| 1 | //===- DeadStoreElimination.cpp - MemorySSA Backed Dead Store Elimination -===// |
|---|---|
| 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 | // The code below implements dead store elimination using MemorySSA. It uses |
| 10 | // the following general approach: given a MemoryDef, walk upwards to find |
| 11 | // clobbering MemoryDefs that may be killed by the starting def. Then check |
| 12 | // that there are no uses that may read the location of the original MemoryDef |
| 13 | // in between both MemoryDefs. A bit more concretely: |
| 14 | // |
| 15 | // For all MemoryDefs StartDef: |
| 16 | // 1. Get the next dominating clobbering MemoryDef (MaybeDeadAccess) by walking |
| 17 | // upwards. |
| 18 | // 2. Check that there are no reads between MaybeDeadAccess and the StartDef by |
| 19 | // checking all uses starting at MaybeDeadAccess and walking until we see |
| 20 | // StartDef. |
| 21 | // 3. For each found CurrentDef, check that: |
| 22 | // 1. There are no barrier instructions between CurrentDef and StartDef (like |
| 23 | // throws or stores with ordering constraints). |
| 24 | // 2. StartDef is executed whenever CurrentDef is executed. |
| 25 | // 3. StartDef completely overwrites CurrentDef. |
| 26 | // 4. Erase CurrentDef from the function and MemorySSA. |
| 27 | // |
| 28 | //===----------------------------------------------------------------------===// |
| 29 | |
| 30 | #include "llvm/Transforms/Scalar/DeadStoreElimination.h" |
| 31 | #include "ScalarOptions.h" |
| 32 | #include "llvm/ADT/APInt.h" |
| 33 | #include "llvm/ADT/DenseMap.h" |
| 34 | #include "llvm/ADT/MapVector.h" |
| 35 | #include "llvm/ADT/PostOrderIterator.h" |
| 36 | #include "llvm/ADT/STLExtras.h" |
| 37 | #include "llvm/ADT/ScopedHashTable.h" |
| 38 | #include "llvm/ADT/SetVector.h" |
| 39 | #include "llvm/ADT/SmallPtrSet.h" |
| 40 | #include "llvm/ADT/SmallVector.h" |
| 41 | #include "llvm/ADT/Statistic.h" |
| 42 | #include "llvm/ADT/StringRef.h" |
| 43 | #include "llvm/Analysis/AliasAnalysis.h" |
| 44 | #include "llvm/Analysis/AssumptionCache.h" |
| 45 | #include "llvm/Analysis/CaptureTracking.h" |
| 46 | #include "llvm/Analysis/CycleAnalysis.h" |
| 47 | #include "llvm/Analysis/GlobalsModRef.h" |
| 48 | #include "llvm/Analysis/Loads.h" |
| 49 | #include "llvm/Analysis/MemoryBuiltins.h" |
| 50 | #include "llvm/Analysis/MemoryLocation.h" |
| 51 | #include "llvm/Analysis/MemorySSA.h" |
| 52 | #include "llvm/Analysis/MemorySSAUpdater.h" |
| 53 | #include "llvm/Analysis/MustExecute.h" |
| 54 | #include "llvm/Analysis/PostDominators.h" |
| 55 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 56 | #include "llvm/Analysis/ValueTracking.h" |
| 57 | #include "llvm/IR/Argument.h" |
| 58 | #include "llvm/IR/AttributeMask.h" |
| 59 | #include "llvm/IR/BasicBlock.h" |
| 60 | #include "llvm/IR/Constant.h" |
| 61 | #include "llvm/IR/ConstantRangeList.h" |
| 62 | #include "llvm/IR/Constants.h" |
| 63 | #include "llvm/IR/DataLayout.h" |
| 64 | #include "llvm/IR/DebugInfo.h" |
| 65 | #include "llvm/IR/Dominators.h" |
| 66 | #include "llvm/IR/Function.h" |
| 67 | #include "llvm/IR/IRBuilder.h" |
| 68 | #include "llvm/IR/InstIterator.h" |
| 69 | #include "llvm/IR/InstrTypes.h" |
| 70 | #include "llvm/IR/Instruction.h" |
| 71 | #include "llvm/IR/Instructions.h" |
| 72 | #include "llvm/IR/IntrinsicInst.h" |
| 73 | #include "llvm/IR/Module.h" |
| 74 | #include "llvm/IR/PassManager.h" |
| 75 | #include "llvm/IR/PatternMatch.h" |
| 76 | #include "llvm/IR/Value.h" |
| 77 | #include "llvm/InitializePasses.h" |
| 78 | #include "llvm/Support/Casting.h" |
| 79 | #include "llvm/Support/Debug.h" |
| 80 | #include "llvm/Support/DebugCounter.h" |
| 81 | #include "llvm/Support/ErrorHandling.h" |
| 82 | #include "llvm/Support/raw_ostream.h" |
| 83 | #include "llvm/Transforms/Scalar.h" |
| 84 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" |
| 85 | #include "llvm/Transforms/Utils/BuildLibCalls.h" |
| 86 | #include "llvm/Transforms/Utils/Local.h" |
| 87 | #include <algorithm> |
| 88 | #include <cassert> |
| 89 | #include <cstdint> |
| 90 | #include <map> |
| 91 | #include <optional> |
| 92 | #include <utility> |
| 93 | |
| 94 | using namespace llvm; |
| 95 | using namespace PatternMatch; |
| 96 | |
| 97 | #define DEBUG_TYPE "dse" |
| 98 | |
| 99 | STATISTIC(NumRemainingStores, "Number of stores remaining after DSE"); |
| 100 | STATISTIC(NumRedundantStores, "Number of redundant stores deleted"); |
| 101 | STATISTIC(NumFastStores, "Number of stores deleted"); |
| 102 | STATISTIC(NumFastOther, "Number of other instrs removed"); |
| 103 | STATISTIC(NumCompletePartials, "Number of stores dead by later partials"); |
| 104 | STATISTIC(NumModifiedStores, "Number of stores modified"); |
| 105 | STATISTIC(NumCFGChecks, "Number of stores modified"); |
| 106 | STATISTIC(NumCFGTries, "Number of stores modified"); |
| 107 | STATISTIC(NumCFGSuccess, "Number of stores modified"); |
| 108 | STATISTIC(NumGetDomMemoryDefPassed, |
| 109 | "Number of times a valid candidate is returned from getDomMemoryDef"); |
| 110 | STATISTIC(NumDomMemDefChecks, |
| 111 | "Number iterations check for reads in getDomMemoryDef"); |
| 112 | |
| 113 | DEBUG_COUNTER(MemorySSACounter, "dse-memoryssa", |
| 114 | "Controls which MemoryDefs are eliminated."); |
| 115 | |
| 116 | //===----------------------------------------------------------------------===// |
| 117 | // Helper functions |
| 118 | //===----------------------------------------------------------------------===// |
| 119 | using OverlapIntervalsTy = std::map<int64_t, int64_t>; |
| 120 | using InstOverlapIntervalsTy = MapVector<Instruction *, OverlapIntervalsTy>; |
| 121 | |
| 122 | /// Returns true if the end of this instruction can be safely shortened in |
| 123 | /// length. |
| 124 | static bool isShortenableAtTheEnd(Instruction *I) { |
| 125 | // Don't shorten stores for now |
| 126 | if (isa<StoreInst>(Val: I)) |
| 127 | return false; |
| 128 | |
| 129 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: I)) { |
| 130 | switch (II->getIntrinsicID()) { |
| 131 | default: return false; |
| 132 | case Intrinsic::memset: |
| 133 | case Intrinsic::memcpy: |
| 134 | case Intrinsic::memcpy_element_unordered_atomic: |
| 135 | case Intrinsic::memset_element_unordered_atomic: |
| 136 | // Do shorten memory intrinsics. |
| 137 | // FIXME: Add memmove if it's also safe to transform. |
| 138 | return true; |
| 139 | } |
| 140 | } |
| 141 | |
| 142 | // Don't shorten libcalls calls for now. |
| 143 | |
| 144 | return false; |
| 145 | } |
| 146 | |
| 147 | /// Returns true if the beginning of this instruction can be safely shortened |
| 148 | /// in length. |
| 149 | static bool isShortenableAtTheBeginning(Instruction *I) { |
| 150 | // FIXME: Handle only memset for now. Supporting memcpy/memmove should be |
| 151 | // easily done by offsetting the source address. |
| 152 | return isa<AnyMemSetInst>(Val: I); |
| 153 | } |
| 154 | |
| 155 | static std::optional<TypeSize> getPointerSize(const Value *V, |
| 156 | const DataLayout &DL, |
| 157 | const TargetLibraryInfo &TLI, |
| 158 | const Function *F) { |
| 159 | uint64_t Size; |
| 160 | ObjectSizeOpts Opts; |
| 161 | Opts.NullIsUnknownSize = NullPointerIsDefined(F); |
| 162 | |
| 163 | if (getObjectSize(Ptr: V, Size, DL, TLI: &TLI, Opts)) |
| 164 | return TypeSize::getFixed(ExactSize: Size); |
| 165 | return std::nullopt; |
| 166 | } |
| 167 | |
| 168 | namespace { |
| 169 | |
| 170 | enum OverwriteResult { |
| 171 | OW_Begin, |
| 172 | OW_Complete, |
| 173 | OW_End, |
| 174 | OW_PartialEarlierWithFullLater, |
| 175 | OW_MaybePartial, |
| 176 | OW_None, |
| 177 | OW_Unknown |
| 178 | }; |
| 179 | |
| 180 | } // end anonymous namespace |
| 181 | |
| 182 | /// Check if two instruction are masked stores that completely |
| 183 | /// overwrite one another. More specifically, \p KillingI has to |
| 184 | /// overwrite \p DeadI. |
| 185 | static OverwriteResult isMaskedStoreOverwrite(const Instruction *KillingI, |
| 186 | const Instruction *DeadI, |
| 187 | BatchAAResults &AA) { |
| 188 | const auto *KillingII = dyn_cast<IntrinsicInst>(Val: KillingI); |
| 189 | const auto *DeadII = dyn_cast<IntrinsicInst>(Val: DeadI); |
| 190 | if (KillingII == nullptr || DeadII == nullptr) |
| 191 | return OW_Unknown; |
| 192 | if (KillingII->getIntrinsicID() != DeadII->getIntrinsicID()) |
| 193 | return OW_Unknown; |
| 194 | |
| 195 | switch (KillingII->getIntrinsicID()) { |
| 196 | case Intrinsic::masked_store: |
| 197 | case Intrinsic::vp_store: { |
| 198 | const DataLayout &DL = KillingII->getDataLayout(); |
| 199 | auto *KillingTy = KillingII->getArgOperand(i: 0)->getType(); |
| 200 | auto *DeadTy = DeadII->getArgOperand(i: 0)->getType(); |
| 201 | if (DL.getTypeSizeInBits(Ty: KillingTy) != DL.getTypeSizeInBits(Ty: DeadTy)) |
| 202 | return OW_Unknown; |
| 203 | // Element count. |
| 204 | if (cast<VectorType>(Val: KillingTy)->getElementCount() != |
| 205 | cast<VectorType>(Val: DeadTy)->getElementCount()) |
| 206 | return OW_Unknown; |
| 207 | // Pointers. |
| 208 | Value *KillingPtr = KillingII->getArgOperand(i: 1); |
| 209 | Value *DeadPtr = DeadII->getArgOperand(i: 1); |
| 210 | if (KillingPtr != DeadPtr && !AA.isMustAlias(V1: KillingPtr, V2: DeadPtr)) |
| 211 | return OW_Unknown; |
| 212 | if (KillingII->getIntrinsicID() == Intrinsic::masked_store) { |
| 213 | // Masks. |
| 214 | // TODO: check that KillingII's mask is a superset of the DeadII's mask. |
| 215 | if (KillingII->getArgOperand(i: 2) != DeadII->getArgOperand(i: 2)) |
| 216 | return OW_Unknown; |
| 217 | } else if (KillingII->getIntrinsicID() == Intrinsic::vp_store) { |
| 218 | // Masks. |
| 219 | // TODO: check that KillingII's mask is a superset of the DeadII's mask. |
| 220 | if (KillingII->getArgOperand(i: 2) != DeadII->getArgOperand(i: 2)) |
| 221 | return OW_Unknown; |
| 222 | // Lengths. |
| 223 | if (KillingII->getArgOperand(i: 3) != DeadII->getArgOperand(i: 3)) |
| 224 | return OW_Unknown; |
| 225 | } |
| 226 | return OW_Complete; |
| 227 | } |
| 228 | default: |
| 229 | return OW_Unknown; |
| 230 | } |
| 231 | } |
| 232 | |
| 233 | /// Return 'OW_Complete' if a store to the 'KillingLoc' location completely |
| 234 | /// overwrites a store to the 'DeadLoc' location, 'OW_End' if the end of the |
| 235 | /// 'DeadLoc' location is completely overwritten by 'KillingLoc', 'OW_Begin' |
| 236 | /// if the beginning of the 'DeadLoc' location is overwritten by 'KillingLoc'. |
| 237 | /// 'OW_PartialEarlierWithFullLater' means that a dead (big) store was |
| 238 | /// overwritten by a killing (smaller) store which doesn't write outside the big |
| 239 | /// store's memory locations. Returns 'OW_Unknown' if nothing can be determined. |
| 240 | /// NOTE: This function must only be called if both \p KillingLoc and \p |
| 241 | /// DeadLoc belong to the same underlying object with valid \p KillingOff and |
| 242 | /// \p DeadOff. |
| 243 | static OverwriteResult isPartialOverwrite(const ScalarOptions &Opts, |
| 244 | const MemoryLocation &KillingLoc, |
| 245 | const MemoryLocation &DeadLoc, |
| 246 | int64_t KillingOff, int64_t DeadOff, |
| 247 | Instruction *DeadI, |
| 248 | InstOverlapIntervalsTy &IOL) { |
| 249 | const uint64_t KillingSize = KillingLoc.Size.getValue(); |
| 250 | const uint64_t DeadSize = DeadLoc.Size.getValue(); |
| 251 | // We may now overlap, although the overlap is not complete. There might also |
| 252 | // be other incomplete overlaps, and together, they might cover the complete |
| 253 | // dead store. |
| 254 | // Note: The correctness of this logic depends on the fact that this function |
| 255 | // is not even called providing DepWrite when there are any intervening reads. |
| 256 | if (Opts.enable_dse_partial_overwrite_tracking && |
| 257 | KillingOff < int64_t(DeadOff + DeadSize) && |
| 258 | int64_t(KillingOff + KillingSize) >= DeadOff) { |
| 259 | |
| 260 | // Insert our part of the overlap into the map. |
| 261 | auto &IM = IOL[DeadI]; |
| 262 | LLVM_DEBUG(dbgs() << "DSE: Partial overwrite: DeadLoc ["<< DeadOff << ", " |
| 263 | << int64_t(DeadOff + DeadSize) << ") KillingLoc [" |
| 264 | << KillingOff << ", "<< int64_t(KillingOff + KillingSize) |
| 265 | << ")\n"); |
| 266 | |
| 267 | // Make sure that we only insert non-overlapping intervals and combine |
| 268 | // adjacent intervals. The intervals are stored in the map with the ending |
| 269 | // offset as the key (in the half-open sense) and the starting offset as |
| 270 | // the value. |
| 271 | int64_t KillingIntStart = KillingOff; |
| 272 | int64_t KillingIntEnd = KillingOff + KillingSize; |
| 273 | |
| 274 | // Find any intervals ending at, or after, KillingIntStart which start |
| 275 | // before KillingIntEnd. |
| 276 | auto ILI = IM.lower_bound(x: KillingIntStart); |
| 277 | if (ILI != IM.end() && ILI->second <= KillingIntEnd) { |
| 278 | // This existing interval is overlapped with the current store somewhere |
| 279 | // in [KillingIntStart, KillingIntEnd]. Merge them by erasing the existing |
| 280 | // intervals and adjusting our start and end. |
| 281 | KillingIntStart = std::min(a: KillingIntStart, b: ILI->second); |
| 282 | KillingIntEnd = std::max(a: KillingIntEnd, b: ILI->first); |
| 283 | ILI = IM.erase(position: ILI); |
| 284 | |
| 285 | // Continue erasing and adjusting our end in case other previous |
| 286 | // intervals are also overlapped with the current store. |
| 287 | // |
| 288 | // |--- dead 1 ---| |--- dead 2 ---| |
| 289 | // |------- killing---------| |
| 290 | // |
| 291 | while (ILI != IM.end() && ILI->second <= KillingIntEnd) { |
| 292 | assert(ILI->second > KillingIntStart && "Unexpected interval"); |
| 293 | KillingIntEnd = std::max(a: KillingIntEnd, b: ILI->first); |
| 294 | ILI = IM.erase(position: ILI); |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | IM[KillingIntEnd] = KillingIntStart; |
| 299 | |
| 300 | ILI = IM.begin(); |
| 301 | if (ILI->second <= DeadOff && ILI->first >= int64_t(DeadOff + DeadSize)) { |
| 302 | LLVM_DEBUG(dbgs() << "DSE: Full overwrite from partials: DeadLoc [" |
| 303 | << DeadOff << ", "<< int64_t(DeadOff + DeadSize) |
| 304 | << ") Composite KillingLoc ["<< ILI->second << ", " |
| 305 | << ILI->first << ")\n"); |
