| 1 | //===- JumpThreading.cpp - Thread control through conditional blocks ------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file implements the Jump Threading pass. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "llvm/Transforms/Scalar/JumpThreading.h" |
| 14 | #include "llvm/ADT/DenseMap.h" |
| 15 | #include "llvm/ADT/MapVector.h" |
| 16 | #include "llvm/ADT/STLExtras.h" |
| 17 | #include "llvm/ADT/ScopeExit.h" |
| 18 | #include "llvm/ADT/SmallPtrSet.h" |
| 19 | #include "llvm/ADT/SmallVector.h" |
| 20 | #include "llvm/ADT/Statistic.h" |
| 21 | #include "llvm/Analysis/AliasAnalysis.h" |
| 22 | #include "llvm/Analysis/BlockFrequencyInfo.h" |
| 23 | #include "llvm/Analysis/BranchProbabilityInfo.h" |
| 24 | #include "llvm/Analysis/CFG.h" |
| 25 | #include "llvm/Analysis/ConstantFolding.h" |
| 26 | #include "llvm/Analysis/GlobalsModRef.h" |
| 27 | #include "llvm/Analysis/GuardUtils.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/MemoryLocation.h" |
| 33 | #include "llvm/Analysis/PostDominators.h" |
| 34 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 35 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 36 | #include "llvm/Analysis/ValueTracking.h" |
| 37 | #include "llvm/IR/BasicBlock.h" |
| 38 | #include "llvm/IR/CFG.h" |
| 39 | #include "llvm/IR/Constant.h" |
| 40 | #include "llvm/IR/ConstantRange.h" |
| 41 | #include "llvm/IR/Constants.h" |
| 42 | #include "llvm/IR/DataLayout.h" |
| 43 | #include "llvm/IR/DebugInfo.h" |
| 44 | #include "llvm/IR/Dominators.h" |
| 45 | #include "llvm/IR/Function.h" |
| 46 | #include "llvm/IR/InstrTypes.h" |
| 47 | #include "llvm/IR/Instruction.h" |
| 48 | #include "llvm/IR/Instructions.h" |
| 49 | #include "llvm/IR/IntrinsicInst.h" |
| 50 | #include "llvm/IR/Intrinsics.h" |
| 51 | #include "llvm/IR/LLVMContext.h" |
| 52 | #include "llvm/IR/MDBuilder.h" |
| 53 | #include "llvm/IR/Metadata.h" |
| 54 | #include "llvm/IR/Module.h" |
| 55 | #include "llvm/IR/PassManager.h" |
| 56 | #include "llvm/IR/PatternMatch.h" |
| 57 | #include "llvm/IR/ProfDataUtils.h" |
| 58 | #include "llvm/IR/Type.h" |
| 59 | #include "llvm/IR/Use.h" |
| 60 | #include "llvm/IR/Value.h" |
| 61 | #include "llvm/Support/BlockFrequency.h" |
| 62 | #include "llvm/Support/BranchProbability.h" |
| 63 | #include "llvm/Support/Casting.h" |
| 64 | #include "llvm/Support/CommandLine.h" |
| 65 | #include "llvm/Support/Debug.h" |
| 66 | #include "llvm/Support/raw_ostream.h" |
| 67 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
| 68 | #include "llvm/Transforms/Utils/Cloning.h" |
| 69 | #include "llvm/Transforms/Utils/Local.h" |
| 70 | #include "llvm/Transforms/Utils/SSAUpdater.h" |
| 71 | #include "llvm/Transforms/Utils/ValueMapper.h" |
| 72 | #include <cassert> |
| 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(Val: 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(Val: 3), cl::Hidden); |
| 98 | |
| 99 | static cl::opt<unsigned> PhiDuplicateThreshold( |
| 100 | "jump-threading-phi-threshold" , |
| 101 | cl::desc("Max PHIs in BB to duplicate for jump threading" ), cl::init(Val: 76), |
| 102 | cl::Hidden); |
| 103 | |
| 104 | static cl::opt<bool> ( |
| 105 | "jump-threading-across-loop-headers" , |
| 106 | cl::desc("Allow JumpThreading to thread across loop headers, for testing" ), |
| 107 | cl::init(Val: false), cl::Hidden); |
| 108 | |
| 109 | JumpThreadingPass::JumpThreadingPass(int T) { |
| 110 | DefaultBBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T); |
| 111 | } |
| 112 | |
| 113 | // Update branch probability information according to conditional |
| 114 | // branch probability. This is usually made possible for cloned branches |
| 115 | // in inline instances by the context specific profile in the caller. |
| 116 | // For instance, |
| 117 | // |
| 118 | // [Block PredBB] |
| 119 | // [Branch PredBr] |
| 120 | // if (t) { |
| 121 | // Block A; |
| 122 | // } else { |
| 123 | // Block B; |
| 124 | // } |
| 125 | // |
| 126 | // [Block BB] |
| 127 | // cond = PN([true, %A], [..., %B]); // PHI node |
| 128 | // [Branch CondBr] |
| 129 | // if (cond) { |
| 130 | // ... // P(cond == true) = 1% |
| 131 | // } |
| 132 | // |
| 133 | // Here we know that when block A is taken, cond must be true, which means |
| 134 | // P(cond == true | A) = 1 |
| 135 | // |
| 136 | // Given that P(cond == true) = P(cond == true | A) * P(A) + |
| 137 | // P(cond == true | B) * P(B) |
| 138 | // we get: |
| 139 | // P(cond == true ) = P(A) + P(cond == true | B) * P(B) |
| 140 | // |
| 141 | // which gives us: |
| 142 | // P(A) is less than P(cond == true), i.e. |
| 143 | // P(t == true) <= P(cond == true) |
| 144 | // |
| 145 | // In other words, if we know P(cond == true) is unlikely, we know |
| 146 | // that P(t == true) is also unlikely. |
| 147 | // |
| 148 | static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) { |
| 149 | CondBrInst *CondBr = dyn_cast<CondBrInst>(Val: BB->getTerminator()); |
| 150 | if (!CondBr) |
| 151 | return; |
| 152 | |
| 153 | uint64_t TrueWeight, FalseWeight; |
| 154 | if (!extractBranchWeights(I: *CondBr, TrueVal&: TrueWeight, FalseVal&: FalseWeight)) |
| 155 | return; |
| 156 | |
| 157 | if (TrueWeight + FalseWeight == 0) |
| 158 | // Zero branch_weights do not give a hint for getting branch probabilities. |
| 159 | // Technically it would result in division by zero denominator, which is |
| 160 | // TrueWeight + FalseWeight. |
| 161 | return; |
| 162 | |
| 163 | // Returns the outgoing edge of the dominating predecessor block |
| 164 | // that leads to the PhiNode's incoming block: |
| 165 | auto GetPredOutEdge = |
| 166 | [](BasicBlock *IncomingBB, |
| 167 | BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> { |
| 168 | auto *PredBB = IncomingBB; |
| 169 | auto *SuccBB = PhiBB; |
| 170 | SmallPtrSet<BasicBlock *, 16> Visited; |
| 171 | while (true) { |
| 172 | if (isa<CondBrInst>(Val: PredBB->getTerminator())) |
| 173 | return {PredBB, SuccBB}; |
| 174 | Visited.insert(Ptr: PredBB); |
| 175 | auto *SinglePredBB = PredBB->getSinglePredecessor(); |
| 176 | if (!SinglePredBB) |
| 177 | return {nullptr, nullptr}; |
| 178 | |
| 179 | // Stop searching when SinglePredBB has been visited. It means we see |
| 180 | // an unreachable loop. |
| 181 | if (Visited.count(Ptr: SinglePredBB)) |
| 182 | return {nullptr, nullptr}; |
| 183 | |
| 184 | SuccBB = PredBB; |
| 185 | PredBB = SinglePredBB; |
| 186 | } |
| 187 | }; |
| 188 | |
| 189 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { |
| 190 | Value *PhiOpnd = PN->getIncomingValue(i); |
| 191 | ConstantInt *CI = dyn_cast<ConstantInt>(Val: PhiOpnd); |
| 192 | |
| 193 | if (!CI || !CI->getType()->isIntegerTy(Bitwidth: 1)) |
| 194 | continue; |
| 195 | |
| 196 | BranchProbability BP = |
| 197 | (CI->isOne() ? BranchProbability::getBranchProbability( |
| 198 | Numerator: TrueWeight, Denominator: TrueWeight + FalseWeight) |
| 199 | : BranchProbability::getBranchProbability( |
| 200 | Numerator: FalseWeight, Denominator: TrueWeight + FalseWeight)); |
| 201 | |
| 202 | auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB); |
| 203 | if (!PredOutEdge.first) |
| 204 | return; |
| 205 | |
| 206 | BasicBlock *PredBB = PredOutEdge.first; |
| 207 | CondBrInst *PredBr = dyn_cast<CondBrInst>(Val: PredBB->getTerminator()); |
| 208 | if (!PredBr) |
| 209 | return; |
| 210 | |
| 211 | uint64_t PredTrueWeight, PredFalseWeight; |
| 212 | // FIXME: We currently only set the profile data when it is missing. |
| 213 | // With PGO, this can be used to refine even existing profile data with |
| 214 | // context information. This needs to be done after more performance |
| 215 | // testing. |
| 216 | if (extractBranchWeights(I: *PredBr, TrueVal&: PredTrueWeight, FalseVal&: PredFalseWeight)) |
| 217 | continue; |
| 218 | |
| 219 | // We can not infer anything useful when BP >= 50%, because BP is the |
| 220 | // upper bound probability value. |
| 221 | if (BP >= BranchProbability(50, 100)) |
| 222 | continue; |
| 223 | |
| 224 | uint32_t Weights[2]; |
| 225 | if (PredBr->getSuccessor(i: 0) == PredOutEdge.second) { |
| 226 | Weights[0] = BP.getNumerator(); |
| 227 | Weights[1] = BP.getCompl().getNumerator(); |
| 228 | } else { |
| 229 | Weights[0] = BP.getCompl().getNumerator(); |
| 230 | Weights[1] = BP.getNumerator(); |
| 231 | } |
| 232 | setBranchWeights(I&: *PredBr, Weights, IsExpected: hasBranchWeightOrigin(I: *PredBr)); |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | PreservedAnalyses JumpThreadingPass::run(Function &F, |
| 237 | FunctionAnalysisManager &AM) { |
| 238 | auto &TTI = AM.getResult<TargetIRAnalysis>(IR&: F); |
| 239 | // Jump Threading has no sense for the targets with divergent CF |
| 240 | if (TTI.hasBranchDivergence(F: &F)) |
| 241 | return PreservedAnalyses::all(); |
| 242 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F); |
| 243 | auto &LVI = AM.getResult<LazyValueAnalysis>(IR&: F); |
| 244 | auto &AA = AM.getResult<AAManager>(IR&: F); |
| 245 | auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 246 | |
| 247 | bool Changed = |
| 248 | runImpl(F, FAM: &AM, TLI: &TLI, TTI: &TTI, LVI: &LVI, AA: &AA, |
| 249 | DTU: std::make_unique<DomTreeUpdater>( |
| 250 | args: &DT, args: nullptr, args: DomTreeUpdater::UpdateStrategy::Lazy), |
| 251 | BFI: nullptr, BPI: nullptr); |
| 252 | |
| 253 | if (!Changed) |
| 254 | return PreservedAnalyses::all(); |
| 255 | |
| 256 | |
| 257 | getDomTreeUpdater()->flush(); |
| 258 | |
| 259 | #if defined(EXPENSIVE_CHECKS) |
| 260 | assert(getDomTreeUpdater()->getDomTree().verify( |
| 261 | DominatorTree::VerificationLevel::Full) && |
| 262 | "DT broken after JumpThreading" ); |
| 263 | assert((!getDomTreeUpdater()->hasPostDomTree() || |
| 264 | getDomTreeUpdater()->getPostDomTree().verify( |
| 265 | PostDominatorTree::VerificationLevel::Full)) && |
| 266 | "PDT broken after JumpThreading" ); |
| 267 | #else |
| 268 | assert(getDomTreeUpdater()->getDomTree().verify( |
| 269 | DominatorTree::VerificationLevel::Fast) && |
| 270 | "DT broken after JumpThreading" ); |
| 271 | assert((!getDomTreeUpdater()->hasPostDomTree() || |
| 272 | getDomTreeUpdater()->getPostDomTree().verify( |
| 273 | PostDominatorTree::VerificationLevel::Fast)) && |
| 274 | "PDT broken after JumpThreading" ); |
| 275 | #endif |
| 276 | |
| 277 | return getPreservedAnalysis(); |
| 278 | } |
| 279 | |
| 280 | bool JumpThreadingPass::runImpl(Function &F_, FunctionAnalysisManager *FAM_, |
| 281 | TargetLibraryInfo *TLI_, |
| 282 | TargetTransformInfo *TTI_, LazyValueInfo *LVI_, |
| 283 | AliasAnalysis *AA_, |
| 284 | std::unique_ptr<DomTreeUpdater> DTU_, |
| 285 | BlockFrequencyInfo *BFI_, |
| 286 | BranchProbabilityInfo *BPI_) { |
| 287 | LLVM_DEBUG(dbgs() << "Jump threading on function '" << F_.getName() << "'\n" ); |
| 288 | F = &F_; |
| 289 | FAM = FAM_; |
| 290 | TLI = TLI_; |
| 291 | TTI = TTI_; |
| 292 | LVI = LVI_; |
| 293 | AA = AA_; |
| 294 | DTU = std::move(DTU_); |
| 295 | BFI = BFI_; |
| 296 | BPI = BPI_; |
| 297 | auto *GuardDecl = Intrinsic::getDeclarationIfExists( |
| 298 | M: F->getParent(), id: Intrinsic::experimental_guard); |
| 299 | HasGuards = GuardDecl && !GuardDecl->use_empty(); |
| 300 | |
| 301 | // Reduce the number of instructions duplicated when optimizing strictly for |
| 302 | // size. |
| 303 | if (BBDuplicateThreshold.getNumOccurrences()) |
| 304 | BBDupThreshold = BBDuplicateThreshold; |
| 305 | else if (F->hasMinSize()) |
| 306 | BBDupThreshold = 3; |
| 307 | else |
| 308 | BBDupThreshold = DefaultBBDupThreshold; |
| 309 | |
| 310 | assert(DTU && "DTU isn't passed into JumpThreading before using it." ); |
| 311 | assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed." ); |
| 312 | DominatorTree &DT = DTU->getDomTree(); |
| 313 | |
| 314 | Unreachable.clear(); |
| 315 | for (auto &BB : *F) |
| 316 | if (!DT.isReachableFromEntry(A: &BB)) |
| 317 | Unreachable.insert(Ptr: &BB); |
| 318 | |
| 319 | if (!ThreadAcrossLoopHeaders) |
| 320 | findLoopHeaders(F&: *F); |
| 321 | |
| 322 | bool EverChanged = false; |
| 323 | bool Changed; |
| 324 | do { |
| 325 | Changed = false; |
| 326 | for (auto &BB : *F) { |
| 327 | if (Unreachable.count(Ptr: &BB)) |
| 328 | continue; |
| 329 | while (processBlock(BB: &BB)) // Thread all of the branches we can over BB. |
| 330 | Changed = ChangedSinceLastAnalysisUpdate = true; |
| 331 | |
| 332 | // Stop processing BB if it's the entry or is now deleted. The following |
| 333 | // routines attempt to eliminate BB and locating a suitable replacement |
| 334 | // for the entry is non-trivial. |
| 335 | if (&BB == &F->getEntryBlock() || DTU->isBBPendingDeletion(DelBB: &BB)) |
| 336 | continue; |
| 337 | |
| 338 | if (pred_empty(BB: &BB)) { |
| 339 | // When processBlock makes BB unreachable it doesn't bother to fix up |
| 340 | // the instructions in it. We must remove BB to prevent invalid IR. |
| 341 | LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName() |
| 342 | << "' with terminator: " << *BB.getTerminator() |
| 343 | << '\n'); |
| 344 | LoopHeaders.erase(Ptr: &BB); |
| 345 | LVI->eraseBlock(BB: &BB); |
| 346 | DeleteDeadBlock(BB: &BB, DTU: DTU.get()); |
| 347 | Changed = ChangedSinceLastAnalysisUpdate = true; |
| 348 | continue; |
| 349 | } |
| 350 | |
| 351 | // processBlock doesn't thread BBs with unconditional TIs. However, if BB |
| 352 | // is "almost empty", we attempt to merge BB with its sole successor. |
| 353 | if (auto *BI = dyn_cast<UncondBrInst>(Val: BB.getTerminator())) { |
| 354 | BasicBlock *Succ = BI->getSuccessor(); |
| 355 | if ( |
| 356 | // The terminator must be the only non-phi instruction in BB. |
| 357 | BB.getFirstNonPHIOrDbg(SkipPseudoOp: true)->isTerminator() && |
| 358 | // Don't alter Loop headers and latches to ensure another pass can |
| 359 | // detect and transform nested loops later. |
| 360 | !LoopHeaders.count(Ptr: &BB) && !LoopHeaders.count(Ptr: Succ) && |
| 361 | TryToSimplifyUncondBranchFromEmptyBlock(BB: &BB, DTU: DTU.get())) { |
| 362 | // BB is valid for cleanup here because we passed in DTU. F remains |
| 363 | // BB's parent until a DTU->getDomTree() event. |
| 364 | LVI->eraseBlock(BB: &BB); |
| 365 | Changed = ChangedSinceLastAnalysisUpdate = true; |
| 366 | } |
| 367 | } |
| 368 | } |
| 369 | EverChanged |= Changed; |
| 370 | } while (Changed); |
| 371 | |
| 372 | // Jump threading may have introduced redundant debug values into F which |
| 373 | // should be removed. |
| 374 | if (EverChanged) |
| 375 | for (auto &BB : *F) { |
| 376 | RemoveRedundantDbgInstrs(BB: &BB); |
| 377 | } |
| 378 | |
| 379 | LoopHeaders.clear(); |
| 380 | return EverChanged; |
| 381 | } |
| 382 | |
| 383 | // Replace uses of Cond with ToVal when safe to do so. If all uses are |
| 384 | // replaced, we can remove Cond. We cannot blindly replace all uses of Cond |
| 385 | // because we may incorrectly replace uses when guards/assumes are uses of |
| 386 | // of `Cond` and we used the guards/assume to reason about the `Cond` value |
| 387 | // at the end of block. RAUW unconditionally replaces all uses |
| 388 | // including the guards/assumes themselves and the uses before the |
| 389 | // guard/assume. |
| 390 | static bool replaceFoldableUses(Instruction *Cond, Value *ToVal, |
| 391 | BasicBlock *KnownAtEndOfBB) { |
| 392 | bool Changed = false; |
| 393 | assert(Cond->getType() == ToVal->getType()); |
| 394 | // We can unconditionally replace all uses in non-local blocks (i.e. uses |
| 395 | // strictly dominated by BB), since LVI information is true from the |
| 396 | // terminator of BB. |
| 397 | if (Cond->getParent() == KnownAtEndOfBB) |
| 398 | Changed |= replaceNonLocalUsesWith(From: Cond, To: ToVal); |
| 399 | for (Instruction &I : reverse(C&: *KnownAtEndOfBB)) { |
| 400 | // Replace any debug-info record users of Cond with ToVal. |
| 401 | for (DbgVariableRecord &DVR : filterDbgVars(R: I.getDbgRecordRange())) |
| 402 | DVR.replaceVariableLocationOp(OldValue: Cond, NewValue: ToVal, AllowEmpty: true); |
| 403 | |
| 404 | // Reached the Cond whose uses we are trying to replace, so there are no |
| 405 | // more uses. |
| 406 | if (&I == Cond) |
| 407 | break; |
| 408 | // We only replace uses in instructions that are guaranteed to reach the end |
| 409 | // of BB, where we know Cond is ToVal. |
| 410 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
| 411 | break; |
| 412 | Changed |= I.replaceUsesOfWith(From: Cond, To: ToVal); |
| 413 | } |
| 414 | if (Cond->use_empty() && !Cond->mayHaveSideEffects()) { |
| 415 | Cond->eraseFromParent(); |
| 416 | Changed = true; |
| 417 | } |
| 418 | return Changed; |
| 419 | } |
| 420 | |
| 421 | /// Return the cost of duplicating a piece of this block from first non-phi |
| 422 | /// and before StopAt instruction to thread across it. Stop scanning the block |
| 423 | /// when exceeding the threshold. If duplication is impossible, returns ~0U. |
| 424 | static unsigned getJumpThreadDuplicationCost(const TargetTransformInfo *TTI, |
| 425 | BasicBlock *BB, |
| 426 | Instruction *StopAt, |
| 427 | unsigned Threshold) { |
| 428 | assert(StopAt->getParent() == BB && "Not an instruction from proper BB?" ); |
| 429 | |
| 430 | // Do not duplicate the BB if it has a lot of PHI nodes. |
| 431 | // If a threadable chain is too long then the number of PHI nodes can add up, |
| 432 | // leading to a substantial increase in compile time when rewriting the SSA. |
| 433 | unsigned PhiCount = 0; |
| 434 | Instruction *FirstNonPHI = nullptr; |
| 435 | for (Instruction &I : *BB) { |
| 436 | if (!isa<PHINode>(Val: &I)) { |
| 437 | FirstNonPHI = &I; |
| 438 | break; |
| 439 | } |
| 440 | if (++PhiCount > PhiDuplicateThreshold) |
| 441 | return ~0U; |
| 442 | } |
| 443 | |
| 444 | /// Ignore PHI nodes, these will be flattened when duplication happens. |
| 445 | BasicBlock::const_iterator I(FirstNonPHI); |
| 446 | |
| 447 | // FIXME: THREADING will delete values that are just used to compute the |
| 448 | // branch, so they shouldn't count against the duplication cost. |
| 449 | |
| 450 | unsigned Bonus = 0; |
| 451 | if (BB->getTerminator() == StopAt) { |
| 452 | // Threading through a switch statement is particularly profitable. If this |
| 453 | // block ends in a switch, decrease its cost to make it more likely to |
| 454 | // happen. |
| 455 | if (isa<SwitchInst>(Val: StopAt)) |
| 456 | Bonus = 6; |
| 457 | |
| 458 | // The same holds for indirect branches, but slightly more so. |
| 459 | if (isa<IndirectBrInst>(Val: StopAt)) |
