some* uses of the CondInst if it has 1729 // exactly one value as returned by LVI. RAUW is incorrect in the 1730 // presence of guards and assumes, that have the `Cond` as the use. This 1731 // is because we use the guards/assume to reason about the `Cond` value 1732 // at the end of block, but RAUW unconditionally replaces all uses 1733 // including the guards/assumes themselves and the uses before the 1734 // guard/assume. 1735 else if (OnlyVal && OnlyVal != MultipleVal) 1736 replaceFoldableUses(CondInst, OnlyVal, BB); 1737 } 1738 return true; 1739 } 1740 } 1741 1742 // Determine which is the most common successor. If we have many inputs and 1743 // this block is a switch, we want to start by threading the batch that goes 1744 // to the most popular destination first. If we only know about one 1745 // threadable destination (the common case) we can avoid this. 1746 BasicBlock *MostPopularDest = OnlyDest; 1747 1748 if (MostPopularDest == MultipleDestSentinel) { 1749 // Remove any loop headers from the Dest list, threadEdge conservatively 1750 // won't process them, but we might have other destination that are eligible 1751 // and we still want to process. 1752 erase_if(PredToDestList, 1753 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) { 1754 return LoopHeaders.contains(PredToDest.second); 1755 }); 1756 1757 if (PredToDestList.empty()) 1758 return false; 1759 1760 MostPopularDest = findMostPopularDest(BB, PredToDestList); 1761 } 1762 1763 // Now that we know what the most popular destination is, factor all 1764 // predecessors that will jump to it into a single predecessor. 1765 SmallVector<BasicBlock*, 16> PredsToFactor; 1766 for (const auto &PredToDest : PredToDestList) 1767 if (PredToDest.second == MostPopularDest) { 1768 BasicBlock *Pred = PredToDest.first; 1769 1770 // This predecessor may be a switch or something else that has multiple 1771 // edges to the block. Factor each of these edges by listing them 1772 // according to # occurrences in PredsToFactor. 1773 for (BasicBlock *Succ : successors(Pred)) 1774 if (Succ == BB) 1775 PredsToFactor.push_back(Pred); 1776 } 1777 1778 // If the threadable edges are branching on an undefined value, we get to pick 1779 // the destination that these predecessors should get to. 1780 if (!MostPopularDest) 1781 MostPopularDest = BB->getTerminator()-> 1782 getSuccessor(getBestDestForJumpOnUndef(BB)); 1783 1784 // Ok, try to thread it! 1785 return tryThreadEdge(BB, PredsToFactor, MostPopularDest); 1786 } 1787 1788 /// processBranchOnPHI - We have an otherwise unthreadable conditional branch on 1789 /// a PHI node (or freeze PHI) in the current block. See if there are any 1790 /// simplifications we can do based on inputs to the phi node. 1791 bool JumpThreadingPass::processBranchOnPHI(PHINode *PN) { 1792 BasicBlock *BB = PN->getParent(); 1793 1794 // TODO: We could make use of this to do it once for blocks with common PHI 1795 // values. 1796 SmallVector<BasicBlock*, 1> PredBBs; 1797 PredBBs.resize(1); 1798 1799 // If any of the predecessor blocks end in an unconditional branch, we can 1800 // *duplicate* the conditional branch into that block in order to further 1801 // encourage jump threading and to eliminate cases where we have branch on a 1802 // phi of an icmp (branch on icmp is much better). 1803 // This is still beneficial when a frozen phi is used as the branch condition 1804 // because it allows CodeGenPrepare to further canonicalize br(freeze(icmp)) 1805 // to br(icmp(freeze ...)). 1806 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1807 BasicBlock *PredBB = PN->getIncomingBlock(i); 1808 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) 1809 if (PredBr->isUnconditional()) { 1810 PredBBs[0] = PredBB; 1811 // Try to duplicate BB into PredBB. 1812 if (duplicateCondBranchOnPHIIntoPred(BB, PredBBs)) 1813 return true; 1814 } 1815 } 1816 1817 return false; 1818 } 1819 1820 /// processBranchOnXOR - We have an otherwise unthreadable conditional branch on 1821 /// a xor instruction in the current block. See if there are any 1822 /// simplifications we can do based on inputs to the xor. 1823 bool JumpThreadingPass::processBranchOnXOR(BinaryOperator *BO) { 1824 BasicBlock *BB = BO->getParent(); 1825 1826 // If either the LHS or RHS of the xor is a constant, don't do this 1827 // optimization. 1828 if (isa<ConstantInt>(BO->getOperand(0)) || 1829 isa<ConstantInt>(BO->getOperand(1))) 1830 return false; 1831 1832 // If the first instruction in BB isn't a phi, we won't be able to infer 1833 // anything special about any particular predecessor. 1834 if (!isa<PHINode>(BB->front())) 1835 return false; 1836 1837 // If this BB is a landing pad, we won't be able to split the edge into it. 1838 if (BB->isEHPad()) 1839 return false; 1840 1841 // If we have a xor as the branch input to this block, and we know that the 1842 // LHS or RHS of the xor in any predecessor is true/false, then we can clone 1843 // the condition into the predecessor and fix that value to true, saving some 1844 // logical ops on that path and encouraging other paths to simplify. 1845 // 1846 // This copies something like this: 1847 // 1848 // BB: 1849 // %X = phi i1 [1], [%X'] 1850 // %Y = icmp eq i32 %A, %B 1851 // %Z = xor i1 %X, %Y 1852 // br i1 %Z, ... 1853 // 1854 // Into: 1855 // BB': 1856 // %Y = icmp ne i32 %A, %B 1857 // br i1 %Y, ... 1858 1859 PredValueInfoTy XorOpValues; 1860 bool isLHS = true; 1861 if (!computeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, 1862 WantInteger, BO)) { 1863 assert(XorOpValues.empty()); 1864 if (!computeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, 1865 WantInteger, BO)) 1866 return false; 1867 isLHS = false; 1868 } 1869 1870 assert(!XorOpValues.empty() && 1871 "computeValueKnownInPredecessors returned true with no values"); 1872 1873 // Scan the information to see which is most popular: true or false. The 1874 // predecessors can be of the set true, false, or undef. 1875 unsigned NumTrue = 0, NumFalse = 0; 1876 for (const auto &XorOpValue : XorOpValues) { 1877 if (isa<UndefValue>(XorOpValue.first)) 1878 // Ignore undefs for the count. 1879 continue; 1880 if (cast<ConstantInt>(XorOpValue.first)->isZero()) 1881 ++NumFalse; 1882 else 1883 ++NumTrue; 1884 } 1885 1886 // Determine which value to split on, true, false, or undef if neither. 