| 306 | ++NumCompletePartials; |
| 307 | return OW_Complete; |
| 308 | } |
| 309 | } |
| 310 | |
| 311 | // Check for a dead store which writes to all the memory locations that |
| 312 | // the killing store writes to. |
| 313 | if (Opts.enable_dse_partial_store_merging && KillingOff >= DeadOff && |
| 314 | int64_t(DeadOff + DeadSize) > KillingOff && |
| 315 | uint64_t(KillingOff - DeadOff) + KillingSize <= DeadSize) { |
| 316 | LLVM_DEBUG(dbgs() << "DSE: Partial overwrite a dead load ["<< DeadOff |
| 317 | << ", "<< int64_t(DeadOff + DeadSize) |
| 318 | << ") by a killing store ["<< KillingOff << ", " |
| 319 | << int64_t(KillingOff + KillingSize) << ")\n"); |
| 320 | // TODO: Maybe come up with a better name? |
| 321 | return OW_PartialEarlierWithFullLater; |
| 322 | } |
| 323 | |
| 324 | // Another interesting case is if the killing store overwrites the end of the |
| 325 | // dead store. |
| 326 | // |
| 327 | // |--dead--| |
| 328 | // |-- killing --| |
| 329 | // |
| 330 | // In this case we may want to trim the size of dead store to avoid |
| 331 | // generating stores to addresses which will definitely be overwritten killing |
| 332 | // store. |
| 333 | if (!Opts.enable_dse_partial_overwrite_tracking && |
| 334 | (KillingOff > DeadOff && KillingOff < int64_t(DeadOff + DeadSize) && |
| 335 | int64_t(KillingOff + KillingSize) >= int64_t(DeadOff + DeadSize))) |
| 336 | return OW_End; |
| 337 | |
| 338 | // Finally, we also need to check if the killing store overwrites the |
| 339 | // beginning of the dead store. |
| 340 | // |
| 341 | // |--dead--| |
| 342 | // |-- killing --| |
| 343 | // |
| 344 | // In this case we may want to move the destination address and trim the size |
| 345 | // of dead store to avoid generating stores to addresses which will definitely |
| 346 | // be overwritten killing store. |
| 347 | if (!Opts.enable_dse_partial_overwrite_tracking && |
| 348 | (KillingOff <= DeadOff && int64_t(KillingOff + KillingSize) > DeadOff)) { |
| 349 | assert(int64_t(KillingOff + KillingSize) < int64_t(DeadOff + DeadSize) && |
| 350 | "Expect to be handled as OW_Complete"); |
| 351 | return OW_Begin; |
| 352 | } |
| 353 | // Otherwise, they don't completely overlap. |
| 354 | return OW_Unknown; |
| 355 | } |
| 356 | |
| 357 | /// Returns true if the memory which is accessed by the second instruction is not |
| 358 | /// modified between the first and the second instruction. |
| 359 | /// Precondition: Second instruction must be dominated by the first |
| 360 | /// instruction. |
| 361 | static bool |
| 362 | memoryIsNotModifiedBetween(Instruction *FirstI, Instruction *SecondI, |
| 363 | BatchAAResults &AA, const DataLayout &DL, |
| 364 | DominatorTree *DT) { |
| 365 | // Do a backwards scan through the CFG from SecondI to FirstI. Look for |
| 366 | // instructions which can modify the memory location accessed by SecondI. |
| 367 | // |
| 368 | // While doing the walk keep track of the address to check. It might be |
| 369 | // different in different basic blocks due to PHI translation. |
| 370 | using BlockAddressPair = std::pair<BasicBlock *, PHITransAddr>; |
| 371 | SmallVector<BlockAddressPair, 16> WorkList; |
| 372 | // Keep track of the address we visited each block with. Bail out if we |
| 373 | // visit a block with different addresses. |
| 374 | DenseMap<BasicBlock *, Value *> Visited; |
| 375 | |
| 376 | BasicBlock::iterator FirstBBI(FirstI); |
| 377 | ++FirstBBI; |
| 378 | BasicBlock::iterator SecondBBI(SecondI); |
| 379 | BasicBlock *FirstBB = FirstI->getParent(); |
| 380 | BasicBlock *SecondBB = SecondI->getParent(); |
| 381 | MemoryLocation MemLoc; |
| 382 | if (auto *MemSet = dyn_cast<MemSetInst>(Val: SecondI)) |
| 383 | MemLoc = MemoryLocation::getForDest(MI: MemSet); |
| 384 | else |
| 385 | MemLoc = MemoryLocation::get(Inst: SecondI); |
| 386 | |
| 387 | auto *MemLocPtr = const_cast<Value *>(MemLoc.Ptr); |
| 388 | |
| 389 | // Start checking the SecondBB. |
| 390 | WorkList.push_back( |
| 391 | Elt: std::make_pair(x&: SecondBB, y: PHITransAddr(MemLocPtr, DL, nullptr))); |
| 392 | bool isFirstBlock = true; |
| 393 | |
| 394 | // Check all blocks going backward until we reach the FirstBB. |
| 395 | while (!WorkList.empty()) { |
| 396 | BlockAddressPair Current = WorkList.pop_back_val(); |
| 397 | BasicBlock *B = Current.first; |
| 398 | PHITransAddr &Addr = Current.second; |
| 399 | Value *Ptr = Addr.getAddr(); |
| 400 | |
| 401 | // Ignore instructions before FirstI if this is the FirstBB. |
| 402 | BasicBlock::iterator BI = (B == FirstBB ? FirstBBI : B->begin()); |
| 403 | |
| 404 | BasicBlock::iterator EI; |
| 405 | if (isFirstBlock) { |
| 406 | // Ignore instructions after SecondI if this is the first visit of SecondBB. |
| 407 | assert(B == SecondBB && "first block is not the store block"); |
| 408 | EI = SecondBBI; |
| 409 | isFirstBlock = false; |
| 410 | } else { |
| 411 | // It's not SecondBB or (in case of a loop) the second visit of SecondBB. |
| 412 | // In this case we also have to look at instructions after SecondI. |
| 413 | EI = B->end(); |
| 414 | } |
| 415 | for (; BI != EI; ++BI) { |
| 416 | Instruction *I = &*BI; |
| 417 | if (I->mayWriteToMemory() && I != SecondI) |
| 418 | if (isModSet(MRI: AA.getModRefInfo(I, OptLoc: MemLoc.getWithNewPtr(NewPtr: Ptr)))) |
| 419 | return false; |
| 420 | } |
| 421 | if (B != FirstBB) { |
| 422 | assert(B != &FirstBB->getParent()->getEntryBlock() && |
| 423 | "Should not hit the entry block because SI must be dominated by LI"); |
| 424 | for (BasicBlock *Pred : predecessors(BB: B)) { |
| 425 | PHITransAddr PredAddr = Addr; |
| 426 | if (PredAddr.needsPHITranslationFromBlock(BB: B)) { |
| 427 | if (!PredAddr.isPotentiallyPHITranslatable()) |
| 428 | return false; |
| 429 | if (!PredAddr.translateValue(CurBB: B, PredBB: Pred, DT, MustDominate: false)) |
| 430 | return false; |
| 431 | } |
| 432 | Value *TranslatedPtr = PredAddr.getAddr(); |
| 433 | auto Inserted = Visited.insert(KV: std::make_pair(x&: Pred, y&: TranslatedPtr)); |
| 434 | if (!Inserted.second) { |
| 435 | // We already visited this block before. If it was with a different |
| 436 | // address - bail out! |
| 437 | if (TranslatedPtr != Inserted.first->second) |
| 438 | return false; |
| 439 | // ... otherwise just skip it. |
| 440 | continue; |
| 441 | } |
| 442 | WorkList.push_back(Elt: std::make_pair(x&: Pred, y&: PredAddr)); |
| 443 | } |
| 444 | } |
| 445 | } |
| 446 | return true; |
| 447 | } |
| 448 | |
| 449 | static void shortenAssignment(Instruction *Inst, Value *OriginalDest, |
| 450 | uint64_t OldSizeInBits, uint64_t NewSizeInBits, |
| 451 | bool IsOverwriteEnd) { |
| 452 | const DataLayout &DL = Inst->getDataLayout(); |
| 453 | uint64_t DeadSliceSizeInBits = OldSizeInBits - NewSizeInBits; |
| 454 | // The dead slice offset is relative to OriginalDest, the slice start we hand |
| 455 | // calculateFragmentIntersect. Shortening the end keeps the front of the |
| 456 | // store, so the dead bits start where the new store ends; shortening the |
| 457 | // beginning kills the bits at OriginalDest itself. Don't add OriginalDest's |
| 458 | // offset from its base object here. calculateFragmentIntersect already |
| 459 | // measures that pointer against the marker's address. |
| 460 | uint64_t DeadSliceOffsetInBits = IsOverwriteEnd ? NewSizeInBits : 0; |
| 461 | auto SetDeadFragExpr = [](auto *Assign, |
| 462 | DIExpression::FragmentInfo DeadFragment) { |
| 463 | // createFragmentExpression expects an offset relative to the existing |
| 464 | // fragment offset if there is one. |
| 465 | uint64_t RelativeOffset = DeadFragment.OffsetInBits - |
| 466 | Assign->getExpression() |
| 467 | ->getFragmentInfo() |
| 468 | .value_or(DIExpression::FragmentInfo(0, 0)) |
| 469 | .OffsetInBits; |
| 470 | if (auto NewExpr = DIExpression::createFragmentExpression( |
| 471 | Expr: Assign->getExpression(), OffsetInBits: RelativeOffset, SizeInBits: DeadFragment.SizeInBits)) { |
| 472 | Assign->setExpression(*NewExpr); |
| 473 | return; |
| 474 | } |
| 475 | // Failed to create a fragment expression for this so discard the value, |
| 476 | // making this a kill location. |
| 477 | auto *Expr = *DIExpression::createFragmentExpression( |
| 478 | Expr: DIExpression::get(Context&: Assign->getContext(), Elements: {}), OffsetInBits: DeadFragment.OffsetInBits, |
| 479 | SizeInBits: DeadFragment.SizeInBits); |
| 480 | Assign->setExpression(Expr); |
| 481 | Assign->setKillLocation(); |
| 482 | }; |
| 483 | |
| 484 | // A DIAssignID to use so that the inserted dbg.assign intrinsics do not |
| 485 | // link to any instructions. Created in the loop below (once). |
| 486 | DIAssignID *LinkToNothing = nullptr; |
| 487 | LLVMContext &Ctx = Inst->getContext(); |
| 488 | auto GetDeadLink = [&Ctx, &LinkToNothing]() { |
| 489 | if (!LinkToNothing) |
| 490 | LinkToNothing = DIAssignID::getDistinct(Context&: Ctx); |
| 491 | return LinkToNothing; |
| 492 | }; |
| 493 | |
| 494 | // Insert an unlinked dbg.assign intrinsic for the dead fragment after each |
| 495 | // overlapping dbg.assign intrinsic. |
| 496 | for (DbgVariableRecord *Assign : at::getDVRAssignmentMarkers(Inst)) { |
| 497 | std::optional<DIExpression::FragmentInfo> NewFragment; |
| 498 | if (!at::calculateFragmentIntersect(DL, Dest: OriginalDest, SliceOffsetInBits: DeadSliceOffsetInBits, |
| 499 | SliceSizeInBits: DeadSliceSizeInBits, DVRAssign: Assign, |
| 500 | Result&: NewFragment) || |
| 501 | !NewFragment) { |
| 502 | // Either the intersection couldn't be worked out, or it covers the |
| 503 | // entire variable region described by the record. Full coverage leaves |
| 504 | // NewFragment empty rather than making calculateFragmentIntersect fail, |
| 505 | // so unlink the whole assignment from the store in both cases. |
| 506 | Assign->setKillAddress(); |
| 507 | Assign->setAssignId(GetDeadLink()); |
| 508 | continue; |
| 509 | } |
| 510 | // No intersect. |
| 511 | if (NewFragment->SizeInBits == 0) |
| 512 | continue; |
| 513 | |
| 514 | // Fragments overlap: insert a new dbg.assign for this dead part. |
| 515 | auto *NewAssign = static_cast<decltype(Assign)>(Assign->clone()); |
| 516 | NewAssign->insertAfter(InsertAfter: Assign->getIterator()); |
| 517 | NewAssign->setAssignId(GetDeadLink()); |
| 518 | if (NewFragment) |
| 519 | SetDeadFragExpr(NewAssign, *NewFragment); |
| 520 | NewAssign->setKillAddress(); |
| 521 | } |
| 522 | } |
| 523 | |
| 524 | /// Update the attributes given that a memory access is updated (the |
| 525 | /// dereferenced pointer could be moved forward when shortening a |
| 526 | /// mem intrinsic). |
| 527 | static void adjustArgAttributes(AnyMemIntrinsic *Intrinsic, unsigned ArgNo, |
| 528 | uint64_t PtrOffset) { |
| 529 | // Remember old attributes. |
| 530 | AttributeSet OldAttrs = Intrinsic->getParamAttributes(ArgNo); |
| 531 | |
| 532 | // Find attributes that should be kept, and remove the rest. |
| 533 | AttributeMask AttrsToRemove; |
| 534 | for (auto &Attr : OldAttrs) { |
| 535 | if (Attr.hasKindAsEnum()) { |
| 536 | switch (Attr.getKindAsEnum()) { |
| 537 | default: |
| 538 | break; |
| 539 | case Attribute::Alignment: |
| 540 | // Only keep alignment if PtrOffset satisfy the alignment. |
| 541 | if (isAligned(Lhs: Attr.getAlignment().valueOrOne(), SizeInBytes: PtrOffset)) |
| 542 | continue; |
| 543 | break; |
| 544 | case Attribute::Dereferenceable: |
| 545 | case Attribute::DereferenceableOrNull: |
| 546 | // We could reduce the size of these attributes according to |
| 547 | // PtrOffset. But we simply drop these for now. |
| 548 | break; |
| 549 | case Attribute::NonNull: |
| 550 | case Attribute::NoUndef: |
| 551 | continue; |
| 552 | } |
| 553 | } |
| 554 | AttrsToRemove.addAttribute(A: Attr); |
| 555 | } |
| 556 | |
| 557 | // Remove the attributes that should be dropped. |
| 558 | Intrinsic->removeParamAttrs(ArgNo, AttrsToRemove); |
| 559 | } |
| 560 | |
| 561 | static bool tryToShorten(Instruction *DeadI, int64_t &DeadStart, |
| 562 | uint64_t &DeadSize, int64_t KillingStart, |
| 563 | uint64_t KillingSize, bool IsOverwriteEnd) { |
| 564 | auto *DeadIntrinsic = cast<AnyMemIntrinsic>(Val: DeadI); |
| 565 | Align PrefAlign = DeadIntrinsic->getDestAlign().valueOrOne(); |
| 566 | |
| 567 | // We assume that memet/memcpy operates in chunks of the "largest" native |
| 568 | // type size and aligned on the same value. That means optimal start and size |
| 569 | // of memset/memcpy should be modulo of preferred alignment of that type. That |
| 570 | // is it there is no any sense in trying to reduce store size any further |
| 571 | // since any "extra" stores comes for free anyway. |
| 572 | // On the other hand, maximum alignment we can achieve is limited by alignment |
| 573 | // of initial store. |
| 574 | |
| 575 | // TODO: Limit maximum alignment by preferred (or abi?) alignment of the |
| 576 | // "largest" native type. |
| 577 | // Note: What is the proper way to get that value? |
| 578 | // Should TargetTransformInfo::getRegisterBitWidth be used or anything else? |
| 579 | // PrefAlign = std::min(DL.getPrefTypeAlign(LargestType), PrefAlign); |
| 580 | |
| 581 | int64_t ToRemoveStart = 0; |
| 582 | uint64_t ToRemoveSize = 0; |
| 583 | // Compute start and size of the region to remove. Make sure 'PrefAlign' is |
| 584 | // maintained on the remaining store. |
| 585 | if (IsOverwriteEnd) { |
| 586 | // Calculate required adjustment for 'KillingStart' in order to keep |
| 587 | // remaining store size aligned on 'PerfAlign'. |
| 588 | uint64_t Off = |
| 589 | offsetToAlignment(Value: uint64_t(KillingStart - DeadStart), Alignment: PrefAlign); |
| 590 | ToRemoveStart = KillingStart + Off; |
| 591 | if (DeadSize <= uint64_t(ToRemoveStart - DeadStart)) |
| 592 | return false; |
| 593 | ToRemoveSize = DeadSize - uint64_t(ToRemoveStart - DeadStart); |
| 594 | } else { |
| 595 | ToRemoveStart = DeadStart; |
| 596 | assert(KillingSize >= uint64_t(DeadStart - KillingStart) && |
| 597 | "Not overlapping accesses?"); |
| 598 | ToRemoveSize = KillingSize - uint64_t(DeadStart - KillingStart); |
| 599 | // Calculate required adjustment for 'ToRemoveSize'in order to keep |
| 600 | // start of the remaining store aligned on 'PerfAlign'. |