| 460 | Bonus = 8; |
| 461 | } |
| 462 | |
| 463 | // Bump the threshold up so the early exit from the loop doesn't skip the |
| 464 | // terminator-based Size adjustment at the end. |
| 465 | Threshold += Bonus; |
| 466 | |
| 467 | // Sum up the cost of each instruction until we get to the terminator. Don't |
| 468 | // include the terminator because the copy won't include it. |
| 469 | unsigned Size = 0; |
| 470 | for (; &*I != StopAt; ++I) { |
| 471 | |
| 472 | // Stop scanning the block if we've reached the threshold. |
| 473 | if (Size > Threshold) |
| 474 | return Size; |
| 475 | |
| 476 | // Bail out if this instruction gives back a token type, it is not possible |
| 477 | // to duplicate it if it is used outside this BB. |
| 478 | if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB)) |
| 479 | return ~0U; |
| 480 | |
| 481 | // Blocks with NoDuplicate are modelled as having infinite cost, so they |
| 482 | // are never duplicated. |
| 483 | if (const CallInst *CI = dyn_cast<CallInst>(Val&: I)) |
| 484 | if (CI->cannotDuplicate() || CI->isConvergent()) |
| 485 | return ~0U; |
| 486 | |
| 487 | if (TTI->getInstructionCost(U: &*I, CostKind: TargetTransformInfo::TCK_SizeAndLatency) == |
| 488 | TargetTransformInfo::TCC_Free) |
| 489 | continue; |
| 490 | |
| 491 | // All other instructions count for at least one unit. |
| 492 | ++Size; |
| 493 | |
| 494 | // Calls are more expensive. If they are non-intrinsic calls, we model them |
| 495 | // as having cost of 4. If they are a non-vector intrinsic, we model them |
| 496 | // as having cost of 2 total, and if they are a vector intrinsic, we model |
| 497 | // them as having cost 1. |
| 498 | if (const CallInst *CI = dyn_cast<CallInst>(Val&: I)) { |
| 499 | if (!isa<IntrinsicInst>(Val: CI)) |
| 500 | Size += 3; |
| 501 | else if (!CI->getType()->isVectorTy()) |
| 502 | Size += 1; |
| 503 | } |
| 504 | } |
| 505 | |
| 506 | return Size > Bonus ? Size - Bonus : 0; |
| 507 | } |
| 508 | |
| 509 | /// findLoopHeaders - We do not want jump threading to turn proper loop |
| 510 | /// structures into irreducible loops. Doing this breaks up the loop nesting |
| 511 | /// hierarchy and pessimizes later transformations. To prevent this from |
| 512 | /// happening, we first have to find the loop headers. Here we approximate this |
| 513 | /// by finding targets of backedges in the CFG. |
| 514 | /// |
| 515 | /// Note that there definitely are cases when we want to allow threading of |
| 516 | /// edges across a loop header. For example, threading a jump from outside the |
| 517 | /// loop (the preheader) to an exit block of the loop is definitely profitable. |
| 518 | /// It is also almost always profitable to thread backedges from within the loop |
| 519 | /// to exit blocks, and is often profitable to thread backedges to other blocks |
| 520 | /// within the loop (forming a nested loop). This simple analysis is not rich |
| 521 | /// enough to track all of these properties and keep it up-to-date as the CFG |
| 522 | /// mutates, so we don't allow any of these transformations. |
| 523 | void JumpThreadingPass::(Function &F) { |
| 524 | SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; |
| 525 | FindFunctionBackedges(F, Result&: Edges); |
| 526 | LoopHeaders.insert_range(R: llvm::make_second_range(c&: Edges)); |
| 527 | } |
| 528 | |
| 529 | /// getKnownConstant - Helper method to determine if we can thread over a |
| 530 | /// terminator with the given value as its condition, and if so what value to |
| 531 | /// use for that. What kind of value this is depends on whether we want an |
| 532 | /// integer or a block address, but an undef is always accepted. |
| 533 | /// Returns null if Val is null or not an appropriate constant. |
| 534 | static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { |
| 535 | if (!Val) |
| 536 | return nullptr; |
| 537 | |
| 538 | // Undef is "known" enough. |
| 539 | if (UndefValue *U = dyn_cast<UndefValue>(Val)) |
| 540 | return U; |
| 541 | |
| 542 | if (Preference == WantBlockAddress) |
| 543 | return dyn_cast<BlockAddress>(Val: Val->stripPointerCasts()); |
| 544 | |
| 545 | return dyn_cast<ConstantInt>(Val); |
| 546 | } |
| 547 | |
| 548 | /// computeValueKnownInPredecessors - Given a basic block BB and a value V, see |
| 549 | /// if we can infer that the value is a known ConstantInt/BlockAddress or undef |
| 550 | /// in any of our predecessors. If so, return the known list of value and pred |
| 551 | /// BB in the result vector. |
| 552 | /// |
| 553 | /// This returns true if there were any known values. |
| 554 | bool JumpThreadingPass::computeValueKnownInPredecessorsImpl( |
| 555 | Value *V, BasicBlock *BB, PredValueInfo &Result, |
| 556 | ConstantPreference Preference, SmallPtrSet<Value *, 4> &RecursionSet, |
| 557 | Instruction *CxtI) { |
| 558 | const DataLayout &DL = BB->getDataLayout(); |
| 559 | |
| 560 | // This method walks up use-def chains recursively. Because of this, we could |
| 561 | // get into an infinite loop going around loops in the use-def chain. To |
| 562 | // prevent this, keep track of what (value, block) pairs we've already visited |
| 563 | // and terminate the search if we loop back to them |
| 564 | if (!RecursionSet.insert(Ptr: V).second) |
| 565 | return false; |
| 566 | |
| 567 | // If V is a constant, then it is known in all predecessors. |
| 568 | if (Constant *KC = getKnownConstant(Val: V, Preference)) { |
| 569 | for (BasicBlock *Pred : predecessors(BB)) |
| 570 | Result.emplace_back(Args&: KC, Args&: Pred); |
| 571 | |
| 572 | return !Result.empty(); |
| 573 | } |
| 574 | |
| 575 | // If V is a non-instruction value, or an instruction in a different block, |
| 576 | // then it can't be derived from a PHI. |
| 577 | Instruction *I = dyn_cast<Instruction>(Val: V); |
| 578 | if (!I || I->getParent() != BB) { |
| 579 | |
| 580 | // Okay, if this is a live-in value, see if it has a known value at the any |
| 581 | // edge from our predecessors. |
| 582 | for (BasicBlock *P : predecessors(BB)) { |
| 583 | using namespace PatternMatch; |
| 584 | // If the value is known by LazyValueInfo to be a constant in a |
| 585 | // predecessor, use that information to try to thread this block. |
| 586 | Constant *PredCst = LVI->getConstantOnEdge(V, FromBB: P, ToBB: BB, CxtI); |
| 587 | // If I is a non-local compare-with-constant instruction, use more-rich |
| 588 | // 'getPredicateOnEdge' method. This would be able to handle value |
| 589 | // inequalities better, for example if the compare is "X < 4" and "X < 3" |
| 590 | // is known true but "X < 4" itself is not available. |
| 591 | CmpPredicate Pred; |
| 592 | Value *Val; |
| 593 | Constant *Cst; |
| 594 | if (!PredCst && match(V, P: m_Cmp(Pred, L: m_Value(V&: Val), R: m_Constant(C&: Cst)))) |
| 595 | PredCst = LVI->getPredicateOnEdge(Pred, V: Val, C: Cst, FromBB: P, ToBB: BB, CxtI); |
| 596 | if (Constant *KC = getKnownConstant(Val: PredCst, Preference)) |
| 597 | Result.emplace_back(Args&: KC, Args&: P); |
| 598 | } |
| 599 | |
| 600 | return !Result.empty(); |
| 601 | } |
| 602 | |
| 603 | /// If I is a PHI node, then we know the incoming values for any constants. |
| 604 | if (PHINode *PN = dyn_cast<PHINode>(Val: I)) { |
| 605 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { |
| 606 | Value *InVal = PN->getIncomingValue(i); |
| 607 | if (Constant *KC = getKnownConstant(Val: InVal, Preference)) { |
| 608 | Result.emplace_back(Args&: KC, Args: PN->getIncomingBlock(i)); |
| 609 | } else { |
| 610 | Constant *CI = LVI->getConstantOnEdge(V: InVal, |
| 611 | FromBB: PN->getIncomingBlock(i), |
| 612 | ToBB: BB, CxtI); |
| 613 | if (Constant *KC = getKnownConstant(Val: CI, Preference)) |
| 614 | Result.emplace_back(Args&: KC, Args: PN->getIncomingBlock(i)); |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | return !Result.empty(); |
| 619 | } |
| 620 | |
| 621 | // Handle Cast instructions. |
| 622 | if (CastInst *CI = dyn_cast<CastInst>(Val: I)) { |
| 623 | Value *Source = CI->getOperand(i_nocapture: 0); |
| 624 | PredValueInfoTy Vals; |
| 625 | computeValueKnownInPredecessorsImpl(V: Source, BB, Result&: Vals, Preference, |
| 626 | RecursionSet, CxtI); |
| 627 | if (Vals.empty()) |
| 628 | return false; |
| 629 | |
| 630 | // Convert the known values. |
| 631 | for (auto &Val : Vals) |
| 632 | if (Constant *Folded = ConstantFoldCastOperand(Opcode: CI->getOpcode(), C: Val.first, |
| 633 | DestTy: CI->getType(), DL)) |
| 634 | Result.emplace_back(Args&: Folded, Args&: Val.second); |
| 635 | |
| 636 | return !Result.empty(); |
| 637 | } |
| 638 | |
| 639 | if (FreezeInst *FI = dyn_cast<FreezeInst>(Val: I)) { |
| 640 | Value *Source = FI->getOperand(i_nocapture: 0); |
| 641 | computeValueKnownInPredecessorsImpl(V: Source, BB, Result, Preference, |
| 642 | RecursionSet, CxtI); |
| 643 | |
| 644 | erase_if(C&: Result, P: [](auto &Pair) { |
| 645 | return !isGuaranteedNotToBeUndefOrPoison(Pair.first); |
| 646 | }); |
| 647 | |
| 648 | return !Result.empty(); |
| 649 | } |
| 650 | |
| 651 | // Handle some boolean conditions. |
| 652 | if (I->getType()->getPrimitiveSizeInBits() == 1) { |
| 653 | using namespace PatternMatch; |
| 654 | if (Preference != WantInteger) |
| 655 | return false; |
| 656 | // X | true -> true |
| 657 | // X & false -> false |
| 658 | Value *Op0, *Op1; |
| 659 | if (match(V: I, P: m_LogicalOr(L: m_Value(V&: Op0), R: m_Value(V&: Op1))) || |
| 660 | match(V: I, P: m_LogicalAnd(L: m_Value(V&: Op0), R: m_Value(V&: Op1)))) { |
| 661 | PredValueInfoTy LHSVals, RHSVals; |
| 662 | |
| 663 | computeValueKnownInPredecessorsImpl(V: Op0, BB, Result&: LHSVals, Preference: WantInteger, |
| 664 | RecursionSet, CxtI); |
| 665 | computeValueKnownInPredecessorsImpl(V: Op1, BB, Result&: RHSVals, Preference: WantInteger, |
| 666 | RecursionSet, CxtI); |
| 667 | |
| 668 | if (LHSVals.empty() && RHSVals.empty()) |
| 669 | return false; |
| 670 | |
| 671 | ConstantInt *InterestingVal; |
| 672 | if (match(V: I, P: m_LogicalOr())) |
| 673 | InterestingVal = ConstantInt::getTrue(Context&: I->getContext()); |
| 674 | else |
| 675 | InterestingVal = ConstantInt::getFalse(Context&: I->getContext()); |
| 676 | |
| 677 | SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; |
| 678 | |
| 679 | // Scan for the sentinel. If we find an undef, force it to the |
| 680 | // interesting value: x|undef -> true and x&undef -> false. |
| 681 | for (const auto &LHSVal : LHSVals) |
| 682 | if (LHSVal.first == InterestingVal || isa<UndefValue>(Val: LHSVal.first)) { |
| 683 | Result.emplace_back(Args&: InterestingVal, Args: LHSVal.second); |
| 684 | LHSKnownBBs.insert(Ptr: LHSVal.second); |
| 685 | } |
| 686 | for (const auto &RHSVal : RHSVals) |
| 687 | if (RHSVal.first == InterestingVal || isa<UndefValue>(Val: RHSVal.first)) { |
| 688 | // If we already inferred a value for this block on the LHS, don't |
| 689 | // re-add it. |
| 690 | if (!LHSKnownBBs.count(Ptr: RHSVal.second)) |
| 691 | Result.emplace_back(Args&: InterestingVal, Args: RHSVal.second); |
| 692 | } |
| 693 | |
| 694 | return !Result.empty(); |
| 695 | } |
| 696 | |
| 697 | // Handle the NOT form of XOR. |
| 698 | if (I->getOpcode() == Instruction::Xor && |
| 699 | isa<ConstantInt>(Val: I->getOperand(i: 1)) && |
| 700 | cast<ConstantInt>(Val: I->getOperand(i: 1))->isOne()) { |
| 701 | computeValueKnownInPredecessorsImpl(V: I->getOperand(i: 0), BB, Result, |
| 702 | Preference: WantInteger, RecursionSet, CxtI); |
| 703 | if (Result.empty()) |
| 704 | return false; |
| 705 | |
| 706 | // Invert the known values. |
| 707 | for (auto &R : Result) |
| 708 | R.first = ConstantExpr::getNot(C: R.first); |
| 709 | |
| 710 | return true; |
| 711 | } |
| 712 | |
| 713 | // Try to simplify some other binary operator values. |
| 714 | } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: I)) { |
| 715 | if (Preference != WantInteger) |
| 716 | return false; |
| 717 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1))) { |
| 718 | PredValueInfoTy LHSVals; |
| 719 | computeValueKnownInPredecessorsImpl(V: BO->getOperand(i_nocapture: 0), BB, Result&: LHSVals, |
| 720 | Preference: WantInteger, RecursionSet, CxtI); |
| 721 | |
| 722 | // Try to use constant folding to simplify the binary operator. |
| 723 | for (const auto &LHSVal : LHSVals) { |
| 724 | Constant *V = LHSVal.first; |
| 725 | Constant *Folded = |
| 726 | ConstantFoldBinaryOpOperands(Opcode: BO->getOpcode(), LHS: V, RHS: CI, DL); |
| 727 | |
| 728 | if (Constant *KC = getKnownConstant(Val: Folded, Preference: WantInteger)) |
| 729 | Result.emplace_back(Args&: KC, Args: LHSVal.second); |
| 730 | } |
| 731 | } |
| 732 | |
| 733 | return !Result.empty(); |
| 734 | } |
| 735 | |
| 736 | // Handle compare with phi operand, where the PHI is defined in this block. |
| 737 | if (CmpInst *Cmp = dyn_cast<CmpInst>(Val: I)) { |
| 738 | if (Preference != WantInteger) |
| 739 | return false; |
| 740 | Type *CmpType = Cmp->getType(); |
| 741 | Value *CmpLHS = Cmp->getOperand(i_nocapture: 0); |
| 742 | Value *CmpRHS = Cmp->getOperand(i_nocapture: 1); |
| 743 | CmpInst::Predicate Pred = Cmp->getPredicate(); |
| 744 | |
| 745 | PHINode *PN = dyn_cast<PHINode>(Val: CmpLHS); |
| 746 | if (!PN) |
| 747 | PN = dyn_cast<PHINode>(Val: CmpRHS); |
| 748 | // Do not perform phi translation across a loop header phi, because this |
| 749 | // may result in comparison of values from two different loop iterations. |
| 750 | // FIXME: This check is broken if LoopHeaders is not populated. |
| 751 | if (PN && PN->getParent() == BB && !LoopHeaders.contains(Ptr: BB)) { |
| 752 | const DataLayout &DL = PN->getDataLayout(); |
| 753 | // We can do this simplification if any comparisons fold to true or false. |
| 754 | // See if any do. |
| 755 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { |
| 756 | BasicBlock *PredBB = PN->getIncomingBlock(i); |
| 757 | Value *LHS, *RHS; |
| 758 | if (PN == CmpLHS) { |
| 759 | LHS = PN->getIncomingValue(i); |
| 760 | RHS = CmpRHS->DoPHITranslation(CurBB: BB, PredBB); |
| 761 | } else { |
| 762 | LHS = CmpLHS->DoPHITranslation(CurBB: BB, PredBB); |
| 763 | RHS = PN->getIncomingValue(i); |
| 764 | } |
| 765 | Value *Res = simplifyCmpInst(Predicate: Pred, LHS, RHS, Q: {DL}); |
| 766 | if (!Res) { |
| 767 | if (!isa<Constant>(Val: RHS)) |
| 768 | continue; |
| 769 | |
| 770 | // getPredicateOnEdge call will make no sense if LHS is defined in BB. |
| 771 | auto LHSInst = dyn_cast<Instruction>(Val: LHS); |
| 772 | if (LHSInst && LHSInst->getParent() == BB) |
| 773 | continue; |
| 774 | |
| 775 | Res = LVI->getPredicateOnEdge(Pred, V: LHS, C: cast<Constant>(Val: RHS), FromBB: PredBB, |
| 776 | ToBB: BB, CxtI: CxtI ? CxtI : Cmp); |
| 777 | } |
| 778 | |
| 779 | if (Constant *KC = getKnownConstant(Val: Res, Preference: WantInteger)) |
| 780 | Result.emplace_back(Args&: KC, Args&: PredBB); |
| 781 | } |
| 782 | |
| 783 | return !Result.empty(); |
| 784 | } |
| 785 | |
| 786 | // If comparing a live-in value against a constant, see if we know the |
| 787 | // live-in value on any predecessors. |
| 788 | if (isa<Constant>(Val: CmpRHS) && !CmpType->isVectorTy()) { |
| 789 | Constant *CmpConst = cast<Constant>(Val: CmpRHS); |
| 790 | |
| 791 | if (!isa<Instruction>(Val: CmpLHS) || |
| 792 | cast<Instruction>(Val: CmpLHS)->getParent() != BB) { |
| 793 | for (BasicBlock *P : predecessors(BB)) { |
| 794 | // If the value is known by LazyValueInfo to be a constant in a |
| 795 | // predecessor, use that information to try to thread this block. |
| 796 | Constant *Res = LVI->getPredicateOnEdge(Pred, V: CmpLHS, C: CmpConst, FromBB: P, ToBB: BB, |
| 797 | CxtI: CxtI ? CxtI : Cmp); |
| 798 | if (Constant *KC = getKnownConstant(Val: Res, Preference: WantInteger)) |
| 799 | Result.emplace_back(Args&: KC, Args&: P); |
| 800 | } |
| 801 | |
| 802 | return !Result.empty(); |
| 803 | } |
| 804 | |
| 805 | // InstCombine can fold some forms of constant range checks into |
| 806 | // (icmp (add (x, C1)), C2). See if we have we have such a thing with |
| 807 | // x as a live-in. |
| 808 | { |
| 809 | using namespace PatternMatch; |
| 810 | |
| 811 | Value *AddLHS; |
| 812 | ConstantInt *AddConst; |
| 813 | if (isa<ConstantInt>(Val: CmpConst) && |
| 814 | match(V: CmpLHS, P: m_Add(L: m_Value(V&: AddLHS), R: m_ConstantInt(CI&: AddConst)))) { |
| 815 | if (!isa<Instruction>(Val: AddLHS) || |
| 816 | cast<Instruction>(Val: AddLHS)->getParent() != BB) { |
| 817 | for (BasicBlock *P : predecessors(BB)) { |
| 818 | // If the value is known by LazyValueInfo to be a ConstantRange in |
| 819 | // a predecessor, use that information to try to thread this |
| 820 | // block. |
| 821 | ConstantRange CR = LVI->getConstantRangeOnEdge( |
| 822 | V: AddLHS, FromBB: P, ToBB: BB, CxtI: CxtI ? CxtI : cast<Instruction>(Val: CmpLHS)); |
| 823 | // Propagate the range through the addition. |
| 824 | CR = CR.add(Other: AddConst->getValue()); |
| 825 | |
| 826 | // Get the range where the compare returns true. |
| 827 | ConstantRange CmpRange = ConstantRange::makeExactICmpRegion( |
| 828 | Pred, Other: cast<ConstantInt>(Val: CmpConst)->getValue()); |
| 829 | |
| 830 | Constant *ResC; |
| 831 | if (CmpRange.contains(CR)) |
| 832 | ResC = ConstantInt::getTrue(Ty: CmpType); |
| 833 | else if (CmpRange.inverse().contains(CR)) |
| 834 | ResC = ConstantInt::getFalse(Ty: CmpType); |
| 835 | else |
| 836 | continue; |
| 837 | |
| 838 | Result.emplace_back(Args&: ResC, Args&: P); |
| 839 | } |
| 840 | |
| 841 | return !Result.empty(); |
| 842 | } |
| 843 | } |
| 844 | } |
| 845 | |
| 846 | // Try to find a constant value for the LHS of a comparison, |
| 847 | // and evaluate it statically if we can. |
| 848 | PredValueInfoTy LHSVals; |
| 849 | computeValueKnownInPredecessorsImpl(V: I->getOperand(i: 0), BB, Result&: LHSVals, |
| 850 | Preference: WantInteger, RecursionSet, CxtI); |
| 851 | |
| 852 | for (const auto &LHSVal : LHSVals) { |
| 853 | Constant *V = LHSVal.first; |
| 854 | Constant *Folded = |
| 855 | ConstantFoldCompareInstOperands(Predicate: Pred, LHS: V, RHS: CmpConst, DL); |
| 856 | if (Constant *KC = getKnownConstant(Val: Folded, Preference: WantInteger)) |
| 857 | Result.emplace_back(Args&: KC, Args: LHSVal.second); |
| 858 | } |
| 859 | |
| 860 | return !Result.empty(); |
| 861 | } |
| 862 | } |
| 863 | |
| 864 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: I)) { |
| 865 | // Handle select instructions where at least one operand is a known constant |