1887 ConstantInt *SplitVal = nullptr; 1888 if (NumTrue > NumFalse) 1889 SplitVal = ConstantInt::getTrue(BB->getContext()); 1890 else if (NumTrue != 0 || NumFalse != 0) 1891 SplitVal = ConstantInt::getFalse(BB->getContext()); 1892 1893 // Collect all of the blocks that this can be folded into so that we can 1894 // factor this once and clone it once. 1895 SmallVector<BasicBlock*, 8> BlocksToFoldInto; 1896 for (const auto &XorOpValue : XorOpValues) { 1897 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first)) 1898 continue; 1899 1900 BlocksToFoldInto.push_back(XorOpValue.second); 1901 } 1902 1903 // If we inferred a value for all of the predecessors, then duplication won't 1904 // help us. However, we can just replace the LHS or RHS with the constant. 1905 if (BlocksToFoldInto.size() == 1906 cast<PHINode>(BB->front()).getNumIncomingValues()) { 1907 if (!SplitVal) { 1908 // If all preds provide undef, just nuke the xor, because it is undef too. 1909 BO->replaceAllUsesWith(UndefValue::get(BO->getType())); 1910 BO->eraseFromParent(); 1911 } else if (SplitVal->isZero()) { 1912 // If all preds provide 0, replace the xor with the other input. 1913 BO->replaceAllUsesWith(BO->getOperand(isLHS)); 1914 BO->eraseFromParent(); 1915 } else { 1916 // If all preds provide 1, set the computed value to 1. 1917 BO->setOperand(!isLHS, SplitVal); 1918 } 1919 1920 return true; 1921 } 1922 1923 // If any of predecessors end with an indirect goto, we can't change its 1924 // destination. 1925 if (any_of(BlocksToFoldInto, [](BasicBlock *Pred) { 1926 return isa<IndirectBrInst>(Pred->getTerminator()); 1927 })) 1928 return false; 1929 1930 // Try to duplicate BB into PredBB. 1931 return duplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); 1932 } 1933 1934 /// addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new 1935 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for 1936 /// NewPred using the entries from OldPred (suitably mapped). 1937 static void addPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, 1938 BasicBlock *OldPred, 1939 BasicBlock *NewPred, 1940 DenseMap<Instruction*, Value*> &ValueMap) { 1941 for (PHINode &PN : PHIBB->phis()) { 1942 // Ok, we have a PHI node. Figure out what the incoming value was for the 1943 // DestBlock. 1944 Value *IV = PN.getIncomingValueForBlock(OldPred); 1945 1946 // Remap the value if necessary. 1947 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 1948 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); 1949 if (I != ValueMap.end()) 1950 IV = I->second; 1951 } 1952 1953 PN.addIncoming(IV, NewPred); 1954 } 1955 } 1956 1957 /// Merge basic block BB into its sole predecessor if possible. 1958 bool JumpThreadingPass::maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB) { 1959 BasicBlock *SinglePred = BB->getSinglePredecessor(); 1960 if (!SinglePred) 1961 return false; 1962 1963 const Instruction *TI = SinglePred->getTerminator(); 1964 if (TI->isExceptionalTerminator() || TI->getNumSuccessors() != 1 || 1965 SinglePred == BB || hasAddressTakenAndUsed(BB)) 1966 return false; 1967 1968 // If SinglePred was a loop header, BB becomes one. 1969 if (LoopHeaders.erase(SinglePred)) 1970 LoopHeaders.insert(BB); 1971 1972 LVI->eraseBlock(SinglePred); 1973 MergeBasicBlockIntoOnlyPred(BB, DTU); 1974 1975 // Now that BB is merged into SinglePred (i.e. SinglePred code followed by 1976 // BB code within one basic block `BB`), we need to invalidate the LVI 1977 // information associated with BB, because the LVI information need not be 1978 // true for all of BB after the merge. For example, 1979 // Before the merge, LVI info and code is as follows: 1980 // SinglePred: <LVI info1 for %p val> 1981 // %y = use of %p 1982 // call @exit() // need not transfer execution to successor. 1983 // assume(%p) // from this point on %p is true 1984 // br label %BB 1985 // BB: <LVI info2 for %p val, i.e. %p is true> 1986 // %x = use of %p 1987 // br label exit 1988 // 1989 // Note that this LVI info for blocks BB and SinglPred is correct for %p 1990 // (info2 and info1 respectively). After the merge and the deletion of the 1991 // LVI info1 for SinglePred. We have the following code: 1992 // BB: <LVI info2 for %p val> 1993 // %y = use of %p 1994 // call @exit() 1995 // assume(%p) 1996 // %x = use of %p <-- LVI info2 is correct from here onwards. 1997 // br label exit 1998 // LVI info2 for BB is incorrect at the beginning of BB. 1999 2000 // Invalidate LVI information for BB if the LVI is not provably true for 2001 // all of BB. 2002 if (!isGuaranteedToTransferExecutionToSuccessor(BB)) 2003 LVI->eraseBlock(BB); 2004 return true; 2005 } 2006 2007 /// Update the SSA form. NewBB contains instructions that are copied from BB. 2008 /// ValueMapping maps old values in BB to new ones in NewBB. 2009 void JumpThreadingPass::updateSSA( 2010 BasicBlock *BB, BasicBlock *NewBB, 2011 DenseMap<Instruction *, Value *> &ValueMapping) { 2012 // If there were values defined in BB that are used outside the block, then we 2013 // now have to update all uses of the value to use either the original value, 2014 // the cloned value, or some PHI derived value. This can require arbitrary 2015 // PHI insertion, of which we are prepared to do, clean these up now. 2016 SSAUpdater SSAUpdate; 2017 SmallVector<Use *, 16> UsesToRename; 2018 2019 for (Instruction &I : *BB) { 2020 // Scan all uses of this instruction to see if it is used outside of its 2021 // block, and if so, record them in UsesToRename. 2022 for (Use &U : I.uses()) { 2023 Instruction *User = cast<Instruction>(U.getUser()); 2024 if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 2025 if (UserPN->getIncomingBlock(U) == BB) 2026 continue; 2027 } else if (User->getParent() == BB) 2028 continue; 2029 2030 UsesToRename.push_back(&U); 2031 } 2032 2033 // If there are no uses outside the block, we're done with this instruction. 2034 if (UsesToRename.empty()) 2035 continue; 2036 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n"); 2037 2038 // We found a use of I outside of BB. Rename all uses of I that are outside 2039 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks 2040 // with the two values we know. 2041 SSAUpdate.Initialize(I.getType(), I.getName()); 2042 SSAUpdate.AddAvailableValue(BB, &I); 2043 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]); 2044 2045 while (!UsesToRename.empty()) 2046 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); 2047 LLVM_DEBUG(dbgs() << "\n"); 2048 } 2049 } 2050 2051 /// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone 2052 /// arguments that come from PredBB. Return the map from the variables in the 2053 /// source basic block to the variables in the newly created basic block. 2054 DenseMap<Instruction *, Value *> 2055 JumpThreadingPass::cloneInstructions(BasicBlock::iterator BI, 2056 BasicBlock::iterator BE, BasicBlock *NewBB, 2057 BasicBlock *PredBB) { 2058 // We are going to have to map operands from the source basic block to the new 2059 // copy of the block 'NewBB'. If there are PHI nodes in the source basic 2060 // block, evaluate them to account for entry from PredBB. 