| 1 | //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// |
| 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 pass performs loop invariant code motion, attempting to remove as much |
| 10 | // code from the body of a loop as possible. It does this by either hoisting |
| 11 | // code into the preheader block, or by sinking code to the exit blocks if it is |
| 12 | // safe. This pass also promotes must-aliased memory locations in the loop to |
| 13 | // live in registers, thus hoisting and sinking "invariant" loads and stores. |
| 14 | // |
| 15 | // Hoisting operations out of loops is a canonicalization transform. It |
| 16 | // enables and simplifies subsequent optimizations in the middle-end. |
| 17 | // Rematerialization of hoisted instructions to reduce register pressure is the |
| 18 | // responsibility of the back-end, which has more accurate information about |
| 19 | // register pressure and also handles other optimizations than LICM that |
| 20 | // increase live-ranges. |
| 21 | // |
| 22 | // This pass uses alias analysis for two purposes: |
| 23 | // |
| 24 | // 1. Moving loop invariant loads and calls out of loops. If we can determine |
| 25 | // that a load or call inside of a loop never aliases anything stored to, |
| 26 | // we can hoist it or sink it like any other instruction. |
| 27 | // 2. Scalar Promotion of Memory - If there is a store instruction inside of |
| 28 | // the loop, we try to move the store to happen AFTER the loop instead of |
| 29 | // inside of the loop. This can only happen if a few conditions are true: |
| 30 | // A. The pointer stored through is loop invariant |
| 31 | // B. There are no stores or loads in the loop which _may_ alias the |
| 32 | // pointer. There are no calls in the loop which mod/ref the pointer. |
| 33 | // If these conditions are true, we can promote the loads and stores in the |
| 34 | // loop of the pointer to use a temporary alloca'd variable. We then use |
| 35 | // the SSAUpdater to construct the appropriate SSA form for the value. |
| 36 | // |
| 37 | //===----------------------------------------------------------------------===// |
| 38 | |
| 39 | #include "llvm/Transforms/Scalar/LICM.h" |
| 40 | #include "llvm/ADT/PriorityWorklist.h" |
| 41 | #include "llvm/ADT/SetOperations.h" |
| 42 | #include "llvm/ADT/Statistic.h" |
| 43 | #include "llvm/Analysis/AliasAnalysis.h" |
| 44 | #include "llvm/Analysis/AliasSetTracker.h" |
| 45 | #include "llvm/Analysis/AssumptionCache.h" |
| 46 | #include "llvm/Analysis/CaptureTracking.h" |
| 47 | #include "llvm/Analysis/DomTreeUpdater.h" |
| 48 | #include "llvm/Analysis/GuardUtils.h" |
| 49 | #include "llvm/Analysis/LazyBlockFrequencyInfo.h" |
| 50 | #include "llvm/Analysis/Loads.h" |
| 51 | #include "llvm/Analysis/LoopInfo.h" |
| 52 | #include "llvm/Analysis/LoopIterator.h" |
| 53 | #include "llvm/Analysis/LoopNestAnalysis.h" |
| 54 | #include "llvm/Analysis/LoopPass.h" |
| 55 | #include "llvm/Analysis/MemorySSA.h" |
| 56 | #include "llvm/Analysis/MemorySSAUpdater.h" |
| 57 | #include "llvm/Analysis/MustExecute.h" |
| 58 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 59 | #include "llvm/Analysis/ScalarEvolution.h" |
| 60 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 61 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 62 | #include "llvm/Analysis/ValueTracking.h" |
| 63 | #include "llvm/IR/CFG.h" |
| 64 | #include "llvm/IR/Constants.h" |
| 65 | #include "llvm/IR/DataLayout.h" |
| 66 | #include "llvm/IR/DebugInfoMetadata.h" |
| 67 | #include "llvm/IR/DerivedTypes.h" |
| 68 | #include "llvm/IR/Dominators.h" |
| 69 | #include "llvm/IR/IRBuilder.h" |
| 70 | #include "llvm/IR/Instructions.h" |
| 71 | #include "llvm/IR/IntrinsicInst.h" |
| 72 | #include "llvm/IR/LLVMContext.h" |
| 73 | #include "llvm/IR/Metadata.h" |
| 74 | #include "llvm/IR/PatternMatch.h" |
| 75 | #include "llvm/IR/PredIteratorCache.h" |
| 76 | #include "llvm/InitializePasses.h" |
| 77 | #include "llvm/Support/CommandLine.h" |
| 78 | #include "llvm/Support/Debug.h" |
| 79 | #include "llvm/Support/raw_ostream.h" |
| 80 | #include "llvm/Transforms/Scalar.h" |
| 81 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" |
| 82 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
| 83 | #include "llvm/Transforms/Utils/Local.h" |
| 84 | #include "llvm/Transforms/Utils/LoopUtils.h" |
| 85 | #include "llvm/Transforms/Utils/SSAUpdater.h" |
| 86 | #include <algorithm> |
| 87 | #include <utility> |
| 88 | using namespace llvm; |
| 89 | |
| 90 | namespace llvm { |
| 91 | class LPMUpdater; |
| 92 | } // namespace llvm |
| 93 | |
| 94 | #define DEBUG_TYPE "licm" |
| 95 | |
| 96 | STATISTIC(NumCreatedBlocks, "Number of blocks created" ); |
| 97 | STATISTIC(NumClonedBranches, "Number of branches cloned" ); |
| 98 | STATISTIC(NumSunk, "Number of instructions sunk out of loop" ); |
| 99 | STATISTIC(NumHoisted, "Number of instructions hoisted out of loop" ); |
| 100 | STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk" ); |
| 101 | STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk" ); |
| 102 | STATISTIC(NumPromotionCandidates, "Number of promotion candidates" ); |
| 103 | STATISTIC(NumLoadPromoted, "Number of load-only promotions" ); |
| 104 | STATISTIC(NumLoadStorePromoted, "Number of load and store promotions" ); |
| 105 | STATISTIC(NumMinMaxHoisted, |
| 106 | "Number of min/max expressions hoisted out of the loop" ); |
| 107 | STATISTIC(NumGEPsHoisted, |
| 108 | "Number of geps reassociated and hoisted out of the loop" ); |
| 109 | STATISTIC(NumAddSubHoisted, "Number of add/subtract expressions reassociated " |
| 110 | "and hoisted out of the loop" ); |
| 111 | STATISTIC(NumFPAssociationsHoisted, "Number of invariant FP expressions " |
| 112 | "reassociated and hoisted out of the loop" ); |
| 113 | STATISTIC(NumIntAssociationsHoisted, |
| 114 | "Number of invariant int expressions " |
| 115 | "reassociated and hoisted out of the loop" ); |
| 116 | STATISTIC(NumBOAssociationsHoisted, "Number of invariant BinaryOp expressions " |
| 117 | "reassociated and hoisted out of the loop" ); |
| 118 | |
| 119 | /// Memory promotion is enabled by default. |
| 120 | static cl::opt<bool> |
| 121 | DisablePromotion("disable-licm-promotion" , cl::Hidden, cl::init(Val: false), |
| 122 | cl::desc("Disable memory promotion in LICM pass" )); |
| 123 | |
| 124 | static cl::opt<bool> ControlFlowHoisting( |
| 125 | "licm-control-flow-hoisting" , cl::Hidden, cl::init(Val: false), |
| 126 | cl::desc("Enable control flow (and PHI) hoisting in LICM" )); |
| 127 | |
| 128 | static cl::opt<bool> |
| 129 | SingleThread("licm-force-thread-model-single" , cl::Hidden, cl::init(Val: false), |
| 130 | cl::desc("Force thread model single in LICM pass" )); |
| 131 | |
| 132 | static cl::opt<uint32_t> MaxNumUsesTraversed( |
| 133 | "licm-max-num-uses-traversed" , cl::Hidden, cl::init(Val: 8), |
| 134 | cl::desc("Max num uses visited for identifying load " |
| 135 | "invariance in loop using invariant start (default = 8)" )); |
| 136 | |
| 137 | static cl::opt<unsigned> FPAssociationUpperLimit( |
| 138 | "licm-max-num-fp-reassociations" , cl::init(Val: 5U), cl::Hidden, |
| 139 | cl::desc( |
| 140 | "Set upper limit for the number of transformations performed " |
| 141 | "during a single round of hoisting the reassociated expressions." )); |
| 142 | |
| 143 | static cl::opt<unsigned> IntAssociationUpperLimit( |
| 144 | "licm-max-num-int-reassociations" , cl::init(Val: 5U), cl::Hidden, |
| 145 | cl::desc( |
| 146 | "Set upper limit for the number of transformations performed " |
| 147 | "during a single round of hoisting the reassociated expressions." )); |
| 148 | |
| 149 | // Experimental option to allow imprecision in LICM in pathological cases, in |
| 150 | // exchange for faster compile. This is to be removed if MemorySSA starts to |
| 151 | // address the same issue. LICM calls MemorySSAWalker's |
| 152 | // getClobberingMemoryAccess, up to the value of the Cap, getting perfect |
| 153 | // accuracy. Afterwards, LICM will call into MemorySSA's getDefiningAccess, |
| 154 | // which may not be precise, since optimizeUses is capped. The result is |
| 155 | // correct, but we may not get as "far up" as possible to get which access is |
| 156 | // clobbering the one queried. |
| 157 | cl::opt<unsigned> llvm::SetLicmMssaOptCap( |
| 158 | "licm-mssa-optimization-cap" , cl::init(Val: 100), cl::Hidden, |
| 159 | cl::desc("Enable imprecision in LICM in pathological cases, in exchange " |
| 160 | "for faster compile. Caps the MemorySSA clobbering calls." )); |
| 161 | |
| 162 | // Experimentally, memory promotion carries less importance than sinking and |
| 163 | // hoisting. Limit when we do promotion when using MemorySSA, in order to save |
| 164 | // compile time. |
| 165 | cl::opt<unsigned> llvm::SetLicmMssaNoAccForPromotionCap( |
| 166 | "licm-mssa-max-acc-promotion" , cl::init(Val: 250), cl::Hidden, |
| 167 | cl::desc("[LICM & MemorySSA] When MSSA in LICM is disabled, this has no " |
| 168 | "effect. When MSSA in LICM is enabled, then this is the maximum " |
| 169 | "number of accesses allowed to be present in a loop in order to " |
| 170 | "enable memory promotion." )); |
| 171 | |
| 172 | namespace llvm { |
| 173 | extern cl::opt<bool> ProfcheckDisableMetadataFixes; |
| 174 | } // end namespace llvm |
| 175 | |
| 176 | static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI); |
| 177 | static bool isNotUsedOrFoldableInLoop(const Instruction &I, const Loop *CurLoop, |
| 178 | const LoopSafetyInfo *SafetyInfo, |
| 179 | TargetTransformInfo *TTI, |
| 180 | bool &FoldableInLoop, bool LoopNestMode); |
| 181 | static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop, |
| 182 | BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo, |
| 183 | MemorySSAUpdater &MSSAU, ScalarEvolution *SE, |
| 184 | OptimizationRemarkEmitter *ORE); |
| 185 | static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT, |
| 186 | const Loop *CurLoop, ICFLoopSafetyInfo *SafetyInfo, |
| 187 | MemorySSAUpdater &MSSAU, OptimizationRemarkEmitter *ORE); |
| 188 | static bool isSafeToExecuteUnconditionally( |
| 189 | Instruction &Inst, const DominatorTree *DT, const TargetLibraryInfo *TLI, |
| 190 | const Loop *CurLoop, const LoopSafetyInfo *SafetyInfo, |
| 191 | OptimizationRemarkEmitter *ORE, const Instruction *CtxI, |
| 192 | AssumptionCache *AC, bool AllowSpeculation); |
| 193 | static bool noConflictingReadWrites(Instruction *I, MemorySSA *MSSA, |
| 194 | AAResults *AA, Loop *CurLoop, |
| 195 | SinkAndHoistLICMFlags &Flags); |
| 196 | static bool pointerInvalidatedByLoop(MemorySSA *MSSA, MemoryUse *MU, |
| 197 | Loop *CurLoop, Instruction &I, |
| 198 | SinkAndHoistLICMFlags &Flags, |
| 199 | bool InvariantGroup); |
| 200 | static bool pointerInvalidatedByBlock(BasicBlock &BB, MemorySSA &MSSA, |
| 201 | MemoryUse &MU); |
| 202 | /// Aggregates various functions for hoisting computations out of loop. |
| 203 | static bool hoistArithmetics(Instruction &I, Loop &L, |
| 204 | ICFLoopSafetyInfo &SafetyInfo, |
| 205 | MemorySSAUpdater &MSSAU, AssumptionCache *AC, |
| 206 | DominatorTree *DT); |
| 207 | static Instruction *cloneInstructionInExitBlock( |
| 208 | Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI, |
| 209 | const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater &MSSAU); |
| 210 | |
| 211 | static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo, |
| 212 | MemorySSAUpdater &MSSAU); |
| 213 | |
| 214 | static void moveInstructionBefore(Instruction &I, BasicBlock::iterator Dest, |
| 215 | ICFLoopSafetyInfo &SafetyInfo, |
| 216 | MemorySSAUpdater &MSSAU, ScalarEvolution *SE); |
| 217 | |
| 218 | static void foreachMemoryAccess(MemorySSA *MSSA, Loop *L, |
| 219 | function_ref<void(Instruction *)> Fn); |
| 220 | using PointersAndHasReadsOutsideSet = |
| 221 | std::pair<SmallSetVector<Value *, 8>, bool>; |
| 222 | static SmallVector<PointersAndHasReadsOutsideSet, 0> |
| 223 | collectPromotionCandidates(MemorySSA *MSSA, AliasAnalysis *AA, Loop *L); |
| 224 | |
| 225 | namespace { |
| 226 | struct LoopInvariantCodeMotion { |
| 227 | bool runOnLoop(Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT, |
| 228 | AssumptionCache *AC, TargetLibraryInfo *TLI, |
| 229 | TargetTransformInfo *TTI, ScalarEvolution *SE, MemorySSA *MSSA, |
| 230 | OptimizationRemarkEmitter *ORE, bool LoopNestMode = false); |
| 231 | |
| 232 | LoopInvariantCodeMotion(unsigned LicmMssaOptCap, |
| 233 | unsigned LicmMssaNoAccForPromotionCap, |
| 234 | bool LicmAllowSpeculation) |
| 235 | : LicmMssaOptCap(LicmMssaOptCap), |
| 236 | LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap), |
| 237 | LicmAllowSpeculation(LicmAllowSpeculation) {} |
| 238 | |
| 239 | private: |
| 240 | unsigned LicmMssaOptCap; |
| 241 | unsigned LicmMssaNoAccForPromotionCap; |
| 242 | bool LicmAllowSpeculation; |
| 243 | }; |
| 244 | |
| 245 | struct LegacyLICMPass : public LoopPass { |
| 246 | static char ID; // Pass identification, replacement for typeid |
| 247 | LegacyLICMPass( |
| 248 | unsigned LicmMssaOptCap = SetLicmMssaOptCap, |
| 249 | unsigned LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap, |
| 250 | bool LicmAllowSpeculation = true) |
| 251 | : LoopPass(ID), LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap, |
| 252 | LicmAllowSpeculation) { |
| 253 | initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry()); |
| 254 | } |
| 255 | |
| 256 | bool runOnLoop(Loop *L, LPPassManager &LPM) override { |
| 257 | if (skipLoop(L)) |
| 258 | return false; |
| 259 | |
| 260 | LLVM_DEBUG(dbgs() << "Perform LICM on Loop with header at block " |
| 261 | << L->getHeader()->getNameOrAsOperand() << "\n" ); |
| 262 | |
| 263 | Function *F = L->getHeader()->getParent(); |
| 264 | |
| 265 | auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>(); |
| 266 | MemorySSA *MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA(); |
| 267 | // For the old PM, we can't use OptimizationRemarkEmitter as an analysis |
| 268 | // pass. Function analyses need to be preserved across loop transformations |
| 269 | // but ORE cannot be preserved (see comment before the pass definition). |
| 270 | OptimizationRemarkEmitter ORE(L->getHeader()->getParent()); |
| 271 | return LICM.runOnLoop( |
| 272 | L, AA: &getAnalysis<AAResultsWrapperPass>().getAAResults(), |
| 273 | LI: &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(), |
| 274 | DT: &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), |
| 275 | AC: &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F&: *F), |
| 276 | TLI: &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F: *F), |
| 277 | TTI: &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F: *F), |
| 278 | SE: SE ? &SE->getSE() : nullptr, MSSA, ORE: &ORE); |
| 279 | } |
| 280 | |
| 281 | /// This transformation requires natural loop information & requires that |
| 282 | /// loop preheaders be inserted into the CFG... |
| 283 | /// |
| 284 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 285 | AU.addPreserved<DominatorTreeWrapperPass>(); |
| 286 | AU.addPreserved<LoopInfoWrapperPass>(); |
| 287 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
| 288 | AU.addRequired<MemorySSAWrapperPass>(); |
| 289 | AU.addPreserved<MemorySSAWrapperPass>(); |
| 290 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
| 291 | AU.addRequired<AssumptionCacheTracker>(); |
| 292 | getLoopAnalysisUsage(AU); |
| 293 | LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU); |
| 294 | AU.addPreserved<LazyBlockFrequencyInfoPass>(); |
| 295 | AU.addPreserved<LazyBranchProbabilityInfoPass>(); |
| 296 | } |
| 297 | |
| 298 | private: |
| 299 | LoopInvariantCodeMotion LICM; |
| 300 | }; |
| 301 | } // namespace |
| 302 | |
| 303 | PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM, |
| 304 | LoopStandardAnalysisResults &AR, LPMUpdater &) { |
| 305 | if (!AR.MSSA) |
| 306 | reportFatalUsageError(reason: "LICM requires MemorySSA (loop-mssa)" ); |
| 307 | |
| 308 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis |
| 309 | // pass. Function analyses need to be preserved across loop transformations |
| 310 | // but ORE cannot be preserved (see comment before the pass definition). |
| 311 | OptimizationRemarkEmitter ORE(L.getHeader()->getParent()); |
| 312 | |
| 313 | LoopInvariantCodeMotion LICM(Opts.MssaOptCap, Opts.MssaNoAccForPromotionCap, |
| 314 | Opts.AllowSpeculation); |
| 315 | if (!LICM.runOnLoop(L: &L, AA: &AR.AA, LI: &AR.LI, DT: &AR.DT, AC: &AR.AC, TLI: &AR.TLI, TTI: &AR.TTI, |
| 316 | SE: &AR.SE, MSSA: AR.MSSA, ORE: &ORE)) |
| 317 | return PreservedAnalyses::all(); |
| 318 | |
| 319 | auto PA = getLoopPassPreservedAnalyses(); |
| 320 | PA.preserve<MemorySSAAnalysis>(); |
| 321 | |
| 322 | return PA; |
| 323 | } |
| 324 | |
| 325 | void LICMPass::printPipeline( |
| 326 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
| 327 | static_cast<PassInfoMixin<LICMPass> *>(this)->printPipeline( |
| 328 | OS, MapClassName2PassName); |
| 329 | |
| 330 | OS << '<'; |
| 331 | OS << (Opts.AllowSpeculation ? "" : "no-" ) << "allowspeculation" ; |
| 332 | OS << '>'; |
| 333 | } |
| 334 | |
| 335 | PreservedAnalyses LNICMPass::run(LoopNest &LN, LoopAnalysisManager &AM, |
| 336 | LoopStandardAnalysisResults &AR, |
| 337 | LPMUpdater &) { |
| 338 | if (!AR.MSSA) |
| 339 | reportFatalUsageError(reason: "LNICM requires MemorySSA (loop-mssa)" ); |
| 340 | |
| 341 | // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis |
| 342 | // pass. Function analyses need to be preserved across loop transformations |
| 343 | // but ORE cannot be preserved (see comment before the pass definition). |
| 344 | OptimizationRemarkEmitter ORE(LN.getParent()); |
| 345 | |
| 346 | LoopInvariantCodeMotion LICM(Opts.MssaOptCap, Opts.MssaNoAccForPromotionCap, |
| 347 | Opts.AllowSpeculation); |
| 348 | |
| 349 | Loop &OutermostLoop = LN.getOutermostLoop(); |
| 350 | bool Changed = LICM.runOnLoop(L: &OutermostLoop, AA: &AR.AA, LI: &AR.LI, DT: &AR.DT, AC: &AR.AC, |
| 351 | TLI: &AR.TLI, TTI: &AR.TTI, SE: &AR.SE, MSSA: AR.MSSA, ORE: &ORE, LoopNestMode: true); |
| 352 | |