| 866 | // and we can figure out the condition value for any predecessor block. |
| 867 | Constant *TrueVal = getKnownConstant(Val: SI->getTrueValue(), Preference); |
| 868 | Constant *FalseVal = getKnownConstant(Val: SI->getFalseValue(), Preference); |
| 869 | PredValueInfoTy Conds; |
| 870 | if ((TrueVal || FalseVal) && |
| 871 | computeValueKnownInPredecessorsImpl(V: SI->getCondition(), BB, Result&: Conds, |
| 872 | Preference: WantInteger, RecursionSet, CxtI)) { |
| 873 | for (auto &C : Conds) { |
| 874 | Constant *Cond = C.first; |
| 875 | |
| 876 | // Figure out what value to use for the condition. |
| 877 | bool KnownCond; |
| 878 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Cond)) { |
| 879 | // A known boolean. |
| 880 | KnownCond = CI->isOne(); |
| 881 | } else { |
| 882 | assert(isa<UndefValue>(Cond) && "Unexpected condition value" ); |
| 883 | // Either operand will do, so be sure to pick the one that's a known |
| 884 | // constant. |
| 885 | // FIXME: Do this more cleverly if both values are known constants? |
| 886 | KnownCond = (TrueVal != nullptr); |
| 887 | } |
| 888 | |
| 889 | // See if the select has a known constant value for this predecessor. |
| 890 | if (Constant *Val = KnownCond ? TrueVal : FalseVal) |
| 891 | Result.emplace_back(Args&: Val, Args&: C.second); |
| 892 | } |
| 893 | |
| 894 | return !Result.empty(); |
| 895 | } |
| 896 | } |
| 897 | |
| 898 | // If all else fails, see if LVI can figure out a constant value for us. |
| 899 | assert(CxtI->getParent() == BB && "CxtI should be in BB" ); |
| 900 | Constant *CI = LVI->getConstant(V, CxtI); |
| 901 | if (Constant *KC = getKnownConstant(Val: CI, Preference)) { |
| 902 | for (BasicBlock *Pred : predecessors(BB)) |
| 903 | Result.emplace_back(Args&: KC, Args&: Pred); |
| 904 | } |
| 905 | |
| 906 | return !Result.empty(); |
| 907 | } |
| 908 | |
| 909 | /// GetBestDestForBranchOnUndef - If we determine that the specified block ends |
| 910 | /// in an undefined jump, decide which block is best to revector to. |
| 911 | /// |
| 912 | /// Since we can pick an arbitrary destination, we pick the successor with the |
| 913 | /// fewest predecessors. This should reduce the in-degree of the others. |
| 914 | static unsigned getBestDestForJumpOnUndef(BasicBlock *BB) { |
| 915 | Instruction *BBTerm = BB->getTerminator(); |
| 916 | unsigned MinSucc = 0; |
| 917 | BasicBlock *TestBB = BBTerm->getSuccessor(Idx: MinSucc); |
| 918 | // Compute the successor with the minimum number of predecessors. |
| 919 | unsigned MinNumPreds = pred_size(BB: TestBB); |
| 920 | for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { |
| 921 | TestBB = BBTerm->getSuccessor(Idx: i); |
| 922 | unsigned NumPreds = pred_size(BB: TestBB); |
| 923 | if (NumPreds < MinNumPreds) { |
| 924 | MinSucc = i; |
| 925 | MinNumPreds = NumPreds; |
| 926 | } |
| 927 | } |
| 928 | |
| 929 | return MinSucc; |
| 930 | } |
| 931 | |
| 932 | static bool hasAddressTakenAndUsed(BasicBlock *BB) { |
| 933 | if (!BB->hasAddressTaken()) return false; |
| 934 | |
| 935 | // If the block has its address taken, it may be a tree of dead constants |
| 936 | // hanging off of it. These shouldn't keep the block alive. |
| 937 | BlockAddress *BA = BlockAddress::get(BB); |
| 938 | BA->removeDeadConstantUsers(); |
| 939 | return !BA->use_empty(); |
| 940 | } |
| 941 | |
| 942 | /// processBlock - If there are any predecessors whose control can be threaded |
| 943 | /// through to a successor, transform them now. |
| 944 | bool JumpThreadingPass::processBlock(BasicBlock *BB) { |
| 945 | // If the block is trivially dead, just return and let the caller nuke it. |
| 946 | // This simplifies other transformations. |
| 947 | if (DTU->isBBPendingDeletion(DelBB: BB) || |
| 948 | (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock())) |
| 949 | return false; |
| 950 | |
| 951 | // If this block has a single predecessor, and if that pred has a single |
| 952 | // successor, merge the blocks. This encourages recursive jump threading |
| 953 | // because now the condition in this block can be threaded through |
| 954 | // predecessors of our predecessor block. |
| 955 | if (maybeMergeBasicBlockIntoOnlyPred(BB)) |
| 956 | return true; |
| 957 | |
| 958 | if (tryToUnfoldSelectInCurrBB(BB)) |
| 959 | return true; |
| 960 | |
| 961 | // Look if we can propagate guards to predecessors. |
| 962 | if (HasGuards && processGuards(BB)) |
| 963 | return true; |
| 964 | |
| 965 | // What kind of constant we're looking for. |
| 966 | ConstantPreference Preference = WantInteger; |
| 967 | |
| 968 | // Look to see if the terminator is a conditional branch, switch or indirect |
| 969 | // branch, if not we can't thread it. |
| 970 | Value *Condition; |
| 971 | Instruction *Terminator = BB->getTerminator(); |
| 972 | if (CondBrInst *BI = dyn_cast<CondBrInst>(Val: Terminator)) { |
| 973 | Condition = BI->getCondition(); |
| 974 | } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: Terminator)) { |
| 975 | Condition = SI->getCondition(); |
| 976 | } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Val: Terminator)) { |
| 977 | // Can't thread indirect branch with no successors. |
| 978 | if (IB->getNumSuccessors() == 0) return false; |
| 979 | Condition = IB->getAddress()->stripPointerCasts(); |
| 980 | Preference = WantBlockAddress; |
| 981 | } else { |
| 982 | return false; // Must be an invoke or callbr. |
| 983 | } |
| 984 | |
| 985 | // Keep track if we constant folded the condition in this invocation. |
| 986 | bool ConstantFolded = false; |
| 987 | |
| 988 | // Run constant folding to see if we can reduce the condition to a simple |
| 989 | // constant. |
| 990 | if (Instruction *I = dyn_cast<Instruction>(Val: Condition)) { |
| 991 | Value *SimpleVal = |
| 992 | ConstantFoldInstruction(I, DL: BB->getDataLayout(), TLI); |
| 993 | if (SimpleVal) { |
| 994 | I->replaceAllUsesWith(V: SimpleVal); |
| 995 | if (isInstructionTriviallyDead(I, TLI)) |
| 996 | I->eraseFromParent(); |
| 997 | Condition = SimpleVal; |
| 998 | ConstantFolded = true; |
| 999 | } |
| 1000 | } |
| 1001 | |
| 1002 | // If the terminator is branching on an undef or freeze undef, we can pick any |
| 1003 | // of the successors to branch to. Let getBestDestForJumpOnUndef decide. |
| 1004 | auto *FI = dyn_cast<FreezeInst>(Val: Condition); |
| 1005 | if (isa<UndefValue>(Val: Condition) || |
| 1006 | (FI && isa<UndefValue>(Val: FI->getOperand(i_nocapture: 0)) && FI->hasOneUse())) { |
| 1007 | unsigned BestSucc = getBestDestForJumpOnUndef(BB); |
| 1008 | std::vector<DominatorTree::UpdateType> Updates; |
| 1009 | |
| 1010 | // Fold the branch/switch. |
| 1011 | Instruction *BBTerm = BB->getTerminator(); |
| 1012 | Updates.reserve(n: BBTerm->getNumSuccessors()); |
| 1013 | for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { |
| 1014 | if (i == BestSucc) continue; |
| 1015 | BasicBlock *Succ = BBTerm->getSuccessor(Idx: i); |
| 1016 | Succ->removePredecessor(Pred: BB, KeepOneInputPHIs: true); |
| 1017 | Updates.push_back(x: {DominatorTree::Delete, BB, Succ}); |
| 1018 | } |
| 1019 | |
| 1020 | LLVM_DEBUG(dbgs() << " In block '" << BB->getName() |
| 1021 | << "' folding undef terminator: " << *BBTerm << '\n'); |
| 1022 | Instruction *NewBI = UncondBrInst::Create(IfTrue: BBTerm->getSuccessor(Idx: BestSucc), |
| 1023 | InsertBefore: BBTerm->getIterator()); |
| 1024 | NewBI->setDebugLoc(BBTerm->getDebugLoc()); |
| 1025 | ++NumFolds; |
| 1026 | BBTerm->eraseFromParent(); |
| 1027 | DTU->applyUpdatesPermissive(Updates); |
| 1028 | if (FI) |
| 1029 | FI->eraseFromParent(); |
| 1030 | return true; |
| 1031 | } |
| 1032 | |
| 1033 | // If the terminator of this block is branching on a constant, simplify the |
| 1034 | // terminator to an unconditional branch. This can occur due to threading in |
| 1035 | // other blocks. |
| 1036 | if (getKnownConstant(Val: Condition, Preference)) { |
| 1037 | LLVM_DEBUG(dbgs() << " In block '" << BB->getName() |
| 1038 | << "' folding terminator: " << *BB->getTerminator() |
| 1039 | << '\n'); |
| 1040 | ++NumFolds; |
| 1041 | ConstantFoldTerminator(BB, DeleteDeadConditions: true, TLI: nullptr, DTU: DTU.get()); |
| 1042 | if (auto *BPI = getBPI()) |
| 1043 | BPI->eraseBlock(BB); |
| 1044 | return true; |
| 1045 | } |
| 1046 | |
| 1047 | Instruction *CondInst = dyn_cast<Instruction>(Val: Condition); |
| 1048 | |
| 1049 | // All the rest of our checks depend on the condition being an instruction. |
| 1050 | if (!CondInst) { |
| 1051 | // FIXME: Unify this with code below. |
| 1052 | if (processThreadableEdges(Cond: Condition, BB, Preference, CxtI: Terminator)) |
| 1053 | return true; |
| 1054 | return ConstantFolded; |
| 1055 | } |
| 1056 | |
| 1057 | // Some of the following optimization can safely work on the unfrozen cond. |
| 1058 | Value *CondWithoutFreeze = CondInst; |
| 1059 | if (auto *FI = dyn_cast<FreezeInst>(Val: CondInst)) |
| 1060 | CondWithoutFreeze = FI->getOperand(i_nocapture: 0); |
| 1061 | |
| 1062 | if (CmpInst *CondCmp = dyn_cast<CmpInst>(Val: CondWithoutFreeze)) { |
| 1063 | // If we're branching on a conditional, LVI might be able to determine |
| 1064 | // it's value at the branch instruction. We only handle comparisons |
| 1065 | // against a constant at this time. |
| 1066 | if (Constant *CondConst = dyn_cast<Constant>(Val: CondCmp->getOperand(i_nocapture: 1))) { |
| 1067 | Constant *Res = |
| 1068 | LVI->getPredicateAt(Pred: CondCmp->getPredicate(), V: CondCmp->getOperand(i_nocapture: 0), |
| 1069 | C: CondConst, CxtI: BB->getTerminator(), |
| 1070 | /*UseBlockValue=*/false); |
| 1071 | if (Res) { |
| 1072 | // We can safely replace *some* uses of the CondInst if it has |
| 1073 | // exactly one value as returned by LVI. RAUW is incorrect in the |
| 1074 | // presence of guards and assumes, that have the `Cond` as the use. This |
| 1075 | // is because we use the guards/assume to reason about the `Cond` value |
| 1076 | // at the end of block, but RAUW unconditionally replaces all uses |
| 1077 | // including the guards/assumes themselves and the uses before the |
| 1078 | // guard/assume. |
| 1079 | if (replaceFoldableUses(Cond: CondCmp, ToVal: Res, KnownAtEndOfBB: BB)) |
| 1080 | return true; |
| 1081 | } |
| 1082 | |
| 1083 | // We did not manage to simplify this branch, try to see whether |
| 1084 | // CondCmp depends on a known phi-select pattern. |
| 1085 | if (tryToUnfoldSelect(CondCmp, BB)) |
| 1086 | return true; |
| 1087 | } |
| 1088 | } |
| 1089 | |
| 1090 | if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: BB->getTerminator())) |
| 1091 | if (tryToUnfoldSelect(SI, BB)) |
| 1092 | return true; |
| 1093 | |
| 1094 | // Check for some cases that are worth simplifying. Right now we want to look |
| 1095 | // for loads that are used by a switch or by the condition for the branch. If |
| 1096 | // we see one, check to see if it's partially redundant. If so, insert a PHI |
| 1097 | // which can then be used to thread the values. |
| 1098 | Value *SimplifyValue = CondWithoutFreeze; |
| 1099 | |
| 1100 | if (CmpInst *CondCmp = dyn_cast<CmpInst>(Val: SimplifyValue)) |
| 1101 | if (isa<Constant>(Val: CondCmp->getOperand(i_nocapture: 1))) |
| 1102 | SimplifyValue = CondCmp->getOperand(i_nocapture: 0); |
| 1103 | |
| 1104 | // TODO: There are other places where load PRE would be profitable, such as |
| 1105 | // more complex comparisons. |
| 1106 | if (LoadInst *LoadI = dyn_cast<LoadInst>(Val: SimplifyValue)) |
| 1107 | if (simplifyPartiallyRedundantLoad(LI: LoadI)) |
| 1108 | return true; |
| 1109 | |
| 1110 | // Before threading, try to propagate profile data backwards: |
| 1111 | if (PHINode *PN = dyn_cast<PHINode>(Val: CondInst)) |
| 1112 | if (PN->getParent() == BB && isa<CondBrInst>(Val: BB->getTerminator())) |
| 1113 | updatePredecessorProfileMetadata(PN, BB); |
| 1114 | |
| 1115 | // Handle a variety of cases where we are branching on something derived from |
| 1116 | // a PHI node in the current block. If we can prove that any predecessors |
| 1117 | // compute a predictable value based on a PHI node, thread those predecessors. |
| 1118 | if (processThreadableEdges(Cond: CondInst, BB, Preference, CxtI: Terminator)) |
| 1119 | return true; |
| 1120 | |
| 1121 | // If this is an otherwise-unfoldable branch on a phi node or freeze(phi) in |
| 1122 | // the current block, see if we can simplify. |
| 1123 | PHINode *PN = dyn_cast<PHINode>(Val: CondWithoutFreeze); |
| 1124 | if (PN && PN->getParent() == BB && isa<CondBrInst>(Val: BB->getTerminator())) |
| 1125 | return processBranchOnPHI(PN); |
| 1126 | |
| 1127 | // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. |
| 1128 | if (CondInst->getOpcode() == Instruction::Xor && |
| 1129 | CondInst->getParent() == BB && isa<CondBrInst>(Val: BB->getTerminator())) |
| 1130 | return processBranchOnXOR(BO: cast<BinaryOperator>(Val: CondInst)); |
| 1131 | |
| 1132 | // Search for a stronger dominating condition that can be used to simplify a |
| 1133 | // conditional branch leaving BB. |
| 1134 | if (processImpliedCondition(BB)) |
| 1135 | return true; |
| 1136 | |
| 1137 | return false; |
| 1138 | } |
| 1139 | |
| 1140 | bool JumpThreadingPass::processImpliedCondition(BasicBlock *BB) { |
| 1141 | auto *BI = dyn_cast<CondBrInst>(Val: BB->getTerminator()); |
| 1142 | if (!BI) |
| 1143 | return false; |
| 1144 | |
| 1145 | Value *Cond = BI->getCondition(); |
| 1146 | // Assuming that predecessor's branch was taken, if pred's branch condition |
| 1147 | // (V) implies Cond, Cond can be either true, undef, or poison. In this case, |
| 1148 | // freeze(Cond) is either true or a nondeterministic value. |
| 1149 | // If freeze(Cond) has only one use, we can freely fold freeze(Cond) to true |
| 1150 | // without affecting other instructions. |
| 1151 | auto *FICond = dyn_cast<FreezeInst>(Val: Cond); |
| 1152 | if (FICond && FICond->hasOneUse()) |
| 1153 | Cond = FICond->getOperand(i_nocapture: 0); |
| 1154 | else |
| 1155 | FICond = nullptr; |
| 1156 | |
| 1157 | BasicBlock *CurrentBB = BB; |
| 1158 | BasicBlock *CurrentPred = BB->getSinglePredecessor(); |
| 1159 | unsigned Iter = 0; |
| 1160 | |
| 1161 | auto &DL = BB->getDataLayout(); |
| 1162 | |
| 1163 | while (CurrentPred && Iter++ < ImplicationSearchThreshold) { |
| 1164 | auto *PBI = dyn_cast<CondBrInst>(Val: CurrentPred->getTerminator()); |
| 1165 | if (!PBI) |
| 1166 | return false; |
| 1167 | if (PBI->getSuccessor(i: 0) != CurrentBB && PBI->getSuccessor(i: 1) != CurrentBB) |
| 1168 | return false; |
| 1169 | |
| 1170 | bool CondIsTrue = PBI->getSuccessor(i: 0) == CurrentBB; |
| 1171 | std::optional<bool> Implication = |
| 1172 | isImpliedCondition(LHS: PBI->getCondition(), RHS: Cond, DL, LHSIsTrue: CondIsTrue); |
| 1173 | |
| 1174 | // If the branch condition of BB (which is Cond) and CurrentPred are |
| 1175 | // exactly the same freeze instruction, Cond can be folded into CondIsTrue. |
| 1176 | if (!Implication && FICond && isa<FreezeInst>(Val: PBI->getCondition())) { |
| 1177 | if (cast<FreezeInst>(Val: PBI->getCondition())->getOperand(i_nocapture: 0) == |
| 1178 | FICond->getOperand(i_nocapture: 0)) |
| 1179 | Implication = CondIsTrue; |
| 1180 | } |
| 1181 | |
| 1182 | if (Implication) { |
| 1183 | BasicBlock *KeepSucc = BI->getSuccessor(i: *Implication ? 0 : 1); |
| 1184 | BasicBlock *RemoveSucc = BI->getSuccessor(i: *Implication ? 1 : 0); |
| 1185 | RemoveSucc->removePredecessor(Pred: BB); |
| 1186 | UncondBrInst *UncondBI = |
| 1187 | UncondBrInst::Create(IfTrue: KeepSucc, InsertBefore: BI->getIterator()); |
| 1188 | UncondBI->setDebugLoc(BI->getDebugLoc()); |
| 1189 | ++NumFolds; |
| 1190 | BI->eraseFromParent(); |
| 1191 | if (FICond) |
| 1192 | FICond->eraseFromParent(); |
| 1193 | |
| 1194 | DTU->applyUpdatesPermissive(Updates: {{DominatorTree::Delete, BB, RemoveSucc}}); |
| 1195 | if (auto *BPI = getBPI()) |
| 1196 | BPI->eraseBlock(BB); |
| 1197 | return true; |
| 1198 | } |
| 1199 | CurrentBB = CurrentPred; |
| 1200 | CurrentPred = CurrentBB->getSinglePredecessor(); |
| 1201 | } |
| 1202 | |
| 1203 | return false; |
| 1204 | } |
| 1205 | |
| 1206 | /// Return true if Op is an instruction defined in the given block. |
| 1207 | static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) { |
| 1208 | if (Instruction *OpInst = dyn_cast<Instruction>(Val: Op)) |
| 1209 | if (OpInst->getParent() == BB) |
| 1210 | return true; |
| 1211 | return false; |
| 1212 | } |
| 1213 | |
| 1214 | /// simplifyPartiallyRedundantLoad - If LoadI is an obviously partially |
| 1215 | /// redundant load instruction, eliminate it by replacing it with a PHI node. |
| 1216 | /// This is an important optimization that encourages jump threading, and needs |
| 1217 | /// to be run interlaced with other jump threading tasks. |
| 1218 | bool JumpThreadingPass::simplifyPartiallyRedundantLoad(LoadInst *LoadI) { |
| 1219 | // Don't hack volatile and ordered loads. |
| 1220 | if (!LoadI->isUnordered()) return false; |
| 1221 | |
| 1222 | // If the load is defined in a block with exactly one predecessor, it can't be |
| 1223 | // partially redundant. |
| 1224 | BasicBlock *LoadBB = LoadI->getParent(); |
| 1225 | if (LoadBB->getSinglePredecessor()) |
| 1226 | return false; |
| 1227 | |
| 1228 | // If the load is defined in an EH pad, it can't be partially redundant, |
| 1229 | // because the edges between the invoke and the EH pad cannot have other |
| 1230 | // instructions between them. |
| 1231 | if (LoadBB->isEHPad()) |
| 1232 | return false; |
| 1233 | |
| 1234 | Value *LoadedPtr = LoadI->getOperand(i_nocapture: 0); |
| 1235 | |
| 1236 | // If the loaded operand is defined in the LoadBB and its not a phi, |
| 1237 | // it can't be available in predecessors. |
| 1238 | if (isOpDefinedInBlock(Op: LoadedPtr, BB: LoadBB) && !isa<PHINode>(Val: LoadedPtr)) |
| 1239 | return false; |
| 1240 | |
| 1241 | // Scan a few instructions up from the load, to see if it is obviously live at |
| 1242 | // the entry to its block. |
| 1243 | BasicBlock::iterator BBIt(LoadI); |
| 1244 | bool IsLoadCSE; |
| 1245 | BatchAAResults BatchAA(*AA); |
| 1246 | // The dominator tree is updated lazily and may not be valid at this point. |
| 1247 | BatchAA.disableDominatorTree(); |
| 1248 | if (Value *AvailableVal = FindAvailableLoadedValue( |
| 1249 | Load: LoadI, ScanBB: LoadBB, ScanFrom&: BBIt, MaxInstsToScan: DefMaxInstsToScan, AA: &BatchAA, IsLoadCSE: &IsLoadCSE)) { |