2061 DenseMap<Instruction *, Value *> ValueMapping; 2062 2063 // Clone the phi nodes of the source basic block into NewBB. The resulting 2064 // phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater 2065 // might need to rewrite the operand of the cloned phi. 2066 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) { 2067 PHINode *NewPN = PHINode::Create(PN->getType(), 1, PN->getName(), NewBB); 2068 NewPN->addIncoming(PN->getIncomingValueForBlock(PredBB), PredBB); 2069 ValueMapping[PN] = NewPN; 2070 } 2071 2072 // Clone noalias scope declarations in the threaded block. When threading a 2073 // loop exit, we would otherwise end up with two idential scope declarations 2074 // visible at the same time. 2075 SmallVector<MDNode *> NoAliasScopes; 2076 DenseMap<MDNode *, MDNode *> ClonedScopes; 2077 LLVMContext &Context = PredBB->getContext(); 2078 identifyNoAliasScopesToClone(BI, BE, NoAliasScopes); 2079 cloneNoAliasScopes(NoAliasScopes, ClonedScopes, "thread", Context); 2080 2081 // Clone the non-phi instructions of the source basic block into NewBB, 2082 // keeping track of the mapping and using it to remap operands in the cloned 2083 // instructions. 2084 for (; BI != BE; ++BI) { 2085 Instruction *New = BI->clone(); 2086 New->setName(BI->getName()); 2087 NewBB->getInstList().push_back(New); 2088 ValueMapping[&*BI] = New; 2089 adaptNoAliasScopes(New, ClonedScopes, Context); 2090 2091 // Remap operands to patch up intra-block references. 2092 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 2093 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 2094 DenseMap<Instruction *, Value *>::iterator I = ValueMapping.find(Inst); 2095 if (I != ValueMapping.end()) 2096 New->setOperand(i, I->second); 2097 } 2098 } 2099 2100 return ValueMapping; 2101 } 2102 2103 /// Attempt to thread through two successive basic blocks. 2104 bool JumpThreadingPass::maybethreadThroughTwoBasicBlocks(BasicBlock *BB, 2105 Value *Cond) { 2106 // Consider: 2107 // 2108 // PredBB: 2109 // %var = phi i32* [ null, %bb1 ], [ @a, %bb2 ] 2110 // %tobool = icmp eq i32 %cond, 0 2111 // br i1 %tobool, label %BB, label ... 2112 // 2113 // BB: 2114 // %cmp = icmp eq i32* %var, null 2115 // br i1 %cmp, label ..., label ... 2116 // 2117 // We don't know the value of %var at BB even if we know which incoming edge 2118 // we take to BB. However, once we duplicate PredBB for each of its incoming 2119 // edges (say, PredBB1 and PredBB2), we know the value of %var in each copy of 2120 // PredBB. Then we can thread edges PredBB1->BB and PredBB2->BB through BB. 2121 2122 // Require that BB end with a Branch for simplicity. 2123 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 2124 if (!CondBr) 2125 return false; 2126 2127 // BB must have exactly one predecessor. 2128 BasicBlock *PredBB = BB->getSinglePredecessor(); 2129 if (!PredBB) 2130 return false; 2131 2132 // Require that PredBB end with a conditional Branch. If PredBB ends with an 2133 // unconditional branch, we should be merging PredBB and BB instead. For 2134 // simplicity, we don't deal with a switch. 2135 BranchInst *PredBBBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 2136 if (!PredBBBranch || PredBBBranch->isUnconditional()) 2137 return false; 2138 2139 // If PredBB has exactly one incoming edge, we don't gain anything by copying 2140 // PredBB. 2141 if (PredBB->getSinglePredecessor()) 2142 return false; 2143 2144 // Don't thread through PredBB if it contains a successor edge to itself, in 2145 // which case we would infinite loop. Suppose we are threading an edge from 2146 // PredPredBB through PredBB and BB to SuccBB with PredBB containing a 2147 // successor edge to itself. If we allowed jump threading in this case, we 2148 // could duplicate PredBB and BB as, say, PredBB.thread and BB.thread. Since 2149 // PredBB.thread has a successor edge to PredBB, we would immediately come up 2150 // with another jump threading opportunity from PredBB.thread through PredBB 2151 // and BB to SuccBB. This jump threading would repeatedly occur. That is, we 2152 // would keep peeling one iteration from PredBB. 2153 if (llvm::is_contained(successors(PredBB), PredBB)) 2154 return false; 2155 2156 // Don't thread across a loop header. 2157 if (LoopHeaders.count(PredBB)) 2158 return false; 2159 2160 // Avoid complication with duplicating EH pads. 2161 if (PredBB->isEHPad()) 2162 return false; 2163 2164 // Find a predecessor that we can thread. For simplicity, we only consider a 2165 // successor edge out of BB to which we thread exactly one incoming edge into 2166 // PredBB. 2167 unsigned ZeroCount = 0; 2168 unsigned OneCount = 0; 2169 BasicBlock *ZeroPred = nullptr; 2170 BasicBlock *OnePred = nullptr; 2171 for (BasicBlock *P : predecessors(PredBB)) { 2172 // If PredPred ends with IndirectBrInst, we can't handle it. 2173 if (isa<IndirectBrInst>(P->getTerminator())) 2174 continue; 2175 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>( 2176 evaluateOnPredecessorEdge(BB, P, Cond))) { 2177 if (CI->isZero()) { 2178 ZeroCount++; 2179 ZeroPred = P; 2180 } else if (CI->isOne()) { 2181 OneCount++; 2182 OnePred = P; 2183 } 2184 } 2185 } 2186 2187 // Disregard complicated cases where we have to thread multiple edges. 2188 BasicBlock *PredPredBB; 2189 if (ZeroCount == 1) { 2190 PredPredBB = ZeroPred; 2191 } else if (OneCount == 1) { 2192 PredPredBB = OnePred; 2193 } else { 2194 return false; 2195 } 2196 2197 BasicBlock *SuccBB = CondBr->getSuccessor(PredPredBB == ZeroPred); 2198 2199 // If threading to the same block as we come from, we would infinite loop. 2200 if (SuccBB == BB) { 2201 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 2202 << "' - would thread to self!\n"); 2203 return false; 2204 } 2205 2206 // If threading this would thread across a loop header, don't thread the edge. 2207 // See the comments above findLoopHeaders for justifications and caveats. 2208 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) { 2209 LLVM_DEBUG({ 2210 bool BBIsHeader = LoopHeaders.count(BB); 2211 bool SuccIsHeader = LoopHeaders.count(SuccBB); 2212 dbgs() << " Not threading across " 2213 << (BBIsHeader ? "loop header BB '" : "block BB '") 2214 << BB->getName() << "' to dest " 2215 << (SuccIsHeader ? "loop header BB '" : "block BB '") 2216 << SuccBB->getName() 2217 << "' - it might create an irreducible loop!\n"; 2218 }); 2219 return false; 2220 } 2221 2222 // Compute the cost of duplicating BB and PredBB. 