| 353 | if (!Changed) |
| 354 | return PreservedAnalyses::all(); |
| 355 | |
| 356 | auto PA = getLoopPassPreservedAnalyses(); |
| 357 | |
| 358 | PA.preserve<DominatorTreeAnalysis>(); |
| 359 | PA.preserve<LoopAnalysis>(); |
| 360 | PA.preserve<MemorySSAAnalysis>(); |
| 361 | |
| 362 | return PA; |
| 363 | } |
| 364 | |
| 365 | void LNICMPass::printPipeline( |
| 366 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
| 367 | static_cast<PassInfoMixin<LNICMPass> *>(this)->printPipeline( |
| 368 | OS, MapClassName2PassName); |
| 369 | |
| 370 | OS << '<'; |
| 371 | OS << (Opts.AllowSpeculation ? "" : "no-" ) << "allowspeculation" ; |
| 372 | OS << '>'; |
| 373 | } |
| 374 | |
| 375 | char LegacyLICMPass::ID = 0; |
| 376 | INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm" , "Loop Invariant Code Motion" , |
| 377 | false, false) |
| 378 | INITIALIZE_PASS_DEPENDENCY(LoopPass) |
| 379 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
| 380 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) |
| 381 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) |
| 382 | INITIALIZE_PASS_DEPENDENCY(LazyBFIPass) |
| 383 | INITIALIZE_PASS_END(LegacyLICMPass, "licm" , "Loop Invariant Code Motion" , false, |
| 384 | false) |
| 385 | |
| 386 | Pass *llvm::createLICMPass() { return new LegacyLICMPass(); } |
| 387 | |
| 388 | llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags(bool IsSink, Loop &L, |
| 389 | MemorySSA &MSSA) |
| 390 | : SinkAndHoistLICMFlags(SetLicmMssaOptCap, SetLicmMssaNoAccForPromotionCap, |
| 391 | IsSink, L, MSSA) {} |
| 392 | |
| 393 | llvm::SinkAndHoistLICMFlags::SinkAndHoistLICMFlags( |
| 394 | unsigned LicmMssaOptCap, unsigned LicmMssaNoAccForPromotionCap, bool IsSink, |
| 395 | Loop &L, MemorySSA &MSSA) |
| 396 | : LicmMssaOptCap(LicmMssaOptCap), |
| 397 | LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap), |
| 398 | IsSink(IsSink) { |
| 399 | unsigned AccessCapCount = 0; |
| 400 | for (auto *BB : L.getBlocks()) |
| 401 | if (const auto *Accesses = MSSA.getBlockAccesses(BB)) |
| 402 | for (const auto &MA : *Accesses) { |
| 403 | (void)MA; |
| 404 | ++AccessCapCount; |
| 405 | if (AccessCapCount > LicmMssaNoAccForPromotionCap) { |
| 406 | NoOfMemAccTooLarge = true; |
| 407 | return; |
| 408 | } |
| 409 | } |
| 410 | } |
| 411 | |
| 412 | /// Hoist expressions out of the specified loop. Note, alias info for inner |
| 413 | /// loop is not preserved so it is not a good idea to run LICM multiple |
| 414 | /// times on one loop. |
| 415 | bool LoopInvariantCodeMotion::runOnLoop(Loop *L, AAResults *AA, LoopInfo *LI, |
| 416 | DominatorTree *DT, AssumptionCache *AC, |
| 417 | TargetLibraryInfo *TLI, |
| 418 | TargetTransformInfo *TTI, |
| 419 | ScalarEvolution *SE, MemorySSA *MSSA, |
| 420 | OptimizationRemarkEmitter *ORE, |
| 421 | bool LoopNestMode) { |
| 422 | bool Changed = false; |
| 423 | |
| 424 | assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form." ); |
| 425 | |
| 426 | // If this loop has metadata indicating that LICM is not to be performed then |
| 427 | // just exit. |
| 428 | if (hasDisableLICMTransformsHint(L)) { |
| 429 | return false; |
| 430 | } |
| 431 | |
| 432 | // Don't sink stores from loops with coroutine suspend instructions. |
| 433 | // LICM would sink instructions into the default destination of |
| 434 | // the coroutine switch. The default destination of the switch is to |
| 435 | // handle the case where the coroutine is suspended, by which point the |
| 436 | // coroutine frame may have been destroyed. No instruction can be sunk there. |
| 437 | // FIXME: This would unfortunately hurt the performance of coroutines, however |
| 438 | // there is currently no general solution for this. Similar issues could also |
| 439 | // potentially happen in other passes where instructions are being moved |
| 440 | // across that edge. |
| 441 | bool HasCoroSuspendInst = llvm::any_of(Range: L->getBlocks(), P: [](BasicBlock *BB) { |
| 442 | using namespace PatternMatch; |
| 443 | return any_of(Range: make_pointer_range(Range&: *BB), |
| 444 | P: match_fn(P: m_Intrinsic<Intrinsic::coro_suspend>())); |
| 445 | }); |
| 446 | |
| 447 | MemorySSAUpdater MSSAU(MSSA); |
| 448 | SinkAndHoistLICMFlags Flags(LicmMssaOptCap, LicmMssaNoAccForPromotionCap, |
| 449 | /*IsSink=*/true, *L, *MSSA); |
| 450 | |
| 451 | // Get the preheader block to move instructions into... |
| 452 | BasicBlock * = L->getLoopPreheader(); |
| 453 | |
| 454 | // Compute loop safety information. |
| 455 | ICFLoopSafetyInfo SafetyInfo; |
| 456 | SafetyInfo.computeLoopSafetyInfo(CurLoop: L); |
| 457 | |
| 458 | // We want to visit all of the instructions in this loop... that are not parts |
| 459 | // of our subloops (they have already had their invariants hoisted out of |
| 460 | // their loop, into this loop, so there is no need to process the BODIES of |
| 461 | // the subloops). |
| 462 | // |
| 463 | // Traverse the body of the loop in depth first order on the dominator tree so |
| 464 | // that we are guaranteed to see definitions before we see uses. This allows |
| 465 | // us to sink instructions in one pass, without iteration. After sinking |
| 466 | // instructions, we perform another pass to hoist them out of the loop. |
| 467 | if (L->hasDedicatedExits()) |
| 468 | Changed |= |
| 469 | LoopNestMode |
| 470 | ? sinkRegionForLoopNest(DT->getNode(BB: L->getHeader()), AA, LI, DT, |
| 471 | TLI, TTI, L, MSSAU, &SafetyInfo, Flags, ORE) |
| 472 | : sinkRegion(DT->getNode(BB: L->getHeader()), AA, LI, DT, TLI, TTI, CurLoop: L, |
| 473 | MSSAU, &SafetyInfo, Flags, ORE); |
| 474 | Flags.setIsSink(false); |
| 475 | if (Preheader) |
| 476 | Changed |= hoistRegion(DT->getNode(BB: L->getHeader()), AA, LI, DT, AC, TLI, L, |
| 477 | MSSAU, SE, &SafetyInfo, Flags, ORE, LoopNestMode, |
| 478 | AllowSpeculation: LicmAllowSpeculation); |
| 479 | |
| 480 | // Now that all loop invariants have been removed from the loop, promote any |
| 481 | // memory references to scalars that we can. |
| 482 | // Don't sink stores from loops without dedicated block exits. Exits |
| 483 | // containing indirect branches are not transformed by loop simplify, |
| 484 | // make sure we catch that. An additional load may be generated in the |
| 485 | // preheader for SSA updater, so also avoid sinking when no preheader |
| 486 | // is available. |
| 487 | if (!DisablePromotion && Preheader && L->hasDedicatedExits() && |
| 488 | !Flags.tooManyMemoryAccesses() && !HasCoroSuspendInst) { |
| 489 | // Figure out the loop exits and their insertion points |
| 490 | SmallVector<BasicBlock *, 8> ExitBlocks; |
| 491 | L->getUniqueExitBlocks(ExitBlocks); |
| 492 | |
| 493 | // We can't insert into a catchswitch. |
| 494 | bool HasCatchSwitch = llvm::any_of(Range&: ExitBlocks, P: [](BasicBlock *Exit) { |
| 495 | return isa<CatchSwitchInst>(Val: Exit->getTerminator()); |
| 496 | }); |
| 497 | |
| 498 | if (!HasCatchSwitch) { |
| 499 | SmallVector<BasicBlock::iterator, 8> InsertPts; |
| 500 | SmallVector<MemoryAccess *, 8> MSSAInsertPts; |
| 501 | InsertPts.reserve(N: ExitBlocks.size()); |
| 502 | MSSAInsertPts.reserve(N: ExitBlocks.size()); |
| 503 | for (BasicBlock *ExitBlock : ExitBlocks) { |
| 504 | InsertPts.push_back(Elt: ExitBlock->getFirstInsertionPt()); |
| 505 | MSSAInsertPts.push_back(Elt: nullptr); |
| 506 | } |
| 507 | |
| 508 | PredIteratorCache PIC; |
| 509 | |
| 510 | // Promoting one set of accesses may make the pointers for another set |
| 511 | // loop invariant, so run this in a loop. |
| 512 | bool Promoted = false; |
| 513 | bool LocalPromoted; |
| 514 | do { |
| 515 | LocalPromoted = false; |
| 516 | for (auto [PointerMustAliases, HasReadsOutsideSet] : |
| 517 | collectPromotionCandidates(MSSA, AA, L)) { |
| 518 | LocalPromoted |= promoteLoopAccessesToScalars( |
| 519 | PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, LI, |
| 520 | DT, AC, TLI, TTI, L, MSSAU, &SafetyInfo, ORE, |
| 521 | AllowSpeculation: LicmAllowSpeculation, HasReadsOutsideSet); |
| 522 | } |
| 523 | Promoted |= LocalPromoted; |
| 524 | } while (LocalPromoted); |
| 525 | |
| 526 | // Once we have promoted values across the loop body we have to |
| 527 | // recursively reform LCSSA as any nested loop may now have values defined |
| 528 | // within the loop used in the outer loop. |
| 529 | // FIXME: This is really heavy handed. It would be a bit better to use an |
| 530 | // SSAUpdater strategy during promotion that was LCSSA aware and reformed |
| 531 | // it as it went. |
| 532 | if (Promoted) |
| 533 | formLCSSARecursively(L&: *L, DT: *DT, LI, SE); |
| 534 | |
| 535 | Changed |= Promoted; |
| 536 | } |
| 537 | } |
| 538 | |
| 539 | // Check that neither this loop nor its parent have had LCSSA broken. LICM is |
| 540 | // specifically moving instructions across the loop boundary and so it is |
| 541 | // especially in need of basic functional correctness checking here. |
| 542 | assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!" ); |
| 543 | assert((L->isOutermost() || L->getParentLoop()->isLCSSAForm(*DT)) && |
| 544 | "Parent loop not left in LCSSA form after LICM!" ); |
| 545 | |
| 546 | if (VerifyMemorySSA) |
| 547 | MSSA->verifyMemorySSA(); |
| 548 | |
| 549 | if (Changed && SE) |
| 550 | SE->forgetLoopDispositions(); |
| 551 | return Changed; |
| 552 | } |
| 553 | |
| 554 | /// Walk the specified region of the CFG (defined by all blocks dominated by |
| 555 | /// the specified block, and that are in the current loop) in reverse depth |
| 556 | /// first order w.r.t the DominatorTree. This allows us to visit uses before |
| 557 | /// definitions, allowing us to sink a loop body in one pass without iteration. |
| 558 | /// |
| 559 | bool llvm::sinkRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI, |
| 560 | DominatorTree *DT, TargetLibraryInfo *TLI, |
| 561 | TargetTransformInfo *TTI, Loop *CurLoop, |
| 562 | MemorySSAUpdater &MSSAU, ICFLoopSafetyInfo *SafetyInfo, |
| 563 | SinkAndHoistLICMFlags &Flags, |
| 564 | OptimizationRemarkEmitter *ORE, Loop *OutermostLoop) { |
| 565 | |
| 566 | // Verify inputs. |
| 567 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && |
| 568 | CurLoop != nullptr && SafetyInfo != nullptr && |
| 569 | "Unexpected input to sinkRegion." ); |
| 570 | |
| 571 | // We want to visit children before parents. We will enqueue all the parents |
| 572 | // before their children in the worklist and process the worklist in reverse |
| 573 | // order. |
| 574 | SmallVector<BasicBlock *, 16> Worklist = |
| 575 | collectChildrenInLoop(DT, N, CurLoop); |
| 576 | |
| 577 | bool Changed = false; |
| 578 | for (BasicBlock *BB : reverse(C&: Worklist)) { |
| 579 | // subloop (which would already have been processed). |
| 580 | if (inSubLoop(BB, CurLoop, LI)) |
| 581 | continue; |
| 582 | |
| 583 | for (BasicBlock::iterator II = BB->end(); II != BB->begin();) { |
| 584 | Instruction &I = *--II; |
| 585 | |
| 586 | // The instruction is not used in the loop if it is dead. In this case, |
| 587 | // we just delete it instead of sinking it. |
| 588 | if (isInstructionTriviallyDead(I: &I, TLI)) { |
| 589 | LLVM_DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n'); |
| 590 | salvageKnowledge(I: &I); |
| 591 | salvageDebugInfo(I); |
| 592 | ++II; |
| 593 | eraseInstruction(I, SafetyInfo&: *SafetyInfo, MSSAU); |
| 594 | Changed = true; |
| 595 | continue; |
| 596 | } |
| 597 | |
| 598 | // Check to see if we can sink this instruction to the exit blocks |
| 599 | // of the loop. We can do this if the all users of the instruction are |
| 600 | // outside of the loop. In this case, it doesn't even matter if the |
| 601 | // operands of the instruction are loop invariant. |
| 602 | // |
| 603 | bool FoldableInLoop = false; |
| 604 | bool LoopNestMode = OutermostLoop != nullptr; |
| 605 | if (!I.mayHaveSideEffects() && |
| 606 | isNotUsedOrFoldableInLoop(I, CurLoop: LoopNestMode ? OutermostLoop : CurLoop, |
| 607 | SafetyInfo, TTI, FoldableInLoop, |
| 608 | LoopNestMode) && |
| 609 | canSinkOrHoistInst(I, AA, DT, CurLoop, MSSAU, TargetExecutesOncePerLoop: true, LICMFlags&: Flags, ORE)) { |
| 610 | if (sink(I, LI, DT, CurLoop, SafetyInfo, MSSAU, ORE)) { |
| 611 | if (!FoldableInLoop) { |
| 612 | ++II; |
| 613 | salvageDebugInfo(I); |
| 614 | eraseInstruction(I, SafetyInfo&: *SafetyInfo, MSSAU); |
| 615 | } |
| 616 | Changed = true; |
| 617 | } |
| 618 | } |
| 619 | } |
| 620 | } |
| 621 | if (VerifyMemorySSA) |
| 622 | MSSAU.getMemorySSA()->verifyMemorySSA(); |
| 623 | return Changed; |
| 624 | } |
| 625 | |
| 626 | bool llvm::sinkRegionForLoopNest(DomTreeNode *N, AAResults *AA, LoopInfo *LI, |
| 627 | DominatorTree *DT, TargetLibraryInfo *TLI, |
| 628 | TargetTransformInfo *TTI, Loop *CurLoop, |
| 629 | MemorySSAUpdater &MSSAU, |
| 630 | ICFLoopSafetyInfo *SafetyInfo, |
| 631 | SinkAndHoistLICMFlags &Flags, |
| 632 | OptimizationRemarkEmitter *ORE) { |
| 633 | |
| 634 | bool Changed = false; |
| 635 | SmallPriorityWorklist<Loop *, 4> Worklist; |
| 636 | Worklist.insert(X: CurLoop); |
| 637 | appendLoopsToWorklist(*CurLoop, Worklist); |
| 638 | while (!Worklist.empty()) { |
| 639 | Loop *L = Worklist.pop_back_val(); |
| 640 | Changed |= sinkRegion(N: DT->getNode(BB: L->getHeader()), AA, LI, DT, TLI, TTI, CurLoop: L, |
| 641 | MSSAU, SafetyInfo, Flags, ORE, OutermostLoop: CurLoop); |
| 642 | } |
| 643 | return Changed; |
| 644 | } |
| 645 | |
| 646 | namespace { |
| 647 | // This is a helper class for hoistRegion to make it able to hoist control flow |
| 648 | // in order to be able to hoist phis. The way this works is that we initially |
| 649 | // start hoisting to the loop preheader, and when we see a loop invariant branch |
| 650 | // we make note of this. When we then come to hoist an instruction that's |
| 651 | // conditional on such a branch we duplicate the branch and the relevant control |
| 652 | // flow, then hoist the instruction into the block corresponding to its original |
| 653 | // block in the duplicated control flow. |
| 654 | class ControlFlowHoister { |
| 655 | private: |
| 656 | // Information about the loop we are hoisting from |
| 657 | LoopInfo *LI; |
| 658 | DominatorTree *DT; |
| 659 | Loop *CurLoop; |
| 660 | MemorySSAUpdater &MSSAU; |
| 661 | |
| 662 | // A map of blocks in the loop to the block their instructions will be hoisted |
| 663 | // to. |
| 664 | DenseMap<BasicBlock *, BasicBlock *> HoistDestinationMap; |
| 665 | |
| 666 | // The branches that we can hoist, mapped to the block that marks a |
| 667 | // convergence point of their control flow. |
| 668 | DenseMap<BranchInst *, BasicBlock *> HoistableBranches; |
| 669 | |
| 670 | public: |
| 671 | ControlFlowHoister(LoopInfo *LI, DominatorTree *DT, Loop *CurLoop, |
| 672 | MemorySSAUpdater &MSSAU) |
| 673 | : LI(LI), DT(DT), CurLoop(CurLoop), MSSAU(MSSAU) {} |
| 674 | |
| 675 | void registerPossiblyHoistableBranch(BranchInst *BI) { |
| 676 | // We can only hoist conditional branches with loop invariant operands. |
| 677 | if (!ControlFlowHoisting || !BI->isConditional() || |
| 678 | !CurLoop->hasLoopInvariantOperands(I: BI)) |
| 679 | return; |
| 680 | |
| 681 | // The branch destinations need to be in the loop, and we don't gain |
| 682 | // anything by duplicating conditional branches with duplicate successors, |
| 683 | // as it's essentially the same as an unconditional branch. |
| 684 | BasicBlock *TrueDest = BI->getSuccessor(i: 0); |
| 685 | BasicBlock *FalseDest = BI->getSuccessor(i: 1); |
| 686 | if (!CurLoop->contains(BB: TrueDest) || !CurLoop->contains(BB: FalseDest) || |
| 687 | TrueDest == FalseDest) |
| 688 | return; |
| 689 | |
| 690 | // We can hoist BI if one branch destination is the successor of the other, |
| 691 | // or both have common successor which we check by seeing if the |
| 692 | // intersection of their successors is non-empty. |
| 693 | // TODO: This could be expanded to allowing branches where both ends |
| 694 | // eventually converge to a single block. |
| 695 | SmallPtrSet<BasicBlock *, 4> TrueDestSucc(llvm::from_range, |
| 696 | successors(BB: TrueDest)); |
| 697 | SmallPtrSet<BasicBlock *, 4> FalseDestSucc(llvm::from_range, |
| 698 | successors(BB: FalseDest)); |
| 699 | BasicBlock *CommonSucc = nullptr; |
| 700 | if (TrueDestSucc.count(Ptr: FalseDest)) { |
| 701 | CommonSucc = FalseDest; |
| 702 | } else if (FalseDestSucc.count(Ptr: TrueDest)) { |
| 703 | CommonSucc = TrueDest; |
| 704 | } else { |
| 705 | set_intersect(S1&: TrueDestSucc, S2: FalseDestSucc); |
| 706 | // If there's one common successor use that. |
| 707 | if (TrueDestSucc.size() == 1) |
| 708 | CommonSucc = *TrueDestSucc.begin(); |
| 709 | // If there's more than one pick whichever appears first in the block list |
| 710 | // (we can't use the value returned by TrueDestSucc.begin() as it's |
| 711 | // unpredicatable which element gets returned). |
| 712 | else if (!TrueDestSucc.empty()) { |
| 713 | Function *F = TrueDest->getParent(); |
| 714 | auto IsSucc = [&](BasicBlock &BB) { return TrueDestSucc.count(Ptr: &BB); }; |
| 715 | auto It = llvm::find_if(Range&: *F, P: IsSucc); |
| 716 | assert(It != F->end() && "Could not find successor in function" ); |
| 717 | CommonSucc = &*It; |
| 718 | } |
| 719 | } |
| 720 | // The common successor has to be dominated by the branch, as otherwise |
| 721 | // there will be some other path to the successor that will not be |
| 722 | // controlled by this branch so any phi we hoist would be controlled by the |
| 723 | // wrong condition. This also takes care of avoiding hoisting of loop back |
| 724 | // edges. |
| 725 | // TODO: In some cases this could be relaxed if the successor is dominated |
| 726 | // by another block that's been hoisted and we can guarantee that the |
| 727 | // control flow has been replicated exactly. |
| 728 | if (CommonSucc && DT->dominates(Def: BI, BB: CommonSucc)) |
| 729 | HoistableBranches[BI] = CommonSucc; |
| 730 | } |
| 731 | |
| 732 | bool canHoistPHI(PHINode *PN) { |
| 733 | // The phi must have loop invariant operands. |
| 734 | if (!ControlFlowHoisting || !CurLoop->hasLoopInvariantOperands(I: PN)) |