| 1250 | // If the value of the load is locally available within the block, just use |
| 1251 | // it. This frequently occurs for reg2mem'd allocas. |
| 1252 | |
| 1253 | if (IsLoadCSE) { |
| 1254 | LoadInst *NLoadI = cast<LoadInst>(Val: AvailableVal); |
| 1255 | combineMetadataForCSE(K: NLoadI, J: LoadI, DoesKMove: false); |
| 1256 | LVI->forgetValue(V: NLoadI); |
| 1257 | }; |
| 1258 | |
| 1259 | // If the returned value is the load itself, replace with poison. This can |
| 1260 | // only happen in dead loops. |
| 1261 | if (AvailableVal == LoadI) |
| 1262 | AvailableVal = PoisonValue::get(T: LoadI->getType()); |
| 1263 | if (AvailableVal->getType() != LoadI->getType()) { |
| 1264 | AvailableVal = CastInst::CreateBitOrPointerCast( |
| 1265 | S: AvailableVal, Ty: LoadI->getType(), Name: "" , InsertBefore: LoadI->getIterator()); |
| 1266 | cast<Instruction>(Val: AvailableVal)->setDebugLoc(LoadI->getDebugLoc()); |
| 1267 | } |
| 1268 | LoadI->replaceAllUsesWith(V: AvailableVal); |
| 1269 | LoadI->eraseFromParent(); |
| 1270 | return true; |
| 1271 | } |
| 1272 | |
| 1273 | // Otherwise, if we scanned the whole block and got to the top of the block, |
| 1274 | // we know the block is locally transparent to the load. If not, something |
| 1275 | // might clobber its value. |
| 1276 | if (BBIt != LoadBB->begin()) |
| 1277 | return false; |
| 1278 | |
| 1279 | // If all of the loads and stores that feed the value have the same AA tags, |
| 1280 | // then we can propagate them onto any newly inserted loads. |
| 1281 | AAMDNodes AATags = LoadI->getAAMetadata(); |
| 1282 | |
| 1283 | SmallPtrSet<BasicBlock*, 8> PredsScanned; |
| 1284 | |
| 1285 | using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>; |
| 1286 | |
| 1287 | AvailablePredsTy AvailablePreds; |
| 1288 | BasicBlock *OneUnavailablePred = nullptr; |
| 1289 | SmallVector<LoadInst*, 8> CSELoads; |
| 1290 | |
| 1291 | // If we got here, the loaded value is transparent through to the start of the |
| 1292 | // block. Check to see if it is available in any of the predecessor blocks. |
| 1293 | for (BasicBlock *PredBB : predecessors(BB: LoadBB)) { |
| 1294 | // If we already scanned this predecessor, skip it. |
| 1295 | if (!PredsScanned.insert(Ptr: PredBB).second) |
| 1296 | continue; |
| 1297 | |
| 1298 | BBIt = PredBB->end(); |
| 1299 | unsigned NumScanedInst = 0; |
| 1300 | Value *PredAvailable = nullptr; |
| 1301 | // NOTE: We don't CSE load that is volatile or anything stronger than |
| 1302 | // unordered, that should have been checked when we entered the function. |
| 1303 | assert(LoadI->isUnordered() && |
| 1304 | "Attempting to CSE volatile or atomic loads" ); |
| 1305 | // If this is a load on a phi pointer, phi-translate it and search |
| 1306 | // for available load/store to the pointer in predecessors. |
| 1307 | Type *AccessTy = LoadI->getType(); |
| 1308 | const auto &DL = LoadI->getDataLayout(); |
| 1309 | MemoryLocation Loc(LoadedPtr->DoPHITranslation(CurBB: LoadBB, PredBB), |
| 1310 | LocationSize::precise(Value: DL.getTypeStoreSize(Ty: AccessTy)), |
| 1311 | AATags); |
| 1312 | PredAvailable = findAvailablePtrLoadStore( |
| 1313 | Loc, AccessTy, AtLeastAtomic: LoadI->isAtomic(), ScanBB: PredBB, ScanFrom&: BBIt, MaxInstsToScan: DefMaxInstsToScan, |
| 1314 | AA: &BatchAA, IsLoadCSE: &IsLoadCSE, NumScanedInst: &NumScanedInst); |
| 1315 | |
| 1316 | // If PredBB has a single predecessor, continue scanning through the |
| 1317 | // single predecessor. |
| 1318 | BasicBlock *SinglePredBB = PredBB; |
| 1319 | while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() && |
| 1320 | NumScanedInst < DefMaxInstsToScan) { |
| 1321 | SinglePredBB = SinglePredBB->getSinglePredecessor(); |
| 1322 | if (SinglePredBB) { |
| 1323 | BBIt = SinglePredBB->end(); |
| 1324 | PredAvailable = findAvailablePtrLoadStore( |
| 1325 | Loc, AccessTy, AtLeastAtomic: LoadI->isAtomic(), ScanBB: SinglePredBB, ScanFrom&: BBIt, |
| 1326 | MaxInstsToScan: (DefMaxInstsToScan - NumScanedInst), AA: &BatchAA, IsLoadCSE: &IsLoadCSE, |
| 1327 | NumScanedInst: &NumScanedInst); |
| 1328 | } |
| 1329 | } |
| 1330 | |
| 1331 | if (!PredAvailable) { |
| 1332 | OneUnavailablePred = PredBB; |
| 1333 | continue; |
| 1334 | } |
| 1335 | |
| 1336 | if (IsLoadCSE) |
| 1337 | CSELoads.push_back(Elt: cast<LoadInst>(Val: PredAvailable)); |
| 1338 | |
| 1339 | // If so, this load is partially redundant. Remember this info so that we |
| 1340 | // can create a PHI node. |
| 1341 | AvailablePreds.emplace_back(Args&: PredBB, Args&: PredAvailable); |
| 1342 | } |
| 1343 | |
| 1344 | // If the loaded value isn't available in any predecessor, it isn't partially |
| 1345 | // redundant. |
| 1346 | if (AvailablePreds.empty()) return false; |
| 1347 | |
| 1348 | // Okay, the loaded value is available in at least one (and maybe all!) |
| 1349 | // predecessors. If the value is unavailable in more than one unique |
| 1350 | // predecessor, we want to insert a merge block for those common predecessors. |
| 1351 | // This ensures that we only have to insert one reload, thus not increasing |
| 1352 | // code size. |
| 1353 | BasicBlock *UnavailablePred = nullptr; |
| 1354 | |
| 1355 | // If the value is unavailable in one of predecessors, we will end up |
| 1356 | // inserting a new instruction into them. It is only valid if all the |
| 1357 | // instructions before LoadI are guaranteed to pass execution to its |
| 1358 | // successor, or if LoadI is safe to speculate. |
| 1359 | // TODO: If this logic becomes more complex, and we will perform PRE insertion |
| 1360 | // farther than to a predecessor, we need to reuse the code from GVN's PRE. |
| 1361 | // It requires domination tree analysis, so for this simple case it is an |
| 1362 | // overkill. |
| 1363 | if (PredsScanned.size() != AvailablePreds.size() && |
| 1364 | !isSafeToSpeculativelyExecute(I: LoadI)) |
| 1365 | for (auto I = LoadBB->begin(); &*I != LoadI; ++I) |
| 1366 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &*I)) |
| 1367 | return false; |
| 1368 | |
| 1369 | // If there is exactly one predecessor where the value is unavailable, the |
| 1370 | // already computed 'OneUnavailablePred' block is it. If it ends in an |
| 1371 | // unconditional branch, we know that it isn't a critical edge. |
| 1372 | if (PredsScanned.size() == AvailablePreds.size()+1 && |
| 1373 | OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { |
| 1374 | UnavailablePred = OneUnavailablePred; |
| 1375 | } else if (PredsScanned.size() != AvailablePreds.size()) { |
| 1376 | // Otherwise, we had multiple unavailable predecessors or we had a critical |
| 1377 | // edge from the one. |
| 1378 | SmallVector<BasicBlock*, 8> PredsToSplit; |
| 1379 | SmallPtrSet<BasicBlock *, 8> AvailablePredSet( |
| 1380 | llvm::from_range, llvm::make_first_range(c&: AvailablePreds)); |
| 1381 | |
| 1382 | // Add all the unavailable predecessors to the PredsToSplit list. |
| 1383 | for (BasicBlock *P : predecessors(BB: LoadBB)) { |
| 1384 | // If the predecessor is an indirect goto, we can't split the edge. |
| 1385 | if (isa<IndirectBrInst>(Val: P->getTerminator())) |
| 1386 | return false; |
| 1387 | |
| 1388 | if (!AvailablePredSet.count(Ptr: P)) |
| 1389 | PredsToSplit.push_back(Elt: P); |
| 1390 | } |
| 1391 | |
| 1392 | // Split them out to their own block. |
| 1393 | UnavailablePred = splitBlockPreds(BB: LoadBB, Preds: PredsToSplit, Suffix: "thread-pre-split" ); |
| 1394 | } |
| 1395 | |
| 1396 | // If the value isn't available in all predecessors, then there will be |
| 1397 | // exactly one where it isn't available. Insert a load on that edge and add |
| 1398 | // it to the AvailablePreds list. |
| 1399 | if (UnavailablePred) { |
| 1400 | assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && |
| 1401 | "Can't handle critical edge here!" ); |
| 1402 | LoadInst *NewVal = new LoadInst( |
| 1403 | LoadI->getType(), LoadedPtr->DoPHITranslation(CurBB: LoadBB, PredBB: UnavailablePred), |
| 1404 | LoadI->getName() + ".pr" , false, LoadI->getAlign(), |
| 1405 | LoadI->getOrdering(), LoadI->getSyncScopeID(), |
| 1406 | UnavailablePred->getTerminator()->getIterator()); |
| 1407 | NewVal->setDebugLoc(LoadI->getDebugLoc()); |
| 1408 | if (AATags) |
| 1409 | NewVal->setAAMetadata(AATags); |
| 1410 | |
| 1411 | AvailablePreds.emplace_back(Args&: UnavailablePred, Args&: NewVal); |
| 1412 | } |
| 1413 | |
| 1414 | // Now we know that each predecessor of this block has a value in |
| 1415 | // AvailablePreds, sort them for efficient access as we're walking the preds. |
| 1416 | array_pod_sort(Start: AvailablePreds.begin(), End: AvailablePreds.end()); |
| 1417 | |
| 1418 | // Create a PHI node at the start of the block for the PRE'd load value. |
| 1419 | PHINode *PN = PHINode::Create(Ty: LoadI->getType(), NumReservedValues: pred_size(BB: LoadBB), NameStr: "" ); |
| 1420 | PN->insertBefore(InsertPos: LoadBB->begin()); |
| 1421 | PN->takeName(V: LoadI); |
| 1422 | PN->setDebugLoc(LoadI->getDebugLoc()); |
| 1423 | |
| 1424 | // Insert new entries into the PHI for each predecessor. A single block may |
| 1425 | // have multiple entries here. |
| 1426 | for (BasicBlock *P : predecessors(BB: LoadBB)) { |
| 1427 | AvailablePredsTy::iterator I = |
| 1428 | llvm::lower_bound(Range&: AvailablePreds, Value: std::make_pair(x&: P, y: (Value *)nullptr)); |
| 1429 | |
| 1430 | assert(I != AvailablePreds.end() && I->first == P && |
| 1431 | "Didn't find entry for predecessor!" ); |
| 1432 | |
| 1433 | // If we have an available predecessor but it requires casting, insert the |
| 1434 | // cast in the predecessor and use the cast. Note that we have to update the |
| 1435 | // AvailablePreds vector as we go so that all of the PHI entries for this |
| 1436 | // predecessor use the same bitcast. |
| 1437 | Value *&PredV = I->second; |
| 1438 | if (PredV->getType() != LoadI->getType()) { |
| 1439 | PredV = CastInst::CreateBitOrPointerCast( |
| 1440 | S: PredV, Ty: LoadI->getType(), Name: "" , InsertBefore: P->getTerminator()->getIterator()); |
| 1441 | // The new cast is producing the value used to replace the load |
| 1442 | // instruction, so uses the load's debug location. If P does not always |
| 1443 | // branch to the load BB however then the debug location must be dropped, |
| 1444 | // as it is hoisted past a conditional branch. |
| 1445 | DebugLoc DL = P->getTerminator()->getNumSuccessors() == 1 |
| 1446 | ? LoadI->getDebugLoc() |
| 1447 | : DebugLoc::getDropped(); |
| 1448 | cast<CastInst>(Val: PredV)->setDebugLoc(DL); |
| 1449 | } |
| 1450 | |
| 1451 | PN->addIncoming(V: PredV, BB: I->first); |
| 1452 | } |
| 1453 | |
| 1454 | for (LoadInst *PredLoadI : CSELoads) { |
| 1455 | combineMetadataForCSE(K: PredLoadI, J: LoadI, DoesKMove: true); |
| 1456 | LVI->forgetValue(V: PredLoadI); |
| 1457 | } |
| 1458 | |
| 1459 | LoadI->replaceAllUsesWith(V: PN); |
| 1460 | LoadI->eraseFromParent(); |
| 1461 | |
| 1462 | return true; |
| 1463 | } |
| 1464 | |
| 1465 | /// findMostPopularDest - The specified list contains multiple possible |
| 1466 | /// threadable destinations. Pick the one that occurs the most frequently in |
| 1467 | /// the list. |
| 1468 | static BasicBlock * |
| 1469 | findMostPopularDest(BasicBlock *BB, |
| 1470 | const SmallVectorImpl<std::pair<BasicBlock *, |
| 1471 | BasicBlock *>> &PredToDestList) { |
| 1472 | assert(!PredToDestList.empty()); |
| 1473 | |
| 1474 | // Determine popularity. If there are multiple possible destinations, we |
| 1475 | // explicitly choose to ignore 'undef' destinations. We prefer to thread |
| 1476 | // blocks with known and real destinations to threading undef. We'll handle |
| 1477 | // them later if interesting. |
| 1478 | MapVector<BasicBlock *, unsigned> DestPopularity; |
| 1479 | |
| 1480 | // Populate DestPopularity with the successors in the order they appear in the |
| 1481 | // successor list. This way, we ensure determinism by iterating it in the |
| 1482 | // same order in llvm::max_element below. We map nullptr to 0 so that we can |
| 1483 | // return nullptr when PredToDestList contains nullptr only. |
| 1484 | DestPopularity[nullptr] = 0; |
| 1485 | for (auto *SuccBB : successors(BB)) |
| 1486 | DestPopularity[SuccBB] = 0; |
| 1487 | |
| 1488 | for (const auto &PredToDest : PredToDestList) |
| 1489 | if (PredToDest.second) |
| 1490 | DestPopularity[PredToDest.second]++; |
| 1491 | |
| 1492 | // Find the most popular dest. |
| 1493 | auto MostPopular = llvm::max_element(Range&: DestPopularity, C: llvm::less_second()); |
| 1494 | |
| 1495 | // Okay, we have finally picked the most popular destination. |
| 1496 | return MostPopular->first; |
| 1497 | } |
| 1498 | |
| 1499 | // Try to evaluate the value of V when the control flows from PredPredBB to |
| 1500 | // BB->getSinglePredecessor() and then on to BB. |
| 1501 | Constant *JumpThreadingPass::evaluateOnPredecessorEdge(BasicBlock *BB, |
| 1502 | BasicBlock *PredPredBB, |
| 1503 | Value *V, |
| 1504 | const DataLayout &DL) { |
| 1505 | SmallPtrSet<Value *, 8> Visited; |
| 1506 | return evaluateOnPredecessorEdge(BB, PredPredBB, cond: V, DL, Visited); |
| 1507 | } |
| 1508 | |
| 1509 | Constant *JumpThreadingPass::evaluateOnPredecessorEdge( |
| 1510 | BasicBlock *BB, BasicBlock *PredPredBB, Value *V, const DataLayout &DL, |
| 1511 | SmallPtrSet<Value *, 8> &Visited) { |
| 1512 | if (!Visited.insert(Ptr: V).second) |
| 1513 | return nullptr; |
| 1514 | llvm::scope_exit _([&Visited, V]() { Visited.erase(Ptr: V); }); |
| 1515 | |
| 1516 | BasicBlock *PredBB = BB->getSinglePredecessor(); |
| 1517 | assert(PredBB && "Expected a single predecessor" ); |
| 1518 | |
| 1519 | if (Constant *Cst = dyn_cast<Constant>(Val: V)) { |
| 1520 | return Cst; |
| 1521 | } |
| 1522 | |
| 1523 | // Consult LVI if V is not an instruction in BB or PredBB. |
| 1524 | Instruction *I = dyn_cast<Instruction>(Val: V); |
| 1525 | if (!I || (I->getParent() != BB && I->getParent() != PredBB)) { |
| 1526 | return LVI->getConstantOnEdge(V, FromBB: PredPredBB, ToBB: PredBB, CxtI: nullptr); |
| 1527 | } |
| 1528 | |
| 1529 | // Look into a PHI argument. |
| 1530 | if (PHINode *PHI = dyn_cast<PHINode>(Val: V)) { |
| 1531 | if (PHI->getParent() == PredBB) |
| 1532 | return dyn_cast<Constant>(Val: PHI->getIncomingValueForBlock(BB: PredPredBB)); |
| 1533 | return nullptr; |
| 1534 | } |
| 1535 | |
| 1536 | // If we have a CmpInst, try to fold it for each incoming edge into PredBB. |
| 1537 | // Note that during the execution of the pass, phi nodes may become constant |
| 1538 | // and may be removed, which can lead to self-referencing instructions in |
| 1539 | // code that becomes unreachable. Consequently, we need to handle those |
| 1540 | // instructions in unreachable code and check before going into recursion. |
| 1541 | if (CmpInst *CondCmp = dyn_cast<CmpInst>(Val: V)) { |
| 1542 | if (CondCmp->getParent() == BB) { |
| 1543 | Constant *Op0 = evaluateOnPredecessorEdge( |
| 1544 | BB, PredPredBB, V: CondCmp->getOperand(i_nocapture: 0), DL, Visited); |
| 1545 | Constant *Op1 = evaluateOnPredecessorEdge( |
| 1546 | BB, PredPredBB, V: CondCmp->getOperand(i_nocapture: 1), DL, Visited); |
| 1547 | if (Op0 && Op1) { |
| 1548 | return ConstantFoldCompareInstOperands(Predicate: CondCmp->getPredicate(), LHS: Op0, |
| 1549 | RHS: Op1, DL); |
| 1550 | } |
| 1551 | } |
| 1552 | return nullptr; |
| 1553 | } |
| 1554 | |
| 1555 | return nullptr; |
| 1556 | } |
| 1557 | |
| 1558 | bool JumpThreadingPass::processThreadableEdges(Value *Cond, BasicBlock *BB, |
| 1559 | ConstantPreference Preference, |
| 1560 | Instruction *CxtI) { |
| 1561 | // If threading this would thread across a loop header, don't even try to |
| 1562 | // thread the edge. |
| 1563 | if (LoopHeaders.count(Ptr: BB)) |
| 1564 | return false; |
| 1565 | |
| 1566 | PredValueInfoTy PredValues; |
| 1567 | if (!computeValueKnownInPredecessors(V: Cond, BB, Result&: PredValues, Preference, |
| 1568 | CxtI)) { |
| 1569 | // We don't have known values in predecessors. See if we can thread through |
| 1570 | // BB and its sole predecessor. |
| 1571 | return maybethreadThroughTwoBasicBlocks(BB, Cond); |
| 1572 | } |
| 1573 | |
| 1574 | assert(!PredValues.empty() && |
| 1575 | "computeValueKnownInPredecessors returned true with no values" ); |
| 1576 | |
| 1577 | LLVM_DEBUG(dbgs() << "IN BB: " << *BB; |
| 1578 | for (const auto &PredValue : PredValues) { |
| 1579 | dbgs() << " BB '" << BB->getName() |
| 1580 | << "': FOUND condition = " << *PredValue.first |
| 1581 | << " for pred '" << PredValue.second->getName() << "'.\n" ; |
| 1582 | }); |
| 1583 | |
| 1584 | // Decide what we want to thread through. Convert our list of known values to |
| 1585 | // a list of known destinations for each pred. This also discards duplicate |
| 1586 | // predecessors and keeps track of the undefined inputs (which are represented |
| 1587 | // as a null dest in the PredToDestList). |
| 1588 | SmallPtrSet<BasicBlock*, 16> SeenPreds; |
| 1589 | SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; |
| 1590 | |
| 1591 | BasicBlock *OnlyDest = nullptr; |
| 1592 | BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; |
| 1593 | Constant *OnlyVal = nullptr; |
| 1594 | Constant *MultipleVal = (Constant *)(intptr_t)~0ULL; |
| 1595 | |
| 1596 | for (const auto &PredValue : PredValues) { |
| 1597 | BasicBlock *Pred = PredValue.second; |
| 1598 | if (!SeenPreds.insert(Ptr: Pred).second) |
| 1599 | continue; // Duplicate predecessor entry. |
| 1600 | |
| 1601 | Constant *Val = PredValue.first; |
| 1602 | |
| 1603 | BasicBlock *DestBB; |
| 1604 | if (isa<UndefValue>(Val)) |
| 1605 | DestBB = nullptr; |
| 1606 | else if (CondBrInst *BI = dyn_cast<CondBrInst>(Val: BB->getTerminator())) { |
| 1607 | assert(isa<ConstantInt>(Val) && "Expecting a constant integer" ); |
| 1608 | DestBB = BI->getSuccessor(i: cast<ConstantInt>(Val)->isZero()); |
| 1609 | } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: BB->getTerminator())) { |
| 1610 | assert(isa<ConstantInt>(Val) && "Expecting a constant integer" ); |
| 1611 | DestBB = SI->findCaseValue(C: cast<ConstantInt>(Val))->getCaseSuccessor(); |
| 1612 | } else { |
| 1613 | assert(isa<IndirectBrInst>(BB->getTerminator()) |
| 1614 | && "Unexpected terminator" ); |
| 1615 | assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress" ); |
| 1616 | DestBB = cast<BlockAddress>(Val)->getBasicBlock(); |
| 1617 | } |
| 1618 | |
| 1619 | // If we have exactly one destination, remember it for efficiency below. |