| 601 | uint64_t Off = offsetToAlignment(Value: ToRemoveSize, Alignment: PrefAlign); |
| 602 | if (Off != 0) { |
| 603 | if (ToRemoveSize <= (PrefAlign.value() - Off)) |
| 604 | return false; |
| 605 | ToRemoveSize -= PrefAlign.value() - Off; |
| 606 | } |
| 607 | assert(isAligned(PrefAlign, ToRemoveSize) && |
| 608 | "Should preserve selected alignment"); |
| 609 | } |
| 610 | |
| 611 | assert(ToRemoveSize > 0 && "Shouldn't reach here if nothing to remove"); |
| 612 | assert(DeadSize > ToRemoveSize && "Can't remove more than original size"); |
| 613 | |
| 614 | uint64_t NewSize = DeadSize - ToRemoveSize; |
| 615 | if (DeadIntrinsic->isAtomic()) { |
| 616 | // When shortening an atomic memory intrinsic, the newly shortened |
| 617 | // length must remain an integer multiple of the element size. |
| 618 | const uint32_t ElementSize = DeadIntrinsic->getElementSizeInBytes(); |
| 619 | if (0 != NewSize % ElementSize) |
| 620 | return false; |
| 621 | } |
| 622 | |
| 623 | LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n OW " |
| 624 | << (IsOverwriteEnd ? "END": "BEGIN") << ": "<< *DeadI |
| 625 | << "\n KILLER ["<< ToRemoveStart << ", " |
| 626 | << int64_t(ToRemoveStart + ToRemoveSize) << ")\n"); |
| 627 | |
| 628 | DeadIntrinsic->setLength(NewSize); |
| 629 | DeadIntrinsic->setDestAlignment(PrefAlign); |
| 630 | |
| 631 | Value *OrigDest = DeadIntrinsic->getRawDest(); |
| 632 | if (!IsOverwriteEnd) { |
| 633 | Value *Indices[1] = { |
| 634 | ConstantInt::get(Ty: DeadIntrinsic->getLength()->getType(), V: ToRemoveSize)}; |
| 635 | Instruction *NewDestGEP = GetElementPtrInst::CreateInBounds( |
| 636 | PointeeType: Type::getInt8Ty(C&: DeadIntrinsic->getContext()), Ptr: OrigDest, IdxList: Indices, NameStr: "", |
| 637 | InsertBefore: DeadI->getIterator()); |
| 638 | NewDestGEP->setDebugLoc(DeadIntrinsic->getDebugLoc()); |
| 639 | DeadIntrinsic->setDest(NewDestGEP); |
| 640 | adjustArgAttributes(Intrinsic: DeadIntrinsic, ArgNo: 0, PtrOffset: ToRemoveSize); |
| 641 | } |
| 642 | |
| 643 | // Update attached dbg.assign intrinsics. Assume 8-bit byte. |
| 644 | shortenAssignment(Inst: DeadI, OriginalDest: OrigDest, OldSizeInBits: DeadSize * 8, NewSizeInBits: NewSize * 8, IsOverwriteEnd); |
| 645 | |
| 646 | // Finally update start and size of dead access. |
| 647 | if (!IsOverwriteEnd) |
| 648 | DeadStart += ToRemoveSize; |
| 649 | DeadSize = NewSize; |
| 650 | |
| 651 | return true; |
| 652 | } |
| 653 | |
| 654 | static bool tryToShortenEnd(Instruction *DeadI, OverlapIntervalsTy &IntervalMap, |
| 655 | int64_t &DeadStart, uint64_t &DeadSize) { |
| 656 | if (IntervalMap.empty() || !isShortenableAtTheEnd(I: DeadI)) |
| 657 | return false; |
| 658 | |
| 659 | OverlapIntervalsTy::iterator OII = --IntervalMap.end(); |
| 660 | int64_t KillingStart = OII->second; |
| 661 | uint64_t KillingSize = OII->first - KillingStart; |
| 662 | |
| 663 | assert(OII->first - KillingStart >= 0 && "Size expected to be positive"); |
| 664 | |
| 665 | if (KillingStart > DeadStart && |
| 666 | // Note: "KillingStart - KillingStart" is known to be positive due to |
| 667 | // preceding check. |
| 668 | (uint64_t)(KillingStart - DeadStart) < DeadSize && |
| 669 | // Note: "DeadSize - (uint64_t)(KillingStart - DeadStart)" is known to |
| 670 | // be non negative due to preceding checks. |
| 671 | KillingSize >= DeadSize - (uint64_t)(KillingStart - DeadStart)) { |
| 672 | if (tryToShorten(DeadI, DeadStart, DeadSize, KillingStart, KillingSize, |
| 673 | IsOverwriteEnd: true)) { |
| 674 | IntervalMap.erase(position: OII); |
| 675 | return true; |
| 676 | } |
| 677 | } |
| 678 | return false; |
| 679 | } |
| 680 | |
| 681 | static bool tryToShortenBegin(Instruction *DeadI, |
| 682 | OverlapIntervalsTy &IntervalMap, |
| 683 | int64_t &DeadStart, uint64_t &DeadSize) { |
| 684 | if (IntervalMap.empty() || !isShortenableAtTheBeginning(I: DeadI)) |
| 685 | return false; |
| 686 | |
| 687 | OverlapIntervalsTy::iterator OII = IntervalMap.begin(); |
| 688 | int64_t KillingStart = OII->second; |
| 689 | uint64_t KillingSize = OII->first - KillingStart; |
| 690 | |
| 691 | assert(OII->first - KillingStart >= 0 && "Size expected to be positive"); |
| 692 | |
| 693 | if (KillingStart <= DeadStart && |
| 694 | // Note: "DeadStart - KillingStart" is known to be non negative due to |
| 695 | // preceding check. |
| 696 | KillingSize > (uint64_t)(DeadStart - KillingStart)) { |
| 697 | // Note: "KillingSize - (uint64_t)(DeadStart - DeadStart)" is known to |
| 698 | // be positive due to preceding checks. |
| 699 | assert(KillingSize - (uint64_t)(DeadStart - KillingStart) < DeadSize && |
| 700 | "Should have been handled as OW_Complete"); |
| 701 | if (tryToShorten(DeadI, DeadStart, DeadSize, KillingStart, KillingSize, |
| 702 | IsOverwriteEnd: false)) { |
| 703 | IntervalMap.erase(position: OII); |
| 704 | return true; |
| 705 | } |
| 706 | } |
| 707 | return false; |
| 708 | } |
| 709 | |
| 710 | static Constant * |
| 711 | tryToMergePartialOverlappingStores(StoreInst *KillingI, StoreInst *DeadI, |
| 712 | int64_t KillingOffset, int64_t DeadOffset, |
| 713 | const DataLayout &DL, BatchAAResults &AA, |
| 714 | DominatorTree *DT) { |
| 715 | assert(KillingI); |
| 716 | assert(DeadI); |
| 717 | |
| 718 | // If the store we find is: |
| 719 | // a) partially overwritten by the store to 'Loc' |
| 720 | // b) the killing store is fully contained in the dead one and |
| 721 | // c) they both have a constant value |
| 722 | // d) none of the two stores need padding |
| 723 | // Merge the two stores, replacing the dead store's value with a |
| 724 | // merge of both values. |
| 725 | // |
| 726 | // TODO: Deal with other constant types (vectors, etc), and probably |
| 727 | // some mem intrinsics (if needed) |
| 728 | if (!isa<ConstantInt>(Val: DeadI->getValueOperand()) || |
| 729 | !DL.typeSizeEqualsStoreSize(Ty: DeadI->getValueOperand()->getType()) || |
| 730 | !isa<ConstantInt>(Val: KillingI->getValueOperand()) || |
| 731 | !DL.typeSizeEqualsStoreSize(Ty: KillingI->getValueOperand()->getType()) || |
| 732 | !memoryIsNotModifiedBetween(FirstI: DeadI, SecondI: KillingI, AA, DL, DT)) |
| 733 | return nullptr; |
| 734 | |
| 735 | // The merge erases KillingI and writes its bytes via DeadI. For that to be |
| 736 | // safe: |
| 737 | // - KillingI must be deletable (not volatile, ordering at most unordered), |
| 738 | // - DeadI must be safe to rewrite, and |
| 739 | // - their orderings must match, so the bytes originally written by |
| 740 | // KillingI keep the same atomicity after they are folded into DeadI. |
| 741 | // This allows merging two simple stores or two unordered-atomic stores with |
| 742 | // matching ordering, while leaving volatile and ordered-atomic stores in |
| 743 | // place. |
| 744 | if (!KillingI->isUnordered() || !DeadI->isUnordered() || |
| 745 | KillingI->getOrdering() != DeadI->getOrdering()) |
| 746 | return nullptr; |
| 747 | |
| 748 | APInt DeadValue = cast<ConstantInt>(Val: DeadI->getValueOperand())->getValue(); |
| 749 | APInt KillingValue = |
| 750 | cast<ConstantInt>(Val: KillingI->getValueOperand())->getValue(); |
| 751 | unsigned KillingBits = KillingValue.getBitWidth(); |
| 752 | assert(DeadValue.getBitWidth() > KillingValue.getBitWidth()); |
| 753 | KillingValue = KillingValue.zext(width: DeadValue.getBitWidth()); |
| 754 | |
| 755 | // Offset of the smaller store inside the larger store |
| 756 | unsigned BitOffsetDiff = (KillingOffset - DeadOffset) * 8; |
| 757 | unsigned LShiftAmount = |
| 758 | DL.isBigEndian() ? DeadValue.getBitWidth() - BitOffsetDiff - KillingBits |
| 759 | : BitOffsetDiff; |
| 760 | APInt Mask = APInt::getBitsSet(numBits: DeadValue.getBitWidth(), loBit: LShiftAmount, |
| 761 | hiBit: LShiftAmount + KillingBits); |
| 762 | // Clear the bits we'll be replacing, then OR with the smaller |
| 763 | // store, shifted appropriately. |
| 764 | APInt Merged = (DeadValue & ~Mask) | (KillingValue << LShiftAmount); |
| 765 | LLVM_DEBUG(dbgs() << "DSE: Merge Stores:\n Dead: "<< *DeadI |
| 766 | << "\n Killing: "<< *KillingI |
| 767 | << "\n Merged Value: "<< Merged << '\n'); |
| 768 | return ConstantInt::get(Ty: DeadI->getValueOperand()->getType(), V: Merged); |
| 769 | } |
| 770 | |
| 771 | // Returns true if \p I is an intrinsic that does not read or write memory. |
| 772 | static bool isNoopIntrinsic(Instruction *I) { |
| 773 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: I)) { |
| 774 | switch (II->getIntrinsicID()) { |
| 775 | case Intrinsic::lifetime_start: |
| 776 | case Intrinsic::lifetime_end: |
| 777 | case Intrinsic::invariant_end: |
| 778 | case Intrinsic::launder_invariant_group: |
| 779 | case Intrinsic::assume: |
| 780 | return true; |
| 781 | default: |
| 782 | return false; |
| 783 | } |
| 784 | } |
| 785 | return false; |
| 786 | } |
| 787 | |
| 788 | // Check if we can ignore \p D for DSE. |
| 789 | static bool canSkipDef(MemoryDef *D, bool DefVisibleToCaller) { |
| 790 | Instruction *DI = D->getMemoryInst(); |
| 791 | // Calls that only access inaccessible memory cannot read or write any memory |
| 792 | // locations we consider for elimination. |
| 793 | if (auto *CB = dyn_cast<CallBase>(Val: DI)) |
| 794 | if (CB->onlyAccessesInaccessibleMemory()) |
| 795 | return true; |
| 796 | |
| 797 | // We can eliminate stores to locations not visible to the caller across |
| 798 | // throwing instructions. |
| 799 | if (DI->mayThrow() && !DefVisibleToCaller) |
| 800 | return true; |
| 801 | |
| 802 | // We can remove the dead stores, irrespective of the fence and its ordering |
| 803 | // (release/acquire/seq_cst). Fences only constraints the ordering of |
| 804 | // already visible stores, it does not make a store visible to other |
| 805 | // threads. So, skipping over a fence does not change a store from being |
| 806 | // dead. |
| 807 | if (isa<FenceInst>(Val: DI)) |
| 808 | return true; |
| 809 | |
| 810 | // Skip intrinsics that do not really read or modify memory. |
| 811 | if (isNoopIntrinsic(I: DI)) |
| 812 | return true; |
| 813 | |
| 814 | return false; |
| 815 | } |
| 816 | |
| 817 | namespace { |
| 818 | |
| 819 | // A memory location wrapper that represents a MemoryLocation, `MemLoc`, |
| 820 | // defined by `MemDef`. |
| 821 | struct MemoryLocationWrapper { |
| 822 | MemoryLocationWrapper(MemoryLocation MemLoc, MemoryDef *MemDef, |
| 823 | bool DefByInitializesAttr) |
| 824 | : MemLoc(MemLoc), MemDef(MemDef), |
| 825 | DefByInitializesAttr(DefByInitializesAttr) { |
| 826 | assert(MemLoc.Ptr && "MemLoc should be not null"); |
| 827 | UnderlyingObject = getUnderlyingObject(V: MemLoc.Ptr); |
| 828 | DefInst = MemDef->getMemoryInst(); |
| 829 | } |
| 830 | |
| 831 | MemoryLocation MemLoc; |
| 832 | const Value *UnderlyingObject; |
| 833 | MemoryDef *MemDef; |
| 834 | Instruction *DefInst; |
| 835 | bool DefByInitializesAttr = false; |
| 836 | }; |
| 837 | |
| 838 | // A memory def wrapper that represents a MemoryDef and the MemoryLocation(s) |
| 839 | // defined by this MemoryDef. |
| 840 | struct MemoryDefWrapper { |
| 841 | MemoryDefWrapper(MemoryDef *MemDef, |
| 842 | ArrayRef<std::pair<MemoryLocation, bool>> MemLocations) { |
| 843 | DefInst = MemDef->getMemoryInst(); |
| 844 | for (auto &[MemLoc, DefByInitializesAttr] : MemLocations) |
| 845 | DefinedLocations.push_back( |
| 846 | Elt: MemoryLocationWrapper(MemLoc, MemDef, DefByInitializesAttr)); |
| 847 | } |
| 848 | Instruction *DefInst; |
| 849 | SmallVector<MemoryLocationWrapper, 1> DefinedLocations; |
| 850 | }; |
| 851 | |
| 852 | struct ArgumentInitInfo { |
| 853 | unsigned Idx; |
| 854 | bool IsDeadOrInvisibleOnUnwind; |
| 855 | ConstantRangeList Inits; |
| 856 | }; |
| 857 | } // namespace |
| 858 | |
| 859 | static bool hasInitializesAttr(Instruction *I) { |
| 860 | CallBase *CB = dyn_cast<CallBase>(Val: I); |
| 861 | return CB && CB->getArgOperandWithAttribute(Kind: Attribute::Initializes); |
| 862 | } |
| 863 | |
| 864 | // Return the intersected range list of the initializes attributes of "Args". |
| 865 | // "Args" are call arguments that alias to each other. |
| 866 | // If any argument in "Args" doesn't have dead_on_unwind attr and |
| 867 | // "CallHasNoUnwindAttr" is false, return empty. |
| 868 | static ConstantRangeList |
| 869 | getIntersectedInitRangeList(ArrayRef<ArgumentInitInfo> Args, |
| 870 | bool CallHasNoUnwindAttr) { |
| 871 | if (Args.empty()) |
| 872 | return {}; |
| 873 | |
| 874 | // To address unwind, the function should have nounwind attribute or the |
| 875 | // arguments have dead or invisible on unwind. Otherwise, return empty. |
| 876 | for (const auto &Arg : Args) { |
| 877 | if (!CallHasNoUnwindAttr && !Arg.IsDeadOrInvisibleOnUnwind) |
| 878 | return {}; |
| 879 | if (Arg.Inits.empty()) |
| 880 | return {}; |
| 881 | } |
| 882 | |
| 883 | ConstantRangeList IntersectedIntervals = Args.front().Inits; |
| 884 | for (auto &Arg : Args.drop_front()) |
| 885 | IntersectedIntervals = IntersectedIntervals.intersectWith(CRL: Arg.Inits); |
| 886 | |
| 887 | return IntersectedIntervals; |
| 888 | } |
| 889 | |
| 890 | namespace { |
| 891 | |
| 892 | struct DSEState { |
| 893 | const ScalarOptions &Opts; |
| 894 | Function &F; |
| 895 | AliasAnalysis &AA; |
| 896 | EarliestEscapeAnalysis EA; |
| 897 | |
| 898 | /// The single BatchAA instance that is used to cache AA queries. It will |
| 899 | /// not be invalidated over the whole run. This is safe, because: |
| 900 | /// 1. Only memory writes are removed, so the alias cache for memory |
| 901 | /// locations remains valid. |
| 902 | /// 2. No new instructions are added (only instructions removed), so cached |
| 903 | /// information for a deleted value cannot be accessed by a re-used new |
| 904 | /// value pointer. |
| 905 | BatchAAResults BatchAA; |
| 906 | |
| 907 | MemorySSA &MSSA; |
| 908 | DominatorTree &DT; |
| 909 | PostDominatorTree &PDT; |
| 910 | const TargetLibraryInfo &TLI; |
| 911 | const DataLayout &DL; |
| 912 | const CycleInfo &CI; |
| 913 | |
| 914 | // All MemoryDefs that potentially could kill other MemDefs. |
| 915 | SmallVector<MemoryDef *, 64> MemDefs; |
| 916 | // Any that should be skipped as they are already deleted |
| 917 | SmallPtrSet<MemoryAccess *, 4> SkipStores; |
| 918 | // Keep track whether a given object is captured before return or not. |