| 735 | return false; |
| 736 | // We can hoist phis if the block they are in is the target of hoistable |
| 737 | // branches which cover all of the predecessors of the block. |
| 738 | BasicBlock *BB = PN->getParent(); |
| 739 | SmallPtrSet<BasicBlock *, 8> PredecessorBlocks(llvm::from_range, |
| 740 | predecessors(BB)); |
| 741 | // If we have less predecessor blocks than predecessors then the phi will |
| 742 | // have more than one incoming value for the same block which we can't |
| 743 | // handle. |
| 744 | // TODO: This could be handled be erasing some of the duplicate incoming |
| 745 | // values. |
| 746 | if (PredecessorBlocks.size() != pred_size(BB)) |
| 747 | return false; |
| 748 | for (auto &Pair : HoistableBranches) { |
| 749 | if (Pair.second == BB) { |
| 750 | // Which blocks are predecessors via this branch depends on if the |
| 751 | // branch is triangle-like or diamond-like. |
| 752 | if (Pair.first->getSuccessor(i: 0) == BB) { |
| 753 | PredecessorBlocks.erase(Ptr: Pair.first->getParent()); |
| 754 | PredecessorBlocks.erase(Ptr: Pair.first->getSuccessor(i: 1)); |
| 755 | } else if (Pair.first->getSuccessor(i: 1) == BB) { |
| 756 | PredecessorBlocks.erase(Ptr: Pair.first->getParent()); |
| 757 | PredecessorBlocks.erase(Ptr: Pair.first->getSuccessor(i: 0)); |
| 758 | } else { |
| 759 | PredecessorBlocks.erase(Ptr: Pair.first->getSuccessor(i: 0)); |
| 760 | PredecessorBlocks.erase(Ptr: Pair.first->getSuccessor(i: 1)); |
| 761 | } |
| 762 | } |
| 763 | } |
| 764 | // PredecessorBlocks will now be empty if for every predecessor of BB we |
| 765 | // found a hoistable branch source. |
| 766 | return PredecessorBlocks.empty(); |
| 767 | } |
| 768 | |
| 769 | BasicBlock *getOrCreateHoistedBlock(BasicBlock *BB) { |
| 770 | if (!ControlFlowHoisting) |
| 771 | return CurLoop->getLoopPreheader(); |
| 772 | // If BB has already been hoisted, return that |
| 773 | if (auto It = HoistDestinationMap.find(Val: BB); It != HoistDestinationMap.end()) |
| 774 | return It->second; |
| 775 | |
| 776 | // Check if this block is conditional based on a pending branch |
| 777 | auto HasBBAsSuccessor = |
| 778 | [&](DenseMap<BranchInst *, BasicBlock *>::value_type &Pair) { |
| 779 | return BB != Pair.second && (Pair.first->getSuccessor(i: 0) == BB || |
| 780 | Pair.first->getSuccessor(i: 1) == BB); |
| 781 | }; |
| 782 | auto It = llvm::find_if(Range&: HoistableBranches, P: HasBBAsSuccessor); |
| 783 | |
| 784 | // If not involved in a pending branch, hoist to preheader |
| 785 | BasicBlock * = CurLoop->getLoopPreheader(); |
| 786 | if (It == HoistableBranches.end()) { |
| 787 | LLVM_DEBUG(dbgs() << "LICM using " |
| 788 | << InitialPreheader->getNameOrAsOperand() |
| 789 | << " as hoist destination for " |
| 790 | << BB->getNameOrAsOperand() << "\n" ); |
| 791 | HoistDestinationMap[BB] = InitialPreheader; |
| 792 | return InitialPreheader; |
| 793 | } |
| 794 | BranchInst *BI = It->first; |
| 795 | assert(std::none_of(std::next(It), HoistableBranches.end(), |
| 796 | HasBBAsSuccessor) && |
| 797 | "BB is expected to be the target of at most one branch" ); |
| 798 | |
| 799 | LLVMContext &C = BB->getContext(); |
| 800 | BasicBlock *TrueDest = BI->getSuccessor(i: 0); |
| 801 | BasicBlock *FalseDest = BI->getSuccessor(i: 1); |
| 802 | BasicBlock *CommonSucc = HoistableBranches[BI]; |
| 803 | BasicBlock *HoistTarget = getOrCreateHoistedBlock(BB: BI->getParent()); |
| 804 | |
| 805 | // Create hoisted versions of blocks that currently don't have them |
| 806 | auto CreateHoistedBlock = [&](BasicBlock *Orig) { |
| 807 | auto [It, Inserted] = HoistDestinationMap.try_emplace(Key: Orig); |
| 808 | if (!Inserted) |
| 809 | return It->second; |
| 810 | BasicBlock *New = |
| 811 | BasicBlock::Create(Context&: C, Name: Orig->getName() + ".licm" , Parent: Orig->getParent()); |
| 812 | It->second = New; |
| 813 | DT->addNewBlock(BB: New, DomBB: HoistTarget); |
| 814 | if (CurLoop->getParentLoop()) |
| 815 | CurLoop->getParentLoop()->addBasicBlockToLoop(NewBB: New, LI&: *LI); |
| 816 | ++NumCreatedBlocks; |
| 817 | LLVM_DEBUG(dbgs() << "LICM created " << New->getName() |
| 818 | << " as hoist destination for " << Orig->getName() |
| 819 | << "\n" ); |
| 820 | return New; |
| 821 | }; |
| 822 | BasicBlock *HoistTrueDest = CreateHoistedBlock(TrueDest); |
| 823 | BasicBlock *HoistFalseDest = CreateHoistedBlock(FalseDest); |
| 824 | BasicBlock *HoistCommonSucc = CreateHoistedBlock(CommonSucc); |
| 825 | |
| 826 | // Link up these blocks with branches. |
| 827 | if (!HoistCommonSucc->getTerminator()) { |
| 828 | // The new common successor we've generated will branch to whatever that |
| 829 | // hoist target branched to. |
| 830 | BasicBlock *TargetSucc = HoistTarget->getSingleSuccessor(); |
| 831 | assert(TargetSucc && "Expected hoist target to have a single successor" ); |
| 832 | HoistCommonSucc->moveBefore(MovePos: TargetSucc); |
| 833 | BranchInst::Create(IfTrue: TargetSucc, InsertBefore: HoistCommonSucc); |
| 834 | } |
| 835 | if (!HoistTrueDest->getTerminator()) { |
| 836 | HoistTrueDest->moveBefore(MovePos: HoistCommonSucc); |
| 837 | BranchInst::Create(IfTrue: HoistCommonSucc, InsertBefore: HoistTrueDest); |
| 838 | } |
| 839 | if (!HoistFalseDest->getTerminator()) { |
| 840 | HoistFalseDest->moveBefore(MovePos: HoistCommonSucc); |
| 841 | BranchInst::Create(IfTrue: HoistCommonSucc, InsertBefore: HoistFalseDest); |
| 842 | } |
| 843 | |
| 844 | // If BI is being cloned to what was originally the preheader then |
| 845 | // HoistCommonSucc will now be the new preheader. |
| 846 | if (HoistTarget == InitialPreheader) { |
| 847 | // Phis in the loop header now need to use the new preheader. |
| 848 | InitialPreheader->replaceSuccessorsPhiUsesWith(New: HoistCommonSucc); |
| 849 | MSSAU.wireOldPredecessorsToNewImmediatePredecessor( |
| 850 | Old: HoistTarget->getSingleSuccessor(), New: HoistCommonSucc, Preds: {HoistTarget}); |
| 851 | // The new preheader dominates the loop header. |
| 852 | DomTreeNode * = DT->getNode(BB: HoistCommonSucc); |
| 853 | DomTreeNode * = DT->getNode(BB: CurLoop->getHeader()); |
| 854 | DT->changeImmediateDominator(N: HeaderNode, NewIDom: PreheaderNode); |
| 855 | // The preheader hoist destination is now the new preheader, with the |
| 856 | // exception of the hoist destination of this branch. |
| 857 | for (auto &Pair : HoistDestinationMap) |
| 858 | if (Pair.second == InitialPreheader && Pair.first != BI->getParent()) |
| 859 | Pair.second = HoistCommonSucc; |
| 860 | } |
| 861 | |
| 862 | // Now finally clone BI. |
| 863 | auto *NewBI = |
| 864 | BranchInst::Create(IfTrue: HoistTrueDest, IfFalse: HoistFalseDest, Cond: BI->getCondition(), |
| 865 | InsertBefore: HoistTarget->getTerminator()->getIterator()); |
| 866 | HoistTarget->getTerminator()->eraseFromParent(); |
| 867 | // md_prof should also come from the original branch - since the |
| 868 | // condition was hoisted, the branch probabilities shouldn't change. |
| 869 | if (!ProfcheckDisableMetadataFixes) |
| 870 | NewBI->copyMetadata(SrcInst: *BI, WL: {LLVMContext::MD_prof}); |
| 871 | // FIXME: Issue #152767: debug info should also be the same as the |
| 872 | // original branch, **if** the user explicitly indicated that. |
| 873 | NewBI->setDebugLoc(HoistTarget->getTerminator()->getDebugLoc()); |
| 874 | |
| 875 | ++NumClonedBranches; |
| 876 | |
| 877 | assert(CurLoop->getLoopPreheader() && |
| 878 | "Hoisting blocks should not have destroyed preheader" ); |
| 879 | return HoistDestinationMap[BB]; |
| 880 | } |
| 881 | }; |
| 882 | } // namespace |
| 883 | |
| 884 | /// Walk the specified region of the CFG (defined by all blocks dominated by |
| 885 | /// the specified block, and that are in the current loop) in depth first |
| 886 | /// order w.r.t the DominatorTree. This allows us to visit definitions before |
| 887 | /// uses, allowing us to hoist a loop body in one pass without iteration. |
| 888 | /// |
| 889 | bool llvm::hoistRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI, |
| 890 | DominatorTree *DT, AssumptionCache *AC, |
| 891 | TargetLibraryInfo *TLI, Loop *CurLoop, |
| 892 | MemorySSAUpdater &MSSAU, ScalarEvolution *SE, |
| 893 | ICFLoopSafetyInfo *SafetyInfo, |
| 894 | SinkAndHoistLICMFlags &Flags, |
| 895 | OptimizationRemarkEmitter *ORE, bool LoopNestMode, |
| 896 | bool AllowSpeculation) { |
| 897 | // Verify inputs. |
| 898 | assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && |
| 899 | CurLoop != nullptr && SafetyInfo != nullptr && |
| 900 | "Unexpected input to hoistRegion." ); |
| 901 | |
| 902 | ControlFlowHoister CFH(LI, DT, CurLoop, MSSAU); |
| 903 | |
| 904 | // Keep track of instructions that have been hoisted, as they may need to be |
| 905 | // re-hoisted if they end up not dominating all of their uses. |
| 906 | SmallVector<Instruction *, 16> HoistedInstructions; |
| 907 | |
| 908 | // For PHI hoisting to work we need to hoist blocks before their successors. |
| 909 | // We can do this by iterating through the blocks in the loop in reverse |
| 910 | // post-order. |
| 911 | LoopBlocksRPO Worklist(CurLoop); |
| 912 | Worklist.perform(LI); |
| 913 | bool Changed = false; |
| 914 | BasicBlock * = CurLoop->getLoopPreheader(); |
| 915 | for (BasicBlock *BB : Worklist) { |
| 916 | // Only need to process the contents of this block if it is not part of a |
| 917 | // subloop (which would already have been processed). |
| 918 | if (!LoopNestMode && inSubLoop(BB, CurLoop, LI)) |
| 919 | continue; |
| 920 | |
| 921 | for (Instruction &I : llvm::make_early_inc_range(Range&: *BB)) { |
| 922 | // Try hoisting the instruction out to the preheader. We can only do |
| 923 | // this if all of the operands of the instruction are loop invariant and |
| 924 | // if it is safe to hoist the instruction. We also check block frequency |
| 925 | // to make sure instruction only gets hoisted into colder blocks. |
| 926 | // TODO: It may be safe to hoist if we are hoisting to a conditional block |
| 927 | // and we have accurately duplicated the control flow from the loop header |
| 928 | // to that block. |
| 929 | if (CurLoop->hasLoopInvariantOperands(I: &I) && |
| 930 | canSinkOrHoistInst(I, AA, DT, CurLoop, MSSAU, TargetExecutesOncePerLoop: true, LICMFlags&: Flags, ORE) && |
| 931 | isSafeToExecuteUnconditionally(Inst&: I, DT, TLI, CurLoop, SafetyInfo, ORE, |
| 932 | CtxI: Preheader->getTerminator(), AC, |
| 933 | AllowSpeculation)) { |
| 934 | hoist(I, DT, CurLoop, Dest: CFH.getOrCreateHoistedBlock(BB), SafetyInfo, |
| 935 | MSSAU, SE, ORE); |
| 936 | HoistedInstructions.push_back(Elt: &I); |
| 937 | Changed = true; |
| 938 | continue; |
| 939 | } |
| 940 | |
| 941 | // Attempt to remove floating point division out of the loop by |
| 942 | // converting it to a reciprocal multiplication. |
| 943 | if (I.getOpcode() == Instruction::FDiv && I.hasAllowReciprocal() && |
| 944 | CurLoop->isLoopInvariant(V: I.getOperand(i: 1))) { |
| 945 | auto Divisor = I.getOperand(i: 1); |
| 946 | auto One = llvm::ConstantFP::get(Ty: Divisor->getType(), V: 1.0); |
| 947 | auto ReciprocalDivisor = BinaryOperator::CreateFDiv(V1: One, V2: Divisor); |
| 948 | ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags()); |
| 949 | SafetyInfo->insertInstructionTo(Inst: ReciprocalDivisor, BB: I.getParent()); |
| 950 | ReciprocalDivisor->insertBefore(InsertPos: I.getIterator()); |
| 951 | ReciprocalDivisor->setDebugLoc(I.getDebugLoc()); |
| 952 | |
| 953 | auto Product = |
| 954 | BinaryOperator::CreateFMul(V1: I.getOperand(i: 0), V2: ReciprocalDivisor); |
| 955 | Product->setFastMathFlags(I.getFastMathFlags()); |
| 956 | SafetyInfo->insertInstructionTo(Inst: Product, BB: I.getParent()); |
| 957 | Product->insertAfter(InsertPos: I.getIterator()); |
| 958 | Product->setDebugLoc(I.getDebugLoc()); |
| 959 | I.replaceAllUsesWith(V: Product); |
| 960 | eraseInstruction(I, SafetyInfo&: *SafetyInfo, MSSAU); |
| 961 | |
| 962 | hoist(I&: *ReciprocalDivisor, DT, CurLoop, Dest: CFH.getOrCreateHoistedBlock(BB), |
| 963 | SafetyInfo, MSSAU, SE, ORE); |
| 964 | HoistedInstructions.push_back(Elt: ReciprocalDivisor); |
| 965 | Changed = true; |
| 966 | continue; |
| 967 | } |
| 968 | |
| 969 | auto IsInvariantStart = [&](Instruction &I) { |
| 970 | using namespace PatternMatch; |
| 971 | return I.use_empty() && |
| 972 | match(V: &I, P: m_Intrinsic<Intrinsic::invariant_start>()); |
| 973 | }; |
| 974 | auto MustExecuteWithoutWritesBefore = [&](Instruction &I) { |
| 975 | return SafetyInfo->isGuaranteedToExecute(Inst: I, DT, CurLoop) && |
| 976 | SafetyInfo->doesNotWriteMemoryBefore(I, CurLoop); |
| 977 | }; |
| 978 | if ((IsInvariantStart(I) || isGuard(U: &I)) && |
| 979 | CurLoop->hasLoopInvariantOperands(I: &I) && |
| 980 | MustExecuteWithoutWritesBefore(I)) { |
| 981 | hoist(I, DT, CurLoop, Dest: CFH.getOrCreateHoistedBlock(BB), SafetyInfo, |
| 982 | MSSAU, SE, ORE); |
| 983 | HoistedInstructions.push_back(Elt: &I); |
| 984 | Changed = true; |
| 985 | continue; |
| 986 | } |
| 987 | |
| 988 | if (PHINode *PN = dyn_cast<PHINode>(Val: &I)) { |
| 989 | if (CFH.canHoistPHI(PN)) { |
| 990 | // Redirect incoming blocks first to ensure that we create hoisted |
| 991 | // versions of those blocks before we hoist the phi. |
| 992 | for (unsigned int i = 0; i < PN->getNumIncomingValues(); ++i) |
| 993 | PN->setIncomingBlock( |
| 994 | i, BB: CFH.getOrCreateHoistedBlock(BB: PN->getIncomingBlock(i))); |
| 995 | hoist(I&: *PN, DT, CurLoop, Dest: CFH.getOrCreateHoistedBlock(BB), SafetyInfo, |
| 996 | MSSAU, SE, ORE); |
| 997 | assert(DT->dominates(PN, BB) && "Conditional PHIs not expected" ); |
| 998 | Changed = true; |
| 999 | continue; |
| 1000 | } |
| 1001 | } |
| 1002 | |
| 1003 | // Try to reassociate instructions so that part of computations can be |
| 1004 | // done out of loop. |
| 1005 | if (hoistArithmetics(I, L&: *CurLoop, SafetyInfo&: *SafetyInfo, MSSAU, AC, DT)) { |
| 1006 | Changed = true; |
| 1007 | continue; |
| 1008 | } |
| 1009 | |
| 1010 | // Remember possibly hoistable branches so we can actually hoist them |
| 1011 | // later if needed. |
| 1012 | if (BranchInst *BI = dyn_cast<BranchInst>(Val: &I)) |
| 1013 | CFH.registerPossiblyHoistableBranch(BI); |
| 1014 | } |
| 1015 | } |
| 1016 | |
| 1017 | // If we hoisted instructions to a conditional block they may not dominate |
| 1018 | // their uses that weren't hoisted (such as phis where some operands are not |
| 1019 | // loop invariant). If so make them unconditional by moving them to their |
| 1020 | // immediate dominator. We iterate through the instructions in reverse order |
| 1021 | // which ensures that when we rehoist an instruction we rehoist its operands, |
| 1022 | // and also keep track of where in the block we are rehoisting to make sure |
| 1023 | // that we rehoist instructions before the instructions that use them. |
| 1024 | Instruction *HoistPoint = nullptr; |
| 1025 | if (ControlFlowHoisting) { |
| 1026 | for (Instruction *I : reverse(C&: HoistedInstructions)) { |
| 1027 | if (!llvm::all_of(Range: I->uses(), |
| 1028 | P: [&](Use &U) { return DT->dominates(Def: I, U); })) { |
| 1029 | BasicBlock *Dominator = |
| 1030 | DT->getNode(BB: I->getParent())->getIDom()->getBlock(); |
| 1031 | if (!HoistPoint || !DT->dominates(A: HoistPoint->getParent(), B: Dominator)) { |
| 1032 | if (HoistPoint) |
| 1033 | assert(DT->dominates(Dominator, HoistPoint->getParent()) && |
| 1034 | "New hoist point expected to dominate old hoist point" ); |
| 1035 | HoistPoint = Dominator->getTerminator(); |
| 1036 | } |
| 1037 | LLVM_DEBUG(dbgs() << "LICM rehoisting to " |
| 1038 | << HoistPoint->getParent()->getNameOrAsOperand() |
| 1039 | << ": " << *I << "\n" ); |
| 1040 | moveInstructionBefore(I&: *I, Dest: HoistPoint->getIterator(), SafetyInfo&: *SafetyInfo, MSSAU, |
| 1041 | SE); |
| 1042 | HoistPoint = I; |
| 1043 | Changed = true; |
| 1044 | } |
| 1045 | } |
| 1046 | } |
| 1047 | if (VerifyMemorySSA) |
| 1048 | MSSAU.getMemorySSA()->verifyMemorySSA(); |
| 1049 | |
| 1050 | // Now that we've finished hoisting make sure that LI and DT are still |
| 1051 | // valid. |
| 1052 | #ifdef EXPENSIVE_CHECKS |
| 1053 | if (Changed) { |
| 1054 | assert(DT->verify(DominatorTree::VerificationLevel::Fast) && |
| 1055 | "Dominator tree verification failed" ); |
| 1056 | LI->verify(*DT); |
| 1057 | } |
| 1058 | #endif |
| 1059 | |
| 1060 | return Changed; |
| 1061 | } |
| 1062 | |
| 1063 | // Return true if LI is invariant within scope of the loop. LI is invariant if |
| 1064 | // CurLoop is dominated by an invariant.start representing the same memory |
| 1065 | // location and size as the memory location LI loads from, and also the |
| 1066 | // invariant.start has no uses. |
| 1067 | static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT, |
| 1068 | Loop *CurLoop) { |
| 1069 | Value *Addr = LI->getPointerOperand(); |
| 1070 | const DataLayout &DL = LI->getDataLayout(); |
| 1071 | const TypeSize LocSizeInBits = DL.getTypeSizeInBits(Ty: LI->getType()); |
| 1072 | |
| 1073 | // It is not currently possible for clang to generate an invariant.start |
| 1074 | // intrinsic with scalable vector types because we don't support thread local |
| 1075 | // sizeless types and we don't permit sizeless types in structs or classes. |
| 1076 | // Furthermore, even if support is added for this in future the intrinsic |
| 1077 | // itself is defined to have a size of -1 for variable sized objects. This |
| 1078 | // makes it impossible to verify if the intrinsic envelops our region of |
| 1079 | // interest. For example, both <vscale x 32 x i8> and <vscale x 16 x i8> |
| 1080 | // types would have a -1 parameter, but the former is clearly double the size |
| 1081 | // of the latter. |
| 1082 | if (LocSizeInBits.isScalable()) |
| 1083 | return false; |
| 1084 | |
| 1085 | // If we've ended up at a global/constant, bail. We shouldn't be looking at |
| 1086 | // uselists for non-local Values in a loop pass. |
| 1087 | if (isa<Constant>(Val: Addr)) |
| 1088 | return false; |
| 1089 | |
| 1090 | unsigned UsesVisited = 0; |
| 1091 | // Traverse all uses of the load operand value, to see if invariant.start is |