2223 unsigned BBCost = getJumpThreadDuplicationCost( 2224 TTI, BB, BB->getTerminator(), BBDupThreshold); 2225 unsigned PredBBCost = getJumpThreadDuplicationCost( 2226 TTI, PredBB, PredBB->getTerminator(), BBDupThreshold); 2227 2228 // Give up if costs are too high. We need to check BBCost and PredBBCost 2229 // individually before checking their sum because getJumpThreadDuplicationCost 2230 // return (unsigned)~0 for those basic blocks that cannot be duplicated. 2231 if (BBCost > BBDupThreshold || PredBBCost > BBDupThreshold || 2232 BBCost + PredBBCost > BBDupThreshold) { 2233 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName() 2234 << "' - Cost is too high: " << PredBBCost 2235 << " for PredBB, " << BBCost << "for BB\n"); 2236 return false; 2237 } 2238 2239 // Now we are ready to duplicate PredBB. 2240 threadThroughTwoBasicBlocks(PredPredBB, PredBB, BB, SuccBB); 2241 return true; 2242 } 2243 2244 void JumpThreadingPass::threadThroughTwoBasicBlocks(BasicBlock *PredPredBB, 2245 BasicBlock *PredBB, 2246 BasicBlock *BB, 2247 BasicBlock *SuccBB) { 2248 LLVM_DEBUG(dbgs() << " Threading through '" << PredBB->getName() << "' and '" 2249 << BB->getName() << "'\n"); 2250 2251 BranchInst *CondBr = cast<BranchInst>(BB->getTerminator()); 2252 BranchInst *PredBBBranch = cast<BranchInst>(PredBB->getTerminator()); 2253 2254 BasicBlock *NewBB = 2255 BasicBlock::Create(PredBB->getContext(), PredBB->getName() + ".thread", 2256 PredBB->getParent(), PredBB); 2257 NewBB->moveAfter(PredBB); 2258 2259 // Set the block frequency of NewBB. 2260 if (HasProfileData) { 2261 auto NewBBFreq = BFI->getBlockFreq(PredPredBB) * 2262 BPI->getEdgeProbability(PredPredBB, PredBB); 2263 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency()); 2264 } 2265 2266 // We are going to have to map operands from the original BB block to the new 2267 // copy of the block 'NewBB'. If there are PHI nodes in PredBB, evaluate them 2268 // to account for entry from PredPredBB. 2269 DenseMap<Instruction *, Value *> ValueMapping = 2270 cloneInstructions(PredBB->begin(), PredBB->end(), NewBB, PredPredBB); 2271 2272 // Copy the edge probabilities from PredBB to NewBB. 2273 if (HasProfileData) 2274 BPI->copyEdgeProbabilities(PredBB, NewBB); 2275 2276 // Update the terminator of PredPredBB to jump to NewBB instead of PredBB. 2277 // This eliminates predecessors from PredPredBB, which requires us to simplify 2278 // any PHI nodes in PredBB. 2279 Instruction *PredPredTerm = PredPredBB->getTerminator(); 2280 for (unsigned i = 0, e = PredPredTerm->getNumSuccessors(); i != e; ++i) 2281 if (PredPredTerm->getSuccessor(i) == PredBB) { 2282 PredBB->removePredecessor(PredPredBB, true); 2283 PredPredTerm->setSuccessor(i, NewBB); 2284 } 2285 2286 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(0), PredBB, NewBB, 2287 ValueMapping); 2288 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(1), PredBB, NewBB, 2289 ValueMapping); 2290 2291 DTU->applyUpdatesPermissive( 2292 {{DominatorTree::Insert, NewBB, CondBr->getSuccessor(0)}, 2293 {DominatorTree::Insert, NewBB, CondBr->getSuccessor(1)}, 2294 {DominatorTree::Insert, PredPredBB, NewBB}, 2295 {DominatorTree::Delete, PredPredBB, PredBB}}); 2296 2297 updateSSA(PredBB, NewBB, ValueMapping); 2298 2299 // Clean up things like PHI nodes with single operands, dead instructions, 2300 // etc. 2301 SimplifyInstructionsInBlock(NewBB, TLI); 2302 SimplifyInstructionsInBlock(PredBB, TLI); 2303 2304 SmallVector<BasicBlock *, 1> PredsToFactor; 2305 PredsToFactor.push_back(NewBB); 2306 threadEdge(BB, PredsToFactor, SuccBB); 2307 } 2308 2309 /// tryThreadEdge - Thread an edge if it's safe and profitable to do so. 2310 bool JumpThreadingPass::tryThreadEdge( 2311 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs, 2312 BasicBlock *SuccBB) { 2313 // If threading to the same block as we come from, we would infinite loop. 2314 if (SuccBB == BB) { 2315 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName() 2316 << "' - would thread to self!\n"); 2317 return false; 2318 } 2319 2320 // If threading this would thread across a loop header, don't thread the edge. 2321 // See the comments above findLoopHeaders for justifications and caveats. 2322 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) { 2323 LLVM_DEBUG({ 2324 bool BBIsHeader = LoopHeaders.count(BB); 2325 bool SuccIsHeader = LoopHeaders.count(SuccBB); 2326 dbgs() << " Not threading across " 2327 << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() 2328 << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") 2329 << SuccBB->getName() << "' - it might create an irreducible loop!\n"; 2330 }); 2331 return false; 2332 } 2333 2334 unsigned JumpThreadCost = getJumpThreadDuplicationCost( 2335 TTI, BB, BB->getTerminator(), BBDupThreshold); 2336 if (JumpThreadCost > BBDupThreshold) { 2337 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName() 2338 << "' - Cost is too high: " << JumpThreadCost << "\n"); 2339 return false; 2340 } 2341 2342 threadEdge(BB, PredBBs, SuccBB); 2343 return true; 2344 } 2345 2346 /// threadEdge - We have decided that it is safe and profitable to factor the 2347 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB 2348 /// across BB. Transform the IR to reflect this change. 2349 void JumpThreadingPass::threadEdge(BasicBlock *BB, 2350 const SmallVectorImpl<BasicBlock *> &PredBBs, 2351 BasicBlock *SuccBB) { 2352 assert(SuccBB != BB && "Don't create an infinite loop"); 2353 2354 assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) && 2355 "Don't thread across loop headers"); 2356 2357 // And finally, do it! Start by factoring the predecessors if needed. 2358 BasicBlock *PredBB; 2359 if (PredBBs.size() == 1) 2360 PredBB = PredBBs[0]; 2361 else { 2362 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size() 2363 << " common predecessors.\n"); 2364 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm"); 2365 } 2366 2367 // And finally, do it! 2368 LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() 2369 << "' to '" << SuccBB->getName() 2370 << ", across block:\n " << *BB << "\n"); 2371 2372 LVI->threadEdge(PredBB, BB, SuccBB); 2373 2374 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), 2375 BB->getName()+".thread", 2376 BB->getParent(), BB); 2377 NewBB->moveAfter(PredBB); 2378 2379 // Set the block frequency of NewBB. 2380 if (HasProfileData) { 2381 auto NewBBFreq = 2382 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB); 2383 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency()); 2384 } 2385 2386 // Copy all the instructions from BB to NewBB except the terminator. 