| 1620 | if (PredToDestList.empty()) { |
| 1621 | OnlyDest = DestBB; |
| 1622 | OnlyVal = Val; |
| 1623 | } else { |
| 1624 | if (OnlyDest != DestBB) |
| 1625 | OnlyDest = MultipleDestSentinel; |
| 1626 | // It possible we have same destination, but different value, e.g. default |
| 1627 | // case in switchinst. |
| 1628 | if (Val != OnlyVal) |
| 1629 | OnlyVal = MultipleVal; |
| 1630 | } |
| 1631 | |
| 1632 | // If the predecessor ends with an indirect goto, we can't change its |
| 1633 | // destination. |
| 1634 | if (isa<IndirectBrInst>(Val: Pred->getTerminator())) |
| 1635 | continue; |
| 1636 | |
| 1637 | PredToDestList.emplace_back(Args&: Pred, Args&: DestBB); |
| 1638 | } |
| 1639 | |
| 1640 | // If all edges were unthreadable, we fail. |
| 1641 | if (PredToDestList.empty()) |
| 1642 | return false; |
| 1643 | |
| 1644 | // If all the predecessors go to a single known successor, we want to fold, |
| 1645 | // not thread. By doing so, we do not need to duplicate the current block and |
| 1646 | // also miss potential opportunities in case we dont/cant duplicate. |
| 1647 | if (OnlyDest && OnlyDest != MultipleDestSentinel) { |
| 1648 | if (BB->hasNPredecessors(N: PredToDestList.size())) { |
| 1649 | bool SeenFirstBranchToOnlyDest = false; |
| 1650 | std::vector <DominatorTree::UpdateType> Updates; |
| 1651 | Updates.reserve(n: BB->getTerminator()->getNumSuccessors() - 1); |
| 1652 | for (BasicBlock *SuccBB : successors(BB)) { |
| 1653 | if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) { |
| 1654 | SeenFirstBranchToOnlyDest = true; // Don't modify the first branch. |
| 1655 | } else { |
| 1656 | SuccBB->removePredecessor(Pred: BB, KeepOneInputPHIs: true); // This is unreachable successor. |
| 1657 | Updates.push_back(x: {DominatorTree::Delete, BB, SuccBB}); |
| 1658 | } |
| 1659 | } |
| 1660 | |
| 1661 | // Finally update the terminator. |
| 1662 | Instruction *Term = BB->getTerminator(); |
| 1663 | Instruction *NewBI = UncondBrInst::Create(IfTrue: OnlyDest, InsertBefore: Term->getIterator()); |
| 1664 | NewBI->setDebugLoc(Term->getDebugLoc()); |
| 1665 | ++NumFolds; |
| 1666 | Term->eraseFromParent(); |
| 1667 | DTU->applyUpdatesPermissive(Updates); |
| 1668 | if (auto *BPI = getBPI()) |
| 1669 | BPI->eraseBlock(BB); |
| 1670 | |
| 1671 | // If the condition is now dead due to the removal of the old terminator, |
| 1672 | // erase it. |
| 1673 | if (auto *CondInst = dyn_cast<Instruction>(Val: Cond)) { |
| 1674 | if (CondInst->use_empty() && !CondInst->mayHaveSideEffects()) |
| 1675 | CondInst->eraseFromParent(); |
| 1676 | // We can safely replace *some* uses of the CondInst if it has |
| 1677 | // exactly one value as returned by LVI. RAUW is incorrect in the |
| 1678 | // presence of guards and assumes, that have the `Cond` as the use. This |
| 1679 | // is because we use the guards/assume to reason about the `Cond` value |
| 1680 | // at the end of block, but RAUW unconditionally replaces all uses |
| 1681 | // including the guards/assumes themselves and the uses before the |
| 1682 | // guard/assume. |
| 1683 | else if (OnlyVal && OnlyVal != MultipleVal) |
| 1684 | replaceFoldableUses(Cond: CondInst, ToVal: OnlyVal, KnownAtEndOfBB: BB); |
| 1685 | } |
| 1686 | return true; |
| 1687 | } |
| 1688 | } |
| 1689 | |
| 1690 | // Determine which is the most common successor. If we have many inputs and |
| 1691 | // this block is a switch, we want to start by threading the batch that goes |
| 1692 | // to the most popular destination first. If we only know about one |
| 1693 | // threadable destination (the common case) we can avoid this. |
| 1694 | BasicBlock *MostPopularDest = OnlyDest; |
| 1695 | |
| 1696 | if (MostPopularDest == MultipleDestSentinel) { |
| 1697 | // Remove any loop headers from the Dest list, threadEdge conservatively |
| 1698 | // won't process them, but we might have other destination that are eligible |
| 1699 | // and we still want to process. |
| 1700 | erase_if(C&: PredToDestList, |
| 1701 | P: [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) { |
| 1702 | return LoopHeaders.contains(Ptr: PredToDest.second); |
| 1703 | }); |
| 1704 | |
| 1705 | if (PredToDestList.empty()) |
| 1706 | return false; |
| 1707 | |
| 1708 | MostPopularDest = findMostPopularDest(BB, PredToDestList); |
| 1709 | } |
| 1710 | |
| 1711 | // Now that we know what the most popular destination is, factor all |
| 1712 | // predecessors that will jump to it into a single predecessor. |
| 1713 | SmallVector<BasicBlock*, 16> PredsToFactor; |
| 1714 | for (const auto &PredToDest : PredToDestList) |
| 1715 | if (PredToDest.second == MostPopularDest) { |
| 1716 | BasicBlock *Pred = PredToDest.first; |
| 1717 | |
| 1718 | // This predecessor may be a switch or something else that has multiple |
| 1719 | // edges to the block. Factor each of these edges by listing them |
| 1720 | // according to # occurrences in PredsToFactor. |
| 1721 | for (BasicBlock *Succ : successors(BB: Pred)) |
| 1722 | if (Succ == BB) |
| 1723 | PredsToFactor.push_back(Elt: Pred); |
| 1724 | } |
| 1725 | |
| 1726 | // If the threadable edges are branching on an undefined value, we get to pick |
| 1727 | // the destination that these predecessors should get to. |
| 1728 | if (!MostPopularDest) |
| 1729 | MostPopularDest = BB->getTerminator()-> |
| 1730 | getSuccessor(Idx: getBestDestForJumpOnUndef(BB)); |
| 1731 | |
| 1732 | // Ok, try to thread it! |
| 1733 | return tryThreadEdge(BB, PredBBs: PredsToFactor, SuccBB: MostPopularDest); |
| 1734 | } |
| 1735 | |
| 1736 | /// processBranchOnPHI - We have an otherwise unthreadable conditional branch on |
| 1737 | /// a PHI node (or freeze PHI) in the current block. See if there are any |
| 1738 | /// simplifications we can do based on inputs to the phi node. |
| 1739 | bool JumpThreadingPass::processBranchOnPHI(PHINode *PN) { |
| 1740 | BasicBlock *BB = PN->getParent(); |
| 1741 | |
| 1742 | // TODO: We could make use of this to do it once for blocks with common PHI |
| 1743 | // values. |
| 1744 | SmallVector<BasicBlock*, 1> PredBBs; |
| 1745 | PredBBs.resize(N: 1); |
| 1746 | |
| 1747 | // If any of the predecessor blocks end in an unconditional branch, we can |
| 1748 | // *duplicate* the conditional branch into that block in order to further |
| 1749 | // encourage jump threading and to eliminate cases where we have branch on a |
| 1750 | // phi of an icmp (branch on icmp is much better). |
| 1751 | // This is still beneficial when a frozen phi is used as the branch condition |
| 1752 | // because it allows CodeGenPrepare to further canonicalize br(freeze(icmp)) |
| 1753 | // to br(icmp(freeze ...)). |
| 1754 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { |
| 1755 | BasicBlock *PredBB = PN->getIncomingBlock(i); |
| 1756 | if (isa<UncondBrInst>(Val: PredBB->getTerminator())) { |
| 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>(Val: BO->getOperand(i_nocapture: 0)) || |
| 1776 | isa<ConstantInt>(Val: BO->getOperand(i_nocapture: 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>(Val: 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(V: BO->getOperand(i_nocapture: 0), BB, Result&: XorOpValues, |
| 1809 | Preference: WantInteger, CxtI: BO)) { |
| 1810 | assert(XorOpValues.empty()); |
| 1811 | if (!computeValueKnownInPredecessors(V: BO->getOperand(i_nocapture: 1), BB, Result&: XorOpValues, |
| 1812 | Preference: WantInteger, CxtI: 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>(Val: XorOpValue.first)) |
| 1825 | // Ignore undefs for the count. |
| 1826 | continue; |
| 1827 | if (cast<ConstantInt>(Val: 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(Context&: BB->getContext()); |
| 1837 | else if (NumTrue != 0 || NumFalse != 0) |
| 1838 | SplitVal = ConstantInt::getFalse(Context&: 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>(Val: XorOpValue.first)) |
| 1845 | continue; |
| 1846 | |
| 1847 | BlocksToFoldInto.push_back(Elt: 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>(Val&: BB->front()).getNumIncomingValues()) { |
| 1854 | if (!SplitVal) { |
| 1855 | // If all preds provide undef, just nuke the xor, because it is undef too. |
| 1856 | BO->replaceAllUsesWith(V: UndefValue::get(T: BO->getType())); |
| 1857 | BO->eraseFromParent(); |
| 1858 | } else if (SplitVal->isZero() && BO != BO->getOperand(i_nocapture: isLHS)) { |
| 1859 | // If all preds provide 0, replace the xor with the other input. |
| 1860 | BO->replaceAllUsesWith(V: BO->getOperand(i_nocapture: isLHS)); |
| 1861 | BO->eraseFromParent(); |
| 1862 | } else { |
| 1863 | // If all preds provide 1, set the computed value to 1. |
| 1864 | BO->setOperand(i_nocapture: !isLHS, Val_nocapture: SplitVal); |
| 1865 | } |
| 1866 | |
| 1867 | return true; |
| 1868 | } |
| 1869 | |
| 1870 | // If any of predecessors end with an indirect goto, we can't change its |
| 1871 | // destination. |
| 1872 | if (any_of(Range&: BlocksToFoldInto, P: [](BasicBlock *Pred) { |
| 1873 | return isa<IndirectBrInst>(Val: Pred->getTerminator()); |
| 1874 | })) |
| 1875 | return false; |
| 1876 | |
| 1877 | // Try to duplicate BB into PredBB. |
| 1878 | return duplicateCondBranchOnPHIIntoPred(BB, PredBBs: BlocksToFoldInto); |
| 1879 | } |
| 1880 | |
| 1881 | /// addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new |
| 1882 | /// predecessor to the PHIBB block. If it has PHI nodes, add entries for |
| 1883 | /// NewPred using the entries from OldPred (suitably mapped). |
| 1884 | static void addPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, |
| 1885 | BasicBlock *OldPred, |
| 1886 | BasicBlock *NewPred, |
| 1887 | ValueToValueMapTy &ValueMap) { |
| 1888 | for (PHINode &PN : PHIBB->phis()) { |
| 1889 | // Ok, we have a PHI node. Figure out what the incoming value was for the |
| 1890 | // DestBlock. |
| 1891 | Value *IV = PN.getIncomingValueForBlock(BB: OldPred); |
| 1892 | |
| 1893 | // Remap the value if necessary. |
| 1894 | if (Instruction *Inst = dyn_cast<Instruction>(Val: IV)) { |
| 1895 | ValueToValueMapTy::iterator I = ValueMap.find(Val: Inst); |
| 1896 | if (I != ValueMap.end()) |
| 1897 | IV = I->second; |
| 1898 | } |
| 1899 | |
| 1900 | PN.addIncoming(V: IV, BB: NewPred); |
| 1901 | } |
| 1902 | } |
| 1903 | |
| 1904 | /// Merge basic block BB into its sole predecessor if possible. |
| 1905 | bool JumpThreadingPass::maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB) { |
| 1906 | BasicBlock *SinglePred = BB->getSinglePredecessor(); |
| 1907 | if (!SinglePred) |
| 1908 | return false; |
| 1909 | |
| 1910 | const Instruction *TI = SinglePred->getTerminator(); |
| 1911 | if (TI->isSpecialTerminator() || TI->getNumSuccessors() != 1 || |
| 1912 | SinglePred == BB || hasAddressTakenAndUsed(BB)) |
| 1913 | return false; |
| 1914 | |
| 1915 | // MergeBasicBlockIntoOnlyPred may delete SinglePred, we need to avoid |
| 1916 | // deleting a BB pointer from Unreachable. |
| 1917 | if (Unreachable.count(Ptr: SinglePred)) |
| 1918 | return false; |
| 1919 | |
| 1920 | // Don't merge if both the basic block and the predecessor contain loop or |
| 1921 | // entry convergent intrinsics, since there may only be one convergence token |
| 1922 | // per block. |
| 1923 | if (HasLoopOrEntryConvergenceToken(BB) && |
| 1924 | HasLoopOrEntryConvergenceToken(BB: SinglePred)) |
| 1925 | return false; |
| 1926 | |
| 1927 | // If SinglePred was a loop header, BB becomes one. |
| 1928 | if (LoopHeaders.erase(Ptr: SinglePred)) |
| 1929 | LoopHeaders.insert(Ptr: BB); |
| 1930 | |
| 1931 | LVI->eraseBlock(BB: SinglePred); |
| 1932 | MergeBasicBlockIntoOnlyPred(BB, DTU: DTU.get()); |
| 1933 | |
| 1934 | // Now that BB is merged into SinglePred (i.e. SinglePred code followed by |
| 1935 | // BB code within one basic block `BB`), we need to invalidate the LVI |
| 1936 | // information associated with BB, because the LVI information need not be |
| 1937 | // true for all of BB after the merge. For example, |
| 1938 | // Before the merge, LVI info and code is as follows: |
| 1939 | // SinglePred: <LVI info1 for %p val> |
| 1940 | // %y = use of %p |
| 1941 | // call @exit() // need not transfer execution to successor. |
| 1942 | // assume(%p) // from this point on %p is true |
| 1943 | // br label %BB |
| 1944 | // BB: <LVI info2 for %p val, i.e. %p is true> |
| 1945 | // %x = use of %p |
| 1946 | // br label exit |
| 1947 | // |
| 1948 | // Note that this LVI info for blocks BB and SinglPred is correct for %p |
| 1949 | // (info2 and info1 respectively). After the merge and the deletion of the |
| 1950 | // LVI info1 for SinglePred. We have the following code: |
| 1951 | // BB: <LVI info2 for %p val> |
| 1952 | // %y = use of %p |
| 1953 | // call @exit() |
| 1954 | // assume(%p) |
| 1955 | // %x = use of %p <-- LVI info2 is correct from here onwards. |
| 1956 | // br label exit |
| 1957 | // LVI info2 for BB is incorrect at the beginning of BB. |
| 1958 | |
| 1959 | // Invalidate LVI information for BB if the LVI is not provably true for |
| 1960 | // all of BB. |
| 1961 | if (!isGuaranteedToTransferExecutionToSuccessor(BB)) |
| 1962 | LVI->eraseBlock(BB); |
| 1963 | return true; |
| 1964 | } |
| 1965 | |
| 1966 | /// Update the SSA form. NewBB contains instructions that are copied from BB. |
| 1967 | /// ValueMapping maps old values in BB to new ones in NewBB. |
| 1968 | void JumpThreadingPass::updateSSA(BasicBlock *BB, BasicBlock *NewBB, |
| 1969 | ValueToValueMapTy &ValueMapping) { |
| 1970 | // If there were values defined in BB that are used outside the block, then we |
| 1971 | // now have to update all uses of the value to use either the original value, |
| 1972 | // the cloned value, or some PHI derived value. This can require arbitrary |
| 1973 | // PHI insertion, of which we are prepared to do, clean these up now. |
| 1974 | SSAUpdater SSAUpdate; |
| 1975 | SmallVector<Use *, 16> UsesToRename; |
| 1976 | SmallVector<DbgVariableRecord *, 4> DbgVariableRecords; |
| 1977 | |
| 1978 | for (Instruction &I : *BB) { |
| 1979 | // Scan all uses of this instruction to see if it is used outside of its |
| 1980 | // block, and if so, record them in UsesToRename. |
| 1981 | for (Use &U : I.uses()) { |
| 1982 | Instruction *User = cast<Instruction>(Val: U.getUser()); |
| 1983 | if (PHINode *UserPN = dyn_cast<PHINode>(Val: User)) { |
| 1984 | if (UserPN->getIncomingBlock(U) == BB) |
| 1985 | continue; |
| 1986 | } else if (User->getParent() == BB) |
| 1987 | continue; |
| 1988 | |
| 1989 | UsesToRename.push_back(Elt: &U); |
| 1990 | } |
| 1991 | |
| 1992 | // Find debug values outside of the block |
| 1993 | findDbgValues(V: &I, DbgVariableRecords); |
| 1994 | llvm::erase_if(C&: DbgVariableRecords, P: [&](const DbgVariableRecord *DbgVarRec) { |
| 1995 | return DbgVarRec->getParent() == BB; |
| 1996 | }); |
| 1997 | |
| 1998 | // If there are no uses outside the block, we're done with this instruction. |
| 1999 | if (UsesToRename.empty() && DbgVariableRecords.empty()) |
| 2000 | continue; |
| 2001 | LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n" ); |
| 2002 | |
| 2003 | // We found a use of I outside of BB. Rename all uses of I that are outside |
| 2004 | // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks |
| 2005 | // with the two values we know. |
| 2006 | SSAUpdate.Initialize(Ty: I.getType(), Name: I.getName()); |
| 2007 | SSAUpdate.AddAvailableValue(BB, V: &I); |
| 2008 | SSAUpdate.AddAvailableValue(BB: NewBB, V: ValueMapping[&I]); |
| 2009 | |
| 2010 | while (!UsesToRename.empty()) |
| 2011 | SSAUpdate.RewriteUse(U&: *UsesToRename.pop_back_val()); |
| 2012 | if (!DbgVariableRecords.empty()) { |
| 2013 | SSAUpdate.UpdateDebugValues(I: &I, DbgValues&: DbgVariableRecords); |
| 2014 | DbgVariableRecords.clear(); |
| 2015 | } |
| 2016 | |
| 2017 | LLVM_DEBUG(dbgs() << "\n" ); |
| 2018 | } |
| 2019 | } |
| 2020 | |
| 2021 | static void remapSourceAtoms(ValueToValueMapTy &VM, BasicBlock::iterator Begin, |
| 2022 | BasicBlock::iterator End) { |
| 2023 | if (VM.AtomMap.empty()) |
| 2024 | return; |
| 2025 | for (auto It = Begin; It != End; ++It) |
| 2026 | RemapSourceAtom(I: &*It, VM); |
| 2027 | } |
| 2028 | |
| 2029 | /// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone |
| 2030 | /// arguments that come from PredBB. Return the map from the variables in the |
| 2031 | /// source basic block to the variables in the newly created basic block. |
| 2032 | |
| 2033 | void JumpThreadingPass::cloneInstructions(ValueToValueMapTy &ValueMapping, |
| 2034 | BasicBlock::iterator BI, |
| 2035 | BasicBlock::iterator BE, |
| 2036 | BasicBlock *NewBB, |
| 2037 | BasicBlock *PredBB) { |
| 2038 | // We are going to have to map operands from the source basic block to the new |
| 2039 | // copy of the block 'NewBB'. If there are PHI nodes in the source basic |
| 2040 | // block, evaluate them to account for entry from PredBB. |
| 2041 | |
| 2042 | // Retargets dbg.value to any renamed variables. |
| 2043 | auto RetargetDbgVariableRecordIfPossible = [&](DbgVariableRecord *DVR) { |
| 2044 | SmallSet<std::pair<Value *, Value *>, 16> OperandsToRemap; |
| 2045 | for (auto *Op : DVR->location_ops()) { |
| 2046 | Instruction *OpInst = dyn_cast<Instruction>(Val: Op); |
| 2047 | if (!OpInst) |
| 2048 | continue; |
| 2049 | |
| 2050 | auto I = ValueMapping.find(Val: OpInst); |
| 2051 | if (I != ValueMapping.end()) |
| 2052 | OperandsToRemap.insert(V: {OpInst, I->second}); |
| 2053 | } |
| 2054 | |
| 2055 | for (auto &[OldOp, MappedOp] : OperandsToRemap) |
| 2056 | DVR->replaceVariableLocationOp(OldValue: OldOp, NewValue: MappedOp); |
| 2057 | }; |
| 2058 | |
| 2059 | BasicBlock *RangeBB = BI->getParent(); |
| 2060 | |
| 2061 | // Clone the phi nodes of the source basic block into NewBB. The resulting |
| 2062 | // phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater |
| 2063 | // might need to rewrite the operand of the cloned phi. |
| 2064 | for (; PHINode *PN = dyn_cast<PHINode>(Val&: BI); ++BI) { |
| 2065 | PHINode *NewPN = PHINode::Create(Ty: PN->getType(), NumReservedValues: 1, NameStr: PN->getName(), InsertBefore: NewBB); |
| 2066 | NewPN->addIncoming(V: PN->getIncomingValueForBlock(BB: PredBB), BB: PredBB); |
| 2067 | ValueMapping[PN] = NewPN; |
| 2068 | if (const DebugLoc &DL = PN->getDebugLoc()) |
| 2069 | mapAtomInstance(DL, VMap&: ValueMapping); |
| 2070 | } |
| 2071 | |
| 2072 | // Clone noalias scope declarations in the threaded block. When threading a |