| 919 | DenseMap<const Value *, bool> CapturedBeforeReturn; |
| 920 | // Keep track of all of the objects that are invisible to the caller after |
| 921 | // the function returns. |
| 922 | DenseMap<const Value *, bool> InvisibleToCallerAfterRet; |
| 923 | DenseMap<const Value *, uint64_t> InvisibleToCallerAfterRetBounded; |
| 924 | // Keep track of blocks with throwing instructions not modeled in MemorySSA. |
| 925 | SmallPtrSet<BasicBlock *, 16> ThrowingBlocks; |
| 926 | // Post-order numbers for each basic block. Used to figure out if memory |
| 927 | // accesses are executed before another access. |
| 928 | DenseMap<BasicBlock *, unsigned> PostOrderNumbers; |
| 929 | |
| 930 | /// Keep track of instructions (partly) overlapping with killing MemoryDefs per |
| 931 | /// basic block. |
| 932 | MapVector<BasicBlock *, InstOverlapIntervalsTy> IOLs; |
| 933 | // Check if there are root nodes that are terminated by UnreachableInst. |
| 934 | // Those roots pessimize post-dominance queries. If there are such roots, |
| 935 | // fall back to CFG scan starting from all non-unreachable roots. |
| 936 | bool AnyUnreachableExit; |
| 937 | |
| 938 | // Whether or not we should iterate on removing dead stores at the end of the |
| 939 | // function due to removing a store causing a previously captured pointer to |
| 940 | // no longer be captured. |
| 941 | bool ShouldIterateEndOfFunctionDSE; |
| 942 | |
| 943 | /// Dead instructions to be removed at the end of DSE. |
| 944 | SmallVector<Instruction *> ToRemove; |
| 945 | |
| 946 | // Class contains self-reference, make sure it's not copied/moved. |
| 947 | DSEState(const ScalarOptions &Opts, Function &F, AliasAnalysis &AA, |
| 948 | MemorySSA &MSSA, DominatorTree &DT, PostDominatorTree &PDT, |
| 949 | const TargetLibraryInfo &TLI, const CycleInfo &CI); |
| 950 | DSEState(const DSEState &) = delete; |
| 951 | DSEState &operator=(const DSEState &) = delete; |
| 952 | |
| 953 | LocationSize strengthenLocationSize(const Instruction *I, |
| 954 | LocationSize Size) const; |
| 955 | |
| 956 | /// Return 'OW_Complete' if a store to the 'KillingLoc' location (by \p |
| 957 | /// KillingI instruction) completely overwrites a store to the 'DeadLoc' |
| 958 | /// location (by \p DeadI instruction). |
| 959 | /// Return OW_MaybePartial if \p KillingI does not completely overwrite |
| 960 | /// \p DeadI, but they both write to the same underlying object. In that |
| 961 | /// case, use isPartialOverwrite to check if \p KillingI partially overwrites |
| 962 | /// \p DeadI. Returns 'OR_None' if \p KillingI is known to not overwrite the |
| 963 | /// \p DeadI. Returns 'OW_Unknown' if nothing can be determined. |
| 964 | OverwriteResult isOverwrite(const Instruction *KillingI, |
| 965 | const Instruction *DeadI, |
| 966 | const MemoryLocation &KillingLoc, |
| 967 | const MemoryLocation &DeadLoc, |
| 968 | int64_t &KillingOff, int64_t &DeadOff); |
| 969 | |
| 970 | bool isInvisibleToCallerAfterRet(const Value *V, const Value *Ptr, |
| 971 | const LocationSize StoreSize); |
| 972 | |
| 973 | bool isInvisibleToCallerOnUnwind(const Value *V); |
| 974 | |
| 975 | std::optional<MemoryLocation> getLocForWrite(Instruction *I) const; |
| 976 | |
| 977 | // Returns a list of <MemoryLocation, bool> pairs written by I. |
| 978 | // The bool means whether the write is from Initializes attr. |
| 979 | SmallVector<std::pair<MemoryLocation, bool>, 1> |
| 980 | getLocForInst(Instruction *I, bool ConsiderInitializesAttr); |
| 981 | |
| 982 | /// Assuming this instruction has a dead analyzable write, can we delete |
| 983 | /// this instruction? |
| 984 | bool isRemovable(Instruction *I); |
| 985 | |
| 986 | /// Returns true if \p UseInst completely overwrites \p DefLoc |
| 987 | /// (stored by \p DefInst). |
| 988 | bool isCompleteOverwrite(const MemoryLocation &DefLoc, Instruction *DefInst, |
| 989 | Instruction *UseInst); |
| 990 | |
| 991 | /// Returns true if \p Def is not read before returning from the function. |
| 992 | bool isWriteAtEndOfFunction(MemoryDef *Def, const MemoryLocation &DefLoc); |
| 993 | |
| 994 | /// If \p I is a memory terminator like llvm.lifetime.end or free, return a |
| 995 | /// pair with the MemoryLocation terminated by \p I and a boolean flag |
| 996 | /// indicating whether \p I is a free-like call. |
| 997 | std::optional<std::pair<MemoryLocation, bool>> |
| 998 | getLocForTerminator(Instruction *I) const; |
| 999 | |
| 1000 | /// Returns true if \p I is a memory terminator instruction like |
| 1001 | /// llvm.lifetime.end or free. |
| 1002 | bool isMemTerminatorInst(Instruction *I) const; |
| 1003 | |
| 1004 | /// Returns true if \p MaybeTerm is a memory terminator for \p Loc from |
| 1005 | /// instruction \p AccessI. |
| 1006 | bool isMemTerminator(const MemoryLocation &Loc, Instruction *AccessI, |
| 1007 | Instruction *MaybeTerm); |
| 1008 | |
| 1009 | // Returns true if \p Use may read from \p DefLoc. |
| 1010 | bool isReadClobber(const MemoryLocation &DefLoc, Instruction *UseInst); |
| 1011 | |
| 1012 | /// Returns true if a dependency between \p Current and \p KillingDef is |
| 1013 | /// guaranteed to be loop invariant for the loops that they are in. Either |
| 1014 | /// because they are known to be in the same block, in the same loop level or |
| 1015 | /// by guaranteeing that \p CurrentLoc only references a single MemoryLocation |
| 1016 | /// during execution of the containing function. |
| 1017 | bool isGuaranteedLoopIndependent(const Instruction *Current, |
| 1018 | const Instruction *KillingDef, |
| 1019 | const MemoryLocation &CurrentLoc); |
| 1020 | |
| 1021 | /// Returns true if \p Ptr is guaranteed to be loop invariant for any possible |
| 1022 | /// loop. In particular, this guarantees that it only references a single |
| 1023 | /// MemoryLocation during execution of the containing function. |
| 1024 | bool isGuaranteedLoopInvariant(const Value *Ptr); |
| 1025 | |
| 1026 | // Find a MemoryDef writing to \p KillingLoc and dominating \p StartAccess, |
| 1027 | // with no read access between them or on any other path to a function exit |
| 1028 | // block if \p KillingLoc is not accessible after the function returns. If |
| 1029 | // there is no such MemoryDef, return std::nullopt. The returned value may not |
| 1030 | // (completely) overwrite \p KillingLoc. Currently we bail out when we |
| 1031 | // encounter an aliasing MemoryUse (read). |
| 1032 | std::optional<MemoryAccess *> |
| 1033 | getDomMemoryDef(MemoryDef *KillingDef, MemoryAccess *StartAccess, |
| 1034 | const MemoryLocation &KillingLoc, const Value *KillingUndObj, |
| 1035 | unsigned &ScanLimit, unsigned &WalkerStepLimit, |
| 1036 | bool IsMemTerm, unsigned &PartialLimit, |
| 1037 | bool IsInitializesAttrMemLoc); |
| 1038 | |
| 1039 | /// Delete dead memory defs and recursively add their operands to ToRemove if |
| 1040 | /// they became dead. |
| 1041 | void |
| 1042 | deleteDeadInstruction(Instruction *SI, |
| 1043 | SmallPtrSetImpl<MemoryAccess *> *Deleted = nullptr); |
| 1044 | |
| 1045 | // Check for any extra throws between \p KillingI and \p DeadI that block |
| 1046 | // DSE. This only checks extra maythrows (those that aren't MemoryDef's). |
| 1047 | // MemoryDef that may throw are handled during the walk from one def to the |
| 1048 | // next. |
| 1049 | bool mayThrowBetween(Instruction *KillingI, Instruction *DeadI, |
| 1050 | const Value *KillingUndObj); |
| 1051 | |
| 1052 | // Check if \p DeadI acts as a DSE barrier for \p KillingI. The following |
| 1053 | // instructions act as barriers: |
| 1054 | // * A memory instruction that may throw and \p KillingI accesses a non-stack |
| 1055 | // object. |
| 1056 | // * Atomic stores stronger that monotonic. |
| 1057 | bool isDSEBarrier(const Value *KillingUndObj, Instruction *DeadI); |
| 1058 | |
| 1059 | /// Eliminate writes to objects that are not visible in the caller and are not |
| 1060 | /// accessed before returning from the function. |
| 1061 | bool eliminateDeadWritesAtEndOfFunction(); |
| 1062 | |
| 1063 | /// If we have a zero initializing memset following a call to malloc, |
| 1064 | /// try folding it into a call to calloc. |
| 1065 | bool tryFoldIntoCalloc(MemoryDef *Def, const Value *DefUO); |
| 1066 | |
| 1067 | /// \returns true if \p Def is a no-op store, either because it |
| 1068 | /// directly stores back a loaded value or stores zero to a calloced object. |
| 1069 | bool storeIsNoop(MemoryDef *Def, const Value *DefUO); |
| 1070 | |
| 1071 | bool removePartiallyOverlappedStores(InstOverlapIntervalsTy &IOL); |
| 1072 | |
| 1073 | /// Eliminates writes to locations where the value that is being written |
| 1074 | /// is already stored at the same location. |
| 1075 | bool eliminateRedundantStoresOfExistingValues(); |
| 1076 | |
| 1077 | /// If there is a dominating condition that implies the value being stored in |
| 1078 | /// a pointer, and such a condition appears in a node that dominates the |
| 1079 | /// store, then the store may be redundant if no write occurs in between. |
| 1080 | bool eliminateRedundantStoresViaDominatingConditions(); |
| 1081 | |
| 1082 | // Return the locations written by the initializes attribute. |
| 1083 | // Note that this function considers: |
| 1084 | // 1. Unwind edge: use "initializes" attribute only if the callee has |
| 1085 | // "nounwind" attribute, or the argument has "dead_on_unwind" attribute, |
| 1086 | // or the argument is invisible to caller on unwind. That is, we don't |
| 1087 | // perform incorrect DSE on unwind edges in the current function. |
| 1088 | // 2. Argument alias: for aliasing arguments, the "initializes" attribute is |
| 1089 | // the intersected range list of their "initializes" attributes. |
| 1090 | SmallVector<MemoryLocation, 1> getInitializesArgMemLoc(const Instruction *I); |
| 1091 | |
| 1092 | // Try to eliminate dead defs that access `KillingLocWrapper.MemLoc` and are |
| 1093 | // killed by `KillingLocWrapper.MemDef`. Return whether |
| 1094 | // any changes were made, and whether `KillingLocWrapper.DefInst` was deleted. |
| 1095 | std::pair<bool, bool> |
| 1096 | eliminateDeadDefs(const MemoryLocationWrapper &KillingLocWrapper); |
| 1097 | |
| 1098 | // Try to eliminate dead defs killed by `KillingDefWrapper` and return the |
| 1099 | // change state: whether make any change. |
| 1100 | bool eliminateDeadDefs(const MemoryDefWrapper &KillingDefWrapper); |
| 1101 | }; |
| 1102 | |
| 1103 | } // end anonymous namespace |
| 1104 | |
| 1105 | static void pushMemUses(MemoryAccess *Acc, |
| 1106 | SmallVectorImpl<MemoryAccess *> &WorkList, |
| 1107 | SmallPtrSetImpl<MemoryAccess *> &Visited) { |
| 1108 | for (Use &U : Acc->uses()) { |
| 1109 | auto *MA = cast<MemoryAccess>(Val: U.getUser()); |
| 1110 | if (Visited.insert(Ptr: MA).second) |
| 1111 | WorkList.push_back(Elt: MA); |
| 1112 | } |
| 1113 | } |
| 1114 | |
| 1115 | // Return true if "Arg" is function local and isn't captured before "CB". |
| 1116 | static bool isFuncLocalAndNotCaptured(Value *Arg, const CallBase *CB, |
| 1117 | EarliestEscapeAnalysis &EA) { |
| 1118 | const Value *UnderlyingObj = getUnderlyingObject(V: Arg); |
| 1119 | return isIdentifiedFunctionLocal(V: UnderlyingObj) && |
| 1120 | capturesNothing(CC: EA.getCapturesBefore(Object: UnderlyingObj, I: CB, /*OrAt=*/true, |
| 1121 | /*ReturnCaptures=*/false)); |
| 1122 | } |
| 1123 | |
| 1124 | DSEState::DSEState(const ScalarOptions &Opts, Function &F, AliasAnalysis &AA, |
| 1125 | MemorySSA &MSSA, DominatorTree &DT, PostDominatorTree &PDT, |
| 1126 | const TargetLibraryInfo &TLI, const CycleInfo &CI) |
| 1127 | : Opts(Opts), F(F), AA(AA), EA(DT, nullptr, &CI), BatchAA(AA, &EA), |
| 1128 | MSSA(MSSA), DT(DT), PDT(PDT), TLI(TLI), DL(F.getDataLayout()), CI(CI) { |
| 1129 | // Collect blocks with throwing instructions not modeled in MemorySSA and |
| 1130 | // alloc-like objects. |
| 1131 | unsigned PO = 0; |
| 1132 | for (BasicBlock *BB : post_order(G: &F)) { |
| 1133 | PostOrderNumbers[BB] = PO++; |
| 1134 | for (Instruction &I : *BB) { |
| 1135 | MemoryAccess *MA = MSSA.getMemoryAccess(I: &I); |
| 1136 | if (I.mayThrow() && !MA) |
| 1137 | ThrowingBlocks.insert(Ptr: I.getParent()); |
| 1138 | |
| 1139 | auto *MD = dyn_cast_or_null<MemoryDef>(Val: MA); |
| 1140 | if (MD && MemDefs.size() < Opts.dse_memoryssa_defs_per_block_limit && |
| 1141 | (getLocForWrite(I: &I) || isMemTerminatorInst(I: &I) || |
| 1142 | (Opts.enable_dse_initializes_attr_improvement && |
| 1143 | hasInitializesAttr(I: &I)))) |
| 1144 | MemDefs.push_back(Elt: MD); |
| 1145 | } |
| 1146 | } |
| 1147 | |
| 1148 | // Treat byval, inalloca or dead on return arguments the same as Allocas, |
| 1149 | // stores to them are dead at the end of the function. |
| 1150 | for (Argument &AI : F.args()) { |
| 1151 | if (AI.hasPassPointeeByValueCopyAttr()) { |
| 1152 | InvisibleToCallerAfterRet.insert(KV: {&AI, true}); |
| 1153 | continue; |
| 1154 | } |
| 1155 | |
| 1156 | if (!AI.getType()->isPointerTy()) |
| 1157 | continue; |
| 1158 | |
| 1159 | const DeadOnReturnInfo &Info = AI.getDeadOnReturnInfo(); |
| 1160 | if (Info.coversAllReachableMemory()) |
| 1161 | InvisibleToCallerAfterRet.insert(KV: {&AI, true}); |
| 1162 | else if (uint64_t DeadBytes = Info.getNumberOfDeadBytes()) |
| 1163 | InvisibleToCallerAfterRetBounded.insert(KV: {&AI, DeadBytes}); |
| 1164 | } |
| 1165 | |
| 1166 | AnyUnreachableExit = any_of(Range: PDT.roots(), P: [](const BasicBlock *E) { |
| 1167 | return isa<UnreachableInst>(Val: E->getTerminator()); |
| 1168 | }); |
| 1169 | } |
| 1170 | |
| 1171 | LocationSize DSEState::strengthenLocationSize(const Instruction *I, |
| 1172 | LocationSize Size) const { |
| 1173 | if (auto *CB = dyn_cast<CallBase>(Val: I)) { |
| 1174 | LibFunc F = TLI.getLibFunc(CB: *CB); |
| 1175 | if (TLI.has(F) && (F == LibFunc_memset_chk || F == LibFunc_memcpy_chk)) { |
| 1176 | // Use the precise location size specified by the 3rd argument |
| 1177 | // for determining KillingI overwrites DeadLoc if it is a memset_chk |
| 1178 | // instruction. memset_chk will write either the amount specified as 3rd |
| 1179 | // argument or the function will immediately abort and exit the program. |