2387 DenseMap<Instruction *, Value *> ValueMapping = 2388 cloneInstructions(BB->begin(), std::prev(BB->end()), NewBB, PredBB); 2389 2390 // We didn't copy the terminator from BB over to NewBB, because there is now 2391 // an unconditional jump to SuccBB. Insert the unconditional jump. 2392 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB); 2393 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc()); 2394 2395 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 2396 // PHI nodes for NewBB now. 2397 addPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); 2398 2399 // Update the terminator of PredBB to jump to NewBB instead of BB. This 2400 // eliminates predecessors from BB, which requires us to simplify any PHI 2401 // nodes in BB. 2402 Instruction *PredTerm = PredBB->getTerminator(); 2403 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 2404 if (PredTerm->getSuccessor(i) == BB) { 2405 BB->removePredecessor(PredBB, true); 2406 PredTerm->setSuccessor(i, NewBB); 2407 } 2408 2409 // Enqueue required DT updates. 2410 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB}, 2411 {DominatorTree::Insert, PredBB, NewBB}, 2412 {DominatorTree::Delete, PredBB, BB}}); 2413 2414 updateSSA(BB, NewBB, ValueMapping); 2415 2416 // At this point, the IR is fully up to date and consistent. Do a quick scan 2417 // over the new instructions and zap any that are constants or dead. This 2418 // frequently happens because of phi translation. 2419 SimplifyInstructionsInBlock(NewBB, TLI); 2420 2421 // Update the edge weight from BB to SuccBB, which should be less than before. 2422 updateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB); 2423 2424 // Threaded an edge! 2425 ++NumThreads; 2426 } 2427 2428 /// Create a new basic block that will be the predecessor of BB and successor of 2429 /// all blocks in Preds. When profile data is available, update the frequency of 2430 /// this new block. 2431 BasicBlock *JumpThreadingPass::splitBlockPreds(BasicBlock *BB, 2432 ArrayRef<BasicBlock *> Preds, 2433 const char *Suffix) { 2434 SmallVector<BasicBlock *, 2> NewBBs; 2435 2436 // Collect the frequencies of all predecessors of BB, which will be used to 2437 // update the edge weight of the result of splitting predecessors. 2438 DenseMap<BasicBlock *, BlockFrequency> FreqMap; 2439 if (HasProfileData) 2440 for (auto Pred : Preds) 2441 FreqMap.insert(std::make_pair( 2442 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB))); 2443 2444 // In the case when BB is a LandingPad block we create 2 new predecessors 2445 // instead of just one. 2446 if (BB->isLandingPad()) { 2447 std::string NewName = std::string(Suffix) + ".split-lp"; 2448 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs); 2449 } else { 2450 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix)); 2451 } 2452 2453 std::vector<DominatorTree::UpdateType> Updates; 2454 Updates.reserve((2 * Preds.size()) + NewBBs.size()); 2455 for (auto NewBB : NewBBs) { 2456 BlockFrequency NewBBFreq(0); 2457 Updates.push_back({DominatorTree::Insert, NewBB, BB}); 2458 for (auto Pred : predecessors(NewBB)) { 2459 Updates.push_back({DominatorTree::Delete, Pred, BB}); 2460 Updates.push_back({DominatorTree::Insert, Pred, NewBB}); 2461 if (HasProfileData) // Update frequencies between Pred -> NewBB. 2462 NewBBFreq += FreqMap.lookup(Pred); 2463 } 2464 if (HasProfileData) // Apply the summed frequency to NewBB. 2465 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency()); 2466 } 2467 2468 DTU->applyUpdatesPermissive(Updates); 2469 return NewBBs[0]; 2470 } 2471 2472 bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) { 2473 const Instruction *TI = BB->getTerminator(); 2474 assert(TI->getNumSuccessors() > 1 && "not a split"); 2475 2476 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof); 2477 if (!WeightsNode) 2478 return false; 2479 2480 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0)); 2481 if (MDName->getString() != "branch_weights") 2482 return false; 2483 2484 // Ensure there are weights for all of the successors. Note that the first 2485 // operand to the metadata node is a name, not a weight. 2486 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1; 2487 } 2488 2489 /// Update the block frequency of BB and branch weight and the metadata on the 2490 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 - 2491 /// Freq(PredBB->BB) / Freq(BB->SuccBB). 2492 void JumpThreadingPass::updateBlockFreqAndEdgeWeight(BasicBlock *PredBB, 2493 BasicBlock *BB, 2494 BasicBlock *NewBB, 2495 BasicBlock *SuccBB) { 2496 if (!HasProfileData) 2497 return; 2498 2499 assert(BFI && BPI && "BFI & BPI should have been created here"); 2500 2501 // As the edge from PredBB to BB is deleted, we have to update the block 2502 // frequency of BB. 2503 auto BBOrigFreq = BFI->getBlockFreq(BB); 2504 auto NewBBFreq = BFI->getBlockFreq(NewBB); 2505 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB); 2506 auto BBNewFreq = BBOrigFreq - NewBBFreq; 2507 BFI->setBlockFreq(BB, BBNewFreq.getFrequency()); 2508 2509 // Collect updated outgoing edges' frequencies from BB and use them to update 2510 // edge probabilities. 2511 SmallVector<uint64_t, 4> BBSuccFreq; 2512 for (BasicBlock *Succ : successors(BB)) { 2513 auto SuccFreq = (Succ == SuccBB) 2514 ? BB2SuccBBFreq - NewBBFreq 2515 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ); 2516 BBSuccFreq.push_back(SuccFreq.getFrequency()); 2517 } 2518 2519 uint64_t MaxBBSuccFreq = 2520 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end()); 2521 2522 SmallVector<BranchProbability, 4> BBSuccProbs; 2523 if (MaxBBSuccFreq == 0) 2524 BBSuccProbs.assign(BBSuccFreq.size(), 2525 {1, static_cast<unsigned>(BBSuccFreq.size())}); 2526 else { 2527 for (uint64_t Freq : BBSuccFreq) 2528 BBSuccProbs.push_back( 2529 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq)); 2530 // Normalize edge probabilities so that they sum up to one. 2531 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(), 2532 BBSuccProbs.end()); 2533 } 2534 2535 // Update edge probabilities in BPI. 2536 BPI->setEdgeProbability(BB, BBSuccProbs); 2537 2538 // Update the profile metadata as well. 2539 // 2540 // Don't do this if the profile of the transformed blocks was statically 2541 // estimated. (This could occur despite the function having an entry 2542 // frequency in completely cold parts of the CFG.) 2543 // 2544 // In this case we don't want to suggest to subsequent passes that the 2545 // calculated weights are fully consistent. Consider this graph: 2546 // 2547 // check_1 2548 // 50% / | 2549 // eq_1 | 50% 2550 // \ | 2551 // check_2 2552 // 50% / | 2553 // eq_2 | 50% 2554 // \ | 2555 // check_3 2556 // 50% / | 2557 // eq_3 | 50% 2558 // \ | 2559 // 2560 // Assuming the blocks check_* all compare the same value against 1, 2 and 3, 2561 // the overall probabilities are inconsistent; the total probability that the 2562 // value is either 1, 2 or 3 is 150%. 