| 2073 | // loop exit, we would otherwise end up with two idential scope declarations |
| 2074 | // visible at the same time. |
| 2075 | SmallVector<MDNode *> NoAliasScopes; |
| 2076 | DenseMap<MDNode *, MDNode *> ClonedScopes; |
| 2077 | LLVMContext &Context = PredBB->getContext(); |
| 2078 | identifyNoAliasScopesToClone(Start: BI, End: BE, NoAliasDeclScopes&: NoAliasScopes); |
| 2079 | cloneNoAliasScopes(NoAliasDeclScopes: NoAliasScopes, ClonedScopes, Ext: "thread" , Context); |
| 2080 | |
| 2081 | auto CloneAndRemapDbgInfo = [&](Instruction *NewInst, Instruction *From) { |
| 2082 | auto DVRRange = NewInst->cloneDebugInfoFrom(From); |
| 2083 | for (DbgVariableRecord &DVR : filterDbgVars(R: DVRRange)) |
| 2084 | RetargetDbgVariableRecordIfPossible(&DVR); |
| 2085 | }; |
| 2086 | |
| 2087 | // Clone the non-phi instructions of the source basic block into NewBB, |
| 2088 | // keeping track of the mapping and using it to remap operands in the cloned |
| 2089 | // instructions. |
| 2090 | for (; BI != BE; ++BI) { |
| 2091 | Instruction *New = BI->clone(); |
| 2092 | New->setName(BI->getName()); |
| 2093 | New->insertInto(ParentBB: NewBB, It: NewBB->end()); |
| 2094 | ValueMapping[&*BI] = New; |
| 2095 | adaptNoAliasScopes(I: New, ClonedScopes, Context); |
| 2096 | |
| 2097 | CloneAndRemapDbgInfo(New, &*BI); |
| 2098 | if (const DebugLoc &DL = New->getDebugLoc()) |
| 2099 | mapAtomInstance(DL, VMap&: ValueMapping); |
| 2100 | |
| 2101 | // Remap operands to patch up intra-block references. |
| 2102 | for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) |
| 2103 | if (Instruction *Inst = dyn_cast<Instruction>(Val: New->getOperand(i))) { |
| 2104 | ValueToValueMapTy::iterator I = ValueMapping.find(Val: Inst); |
| 2105 | if (I != ValueMapping.end()) |
| 2106 | New->setOperand(i, Val: I->second); |
| 2107 | } |
| 2108 | } |
| 2109 | |
| 2110 | // There may be DbgVariableRecords on the terminator, clone directly from |
| 2111 | // marker to marker as there isn't an instruction there. |
| 2112 | if (BE != RangeBB->end() && BE->hasDbgRecords()) { |
| 2113 | // Dump them at the end. |
| 2114 | DbgMarker *Marker = RangeBB->getMarker(It: BE); |
| 2115 | DbgMarker *EndMarker = NewBB->createMarker(It: NewBB->end()); |
| 2116 | auto DVRRange = EndMarker->cloneDebugInfoFrom(From: Marker, FromHere: std::nullopt); |
| 2117 | for (DbgVariableRecord &DVR : filterDbgVars(R: DVRRange)) |
| 2118 | RetargetDbgVariableRecordIfPossible(&DVR); |
| 2119 | } |
| 2120 | } |
| 2121 | |
| 2122 | /// Attempt to thread through two successive basic blocks. |
| 2123 | bool JumpThreadingPass::maybethreadThroughTwoBasicBlocks(BasicBlock *BB, |
| 2124 | Value *Cond) { |
| 2125 | // Consider: |
| 2126 | // |
| 2127 | // PredBB: |
| 2128 | // %var = phi i32* [ null, %bb1 ], [ @a, %bb2 ] |
| 2129 | // %tobool = icmp eq i32 %cond, 0 |
| 2130 | // br i1 %tobool, label %BB, label ... |
| 2131 | // |
| 2132 | // BB: |
| 2133 | // %cmp = icmp eq i32* %var, null |
| 2134 | // br i1 %cmp, label ..., label ... |
| 2135 | // |
| 2136 | // We don't know the value of %var at BB even if we know which incoming edge |
| 2137 | // we take to BB. However, once we duplicate PredBB for each of its incoming |
| 2138 | // edges (say, PredBB1 and PredBB2), we know the value of %var in each copy of |
| 2139 | // PredBB. Then we can thread edges PredBB1->BB and PredBB2->BB through BB. |
| 2140 | |
| 2141 | // Require that BB end with a Branch for simplicity. |
| 2142 | CondBrInst *CondBr = dyn_cast<CondBrInst>(Val: BB->getTerminator()); |
| 2143 | if (!CondBr) |
| 2144 | return false; |
| 2145 | |
| 2146 | // BB must have exactly one predecessor. |
| 2147 | BasicBlock *PredBB = BB->getSinglePredecessor(); |
| 2148 | if (!PredBB) |
| 2149 | return false; |
| 2150 | |
| 2151 | // Require that PredBB end with a conditional Branch. If PredBB ends with an |
| 2152 | // unconditional branch, we should be merging PredBB and BB instead. For |
| 2153 | // simplicity, we don't deal with a switch. |
| 2154 | CondBrInst *PredBBBranch = dyn_cast<CondBrInst>(Val: PredBB->getTerminator()); |
| 2155 | if (!PredBBBranch) |
| 2156 | return false; |
| 2157 | |
| 2158 | // If PredBB has exactly one incoming edge, we don't gain anything by copying |
| 2159 | // PredBB. |
| 2160 | if (PredBB->getSinglePredecessor()) |
| 2161 | return false; |
| 2162 | |
| 2163 | // Don't thread through PredBB if it contains a successor edge to itself, in |
| 2164 | // which case we would infinite loop. Suppose we are threading an edge from |
| 2165 | // PredPredBB through PredBB and BB to SuccBB with PredBB containing a |
| 2166 | // successor edge to itself. If we allowed jump threading in this case, we |
| 2167 | // could duplicate PredBB and BB as, say, PredBB.thread and BB.thread. Since |
| 2168 | // PredBB.thread has a successor edge to PredBB, we would immediately come up |
| 2169 | // with another jump threading opportunity from PredBB.thread through PredBB |
| 2170 | // and BB to SuccBB. This jump threading would repeatedly occur. That is, we |
| 2171 | // would keep peeling one iteration from PredBB. |
| 2172 | if (llvm::is_contained(Range: successors(BB: PredBB), Element: PredBB)) |
| 2173 | return false; |
| 2174 | |
| 2175 | // Don't thread across a loop header. |
| 2176 | if (LoopHeaders.count(Ptr: PredBB)) |
| 2177 | return false; |
| 2178 | |
| 2179 | // Avoid complication with duplicating EH pads. |
| 2180 | if (PredBB->isEHPad()) |
| 2181 | return false; |
| 2182 | |
| 2183 | // Find a predecessor that we can thread. For simplicity, we only consider a |
| 2184 | // successor edge out of BB to which we thread exactly one incoming edge into |
| 2185 | // PredBB. |
| 2186 | unsigned ZeroCount = 0; |
| 2187 | unsigned OneCount = 0; |
| 2188 | BasicBlock *ZeroPred = nullptr; |
| 2189 | BasicBlock *OnePred = nullptr; |
| 2190 | const DataLayout &DL = BB->getDataLayout(); |
| 2191 | for (BasicBlock *P : predecessors(BB: PredBB)) { |
| 2192 | // If PredPred ends with IndirectBrInst, we can't handle it. |
| 2193 | if (isa<IndirectBrInst>(Val: P->getTerminator())) |
| 2194 | continue; |
| 2195 | if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>( |
| 2196 | Val: evaluateOnPredecessorEdge(BB, PredPredBB: P, V: Cond, DL))) { |
| 2197 | if (CI->isZero()) { |
| 2198 | ZeroCount++; |
| 2199 | ZeroPred = P; |
| 2200 | } else if (CI->isOne()) { |
| 2201 | OneCount++; |
| 2202 | OnePred = P; |
| 2203 | } |
| 2204 | } |
| 2205 | } |
| 2206 | |
| 2207 | // Disregard complicated cases where we have to thread multiple edges. |
| 2208 | BasicBlock *PredPredBB; |
| 2209 | if (ZeroCount == 1) { |
| 2210 | PredPredBB = ZeroPred; |
| 2211 | } else if (OneCount == 1) { |
| 2212 | PredPredBB = OnePred; |
| 2213 | } else { |
| 2214 | return false; |
| 2215 | } |
| 2216 | |
| 2217 | BasicBlock *SuccBB = CondBr->getSuccessor(i: PredPredBB == ZeroPred); |
| 2218 | |
| 2219 | // If threading to the same block as we come from, we would infinite loop. |
| 2220 | if (SuccBB == BB) { |
| 2221 | LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName() |
| 2222 | << "' - would thread to self!\n" ); |
| 2223 | return false; |
| 2224 | } |
| 2225 | |
| 2226 | // If threading this would thread across a loop header, don't thread the edge. |
| 2227 | // See the comments above findLoopHeaders for justifications and caveats. |
| 2228 | if (LoopHeaders.count(Ptr: BB) || LoopHeaders.count(Ptr: SuccBB)) { |
| 2229 | LLVM_DEBUG({ |
| 2230 | bool BBIsHeader = LoopHeaders.count(BB); |
| 2231 | bool SuccIsHeader = LoopHeaders.count(SuccBB); |
| 2232 | dbgs() << " Not threading across " |
| 2233 | << (BBIsHeader ? "loop header BB '" : "block BB '" ) |
| 2234 | << BB->getName() << "' to dest " |
| 2235 | << (SuccIsHeader ? "loop header BB '" : "block BB '" ) |
| 2236 | << SuccBB->getName() |
| 2237 | << "' - it might create an irreducible loop!\n" ; |
| 2238 | }); |
| 2239 | return false; |
| 2240 | } |
| 2241 | |
| 2242 | // Compute the cost of duplicating BB and PredBB. |
| 2243 | unsigned BBCost = getJumpThreadDuplicationCost( |
| 2244 | TTI, BB, StopAt: BB->getTerminator(), Threshold: BBDupThreshold); |
| 2245 | unsigned PredBBCost = getJumpThreadDuplicationCost( |
| 2246 | TTI, BB: PredBB, StopAt: PredBB->getTerminator(), Threshold: BBDupThreshold); |
| 2247 | |
| 2248 | // Give up if costs are too high. We need to check BBCost and PredBBCost |
| 2249 | // individually before checking their sum because getJumpThreadDuplicationCost |
| 2250 | // return (unsigned)~0 for those basic blocks that cannot be duplicated. |
| 2251 | if (BBCost > BBDupThreshold || PredBBCost > BBDupThreshold || |
| 2252 | BBCost + PredBBCost > BBDupThreshold) { |
| 2253 | LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName() |
| 2254 | << "' - Cost is too high: " << PredBBCost |
| 2255 | << " for PredBB, " << BBCost << "for BB\n" ); |
| 2256 | return false; |
| 2257 | } |
| 2258 | |
| 2259 | // Now we are ready to duplicate PredBB. |
| 2260 | threadThroughTwoBasicBlocks(PredPredBB, PredBB, BB, SuccBB); |
| 2261 | return true; |
| 2262 | } |
| 2263 | |
| 2264 | void JumpThreadingPass::threadThroughTwoBasicBlocks(BasicBlock *PredPredBB, |
| 2265 | BasicBlock *PredBB, |
| 2266 | BasicBlock *BB, |
| 2267 | BasicBlock *SuccBB) { |
| 2268 | LLVM_DEBUG(dbgs() << " Threading through '" << PredBB->getName() << "' and '" |
| 2269 | << BB->getName() << "'\n" ); |
| 2270 | |
| 2271 | // Build BPI/BFI before any changes are made to IR. |
| 2272 | bool HasProfile = doesBlockHaveProfileData(BB); |
| 2273 | auto *BFI = getOrCreateBFI(Force: HasProfile); |
| 2274 | auto *BPI = getOrCreateBPI(Force: BFI != nullptr); |
| 2275 | |
| 2276 | CondBrInst *CondBr = cast<CondBrInst>(Val: BB->getTerminator()); |
| 2277 | CondBrInst *PredBBBranch = cast<CondBrInst>(Val: PredBB->getTerminator()); |
| 2278 | |
| 2279 | BasicBlock *NewBB = |
| 2280 | BasicBlock::Create(Context&: PredBB->getContext(), Name: PredBB->getName() + ".thread" , |
| 2281 | Parent: PredBB->getParent(), InsertBefore: PredBB); |
| 2282 | NewBB->moveAfter(MovePos: PredBB); |
| 2283 | |
| 2284 | // Set the block frequency of NewBB. |
| 2285 | if (BFI) { |
| 2286 | assert(BPI && "It's expected BPI to exist along with BFI" ); |
| 2287 | auto NewBBFreq = BFI->getBlockFreq(BB: PredPredBB) * |
| 2288 | BPI->getEdgeProbability(Src: PredPredBB, Dst: PredBB); |
| 2289 | BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq); |
| 2290 | } |
| 2291 | |
| 2292 | // We are going to have to map operands from the original BB block to the new |
| 2293 | // copy of the block 'NewBB'. If there are PHI nodes in PredBB, evaluate them |
| 2294 | // to account for entry from PredPredBB. |
| 2295 | ValueToValueMapTy ValueMapping; |
| 2296 | cloneInstructions(ValueMapping, BI: PredBB->begin(), BE: PredBB->end(), NewBB, |
| 2297 | PredBB: PredPredBB); |
| 2298 | |
| 2299 | // Copy the edge probabilities from PredBB to NewBB. |
| 2300 | if (BPI) |
| 2301 | BPI->copyEdgeProbabilities(Src: PredBB, Dst: NewBB); |
| 2302 | |
| 2303 | // Update the terminator of PredPredBB to jump to NewBB instead of PredBB. |
| 2304 | // This eliminates predecessors from PredPredBB, which requires us to simplify |
| 2305 | // any PHI nodes in PredBB. |
| 2306 | Instruction *PredPredTerm = PredPredBB->getTerminator(); |
| 2307 | for (unsigned i = 0, e = PredPredTerm->getNumSuccessors(); i != e; ++i) |
| 2308 | if (PredPredTerm->getSuccessor(Idx: i) == PredBB) { |
| 2309 | PredBB->removePredecessor(Pred: PredPredBB, KeepOneInputPHIs: true); |
| 2310 | PredPredTerm->setSuccessor(Idx: i, BB: NewBB); |
| 2311 | } |
| 2312 | |
| 2313 | addPHINodeEntriesForMappedBlock(PHIBB: PredBBBranch->getSuccessor(i: 0), OldPred: PredBB, NewPred: NewBB, |
| 2314 | ValueMap&: ValueMapping); |
| 2315 | addPHINodeEntriesForMappedBlock(PHIBB: PredBBBranch->getSuccessor(i: 1), OldPred: PredBB, NewPred: NewBB, |
| 2316 | ValueMap&: ValueMapping); |
| 2317 | |
| 2318 | DTU->applyUpdatesPermissive( |
| 2319 | Updates: {{DominatorTree::Insert, NewBB, CondBr->getSuccessor(i: 0)}, |
| 2320 | {DominatorTree::Insert, NewBB, CondBr->getSuccessor(i: 1)}, |
| 2321 | {DominatorTree::Insert, PredPredBB, NewBB}, |
| 2322 | {DominatorTree::Delete, PredPredBB, PredBB}}); |
| 2323 | |
| 2324 | // Remap source location atoms beacuse we're duplicating control flow. |
| 2325 | remapSourceAtoms(VM&: ValueMapping, Begin: NewBB->begin(), End: NewBB->end()); |
| 2326 | |
| 2327 | updateSSA(BB: PredBB, NewBB, ValueMapping); |
| 2328 | |
| 2329 | // Clean up things like PHI nodes with single operands, dead instructions, |
| 2330 | // etc. |
| 2331 | SimplifyInstructionsInBlock(BB: NewBB, TLI); |
| 2332 | SimplifyInstructionsInBlock(BB: PredBB, TLI); |
| 2333 | |
| 2334 | SmallVector<BasicBlock *, 1> PredsToFactor; |
| 2335 | PredsToFactor.push_back(Elt: NewBB); |
| 2336 | threadEdge(BB, PredBBs: PredsToFactor, SuccBB); |
| 2337 | } |
| 2338 | |
| 2339 | /// tryThreadEdge - Thread an edge if it's safe and profitable to do so. |
| 2340 | bool JumpThreadingPass::tryThreadEdge( |
| 2341 | BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs, |
| 2342 | BasicBlock *SuccBB) { |
| 2343 | // If threading to the same block as we come from, we would infinite loop. |
| 2344 | if (SuccBB == BB) { |
| 2345 | LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName() |
| 2346 | << "' - would thread to self!\n" ); |
| 2347 | return false; |
| 2348 | } |
| 2349 | |
| 2350 | // If threading this would thread across a loop header, don't thread the edge. |
| 2351 | // See the comments above findLoopHeaders for justifications and caveats. |
| 2352 | if (LoopHeaders.count(Ptr: BB) || LoopHeaders.count(Ptr: SuccBB)) { |
| 2353 | LLVM_DEBUG({ |
| 2354 | bool BBIsHeader = LoopHeaders.count(BB); |
| 2355 | bool SuccIsHeader = LoopHeaders.count(SuccBB); |
| 2356 | dbgs() << " Not threading across " |
| 2357 | << (BBIsHeader ? "loop header BB '" : "block BB '" ) << BB->getName() |
| 2358 | << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '" ) |
| 2359 | << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; |
| 2360 | }); |
| 2361 | return false; |
| 2362 | } |
| 2363 | |
| 2364 | unsigned JumpThreadCost = getJumpThreadDuplicationCost( |
| 2365 | TTI, BB, StopAt: BB->getTerminator(), Threshold: BBDupThreshold); |
| 2366 | if (JumpThreadCost > BBDupThreshold) { |
| 2367 | LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName() |
| 2368 | << "' - Cost is too high: " << JumpThreadCost << "\n" ); |
| 2369 | return false; |
| 2370 | } |
| 2371 | |
| 2372 | threadEdge(BB, PredBBs, SuccBB); |
| 2373 | return true; |
| 2374 | } |
| 2375 | |
| 2376 | /// threadEdge - We have decided that it is safe and profitable to factor the |
| 2377 | /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB |
| 2378 | /// across BB. Transform the IR to reflect this change. |
| 2379 | void JumpThreadingPass::threadEdge(BasicBlock *BB, |
| 2380 | const SmallVectorImpl<BasicBlock *> &PredBBs, |
| 2381 | BasicBlock *SuccBB) { |
| 2382 | assert(SuccBB != BB && "Don't create an infinite loop" ); |
| 2383 | |
| 2384 | assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) && |
| 2385 | "Don't thread across loop headers" ); |
| 2386 | |
| 2387 | // Build BPI/BFI before any changes are made to IR. |
| 2388 | bool HasProfile = doesBlockHaveProfileData(BB); |
| 2389 | auto *BFI = getOrCreateBFI(Force: HasProfile); |
| 2390 | auto *BPI = getOrCreateBPI(Force: BFI != nullptr); |
| 2391 | |
| 2392 | // And finally, do it! Start by factoring the predecessors if needed. |
| 2393 | BasicBlock *PredBB; |
| 2394 | if (PredBBs.size() == 1) |
| 2395 | PredBB = PredBBs[0]; |
| 2396 | else { |
| 2397 | LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size() |
| 2398 | << " common predecessors.\n" ); |
| 2399 | PredBB = splitBlockPreds(BB, Preds: PredBBs, Suffix: ".thr_comm" ); |
| 2400 | } |
| 2401 | |
| 2402 | // And finally, do it! |
| 2403 | LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() |
| 2404 | << "' to '" << SuccBB->getName() |
| 2405 | << ", across block:\n " << *BB << "\n" ); |
| 2406 | |
| 2407 | LVI->threadEdge(PredBB, OldSucc: BB, NewSucc: SuccBB); |
| 2408 | |
| 2409 | BasicBlock *NewBB = BasicBlock::Create(Context&: BB->getContext(), |
| 2410 | Name: BB->getName()+".thread" , |
| 2411 | Parent: BB->getParent(), InsertBefore: BB); |
| 2412 | NewBB->moveAfter(MovePos: PredBB); |
| 2413 | |
| 2414 | // Set the block frequency of NewBB. |
| 2415 | if (BFI) { |
| 2416 | assert(BPI && "It's expected BPI to exist along with BFI" ); |
| 2417 | auto NewBBFreq = |
| 2418 | BFI->getBlockFreq(BB: PredBB) * BPI->getEdgeProbability(Src: PredBB, Dst: BB); |
| 2419 | BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq); |
| 2420 | } |
| 2421 | |
| 2422 | // Copy all the instructions from BB to NewBB except the terminator. |
| 2423 | ValueToValueMapTy ValueMapping; |
| 2424 | cloneInstructions(ValueMapping, BI: BB->begin(), BE: std::prev(x: BB->end()), NewBB, |
| 2425 | PredBB); |
| 2426 | |
| 2427 | // We didn't copy the terminator from BB over to NewBB, because there is now |
| 2428 | // an unconditional jump to SuccBB. Insert the unconditional jump. |
| 2429 | UncondBrInst *NewBI = UncondBrInst::Create(IfTrue: SuccBB, InsertBefore: NewBB); |
| 2430 | NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc()); |
| 2431 | |
| 2432 | // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the |
| 2433 | // PHI nodes for NewBB now. |
| 2434 | addPHINodeEntriesForMappedBlock(PHIBB: SuccBB, OldPred: BB, NewPred: NewBB, ValueMap&: ValueMapping); |
| 2435 | |
| 2436 | // Update the terminator of PredBB to jump to NewBB instead of BB. This |
| 2437 | // eliminates predecessors from BB, which requires us to simplify any PHI |
| 2438 | // nodes in BB. |
| 2439 | Instruction *PredTerm = PredBB->getTerminator(); |
| 2440 | for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) |
| 2441 | if (PredTerm->getSuccessor(Idx: i) == BB) { |
| 2442 | BB->removePredecessor(Pred: PredBB, KeepOneInputPHIs: true); |
| 2443 | PredTerm->setSuccessor(Idx: i, BB: NewBB); |
| 2444 | } |
| 2445 | |
| 2446 | // Enqueue required DT updates. |
| 2447 | DTU->applyUpdatesPermissive(Updates: {{DominatorTree::Insert, NewBB, SuccBB}, |
| 2448 | {DominatorTree::Insert, PredBB, NewBB}, |
| 2449 | {DominatorTree::Delete, PredBB, BB}}); |
| 2450 | |
| 2451 | remapSourceAtoms(VM&: ValueMapping, Begin: NewBB->begin(), End: NewBB->end()); |
| 2452 | updateSSA(BB, NewBB, ValueMapping); |