| 1180 | // NOTE: AA may determine NoAlias if it can prove that the access size |
| 1181 | // is larger than the allocation size due to that being UB. To avoid |
| 1182 | // returning potentially invalid NoAlias results by AA, limit the use of |
| 1183 | // the precise location size to isOverwrite. |
| 1184 | if (const auto *Len = dyn_cast<ConstantInt>(Val: CB->getArgOperand(i: 2))) |
| 1185 | return LocationSize::precise(Value: Len->getZExtValue()); |
| 1186 | } |
| 1187 | } |
| 1188 | return Size; |
| 1189 | } |
| 1190 | |
| 1191 | OverwriteResult DSEState::isOverwrite(const Instruction *KillingI, |
| 1192 | const Instruction *DeadI, |
| 1193 | const MemoryLocation &KillingLoc, |
| 1194 | const MemoryLocation &DeadLoc, |
| 1195 | int64_t &KillingOff, int64_t &DeadOff) { |
| 1196 | // AliasAnalysis does not always account for loops. Limit overwrite checks |
| 1197 | // to dependencies for which we can guarantee they are independent of any |
| 1198 | // loops they are in. |
| 1199 | if (!isGuaranteedLoopIndependent(Current: DeadI, KillingDef: KillingI, CurrentLoc: DeadLoc)) |
| 1200 | return OW_Unknown; |
| 1201 | |
| 1202 | LocationSize KillingLocSize = |
| 1203 | strengthenLocationSize(I: KillingI, Size: KillingLoc.Size); |
| 1204 | const Value *DeadPtr = DeadLoc.Ptr->stripPointerCasts(); |
| 1205 | const Value *KillingPtr = KillingLoc.Ptr->stripPointerCasts(); |
| 1206 | const Value *DeadUndObj = getUnderlyingObject(V: DeadPtr); |
| 1207 | const Value *KillingUndObj = getUnderlyingObject(V: KillingPtr); |
| 1208 | |
| 1209 | // Check whether the killing store overwrites the whole object, in which |
| 1210 | // case the size/offset of the dead store does not matter. |
| 1211 | if (DeadUndObj == KillingUndObj && KillingLocSize.isPrecise() && |
| 1212 | isIdentifiedObject(V: KillingUndObj)) { |
| 1213 | std::optional<TypeSize> KillingUndObjSize = |
| 1214 | getPointerSize(V: KillingUndObj, DL, TLI, F: &F); |
| 1215 | if (KillingUndObjSize && *KillingUndObjSize == KillingLocSize.getValue()) |
| 1216 | return OW_Complete; |
| 1217 | } |
| 1218 | |
| 1219 | // FIXME: Vet that this works for size upper-bounds. Seems unlikely that we'll |
| 1220 | // get imprecise values here, though (except for unknown sizes). |
| 1221 | if (!KillingLocSize.isPrecise() || !DeadLoc.Size.isPrecise()) { |
| 1222 | // In case no constant size is known, try to an IR values for the number |
| 1223 | // of bytes written and check if they match. |
| 1224 | const auto *KillingMemI = dyn_cast<MemIntrinsic>(Val: KillingI); |
| 1225 | const auto *DeadMemI = dyn_cast<MemIntrinsic>(Val: DeadI); |
| 1226 | if (KillingMemI && DeadMemI) { |
| 1227 | const Value *KillingV = KillingMemI->getLength(); |
| 1228 | const Value *DeadV = DeadMemI->getLength(); |
| 1229 | if (KillingV == DeadV && BatchAA.isMustAlias(LocA: DeadLoc, LocB: KillingLoc)) |
| 1230 | return OW_Complete; |
| 1231 | } |
| 1232 | |
| 1233 | // Masked stores have imprecise locations, but we can reason about them |
| 1234 | // to some extent. |
| 1235 | return isMaskedStoreOverwrite(KillingI, DeadI, AA&: BatchAA); |
| 1236 | } |
| 1237 | |
| 1238 | const TypeSize KillingSize = KillingLocSize.getValue(); |
| 1239 | const TypeSize DeadSize = DeadLoc.Size.getValue(); |
| 1240 | // Bail on doing Size comparison which depends on AA for now |
| 1241 | // TODO: Remove AnyScalable once Alias Analysis deal with scalable vectors |
| 1242 | const bool AnyScalable = DeadSize.isScalable() || KillingLocSize.isScalable(); |
| 1243 | |
| 1244 | if (AnyScalable) |
| 1245 | return OW_Unknown; |
| 1246 | // Query the alias information |
| 1247 | AliasResult AAR = BatchAA.alias(LocA: KillingLoc, LocB: DeadLoc); |
| 1248 | |
| 1249 | // If the start pointers are the same, we just have to compare sizes to see if |
| 1250 | // the killing store was larger than the dead store. |
| 1251 | if (AAR == AliasResult::MustAlias) { |
| 1252 | // Make sure that the KillingSize size is >= the DeadSize size. |
| 1253 | if (KillingSize >= DeadSize) |
| 1254 | return OW_Complete; |
| 1255 | } |
| 1256 | |
| 1257 | // If we hit a partial alias we may have a full overwrite |
| 1258 | if (AAR == AliasResult::PartialAlias && AAR.hasOffset()) { |
| 1259 | int32_t Off = AAR.getOffset(); |
| 1260 | if (Off >= 0 && (uint64_t)Off + DeadSize <= KillingSize) |
| 1261 | return OW_Complete; |
| 1262 | } |
| 1263 | |
| 1264 | // If we can't resolve the same pointers to the same object, then we can't |
| 1265 | // analyze them at all. |
| 1266 | if (DeadUndObj != KillingUndObj) { |
| 1267 | // Non aliasing stores to different objects don't overlap. Note that |
| 1268 | // if the killing store is known to overwrite whole object (out of |
| 1269 | // bounds access overwrites whole object as well) then it is assumed to |
| 1270 | // completely overwrite any store to the same object even if they don't |
| 1271 | // actually alias (see next check). |
| 1272 | if (AAR == AliasResult::NoAlias) |
| 1273 | return OW_None; |
| 1274 | return OW_Unknown; |
| 1275 | } |
| 1276 | |
| 1277 | // Okay, we have stores to two completely different pointers. Try to |
| 1278 | // decompose the pointer into a "base + constant_offset" form. If the base |
| 1279 | // pointers are equal, then we can reason about the two stores. |
| 1280 | DeadOff = 0; |
| 1281 | KillingOff = 0; |
| 1282 | const Value *DeadBasePtr = |
| 1283 | GetPointerBaseWithConstantOffset(Ptr: DeadPtr, Offset&: DeadOff, DL); |
| 1284 | const Value *KillingBasePtr = |
| 1285 | GetPointerBaseWithConstantOffset(Ptr: KillingPtr, Offset&: KillingOff, DL); |
| 1286 | |
| 1287 | // If the base pointers still differ, we have two completely different |
| 1288 | // stores. |
| 1289 | if (DeadBasePtr != KillingBasePtr) |
| 1290 | return OW_Unknown; |
| 1291 | |
| 1292 | // The killing access completely overlaps the dead store if and only if |
| 1293 | // both start and end of the dead one is "inside" the killing one: |
| 1294 | // |<->|--dead--|<->| |
| 1295 | // |-----killing------| |
| 1296 | // Accesses may overlap if and only if start of one of them is "inside" |
| 1297 | // another one: |
| 1298 | // |<->|--dead--|<-------->| |
| 1299 | // |-------killing--------| |
| 1300 | // OR |
| 1301 | // |-------dead-------| |
| 1302 | // |<->|---killing---|<----->| |
| 1303 | // |
| 1304 | // We have to be careful here as *Off is signed while *.Size is unsigned. |
| 1305 | |
| 1306 | // Check if the dead access starts "not before" the killing one. |
| 1307 | if (DeadOff >= KillingOff) { |
| 1308 | // If the dead access ends "not after" the killing access then the |
| 1309 | // dead one is completely overwritten by the killing one. |
| 1310 | if (uint64_t(DeadOff - KillingOff) + DeadSize <= KillingSize) |
| 1311 | return OW_Complete; |
| 1312 | // If start of the dead access is "before" end of the killing access |
| 1313 | // then accesses overlap. |
| 1314 | else if ((uint64_t)(DeadOff - KillingOff) < KillingSize) |
| 1315 | return OW_MaybePartial; |
| 1316 | } |
| 1317 | // If start of the killing access is "before" end of the dead access then |
| 1318 | // accesses overlap. |
| 1319 | else if ((uint64_t)(KillingOff - DeadOff) < DeadSize) { |
| 1320 | return OW_MaybePartial; |
| 1321 | } |
| 1322 | |
| 1323 | // Can reach here only if accesses are known not to overlap. |
| 1324 | return OW_None; |
| 1325 | } |
| 1326 | |
| 1327 | bool DSEState::isInvisibleToCallerAfterRet(const Value *V, const Value *Ptr, |
| 1328 | const LocationSize StoreSize) { |
| 1329 | if (isa<AllocaInst>(Val: V)) |
| 1330 | return true; |
| 1331 | |
| 1332 | auto IBounded = InvisibleToCallerAfterRetBounded.find(Val: V); |
| 1333 | if (IBounded != InvisibleToCallerAfterRetBounded.end()) { |
| 1334 | int64_t ValueOffset; |
| 1335 | [[maybe_unused]] const Value *BaseValue = |
| 1336 | GetPointerBaseWithConstantOffset(Ptr, Offset&: ValueOffset, DL); |
| 1337 | // If we are not able to find a constant offset from the UO, we have to |
| 1338 | // pessimistically assume that the store writes to memory out of the |
| 1339 | // dead_on_return bounds. |
| 1340 | if (BaseValue != V) |
| 1341 | return false; |
| 1342 | // This store is only invisible after return if we are in bounds of the |
| 1343 | // range marked dead. |
| 1344 | if (StoreSize.hasValue() && |
| 1345 | ValueOffset + StoreSize.getValue() <= IBounded->second && |
| 1346 | ValueOffset >= 0) |
| 1347 | return true; |
| 1348 | } |
| 1349 | auto I = InvisibleToCallerAfterRet.insert(KV: {V, false}); |
| 1350 | if (I.second && isInvisibleToCallerOnUnwind(V) && isNoAliasCall(V)) |
| 1351 | I.first->second = capturesNothing( |
| 1352 | CC: PointerMayBeCaptured(V, Mask: CaptureComponents::Provenance).WithRet); |
| 1353 | return I.first->second; |
| 1354 | } |
| 1355 | |
| 1356 | bool DSEState::isInvisibleToCallerOnUnwind(const Value *V) { |
| 1357 | bool RequiresNoCaptureBeforeUnwind; |
| 1358 | if (!isNotVisibleOnUnwind(Object: V, RequiresNoCaptureBeforeUnwind)) |
| 1359 | return false; |
| 1360 | if (!RequiresNoCaptureBeforeUnwind) |
| 1361 | return true; |
| 1362 | |
| 1363 | auto I = CapturedBeforeReturn.insert(KV: {V, true}); |
| 1364 | if (I.second) |
| 1365 | // NOTE: This could be made more precise by PointerMayBeCapturedBefore |
| 1366 | // with the killing MemoryDef. But we refrain from doing so for now to |
| 1367 | // limit compile-time and this does not cause any changes to the number |
| 1368 | // of stores removed on a large test set in practice. |
| 1369 | I.first->second = capturesAnything( |
| 1370 | CC: PointerMayBeCaptured(V, Mask: CaptureComponents::Provenance).WithoutRet); |
| 1371 | return !I.first->second; |
| 1372 | } |
| 1373 | |
| 1374 | std::optional<MemoryLocation> DSEState::getLocForWrite(Instruction *I) const { |
| 1375 | if (!I->mayWriteToMemory()) |
| 1376 | return std::nullopt; |
| 1377 | |
| 1378 | if (auto *CB = dyn_cast<CallBase>(Val: I)) |
| 1379 | return MemoryLocation::getForDest(CI: CB, TLI); |
| 1380 | |
| 1381 | return MemoryLocation::getOrNone(Inst: I); |
| 1382 | } |
| 1383 | |
| 1384 | SmallVector<std::pair<MemoryLocation, bool>, 1> |
| 1385 | DSEState::getLocForInst(Instruction *I, bool ConsiderInitializesAttr) { |
| 1386 | SmallVector<std::pair<MemoryLocation, bool>, 1> Locations; |
| 1387 | if (isMemTerminatorInst(I)) { |
| 1388 | if (auto Loc = getLocForTerminator(I)) |
| 1389 | Locations.push_back(Elt: std::make_pair(x&: Loc->first, y: false)); |
| 1390 | return Locations; |
| 1391 | } |
| 1392 | |
| 1393 | if (auto Loc = getLocForWrite(I)) |
| 1394 | Locations.push_back(Elt: std::make_pair(x&: *Loc, y: false)); |
| 1395 | |
| 1396 | if (ConsiderInitializesAttr) { |
| 1397 | for (auto &MemLoc : getInitializesArgMemLoc(I)) { |
| 1398 | Locations.push_back(Elt: std::make_pair(x&: MemLoc, y: true)); |
| 1399 | } |
| 1400 | } |
| 1401 | return Locations; |
| 1402 | } |
| 1403 | |
| 1404 | bool DSEState::isRemovable(Instruction *I) { |
| 1405 | assert(getLocForWrite(I) && "Must have analyzable write"); |
| 1406 | |
| 1407 | // Don't remove volatile/atomic stores. |
| 1408 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) |
| 1409 | return SI->isUnordered(); |
| 1410 | |
| 1411 | if (auto *CB = dyn_cast<CallBase>(Val: I)) { |
| 1412 | // Don't remove volatile memory intrinsics. |
| 1413 | if (auto *MI = dyn_cast<MemIntrinsic>(Val: CB)) |
| 1414 | return !MI->isVolatile(); |
| 1415 | |
| 1416 | // Never remove dead lifetime intrinsics, e.g. because they are followed |
| 1417 | // by a free. |
| 1418 | if (CB->isLifetimeStartOrEnd()) |
| 1419 | return false; |
| 1420 | |
| 1421 | return CB->use_empty() && CB->willReturn() && CB->doesNotThrow() && |
| 1422 | !CB->isTerminator(); |
| 1423 | } |
| 1424 | |
| 1425 | return false; |
| 1426 | } |
| 1427 | |
| 1428 | bool DSEState::isCompleteOverwrite(const MemoryLocation &DefLoc, |
| 1429 | Instruction *DefInst, Instruction *UseInst) { |
| 1430 | // UseInst has a MemoryDef associated in MemorySSA. It's possible for a |
| 1431 | // MemoryDef to not write to memory, e.g. a volatile load is modeled as a |
| 1432 | // MemoryDef. |
| 1433 | if (!UseInst->mayWriteToMemory()) |
| 1434 | return false; |
| 1435 | |
| 1436 | if (auto *CB = dyn_cast<CallBase>(Val: UseInst)) |
| 1437 | if (CB->onlyAccessesInaccessibleMemory()) |
| 1438 | return false; |
| 1439 | |
| 1440 | int64_t InstWriteOffset, DepWriteOffset; |
| 1441 | if (auto CC = getLocForWrite(I: UseInst)) |
| 1442 | return isOverwrite(KillingI: UseInst, DeadI: DefInst, KillingLoc: *CC, DeadLoc: DefLoc, KillingOff&: InstWriteOffset, |
| 1443 | DeadOff&: DepWriteOffset) == OW_Complete; |
| 1444 | return false; |
| 1445 | } |
| 1446 | |
| 1447 | bool DSEState::isWriteAtEndOfFunction(MemoryDef *Def, |
| 1448 | const MemoryLocation &DefLoc) { |
| 1449 | LLVM_DEBUG(dbgs() << " Check if def "<< *Def << " (" |
| 1450 | << *Def->getMemoryInst() |
| 1451 | << ") is at the end the function \n"); |
| 1452 | SmallVector<MemoryAccess *, 4> WorkList; |
| 1453 | SmallPtrSet<MemoryAccess *, 8> Visited; |
| 1454 | |
| 1455 | pushMemUses(Acc: Def, WorkList, Visited); |
| 1456 | for (unsigned I = 0; I < WorkList.size(); I++) { |
| 1457 | if (WorkList.size() >= Opts.dse_memoryssa_scanlimit) { |
| 1458 | LLVM_DEBUG(dbgs() << " ... hit exploration limit.\n"); |
| 1459 | return false; |
| 1460 | } |
| 1461 | |
| 1462 | MemoryAccess *UseAccess = WorkList[I]; |
| 1463 | if (isa<MemoryPhi>(Val: UseAccess)) { |
| 1464 | // AliasAnalysis does not account for loops. Limit elimination to |
| 1465 | // candidates for which we can guarantee they always store to the same |
| 1466 | // memory location. |
| 1467 | if (!isGuaranteedLoopInvariant(Ptr: DefLoc.Ptr)) |
| 1468 | return false; |
| 1469 | |
| 1470 | pushMemUses(Acc: cast<MemoryPhi>(Val: UseAccess), WorkList, Visited); |
| 1471 | continue; |
| 1472 | } |
| 1473 | // TODO: Checking for aliasing is expensive. Consider reducing the amount |
| 1474 | // of times this is called and/or caching it. |
| 1475 | Instruction *UseInst = cast<MemoryUseOrDef>(Val: UseAccess)->getMemoryInst(); |
| 1476 | if (isReadClobber(DefLoc, UseInst)) { |
| 1477 | LLVM_DEBUG(dbgs() << " ... hit read clobber "<< *UseInst << ".\n"); |
| 1478 | return false; |
| 1479 | } |
| 1480 | |
| 1481 | if (MemoryDef *UseDef = dyn_cast<MemoryDef>(Val: UseAccess)) |