| 1092 | // one of the uses, and whether it dominates the load instruction. |
| 1093 | for (auto *U : Addr->users()) { |
| 1094 | // Avoid traversing for Load operand with high number of users. |
| 1095 | if (++UsesVisited > MaxNumUsesTraversed) |
| 1096 | return false; |
| 1097 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: U); |
| 1098 | // If there are escaping uses of invariant.start instruction, the load maybe |
| 1099 | // non-invariant. |
| 1100 | if (!II || II->getIntrinsicID() != Intrinsic::invariant_start || |
| 1101 | !II->use_empty()) |
| 1102 | continue; |
| 1103 | ConstantInt *InvariantSize = cast<ConstantInt>(Val: II->getArgOperand(i: 0)); |
| 1104 | // The intrinsic supports having a -1 argument for variable sized objects |
| 1105 | // so we should check for that here. |
| 1106 | if (InvariantSize->isNegative()) |
| 1107 | continue; |
| 1108 | uint64_t InvariantSizeInBits = InvariantSize->getSExtValue() * 8; |
| 1109 | // Confirm the invariant.start location size contains the load operand size |
| 1110 | // in bits. Also, the invariant.start should dominate the load, and we |
| 1111 | // should not hoist the load out of a loop that contains this dominating |
| 1112 | // invariant.start. |
| 1113 | if (LocSizeInBits.getFixedValue() <= InvariantSizeInBits && |
| 1114 | DT->properlyDominates(A: II->getParent(), B: CurLoop->getHeader())) |
| 1115 | return true; |
| 1116 | } |
| 1117 | |
| 1118 | return false; |
| 1119 | } |
| 1120 | |
| 1121 | /// Return true if-and-only-if we know how to (mechanically) both hoist and |
| 1122 | /// sink a given instruction out of a loop. Does not address legality |
| 1123 | /// concerns such as aliasing or speculation safety. |
| 1124 | static bool isHoistableAndSinkableInst(Instruction &I) { |
| 1125 | // Only these instructions are hoistable/sinkable. |
| 1126 | return (isa<LoadInst>(Val: I) || isa<StoreInst>(Val: I) || isa<CallInst>(Val: I) || |
| 1127 | isa<FenceInst>(Val: I) || isa<CastInst>(Val: I) || isa<UnaryOperator>(Val: I) || |
| 1128 | isa<BinaryOperator>(Val: I) || isa<SelectInst>(Val: I) || |
| 1129 | isa<GetElementPtrInst>(Val: I) || isa<CmpInst>(Val: I) || |
| 1130 | isa<InsertElementInst>(Val: I) || isa<ExtractElementInst>(Val: I) || |
| 1131 | isa<ShuffleVectorInst>(Val: I) || isa<ExtractValueInst>(Val: I) || |
| 1132 | isa<InsertValueInst>(Val: I) || isa<FreezeInst>(Val: I)); |
| 1133 | } |
| 1134 | |
| 1135 | /// Return true if I is the only Instruction with a MemoryAccess in L. |
| 1136 | static bool isOnlyMemoryAccess(const Instruction *I, const Loop *L, |
| 1137 | const MemorySSAUpdater &MSSAU) { |
| 1138 | for (auto *BB : L->getBlocks()) |
| 1139 | if (auto *Accs = MSSAU.getMemorySSA()->getBlockAccesses(BB)) { |
| 1140 | int NotAPhi = 0; |
| 1141 | for (const auto &Acc : *Accs) { |
| 1142 | if (isa<MemoryPhi>(Val: &Acc)) |
| 1143 | continue; |
| 1144 | const auto *MUD = cast<MemoryUseOrDef>(Val: &Acc); |
| 1145 | if (MUD->getMemoryInst() != I || NotAPhi++ == 1) |
| 1146 | return false; |
| 1147 | } |
| 1148 | } |
| 1149 | return true; |
| 1150 | } |
| 1151 | |
| 1152 | static MemoryAccess *getClobberingMemoryAccess(MemorySSA &MSSA, |
| 1153 | BatchAAResults &BAA, |
| 1154 | SinkAndHoistLICMFlags &Flags, |
| 1155 | MemoryUseOrDef *MA) { |
| 1156 | // See declaration of SetLicmMssaOptCap for usage details. |
| 1157 | if (Flags.tooManyClobberingCalls()) |
| 1158 | return MA->getDefiningAccess(); |
| 1159 | |
| 1160 | MemoryAccess *Source = |
| 1161 | MSSA.getSkipSelfWalker()->getClobberingMemoryAccess(MA, AA&: BAA); |
| 1162 | Flags.incrementClobberingCalls(); |
| 1163 | return Source; |
| 1164 | } |
| 1165 | |
| 1166 | bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT, |
| 1167 | Loop *CurLoop, MemorySSAUpdater &MSSAU, |
| 1168 | bool TargetExecutesOncePerLoop, |
| 1169 | SinkAndHoistLICMFlags &Flags, |
| 1170 | OptimizationRemarkEmitter *ORE) { |
| 1171 | // If we don't understand the instruction, bail early. |
| 1172 | if (!isHoistableAndSinkableInst(I)) |
| 1173 | return false; |
| 1174 | |
| 1175 | MemorySSA *MSSA = MSSAU.getMemorySSA(); |
| 1176 | // Loads have extra constraints we have to verify before we can hoist them. |
| 1177 | if (LoadInst *LI = dyn_cast<LoadInst>(Val: &I)) { |
| 1178 | if (!LI->isUnordered()) |
| 1179 | return false; // Don't sink/hoist volatile or ordered atomic loads! |
| 1180 | |
| 1181 | // Loads from constant memory are always safe to move, even if they end up |
| 1182 | // in the same alias set as something that ends up being modified. |
| 1183 | if (!isModSet(MRI: AA->getModRefInfoMask(P: LI->getOperand(i_nocapture: 0)))) |
| 1184 | return true; |
| 1185 | if (LI->hasMetadata(KindID: LLVMContext::MD_invariant_load)) |
| 1186 | return true; |
| 1187 | |
| 1188 | if (LI->isAtomic() && !TargetExecutesOncePerLoop) |
| 1189 | return false; // Don't risk duplicating unordered loads |
| 1190 | |
| 1191 | // This checks for an invariant.start dominating the load. |
| 1192 | if (isLoadInvariantInLoop(LI, DT, CurLoop)) |
| 1193 | return true; |
| 1194 | |
| 1195 | auto MU = cast<MemoryUse>(Val: MSSA->getMemoryAccess(I: LI)); |
| 1196 | |
| 1197 | bool InvariantGroup = LI->hasMetadata(KindID: LLVMContext::MD_invariant_group); |
| 1198 | |
| 1199 | bool Invalidated = pointerInvalidatedByLoop( |
| 1200 | MSSA, MU, CurLoop, I, Flags, InvariantGroup); |
| 1201 | // Check loop-invariant address because this may also be a sinkable load |
| 1202 | // whose address is not necessarily loop-invariant. |
| 1203 | if (ORE && Invalidated && CurLoop->isLoopInvariant(V: LI->getPointerOperand())) |
| 1204 | ORE->emit(RemarkBuilder: [&]() { |
| 1205 | return OptimizationRemarkMissed( |
| 1206 | DEBUG_TYPE, "LoadWithLoopInvariantAddressInvalidated" , LI) |
| 1207 | << "failed to move load with loop-invariant address " |
| 1208 | "because the loop may invalidate its value" ; |
| 1209 | }); |
| 1210 | |
| 1211 | return !Invalidated; |
| 1212 | } else if (CallInst *CI = dyn_cast<CallInst>(Val: &I)) { |
| 1213 | // Don't sink calls which can throw. |
| 1214 | if (CI->mayThrow()) |
| 1215 | return false; |
| 1216 | |
| 1217 | // Convergent attribute has been used on operations that involve |
| 1218 | // inter-thread communication which results are implicitly affected by the |
| 1219 | // enclosing control flows. It is not safe to hoist or sink such operations |
| 1220 | // across control flow. |
| 1221 | if (CI->isConvergent()) |
| 1222 | return false; |
| 1223 | |
| 1224 | // FIXME: Current LLVM IR semantics don't work well with coroutines and |
| 1225 | // thread local globals. We currently treat getting the address of a thread |
| 1226 | // local global as not accessing memory, even though it may not be a |
| 1227 | // constant throughout a function with coroutines. Remove this check after |
| 1228 | // we better model semantics of thread local globals. |
| 1229 | if (CI->getFunction()->isPresplitCoroutine()) |
| 1230 | return false; |
| 1231 | |
| 1232 | using namespace PatternMatch; |
| 1233 | if (match(V: CI, P: m_Intrinsic<Intrinsic::assume>())) |
| 1234 | // Assumes don't actually alias anything or throw |
| 1235 | return true; |
| 1236 | |
| 1237 | // Handle simple cases by querying alias analysis. |
| 1238 | MemoryEffects Behavior = AA->getMemoryEffects(Call: CI); |
| 1239 | |
| 1240 | if (Behavior.doesNotAccessMemory()) |
| 1241 | return true; |
| 1242 | if (Behavior.onlyReadsMemory()) { |
| 1243 | // Might have stale MemoryDef for call that was later inferred to be |
| 1244 | // read-only. |
| 1245 | auto *MU = dyn_cast<MemoryUse>(Val: MSSA->getMemoryAccess(I: CI)); |
| 1246 | if (!MU) |
| 1247 | return false; |
| 1248 | |
| 1249 | // If we can prove there are no writes to the memory read by the call, we |
| 1250 | // can hoist or sink. |
| 1251 | return !pointerInvalidatedByLoop( |
| 1252 | MSSA, MU, CurLoop, I, Flags, /*InvariantGroup=*/false); |
| 1253 | } |
| 1254 | |
| 1255 | if (Behavior.onlyWritesMemory()) { |
| 1256 | // can hoist or sink if there are no conflicting read/writes to the |
| 1257 | // memory location written to by the call. |
| 1258 | return noConflictingReadWrites(I: CI, MSSA, AA, CurLoop, Flags); |
| 1259 | } |
| 1260 | |
| 1261 | return false; |
| 1262 | } else if (auto *FI = dyn_cast<FenceInst>(Val: &I)) { |
| 1263 | // Fences alias (most) everything to provide ordering. For the moment, |
| 1264 | // just give up if there are any other memory operations in the loop. |
| 1265 | return isOnlyMemoryAccess(I: FI, L: CurLoop, MSSAU); |
| 1266 | } else if (auto *SI = dyn_cast<StoreInst>(Val: &I)) { |
| 1267 | if (!SI->isUnordered()) |
| 1268 | return false; // Don't sink/hoist volatile or ordered atomic store! |
| 1269 | |
| 1270 | // We can only hoist a store that we can prove writes a value which is not |
| 1271 | // read or overwritten within the loop. For those cases, we fallback to |
| 1272 | // load store promotion instead. TODO: We can extend this to cases where |
| 1273 | // there is exactly one write to the location and that write dominates an |
| 1274 | // arbitrary number of reads in the loop. |
| 1275 | if (isOnlyMemoryAccess(I: SI, L: CurLoop, MSSAU)) |
| 1276 | return true; |
| 1277 | return noConflictingReadWrites(I: SI, MSSA, AA, CurLoop, Flags); |
| 1278 | } |
| 1279 | |
| 1280 | assert(!I.mayReadOrWriteMemory() && "unhandled aliasing" ); |
| 1281 | |
| 1282 | // We've established mechanical ability and aliasing, it's up to the caller |
| 1283 | // to check fault safety |
| 1284 | return true; |
| 1285 | } |
| 1286 | |
| 1287 | /// Returns true if a PHINode is a trivially replaceable with an |
| 1288 | /// Instruction. |
| 1289 | /// This is true when all incoming values are that instruction. |
| 1290 | /// This pattern occurs most often with LCSSA PHI nodes. |
| 1291 | /// |
| 1292 | static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) { |
| 1293 | for (const Value *IncValue : PN.incoming_values()) |
| 1294 | if (IncValue != &I) |
| 1295 | return false; |
| 1296 | |
| 1297 | return true; |
| 1298 | } |
| 1299 | |
| 1300 | /// Return true if the instruction is foldable in the loop. |
| 1301 | static bool isFoldableInLoop(const Instruction &I, const Loop *CurLoop, |
| 1302 | const TargetTransformInfo *TTI) { |
| 1303 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: &I)) { |
| 1304 | InstructionCost CostI = |
| 1305 | TTI->getInstructionCost(U: &I, CostKind: TargetTransformInfo::TCK_SizeAndLatency); |
| 1306 | if (CostI != TargetTransformInfo::TCC_Free) |
| 1307 | return false; |
| 1308 | // For a GEP, we cannot simply use getInstructionCost because currently |
| 1309 | // it optimistically assumes that a GEP will fold into addressing mode |
| 1310 | // regardless of its users. |
| 1311 | const BasicBlock *BB = GEP->getParent(); |
| 1312 | for (const User *U : GEP->users()) { |
| 1313 | const Instruction *UI = cast<Instruction>(Val: U); |
| 1314 | if (CurLoop->contains(Inst: UI) && |
| 1315 | (BB != UI->getParent() || |
| 1316 | (!isa<StoreInst>(Val: UI) && !isa<LoadInst>(Val: UI)))) |
| 1317 | return false; |
| 1318 | } |
| 1319 | return true; |
| 1320 | } |
| 1321 | |
| 1322 | return false; |
| 1323 | } |
| 1324 | |
| 1325 | /// Return true if the only users of this instruction are outside of |
| 1326 | /// the loop. If this is true, we can sink the instruction to the exit |
| 1327 | /// blocks of the loop. |
| 1328 | /// |
| 1329 | /// We also return true if the instruction could be folded away in lowering. |
| 1330 | /// (e.g., a GEP can be folded into a load as an addressing mode in the loop). |
| 1331 | static bool isNotUsedOrFoldableInLoop(const Instruction &I, const Loop *CurLoop, |
| 1332 | const LoopSafetyInfo *SafetyInfo, |
| 1333 | TargetTransformInfo *TTI, |
| 1334 | bool &FoldableInLoop, bool LoopNestMode) { |
| 1335 | const auto &BlockColors = SafetyInfo->getBlockColors(); |
| 1336 | bool IsFoldable = isFoldableInLoop(I, CurLoop, TTI); |
| 1337 | for (const User *U : I.users()) { |
| 1338 | const Instruction *UI = cast<Instruction>(Val: U); |
| 1339 | if (const PHINode *PN = dyn_cast<PHINode>(Val: UI)) { |
| 1340 | const BasicBlock *BB = PN->getParent(); |
| 1341 | // We cannot sink uses in catchswitches. |
| 1342 | if (isa<CatchSwitchInst>(Val: BB->getTerminator())) |
| 1343 | return false; |
| 1344 | |
| 1345 | // We need to sink a callsite to a unique funclet. Avoid sinking if the |
| 1346 | // phi use is too muddled. |
| 1347 | if (isa<CallInst>(Val: I)) |
| 1348 | if (!BlockColors.empty() && |
| 1349 | BlockColors.find(Val: const_cast<BasicBlock *>(BB))->second.size() != 1) |
| 1350 | return false; |
| 1351 | |
| 1352 | if (LoopNestMode) { |
| 1353 | while (isa<PHINode>(Val: UI) && UI->hasOneUser() && |
| 1354 | UI->getNumOperands() == 1) { |
| 1355 | if (!CurLoop->contains(Inst: UI)) |
| 1356 | break; |
| 1357 | UI = cast<Instruction>(Val: UI->user_back()); |
| 1358 | } |
| 1359 | } |
| 1360 | } |
| 1361 | |
| 1362 | if (CurLoop->contains(Inst: UI)) { |
| 1363 | if (IsFoldable) { |
| 1364 | FoldableInLoop = true; |
| 1365 | continue; |
| 1366 | } |
| 1367 | return false; |
| 1368 | } |
| 1369 | } |
| 1370 | return true; |
| 1371 | } |
| 1372 | |
| 1373 | static Instruction *cloneInstructionInExitBlock( |
| 1374 | Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI, |
| 1375 | const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater &MSSAU) { |
| 1376 | Instruction *New; |
| 1377 | if (auto *CI = dyn_cast<CallInst>(Val: &I)) { |
| 1378 | const auto &BlockColors = SafetyInfo->getBlockColors(); |
| 1379 | |
| 1380 | // Sinking call-sites need to be handled differently from other |
| 1381 | // instructions. The cloned call-site needs a funclet bundle operand |
| 1382 | // appropriate for its location in the CFG. |
| 1383 | SmallVector<OperandBundleDef, 1> OpBundles; |
| 1384 | for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles(); |
| 1385 | BundleIdx != BundleEnd; ++BundleIdx) { |
| 1386 | OperandBundleUse Bundle = CI->getOperandBundleAt(Index: BundleIdx); |
| 1387 | if (Bundle.getTagID() == LLVMContext::OB_funclet) |
| 1388 | continue; |
| 1389 | |
| 1390 | OpBundles.emplace_back(Args&: Bundle); |
| 1391 | } |
| 1392 | |
| 1393 | if (!BlockColors.empty()) { |
| 1394 | const ColorVector &CV = BlockColors.find(Val: &ExitBlock)->second; |
| 1395 | assert(CV.size() == 1 && "non-unique color for exit block!" ); |
| 1396 | BasicBlock *BBColor = CV.front(); |
| 1397 | BasicBlock::iterator EHPad = BBColor->getFirstNonPHIIt(); |
| 1398 | if (EHPad->isEHPad()) |
| 1399 | OpBundles.emplace_back(Args: "funclet" , Args: &*EHPad); |
| 1400 | } |
| 1401 | |
| 1402 | New = CallInst::Create(CI, Bundles: OpBundles); |
| 1403 | New->copyMetadata(SrcInst: *CI); |
| 1404 | } else { |
| 1405 | New = I.clone(); |
| 1406 | } |
| 1407 | |
| 1408 | New->insertInto(ParentBB: &ExitBlock, It: ExitBlock.getFirstInsertionPt()); |
| 1409 | if (!I.getName().empty()) |
| 1410 | New->setName(I.getName() + ".le" ); |
| 1411 | |
| 1412 | if (MSSAU.getMemorySSA()->getMemoryAccess(I: &I)) { |
| 1413 | // Create a new MemoryAccess and let MemorySSA set its defining access. |
| 1414 | // After running some passes, MemorySSA might be outdated, and the |
| 1415 | // instruction `I` may have become a non-memory touching instruction. |
| 1416 | MemoryAccess *NewMemAcc = MSSAU.createMemoryAccessInBB( |
| 1417 | I: New, Definition: nullptr, BB: New->getParent(), Point: MemorySSA::Beginning, |
| 1418 | /*CreationMustSucceed=*/false); |
| 1419 | if (NewMemAcc) { |
| 1420 | if (auto *MemDef = dyn_cast<MemoryDef>(Val: NewMemAcc)) |
| 1421 | MSSAU.insertDef(Def: MemDef, /*RenameUses=*/true); |
| 1422 | else { |
| 1423 | auto *MemUse = cast<MemoryUse>(Val: NewMemAcc); |
| 1424 | MSSAU.insertUse(Use: MemUse, /*RenameUses=*/true); |
| 1425 | } |
| 1426 | } |
| 1427 | } |
| 1428 | |
| 1429 | // Build LCSSA PHI nodes for any in-loop operands (if legal). Note that |
| 1430 | // this is particularly cheap because we can rip off the PHI node that we're |
| 1431 | // replacing for the number and blocks of the predecessors. |
| 1432 | // OPT: If this shows up in a profile, we can instead finish sinking all |
| 1433 | // invariant instructions, and then walk their operands to re-establish |
| 1434 | // LCSSA. That will eliminate creating PHI nodes just to nuke them when |
| 1435 | // sinking bottom-up. |
| 1436 | for (Use &Op : New->operands()) |
| 1437 | if (LI->wouldBeOutOfLoopUseRequiringLCSSA(V: Op.get(), ExitBB: PN.getParent())) { |
| 1438 | auto *OInst = cast<Instruction>(Val: Op.get()); |
| 1439 | PHINode *OpPN = |
| 1440 | PHINode::Create(Ty: OInst->getType(), NumReservedValues: PN.getNumIncomingValues(), |
| 1441 | NameStr: OInst->getName() + ".lcssa" ); |
| 1442 | OpPN->insertBefore(InsertPos: ExitBlock.begin()); |
| 1443 | for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) |
| 1444 | OpPN->addIncoming(V: OInst, BB: PN.getIncomingBlock(i)); |
| 1445 | Op = OpPN; |
| 1446 | } |
| 1447 | return New; |
| 1448 | } |
| 1449 | |
| 1450 | static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo, |
| 1451 | MemorySSAUpdater &MSSAU) { |
| 1452 | MSSAU.removeMemoryAccess(I: &I); |
| 1453 | SafetyInfo.removeInstruction(Inst: &I); |
| 1454 | I.eraseFromParent(); |
| 1455 | } |
| 1456 | |
| 1457 | static void moveInstructionBefore(Instruction &I, BasicBlock::iterator Dest, |
| 1458 | ICFLoopSafetyInfo &SafetyInfo, |
| 1459 | MemorySSAUpdater &MSSAU, |
| 1460 | ScalarEvolution *SE) { |
| 1461 | SafetyInfo.removeInstruction(Inst: &I); |
| 1462 | SafetyInfo.insertInstructionTo(Inst: &I, BB: Dest->getParent()); |
| 1463 | I.moveBefore(BB&: *Dest->getParent(), I: Dest); |
| 1464 | if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>( |
| 1465 | Val: MSSAU.getMemorySSA()->getMemoryAccess(I: &I))) |
| 1466 | MSSAU.moveToPlace(What: OldMemAcc, BB: Dest->getParent(), |
| 1467 | Where: MemorySSA::BeforeTerminator); |
| 1468 | if (SE) |
| 1469 | SE->forgetBlockAndLoopDispositions(V: &I); |
| 1470 | } |
| 1471 | |
| 1472 | static Instruction *sinkThroughTriviallyReplaceablePHI( |
| 1473 | PHINode *TPN, Instruction *I, LoopInfo *LI, |