2563 // 2564 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3 2565 // becomes 0%. This is even worse if the edge whose probability becomes 0% is 2566 // the loop exit edge. Then based solely on static estimation we would assume 2567 // the loop was extremely hot. 2568 // 2569 // FIXME this locally as well so that BPI and BFI are consistent as well. We 2570 // shouldn't make edges extremely likely or unlikely based solely on static 2571 // estimation. 2572 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) { 2573 SmallVector<uint32_t, 4> Weights; 2574 for (auto Prob : BBSuccProbs) 2575 Weights.push_back(Prob.getNumerator()); 2576 2577 auto TI = BB->getTerminator(); 2578 TI->setMetadata( 2579 LLVMContext::MD_prof, 2580 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights)); 2581 } 2582 } 2583 2584 /// duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch 2585 /// to BB which contains an i1 PHI node and a conditional branch on that PHI. 2586 /// If we can duplicate the contents of BB up into PredBB do so now, this 2587 /// improves the odds that the branch will be on an analyzable instruction like 2588 /// a compare. 2589 bool JumpThreadingPass::duplicateCondBranchOnPHIIntoPred( 2590 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) { 2591 assert(!PredBBs.empty() && "Can't handle an empty set"); 2592 2593 // If BB is a loop header, then duplicating this block outside the loop would 2594 // cause us to transform this into an irreducible loop, don't do this. 2595 // See the comments above findLoopHeaders for justifications and caveats. 2596 if (LoopHeaders.count(BB)) { 2597 LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() 2598 << "' into predecessor block '" << PredBBs[0]->getName() 2599 << "' - it might create an irreducible loop!\n"); 2600 return false; 2601 } 2602 2603 unsigned DuplicationCost = getJumpThreadDuplicationCost( 2604 TTI, BB, BB->getTerminator(), BBDupThreshold); 2605 if (DuplicationCost > BBDupThreshold) { 2606 LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() 2607 << "' - Cost is too high: " << DuplicationCost << "\n"); 2608 return false; 2609 } 2610 2611 // And finally, do it! Start by factoring the predecessors if needed. 2612 std::vector<DominatorTree::UpdateType> Updates; 2613 BasicBlock *PredBB; 2614 if (PredBBs.size() == 1) 2615 PredBB = PredBBs[0]; 2616 else { 2617 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size() 2618 << " common predecessors.\n"); 2619 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm"); 2620 } 2621 Updates.push_back({DominatorTree::Delete, PredBB, BB}); 2622 2623 // Okay, we decided to do this! Clone all the instructions in BB onto the end 2624 // of PredBB. 2625 LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName() 2626 << "' into end of '" << PredBB->getName() 2627 << "' to eliminate branch on phi. Cost: " 2628 << DuplicationCost << " block is:" << *BB << "\n"); 2629 2630 // Unless PredBB ends with an unconditional branch, split the edge so that we 2631 // can just clone the bits from BB into the end of the new PredBB. 2632 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); 2633 2634 if (!OldPredBranch || !OldPredBranch->isUnconditional()) { 2635 BasicBlock *OldPredBB = PredBB; 2636 PredBB = SplitEdge(OldPredBB, BB); 2637 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB}); 2638 Updates.push_back({DominatorTree::Insert, PredBB, BB}); 2639 Updates.push_back({DominatorTree::Delete, OldPredBB, BB}); 2640 OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); 2641 } 2642 2643 // We are going to have to map operands from the original BB block into the 2644 // PredBB block. Evaluate PHI nodes in BB. 2645 DenseMap<Instruction*, Value*> ValueMapping; 2646 2647 BasicBlock::iterator BI = BB->begin(); 2648 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 2649 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 2650 // Clone the non-phi instructions of BB into PredBB, keeping track of the 2651 // mapping and using it to remap operands in the cloned instructions. 2652 for (; BI != BB->end(); ++BI) { 2653 Instruction *New = BI->clone(); 2654 2655 // Remap operands to patch up intra-block references. 2656 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 2657 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 2658 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 2659 if (I != ValueMapping.end()) 2660 New->setOperand(i, I->second); 2661 } 2662 2663 // If this instruction can be simplified after the operands are updated, 2664 // just use the simplified value instead. This frequently happens due to 2665 // phi translation. 2666 if (Value *IV = simplifyInstruction( 2667 New, 2668 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) { 2669 ValueMapping[&*BI] = IV; 2670 if (!New->mayHaveSideEffects()) { 2671 New->deleteValue(); 2672 New = nullptr; 2673 } 2674 } else { 2675 ValueMapping[&*BI] = New; 2676 } 2677 if (New) { 2678 // Otherwise, insert the new instruction into the block. 2679 New->setName(BI->getName()); 2680 PredBB->getInstList().insert(OldPredBranch->getIterator(), New); 2681 // Update Dominance from simplified New instruction operands. 2682 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 2683 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i))) 2684 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB}); 2685 } 2686 } 2687 2688 // Check to see if the targets of the branch had PHI nodes. If so, we need to 2689 // add entries to the PHI nodes for branch from PredBB now. 2690 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); 2691 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, 2692 ValueMapping); 2693 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, 2694 ValueMapping); 2695 2696 updateSSA(BB, PredBB, ValueMapping); 2697 2698 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge 2699 // that we nuked. 2700 BB->removePredecessor(PredBB, true); 2701 2702 // Remove the unconditional branch at the end of the PredBB block. 2703 OldPredBranch->eraseFromParent(); 2704 if (HasProfileData) 2705 BPI->copyEdgeProbabilities(BB, PredBB); 2706 DTU->applyUpdatesPermissive(Updates); 2707 2708 ++NumDupes; 2709 return true; 2710 } 2711 2712 // Pred is a predecessor of BB with an unconditional branch to BB. SI is 2713 // a Select instruction in Pred. BB has other predecessors and SI is used in 2714 // a PHI node in BB. SI has no other use. 2715 // A new basic block, NewBB, is created and SI is converted to compare and 2716 // conditional branch. SI is erased from parent. 