| 2453 | |
| 2454 | // At this point, the IR is fully up to date and consistent. Do a quick scan |
| 2455 | // over the new instructions and zap any that are constants or dead. This |
| 2456 | // frequently happens because of phi translation. |
| 2457 | SimplifyInstructionsInBlock(BB: NewBB, TLI); |
| 2458 | |
| 2459 | // Update the edge weight from BB to SuccBB, which should be less than before. |
| 2460 | updateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB, BFI, BPI, HasProfile); |
| 2461 | |
| 2462 | // Threaded an edge! |
| 2463 | ++NumThreads; |
| 2464 | } |
| 2465 | |
| 2466 | /// Create a new basic block that will be the predecessor of BB and successor of |
| 2467 | /// all blocks in Preds. When profile data is available, update the frequency of |
| 2468 | /// this new block. |
| 2469 | BasicBlock *JumpThreadingPass::splitBlockPreds(BasicBlock *BB, |
| 2470 | ArrayRef<BasicBlock *> Preds, |
| 2471 | const char *Suffix) { |
| 2472 | SmallVector<BasicBlock *, 2> NewBBs; |
| 2473 | |
| 2474 | // Collect the frequencies of all predecessors of BB, which will be used to |
| 2475 | // update the edge weight of the result of splitting predecessors. |
| 2476 | DenseMap<BasicBlock *, BlockFrequency> FreqMap; |
| 2477 | auto *BFI = getBFI(); |
| 2478 | if (BFI) { |
| 2479 | auto *BPI = getOrCreateBPI(Force: true); |
| 2480 | for (auto *Pred : Preds) |
| 2481 | FreqMap.insert(KV: std::make_pair( |
| 2482 | x&: Pred, y: BFI->getBlockFreq(BB: Pred) * BPI->getEdgeProbability(Src: Pred, Dst: BB))); |
| 2483 | } |
| 2484 | |
| 2485 | // In the case when BB is a LandingPad block we create 2 new predecessors |
| 2486 | // instead of just one. |
| 2487 | if (BB->isLandingPad()) { |
| 2488 | std::string NewName = std::string(Suffix) + ".split-lp" ; |
| 2489 | SplitLandingPadPredecessors(OrigBB: BB, Preds, Suffix, Suffix2: NewName.c_str(), NewBBs); |
| 2490 | } else { |
| 2491 | NewBBs.push_back(Elt: SplitBlockPredecessors(BB, Preds, Suffix)); |
| 2492 | } |
| 2493 | |
| 2494 | std::vector<DominatorTree::UpdateType> Updates; |
| 2495 | Updates.reserve(n: (2 * Preds.size()) + NewBBs.size()); |
| 2496 | for (auto *NewBB : NewBBs) { |
| 2497 | BlockFrequency NewBBFreq(0); |
| 2498 | Updates.push_back(x: {DominatorTree::Insert, NewBB, BB}); |
| 2499 | for (auto *Pred : predecessors(BB: NewBB)) { |
| 2500 | Updates.push_back(x: {DominatorTree::Delete, Pred, BB}); |
| 2501 | Updates.push_back(x: {DominatorTree::Insert, Pred, NewBB}); |
| 2502 | if (BFI) // Update frequencies between Pred -> NewBB. |
| 2503 | NewBBFreq += FreqMap.lookup(Val: Pred); |
| 2504 | } |
| 2505 | if (BFI) // Apply the summed frequency to NewBB. |
| 2506 | BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq); |
| 2507 | } |
| 2508 | |
| 2509 | DTU->applyUpdatesPermissive(Updates); |
| 2510 | return NewBBs[0]; |
| 2511 | } |
| 2512 | |
| 2513 | bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) { |
| 2514 | const Instruction *TI = BB->getTerminator(); |
| 2515 | if (!TI || TI->getNumSuccessors() < 2) |
| 2516 | return false; |
| 2517 | |
| 2518 | return hasValidBranchWeightMD(I: *TI); |
| 2519 | } |
| 2520 | |
| 2521 | /// Update the block frequency of BB and branch weight and the metadata on the |
| 2522 | /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 - |
| 2523 | /// Freq(PredBB->BB) / Freq(BB->SuccBB). |
| 2524 | void JumpThreadingPass::updateBlockFreqAndEdgeWeight(BasicBlock *PredBB, |
| 2525 | BasicBlock *BB, |
| 2526 | BasicBlock *NewBB, |
| 2527 | BasicBlock *SuccBB, |
| 2528 | BlockFrequencyInfo *BFI, |
| 2529 | BranchProbabilityInfo *BPI, |
| 2530 | bool HasProfile) { |
| 2531 | assert(((BFI && BPI) || (!BFI && !BFI)) && |
| 2532 | "Both BFI & BPI should either be set or unset" ); |
| 2533 | |
| 2534 | if (!BFI) { |
| 2535 | assert(!HasProfile && |
| 2536 | "It's expected to have BFI/BPI when profile info exists" ); |
| 2537 | return; |
| 2538 | } |
| 2539 | |
| 2540 | // As the edge from PredBB to BB is deleted, we have to update the block |
| 2541 | // frequency of BB. |
| 2542 | auto BBOrigFreq = BFI->getBlockFreq(BB); |
| 2543 | auto NewBBFreq = BFI->getBlockFreq(BB: NewBB); |
| 2544 | auto BBNewFreq = BBOrigFreq - NewBBFreq; |
| 2545 | BFI->setBlockFreq(BB, Freq: BBNewFreq); |
| 2546 | |
| 2547 | // Collect updated outgoing edges' frequencies from BB and use them to update |
| 2548 | // edge probabilities. |
| 2549 | SmallVector<uint64_t, 4> BBSuccFreq; |
| 2550 | for (auto It : enumerate(First: successors(BB))) { |
| 2551 | auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(Src: BB, IndexInSuccessors: It.index()); |
| 2552 | auto SuccFreq = |
| 2553 | (It.value() == SuccBB) ? BB2SuccBBFreq - NewBBFreq : BB2SuccBBFreq; |
| 2554 | BBSuccFreq.push_back(Elt: SuccFreq.getFrequency()); |
| 2555 | } |
| 2556 | |
| 2557 | uint64_t MaxBBSuccFreq = *llvm::max_element(Range&: BBSuccFreq); |
| 2558 | |
| 2559 | SmallVector<BranchProbability, 4> BBSuccProbs; |
| 2560 | if (MaxBBSuccFreq == 0) |
| 2561 | BBSuccProbs.assign(NumElts: BBSuccFreq.size(), |
| 2562 | Elt: {1, static_cast<unsigned>(BBSuccFreq.size())}); |
| 2563 | else { |
| 2564 | for (uint64_t Freq : BBSuccFreq) |
| 2565 | BBSuccProbs.push_back( |
| 2566 | Elt: BranchProbability::getBranchProbability(Numerator: Freq, Denominator: MaxBBSuccFreq)); |
| 2567 | // Normalize edge probabilities so that they sum up to one. |
| 2568 | BranchProbability::normalizeProbabilities(Begin: BBSuccProbs.begin(), |
| 2569 | End: BBSuccProbs.end()); |
| 2570 | } |
| 2571 | |
| 2572 | // Update edge probabilities in BPI. |
| 2573 | BPI->setEdgeProbability(Src: BB, Probs: BBSuccProbs); |
| 2574 | |
| 2575 | // Update the profile metadata as well. |
| 2576 | // |
| 2577 | // Don't do this if the profile of the transformed blocks was statically |
| 2578 | // estimated. (This could occur despite the function having an entry |
| 2579 | // frequency in completely cold parts of the CFG.) |
| 2580 | // |
| 2581 | // In this case we don't want to suggest to subsequent passes that the |
| 2582 | // calculated weights are fully consistent. Consider this graph: |
| 2583 | // |
| 2584 | // check_1 |
| 2585 | // 50% / | |
| 2586 | // eq_1 | 50% |
| 2587 | // \ | |
| 2588 | // check_2 |
| 2589 | // 50% / | |
| 2590 | // eq_2 | 50% |
| 2591 | // \ | |
| 2592 | // check_3 |
| 2593 | // 50% / | |
| 2594 | // eq_3 | 50% |
| 2595 | // \ | |
| 2596 | // |
| 2597 | // Assuming the blocks check_* all compare the same value against 1, 2 and 3, |
| 2598 | // the overall probabilities are inconsistent; the total probability that the |
| 2599 | // value is either 1, 2 or 3 is 150%. |
| 2600 | // |
| 2601 | // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3 |
| 2602 | // becomes 0%. This is even worse if the edge whose probability becomes 0% is |
| 2603 | // the loop exit edge. Then based solely on static estimation we would assume |
| 2604 | // the loop was extremely hot. |
| 2605 | // |
| 2606 | // FIXME this locally as well so that BPI and BFI are consistent as well. We |
| 2607 | // shouldn't make edges extremely likely or unlikely based solely on static |
| 2608 | // estimation. |
| 2609 | if (BBSuccProbs.size() >= 2 && HasProfile) { |
| 2610 | SmallVector<uint32_t, 4> Weights; |
| 2611 | for (auto Prob : BBSuccProbs) |
| 2612 | Weights.push_back(Elt: Prob.getNumerator()); |
| 2613 | |
| 2614 | auto TI = BB->getTerminator(); |
| 2615 | setBranchWeights(I&: *TI, Weights, IsExpected: hasBranchWeightOrigin(I: *TI)); |
| 2616 | } |
| 2617 | } |
| 2618 | |
| 2619 | /// duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch |
| 2620 | /// to BB which contains an i1 PHI node and a conditional branch on that PHI. |
| 2621 | /// If we can duplicate the contents of BB up into PredBB do so now, this |
| 2622 | /// improves the odds that the branch will be on an analyzable instruction like |
| 2623 | /// a compare. |
| 2624 | bool JumpThreadingPass::duplicateCondBranchOnPHIIntoPred( |
| 2625 | BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) { |
| 2626 | assert(!PredBBs.empty() && "Can't handle an empty set" ); |
| 2627 | |
| 2628 | // If BB is a loop header, then duplicating this block outside the loop would |
| 2629 | // cause us to transform this into an irreducible loop, don't do this. |
| 2630 | // See the comments above findLoopHeaders for justifications and caveats. |
| 2631 | if (LoopHeaders.count(Ptr: BB)) { |
| 2632 | LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName() |
| 2633 | << "' into predecessor block '" << PredBBs[0]->getName() |
| 2634 | << "' - it might create an irreducible loop!\n" ); |
| 2635 | return false; |
| 2636 | } |
| 2637 | |
| 2638 | unsigned DuplicationCost = getJumpThreadDuplicationCost( |
| 2639 | TTI, BB, StopAt: BB->getTerminator(), Threshold: BBDupThreshold); |
| 2640 | if (DuplicationCost > BBDupThreshold) { |
| 2641 | LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName() |
| 2642 | << "' - Cost is too high: " << DuplicationCost << "\n" ); |
| 2643 | return false; |
| 2644 | } |
| 2645 | |
| 2646 | // And finally, do it! Start by factoring the predecessors if needed. |
| 2647 | std::vector<DominatorTree::UpdateType> Updates; |
| 2648 | BasicBlock *PredBB; |
| 2649 | if (PredBBs.size() == 1) |
| 2650 | PredBB = PredBBs[0]; |
| 2651 | else { |
| 2652 | LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size() |
| 2653 | << " common predecessors.\n" ); |
| 2654 | PredBB = splitBlockPreds(BB, Preds: PredBBs, Suffix: ".thr_comm" ); |
| 2655 | } |
| 2656 | Updates.push_back(x: {DominatorTree::Delete, PredBB, BB}); |
| 2657 | |
| 2658 | // Okay, we decided to do this! Clone all the instructions in BB onto the end |
| 2659 | // of PredBB. |
| 2660 | LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName() |
| 2661 | << "' into end of '" << PredBB->getName() |
| 2662 | << "' to eliminate branch on phi. Cost: " |
| 2663 | << DuplicationCost << " block is:" << *BB << "\n" ); |
| 2664 | |
| 2665 | // Unless PredBB ends with an unconditional branch, split the edge so that we |
| 2666 | // can just clone the bits from BB into the end of the new PredBB. |
| 2667 | UncondBrInst *OldPredBranch = dyn_cast<UncondBrInst>(Val: PredBB->getTerminator()); |
| 2668 | |
| 2669 | if (!OldPredBranch) { |
| 2670 | BasicBlock *OldPredBB = PredBB; |
| 2671 | PredBB = SplitEdge(From: OldPredBB, To: BB); |
| 2672 | Updates.push_back(x: {DominatorTree::Insert, OldPredBB, PredBB}); |
| 2673 | Updates.push_back(x: {DominatorTree::Insert, PredBB, BB}); |
| 2674 | Updates.push_back(x: {DominatorTree::Delete, OldPredBB, BB}); |
| 2675 | OldPredBranch = cast<UncondBrInst>(Val: PredBB->getTerminator()); |
| 2676 | } |
| 2677 | |
| 2678 | // We are going to have to map operands from the original BB block into the |
| 2679 | // PredBB block. Evaluate PHI nodes in BB. |
| 2680 | ValueToValueMapTy ValueMapping; |
| 2681 | |
| 2682 | // Remember the position before the inserted instructions. |
| 2683 | auto RItBeforeInsertPt = std::next(x: OldPredBranch->getReverseIterator()); |
| 2684 | |
| 2685 | BasicBlock::iterator BI = BB->begin(); |
| 2686 | for (; PHINode *PN = dyn_cast<PHINode>(Val&: BI); ++BI) |
| 2687 | ValueMapping[PN] = PN->getIncomingValueForBlock(BB: PredBB); |
| 2688 | // Clone the non-phi instructions of BB into PredBB, keeping track of the |
| 2689 | // mapping and using it to remap operands in the cloned instructions. |
| 2690 | for (; BI != BB->end(); ++BI) { |
| 2691 | Instruction *New = BI->clone(); |
| 2692 | New->insertInto(ParentBB: PredBB, It: OldPredBranch->getIterator()); |
| 2693 | |
| 2694 | // Remap operands to patch up intra-block references. |
| 2695 | for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) |
| 2696 | if (Instruction *Inst = dyn_cast<Instruction>(Val: New->getOperand(i))) { |
| 2697 | ValueToValueMapTy::iterator I = ValueMapping.find(Val: Inst); |
| 2698 | if (I != ValueMapping.end()) |
| 2699 | New->setOperand(i, Val: I->second); |
| 2700 | } |
| 2701 | |
| 2702 | // Remap debug variable operands. |
| 2703 | remapDebugVariable(Mapping&: ValueMapping, Inst: New); |
| 2704 | if (const DebugLoc &DL = New->getDebugLoc()) |
| 2705 | mapAtomInstance(DL, VMap&: ValueMapping); |
| 2706 | |
| 2707 | // If this instruction can be simplified after the operands are updated, |
| 2708 | // just use the simplified value instead. This frequently happens due to |
| 2709 | // phi translation. |
| 2710 | if (Value *IV = simplifyInstruction( |
| 2711 | I: New, |
| 2712 | Q: {BB->getDataLayout(), TLI, nullptr, nullptr, New})) { |
| 2713 | ValueMapping[&*BI] = IV; |
| 2714 | if (!New->mayHaveSideEffects()) { |
| 2715 | New->eraseFromParent(); |
| 2716 | New = nullptr; |
| 2717 | // Clone debug-info on the elided instruction to the destination |
| 2718 | // position. |
| 2719 | OldPredBranch->cloneDebugInfoFrom(From: &*BI, FromHere: std::nullopt, InsertAtHead: true); |
| 2720 | } |
| 2721 | } else { |
| 2722 | ValueMapping[&*BI] = New; |
| 2723 | } |
| 2724 | if (New) { |
| 2725 | // Otherwise, insert the new instruction into the block. |
| 2726 | New->setName(BI->getName()); |
| 2727 | // Clone across any debug-info attached to the old instruction. |
| 2728 | New->cloneDebugInfoFrom(From: &*BI); |
| 2729 | // Update Dominance from simplified New instruction operands. |
| 2730 | for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) |
| 2731 | if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(Val: New->getOperand(i))) |
| 2732 | Updates.push_back(x: {DominatorTree::Insert, PredBB, SuccBB}); |
| 2733 | } |
| 2734 | } |
| 2735 | |
| 2736 | // Check to see if the targets of the branch had PHI nodes. If so, we need to |
| 2737 | // add entries to the PHI nodes for branch from PredBB now. |
| 2738 | CondBrInst *BBBranch = cast<CondBrInst>(Val: BB->getTerminator()); |
| 2739 | addPHINodeEntriesForMappedBlock(PHIBB: BBBranch->getSuccessor(i: 0), OldPred: BB, NewPred: PredBB, |
| 2740 | ValueMap&: ValueMapping); |
| 2741 | addPHINodeEntriesForMappedBlock(PHIBB: BBBranch->getSuccessor(i: 1), OldPred: BB, NewPred: PredBB, |
| 2742 | ValueMap&: ValueMapping); |
| 2743 | |
| 2744 | // KeyInstructions: Remap the cloned instructions' atoms only. |
| 2745 | remapSourceAtoms(VM&: ValueMapping, Begin: std::prev(x: RItBeforeInsertPt)->getIterator(), |
| 2746 | End: OldPredBranch->getIterator()); |
| 2747 | |
| 2748 | updateSSA(BB, NewBB: PredBB, ValueMapping); |
| 2749 | |
| 2750 | // PredBB no longer jumps to BB, remove entries in the PHI node for the edge |
| 2751 | // that we nuked. |
| 2752 | BB->removePredecessor(Pred: PredBB, KeepOneInputPHIs: true); |
| 2753 | |
| 2754 | // Remove the unconditional branch at the end of the PredBB block. |
| 2755 | OldPredBranch->eraseFromParent(); |
| 2756 | if (auto *BPI = getBPI()) |
| 2757 | BPI->copyEdgeProbabilities(Src: BB, Dst: PredBB); |
| 2758 | DTU->applyUpdatesPermissive(Updates); |
| 2759 | |
| 2760 | ++NumDupes; |
| 2761 | return true; |
| 2762 | } |
| 2763 | |
| 2764 | // Pred is a predecessor of BB with an unconditional branch to BB. SI is |
| 2765 | // a Select instruction in Pred. BB has other predecessors and SI is used in |
| 2766 | // a PHI node in BB. SI has no other use. |
| 2767 | // A new basic block, NewBB, is created and SI is converted to compare and |
| 2768 | // conditional branch. SI is erased from parent. |
| 2769 | void JumpThreadingPass::unfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB, |
| 2770 | SelectInst *SI, PHINode *SIUse, |
| 2771 | unsigned Idx) { |
| 2772 | // Expand the select. |
| 2773 | // |
| 2774 | // Pred -- |
| 2775 | // | v |
| 2776 | // | NewBB |
| 2777 | // | | |
| 2778 | // |----- |
| 2779 | // v |
| 2780 | // BB |
| 2781 | UncondBrInst *PredTerm = cast<UncondBrInst>(Val: Pred->getTerminator()); |
| 2782 | BasicBlock *NewBB = BasicBlock::Create(Context&: BB->getContext(), Name: "select.unfold" , |
| 2783 | Parent: BB->getParent(), InsertBefore: BB); |
| 2784 | // Move the unconditional branch to NewBB. |
| 2785 | PredTerm->removeFromParent(); |
| 2786 | PredTerm->insertInto(ParentBB: NewBB, It: NewBB->end()); |
| 2787 | // Create a conditional branch and update PHI nodes. |
| 2788 | auto *BI = CondBrInst::Create(Cond: SI->getCondition(), IfTrue: NewBB, IfFalse: BB, InsertBefore: Pred); |
| 2789 | BI->applyMergedLocation(LocA: PredTerm->getDebugLoc(), LocB: SI->getDebugLoc()); |
| 2790 | BI->copyMetadata(SrcInst: *SI, WL: {LLVMContext::MD_prof}); |
| 2791 | SIUse->setIncomingValue(i: Idx, V: SI->getFalseValue()); |
| 2792 | SIUse->addIncoming(V: SI->getTrueValue(), BB: NewBB); |
| 2793 | |
| 2794 | uint64_t TrueWeight = 1; |
| 2795 | uint64_t FalseWeight = 1; |
| 2796 | // Copy probabilities from 'SI' to created conditional branch in 'Pred'. |
| 2797 | if (extractBranchWeights(I: *SI, TrueVal&: TrueWeight, FalseVal&: FalseWeight) && |
| 2798 | (TrueWeight + FalseWeight) != 0) { |
| 2799 | SmallVector<BranchProbability, 2> BP; |
| 2800 | BP.emplace_back(Args: BranchProbability::getBranchProbability( |
| 2801 | Numerator: TrueWeight, Denominator: TrueWeight + FalseWeight)); |
| 2802 | BP.emplace_back(Args: BranchProbability::getBranchProbability( |
| 2803 | Numerator: FalseWeight, Denominator: TrueWeight + FalseWeight)); |
| 2804 | // Update BPI if exists. |
| 2805 | if (auto *BPI = getBPI()) |
| 2806 | BPI->setEdgeProbability(Src: Pred, Probs: BP); |
| 2807 | } |
| 2808 | // Set the block frequency of NewBB. |
| 2809 | if (auto *BFI = getBFI()) { |
| 2810 | if ((TrueWeight + FalseWeight) == 0) { |
| 2811 | TrueWeight = 1; |
| 2812 | FalseWeight = 1; |
| 2813 | } |
| 2814 | BranchProbability PredToNewBBProb = BranchProbability::getBranchProbability( |
| 2815 | Numerator: TrueWeight, Denominator: TrueWeight + FalseWeight); |
| 2816 | auto NewBBFreq = BFI->getBlockFreq(BB: Pred) * PredToNewBBProb; |
| 2817 | BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq); |
| 2818 | } |
| 2819 | |
| 2820 | // The select is now dead. |
| 2821 | SI->eraseFromParent(); |
| 2822 | DTU->applyUpdatesPermissive(Updates: {{DominatorTree::Insert, NewBB, BB}, |
| 2823 | {DominatorTree::Insert, Pred, NewBB}}); |
| 2824 | |
| 2825 | // Update any other PHI nodes in BB. |
| 2826 | for (BasicBlock::iterator BI = BB->begin(); |
| 2827 | PHINode *Phi = dyn_cast<PHINode>(Val&: BI); ++BI) |
| 2828 | if (Phi != SIUse) |
| 2829 | Phi->addIncoming(V: Phi->getIncomingValueForBlock(BB: Pred), BB: NewBB); |
| 2830 | } |
| 2831 | |
| 2832 | bool JumpThreadingPass::tryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) { |
