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