| 1474 | SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies, |
| 1475 | const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop, |
| 1476 | MemorySSAUpdater &MSSAU) { |
| 1477 | assert(isTriviallyReplaceablePHI(*TPN, *I) && |
| 1478 | "Expect only trivially replaceable PHI" ); |
| 1479 | BasicBlock *ExitBlock = TPN->getParent(); |
| 1480 | auto [It, Inserted] = SunkCopies.try_emplace(Key: ExitBlock); |
| 1481 | if (Inserted) |
| 1482 | It->second = cloneInstructionInExitBlock(I&: *I, ExitBlock&: *ExitBlock, PN&: *TPN, LI, |
| 1483 | SafetyInfo, MSSAU); |
| 1484 | return It->second; |
| 1485 | } |
| 1486 | |
| 1487 | static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) { |
| 1488 | BasicBlock *BB = PN->getParent(); |
| 1489 | if (!BB->canSplitPredecessors()) |
| 1490 | return false; |
| 1491 | // It's not impossible to split EHPad blocks, but if BlockColors already exist |
| 1492 | // it require updating BlockColors for all offspring blocks accordingly. By |
| 1493 | // skipping such corner case, we can make updating BlockColors after splitting |
| 1494 | // predecessor fairly simple. |
| 1495 | if (!SafetyInfo->getBlockColors().empty() && |
| 1496 | BB->getFirstNonPHIIt()->isEHPad()) |
| 1497 | return false; |
| 1498 | for (BasicBlock *BBPred : predecessors(BB)) { |
| 1499 | if (isa<IndirectBrInst>(Val: BBPred->getTerminator())) |
| 1500 | return false; |
| 1501 | } |
| 1502 | return true; |
| 1503 | } |
| 1504 | |
| 1505 | static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT, |
| 1506 | LoopInfo *LI, const Loop *CurLoop, |
| 1507 | LoopSafetyInfo *SafetyInfo, |
| 1508 | MemorySSAUpdater *MSSAU) { |
| 1509 | #ifndef NDEBUG |
| 1510 | SmallVector<BasicBlock *, 32> ExitBlocks; |
| 1511 | CurLoop->getUniqueExitBlocks(ExitBlocks); |
| 1512 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(llvm::from_range, ExitBlocks); |
| 1513 | #endif |
| 1514 | BasicBlock *ExitBB = PN->getParent(); |
| 1515 | assert(ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block." ); |
| 1516 | |
| 1517 | // Split predecessors of the loop exit to make instructions in the loop are |
| 1518 | // exposed to exit blocks through trivially replaceable PHIs while keeping the |
| 1519 | // loop in the canonical form where each predecessor of each exit block should |
| 1520 | // be contained within the loop. For example, this will convert the loop below |
| 1521 | // from |
| 1522 | // |
| 1523 | // LB1: |
| 1524 | // %v1 = |
| 1525 | // br %LE, %LB2 |
| 1526 | // LB2: |
| 1527 | // %v2 = |
| 1528 | // br %LE, %LB1 |
| 1529 | // LE: |
| 1530 | // %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable |
| 1531 | // |
| 1532 | // to |
| 1533 | // |
| 1534 | // LB1: |
| 1535 | // %v1 = |
| 1536 | // br %LE.split, %LB2 |
| 1537 | // LB2: |
| 1538 | // %v2 = |
| 1539 | // br %LE.split2, %LB1 |
| 1540 | // LE.split: |
| 1541 | // %p1 = phi [%v1, %LB1] <-- trivially replaceable |
| 1542 | // br %LE |
| 1543 | // LE.split2: |
| 1544 | // %p2 = phi [%v2, %LB2] <-- trivially replaceable |
| 1545 | // br %LE |
| 1546 | // LE: |
| 1547 | // %p = phi [%p1, %LE.split], [%p2, %LE.split2] |
| 1548 | // |
| 1549 | const auto &BlockColors = SafetyInfo->getBlockColors(); |
| 1550 | SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(BB: ExitBB), pred_end(BB: ExitBB)); |
| 1551 | DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy); |
| 1552 | while (!PredBBs.empty()) { |
| 1553 | BasicBlock *PredBB = *PredBBs.begin(); |
| 1554 | assert(CurLoop->contains(PredBB) && |
| 1555 | "Expect all predecessors are in the loop" ); |
| 1556 | if (PN->getBasicBlockIndex(BB: PredBB) >= 0) { |
| 1557 | BasicBlock *NewPred = SplitBlockPredecessors( |
| 1558 | BB: ExitBB, Preds: PredBB, Suffix: ".split.loop.exit" , DTU: &DTU, LI, MSSAU, PreserveLCSSA: true); |
| 1559 | // Since we do not allow splitting EH-block with BlockColors in |
| 1560 | // canSplitPredecessors(), we can simply assign predecessor's color to |
| 1561 | // the new block. |
| 1562 | if (!BlockColors.empty()) |
| 1563 | // Grab a reference to the ColorVector to be inserted before getting the |
| 1564 | // reference to the vector we are copying because inserting the new |
| 1565 | // element in BlockColors might cause the map to be reallocated. |
| 1566 | SafetyInfo->copyColors(New: NewPred, Old: PredBB); |
| 1567 | } |
| 1568 | PredBBs.remove(X: PredBB); |
| 1569 | } |
| 1570 | } |
| 1571 | |
| 1572 | /// When an instruction is found to only be used outside of the loop, this |
| 1573 | /// function moves it to the exit blocks and patches up SSA form as needed. |
| 1574 | /// This method is guaranteed to remove the original instruction from its |
| 1575 | /// position, and may either delete it or move it to outside of the loop. |
| 1576 | /// |
| 1577 | static bool (Instruction &I, LoopInfo *LI, DominatorTree *DT, |
| 1578 | const Loop *CurLoop, ICFLoopSafetyInfo *SafetyInfo, |
| 1579 | MemorySSAUpdater &MSSAU, OptimizationRemarkEmitter *ORE) { |
| 1580 | bool Changed = false; |
| 1581 | LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n" ); |
| 1582 | |
| 1583 | // Iterate over users to be ready for actual sinking. Replace users via |
| 1584 | // unreachable blocks with undef and make all user PHIs trivially replaceable. |
| 1585 | SmallPtrSet<Instruction *, 8> VisitedUsers; |
| 1586 | for (Value::user_iterator UI = I.user_begin(), UE = I.user_end(); UI != UE;) { |
| 1587 | auto *User = cast<Instruction>(Val: *UI); |
| 1588 | Use &U = UI.getUse(); |
| 1589 | ++UI; |
| 1590 | |
| 1591 | if (VisitedUsers.count(Ptr: User) || CurLoop->contains(Inst: User)) |
| 1592 | continue; |
| 1593 | |
| 1594 | if (!DT->isReachableFromEntry(A: User->getParent())) { |
| 1595 | U = PoisonValue::get(T: I.getType()); |
| 1596 | Changed = true; |
| 1597 | continue; |
| 1598 | } |
| 1599 | |
| 1600 | // The user must be a PHI node. |
| 1601 | PHINode *PN = cast<PHINode>(Val: User); |
| 1602 | |
| 1603 | // Surprisingly, instructions can be used outside of loops without any |
| 1604 | // exits. This can only happen in PHI nodes if the incoming block is |
| 1605 | // unreachable. |
| 1606 | BasicBlock *BB = PN->getIncomingBlock(U); |
| 1607 | if (!DT->isReachableFromEntry(A: BB)) { |
| 1608 | U = PoisonValue::get(T: I.getType()); |
| 1609 | Changed = true; |
| 1610 | continue; |
| 1611 | } |
| 1612 | |
| 1613 | VisitedUsers.insert(Ptr: PN); |
| 1614 | if (isTriviallyReplaceablePHI(PN: *PN, I)) |
| 1615 | continue; |
| 1616 | |
| 1617 | if (!canSplitPredecessors(PN, SafetyInfo)) |
| 1618 | return Changed; |
| 1619 | |
| 1620 | // Split predecessors of the PHI so that we can make users trivially |
| 1621 | // replaceable. |
| 1622 | splitPredecessorsOfLoopExit(PN, DT, LI, CurLoop, SafetyInfo, MSSAU: &MSSAU); |
| 1623 | |
| 1624 | // Should rebuild the iterators, as they may be invalidated by |
| 1625 | // splitPredecessorsOfLoopExit(). |
| 1626 | UI = I.user_begin(); |
| 1627 | UE = I.user_end(); |
| 1628 | } |
| 1629 | |
| 1630 | if (VisitedUsers.empty()) |
| 1631 | return Changed; |
| 1632 | |
| 1633 | ORE->emit(RemarkBuilder: [&]() { |
| 1634 | return OptimizationRemark(DEBUG_TYPE, "InstSunk" , &I) |
| 1635 | << "sinking " << ore::NV("Inst" , &I); |
| 1636 | }); |
| 1637 | if (isa<LoadInst>(Val: I)) |
| 1638 | ++NumMovedLoads; |
| 1639 | else if (isa<CallInst>(Val: I)) |
| 1640 | ++NumMovedCalls; |
| 1641 | ++NumSunk; |
| 1642 | |
| 1643 | #ifndef NDEBUG |
| 1644 | SmallVector<BasicBlock *, 32> ExitBlocks; |
| 1645 | CurLoop->getUniqueExitBlocks(ExitBlocks); |
| 1646 | SmallPtrSet<BasicBlock *, 32> ExitBlockSet(llvm::from_range, ExitBlocks); |
| 1647 | #endif |
| 1648 | |
| 1649 | // Clones of this instruction. Don't create more than one per exit block! |
| 1650 | SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies; |
| 1651 | |
| 1652 | // If this instruction is only used outside of the loop, then all users are |
| 1653 | // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of |
| 1654 | // the instruction. |
| 1655 | // First check if I is worth sinking for all uses. Sink only when it is worth |
| 1656 | // across all uses. |
| 1657 | SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end()); |
| 1658 | for (auto *UI : Users) { |
| 1659 | auto *User = cast<Instruction>(Val: UI); |
| 1660 | |
| 1661 | if (CurLoop->contains(Inst: User)) |
| 1662 | continue; |
| 1663 | |
| 1664 | PHINode *PN = cast<PHINode>(Val: User); |
| 1665 | assert(ExitBlockSet.count(PN->getParent()) && |
| 1666 | "The LCSSA PHI is not in an exit block!" ); |
| 1667 | |
| 1668 | // The PHI must be trivially replaceable. |
| 1669 | Instruction *New = sinkThroughTriviallyReplaceablePHI( |
| 1670 | TPN: PN, I: &I, LI, SunkCopies, SafetyInfo, CurLoop, MSSAU); |
| 1671 | // As we sink the instruction out of the BB, drop its debug location. |
| 1672 | New->dropLocation(); |
| 1673 | PN->replaceAllUsesWith(V: New); |
| 1674 | eraseInstruction(I&: *PN, SafetyInfo&: *SafetyInfo, MSSAU); |
| 1675 | Changed = true; |
| 1676 | } |
| 1677 | return Changed; |
| 1678 | } |
| 1679 | |
| 1680 | /// When an instruction is found to only use loop invariant operands that |
| 1681 | /// is safe to hoist, this instruction is called to do the dirty work. |
| 1682 | /// |
| 1683 | static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop, |
| 1684 | BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo, |
| 1685 | MemorySSAUpdater &MSSAU, ScalarEvolution *SE, |
| 1686 | OptimizationRemarkEmitter *ORE) { |
| 1687 | LLVM_DEBUG(dbgs() << "LICM hoisting to " << Dest->getNameOrAsOperand() << ": " |
| 1688 | << I << "\n" ); |
| 1689 | ORE->emit(RemarkBuilder: [&]() { |
| 1690 | return OptimizationRemark(DEBUG_TYPE, "Hoisted" , &I) << "hoisting " |
| 1691 | << ore::NV("Inst" , &I); |
| 1692 | }); |
| 1693 | |
| 1694 | // Metadata can be dependent on conditions we are hoisting above. |
| 1695 | // Conservatively strip all metadata on the instruction unless we were |
| 1696 | // guaranteed to execute I if we entered the loop, in which case the metadata |
| 1697 | // is valid in the loop preheader. |
| 1698 | // Similarly, If I is a call and it is not guaranteed to execute in the loop, |
| 1699 | // then moving to the preheader means we should strip attributes on the call |
| 1700 | // that can cause UB since we may be hoisting above conditions that allowed |
| 1701 | // inferring those attributes. They may not be valid at the preheader. |
| 1702 | if ((I.hasMetadataOtherThanDebugLoc() || isa<CallInst>(Val: I)) && |
| 1703 | // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning |
| 1704 | // time in isGuaranteedToExecute if we don't actually have anything to |
| 1705 | // drop. It is a compile time optimization, not required for correctness. |
| 1706 | !SafetyInfo->isGuaranteedToExecute(Inst: I, DT, CurLoop)) { |
| 1707 | I.dropUBImplyingAttrsAndMetadata(); |
| 1708 | } |
| 1709 | |
| 1710 | if (isa<PHINode>(Val: I)) |
| 1711 | // Move the new node to the end of the phi list in the destination block. |
| 1712 | moveInstructionBefore(I, Dest: Dest->getFirstNonPHIIt(), SafetyInfo&: *SafetyInfo, MSSAU, SE); |
| 1713 | else |
| 1714 | // Move the new node to the destination block, before its terminator. |
| 1715 | moveInstructionBefore(I, Dest: Dest->getTerminator()->getIterator(), SafetyInfo&: *SafetyInfo, |
| 1716 | MSSAU, SE); |
| 1717 | |
| 1718 | I.updateLocationAfterHoist(); |
| 1719 | |
| 1720 | if (isa<LoadInst>(Val: I)) |
| 1721 | ++NumMovedLoads; |
| 1722 | else if (isa<CallInst>(Val: I)) |
| 1723 | ++NumMovedCalls; |
| 1724 | ++NumHoisted; |
| 1725 | } |
| 1726 | |
| 1727 | /// Only sink or hoist an instruction if it is not a trapping instruction, |
| 1728 | /// or if the instruction is known not to trap when moved to the preheader. |
| 1729 | /// or if it is a trapping instruction and is guaranteed to execute. |
| 1730 | static bool isSafeToExecuteUnconditionally( |
| 1731 | Instruction &Inst, const DominatorTree *DT, const TargetLibraryInfo *TLI, |
| 1732 | const Loop *CurLoop, const LoopSafetyInfo *SafetyInfo, |
| 1733 | OptimizationRemarkEmitter *ORE, const Instruction *CtxI, |
| 1734 | AssumptionCache *AC, bool AllowSpeculation) { |
| 1735 | if (AllowSpeculation && |
| 1736 | isSafeToSpeculativelyExecute(I: &Inst, CtxI, AC, DT, TLI)) |
| 1737 | return true; |
| 1738 | |
| 1739 | bool GuaranteedToExecute = |
| 1740 | SafetyInfo->isGuaranteedToExecute(Inst, DT, CurLoop); |
| 1741 | |
| 1742 | if (!GuaranteedToExecute) { |
| 1743 | auto *LI = dyn_cast<LoadInst>(Val: &Inst); |
| 1744 | if (LI && CurLoop->isLoopInvariant(V: LI->getPointerOperand())) |
| 1745 | ORE->emit(RemarkBuilder: [&]() { |
| 1746 | return OptimizationRemarkMissed( |
| 1747 | DEBUG_TYPE, "LoadWithLoopInvariantAddressCondExecuted" , LI) |
| 1748 | << "failed to hoist load with loop-invariant address " |
| 1749 | "because load is conditionally executed" ; |
| 1750 | }); |
| 1751 | } |
| 1752 | |
| 1753 | return GuaranteedToExecute; |
| 1754 | } |
| 1755 | |
| 1756 | namespace { |
| 1757 | class LoopPromoter : public LoadAndStorePromoter { |
| 1758 | Value *SomePtr; // Designated pointer to store to. |
| 1759 | SmallVectorImpl<BasicBlock *> &LoopExitBlocks; |
| 1760 | SmallVectorImpl<BasicBlock::iterator> &LoopInsertPts; |
| 1761 | SmallVectorImpl<MemoryAccess *> &MSSAInsertPts; |
| 1762 | PredIteratorCache &PredCache; |
| 1763 | MemorySSAUpdater &MSSAU; |
| 1764 | LoopInfo &LI; |
| 1765 | DebugLoc DL; |
| 1766 | Align Alignment; |
| 1767 | bool UnorderedAtomic; |
| 1768 | AAMDNodes AATags; |
| 1769 | ICFLoopSafetyInfo &SafetyInfo; |
| 1770 | bool CanInsertStoresInExitBlocks; |
| 1771 | ArrayRef<const Instruction *> Uses; |
| 1772 | |
| 1773 | // We're about to add a use of V in a loop exit block. Insert an LCSSA phi |
| 1774 | // (if legal) if doing so would add an out-of-loop use to an instruction |
| 1775 | // defined in-loop. |
| 1776 | Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const { |
| 1777 | if (!LI.wouldBeOutOfLoopUseRequiringLCSSA(V, ExitBB: BB)) |
| 1778 | return V; |
| 1779 | |
| 1780 | Instruction *I = cast<Instruction>(Val: V); |
| 1781 | // We need to create an LCSSA PHI node for the incoming value and |
| 1782 | // store that. |
| 1783 | PHINode *PN = PHINode::Create(Ty: I->getType(), NumReservedValues: PredCache.size(BB), |
| 1784 | NameStr: I->getName() + ".lcssa" ); |
| 1785 | PN->insertBefore(InsertPos: BB->begin()); |
| 1786 | for (BasicBlock *Pred : PredCache.get(BB)) |
| 1787 | PN->addIncoming(V: I, BB: Pred); |
| 1788 | return PN; |
| 1789 | } |
| 1790 | |
| 1791 | public: |
| 1792 | LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S, |
| 1793 | SmallVectorImpl<BasicBlock *> &LEB, |
| 1794 | SmallVectorImpl<BasicBlock::iterator> &LIP, |
| 1795 | SmallVectorImpl<MemoryAccess *> &MSSAIP, PredIteratorCache &PIC, |
| 1796 | MemorySSAUpdater &MSSAU, LoopInfo &li, DebugLoc dl, |
| 1797 | Align Alignment, bool UnorderedAtomic, const AAMDNodes &AATags, |
| 1798 | ICFLoopSafetyInfo &SafetyInfo, bool CanInsertStoresInExitBlocks) |
| 1799 | : LoadAndStorePromoter(Insts, S), SomePtr(SP), LoopExitBlocks(LEB), |
| 1800 | LoopInsertPts(LIP), MSSAInsertPts(MSSAIP), PredCache(PIC), MSSAU(MSSAU), |
| 1801 | LI(li), DL(std::move(dl)), Alignment(Alignment), |
| 1802 | UnorderedAtomic(UnorderedAtomic), AATags(AATags), |
| 1803 | SafetyInfo(SafetyInfo), |
| 1804 | CanInsertStoresInExitBlocks(CanInsertStoresInExitBlocks), Uses(Insts) {} |
| 1805 | |
| 1806 | void insertStoresInLoopExitBlocks() { |
| 1807 | // Insert stores after in the loop exit blocks. Each exit block gets a |
| 1808 | // store of the live-out values that feed them. Since we've already told |
| 1809 | // the SSA updater about the defs in the loop and the preheader |
| 1810 | // definition, it is all set and we can start using it. |
| 1811 | DIAssignID *NewID = nullptr; |
| 1812 | for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { |
| 1813 | BasicBlock *ExitBlock = LoopExitBlocks[i]; |
| 1814 | Value *LiveInValue = SSA.GetValueInMiddleOfBlock(BB: ExitBlock); |
| 1815 | LiveInValue = maybeInsertLCSSAPHI(V: LiveInValue, BB: ExitBlock); |
| 1816 | Value *Ptr = maybeInsertLCSSAPHI(V: SomePtr, BB: ExitBlock); |
| 1817 | BasicBlock::iterator InsertPos = LoopInsertPts[i]; |
| 1818 | StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos); |
| 1819 | if (UnorderedAtomic) |
| 1820 | NewSI->setOrdering(AtomicOrdering::Unordered); |
| 1821 | NewSI->setAlignment(Alignment); |
| 1822 | NewSI->setDebugLoc(DL); |
| 1823 | // Attach DIAssignID metadata to the new store, generating it on the |
| 1824 | // first loop iteration. |
| 1825 | if (i == 0) { |
| 1826 | // NewSI will have its DIAssignID set here if there are any stores in |
| 1827 | // Uses with a DIAssignID attachment. This merged ID will then be |
| 1828 | // attached to the other inserted stores (in the branch below). |
| 1829 | NewSI->mergeDIAssignID(SourceInstructions: Uses); |
| 1830 | NewID = cast_or_null<DIAssignID>( |
| 1831 | Val: NewSI->getMetadata(KindID: LLVMContext::MD_DIAssignID)); |
| 1832 | } else { |
| 1833 | // Attach the DIAssignID (or nullptr) merged from Uses in the branch |
| 1834 | // above. |
| 1835 | NewSI->setMetadata(KindID: LLVMContext::MD_DIAssignID, Node: NewID); |
| 1836 | } |
| 1837 | |
| 1838 | if (AATags) |
| 1839 | NewSI->setAAMetadata(AATags); |
| 1840 | |
| 1841 | MemoryAccess *MSSAInsertPoint = MSSAInsertPts[i]; |
| 1842 | MemoryAccess *NewMemAcc; |
| 1843 | if (!MSSAInsertPoint) { |
| 1844 | NewMemAcc = MSSAU.createMemoryAccessInBB( |
| 1845 | I: NewSI, Definition: nullptr, BB: NewSI->getParent(), Point: MemorySSA::Beginning); |
| 1846 | } else { |
| 1847 | NewMemAcc = |
| 1848 | MSSAU.createMemoryAccessAfter(I: NewSI, Definition: nullptr, InsertPt: MSSAInsertPoint); |
| 1849 | } |
| 1850 | MSSAInsertPts[i] = NewMemAcc; |
| 1851 | MSSAU.insertDef(Def: cast<MemoryDef>(Val: NewMemAcc), RenameUses: true); |
| 1852 | // FIXME: true for safety, false may still be correct. |
| 1853 | } |
| 1854 | } |
| 1855 | |
| 1856 | void () override { |
| 1857 | if (CanInsertStoresInExitBlocks) |
| 1858 | insertStoresInLoopExitBlocks(); |