| 2833 | PHINode *CondPHI = dyn_cast<PHINode>(Val: SI->getCondition()); |
| 2834 | |
| 2835 | if (!CondPHI || CondPHI->getParent() != BB) |
| 2836 | return false; |
| 2837 | |
| 2838 | for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) { |
| 2839 | BasicBlock *Pred = CondPHI->getIncomingBlock(i: I); |
| 2840 | SelectInst *PredSI = dyn_cast<SelectInst>(Val: CondPHI->getIncomingValue(i: I)); |
| 2841 | |
| 2842 | // The second and third condition can be potentially relaxed. Currently |
| 2843 | // the conditions help to simplify the code and allow us to reuse existing |
| 2844 | // code, developed for tryToUnfoldSelect(CmpInst *, BasicBlock *) |
| 2845 | if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse()) |
| 2846 | continue; |
| 2847 | |
| 2848 | UncondBrInst *PredTerm = dyn_cast<UncondBrInst>(Val: Pred->getTerminator()); |
| 2849 | if (!PredTerm) |
| 2850 | continue; |
| 2851 | |
| 2852 | unfoldSelectInstr(Pred, BB, SI: PredSI, SIUse: CondPHI, Idx: I); |
| 2853 | return true; |
| 2854 | } |
| 2855 | return false; |
| 2856 | } |
| 2857 | |
| 2858 | /// tryToUnfoldSelect - Look for blocks of the form |
| 2859 | /// bb1: |
| 2860 | /// %a = select |
| 2861 | /// br bb2 |
| 2862 | /// |
| 2863 | /// bb2: |
| 2864 | /// %p = phi [%a, %bb1] ... |
| 2865 | /// %c = icmp %p |
| 2866 | /// br i1 %c |
| 2867 | /// |
| 2868 | /// And expand the select into a branch structure if one of its arms allows %c |
| 2869 | /// to be folded. This later enables threading from bb1 over bb2. |
| 2870 | bool JumpThreadingPass::tryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) { |
| 2871 | CondBrInst *CondBr = dyn_cast<CondBrInst>(Val: BB->getTerminator()); |
| 2872 | PHINode *CondLHS = dyn_cast<PHINode>(Val: CondCmp->getOperand(i_nocapture: 0)); |
| 2873 | Constant *CondRHS = cast<Constant>(Val: CondCmp->getOperand(i_nocapture: 1)); |
| 2874 | |
| 2875 | if (!CondBr || !CondLHS || CondLHS->getParent() != BB) |
| 2876 | return false; |
| 2877 | |
| 2878 | for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) { |
| 2879 | BasicBlock *Pred = CondLHS->getIncomingBlock(i: I); |
| 2880 | SelectInst *SI = dyn_cast<SelectInst>(Val: CondLHS->getIncomingValue(i: I)); |
| 2881 | |
| 2882 | // Look if one of the incoming values is a select in the corresponding |
| 2883 | // predecessor. |
| 2884 | if (!SI || SI->getParent() != Pred || !SI->hasOneUse()) |
| 2885 | continue; |
| 2886 | |
| 2887 | UncondBrInst *PredTerm = dyn_cast<UncondBrInst>(Val: Pred->getTerminator()); |
| 2888 | if (!PredTerm) |
| 2889 | continue; |
| 2890 | |
| 2891 | // Now check if one of the select values would allow us to constant fold the |
| 2892 | // terminator in BB. We don't do the transform if both sides fold, those |
| 2893 | // cases will be threaded in any case. |
| 2894 | Constant *LHSRes = |
| 2895 | LVI->getPredicateOnEdge(Pred: CondCmp->getPredicate(), V: SI->getOperand(i_nocapture: 1), |
| 2896 | C: CondRHS, FromBB: Pred, ToBB: BB, CxtI: CondCmp); |
| 2897 | Constant *RHSRes = |
| 2898 | LVI->getPredicateOnEdge(Pred: CondCmp->getPredicate(), V: SI->getOperand(i_nocapture: 2), |
| 2899 | C: CondRHS, FromBB: Pred, ToBB: BB, CxtI: CondCmp); |
| 2900 | if ((LHSRes || RHSRes) && LHSRes != RHSRes) { |
| 2901 | unfoldSelectInstr(Pred, BB, SI, SIUse: CondLHS, Idx: I); |
| 2902 | return true; |
| 2903 | } |
| 2904 | } |
| 2905 | return false; |
| 2906 | } |
| 2907 | |
| 2908 | /// tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the |
| 2909 | /// same BB in the form |
| 2910 | /// bb: |
| 2911 | /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ... |
| 2912 | /// %s = select %p, trueval, falseval |
| 2913 | /// |
| 2914 | /// or |
| 2915 | /// |
| 2916 | /// bb: |
| 2917 | /// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ... |
| 2918 | /// %c = cmp %p, 0 |
| 2919 | /// %s = select %c, trueval, falseval |
| 2920 | /// |
| 2921 | /// And expand the select into a branch structure. This later enables |
| 2922 | /// jump-threading over bb in this pass. |
| 2923 | /// |
| 2924 | /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold |
| 2925 | /// select if the associated PHI has at least one constant. If the unfolded |
| 2926 | /// select is not jump-threaded, it will be folded again in the later |
| 2927 | /// optimizations. |
| 2928 | bool JumpThreadingPass::tryToUnfoldSelectInCurrBB(BasicBlock *BB) { |
| 2929 | // This transform would reduce the quality of msan diagnostics. |
| 2930 | // Disable this transform under MemorySanitizer. |
| 2931 | if (BB->getParent()->hasFnAttribute(Kind: Attribute::SanitizeMemory)) |
| 2932 | return false; |
| 2933 | |
| 2934 | // If threading this would thread across a loop header, don't thread the edge. |
| 2935 | // See the comments above findLoopHeaders for justifications and caveats. |
| 2936 | if (LoopHeaders.count(Ptr: BB)) |
| 2937 | return false; |
| 2938 | |
| 2939 | for (BasicBlock::iterator BI = BB->begin(); |
| 2940 | PHINode *PN = dyn_cast<PHINode>(Val&: BI); ++BI) { |
| 2941 | // Look for a Phi having at least one constant incoming value. |
| 2942 | if (llvm::all_of(Range: PN->incoming_values(), |
| 2943 | P: [](Value *V) { return !isa<ConstantInt>(Val: V); })) |
| 2944 | continue; |
| 2945 | |
| 2946 | auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) { |
| 2947 | using namespace PatternMatch; |
| 2948 | |
| 2949 | // Check if SI is in BB and use V as condition. |
| 2950 | if (SI->getParent() != BB) |
| 2951 | return false; |
| 2952 | Value *Cond = SI->getCondition(); |
| 2953 | bool IsAndOr = match(V: SI, P: m_CombineOr(L: m_LogicalAnd(), R: m_LogicalOr())); |
| 2954 | return Cond && Cond == V && Cond->getType()->isIntegerTy(Bitwidth: 1) && !IsAndOr; |
| 2955 | }; |
| 2956 | |
| 2957 | SelectInst *SI = nullptr; |
| 2958 | for (Use &U : PN->uses()) { |
| 2959 | if (ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: U.getUser())) { |
| 2960 | // Look for a ICmp in BB that compares PN with a constant and is the |
| 2961 | // condition of a Select. |
| 2962 | if (Cmp->getParent() == BB && Cmp->hasOneUse() && |
| 2963 | isa<ConstantInt>(Val: Cmp->getOperand(i_nocapture: 1 - U.getOperandNo()))) |
| 2964 | if (SelectInst *SelectI = dyn_cast<SelectInst>(Val: Cmp->user_back())) |
| 2965 | if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) { |
| 2966 | SI = SelectI; |
| 2967 | break; |
| 2968 | } |
| 2969 | } else if (SelectInst *SelectI = dyn_cast<SelectInst>(Val: U.getUser())) { |
| 2970 | // Look for a Select in BB that uses PN as condition. |
| 2971 | if (isUnfoldCandidate(SelectI, U.get())) { |
| 2972 | SI = SelectI; |
| 2973 | break; |
| 2974 | } |
| 2975 | } |
| 2976 | } |
| 2977 | |
| 2978 | if (!SI) |
| 2979 | continue; |
| 2980 | // Expand the select. |
| 2981 | Value *Cond = SI->getCondition(); |
| 2982 | if (!isGuaranteedNotToBeUndefOrPoison(V: Cond, AC: nullptr, CtxI: SI)) { |
| 2983 | Cond = new FreezeInst(Cond, "cond.fr" , SI->getIterator()); |
| 2984 | cast<FreezeInst>(Val: Cond)->setDebugLoc(DebugLoc::getTemporary()); |
| 2985 | } |
| 2986 | MDNode *BranchWeights = getBranchWeightMDNode(I: *SI); |
| 2987 | Instruction *Term = |
| 2988 | SplitBlockAndInsertIfThen(Cond, SplitBefore: SI, Unreachable: false, BranchWeights); |
| 2989 | BasicBlock *SplitBB = SI->getParent(); |
| 2990 | BasicBlock *NewBB = Term->getParent(); |
| 2991 | PHINode *NewPN = PHINode::Create(Ty: SI->getType(), NumReservedValues: 2, NameStr: "" , InsertBefore: SI->getIterator()); |
| 2992 | NewPN->addIncoming(V: SI->getTrueValue(), BB: Term->getParent()); |
| 2993 | NewPN->addIncoming(V: SI->getFalseValue(), BB); |
| 2994 | NewPN->setDebugLoc(SI->getDebugLoc()); |
| 2995 | SI->replaceAllUsesWith(V: NewPN); |
| 2996 | SI->eraseFromParent(); |
| 2997 | // NewBB and SplitBB are newly created blocks which require insertion. |
| 2998 | std::vector<DominatorTree::UpdateType> Updates; |
| 2999 | Updates.reserve(n: (2 * SplitBB->getTerminator()->getNumSuccessors()) + 3); |
| 3000 | Updates.push_back(x: {DominatorTree::Insert, BB, SplitBB}); |
| 3001 | Updates.push_back(x: {DominatorTree::Insert, BB, NewBB}); |
| 3002 | Updates.push_back(x: {DominatorTree::Insert, NewBB, SplitBB}); |
| 3003 | // BB's successors were moved to SplitBB, update DTU accordingly. |
| 3004 | for (auto *Succ : successors(BB: SplitBB)) { |
| 3005 | Updates.push_back(x: {DominatorTree::Delete, BB, Succ}); |
| 3006 | Updates.push_back(x: {DominatorTree::Insert, SplitBB, Succ}); |
| 3007 | } |
| 3008 | DTU->applyUpdatesPermissive(Updates); |
| 3009 | return true; |
| 3010 | } |
| 3011 | return false; |
| 3012 | } |
| 3013 | |
| 3014 | /// Try to propagate a guard from the current BB into one of its predecessors |
| 3015 | /// in case if another branch of execution implies that the condition of this |
| 3016 | /// guard is always true. Currently we only process the simplest case that |
| 3017 | /// looks like: |
| 3018 | /// |
| 3019 | /// Start: |
| 3020 | /// %cond = ... |
| 3021 | /// br i1 %cond, label %T1, label %F1 |
| 3022 | /// T1: |
| 3023 | /// br label %Merge |
| 3024 | /// F1: |
| 3025 | /// br label %Merge |
| 3026 | /// Merge: |
| 3027 | /// %condGuard = ... |
| 3028 | /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ] |
| 3029 | /// |
| 3030 | /// And cond either implies condGuard or !condGuard. In this case all the |
| 3031 | /// instructions before the guard can be duplicated in both branches, and the |
| 3032 | /// guard is then threaded to one of them. |
| 3033 | bool JumpThreadingPass::processGuards(BasicBlock *BB) { |
| 3034 | using namespace PatternMatch; |
| 3035 | |
| 3036 | // We only want to deal with two predecessors. |
| 3037 | BasicBlock *Pred1, *Pred2; |
| 3038 | auto PI = pred_begin(BB), PE = pred_end(BB); |
| 3039 | if (PI == PE) |
| 3040 | return false; |
| 3041 | Pred1 = *PI++; |
| 3042 | if (PI == PE) |
| 3043 | return false; |
| 3044 | Pred2 = *PI++; |
| 3045 | if (PI != PE) |
| 3046 | return false; |
| 3047 | if (Pred1 == Pred2) |
| 3048 | return false; |
| 3049 | |
| 3050 | // Try to thread one of the guards of the block. |
| 3051 | // TODO: Look up deeper than to immediate predecessor? |
| 3052 | auto *Parent = Pred1->getSinglePredecessor(); |
| 3053 | if (!Parent || Parent != Pred2->getSinglePredecessor()) |
| 3054 | return false; |
| 3055 | |
| 3056 | if (auto *BI = dyn_cast<CondBrInst>(Val: Parent->getTerminator())) |
| 3057 | for (auto &I : *BB) |
| 3058 | if (isGuard(U: &I) && threadGuard(BB, Guard: cast<IntrinsicInst>(Val: &I), BI)) |
| 3059 | return true; |
| 3060 | |
| 3061 | return false; |
| 3062 | } |
| 3063 | |
| 3064 | /// Try to propagate the guard from BB which is the lower block of a diamond |
| 3065 | /// to one of its branches, in case if diamond's condition implies guard's |
| 3066 | /// condition. |
| 3067 | bool JumpThreadingPass::threadGuard(BasicBlock *BB, IntrinsicInst *Guard, |
| 3068 | CondBrInst *BI) { |
| 3069 | Value *GuardCond = Guard->getArgOperand(i: 0); |
| 3070 | Value *BranchCond = BI->getCondition(); |
| 3071 | BasicBlock *TrueDest = BI->getSuccessor(i: 0); |
| 3072 | BasicBlock *FalseDest = BI->getSuccessor(i: 1); |
| 3073 | |
| 3074 | auto &DL = BB->getDataLayout(); |
| 3075 | bool TrueDestIsSafe = false; |
| 3076 | bool FalseDestIsSafe = false; |
| 3077 | |
| 3078 | // True dest is safe if BranchCond => GuardCond. |
| 3079 | auto Impl = isImpliedCondition(LHS: BranchCond, RHS: GuardCond, DL); |
| 3080 | if (Impl && *Impl) |
| 3081 | TrueDestIsSafe = true; |
| 3082 | else { |
| 3083 | // False dest is safe if !BranchCond => GuardCond. |
| 3084 | Impl = isImpliedCondition(LHS: BranchCond, RHS: GuardCond, DL, /* LHSIsTrue */ false); |
| 3085 | if (Impl && *Impl) |
| 3086 | FalseDestIsSafe = true; |
| 3087 | } |
| 3088 | |
| 3089 | if (!TrueDestIsSafe && !FalseDestIsSafe) |
| 3090 | return false; |
| 3091 | |
| 3092 | BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest; |
| 3093 | BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest; |
| 3094 | |
| 3095 | ValueToValueMapTy UnguardedMapping, GuardedMapping; |
| 3096 | Instruction *AfterGuard = Guard->getNextNode(); |
| 3097 | unsigned Cost = |
| 3098 | getJumpThreadDuplicationCost(TTI, BB, StopAt: AfterGuard, Threshold: BBDupThreshold); |
| 3099 | if (Cost > BBDupThreshold) |
| 3100 | return false; |
| 3101 | // Duplicate all instructions before the guard and the guard itself to the |
| 3102 | // branch where implication is not proved. |
| 3103 | BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween( |
| 3104 | BB, PredBB: PredGuardedBlock, StopAt: AfterGuard, ValueMapping&: GuardedMapping, DTU&: *DTU); |
| 3105 | assert(GuardedBlock && "Could not create the guarded block?" ); |
| 3106 | // Duplicate all instructions before the guard in the unguarded branch. |
| 3107 | // Since we have successfully duplicated the guarded block and this block |
| 3108 | // has fewer instructions, we expect it to succeed. |
| 3109 | BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween( |
| 3110 | BB, PredBB: PredUnguardedBlock, StopAt: Guard, ValueMapping&: UnguardedMapping, DTU&: *DTU); |
| 3111 | assert(UnguardedBlock && "Could not create the unguarded block?" ); |
| 3112 | LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block " |
| 3113 | << GuardedBlock->getName() << "\n" ); |
| 3114 | // Some instructions before the guard may still have uses. For them, we need |
| 3115 | // to create Phi nodes merging their copies in both guarded and unguarded |
| 3116 | // branches. Those instructions that have no uses can be just removed. |
| 3117 | SmallVector<Instruction *, 4> ToRemove; |
| 3118 | for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI) |
| 3119 | if (!isa<PHINode>(Val: &*BI)) |
| 3120 | ToRemove.push_back(Elt: &*BI); |
| 3121 | |
| 3122 | BasicBlock::iterator InsertionPoint = BB->getFirstInsertionPt(); |
| 3123 | assert(InsertionPoint != BB->end() && "Empty block?" ); |
| 3124 | // Substitute with Phis & remove. |
| 3125 | for (auto *Inst : reverse(C&: ToRemove)) { |
| 3126 | if (!Inst->use_empty()) { |
| 3127 | PHINode *NewPN = PHINode::Create(Ty: Inst->getType(), NumReservedValues: 2); |
| 3128 | NewPN->addIncoming(V: UnguardedMapping[Inst], BB: UnguardedBlock); |
| 3129 | NewPN->addIncoming(V: GuardedMapping[Inst], BB: GuardedBlock); |
| 3130 | NewPN->setDebugLoc(Inst->getDebugLoc()); |
| 3131 | NewPN->insertBefore(InsertPos: InsertionPoint); |
| 3132 | Inst->replaceAllUsesWith(V: NewPN); |
| 3133 | } |
| 3134 | Inst->dropDbgRecords(); |
| 3135 | Inst->eraseFromParent(); |
| 3136 | } |
| 3137 | return true; |
| 3138 | } |
| 3139 | |
| 3140 | PreservedAnalyses JumpThreadingPass::getPreservedAnalysis() const { |
| 3141 | PreservedAnalyses PA; |
| 3142 | PA.preserve<LazyValueAnalysis>(); |
| 3143 | PA.preserve<DominatorTreeAnalysis>(); |
| 3144 | |
| 3145 | // TODO: We would like to preserve BPI/BFI. Enable once all paths update them. |
| 3146 | // TODO: Would be nice to verify BPI/BFI consistency as well. |
| 3147 | return PA; |
| 3148 | } |
| 3149 | |
| 3150 | template <typename AnalysisT> |
| 3151 | typename AnalysisT::Result *JumpThreadingPass::runExternalAnalysis() { |
| 3152 | assert(FAM && "Can't run external analysis without FunctionAnalysisManager" ); |
| 3153 | |
| 3154 | // If there were no changes since last call to 'runExternalAnalysis' then all |
| 3155 | // analysis is either up to date or explicitly invalidated. Just go ahead and |
| 3156 | // run the "external" analysis. |
| 3157 | if (!ChangedSinceLastAnalysisUpdate) { |
| 3158 | assert(!DTU->hasPendingUpdates() && |
| 3159 | "Lost update of 'ChangedSinceLastAnalysisUpdate'?" ); |
| 3160 | // Run the "external" analysis. |
| 3161 | return &FAM->getResult<AnalysisT>(*F); |
| 3162 | } |
| 3163 | ChangedSinceLastAnalysisUpdate = false; |
| 3164 | |
| 3165 | auto PA = getPreservedAnalysis(); |
| 3166 | // TODO: This shouldn't be needed once 'getPreservedAnalysis' reports BPI/BFI |
| 3167 | // as preserved. |
| 3168 | PA.preserve<BranchProbabilityAnalysis>(); |
| 3169 | PA.preserve<BlockFrequencyAnalysis>(); |
| 3170 | // Report everything except explicitly preserved as invalid. |
| 3171 | FAM->invalidate(IR&: *F, PA); |
| 3172 | // Update DT/PDT. |
| 3173 | DTU->flush(); |
| 3174 | // Make sure DT/PDT are valid before running "external" analysis. |
| 3175 | assert(DTU->getDomTree().verify(DominatorTree::VerificationLevel::Fast)); |
| 3176 | assert((!DTU->hasPostDomTree() || |
| 3177 | DTU->getPostDomTree().verify( |
| 3178 | PostDominatorTree::VerificationLevel::Fast))); |
| 3179 | // Run the "external" analysis. |
| 3180 | auto *Result = &FAM->getResult<AnalysisT>(*F); |
| 3181 | // Update analysis JumpThreading depends on and not explicitly preserved. |
| 3182 | TTI = &FAM->getResult<TargetIRAnalysis>(IR&: *F); |
| 3183 | TLI = &FAM->getResult<TargetLibraryAnalysis>(IR&: *F); |
| 3184 | AA = &FAM->getResult<AAManager>(IR&: *F); |
| 3185 | |
| 3186 | return Result; |
| 3187 | } |
| 3188 | |
| 3189 | BranchProbabilityInfo *JumpThreadingPass::getBPI() { |
| 3190 | if (!BPI) { |
| 3191 | assert(FAM && "Can't create BPI without FunctionAnalysisManager" ); |
| 3192 | BPI = FAM->getCachedResult<BranchProbabilityAnalysis>(IR&: *F); |
| 3193 | } |
| 3194 | return BPI; |
| 3195 | } |
| 3196 | |
| 3197 | BlockFrequencyInfo *JumpThreadingPass::getBFI() { |
| 3198 | if (!BFI) { |
| 3199 | assert(FAM && "Can't create BFI without FunctionAnalysisManager" ); |
| 3200 | BFI = FAM->getCachedResult<BlockFrequencyAnalysis>(IR&: *F); |
| 3201 | } |
| 3202 | return BFI; |
| 3203 | } |
| 3204 | |
| 3205 | // Important note on validity of BPI/BFI. JumpThreading tries to preserve |
| 3206 | // BPI/BFI as it goes. Thus if cached instance exists it will be updated. |
| 3207 | // Otherwise, new instance of BPI/BFI is created (up to date by definition). |
| 3208 | BranchProbabilityInfo *JumpThreadingPass::getOrCreateBPI(bool Force) { |
| 3209 | auto *Res = getBPI(); |
| 3210 | if (Res) |
| 3211 | return Res; |
| 3212 | |
| 3213 | if (Force) |
| 3214 | BPI = runExternalAnalysis<BranchProbabilityAnalysis>(); |
| 3215 | |
| 3216 | return BPI; |
| 3217 | } |
| 3218 | |
| 3219 | BlockFrequencyInfo *JumpThreadingPass::getOrCreateBFI(bool Force) { |
| 3220 | auto *Res = getBFI(); |
| 3221 | if (Res) |
| 3222 | return Res; |
| 3223 | |
| 3224 | if (Force) |
| 3225 | BFI = runExternalAnalysis<BlockFrequencyAnalysis>(); |
| 3226 | |
| 3227 | return BFI; |
| 3228 | } |
| 3229 | |