| 1482 | pushMemUses(Acc: UseDef, WorkList, Visited); |
| 1483 | } |
| 1484 | return true; |
| 1485 | } |
| 1486 | |
| 1487 | std::optional<std::pair<MemoryLocation, bool>> |
| 1488 | DSEState::getLocForTerminator(Instruction *I) const { |
| 1489 | if (auto *CB = dyn_cast<CallBase>(Val: I)) { |
| 1490 | if (CB->getIntrinsicID() == Intrinsic::lifetime_end) |
| 1491 | return { |
| 1492 | std::make_pair(x: MemoryLocation::getForArgument(Call: CB, ArgIdx: 0, TLI: &TLI), y: false)}; |
| 1493 | if (Value *FreedOp = getFreedOperand(CB, TLI: &TLI)) |
| 1494 | return {std::make_pair(x: MemoryLocation::getAfter(Ptr: FreedOp), y: true)}; |
| 1495 | } |
| 1496 | |
| 1497 | return std::nullopt; |
| 1498 | } |
| 1499 | |
| 1500 | bool DSEState::isMemTerminatorInst(Instruction *I) const { |
| 1501 | auto *CB = dyn_cast<CallBase>(Val: I); |
| 1502 | return CB && (CB->getIntrinsicID() == Intrinsic::lifetime_end || |
| 1503 | getFreedOperand(CB, TLI: &TLI) != nullptr); |
| 1504 | } |
| 1505 | |
| 1506 | bool DSEState::isMemTerminator(const MemoryLocation &Loc, Instruction *AccessI, |
| 1507 | Instruction *MaybeTerm) { |
| 1508 | std::optional<std::pair<MemoryLocation, bool>> MaybeTermLoc = |
| 1509 | getLocForTerminator(I: MaybeTerm); |
| 1510 | |
| 1511 | if (!MaybeTermLoc) |
| 1512 | return false; |
| 1513 | |
| 1514 | // If the terminator is a free-like call, all accesses to the underlying |
| 1515 | // object can be considered terminated. |
| 1516 | if (getUnderlyingObject(V: Loc.Ptr) != |
| 1517 | getUnderlyingObject(V: MaybeTermLoc->first.Ptr)) |
| 1518 | return false; |
| 1519 | |
| 1520 | auto TermLoc = MaybeTermLoc->first; |
| 1521 | if (MaybeTermLoc->second) { |
| 1522 | const Value *LocUO = getUnderlyingObject(V: Loc.Ptr); |
| 1523 | return BatchAA.isMustAlias(V1: TermLoc.Ptr, V2: LocUO); |
| 1524 | } |
| 1525 | int64_t InstWriteOffset = 0; |
| 1526 | int64_t DepWriteOffset = 0; |
| 1527 | return isOverwrite(KillingI: MaybeTerm, DeadI: AccessI, KillingLoc: TermLoc, DeadLoc: Loc, KillingOff&: InstWriteOffset, |
| 1528 | DeadOff&: DepWriteOffset) == OW_Complete; |
| 1529 | } |
| 1530 | |
| 1531 | bool DSEState::isReadClobber(const MemoryLocation &DefLoc, |
| 1532 | Instruction *UseInst) { |
| 1533 | if (isNoopIntrinsic(I: UseInst)) |
| 1534 | return false; |
| 1535 | |
| 1536 | // Monotonic or weaker atomic stores can be re-ordered and do not need to be |
| 1537 | // treated as read clobber. |
| 1538 | if (auto SI = dyn_cast<StoreInst>(Val: UseInst)) |
| 1539 | return isStrongerThan(AO: SI->getOrdering(), Other: AtomicOrdering::Monotonic); |
| 1540 | |
| 1541 | if (!UseInst->mayReadFromMemory()) |
| 1542 | return false; |
| 1543 | |
| 1544 | if (auto *CB = dyn_cast<CallBase>(Val: UseInst)) |
| 1545 | if (CB->onlyAccessesInaccessibleMemory()) |
| 1546 | return false; |
| 1547 | |
| 1548 | return isRefSet(MRI: BatchAA.getModRefInfo(I: UseInst, OptLoc: DefLoc)); |
| 1549 | } |
| 1550 | |
| 1551 | bool DSEState::isGuaranteedLoopIndependent(const Instruction *Current, |
| 1552 | const Instruction *KillingDef, |
| 1553 | const MemoryLocation &CurrentLoc) { |
| 1554 | // If the dependency is within the same block or loop level (being careful |
| 1555 | // of irreducible loops), we know that AA will return a valid result for the |
| 1556 | // memory dependency. (Both at the function level, outside of any loop, |
| 1557 | // would also be valid but we currently disable that to limit compile time). |
| 1558 | if (Current->getParent() == KillingDef->getParent()) |
| 1559 | return true; |
| 1560 | CycleRef CurrentC = CI.getCycle(Block: Current->getParent()); |
| 1561 | if (CurrentC && CurrentC == CI.getCycle(Block: KillingDef->getParent())) |
| 1562 | return true; |
| 1563 | // Otherwise check the memory location is invariant to any loops. |
| 1564 | return isGuaranteedLoopInvariant(Ptr: CurrentLoc.Ptr); |
| 1565 | } |
| 1566 | |
| 1567 | bool DSEState::isGuaranteedLoopInvariant(const Value *Ptr) { |
| 1568 | Ptr = Ptr->stripPointerCasts(); |
| 1569 | if (auto *GEP = dyn_cast<GEPOperator>(Val: Ptr)) |
| 1570 | if (GEP->hasAllConstantIndices()) |
| 1571 | Ptr = GEP->getPointerOperand()->stripPointerCasts(); |
| 1572 | |
| 1573 | if (auto *I = dyn_cast<Instruction>(Val: Ptr)) { |
| 1574 | return I->getParent()->isEntryBlock() || !CI.getCycle(Block: I->getParent()); |
| 1575 | } |
| 1576 | return true; |
| 1577 | } |
| 1578 | |
| 1579 | std::optional<MemoryAccess *> DSEState::getDomMemoryDef( |
| 1580 | MemoryDef *KillingDef, MemoryAccess *StartAccess, |
| 1581 | const MemoryLocation &KillingLoc, const Value *KillingUndObj, |
| 1582 | unsigned &ScanLimit, unsigned &WalkerStepLimit, bool IsMemTerm, |
| 1583 | unsigned &PartialLimit, bool IsInitializesAttrMemLoc) { |
| 1584 | if (ScanLimit == 0 || WalkerStepLimit == 0) { |
| 1585 | LLVM_DEBUG(dbgs() << "\n ... hit scan limit\n"); |
| 1586 | return std::nullopt; |
| 1587 | } |
| 1588 | |
| 1589 | MemoryAccess *Current = StartAccess; |
| 1590 | Instruction *KillingI = KillingDef->getMemoryInst(); |
| 1591 | LLVM_DEBUG(dbgs() << " trying to get dominating access\n"); |
| 1592 | |
| 1593 | // Only optimize defining access of KillingDef when directly starting at its |
| 1594 | // defining access. The defining access also must only access KillingLoc. At |
| 1595 | // the moment we only support instructions with a single write location, so |
| 1596 | // it should be sufficient to disable optimizations for instructions that |
| 1597 | // also read from memory. |
| 1598 | bool CanOptimize = Opts.dse_optimize_memoryssa && |
| 1599 | KillingDef->getDefiningAccess() == StartAccess && |
| 1600 | !KillingI->mayReadFromMemory(); |
| 1601 | |
| 1602 | // Find the next clobbering Mod access for DefLoc, starting at StartAccess. |
| 1603 | std::optional<MemoryLocation> CurrentLoc; |
| 1604 | for (;; Current = cast<MemoryDef>(Val: Current)->getDefiningAccess()) { |
| 1605 | LLVM_DEBUG({ |
| 1606 | dbgs() << " visiting "<< *Current; |
| 1607 | if (!MSSA.isLiveOnEntryDef(Current) && isa<MemoryUseOrDef>(Current)) |
| 1608 | dbgs() << " ("<< *cast<MemoryUseOrDef>(Current)->getMemoryInst() |
| 1609 | << ")"; |
| 1610 | dbgs() << "\n"; |
| 1611 | }); |
| 1612 | |
| 1613 | // Reached TOP. |
| 1614 | if (MSSA.isLiveOnEntryDef(MA: Current)) { |
| 1615 | LLVM_DEBUG(dbgs() << " ... found LiveOnEntryDef\n"); |
| 1616 | if (CanOptimize && Current != KillingDef->getDefiningAccess()) |
| 1617 | // The first clobbering def is... none. |
| 1618 | KillingDef->setOptimized(Current); |
| 1619 | return std::nullopt; |
| 1620 | } |
| 1621 | |
| 1622 | // Cost of a step. Accesses in the same block are more likely to be valid |
| 1623 | // candidates for elimination, hence consider them cheaper. |
| 1624 | unsigned StepCost = KillingDef->getBlock() == Current->getBlock() |
| 1625 | ? Opts.dse_memoryssa_samebb_cost |
| 1626 | : Opts.dse_memoryssa_otherbb_cost; |
| 1627 | if (WalkerStepLimit <= StepCost) { |
| 1628 | LLVM_DEBUG(dbgs() << " ... hit walker step limit\n"); |
| 1629 | return std::nullopt; |
| 1630 | } |
| 1631 | WalkerStepLimit -= StepCost; |
| 1632 | |
| 1633 | // Return for MemoryPhis. They cannot be eliminated directly and the |
| 1634 | // caller is responsible for traversing them. |
| 1635 | if (isa<MemoryPhi>(Val: Current)) { |
| 1636 | LLVM_DEBUG(dbgs() << " ... found MemoryPhi\n"); |
| 1637 | return Current; |
| 1638 | } |
| 1639 | |
| 1640 | // Below, check if CurrentDef is a valid candidate to be eliminated by |
| 1641 | // KillingDef. If it is not, check the next candidate. |
| 1642 | MemoryDef *CurrentDef = cast<MemoryDef>(Val: Current); |
| 1643 | Instruction *CurrentI = CurrentDef->getMemoryInst(); |
| 1644 | |
| 1645 | if (canSkipDef(D: CurrentDef, DefVisibleToCaller: !isInvisibleToCallerOnUnwind(V: KillingUndObj))) { |
| 1646 | CanOptimize = false; |
| 1647 | continue; |
| 1648 | } |
| 1649 | |
| 1650 | // Before we try to remove anything, check for any extra throwing |
| 1651 | // instructions that block us from DSEing |
| 1652 | if (mayThrowBetween(KillingI, DeadI: CurrentI, KillingUndObj)) { |
| 1653 | LLVM_DEBUG(dbgs() << " ... skip, may throw!\n"); |
| 1654 | return std::nullopt; |
| 1655 | } |
| 1656 | |
| 1657 | // Check for anything that looks like it will be a barrier to further |
| 1658 | // removal |
| 1659 | if (isDSEBarrier(KillingUndObj, DeadI: CurrentI)) { |
| 1660 | LLVM_DEBUG(dbgs() << " ... skip, barrier\n"); |
| 1661 | return std::nullopt; |
| 1662 | } |
| 1663 | |
| 1664 | // If Current is known to be on path that reads DefLoc or is a read |
| 1665 | // clobber, bail out, as the path is not profitable. We skip this check |
| 1666 | // for intrinsic calls, because the code knows how to handle memcpy |
| 1667 | // intrinsics. |
| 1668 | if (!isa<IntrinsicInst>(Val: CurrentI) && isReadClobber(DefLoc: KillingLoc, UseInst: CurrentI)) |
| 1669 | return std::nullopt; |
| 1670 | |
| 1671 | // Quick check if there are direct uses that are read-clobbers. |
| 1672 | if (any_of(Range: Current->uses(), P: [this, &KillingLoc, StartAccess](Use &U) { |
| 1673 | if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(Val: U.getUser())) |
| 1674 | return !MSSA.dominates(A: StartAccess, B: UseOrDef) && |
| 1675 | isReadClobber(DefLoc: KillingLoc, UseInst: UseOrDef->getMemoryInst()); |
| 1676 | return false; |
| 1677 | })) { |
| 1678 | LLVM_DEBUG(dbgs() << " ... found a read clobber\n"); |
| 1679 | return std::nullopt; |
| 1680 | } |
| 1681 | |
| 1682 | // If Current does not have an analyzable write location or is not |
| 1683 | // removable, skip it. |
| 1684 | CurrentLoc = getLocForWrite(I: CurrentI); |
| 1685 | if (!CurrentLoc || !isRemovable(I: CurrentI)) { |
| 1686 | CanOptimize = false; |
| 1687 | continue; |
| 1688 | } |
| 1689 | |
| 1690 | // AliasAnalysis does not account for loops. Limit elimination to |
| 1691 | // candidates for which we can guarantee they always store to the same |
| 1692 | // memory location and not located in different loops. |
| 1693 | if (!isGuaranteedLoopIndependent(Current: CurrentI, KillingDef: KillingI, CurrentLoc: *CurrentLoc)) { |
| 1694 | LLVM_DEBUG(dbgs() << " ... not guaranteed loop independent\n"); |
| 1695 | CanOptimize = false; |
| 1696 | continue; |
| 1697 | } |
| 1698 | |
| 1699 | if (IsMemTerm) { |
| 1700 | // If the killing def is a memory terminator (e.g. lifetime.end), check |
| 1701 | // the next candidate if the current Current does not write the same |
| 1702 | // underlying object as the terminator. |
| 1703 | if (!isMemTerminator(Loc: *CurrentLoc, AccessI: CurrentI, MaybeTerm: KillingI)) { |
| 1704 | CanOptimize = false; |
| 1705 | continue; |
| 1706 | } |
| 1707 | } else { |
| 1708 | int64_t KillingOffset = 0; |
| 1709 | int64_t DeadOffset = 0; |
| 1710 | auto OR = isOverwrite(KillingI, DeadI: CurrentI, KillingLoc, DeadLoc: *CurrentLoc, |
| 1711 | KillingOff&: KillingOffset, DeadOff&: DeadOffset); |
| 1712 | if (CanOptimize) { |
| 1713 | // CurrentDef is the earliest write clobber of KillingDef. Use it as |
| 1714 | // optimized access. Do not optimize if CurrentDef is already the |
| 1715 | // defining access of KillingDef. |
| 1716 | if (CurrentDef != KillingDef->getDefiningAccess() && |
| 1717 | (OR == OW_Complete || OR == OW_MaybePartial)) |
| 1718 | KillingDef->setOptimized(CurrentDef); |
| 1719 | |
| 1720 | // Once a may-aliasing def is encountered do not set an optimized |
| 1721 | // access. |
| 1722 | if (OR != OW_None) |
| 1723 | CanOptimize = false; |
| 1724 | } |
| 1725 | |
| 1726 | // If Current does not write to the same object as KillingDef, check |
| 1727 | // the next candidate. |
| 1728 | if (OR == OW_Unknown || OR == OW_None) |
| 1729 | continue; |
| 1730 | else if (OR == OW_MaybePartial) { |
| 1731 | // If KillingDef only partially overwrites Current, check the next |
| 1732 | // candidate if the partial step limit is exceeded. This aggressively |
| 1733 | // limits the number of candidates for partial store elimination, |
| 1734 | // which are less likely to be removable in the end. |
| 1735 | if (PartialLimit <= 1) { |
| 1736 | WalkerStepLimit -= 1; |
| 1737 | LLVM_DEBUG(dbgs() << " ... reached partial limit ... continue with " |
| 1738 | "next access\n"); |
| 1739 | continue; |
| 1740 | } |
| 1741 | PartialLimit -= 1; |
| 1742 | } |
| 1743 | } |
| 1744 | break; |
| 1745 | }; |
| 1746 | |
| 1747 | // Accesses to objects accessible after the function returns can only be |
| 1748 | // eliminated if the access is dead along all paths to the exit. Collect |
| 1749 | // the blocks with killing (=completely overwriting MemoryDefs) and check if |
| 1750 | // they cover all paths from MaybeDeadAccess to any function exit. |
| 1751 | SmallPtrSet<Instruction *, 16> KillingDefs; |
| 1752 | KillingDefs.insert(Ptr: KillingDef->getMemoryInst()); |
| 1753 | MemoryAccess *MaybeDeadAccess = Current; |
| 1754 | MemoryLocation MaybeDeadLoc = *CurrentLoc; |
| 1755 | Instruction *MaybeDeadI = cast<MemoryDef>(Val: MaybeDeadAccess)->getMemoryInst(); |
| 1756 | LLVM_DEBUG(dbgs() << " Checking for reads of "<< *MaybeDeadAccess << " (" |
| 1757 | << *MaybeDeadI << ")\n"); |
| 1758 | |
| 1759 | SmallVector<MemoryAccess *, 32> WorkList; |
| 1760 | SmallPtrSet<MemoryAccess *, 32> Visited; |
| 1761 | pushMemUses(Acc: MaybeDeadAccess, WorkList, Visited); |
| 1762 | |
| 1763 | // Check if DeadDef may be read. |
| 1764 | for (unsigned I = 0; I < WorkList.size(); I++) { |
| 1765 | MemoryAccess *UseAccess = WorkList[I]; |
| 1766 | |
| 1767 | LLVM_DEBUG(dbgs() << " "<< *UseAccess); |
| 1768 | // Bail out if the number of accesses to check exceeds the scan limit. |
| 1769 | if (ScanLimit < (WorkList.size() - I)) { |
| 1770 | LLVM_DEBUG(dbgs() << "\n ... hit scan limit\n"); |
| 1771 | return std::nullopt; |
| 1772 | } |
| 1773 | --ScanLimit; |
| 1774 | NumDomMemDefChecks++; |
| 1775 | |
| 1776 | if (isa<MemoryPhi>(Val: UseAccess)) { |
| 1777 | if (any_of(Range&: KillingDefs, P: [this, UseAccess](Instruction *KI) { |
| 1778 | return DT.properlyDominates(A: KI->getParent(), B: UseAccess->getBlock()); |
| 1779 | })) { |
| 1780 | LLVM_DEBUG(dbgs() << " ... skipping, dominated by killing block\n"); |
| 1781 | continue; |
| 1782 | } |
| 1783 | LLVM_DEBUG(dbgs() << "\n ... adding PHI uses\n"); |