| 1859 | } |
| 1860 | |
| 1861 | void instructionDeleted(Instruction *I) const override { |
| 1862 | SafetyInfo.removeInstruction(Inst: I); |
| 1863 | MSSAU.removeMemoryAccess(I); |
| 1864 | } |
| 1865 | |
| 1866 | bool shouldDelete(Instruction *I) const override { |
| 1867 | if (isa<StoreInst>(Val: I)) |
| 1868 | return CanInsertStoresInExitBlocks; |
| 1869 | return true; |
| 1870 | } |
| 1871 | }; |
| 1872 | |
| 1873 | bool isNotCapturedBeforeOrInLoop(const Value *V, const Loop *L, |
| 1874 | DominatorTree *DT) { |
| 1875 | // We can perform the captured-before check against any instruction in the |
| 1876 | // loop header, as the loop header is reachable from any instruction inside |
| 1877 | // the loop. |
| 1878 | // TODO: ReturnCaptures=true shouldn't be necessary here. |
| 1879 | return capturesNothing(CC: PointerMayBeCapturedBefore( |
| 1880 | V, /*ReturnCaptures=*/true, I: L->getHeader()->getTerminator(), DT, |
| 1881 | /*IncludeI=*/false, Mask: CaptureComponents::Provenance)); |
| 1882 | } |
| 1883 | |
| 1884 | /// Return true if we can prove that a caller cannot inspect the object if an |
| 1885 | /// unwind occurs inside the loop. |
| 1886 | bool isNotVisibleOnUnwindInLoop(const Value *Object, const Loop *L, |
| 1887 | DominatorTree *DT) { |
| 1888 | bool RequiresNoCaptureBeforeUnwind; |
| 1889 | if (!isNotVisibleOnUnwind(Object, RequiresNoCaptureBeforeUnwind)) |
| 1890 | return false; |
| 1891 | |
| 1892 | return !RequiresNoCaptureBeforeUnwind || |
| 1893 | isNotCapturedBeforeOrInLoop(V: Object, L, DT); |
| 1894 | } |
| 1895 | |
| 1896 | bool isThreadLocalObject(const Value *Object, const Loop *L, DominatorTree *DT, |
| 1897 | TargetTransformInfo *TTI) { |
| 1898 | // The object must be function-local to start with, and then not captured |
| 1899 | // before/in the loop. |
| 1900 | return (isIdentifiedFunctionLocal(V: Object) && |
| 1901 | isNotCapturedBeforeOrInLoop(V: Object, L, DT)) || |
| 1902 | (TTI->isSingleThreaded() || SingleThread); |
| 1903 | } |
| 1904 | |
| 1905 | } // namespace |
| 1906 | |
| 1907 | /// Try to promote memory values to scalars by sinking stores out of the |
| 1908 | /// loop and moving loads to before the loop. We do this by looping over |
| 1909 | /// the stores in the loop, looking for stores to Must pointers which are |
| 1910 | /// loop invariant. |
| 1911 | /// |
| 1912 | bool llvm::promoteLoopAccessesToScalars( |
| 1913 | const SmallSetVector<Value *, 8> &PointerMustAliases, |
| 1914 | SmallVectorImpl<BasicBlock *> &ExitBlocks, |
| 1915 | SmallVectorImpl<BasicBlock::iterator> &InsertPts, |
| 1916 | SmallVectorImpl<MemoryAccess *> &MSSAInsertPts, PredIteratorCache &PIC, |
| 1917 | LoopInfo *LI, DominatorTree *DT, AssumptionCache *AC, |
| 1918 | const TargetLibraryInfo *TLI, TargetTransformInfo *TTI, Loop *CurLoop, |
| 1919 | MemorySSAUpdater &MSSAU, ICFLoopSafetyInfo *SafetyInfo, |
| 1920 | OptimizationRemarkEmitter *ORE, bool AllowSpeculation, |
| 1921 | bool HasReadsOutsideSet) { |
| 1922 | // Verify inputs. |
| 1923 | assert(LI != nullptr && DT != nullptr && CurLoop != nullptr && |
| 1924 | SafetyInfo != nullptr && |
| 1925 | "Unexpected Input to promoteLoopAccessesToScalars" ); |
| 1926 | |
| 1927 | LLVM_DEBUG({ |
| 1928 | dbgs() << "Trying to promote set of must-aliased pointers:\n" ; |
| 1929 | for (Value *Ptr : PointerMustAliases) |
| 1930 | dbgs() << " " << *Ptr << "\n" ; |
| 1931 | }); |
| 1932 | ++NumPromotionCandidates; |
| 1933 | |
| 1934 | Value *SomePtr = *PointerMustAliases.begin(); |
| 1935 | BasicBlock * = CurLoop->getLoopPreheader(); |
| 1936 | |
| 1937 | // It is not safe to promote a load/store from the loop if the load/store is |
| 1938 | // conditional. For example, turning: |
| 1939 | // |
| 1940 | // for () { if (c) *P += 1; } |
| 1941 | // |
| 1942 | // into: |
| 1943 | // |
| 1944 | // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; |
| 1945 | // |
| 1946 | // is not safe, because *P may only be valid to access if 'c' is true. |
| 1947 | // |
| 1948 | // The safety property divides into two parts: |
| 1949 | // p1) The memory may not be dereferenceable on entry to the loop. In this |
| 1950 | // case, we can't insert the required load in the preheader. |
| 1951 | // p2) The memory model does not allow us to insert a store along any dynamic |
| 1952 | // path which did not originally have one. |
| 1953 | // |
| 1954 | // If at least one store is guaranteed to execute, both properties are |
| 1955 | // satisfied, and promotion is legal. |
| 1956 | // |
| 1957 | // This, however, is not a necessary condition. Even if no store/load is |
| 1958 | // guaranteed to execute, we can still establish these properties. |
| 1959 | // We can establish (p1) by proving that hoisting the load into the preheader |
| 1960 | // is safe (i.e. proving dereferenceability on all paths through the loop). We |
| 1961 | // can use any access within the alias set to prove dereferenceability, |
| 1962 | // since they're all must alias. |
| 1963 | // |
| 1964 | // There are two ways establish (p2): |
| 1965 | // a) Prove the location is thread-local. In this case the memory model |
| 1966 | // requirement does not apply, and stores are safe to insert. |
| 1967 | // b) Prove a store dominates every exit block. In this case, if an exit |
| 1968 | // blocks is reached, the original dynamic path would have taken us through |
| 1969 | // the store, so inserting a store into the exit block is safe. Note that this |
| 1970 | // is different from the store being guaranteed to execute. For instance, |
| 1971 | // if an exception is thrown on the first iteration of the loop, the original |
| 1972 | // store is never executed, but the exit blocks are not executed either. |
| 1973 | |
| 1974 | bool DereferenceableInPH = false; |
| 1975 | bool StoreIsGuanteedToExecute = false; |
| 1976 | bool LoadIsGuaranteedToExecute = false; |
| 1977 | bool FoundLoadToPromote = false; |
| 1978 | |
| 1979 | // Goes from Unknown to either Safe or Unsafe, but can't switch between them. |
| 1980 | enum { |
| 1981 | StoreSafe, |
| 1982 | StoreUnsafe, |
| 1983 | StoreSafetyUnknown, |
| 1984 | } StoreSafety = StoreSafetyUnknown; |
| 1985 | |
| 1986 | SmallVector<Instruction *, 64> LoopUses; |
| 1987 | |
| 1988 | // We start with an alignment of one and try to find instructions that allow |
| 1989 | // us to prove better alignment. |
| 1990 | Align Alignment; |
| 1991 | // Keep track of which types of access we see |
| 1992 | bool SawUnorderedAtomic = false; |
| 1993 | bool SawNotAtomic = false; |
| 1994 | AAMDNodes AATags; |
| 1995 | |
| 1996 | const DataLayout &MDL = Preheader->getDataLayout(); |
| 1997 | |
| 1998 | // If there are reads outside the promoted set, then promoting stores is |
| 1999 | // definitely not safe. |
| 2000 | if (HasReadsOutsideSet) |
| 2001 | StoreSafety = StoreUnsafe; |
| 2002 | |
| 2003 | if (StoreSafety == StoreSafetyUnknown && SafetyInfo->anyBlockMayThrow()) { |
| 2004 | // If a loop can throw, we have to insert a store along each unwind edge. |
| 2005 | // That said, we can't actually make the unwind edge explicit. Therefore, |
| 2006 | // we have to prove that the store is dead along the unwind edge. We do |
| 2007 | // this by proving that the caller can't have a reference to the object |
| 2008 | // after return and thus can't possibly load from the object. |
| 2009 | Value *Object = getUnderlyingObject(V: SomePtr); |
| 2010 | if (!isNotVisibleOnUnwindInLoop(Object, L: CurLoop, DT)) |
| 2011 | StoreSafety = StoreUnsafe; |
| 2012 | } |
| 2013 | |
| 2014 | // Check that all accesses to pointers in the alias set use the same type. |
| 2015 | // We cannot (yet) promote a memory location that is loaded and stored in |
| 2016 | // different sizes. While we are at it, collect alignment and AA info. |
| 2017 | Type *AccessTy = nullptr; |
| 2018 | for (Value *ASIV : PointerMustAliases) { |
| 2019 | for (Use &U : ASIV->uses()) { |
| 2020 | // Ignore instructions that are outside the loop. |
| 2021 | Instruction *UI = dyn_cast<Instruction>(Val: U.getUser()); |
| 2022 | if (!UI || !CurLoop->contains(Inst: UI)) |
| 2023 | continue; |
| 2024 | |
| 2025 | // If there is an non-load/store instruction in the loop, we can't promote |
| 2026 | // it. |
| 2027 | if (LoadInst *Load = dyn_cast<LoadInst>(Val: UI)) { |
| 2028 | if (!Load->isUnordered()) |
| 2029 | return false; |
| 2030 | |
| 2031 | SawUnorderedAtomic |= Load->isAtomic(); |
| 2032 | SawNotAtomic |= !Load->isAtomic(); |
| 2033 | FoundLoadToPromote = true; |
| 2034 | |
| 2035 | Align InstAlignment = Load->getAlign(); |
| 2036 | |
| 2037 | if (!LoadIsGuaranteedToExecute) |
| 2038 | LoadIsGuaranteedToExecute = |
| 2039 | SafetyInfo->isGuaranteedToExecute(Inst: *UI, DT, CurLoop); |
| 2040 | |
| 2041 | // Note that proving a load safe to speculate requires proving |
| 2042 | // sufficient alignment at the target location. Proving it guaranteed |
| 2043 | // to execute does as well. Thus we can increase our guaranteed |
| 2044 | // alignment as well. |
| 2045 | if (!DereferenceableInPH || (InstAlignment > Alignment)) |
| 2046 | if (isSafeToExecuteUnconditionally( |
| 2047 | Inst&: *Load, DT, TLI, CurLoop, SafetyInfo, ORE, |
| 2048 | CtxI: Preheader->getTerminator(), AC, AllowSpeculation)) { |
| 2049 | DereferenceableInPH = true; |
| 2050 | Alignment = std::max(a: Alignment, b: InstAlignment); |
| 2051 | } |
| 2052 | } else if (const StoreInst *Store = dyn_cast<StoreInst>(Val: UI)) { |
| 2053 | // Stores *of* the pointer are not interesting, only stores *to* the |
| 2054 | // pointer. |
| 2055 | if (U.getOperandNo() != StoreInst::getPointerOperandIndex()) |
| 2056 | continue; |
| 2057 | if (!Store->isUnordered()) |
| 2058 | return false; |
| 2059 | |
| 2060 | SawUnorderedAtomic |= Store->isAtomic(); |
| 2061 | SawNotAtomic |= !Store->isAtomic(); |
| 2062 | |
| 2063 | // If the store is guaranteed to execute, both properties are satisfied. |
| 2064 | // We may want to check if a store is guaranteed to execute even if we |
| 2065 | // already know that promotion is safe, since it may have higher |
| 2066 | // alignment than any other guaranteed stores, in which case we can |
| 2067 | // raise the alignment on the promoted store. |
| 2068 | Align InstAlignment = Store->getAlign(); |
| 2069 | bool GuaranteedToExecute = |
| 2070 | SafetyInfo->isGuaranteedToExecute(Inst: *UI, DT, CurLoop); |
| 2071 | StoreIsGuanteedToExecute |= GuaranteedToExecute; |
| 2072 | if (GuaranteedToExecute) { |
| 2073 | DereferenceableInPH = true; |
| 2074 | if (StoreSafety == StoreSafetyUnknown) |
| 2075 | StoreSafety = StoreSafe; |
| 2076 | Alignment = std::max(a: Alignment, b: InstAlignment); |
| 2077 | } |
| 2078 | |
| 2079 | // If a store dominates all exit blocks, it is safe to sink. |
| 2080 | // As explained above, if an exit block was executed, a dominating |
| 2081 | // store must have been executed at least once, so we are not |
| 2082 | // introducing stores on paths that did not have them. |
| 2083 | // Note that this only looks at explicit exit blocks. If we ever |
| 2084 | // start sinking stores into unwind edges (see above), this will break. |
| 2085 | if (StoreSafety == StoreSafetyUnknown && |
| 2086 | llvm::all_of(Range&: ExitBlocks, P: [&](BasicBlock *Exit) { |
| 2087 | return DT->dominates(A: Store->getParent(), B: Exit); |
| 2088 | })) |
| 2089 | StoreSafety = StoreSafe; |
| 2090 | |
| 2091 | // If the store is not guaranteed to execute, we may still get |
| 2092 | // deref info through it. |
| 2093 | if (!DereferenceableInPH) { |
| 2094 | DereferenceableInPH = isDereferenceableAndAlignedPointer( |
| 2095 | V: Store->getPointerOperand(), Ty: Store->getValueOperand()->getType(), |
| 2096 | Alignment: Store->getAlign(), DL: MDL, CtxI: Preheader->getTerminator(), AC, DT, TLI); |
| 2097 | } |
| 2098 | } else |
| 2099 | continue; // Not a load or store. |
| 2100 | |
| 2101 | if (!AccessTy) |
| 2102 | AccessTy = getLoadStoreType(I: UI); |
| 2103 | else if (AccessTy != getLoadStoreType(I: UI)) |
| 2104 | return false; |
| 2105 | |
| 2106 | // Merge the AA tags. |
| 2107 | if (LoopUses.empty()) { |
| 2108 | // On the first load/store, just take its AA tags. |
| 2109 | AATags = UI->getAAMetadata(); |
| 2110 | } else if (AATags) { |
| 2111 | AATags = AATags.merge(Other: UI->getAAMetadata()); |
| 2112 | } |
| 2113 | |
| 2114 | LoopUses.push_back(Elt: UI); |
| 2115 | } |
| 2116 | } |
| 2117 | |
| 2118 | // If we found both an unordered atomic instruction and a non-atomic memory |
| 2119 | // access, bail. We can't blindly promote non-atomic to atomic since we |
| 2120 | // might not be able to lower the result. We can't downgrade since that |
| 2121 | // would violate memory model. Also, align 0 is an error for atomics. |
| 2122 | if (SawUnorderedAtomic && SawNotAtomic) |
| 2123 | return false; |
| 2124 | |
| 2125 | // If we're inserting an atomic load in the preheader, we must be able to |
| 2126 | // lower it. We're only guaranteed to be able to lower naturally aligned |
| 2127 | // atomics. |
| 2128 | if (SawUnorderedAtomic && Alignment < MDL.getTypeStoreSize(Ty: AccessTy)) |
| 2129 | return false; |
| 2130 | |
| 2131 | // If we couldn't prove we can hoist the load, bail. |
| 2132 | if (!DereferenceableInPH) { |
| 2133 | LLVM_DEBUG(dbgs() << "Not promoting: Not dereferenceable in preheader\n" ); |
| 2134 | return false; |
| 2135 | } |
| 2136 | |
| 2137 | // We know we can hoist the load, but don't have a guaranteed store. |
| 2138 | // Check whether the location is writable and thread-local. If it is, then we |
| 2139 | // can insert stores along paths which originally didn't have them without |
| 2140 | // violating the memory model. |
| 2141 | if (StoreSafety == StoreSafetyUnknown) { |
| 2142 | Value *Object = getUnderlyingObject(V: SomePtr); |
| 2143 | bool ExplicitlyDereferenceableOnly; |
| 2144 | if (isWritableObject(Object, ExplicitlyDereferenceableOnly) && |
| 2145 | (!ExplicitlyDereferenceableOnly || |
| 2146 | isDereferenceablePointer(V: SomePtr, Ty: AccessTy, DL: MDL)) && |
| 2147 | isThreadLocalObject(Object, L: CurLoop, DT, TTI)) |
| 2148 | StoreSafety = StoreSafe; |
| 2149 | } |
| 2150 | |
| 2151 | // If we've still failed to prove we can sink the store, hoist the load |
| 2152 | // only, if possible. |
| 2153 | if (StoreSafety != StoreSafe && !FoundLoadToPromote) |
| 2154 | // If we cannot hoist the load either, give up. |
| 2155 | return false; |
| 2156 | |
| 2157 | // Lets do the promotion! |
| 2158 | if (StoreSafety == StoreSafe) { |
| 2159 | LLVM_DEBUG(dbgs() << "LICM: Promoting load/store of the value: " << *SomePtr |
| 2160 | << '\n'); |
| 2161 | ++NumLoadStorePromoted; |
| 2162 | } else { |
| 2163 | LLVM_DEBUG(dbgs() << "LICM: Promoting load of the value: " << *SomePtr |
| 2164 | << '\n'); |
| 2165 | ++NumLoadPromoted; |
| 2166 | } |
| 2167 | |
| 2168 | ORE->emit(RemarkBuilder: [&]() { |
| 2169 | return OptimizationRemark(DEBUG_TYPE, "PromoteLoopAccessesToScalar" , |
| 2170 | LoopUses[0]) |
| 2171 | << "Moving accesses to memory location out of the loop" ; |
| 2172 | }); |
| 2173 | |
| 2174 | // Look at all the loop uses, and try to merge their locations. |
| 2175 | std::vector<DebugLoc> LoopUsesLocs; |
| 2176 | for (auto U : LoopUses) |
| 2177 | LoopUsesLocs.push_back(x: U->getDebugLoc()); |
| 2178 | auto DL = DebugLoc::getMergedLocations(Locs: LoopUsesLocs); |
| 2179 | |
| 2180 | // We use the SSAUpdater interface to insert phi nodes as required. |
| 2181 | SmallVector<PHINode *, 16> NewPHIs; |
| 2182 | SSAUpdater SSA(&NewPHIs); |
| 2183 | LoopPromoter Promoter(SomePtr, LoopUses, SSA, ExitBlocks, InsertPts, |
| 2184 | MSSAInsertPts, PIC, MSSAU, *LI, DL, Alignment, |
| 2185 | SawUnorderedAtomic, |
| 2186 | StoreIsGuanteedToExecute ? AATags : AAMDNodes(), |
| 2187 | *SafetyInfo, StoreSafety == StoreSafe); |
| 2188 | |
| 2189 | // Set up the preheader to have a definition of the value. It is the live-out |
| 2190 | // value from the preheader that uses in the loop will use. |
| 2191 | LoadInst * = nullptr; |
| 2192 | if (FoundLoadToPromote || !StoreIsGuanteedToExecute) { |
| 2193 | PreheaderLoad = |
| 2194 | new LoadInst(AccessTy, SomePtr, SomePtr->getName() + ".promoted" , |
| 2195 | Preheader->getTerminator()->getIterator()); |
| 2196 | if (SawUnorderedAtomic) |
| 2197 | PreheaderLoad->setOrdering(AtomicOrdering::Unordered); |
| 2198 | PreheaderLoad->setAlignment(Alignment); |
| 2199 | PreheaderLoad->setDebugLoc(DebugLoc::getDropped()); |
| 2200 | if (AATags && LoadIsGuaranteedToExecute) |
| 2201 | PreheaderLoad->setAAMetadata(AATags); |
| 2202 | |
| 2203 | MemoryAccess * = MSSAU.createMemoryAccessInBB( |
| 2204 | I: PreheaderLoad, Definition: nullptr, BB: PreheaderLoad->getParent(), Point: MemorySSA::End); |
| 2205 | MemoryUse *NewMemUse = cast<MemoryUse>(Val: PreheaderLoadMemoryAccess); |
| 2206 | MSSAU.insertUse(Use: NewMemUse, /*RenameUses=*/true); |
| 2207 | SSA.AddAvailableValue(BB: Preheader, V: PreheaderLoad); |
| 2208 | } else { |
| 2209 | SSA.AddAvailableValue(BB: Preheader, V: PoisonValue::get(T: AccessTy)); |
| 2210 | } |
| 2211 | |
| 2212 | if (VerifyMemorySSA) |
| 2213 | MSSAU.getMemorySSA()->verifyMemorySSA(); |
| 2214 | // Rewrite all the loads in the loop and remember all the definitions from |
| 2215 | // stores in the loop. |
| 2216 | Promoter.run(Insts: LoopUses); |
| 2217 | |
| 2218 | if (VerifyMemorySSA) |
| 2219 | MSSAU.getMemorySSA()->verifyMemorySSA(); |
| 2220 | // If the SSAUpdater didn't use the load in the preheader, just zap it now. |
| 2221 | if (PreheaderLoad && PreheaderLoad->use_empty()) |
| 2222 | eraseInstruction(I&: *PreheaderLoad, SafetyInfo&: *SafetyInfo, MSSAU); |
| 2223 | |
| 2224 | return true; |
| 2225 | } |
| 2226 | |
| 2227 | static void foreachMemoryAccess(MemorySSA *MSSA, Loop *L, |
| 2228 | function_ref<void(Instruction *)> Fn) { |
| 2229 | for (const BasicBlock *BB : L->blocks()) |
| 2230 | if (const auto *Accesses = MSSA->getBlockAccesses(BB)) |
| 2231 | for (const auto &Access : *Accesses) |
| 2232 | if (const auto *MUD = dyn_cast<MemoryUseOrDef>(Val: &Access)) |
| 2233 | Fn(MUD->getMemoryInst()); |
| 2234 | } |
| 2235 | |