2717 void JumpThreadingPass::unfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB, 2718 SelectInst *SI, PHINode *SIUse, 2719 unsigned Idx) { 2720 // Expand the select. 2721 // 2722 // Pred -- 2723 // | v 2724 // | NewBB 2725 // | | 2726 // |----- 2727 // v 2728 // BB 2729 BranchInst *PredTerm = cast<BranchInst>(Pred->getTerminator()); 2730 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold", 2731 BB->getParent(), BB); 2732 // Move the unconditional branch to NewBB. 2733 PredTerm->removeFromParent(); 2734 NewBB->getInstList().insert(NewBB->end(), PredTerm); 2735 // Create a conditional branch and update PHI nodes. 2736 auto *BI = BranchInst::Create(NewBB, BB, SI->getCondition(), Pred); 2737 BI->applyMergedLocation(PredTerm->getDebugLoc(), SI->getDebugLoc()); 2738 SIUse->setIncomingValue(Idx, SI->getFalseValue()); 2739 SIUse->addIncoming(SI->getTrueValue(), NewBB); 2740 2741 // The select is now dead. 2742 SI->eraseFromParent(); 2743 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB}, 2744 {DominatorTree::Insert, Pred, NewBB}}); 2745 2746 // Update any other PHI nodes in BB. 2747 for (BasicBlock::iterator BI = BB->begin(); 2748 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI) 2749 if (Phi != SIUse) 2750 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB); 2751 } 2752 2753 bool JumpThreadingPass::tryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) { 2754 PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition()); 2755 2756 if (!CondPHI || CondPHI->getParent() != BB) 2757 return false; 2758 2759 for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) { 2760 BasicBlock *Pred = CondPHI->getIncomingBlock(I); 2761 SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I)); 2762 2763 // The second and third condition can be potentially relaxed. Currently 2764 // the conditions help to simplify the code and allow us to reuse existing 2765 // code, developed for tryToUnfoldSelect(CmpInst *, BasicBlock *) 2766 if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse()) 2767 continue; 2768 2769 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator()); 2770 if (!PredTerm || !PredTerm->isUnconditional()) 2771 continue; 2772 2773 unfoldSelectInstr(Pred, BB, PredSI, CondPHI, I); 2774 return true; 2775 } 2776 return false; 2777 } 2778 2779 /// tryToUnfoldSelect - Look for blocks of the form 2780 /// bb1: 2781 /// %a = select 2782 /// br bb2 2783 /// 2784 /// bb2: 2785 /// %p = phi [%a, %bb1] ... 2786 /// %c = icmp %p 2787 /// br i1 %c 2788 /// 2789 /// And expand the select into a branch structure if one of its arms allows %c 2790 /// to be folded. This later enables threading from bb1 over bb2. 2791 bool JumpThreadingPass::tryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) { 2792 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); 2793 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0)); 2794 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1)); 2795 2796 if (!CondBr || !CondBr->isConditional() || !CondLHS || 2797 CondLHS->getParent() != BB) 2798 return false; 2799 2800 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) { 2801 BasicBlock *Pred = CondLHS->getIncomingBlock(I); 2802 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I)); 2803 2804 // Look if one of the incoming values is a select in the corresponding 2805 // predecessor. 2806 if (!SI || SI->getParent() != Pred || !SI->hasOneUse()) 2807 continue; 2808 2809 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator()); 2810 if (!PredTerm || !PredTerm->isUnconditional()) 2811 continue; 2812 2813 // Now check if one of the select values would allow us to constant fold the 2814 // terminator in BB. We don't do the transform if both sides fold, those 2815 // cases will be threaded in any case. 2816 LazyValueInfo::Tristate LHSFolds = 2817 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1), 2818 CondRHS, Pred, BB, CondCmp); 2819 LazyValueInfo::Tristate RHSFolds = 2820 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2), 2821 CondRHS, Pred, BB, CondCmp); 2822 if ((LHSFolds != LazyValueInfo::Unknown || 2823 RHSFolds != LazyValueInfo::Unknown) && 2824 LHSFolds != RHSFolds) { 2825 unfoldSelectInstr(Pred, BB, SI, CondLHS, I); 2826 return true; 2827 } 2828 } 2829 return false; 2830 } 2831 2832 /// tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the 2833 /// same BB in the form 2834 /// bb: 2835 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ... 2836 /// %s = select %p, trueval, falseval 2837 /// 2838 /// or 2839 /// 2840 /// bb: 2841 /// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ... 2842 /// %c = cmp %p, 0 2843 /// %s = select %c, trueval, falseval 2844 /// 2845 /// And expand the select into a branch structure. This later enables 2846 /// jump-threading over bb in this pass. 2847 /// 2848 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold 2849 /// select if the associated PHI has at least one constant. If the unfolded 2850 /// select is not jump-threaded, it will be folded again in the later 2851 /// optimizations. 2852 bool JumpThreadingPass::tryToUnfoldSelectInCurrBB(BasicBlock *BB) { 2853 // This transform would reduce the quality of msan diagnostics. 2854 // Disable this transform under MemorySanitizer. 2855 if (BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory)) 2856 return false; 2857 2858 // If threading this would thread across a loop header, don't thread the edge. 2859 // See the comments above findLoopHeaders for justifications and caveats. 2860 if (LoopHeaders.count(BB)) 2861 return false; 2862 2863 for (BasicBlock::iterator BI = BB->begin(); 2864 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) { 2865 // Look for a Phi having at least one constant incoming value. 2866 if (llvm::all_of(PN->incoming_values(), 2867 [](Value *V) { return !isa<ConstantInt>(V); })) 2868 continue; 2869 2870 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) { 2871 using namespace PatternMatch; 2872 2873 // Check if SI is in BB and use V as condition. 2874 if (SI->getParent() != BB) 2875 return false; 2876 Value *Cond = SI->getCondition(); 2877 bool IsAndOr = match(SI, m_CombineOr(m_LogicalAnd(), m_LogicalOr())); 2878 return Cond && Cond == V && Cond->getType()->isIntegerTy(1) && !IsAndOr; 2879 }; 2880 2881 SelectInst *SI = nullptr; 2882 for (Use &U : PN->uses()) { 2883 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) { 2884 // Look for a ICmp in BB that compares PN with a constant and is the 2885 // condition of a Select. 2886 if (Cmp->getParent() == BB && Cmp->hasOneUse() && 2887 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo()))) 2888 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back())) 2889 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) { 2890 SI = SelectI; 2891 break; 2892 } 2893 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) { 2894 // Look for a Select in BB that uses PN as condition. 