| 1784 | pushMemUses(Acc: UseAccess, WorkList, Visited); |
| 1785 | continue; |
| 1786 | } |
| 1787 | |
| 1788 | Instruction *UseInst = cast<MemoryUseOrDef>(Val: UseAccess)->getMemoryInst(); |
| 1789 | LLVM_DEBUG(dbgs() << " ("<< *UseInst << ")\n"); |
| 1790 | |
| 1791 | if (any_of(Range&: KillingDefs, P: [this, UseInst](Instruction *KI) { |
| 1792 | return DT.dominates(Def: KI, User: UseInst); |
| 1793 | })) { |
| 1794 | LLVM_DEBUG(dbgs() << " ... skipping, dominated by killing def\n"); |
| 1795 | continue; |
| 1796 | } |
| 1797 | |
| 1798 | // A memory terminator kills all preceeding MemoryDefs and all succeeding |
| 1799 | // MemoryAccesses. We do not have to check it's users. |
| 1800 | if (isMemTerminator(Loc: MaybeDeadLoc, AccessI: MaybeDeadI, MaybeTerm: UseInst)) { |
| 1801 | LLVM_DEBUG( |
| 1802 | dbgs() |
| 1803 | << " ... skipping, memterminator invalidates following accesses\n"); |
| 1804 | continue; |
| 1805 | } |
| 1806 | |
| 1807 | if (isNoopIntrinsic(I: cast<MemoryUseOrDef>(Val: UseAccess)->getMemoryInst())) { |
| 1808 | LLVM_DEBUG(dbgs() << " ... adding uses of intrinsic\n"); |
| 1809 | pushMemUses(Acc: UseAccess, WorkList, Visited); |
| 1810 | continue; |
| 1811 | } |
| 1812 | |
| 1813 | if (UseInst->mayThrow() && !isInvisibleToCallerOnUnwind(V: KillingUndObj)) { |
| 1814 | LLVM_DEBUG(dbgs() << " ... found throwing instruction\n"); |
| 1815 | return std::nullopt; |
| 1816 | } |
| 1817 | |
| 1818 | // Uses which may read the original MemoryDef mean we cannot eliminate the |
| 1819 | // original MD. Stop walk. |
| 1820 | // If KillingDef is a CallInst with "initializes" attribute, the reads in |
| 1821 | // the callee would be dominated by initializations, so it should be safe. |
| 1822 | bool IsKillingDefFromInitAttr = false; |
| 1823 | if (IsInitializesAttrMemLoc) { |
| 1824 | if (KillingI == UseInst && |
| 1825 | KillingUndObj == getUnderlyingObject(V: MaybeDeadLoc.Ptr)) |
| 1826 | IsKillingDefFromInitAttr = true; |
| 1827 | } |
| 1828 | |
| 1829 | if (isReadClobber(DefLoc: MaybeDeadLoc, UseInst) && !IsKillingDefFromInitAttr) { |
| 1830 | LLVM_DEBUG(dbgs() << " ... found read clobber\n"); |
| 1831 | return std::nullopt; |
| 1832 | } |
| 1833 | |
| 1834 | // If this worklist walks back to the original memory access (and the |
| 1835 | // pointer is not guarenteed loop invariant) then we cannot assume that a |
| 1836 | // store kills itself. |
| 1837 | if (MaybeDeadAccess == UseAccess && |
| 1838 | !isGuaranteedLoopInvariant(Ptr: MaybeDeadLoc.Ptr)) { |
| 1839 | LLVM_DEBUG(dbgs() << " ... found not loop invariant self access\n"); |
| 1840 | return std::nullopt; |
| 1841 | } |
| 1842 | // Otherwise, for the KillingDef and MaybeDeadAccess we only have to check |
| 1843 | // if it reads the memory location. |
| 1844 | // TODO: It would probably be better to check for self-reads before |
| 1845 | // calling the function. |
| 1846 | if (KillingDef == UseAccess || MaybeDeadAccess == UseAccess) { |
| 1847 | LLVM_DEBUG(dbgs() << " ... skipping killing def/dom access\n"); |
| 1848 | continue; |
| 1849 | } |
| 1850 | |
| 1851 | // Check all uses for MemoryDefs, except for defs completely overwriting |
| 1852 | // the original location. Otherwise we have to check uses of *all* |
| 1853 | // MemoryDefs we discover, including non-aliasing ones. Otherwise we might |
| 1854 | // miss cases like the following |
| 1855 | // 1 = Def(LoE) ; <----- DeadDef stores [0,1] |
| 1856 | // 2 = Def(1) ; (2, 1) = NoAlias, stores [2,3] |
| 1857 | // Use(2) ; MayAlias 2 *and* 1, loads [0, 3]. |
| 1858 | // (The Use points to the *first* Def it may alias) |
| 1859 | // 3 = Def(1) ; <---- Current (3, 2) = NoAlias, (3,1) = MayAlias, |
| 1860 | // stores [0,1] |
| 1861 | if (MemoryDef *UseDef = dyn_cast<MemoryDef>(Val: UseAccess)) { |
| 1862 | if (isCompleteOverwrite(DefLoc: MaybeDeadLoc, DefInst: MaybeDeadI, UseInst)) { |
| 1863 | BasicBlock *MaybeKillingBlock = UseInst->getParent(); |
| 1864 | if (PostOrderNumbers.find(Val: MaybeKillingBlock)->second < |
| 1865 | PostOrderNumbers.find(Val: MaybeDeadAccess->getBlock())->second) { |
| 1866 | if (!isInvisibleToCallerAfterRet(V: KillingUndObj, Ptr: KillingLoc.Ptr, |
| 1867 | StoreSize: KillingLoc.Size)) { |
| 1868 | LLVM_DEBUG(dbgs() |
| 1869 | << " ... found killing def "<< *UseInst << "\n"); |
| 1870 | KillingDefs.insert(Ptr: UseInst); |
| 1871 | } |
| 1872 | } else { |
| 1873 | LLVM_DEBUG(dbgs() |
| 1874 | << " ... found preceeding def "<< *UseInst << "\n"); |
| 1875 | return std::nullopt; |
| 1876 | } |
| 1877 | } else |
| 1878 | pushMemUses(Acc: UseDef, WorkList, Visited); |
| 1879 | } |
| 1880 | } |
| 1881 | |
| 1882 | // For accesses to locations visible after the function returns, make sure |
| 1883 | // that the location is dead (=overwritten) along all paths from |
| 1884 | // MaybeDeadAccess to the exit. |
| 1885 | if (!isInvisibleToCallerAfterRet(V: KillingUndObj, Ptr: KillingLoc.Ptr, |
| 1886 | StoreSize: KillingLoc.Size)) { |
| 1887 | SmallPtrSet<BasicBlock *, 16> KillingBlocks; |
| 1888 | for (Instruction *KD : KillingDefs) |
| 1889 | KillingBlocks.insert(Ptr: KD->getParent()); |
| 1890 | assert(!KillingBlocks.empty() && |
| 1891 | "Expected at least a single killing block"); |
| 1892 | |
| 1893 | // Find the common post-dominator of all killing blocks. |
| 1894 | BasicBlock *CommonPred = *KillingBlocks.begin(); |
| 1895 | for (BasicBlock *BB : llvm::drop_begin(RangeOrContainer&: KillingBlocks)) { |
| 1896 | if (!CommonPred) |
| 1897 | break; |
| 1898 | CommonPred = PDT.findNearestCommonDominator(A: CommonPred, B: BB); |
| 1899 | } |
| 1900 | |
| 1901 | // If the common post-dominator does not post-dominate MaybeDeadAccess, |
| 1902 | // there is a path from MaybeDeadAccess to an exit not going through a |
| 1903 | // killing block. |
| 1904 | if (!PDT.dominates(A: CommonPred, B: MaybeDeadAccess->getBlock())) { |
| 1905 | if (!AnyUnreachableExit) |
| 1906 | return std::nullopt; |
| 1907 | |
| 1908 | // Fall back to CFG scan starting at all non-unreachable roots if not |
| 1909 | // all paths to the exit go through CommonPred. |
| 1910 | CommonPred = nullptr; |
| 1911 | } |
| 1912 | |
| 1913 | // If CommonPred itself is in the set of killing blocks, we're done. |
| 1914 | if (KillingBlocks.count(Ptr: CommonPred)) |
| 1915 | return {MaybeDeadAccess}; |
| 1916 | |
| 1917 | SetVector<BasicBlock *> WorkList; |
| 1918 | // If CommonPred is null, there are multiple exits from the function. |
| 1919 | // They all have to be added to the worklist. |
| 1920 | if (CommonPred) |
| 1921 | WorkList.insert(X: CommonPred); |
| 1922 | else |
| 1923 | for (BasicBlock *R : PDT.roots()) { |
| 1924 | if (!isa<UnreachableInst>(Val: R->getTerminator())) |
| 1925 | WorkList.insert(X: R); |
| 1926 | } |
| 1927 | |
| 1928 | NumCFGTries++; |
| 1929 | // Check if all paths starting from an exit node go through one of the |
| 1930 | // killing blocks before reaching MaybeDeadAccess. |
| 1931 | for (unsigned I = 0; I < WorkList.size(); I++) { |
| 1932 | NumCFGChecks++; |
| 1933 | BasicBlock *Current = WorkList[I]; |
| 1934 | if (KillingBlocks.count(Ptr: Current)) |
| 1935 | continue; |
| 1936 | if (Current == MaybeDeadAccess->getBlock()) |
| 1937 | return std::nullopt; |
| 1938 | |
| 1939 | // MaybeDeadAccess is reachable from the entry, so we don't have to |
| 1940 | // explore unreachable blocks further. |
| 1941 | if (!DT.isReachableFromEntry(A: Current)) |
| 1942 | continue; |
| 1943 | |
| 1944 | WorkList.insert_range(R: predecessors(BB: Current)); |
| 1945 | |
| 1946 | if (WorkList.size() >= Opts.dse_memoryssa_path_check_limit) |
| 1947 | return std::nullopt; |
| 1948 | } |
| 1949 | NumCFGSuccess++; |
| 1950 | } |
| 1951 | |
| 1952 | // No aliasing MemoryUses of MaybeDeadAccess found, MaybeDeadAccess is |
| 1953 | // potentially dead. |
| 1954 | return {MaybeDeadAccess}; |
| 1955 | } |
| 1956 | |
| 1957 | void DSEState::deleteDeadInstruction(Instruction *SI, |
| 1958 | SmallPtrSetImpl<MemoryAccess *> *Deleted) { |
| 1959 | MemorySSAUpdater Updater(&MSSA); |
| 1960 | SmallVector<Instruction *, 32> NowDeadInsts; |
| 1961 | NowDeadInsts.push_back(Elt: SI); |
| 1962 | --NumFastOther; |
| 1963 | |
| 1964 | while (!NowDeadInsts.empty()) { |
| 1965 | Instruction *DeadInst = NowDeadInsts.pop_back_val(); |
| 1966 | ++NumFastOther; |
| 1967 | |
| 1968 | // Try to preserve debug information attached to the dead instruction. |
| 1969 | salvageDebugInfo(I&: *DeadInst); |
| 1970 | salvageKnowledge(I: DeadInst); |
| 1971 | |
| 1972 | // Remove the Instruction from MSSA. |
| 1973 | MemoryAccess *MA = MSSA.getMemoryAccess(I: DeadInst); |
| 1974 | bool IsMemDef = MA && isa<MemoryDef>(Val: MA); |
| 1975 | if (MA) { |
| 1976 | if (IsMemDef) { |
| 1977 | auto *MD = cast<MemoryDef>(Val: MA); |
| 1978 | SkipStores.insert(Ptr: MD); |
| 1979 | if (Deleted) |
| 1980 | Deleted->insert(Ptr: MD); |
| 1981 | if (auto *SI = dyn_cast<StoreInst>(Val: MD->getMemoryInst())) { |
| 1982 | if (SI->getValueOperand()->getType()->isPointerTy()) { |
| 1983 | const Value *UO = getUnderlyingObject(V: SI->getValueOperand()); |
| 1984 | if (CapturedBeforeReturn.erase(Val: UO)) |
| 1985 | ShouldIterateEndOfFunctionDSE = true; |
| 1986 | InvisibleToCallerAfterRet.erase(Val: UO); |
| 1987 | InvisibleToCallerAfterRetBounded.erase(Val: UO); |
| 1988 | } |
| 1989 | } |
| 1990 | } |
| 1991 | |
| 1992 | Updater.removeMemoryAccess(MA); |
| 1993 | } |
| 1994 | |
| 1995 | auto I = IOLs.find(Key: DeadInst->getParent()); |
| 1996 | if (I != IOLs.end()) |
| 1997 | I->second.erase(Key: DeadInst); |
| 1998 | // Remove its operands |
| 1999 | for (Use &O : DeadInst->operands()) |
| 2000 | if (Instruction *OpI = dyn_cast<Instruction>(Val&: O)) { |
| 2001 | O.set(PoisonValue::get(T: O->getType())); |
| 2002 | if (isInstructionTriviallyDead(I: OpI, TLI: &TLI)) |
| 2003 | NowDeadInsts.push_back(Elt: OpI); |
| 2004 | } |
| 2005 | |
| 2006 | EA.removeInstruction(I: DeadInst); |
| 2007 | // Remove memory defs directly if they don't produce results, but only |
| 2008 | // queue other dead instructions for later removal. They may have been |
| 2009 | // used as memory locations that have been cached by BatchAA. Removing |
| 2010 | // them here may lead to newly created instructions to be allocated at the |
| 2011 | // same address, yielding stale cache entries. |
| 2012 | if (IsMemDef && DeadInst->getType()->isVoidTy()) |
| 2013 | DeadInst->eraseFromParent(); |
| 2014 | else |
| 2015 | ToRemove.push_back(Elt: DeadInst); |
| 2016 | } |
| 2017 | } |
| 2018 | |
| 2019 | bool DSEState::mayThrowBetween(Instruction *KillingI, Instruction *DeadI, |
| 2020 | const Value *KillingUndObj) { |
| 2021 | // First see if we can ignore it by using the fact that KillingI is an |
| 2022 | // alloca/alloca like object that is not visible to the caller during |
| 2023 | // execution of the function. |
| 2024 | if (KillingUndObj && isInvisibleToCallerOnUnwind(V: KillingUndObj)) |
| 2025 | return false; |
| 2026 | |
| 2027 | if (KillingI->getParent() == DeadI->getParent()) |
| 2028 | return ThrowingBlocks.count(Ptr: KillingI->getParent()); |
| 2029 | return !ThrowingBlocks.empty(); |
| 2030 | } |
| 2031 | |
| 2032 | bool DSEState::isDSEBarrier(const Value *KillingUndObj, Instruction *DeadI) { |
| 2033 | // If DeadI may throw it acts as a barrier, unless we are to an |
| 2034 | // alloca/alloca like object that does not escape. |
| 2035 | if (DeadI->mayThrow() && !isInvisibleToCallerOnUnwind(V: KillingUndObj)) |
| 2036 | return true; |
| 2037 | |
| 2038 | // If DeadI is an atomic load/store stronger than monotonic, do not try to |
| 2039 | // eliminate/reorder it. |
| 2040 | if (DeadI->isAtomic()) { |
| 2041 | if (auto *LI = dyn_cast<LoadInst>(Val: DeadI)) |
| 2042 | return isStrongerThanMonotonic(AO: LI->getOrdering()); |
| 2043 | if (auto *SI = dyn_cast<StoreInst>(Val: DeadI)) |
| 2044 | return isStrongerThanMonotonic(AO: SI->getOrdering()); |
| 2045 | if (auto *ARMW = dyn_cast<AtomicRMWInst>(Val: DeadI)) |
| 2046 | return isStrongerThanMonotonic(AO: ARMW->getOrdering()); |
| 2047 | if (auto *CmpXchg = dyn_cast<AtomicCmpXchgInst>(Val: DeadI)) |
| 2048 | return isStrongerThanMonotonic(AO: CmpXchg->getSuccessOrdering()) || |
| 2049 | isStrongerThanMonotonic(AO: CmpXchg->getFailureOrdering()); |
| 2050 | llvm_unreachable("other instructions should be skipped in MemorySSA"); |
| 2051 | } |
| 2052 | return false; |
| 2053 | } |
| 2054 | |
| 2055 | bool DSEState::eliminateDeadWritesAtEndOfFunction() { |
| 2056 | bool MadeChange = false; |
| 2057 | LLVM_DEBUG( |
| 2058 | dbgs() << "Trying to eliminate MemoryDefs at the end of the function\n"); |
| 2059 | do { |
| 2060 | ShouldIterateEndOfFunctionDSE = false; |
| 2061 | for (MemoryDef *Def : llvm::reverse(C&: MemDefs)) { |
| 2062 | if (SkipStores.contains(Ptr: Def)) |
| 2063 | continue; |
| 2064 | |
| 2065 | Instruction *DefI = Def->getMemoryInst(); |
| 2066 | auto DefLoc = getLocForWrite(I: DefI); |
| 2067 | if (!DefLoc || !isRemovable(I: DefI)) { |
| 2068 | LLVM_DEBUG(dbgs() << " ... could not get location for write or " |
| 2069 | "instruction not removable.\n"); |
| 2070 | continue; |
| 2071 | } |
| 2072 | |
| 2073 | // NOTE: Currently eliminating writes at the end of a function is |
| 2074 | // limited to MemoryDefs with a single underlying object, to save |
| 2075 | // compile-time. In practice it appears the case with multiple |
| 2076 | // underlying objects is very uncommon. If it turns out to be important, |
| 2077 | // we can use getUnderlyingObjects here instead. |
| 2078 | const Value *UO = getUnderlyingObject(V: DefLoc->Ptr); |
| 2079 | if (!isInvisibleToCallerAfterRet(V: UO, Ptr: DefLoc->Ptr, StoreSize: DefLoc->Size)) |
| 2080 | continue; |
| 2081 | |
| 2082 | if (isWriteAtEndOfFunction(Def, DefLoc: *DefLoc)) { |
| 2083 | // See through pointer-to-pointer bitcasts |
| 2084 | LLVM_DEBUG(dbgs() << " ... MemoryDef is not accessed until the end " |
| 2085 | "of the function\n"); |
| 2086 | deleteDeadInstruction(SI: DefI); |
| 2087 | ++NumFastStores; |
| 2088 | MadeChange = true; |
| 2089 | } |
| 2090 | } |
| 2091 | } while (ShouldIterateEndOfFunctionDSE); |
| 2092 | return MadeChange; |
| 2093 | } |
| 2094 | |
| 2095 | bool DSEState::eliminateRedundantStoresViaDominatingConditions() { |