| 2236 | // The bool indicates whether there might be reads outside the set, in which |
| 2237 | // case only loads may be promoted. |
| 2238 | static SmallVector<PointersAndHasReadsOutsideSet, 0> |
| 2239 | collectPromotionCandidates(MemorySSA *MSSA, AliasAnalysis *AA, Loop *L) { |
| 2240 | BatchAAResults BatchAA(*AA); |
| 2241 | AliasSetTracker AST(BatchAA); |
| 2242 | |
| 2243 | auto IsPotentiallyPromotable = [L](const Instruction *I) { |
| 2244 | if (const auto *SI = dyn_cast<StoreInst>(Val: I)) { |
| 2245 | const Value *PtrOp = SI->getPointerOperand(); |
| 2246 | return !isa<ConstantData>(Val: PtrOp) && L->isLoopInvariant(V: PtrOp); |
| 2247 | } |
| 2248 | if (const auto *LI = dyn_cast<LoadInst>(Val: I)) { |
| 2249 | const Value *PtrOp = LI->getPointerOperand(); |
| 2250 | return !isa<ConstantData>(Val: PtrOp) && L->isLoopInvariant(V: PtrOp); |
| 2251 | } |
| 2252 | return false; |
| 2253 | }; |
| 2254 | |
| 2255 | // Populate AST with potentially promotable accesses. |
| 2256 | SmallPtrSet<Value *, 16> AttemptingPromotion; |
| 2257 | foreachMemoryAccess(MSSA, L, Fn: [&](Instruction *I) { |
| 2258 | if (IsPotentiallyPromotable(I)) { |
| 2259 | AttemptingPromotion.insert(Ptr: I); |
| 2260 | AST.add(I); |
| 2261 | } |
| 2262 | }); |
| 2263 | |
| 2264 | // We're only interested in must-alias sets that contain a mod. |
| 2265 | SmallVector<PointerIntPair<const AliasSet *, 1, bool>, 8> Sets; |
| 2266 | for (AliasSet &AS : AST) |
| 2267 | if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias()) |
| 2268 | Sets.push_back(Elt: {&AS, false}); |
| 2269 | |
| 2270 | if (Sets.empty()) |
| 2271 | return {}; // Nothing to promote... |
| 2272 | |
| 2273 | // Discard any sets for which there is an aliasing non-promotable access. |
| 2274 | foreachMemoryAccess(MSSA, L, Fn: [&](Instruction *I) { |
| 2275 | if (AttemptingPromotion.contains(Ptr: I)) |
| 2276 | return; |
| 2277 | |
| 2278 | llvm::erase_if(C&: Sets, P: [&](PointerIntPair<const AliasSet *, 1, bool> &Pair) { |
| 2279 | ModRefInfo MR = Pair.getPointer()->aliasesUnknownInst(Inst: I, AA&: BatchAA); |
| 2280 | // Cannot promote if there are writes outside the set. |
| 2281 | if (isModSet(MRI: MR)) |
| 2282 | return true; |
| 2283 | if (isRefSet(MRI: MR)) { |
| 2284 | // Remember reads outside the set. |
| 2285 | Pair.setInt(true); |
| 2286 | // If this is a mod-only set and there are reads outside the set, |
| 2287 | // we will not be able to promote, so bail out early. |
| 2288 | return !Pair.getPointer()->isRef(); |
| 2289 | } |
| 2290 | return false; |
| 2291 | }); |
| 2292 | }); |
| 2293 | |
| 2294 | SmallVector<std::pair<SmallSetVector<Value *, 8>, bool>, 0> Result; |
| 2295 | for (auto [Set, HasReadsOutsideSet] : Sets) { |
| 2296 | SmallSetVector<Value *, 8> PointerMustAliases; |
| 2297 | for (const auto &MemLoc : *Set) |
| 2298 | PointerMustAliases.insert(X: const_cast<Value *>(MemLoc.Ptr)); |
| 2299 | Result.emplace_back(Args: std::move(PointerMustAliases), Args&: HasReadsOutsideSet); |
| 2300 | } |
| 2301 | |
| 2302 | return Result; |
| 2303 | } |
| 2304 | |
| 2305 | // For a given store instruction or writeonly call instruction, this function |
| 2306 | // checks that there are no read or writes that conflict with the memory |
| 2307 | // access in the instruction |
| 2308 | static bool noConflictingReadWrites(Instruction *I, MemorySSA *MSSA, |
| 2309 | AAResults *AA, Loop *CurLoop, |
| 2310 | SinkAndHoistLICMFlags &Flags) { |
| 2311 | assert(isa<CallInst>(*I) || isa<StoreInst>(*I)); |
| 2312 | // If there are more accesses than the Promotion cap, then give up as we're |
| 2313 | // not walking a list that long. |
| 2314 | if (Flags.tooManyMemoryAccesses()) |
| 2315 | return false; |
| 2316 | |
| 2317 | auto *IMD = MSSA->getMemoryAccess(I); |
| 2318 | BatchAAResults BAA(*AA); |
| 2319 | auto *Source = getClobberingMemoryAccess(MSSA&: *MSSA, BAA, Flags, MA: IMD); |
| 2320 | // Make sure there are no clobbers inside the loop. |
| 2321 | if (!MSSA->isLiveOnEntryDef(MA: Source) && CurLoop->contains(BB: Source->getBlock())) |
| 2322 | return false; |
| 2323 | |
| 2324 | // If there are interfering Uses (i.e. their defining access is in the |
| 2325 | // loop), or ordered loads (stored as Defs!), don't move this store. |
| 2326 | // Could do better here, but this is conservatively correct. |
| 2327 | // TODO: Cache set of Uses on the first walk in runOnLoop, update when |
| 2328 | // moving accesses. Can also extend to dominating uses. |
| 2329 | for (auto *BB : CurLoop->getBlocks()) { |
| 2330 | auto *Accesses = MSSA->getBlockAccesses(BB); |
| 2331 | if (!Accesses) |
| 2332 | continue; |
| 2333 | for (const auto &MA : *Accesses) |
| 2334 | if (const auto *MU = dyn_cast<MemoryUse>(Val: &MA)) { |
| 2335 | auto *MD = getClobberingMemoryAccess(MSSA&: *MSSA, BAA, Flags, |
| 2336 | MA: const_cast<MemoryUse *>(MU)); |
| 2337 | if (!MSSA->isLiveOnEntryDef(MA: MD) && CurLoop->contains(BB: MD->getBlock())) |
| 2338 | return false; |
| 2339 | // Disable hoisting past potentially interfering loads. Optimized |
| 2340 | // Uses may point to an access outside the loop, as getClobbering |
| 2341 | // checks the previous iteration when walking the backedge. |
| 2342 | // FIXME: More precise: no Uses that alias I. |
| 2343 | if (!Flags.getIsSink() && !MSSA->dominates(A: IMD, B: MU)) |
| 2344 | return false; |
| 2345 | } else if (const auto *MD = dyn_cast<MemoryDef>(Val: &MA)) { |
| 2346 | if (auto *LI = dyn_cast<LoadInst>(Val: MD->getMemoryInst())) { |
| 2347 | (void)LI; // Silence warning. |
| 2348 | assert(!LI->isUnordered() && "Expected unordered load" ); |
| 2349 | return false; |
| 2350 | } |
| 2351 | // Any call, while it may not be clobbering I, it may be a use. |
| 2352 | if (auto *CI = dyn_cast<CallInst>(Val: MD->getMemoryInst())) { |
| 2353 | // Check if the call may read from the memory location written |
| 2354 | // to by I. Check CI's attributes and arguments; the number of |
| 2355 | // such checks performed is limited above by NoOfMemAccTooLarge. |
| 2356 | if (auto *SI = dyn_cast<StoreInst>(Val: I)) { |
| 2357 | ModRefInfo MRI = BAA.getModRefInfo(I: CI, OptLoc: MemoryLocation::get(SI)); |
| 2358 | if (isModOrRefSet(MRI)) |
| 2359 | return false; |
| 2360 | } else { |
| 2361 | auto *SCI = cast<CallInst>(Val: I); |
| 2362 | // If the instruction we are wanting to hoist is also a call |
| 2363 | // instruction then we need not check mod/ref info with itself |
| 2364 | if (SCI == CI) |
| 2365 | continue; |
| 2366 | ModRefInfo MRI = BAA.getModRefInfo(I: CI, Call2: SCI); |
| 2367 | if (isModOrRefSet(MRI)) |
| 2368 | return false; |
| 2369 | } |
| 2370 | } |
| 2371 | } |
| 2372 | } |
| 2373 | return true; |
| 2374 | } |
| 2375 | |
| 2376 | static bool pointerInvalidatedByLoop(MemorySSA *MSSA, MemoryUse *MU, |
| 2377 | Loop *CurLoop, Instruction &I, |
| 2378 | SinkAndHoistLICMFlags &Flags, |
| 2379 | bool InvariantGroup) { |
| 2380 | // For hoisting, use the walker to determine safety |
| 2381 | if (!Flags.getIsSink()) { |
| 2382 | // If hoisting an invariant group, we only need to check that there |
| 2383 | // is no store to the loaded pointer between the start of the loop, |
| 2384 | // and the load (since all values must be the same). |
| 2385 | |
| 2386 | // This can be checked in two conditions: |
| 2387 | // 1) if the memoryaccess is outside the loop |
| 2388 | // 2) the earliest access is at the loop header, |
| 2389 | // if the memory loaded is the phi node |
| 2390 | |
| 2391 | BatchAAResults BAA(MSSA->getAA()); |
| 2392 | MemoryAccess *Source = getClobberingMemoryAccess(MSSA&: *MSSA, BAA, Flags, MA: MU); |
| 2393 | return !MSSA->isLiveOnEntryDef(MA: Source) && |
| 2394 | CurLoop->contains(BB: Source->getBlock()) && |
| 2395 | !(InvariantGroup && Source->getBlock() == CurLoop->getHeader() && isa<MemoryPhi>(Val: Source)); |
| 2396 | } |
| 2397 | |
| 2398 | // For sinking, we'd need to check all Defs below this use. The getClobbering |
| 2399 | // call will look on the backedge of the loop, but will check aliasing with |
| 2400 | // the instructions on the previous iteration. |
| 2401 | // For example: |
| 2402 | // for (i ... ) |
| 2403 | // load a[i] ( Use (LoE) |
| 2404 | // store a[i] ( 1 = Def (2), with 2 = Phi for the loop. |
| 2405 | // i++; |
| 2406 | // The load sees no clobbering inside the loop, as the backedge alias check |
| 2407 | // does phi translation, and will check aliasing against store a[i-1]. |
| 2408 | // However sinking the load outside the loop, below the store is incorrect. |
| 2409 | |
| 2410 | // For now, only sink if there are no Defs in the loop, and the existing ones |
| 2411 | // precede the use and are in the same block. |
| 2412 | // FIXME: Increase precision: Safe to sink if Use post dominates the Def; |
| 2413 | // needs PostDominatorTreeAnalysis. |
| 2414 | // FIXME: More precise: no Defs that alias this Use. |
| 2415 | if (Flags.tooManyMemoryAccesses()) |
| 2416 | return true; |
| 2417 | for (auto *BB : CurLoop->getBlocks()) |
| 2418 | if (pointerInvalidatedByBlock(BB&: *BB, MSSA&: *MSSA, MU&: *MU)) |
| 2419 | return true; |
| 2420 | // When sinking, the source block may not be part of the loop so check it. |
| 2421 | if (!CurLoop->contains(Inst: &I)) |
| 2422 | return pointerInvalidatedByBlock(BB&: *I.getParent(), MSSA&: *MSSA, MU&: *MU); |
| 2423 | |
| 2424 | return false; |
| 2425 | } |
| 2426 | |
| 2427 | bool pointerInvalidatedByBlock(BasicBlock &BB, MemorySSA &MSSA, MemoryUse &MU) { |
| 2428 | if (const auto *Accesses = MSSA.getBlockDefs(BB: &BB)) |
| 2429 | for (const auto &MA : *Accesses) |
| 2430 | if (const auto *MD = dyn_cast<MemoryDef>(Val: &MA)) |
| 2431 | if (MU.getBlock() != MD->getBlock() || !MSSA.locallyDominates(A: MD, B: &MU)) |
| 2432 | return true; |
| 2433 | return false; |
| 2434 | } |
| 2435 | |
| 2436 | /// Try to simplify things like (A < INV_1 AND icmp A < INV_2) into (A < |
| 2437 | /// min(INV_1, INV_2)), if INV_1 and INV_2 are both loop invariants and their |
| 2438 | /// minimun can be computed outside of loop, and X is not a loop-invariant. |
| 2439 | static bool hoistMinMax(Instruction &I, Loop &L, ICFLoopSafetyInfo &SafetyInfo, |
| 2440 | MemorySSAUpdater &MSSAU) { |
| 2441 | bool Inverse = false; |
| 2442 | using namespace PatternMatch; |
| 2443 | Value *Cond1, *Cond2; |
| 2444 | if (match(V: &I, P: m_LogicalOr(L: m_Value(V&: Cond1), R: m_Value(V&: Cond2)))) { |
| 2445 | Inverse = true; |
| 2446 | } else if (match(V: &I, P: m_LogicalAnd(L: m_Value(V&: Cond1), R: m_Value(V&: Cond2)))) { |
| 2447 | // Do nothing |
| 2448 | } else |
| 2449 | return false; |
| 2450 | |
| 2451 | auto MatchICmpAgainstInvariant = [&](Value *C, CmpPredicate &P, Value *&LHS, |
| 2452 | Value *&RHS) { |
| 2453 | if (!match(V: C, P: m_OneUse(SubPattern: m_ICmp(Pred&: P, L: m_Value(V&: LHS), R: m_Value(V&: RHS))))) |
| 2454 | return false; |
| 2455 | if (!LHS->getType()->isIntegerTy()) |
| 2456 | return false; |
| 2457 | if (!ICmpInst::isRelational(P)) |
| 2458 | return false; |
| 2459 | if (L.isLoopInvariant(V: LHS)) { |
| 2460 | std::swap(a&: LHS, b&: RHS); |
| 2461 | P = ICmpInst::getSwappedPredicate(pred: P); |
| 2462 | } |
| 2463 | if (L.isLoopInvariant(V: LHS) || !L.isLoopInvariant(V: RHS)) |
| 2464 | return false; |
| 2465 | if (Inverse) |
| 2466 | P = ICmpInst::getInversePredicate(pred: P); |
| 2467 | return true; |
| 2468 | }; |
| 2469 | CmpPredicate P1, P2; |
| 2470 | Value *LHS1, *LHS2, *RHS1, *RHS2; |
| 2471 | if (!MatchICmpAgainstInvariant(Cond1, P1, LHS1, RHS1) || |
| 2472 | !MatchICmpAgainstInvariant(Cond2, P2, LHS2, RHS2)) |
| 2473 | return false; |
| 2474 | auto MatchingPred = CmpPredicate::getMatching(A: P1, B: P2); |
| 2475 | if (!MatchingPred || LHS1 != LHS2) |
| 2476 | return false; |
| 2477 | |
| 2478 | // Everything is fine, we can do the transform. |
| 2479 | bool UseMin = ICmpInst::isLT(P: *MatchingPred) || ICmpInst::isLE(P: *MatchingPred); |
| 2480 | assert( |
| 2481 | (UseMin || ICmpInst::isGT(*MatchingPred) || |
| 2482 | ICmpInst::isGE(*MatchingPred)) && |
| 2483 | "Relational predicate is either less (or equal) or greater (or equal)!" ); |
| 2484 | Intrinsic::ID id = ICmpInst::isSigned(predicate: *MatchingPred) |
| 2485 | ? (UseMin ? Intrinsic::smin : Intrinsic::smax) |
| 2486 | : (UseMin ? Intrinsic::umin : Intrinsic::umax); |
| 2487 | auto * = L.getLoopPreheader(); |
| 2488 | assert(Preheader && "Loop is not in simplify form?" ); |
| 2489 | IRBuilder<> Builder(Preheader->getTerminator()); |
| 2490 | // We are about to create a new guaranteed use for RHS2 which might not exist |
| 2491 | // before (if it was a non-taken input of logical and/or instruction). If it |
| 2492 | // was poison, we need to freeze it. Note that no new use for LHS and RHS1 are |
| 2493 | // introduced, so they don't need this. |
| 2494 | if (isa<SelectInst>(Val: I)) |
| 2495 | RHS2 = Builder.CreateFreeze(V: RHS2, Name: RHS2->getName() + ".fr" ); |
| 2496 | Value *NewRHS = Builder.CreateBinaryIntrinsic( |
| 2497 | ID: id, LHS: RHS1, RHS: RHS2, FMFSource: nullptr, |
| 2498 | Name: StringRef("invariant." ) + |
| 2499 | (ICmpInst::isSigned(predicate: *MatchingPred) ? "s" : "u" ) + |
| 2500 | (UseMin ? "min" : "max" )); |
| 2501 | Builder.SetInsertPoint(&I); |
| 2502 | ICmpInst::Predicate P = *MatchingPred; |
| 2503 | if (Inverse) |
| 2504 | P = ICmpInst::getInversePredicate(pred: P); |
| 2505 | Value *NewCond = Builder.CreateICmp(P, LHS: LHS1, RHS: NewRHS); |
| 2506 | NewCond->takeName(V: &I); |
| 2507 | I.replaceAllUsesWith(V: NewCond); |
| 2508 | eraseInstruction(I, SafetyInfo, MSSAU); |
| 2509 | Instruction &CondI1 = *cast<Instruction>(Val: Cond1); |
| 2510 | Instruction &CondI2 = *cast<Instruction>(Val: Cond2); |
| 2511 | salvageDebugInfo(I&: CondI1); |
| 2512 | salvageDebugInfo(I&: CondI2); |
| 2513 | eraseInstruction(I&: CondI1, SafetyInfo, MSSAU); |
| 2514 | eraseInstruction(I&: CondI2, SafetyInfo, MSSAU); |
| 2515 | return true; |
| 2516 | } |
| 2517 | |
| 2518 | /// Reassociate gep (gep ptr, idx1), idx2 to gep (gep ptr, idx2), idx1 if |
| 2519 | /// this allows hoisting the inner GEP. |
| 2520 | static bool hoistGEP(Instruction &I, Loop &L, ICFLoopSafetyInfo &SafetyInfo, |
| 2521 | MemorySSAUpdater &MSSAU, AssumptionCache *AC, |
| 2522 | DominatorTree *DT) { |
| 2523 | auto *GEP = dyn_cast<GetElementPtrInst>(Val: &I); |
| 2524 | if (!GEP) |
| 2525 | return false; |
| 2526 | |
| 2527 | // Do not try to hoist a constant GEP out of the loop via reassociation. |
| 2528 | // Constant GEPs can often be folded into addressing modes, and reassociating |
| 2529 | // them may inhibit CSE of a common base. |
| 2530 | if (GEP->hasAllConstantIndices()) |
| 2531 | return false; |
| 2532 | |
| 2533 | auto *Src = dyn_cast<GetElementPtrInst>(Val: GEP->getPointerOperand()); |
| 2534 | if (!Src || !Src->hasOneUse() || !L.contains(Inst: Src)) |
| 2535 | return false; |
| 2536 | |
| 2537 | Value *SrcPtr = Src->getPointerOperand(); |
| 2538 | auto LoopInvariant = [&](Value *V) { return L.isLoopInvariant(V); }; |
| 2539 | if (!L.isLoopInvariant(V: SrcPtr) || !all_of(Range: GEP->indices(), P: LoopInvariant)) |
| 2540 | return false; |
| 2541 | |
| 2542 | // This can only happen if !AllowSpeculation, otherwise this would already be |
| 2543 | // handled. |
| 2544 | // FIXME: Should we respect AllowSpeculation in these reassociation folds? |
| 2545 | // The flag exists to prevent metadata dropping, which is not relevant here. |
| 2546 | if (all_of(Range: Src->indices(), P: LoopInvariant)) |
| 2547 | return false; |
| 2548 | |
| 2549 | // The swapped GEPs are inbounds if both original GEPs are inbounds |
| 2550 | // and the sign of the offsets is the same. For simplicity, only |
| 2551 | // handle both offsets being non-negative. |
| 2552 | const DataLayout &DL = GEP->getDataLayout(); |
| 2553 | auto NonNegative = [&](Value *V) { |
| 2554 | return isKnownNonNegative(V, SQ: SimplifyQuery(DL, DT, AC, GEP)); |
| 2555 | }; |
| 2556 | bool IsInBounds = Src->isInBounds() && GEP->isInBounds() && |
| 2557 | all_of(Range: Src->indices(), P: NonNegative) && |
| 2558 | all_of(Range: GEP->indices(), P: NonNegative); |
| 2559 | |
| 2560 | BasicBlock * = L.getLoopPreheader(); |
| 2561 | IRBuilder<> Builder(Preheader->getTerminator()); |
| 2562 | Value *NewSrc = Builder.CreateGEP(Ty: GEP->getSourceElementType(), Ptr: SrcPtr, |
| 2563 | IdxList: SmallVector<Value *>(GEP->indices()), |
| 2564 | Name: "invariant.gep" , NW: IsInBounds); |
| 2565 | Builder.SetInsertPoint(GEP); |
| 2566 | Value *NewGEP = Builder.CreateGEP(Ty: Src->getSourceElementType(), Ptr: NewSrc, |
| 2567 | IdxList: SmallVector<Value *>(Src->indices()), Name: "gep" , |
| 2568 | NW: IsInBounds); |
| 2569 | GEP->replaceAllUsesWith(V: NewGEP); |
| 2570 | eraseInstruction(I&: *GEP, SafetyInfo, MSSAU); |
| 2571 | salvageDebugInfo(I&: *Src); |
| 2572 | eraseInstruction(I&: *Src, SafetyInfo, MSSAU); |
| 2573 | return true; |
| 2574 | } |
| 2575 | |
| 2576 | /// Try to turn things like "LV + C1 < C2" into "LV < C2 - C1". Here |
| 2577 | /// C1 and C2 are loop invariants and LV is a loop-variant. |
| 2578 | static bool hoistAdd(ICmpInst::Predicate Pred, Value *VariantLHS, |
| 2579 | Value *InvariantRHS, ICmpInst &ICmp, Loop &L, |
| 2580 | ICFLoopSafetyInfo &SafetyInfo, MemorySSAUpdater &MSSAU, |
| 2581 | AssumptionCache *AC, DominatorTree *DT) { |
| 2582 | assert(!L.isLoopInvariant(VariantLHS) && "Precondition." ); |
| 2583 | assert(L.isLoopInvariant(InvariantRHS) && "Precondition." ); |
| 2584 | |
| 2585 | bool IsSigned = ICmpInst::isSigned(predicate: Pred); |
| 2586 | |
| 2587 | // Try to represent VariantLHS as sum of invariant and variant operands. |
| 2588 | using namespace PatternMatch; |
| 2589 | Value *VariantOp, *InvariantOp; |
| 2590 | if (IsSigned && |
| 2591 | !match(V: VariantLHS, P: m_NSWAdd(L: m_Value(V&: VariantOp), R: m_Value(V&: InvariantOp)))) |
| 2592 | return false; |
| 2593 | if (!IsSigned && |
| 2594 | !match(V: VariantLHS, P: m_NUWAdd(L: m_Value(V&: VariantOp), R: m_Value(V&: InvariantOp)))) |
| 2595 | return false; |
| 2596 | |
| 2597 | // LHS itself is a loop-variant, try to represent it in the form: |