2895 if (isUnfoldCandidate(SelectI, U.get())) { 2896 SI = SelectI; 2897 break; 2898 } 2899 } 2900 } 2901 2902 if (!SI) 2903 continue; 2904 // Expand the select. 2905 Value *Cond = SI->getCondition(); 2906 if (!isGuaranteedNotToBeUndefOrPoison(Cond, nullptr, SI)) 2907 Cond = new FreezeInst(Cond, "cond.fr", SI); 2908 Instruction *Term = SplitBlockAndInsertIfThen(Cond, SI, false); 2909 BasicBlock *SplitBB = SI->getParent(); 2910 BasicBlock *NewBB = Term->getParent(); 2911 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI); 2912 NewPN->addIncoming(SI->getTrueValue(), Term->getParent()); 2913 NewPN->addIncoming(SI->getFalseValue(), BB); 2914 SI->replaceAllUsesWith(NewPN); 2915 SI->eraseFromParent(); 2916 // NewBB and SplitBB are newly created blocks which require insertion. 2917 std::vector<DominatorTree::UpdateType> Updates; 2918 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3); 2919 Updates.push_back({DominatorTree::Insert, BB, SplitBB}); 2920 Updates.push_back({DominatorTree::Insert, BB, NewBB}); 2921 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB}); 2922 // BB's successors were moved to SplitBB, update DTU accordingly. 2923 for (auto *Succ : successors(SplitBB)) { 2924 Updates.push_back({DominatorTree::Delete, BB, Succ}); 2925 Updates.push_back({DominatorTree::Insert, SplitBB, Succ}); 2926 } 2927 DTU->applyUpdatesPermissive(Updates); 2928 return true; 2929 } 2930 return false; 2931 } 2932 2933 /// Try to propagate a guard from the current BB into one of its predecessors 2934 /// in case if another branch of execution implies that the condition of this 2935 /// guard is always true. Currently we only process the simplest case that 2936 /// looks like: 2937 /// 2938 /// Start: 2939 /// %cond = ... 2940 /// br i1 %cond, label %T1, label %F1 2941 /// T1: 2942 /// br label %Merge 2943 /// F1: 2944 /// br label %Merge 2945 /// Merge: 2946 /// %condGuard = ... 2947 /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ] 2948 /// 2949 /// And cond either implies condGuard or !condGuard. In this case all the 2950 /// instructions before the guard can be duplicated in both branches, and the 2951 /// guard is then threaded to one of them. 2952 bool JumpThreadingPass::processGuards(BasicBlock *BB) { 2953 using namespace PatternMatch; 2954 2955 // We only want to deal with two predecessors. 2956 BasicBlock *Pred1, *Pred2; 2957 auto PI = pred_begin(BB), PE = pred_end(BB); 2958 if (PI == PE) 2959 return false; 2960 Pred1 = *PI++; 2961 if (PI == PE) 2962 return false; 2963 Pred2 = *PI++; 2964 if (PI != PE) 2965 return false; 2966 if (Pred1 == Pred2) 2967 return false; 2968 2969 // Try to thread one of the guards of the block. 2970 // TODO: Look up deeper than to immediate predecessor? 2971 auto *Parent = Pred1->getSinglePredecessor(); 2972 if (!Parent || Parent != Pred2->getSinglePredecessor()) 2973 return false; 2974 2975 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator())) 2976 for (auto &I : *BB) 2977 if (isGuard(&I) && threadGuard(BB, cast<IntrinsicInst>(&I), BI)) 2978 return true; 2979 2980 return false; 2981 } 2982 2983 /// Try to propagate the guard from BB which is the lower block of a diamond 2984 /// to one of its branches, in case if diamond's condition implies guard's 2985 /// condition. 2986 bool JumpThreadingPass::threadGuard(BasicBlock *BB, IntrinsicInst *Guard, 2987 BranchInst *BI) { 2988 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?"); 2989 assert(BI->isConditional() && "Unconditional branch has 2 successors?"); 2990 Value *GuardCond = Guard->getArgOperand(0); 2991 Value *BranchCond = BI->getCondition(); 2992 BasicBlock *TrueDest = BI->getSuccessor(0); 2993 BasicBlock *FalseDest = BI->getSuccessor(1); 2994 2995 auto &DL = BB->getModule()->getDataLayout(); 2996 bool TrueDestIsSafe = false; 2997 bool FalseDestIsSafe = false; 2998 2999 // True dest is safe if BranchCond => GuardCond. 3000 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL); 3001 if (Impl && *Impl) 3002 TrueDestIsSafe = true; 3003 else { 3004 // False dest is safe if !BranchCond => GuardCond. 3005 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false); 3006 if (Impl && *Impl) 3007 FalseDestIsSafe = true; 3008 } 3009 3010 if (!TrueDestIsSafe && !FalseDestIsSafe) 3011 return false; 3012 3013 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest; 3014 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest; 3015 3016 ValueToValueMapTy UnguardedMapping, GuardedMapping; 3017 Instruction *AfterGuard = Guard->getNextNode(); 3018 unsigned Cost = 3019 getJumpThreadDuplicationCost(TTI, BB, AfterGuard, BBDupThreshold); 3020 if (Cost > BBDupThreshold) 3021 return false; 3022 // Duplicate all instructions before the guard and the guard itself to the 3023 // branch where implication is not proved. 3024 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween( 3025 BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU); 3026 assert(GuardedBlock && "Could not create the guarded block?"); 3027 // Duplicate all instructions before the guard in the unguarded branch. 3028 // Since we have successfully duplicated the guarded block and this block 3029 // has fewer instructions, we expect it to succeed. 3030 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween( 3031 BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU); 3032 assert(UnguardedBlock && "Could not create the unguarded block?"); 3033 LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block " 3034 << GuardedBlock->getName() << "\n"); 3035 // Some instructions before the guard may still have uses. For them, we need 3036 // to create Phi nodes merging their copies in both guarded and unguarded 3037 // branches. Those instructions that have no uses can be just removed. 3038 SmallVector<Instruction *, 4> ToRemove; 3039 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI) 3040 if (!isa<PHINode>(&*BI)) 3041 ToRemove.push_back(&*BI); 3042 3043 Instruction *InsertionPoint = &*BB->getFirstInsertionPt(); 3044 assert(InsertionPoint && "Empty block?"); 3045 // Substitute with Phis & remove. 3046 for (auto *Inst : reverse(ToRemove)) { 3047 if (!Inst->use_empty()) { 3048 PHINode *NewPN = PHINode::Create(Inst->getType(), 2); 3049 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock); 3050 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock); 3051 NewPN->insertBefore(InsertionPoint); 3052 Inst->replaceAllUsesWith(NewPN); 3053 } 3054 Inst->eraseFromParent(); 3055 } 3056 return true; 3057 } 3058