| 2598 | // "VariantOp + InvariantOp". If it is possible, then we can reassociate. |
| 2599 | if (L.isLoopInvariant(V: VariantOp)) |
| 2600 | std::swap(a&: VariantOp, b&: InvariantOp); |
| 2601 | if (L.isLoopInvariant(V: VariantOp) || !L.isLoopInvariant(V: InvariantOp)) |
| 2602 | return false; |
| 2603 | |
| 2604 | // In order to turn "LV + C1 < C2" into "LV < C2 - C1", we need to be able to |
| 2605 | // freely move values from left side of inequality to right side (just as in |
| 2606 | // normal linear arithmetics). Overflows make things much more complicated, so |
| 2607 | // we want to avoid this. |
| 2608 | auto &DL = L.getHeader()->getDataLayout(); |
| 2609 | SimplifyQuery SQ(DL, DT, AC, &ICmp); |
| 2610 | if (IsSigned && computeOverflowForSignedSub(LHS: InvariantRHS, RHS: InvariantOp, SQ) != |
| 2611 | llvm::OverflowResult::NeverOverflows) |
| 2612 | return false; |
| 2613 | if (!IsSigned && |
| 2614 | computeOverflowForUnsignedSub(LHS: InvariantRHS, RHS: InvariantOp, SQ) != |
| 2615 | llvm::OverflowResult::NeverOverflows) |
| 2616 | return false; |
| 2617 | auto * = L.getLoopPreheader(); |
| 2618 | assert(Preheader && "Loop is not in simplify form?" ); |
| 2619 | IRBuilder<> Builder(Preheader->getTerminator()); |
| 2620 | Value *NewCmpOp = |
| 2621 | Builder.CreateSub(LHS: InvariantRHS, RHS: InvariantOp, Name: "invariant.op" , |
| 2622 | /*HasNUW*/ !IsSigned, /*HasNSW*/ IsSigned); |
| 2623 | ICmp.setPredicate(Pred); |
| 2624 | ICmp.setOperand(i_nocapture: 0, Val_nocapture: VariantOp); |
| 2625 | ICmp.setOperand(i_nocapture: 1, Val_nocapture: NewCmpOp); |
| 2626 | |
| 2627 | Instruction &DeadI = cast<Instruction>(Val&: *VariantLHS); |
| 2628 | salvageDebugInfo(I&: DeadI); |
| 2629 | eraseInstruction(I&: DeadI, SafetyInfo, MSSAU); |
| 2630 | return true; |
| 2631 | } |
| 2632 | |
| 2633 | /// Try to reassociate and hoist the following two patterns: |
| 2634 | /// LV - C1 < C2 --> LV < C1 + C2, |
| 2635 | /// C1 - LV < C2 --> LV > C1 - C2. |
| 2636 | static bool hoistSub(ICmpInst::Predicate Pred, Value *VariantLHS, |
| 2637 | Value *InvariantRHS, ICmpInst &ICmp, Loop &L, |
| 2638 | ICFLoopSafetyInfo &SafetyInfo, MemorySSAUpdater &MSSAU, |
| 2639 | AssumptionCache *AC, DominatorTree *DT) { |
| 2640 | assert(!L.isLoopInvariant(VariantLHS) && "Precondition." ); |
| 2641 | assert(L.isLoopInvariant(InvariantRHS) && "Precondition." ); |
| 2642 | |
| 2643 | bool IsSigned = ICmpInst::isSigned(predicate: Pred); |
| 2644 | |
| 2645 | // Try to represent VariantLHS as sum of invariant and variant operands. |
| 2646 | using namespace PatternMatch; |
| 2647 | Value *VariantOp, *InvariantOp; |
| 2648 | if (IsSigned && |
| 2649 | !match(V: VariantLHS, P: m_NSWSub(L: m_Value(V&: VariantOp), R: m_Value(V&: InvariantOp)))) |
| 2650 | return false; |
| 2651 | if (!IsSigned && |
| 2652 | !match(V: VariantLHS, P: m_NUWSub(L: m_Value(V&: VariantOp), R: m_Value(V&: InvariantOp)))) |
| 2653 | return false; |
| 2654 | |
| 2655 | bool VariantSubtracted = false; |
| 2656 | // LHS itself is a loop-variant, try to represent it in the form: |
| 2657 | // "VariantOp + InvariantOp". If it is possible, then we can reassociate. If |
| 2658 | // the variant operand goes with minus, we use a slightly different scheme. |
| 2659 | if (L.isLoopInvariant(V: VariantOp)) { |
| 2660 | std::swap(a&: VariantOp, b&: InvariantOp); |
| 2661 | VariantSubtracted = true; |
| 2662 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
| 2663 | } |
| 2664 | if (L.isLoopInvariant(V: VariantOp) || !L.isLoopInvariant(V: InvariantOp)) |
| 2665 | return false; |
| 2666 | |
| 2667 | // In order to turn "LV - C1 < C2" into "LV < C2 + C1", we need to be able to |
| 2668 | // freely move values from left side of inequality to right side (just as in |
| 2669 | // normal linear arithmetics). Overflows make things much more complicated, so |
| 2670 | // we want to avoid this. Likewise, for "C1 - LV < C2" we need to prove that |
| 2671 | // "C1 - C2" does not overflow. |
| 2672 | auto &DL = L.getHeader()->getDataLayout(); |
| 2673 | SimplifyQuery SQ(DL, DT, AC, &ICmp); |
| 2674 | if (VariantSubtracted && IsSigned) { |
| 2675 | // C1 - LV < C2 --> LV > C1 - C2 |
| 2676 | if (computeOverflowForSignedSub(LHS: InvariantOp, RHS: InvariantRHS, SQ) != |
| 2677 | llvm::OverflowResult::NeverOverflows) |
| 2678 | return false; |
| 2679 | } else if (VariantSubtracted && !IsSigned) { |
| 2680 | // C1 - LV < C2 --> LV > C1 - C2 |
| 2681 | if (computeOverflowForUnsignedSub(LHS: InvariantOp, RHS: InvariantRHS, SQ) != |
| 2682 | llvm::OverflowResult::NeverOverflows) |
| 2683 | return false; |
| 2684 | } else if (!VariantSubtracted && IsSigned) { |
| 2685 | // LV - C1 < C2 --> LV < C1 + C2 |
| 2686 | if (computeOverflowForSignedAdd(LHS: InvariantOp, RHS: InvariantRHS, SQ) != |
| 2687 | llvm::OverflowResult::NeverOverflows) |
| 2688 | return false; |
| 2689 | } else { // !VariantSubtracted && !IsSigned |
| 2690 | // LV - C1 < C2 --> LV < C1 + C2 |
| 2691 | if (computeOverflowForUnsignedAdd(LHS: InvariantOp, RHS: InvariantRHS, SQ) != |
| 2692 | llvm::OverflowResult::NeverOverflows) |
| 2693 | return false; |
| 2694 | } |
| 2695 | auto * = L.getLoopPreheader(); |
| 2696 | assert(Preheader && "Loop is not in simplify form?" ); |
| 2697 | IRBuilder<> Builder(Preheader->getTerminator()); |
| 2698 | Value *NewCmpOp = |
| 2699 | VariantSubtracted |
| 2700 | ? Builder.CreateSub(LHS: InvariantOp, RHS: InvariantRHS, Name: "invariant.op" , |
| 2701 | /*HasNUW*/ !IsSigned, /*HasNSW*/ IsSigned) |
| 2702 | : Builder.CreateAdd(LHS: InvariantOp, RHS: InvariantRHS, Name: "invariant.op" , |
| 2703 | /*HasNUW*/ !IsSigned, /*HasNSW*/ IsSigned); |
| 2704 | ICmp.setPredicate(Pred); |
| 2705 | ICmp.setOperand(i_nocapture: 0, Val_nocapture: VariantOp); |
| 2706 | ICmp.setOperand(i_nocapture: 1, Val_nocapture: NewCmpOp); |
| 2707 | |
| 2708 | Instruction &DeadI = cast<Instruction>(Val&: *VariantLHS); |
| 2709 | salvageDebugInfo(I&: DeadI); |
| 2710 | eraseInstruction(I&: DeadI, SafetyInfo, MSSAU); |
| 2711 | return true; |
| 2712 | } |
| 2713 | |
| 2714 | /// Reassociate and hoist add/sub expressions. |
| 2715 | static bool hoistAddSub(Instruction &I, Loop &L, ICFLoopSafetyInfo &SafetyInfo, |
| 2716 | MemorySSAUpdater &MSSAU, AssumptionCache *AC, |
| 2717 | DominatorTree *DT) { |
| 2718 | using namespace PatternMatch; |
| 2719 | CmpPredicate Pred; |
| 2720 | Value *LHS, *RHS; |
| 2721 | if (!match(V: &I, P: m_ICmp(Pred, L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) |
| 2722 | return false; |
| 2723 | |
| 2724 | // Put variant operand to LHS position. |
| 2725 | if (L.isLoopInvariant(V: LHS)) { |
| 2726 | std::swap(a&: LHS, b&: RHS); |
| 2727 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
| 2728 | } |
| 2729 | // We want to delete the initial operation after reassociation, so only do it |
| 2730 | // if it has no other uses. |
| 2731 | if (L.isLoopInvariant(V: LHS) || !L.isLoopInvariant(V: RHS) || !LHS->hasOneUse()) |
| 2732 | return false; |
| 2733 | |
| 2734 | // TODO: We could go with smarter context, taking common dominator of all I's |
| 2735 | // users instead of I itself. |
| 2736 | if (hoistAdd(Pred, VariantLHS: LHS, InvariantRHS: RHS, ICmp&: cast<ICmpInst>(Val&: I), L, SafetyInfo, MSSAU, AC, DT)) |
| 2737 | return true; |
| 2738 | |
| 2739 | if (hoistSub(Pred, VariantLHS: LHS, InvariantRHS: RHS, ICmp&: cast<ICmpInst>(Val&: I), L, SafetyInfo, MSSAU, AC, DT)) |
| 2740 | return true; |
| 2741 | |
| 2742 | return false; |
| 2743 | } |
| 2744 | |
| 2745 | static bool isReassociableOp(Instruction *I, unsigned IntOpcode, |
| 2746 | unsigned FPOpcode) { |
| 2747 | if (I->getOpcode() == IntOpcode) |
| 2748 | return true; |
| 2749 | if (I->getOpcode() == FPOpcode && I->hasAllowReassoc() && |
| 2750 | I->hasNoSignedZeros()) |
| 2751 | return true; |
| 2752 | return false; |
| 2753 | } |
| 2754 | |
| 2755 | /// Try to reassociate expressions like ((A1 * B1) + (A2 * B2) + ...) * C where |
| 2756 | /// A1, A2, ... and C are loop invariants into expressions like |
| 2757 | /// ((A1 * C * B1) + (A2 * C * B2) + ...) and hoist the (A1 * C), (A2 * C), ... |
| 2758 | /// invariant expressions. This functions returns true only if any hoisting has |
| 2759 | /// actually occurred. |
| 2760 | static bool hoistMulAddAssociation(Instruction &I, Loop &L, |
| 2761 | ICFLoopSafetyInfo &SafetyInfo, |
| 2762 | MemorySSAUpdater &MSSAU, AssumptionCache *AC, |
| 2763 | DominatorTree *DT) { |
| 2764 | if (!isReassociableOp(I: &I, IntOpcode: Instruction::Mul, FPOpcode: Instruction::FMul)) |
| 2765 | return false; |
| 2766 | Value *VariantOp = I.getOperand(i: 0); |
| 2767 | Value *InvariantOp = I.getOperand(i: 1); |
| 2768 | if (L.isLoopInvariant(V: VariantOp)) |
| 2769 | std::swap(a&: VariantOp, b&: InvariantOp); |
| 2770 | if (L.isLoopInvariant(V: VariantOp) || !L.isLoopInvariant(V: InvariantOp)) |
| 2771 | return false; |
| 2772 | Value *Factor = InvariantOp; |
| 2773 | |
| 2774 | // First, we need to make sure we should do the transformation. |
| 2775 | SmallVector<Use *> Changes; |
| 2776 | SmallVector<BinaryOperator *> Adds; |
| 2777 | SmallVector<BinaryOperator *> Worklist; |
| 2778 | if (BinaryOperator *VariantBinOp = dyn_cast<BinaryOperator>(Val: VariantOp)) |
| 2779 | Worklist.push_back(Elt: VariantBinOp); |
| 2780 | while (!Worklist.empty()) { |
| 2781 | BinaryOperator *BO = Worklist.pop_back_val(); |
| 2782 | if (!BO->hasOneUse()) |
| 2783 | return false; |
| 2784 | if (isReassociableOp(I: BO, IntOpcode: Instruction::Add, FPOpcode: Instruction::FAdd) && |
| 2785 | isa<BinaryOperator>(Val: BO->getOperand(i_nocapture: 0)) && |
| 2786 | isa<BinaryOperator>(Val: BO->getOperand(i_nocapture: 1))) { |
| 2787 | Worklist.push_back(Elt: cast<BinaryOperator>(Val: BO->getOperand(i_nocapture: 0))); |
| 2788 | Worklist.push_back(Elt: cast<BinaryOperator>(Val: BO->getOperand(i_nocapture: 1))); |
| 2789 | Adds.push_back(Elt: BO); |
| 2790 | continue; |
| 2791 | } |
| 2792 | if (!isReassociableOp(I: BO, IntOpcode: Instruction::Mul, FPOpcode: Instruction::FMul) || |
| 2793 | L.isLoopInvariant(V: BO)) |
| 2794 | return false; |
| 2795 | Use &U0 = BO->getOperandUse(i: 0); |
| 2796 | Use &U1 = BO->getOperandUse(i: 1); |
| 2797 | if (L.isLoopInvariant(V: U0)) |
| 2798 | Changes.push_back(Elt: &U0); |
| 2799 | else if (L.isLoopInvariant(V: U1)) |
| 2800 | Changes.push_back(Elt: &U1); |
| 2801 | else |
| 2802 | return false; |
| 2803 | unsigned Limit = I.getType()->isIntOrIntVectorTy() |
| 2804 | ? IntAssociationUpperLimit |
| 2805 | : FPAssociationUpperLimit; |
| 2806 | if (Changes.size() > Limit) |
| 2807 | return false; |
| 2808 | } |
| 2809 | if (Changes.empty()) |
| 2810 | return false; |
| 2811 | |
| 2812 | // Drop the poison flags for any adds we looked through. |
| 2813 | if (I.getType()->isIntOrIntVectorTy()) { |
| 2814 | for (auto *Add : Adds) |
| 2815 | Add->dropPoisonGeneratingFlags(); |
| 2816 | } |
| 2817 | |
| 2818 | // We know we should do it so let's do the transformation. |
| 2819 | auto * = L.getLoopPreheader(); |
| 2820 | assert(Preheader && "Loop is not in simplify form?" ); |
| 2821 | IRBuilder<> Builder(Preheader->getTerminator()); |
| 2822 | for (auto *U : Changes) { |
| 2823 | assert(L.isLoopInvariant(U->get())); |
| 2824 | auto *Ins = cast<BinaryOperator>(Val: U->getUser()); |
| 2825 | Value *Mul; |
| 2826 | if (I.getType()->isIntOrIntVectorTy()) { |
| 2827 | Mul = Builder.CreateMul(LHS: U->get(), RHS: Factor, Name: "factor.op.mul" ); |
| 2828 | // Drop the poison flags on the original multiply. |
| 2829 | Ins->dropPoisonGeneratingFlags(); |
| 2830 | } else |
| 2831 | Mul = Builder.CreateFMulFMF(L: U->get(), R: Factor, FMFSource: Ins, Name: "factor.op.fmul" ); |
| 2832 | |
| 2833 | // Rewrite the reassociable instruction. |
| 2834 | unsigned OpIdx = U->getOperandNo(); |
| 2835 | auto *LHS = OpIdx == 0 ? Mul : Ins->getOperand(i_nocapture: 0); |
| 2836 | auto *RHS = OpIdx == 1 ? Mul : Ins->getOperand(i_nocapture: 1); |
| 2837 | auto *NewBO = |
| 2838 | BinaryOperator::Create(Op: Ins->getOpcode(), S1: LHS, S2: RHS, |
| 2839 | Name: Ins->getName() + ".reass" , InsertBefore: Ins->getIterator()); |
| 2840 | NewBO->setDebugLoc(DebugLoc::getDropped()); |
| 2841 | NewBO->copyIRFlags(V: Ins); |
| 2842 | if (VariantOp == Ins) |
| 2843 | VariantOp = NewBO; |
| 2844 | Ins->replaceAllUsesWith(V: NewBO); |
| 2845 | eraseInstruction(I&: *Ins, SafetyInfo, MSSAU); |
| 2846 | } |
| 2847 | |
| 2848 | I.replaceAllUsesWith(V: VariantOp); |
| 2849 | eraseInstruction(I, SafetyInfo, MSSAU); |
| 2850 | return true; |
| 2851 | } |
| 2852 | |
| 2853 | /// Reassociate associative binary expressions of the form |
| 2854 | /// |
| 2855 | /// 1. "(LV op C1) op C2" ==> "LV op (C1 op C2)" |
| 2856 | /// 2. "(C1 op LV) op C2" ==> "LV op (C1 op C2)" |
| 2857 | /// 3. "C2 op (C1 op LV)" ==> "LV op (C1 op C2)" |
| 2858 | /// 4. "C2 op (LV op C1)" ==> "LV op (C1 op C2)" |
| 2859 | /// |
| 2860 | /// where op is an associative BinOp, LV is a loop variant, and C1 and C2 are |
| 2861 | /// loop invariants that we want to hoist, noting that associativity implies |
| 2862 | /// commutativity. |
| 2863 | static bool hoistBOAssociation(Instruction &I, Loop &L, |
| 2864 | ICFLoopSafetyInfo &SafetyInfo, |
| 2865 | MemorySSAUpdater &MSSAU, AssumptionCache *AC, |
| 2866 | DominatorTree *DT) { |
| 2867 | auto *BO = dyn_cast<BinaryOperator>(Val: &I); |
| 2868 | if (!BO || !BO->isAssociative()) |
| 2869 | return false; |
| 2870 | |
| 2871 | Instruction::BinaryOps Opcode = BO->getOpcode(); |
| 2872 | bool LVInRHS = L.isLoopInvariant(V: BO->getOperand(i_nocapture: 0)); |
| 2873 | auto *BO0 = dyn_cast<BinaryOperator>(Val: BO->getOperand(i_nocapture: LVInRHS)); |
| 2874 | if (!BO0 || BO0->getOpcode() != Opcode || !BO0->isAssociative() || |
| 2875 | BO0->hasNUsesOrMore(N: BO0->getType()->isIntegerTy() ? 2 : 3)) |
| 2876 | return false; |
| 2877 | |
| 2878 | Value *LV = BO0->getOperand(i_nocapture: 0); |
| 2879 | Value *C1 = BO0->getOperand(i_nocapture: 1); |
| 2880 | Value *C2 = BO->getOperand(i_nocapture: !LVInRHS); |
| 2881 | |
| 2882 | assert(BO->isCommutative() && BO0->isCommutative() && |
| 2883 | "Associativity implies commutativity" ); |
| 2884 | if (L.isLoopInvariant(V: LV) && !L.isLoopInvariant(V: C1)) |
| 2885 | std::swap(a&: LV, b&: C1); |
| 2886 | if (L.isLoopInvariant(V: LV) || !L.isLoopInvariant(V: C1) || !L.isLoopInvariant(V: C2)) |
| 2887 | return false; |
| 2888 | |
| 2889 | auto * = L.getLoopPreheader(); |
| 2890 | assert(Preheader && "Loop is not in simplify form?" ); |
| 2891 | |
| 2892 | IRBuilder<> Builder(Preheader->getTerminator()); |
| 2893 | auto *Inv = Builder.CreateBinOp(Opc: Opcode, LHS: C1, RHS: C2, Name: "invariant.op" ); |
| 2894 | |
| 2895 | auto *NewBO = BinaryOperator::Create( |
| 2896 | Op: Opcode, S1: LV, S2: Inv, Name: BO->getName() + ".reass" , InsertBefore: BO->getIterator()); |
| 2897 | NewBO->setDebugLoc(DebugLoc::getDropped()); |
| 2898 | |
| 2899 | if (Opcode == Instruction::FAdd || Opcode == Instruction::FMul) { |
| 2900 | // Intersect FMF flags for FADD and FMUL. |
| 2901 | FastMathFlags Intersect = BO->getFastMathFlags() & BO0->getFastMathFlags(); |
| 2902 | if (auto *I = dyn_cast<Instruction>(Val: Inv)) |
| 2903 | I->setFastMathFlags(Intersect); |
| 2904 | NewBO->setFastMathFlags(Intersect); |
| 2905 | } else { |
| 2906 | OverflowTracking Flags; |
| 2907 | Flags.AllKnownNonNegative = false; |
| 2908 | Flags.AllKnownNonZero = false; |
| 2909 | Flags.mergeFlags(I&: *BO); |
| 2910 | Flags.mergeFlags(I&: *BO0); |
| 2911 | // If `Inv` was not constant-folded, a new Instruction has been created. |
| 2912 | if (auto *I = dyn_cast<Instruction>(Val: Inv)) |
| 2913 | Flags.applyFlags(I&: *I); |
| 2914 | Flags.applyFlags(I&: *NewBO); |
| 2915 | } |
| 2916 | |
| 2917 | BO->replaceAllUsesWith(V: NewBO); |
| 2918 | eraseInstruction(I&: *BO, SafetyInfo, MSSAU); |
| 2919 | |
| 2920 | // (LV op C1) might not be erased if it has more uses than the one we just |
| 2921 | // replaced. |
| 2922 | if (BO0->use_empty()) { |
| 2923 | salvageDebugInfo(I&: *BO0); |
| 2924 | eraseInstruction(I&: *BO0, SafetyInfo, MSSAU); |
| 2925 | } |
| 2926 | |
| 2927 | return true; |
| 2928 | } |
| 2929 | |
| 2930 | static bool hoistArithmetics(Instruction &I, Loop &L, |
| 2931 | ICFLoopSafetyInfo &SafetyInfo, |
| 2932 | MemorySSAUpdater &MSSAU, AssumptionCache *AC, |
| 2933 | DominatorTree *DT) { |
| 2934 | // Optimize complex patterns, such as (x < INV1 && x < INV2), turning them |
| 2935 | // into (x < min(INV1, INV2)), and hoisting the invariant part of this |
| 2936 | // expression out of the loop. |
| 2937 | if (hoistMinMax(I, L, SafetyInfo, MSSAU)) { |
| 2938 | ++NumHoisted; |
| 2939 | ++NumMinMaxHoisted; |
| 2940 | return true; |
| 2941 | } |
| 2942 | |
| 2943 | // Try to hoist GEPs by reassociation. |
| 2944 | if (hoistGEP(I, L, SafetyInfo, MSSAU, AC, DT)) { |
| 2945 | ++NumHoisted; |
| 2946 | ++NumGEPsHoisted; |
| 2947 | return true; |
| 2948 | } |
| 2949 | |
| 2950 | // Try to hoist add/sub's by reassociation. |
| 2951 | if (hoistAddSub(I, L, SafetyInfo, MSSAU, AC, DT)) { |
| 2952 | ++NumHoisted; |
| 2953 | ++NumAddSubHoisted; |
| 2954 | return true; |
| 2955 | } |
| 2956 | |
| 2957 | bool IsInt = I.getType()->isIntOrIntVectorTy(); |
| 2958 | if (hoistMulAddAssociation(I, L, SafetyInfo, MSSAU, AC, DT)) { |
| 2959 | ++NumHoisted; |
| 2960 | if (IsInt) |
| 2961 | ++NumIntAssociationsHoisted; |
| 2962 | else |
| 2963 | ++NumFPAssociationsHoisted; |
| 2964 | return true; |
| 2965 | } |
| 2966 | |
| 2967 | if (hoistBOAssociation(I, L, SafetyInfo, MSSAU, AC, DT)) { |
| 2968 | ++NumHoisted; |
| 2969 | ++NumBOAssociationsHoisted; |
| 2970 | return true; |
| 2971 | } |
| 2972 | |
| 2973 | return false; |
| 2974 | } |
| 2975 | |
| 2976 | /// Little predicate that returns true if the specified basic block is in |
| 2977 | /// a subloop of the current one, not the current one itself. |
| 2978 | /// |
| 2979 | static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) { |
| 2980 | assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop" ); |
| 2981 | return LI->getLoopFor(BB) != CurLoop; |
| 2982 | } |
| 2983 | |