| 2096 | bool MadeChange = false; |
| 2097 | LLVM_DEBUG(dbgs() << "Trying to eliminate MemoryDefs whose value being " |
| 2098 | "written is implied by a dominating condition\n"); |
| 2099 | |
| 2100 | using ConditionInfo = std::pair<Value *, Value *>; |
| 2101 | using ScopedHTType = ScopedHashTable<ConditionInfo, Instruction *>; |
| 2102 | |
| 2103 | // We maintain a scoped hash table of the active dominating conditions for a |
| 2104 | // given node. |
| 2105 | ScopedHTType ActiveConditions; |
| 2106 | auto GetDominatingCondition = [&](BasicBlock *BB) |
| 2107 | -> std::optional<std::tuple<ConditionInfo, Instruction *, BasicBlock *>> { |
| 2108 | auto *BI = dyn_cast<CondBrInst>(Val: BB->getTerminator()); |
| 2109 | if (!BI) |
| 2110 | return std::nullopt; |
| 2111 | |
| 2112 | // In case both blocks are the same, it is not possible to determine |
| 2113 | // if optimization is possible. (We would not want to optimize a store |
| 2114 | // in the FalseBB if condition is true and vice versa.) |
| 2115 | if (BI->getSuccessor(i: 0) == BI->getSuccessor(i: 1)) |
| 2116 | return std::nullopt; |
| 2117 | |
| 2118 | Instruction *ICmpL; |
| 2119 | CmpPredicate Pred; |
| 2120 | Value *StorePtr, *StoreVal; |
| 2121 | if (!match(V: BI->getCondition(), |
| 2122 | P: m_c_ICmp(Pred, L: m_Instruction(I&: ICmpL, P: m_Load(Op: m_Value(V&: StorePtr))), |
| 2123 | R: m_Value(V&: StoreVal))) || |
| 2124 | !ICmpInst::isEquality(P: Pred)) |
| 2125 | return std::nullopt; |
| 2126 | |
| 2127 | // Ensure the replacement is allowed when comparing pointers, as |
| 2128 | // the equality compares addresses only, not pointers' provenance. |
| 2129 | if (StoreVal->getType()->isPointerTy() && |
| 2130 | !canReplacePointersIfEqual(From: StoreVal, To: ICmpL, DL)) |
| 2131 | return std::nullopt; |
| 2132 | |
| 2133 | unsigned ImpliedSuccIdx = Pred == ICmpInst::ICMP_EQ ? 0 : 1; |
| 2134 | BasicBlock *ImpliedSucc = BI->getSuccessor(i: ImpliedSuccIdx); |
| 2135 | return {{ConditionInfo(StorePtr, StoreVal), ICmpL, ImpliedSucc}}; |
| 2136 | }; |
| 2137 | |
| 2138 | auto VisitNode = [&](DomTreeNode *Node, unsigned Depth, auto &Self) -> void { |
| 2139 | if (Depth > Opts.dse_max_dom_cond_depth) |
| 2140 | return; |
| 2141 | |
| 2142 | BasicBlock *BB = Node->getBlock(); |
| 2143 | // Check for redundant stores against active known conditions. |
| 2144 | if (auto *Accesses = MSSA.getBlockDefs(BB)) { |
| 2145 | for (MemoryDef &Def : |
| 2146 | make_early_inc_range(Range: make_isa_range<MemoryDef>(Range&: *Accesses))) { |
| 2147 | auto *SI = dyn_cast<StoreInst>(Val: Def.getMemoryInst()); |
| 2148 | if (!SI || !SI->isUnordered()) |
| 2149 | continue; |
| 2150 | |
| 2151 | Instruction *LI = ActiveConditions.lookup( |
| 2152 | Key: {SI->getPointerOperand(), SI->getValueOperand()}); |
| 2153 | if (!LI) |
| 2154 | continue; |
| 2155 | |
| 2156 | // Found a dominating condition that may imply the value being stored. |
| 2157 | // Make sure there does not exist any clobbering access between the |
| 2158 | // load and the potential redundant store. |
| 2159 | MemoryAccess *LoadAccess = MSSA.getMemoryAccess(I: LI); |
| 2160 | MemoryAccess *ClobberingAccess = |
| 2161 | MSSA.getSkipSelfWalker()->getClobberingMemoryAccess(&Def, AA&: BatchAA); |
| 2162 | if (MSSA.dominates(A: ClobberingAccess, B: LoadAccess)) { |
| 2163 | LLVM_DEBUG(dbgs() |
| 2164 | << "Removing No-Op Store:\n DEAD: "<< *SI << '\n'); |
| 2165 | deleteDeadInstruction(SI); |
| 2166 | NumRedundantStores++; |
| 2167 | MadeChange = true; |
| 2168 | } |
| 2169 | } |
| 2170 | } |
| 2171 | |
| 2172 | // See whether this basic block establishes a dominating condition. |
| 2173 | auto MaybeCondition = GetDominatingCondition(BB); |
| 2174 | |
| 2175 | for (DomTreeNode *Child : Node->children()) { |
| 2176 | // RAII scope for the active conditions. |
| 2177 | ScopedHTType::ScopeTy Scope(ActiveConditions); |
| 2178 | if (MaybeCondition) { |
| 2179 | const auto &[Cond, LI, ImpliedSucc] = *MaybeCondition; |
| 2180 | if (DT.dominates(BBE: BasicBlockEdge(BB, ImpliedSucc), BB: Child->getBlock())) { |
| 2181 | // Found a condition that holds for this child, dominated by the |
| 2182 | // current node via the equality edge. Propagate the condition to |
| 2183 | // the children by pushing it onto the table. |
| 2184 | ActiveConditions.insert(Key: Cond, Val: LI); |
| 2185 | } |
| 2186 | } |
| 2187 | |
| 2188 | // Recursively visit the children of this node. Upon destruction, the no |
| 2189 | // longer active condition before visiting any sibling nodes is popped |
| 2190 | // from the active scope. |
| 2191 | Self(Child, Depth + 1, Self); |
| 2192 | } |
| 2193 | }; |
| 2194 | |
| 2195 | // Do a DFS walk of the dom-tree. |
| 2196 | VisitNode(DT.getRootNode(), 0, VisitNode); |
| 2197 | |
| 2198 | return MadeChange; |
| 2199 | } |
| 2200 | |
| 2201 | bool DSEState::tryFoldIntoCalloc(MemoryDef *Def, const Value *DefUO) { |
| 2202 | Instruction *DefI = Def->getMemoryInst(); |
| 2203 | MemSetInst *MemSet = dyn_cast<MemSetInst>(Val: DefI); |
| 2204 | if (!MemSet) |
| 2205 | // TODO: Could handle zero store to small allocation as well. |
| 2206 | return false; |
| 2207 | Constant *StoredConstant = dyn_cast<Constant>(Val: MemSet->getValue()); |
| 2208 | if (!StoredConstant || !StoredConstant->isNullValue()) |
| 2209 | return false; |
| 2210 | |
| 2211 | if (!isRemovable(I: DefI)) |
| 2212 | // The memset might be volatile.. |
| 2213 | return false; |
| 2214 | |
| 2215 | if (F.hasFnAttribute(Kind: Attribute::SanitizeMemory) || |
| 2216 | F.hasFnAttribute(Kind: Attribute::SanitizeAddress) || |
| 2217 | F.hasFnAttribute(Kind: Attribute::SanitizeHWAddress) || F.getName() == "calloc") |
| 2218 | return false; |
| 2219 | auto *Malloc = const_cast<CallInst *>(dyn_cast<CallInst>(Val: DefUO)); |
| 2220 | if (!Malloc) |
| 2221 | return false; |
| 2222 | auto *InnerCallee = Malloc->getCalledFunction(); |
| 2223 | if (!InnerCallee) |
| 2224 | return false; |
| 2225 | LibFunc Func = TLI.getLibFunc(FDecl: *InnerCallee); |
| 2226 | StringRef ZeroedVariantName; |
| 2227 | if (Func != LibFunc_malloc || !TLI.has(F: Func)) { |
| 2228 | Attribute Attr = Malloc->getFnAttr(Kind: "alloc-variant-zeroed"); |
| 2229 | if (!Attr.isValid()) |
| 2230 | return false; |
| 2231 | ZeroedVariantName = Attr.getValueAsString(); |
| 2232 | if (ZeroedVariantName.empty()) |
| 2233 | return false; |
| 2234 | } |
| 2235 | |
| 2236 | // Gracefully handle malloc with unexpected memory attributes. |
| 2237 | auto *MallocDef = dyn_cast_or_null<MemoryDef>(Val: MSSA.getMemoryAccess(I: Malloc)); |
| 2238 | if (!MallocDef) |
| 2239 | return false; |
| 2240 | |
| 2241 | auto shouldCreateCalloc = [](CallInst *Malloc, CallInst *Memset) { |
| 2242 | // Check for br(icmp ptr, null), truebb, falsebb) pattern at the end |
| 2243 | // of malloc block |
| 2244 | auto *MallocBB = Malloc->getParent(), *MemsetBB = Memset->getParent(); |
| 2245 | if (MallocBB == MemsetBB) |
| 2246 | return true; |
| 2247 | auto *Ptr = Memset->getArgOperand(i: 0); |
| 2248 | auto *TI = MallocBB->getTerminator(); |
| 2249 | BasicBlock *TrueBB, *FalseBB; |
| 2250 | if (!match(V: TI, P: m_Br(C: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Specific(V: Ptr), |
| 2251 | R: m_Zero()), |
| 2252 | T&: TrueBB, F&: FalseBB))) |
| 2253 | return false; |
| 2254 | if (MemsetBB != FalseBB) |
| 2255 | return false; |
| 2256 | return true; |
| 2257 | }; |
| 2258 | |
| 2259 | if (Malloc->getOperand(i_nocapture: 0) != MemSet->getLength()) |
| 2260 | return false; |
| 2261 | if (!shouldCreateCalloc(Malloc, MemSet) || !DT.dominates(Def: Malloc, User: MemSet) || |
| 2262 | !memoryIsNotModifiedBetween(FirstI: Malloc, SecondI: MemSet, AA&: BatchAA, DL, DT: &DT)) |
| 2263 | return false; |
| 2264 | IRBuilder<> IRB(Malloc); |
| 2265 | assert(Func == LibFunc_malloc || !ZeroedVariantName.empty()); |
| 2266 | Value *Calloc = nullptr; |
| 2267 | if (!ZeroedVariantName.empty()) { |
| 2268 | LLVMContext &Ctx = Malloc->getContext(); |
| 2269 | AttributeList Attrs = InnerCallee->getAttributes(); |
| 2270 | AllocFnKind AllocKind = |
| 2271 | Attrs.getFnAttr(Kind: Attribute::AllocKind).getAllocKind() | |
| 2272 | AllocFnKind::Zeroed; |
| 2273 | AllocKind &= ~AllocFnKind::Uninitialized; |
| 2274 | Attrs = |
| 2275 | Attrs.addFnAttribute(C&: Ctx, Attr: Attribute::getWithAllocKind(Context&: Ctx, Kind: AllocKind)) |
| 2276 | .removeFnAttribute(C&: Ctx, Kind: "alloc-variant-zeroed"); |
| 2277 | FunctionCallee ZeroedVariant = Malloc->getModule()->getOrInsertFunction( |
| 2278 | Name: ZeroedVariantName, T: InnerCallee->getFunctionType(), AttributeList: Attrs); |
| 2279 | cast<Function>(Val: ZeroedVariant.getCallee()) |
| 2280 | ->setCallingConv(Malloc->getCallingConv()); |
| 2281 | SmallVector<Value *, 3> Args; |
| 2282 | Args.append(in_start: Malloc->arg_begin(), in_end: Malloc->arg_end()); |
| 2283 | CallInst *CI = IRB.CreateCall(Callee: ZeroedVariant, Args, Name: ZeroedVariantName); |
| 2284 | CI->setCallingConv(Malloc->getCallingConv()); |
| 2285 | Calloc = CI; |
| 2286 | } else { |
| 2287 | Type *SizeTTy = Malloc->getArgOperand(i: 0)->getType(); |
| 2288 | Calloc = emitCalloc(Num: ConstantInt::get(Ty: SizeTTy, V: 1), Size: Malloc->getArgOperand(i: 0), |
| 2289 | B&: IRB, TLI, AddrSpace: Malloc->getType()->getPointerAddressSpace()); |
| 2290 | } |
| 2291 | if (!Calloc) |
| 2292 | return false; |
| 2293 | |
| 2294 | if (MDNode *MD = Malloc->getMetadata(KindID: LLVMContext::MD_alloc_token)) |
| 2295 | cast<Instruction>(Val: Calloc)->setMetadata(KindID: LLVMContext::MD_alloc_token, Node: MD); |
| 2296 | |
| 2297 | MemorySSAUpdater Updater(&MSSA); |
| 2298 | auto *NewAccess = Updater.createMemoryAccessAfter(I: cast<Instruction>(Val: Calloc), |
| 2299 | Definition: nullptr, InsertPt: MallocDef); |
| 2300 | auto *NewAccessMD = cast<MemoryDef>(Val: NewAccess); |
| 2301 | Updater.insertDef(Def: NewAccessMD, /*RenameUses=*/true); |
| 2302 | Malloc->replaceAllUsesWith(V: Calloc); |
| 2303 | deleteDeadInstruction(SI: Malloc); |
| 2304 | return true; |
| 2305 | } |
| 2306 | |
| 2307 | bool DSEState::storeIsNoop(MemoryDef *Def, const Value *DefUO) { |
| 2308 | Instruction *DefI = Def->getMemoryInst(); |
| 2309 | StoreInst *Store = dyn_cast<StoreInst>(Val: DefI); |
| 2310 | MemSetInst *MemSet = dyn_cast<MemSetInst>(Val: DefI); |
| 2311 | Constant *StoredConstant = nullptr; |
| 2312 | if (Store) |
| 2313 | StoredConstant = dyn_cast<Constant>(Val: Store->getOperand(i_nocapture: 0)); |
| 2314 | else if (MemSet) |
| 2315 | StoredConstant = dyn_cast<Constant>(Val: MemSet->getValue()); |
| 2316 | else |
| 2317 | return false; |
| 2318 | |
| 2319 | if (!isRemovable(I: DefI)) |
| 2320 | return false; |
| 2321 | |
| 2322 | if (StoredConstant) { |
| 2323 | Constant *InitC = |
| 2324 | getInitialValueOfAllocation(V: DefUO, TLI: &TLI, Ty: StoredConstant->getType()); |
| 2325 | // If the clobbering access is LiveOnEntry, no instructions between them |
| 2326 | // can modify the memory location. |
| 2327 | if (InitC && InitC == StoredConstant) |
| 2328 | return MSSA.isLiveOnEntryDef( |
| 2329 | MA: MSSA.getSkipSelfWalker()->getClobberingMemoryAccess(Def, AA&: BatchAA)); |
| 2330 | } |
| 2331 | |
| 2332 | if (!Store) |
| 2333 | return false; |
| 2334 | |
| 2335 | if (auto *LoadI = dyn_cast<LoadInst>(Val: Store->getOperand(i_nocapture: 0))) { |
| 2336 | if (LoadI->getPointerOperand() == Store->getOperand(i_nocapture: 1)) { |
| 2337 | // Get the defining access for the load. |
| 2338 | auto *LoadAccess = MSSA.getMemoryAccess(I: LoadI)->getDefiningAccess(); |
| 2339 | // Fast path: the defining accesses are the same. |
| 2340 | if (LoadAccess == Def->getDefiningAccess()) |
| 2341 | return true; |
| 2342 | |
| 2343 | // Look through phi accesses. Recursively scan all phi accesses by |
| 2344 | // adding them to a worklist. Bail when we run into a memory def that |
| 2345 | // does not match LoadAccess. |
| 2346 | SetVector<MemoryAccess *> ToCheck; |
| 2347 | MemoryAccess *Current = |
| 2348 | MSSA.getWalker()->getClobberingMemoryAccess(Def, AA&: BatchAA); |
| 2349 | // We don't want to bail when we run into the store memory def. But, |
| 2350 | // the phi access may point to it. So, pretend like we've already |
| 2351 | // checked it. |
| 2352 | ToCheck.insert(X: Def); |
| 2353 | ToCheck.insert(X: Current); |
| 2354 | // Start at current (1) to simulate already having checked Def. |
| 2355 | for (unsigned I = 1; I < ToCheck.size(); ++I) { |
| 2356 | Current = ToCheck[I]; |
| 2357 | if (auto PhiAccess = dyn_cast<MemoryPhi>(Val: Current)) { |
| 2358 | // Check all the operands. |
| 2359 | for (auto &Use : PhiAccess->incoming_values()) |
| 2360 | ToCheck.insert(X: cast<MemoryAccess>(Val: &Use)); |
| 2361 | continue; |
| 2362 | } |
| 2363 | |
| 2364 | // If we found a memory def, bail. This happens when we have an |
| 2365 | // unrelated write in between an otherwise noop store. |
| 2366 | assert(isa<MemoryDef>(Current) && "Only MemoryDefs should reach here."); |
| 2367 | // TODO: Skip no alias MemoryDefs that have no aliasing reads. |
| 2368 | // We are searching for the definition of the store's destination. |
| 2369 | // So, if that is the same definition as the load, then this is a |
| 2370 | // noop. Otherwise, fail. |
| 2371 | if (LoadAccess != Current) |
| 2372 | return false; |
| 2373 | } |
| 2374 | return true; |
| 2375 | } |
| 2376 | } |
| 2377 | |
| 2378 | return false; |
| 2379 | } |
| 2380 | |
| 2381 | bool DSEState::removePartiallyOverlappedStores(InstOverlapIntervalsTy &IOL) { |
| 2382 | bool Changed = false; |
| 2383 | for (auto OI : IOL) { |
| 2384 | Instruction *DeadI = OI.first; |
| 2385 | MemoryLocation Loc = *getLocForWrite(I: DeadI); |
| 2386 | assert(isRemovable(DeadI) && "Expect only removable instruction"); |
| 2387 | |
| 2388 | const Value *Ptr = Loc.Ptr->stripPointerCasts(); |
| 2389 | int64_t DeadStart = 0; |
| 2390 | uint64_t DeadSize = Loc.Size.getValue(); |
| 2391 | GetPointerBaseWithConstantOffset(Ptr, Offset&: DeadStart, DL); |
| 2392 | OverlapIntervalsTy &IntervalMap = OI.second; |
| 2393 | Changed |= tryToShortenEnd(DeadI, IntervalMap, DeadStart, DeadSize); |
| 2394 | if (IntervalMap.empty()) |
| 2395 | continue; |
| 2396 | Changed |= tryToShortenBegin(DeadI, IntervalMap, DeadStart, DeadSize); |
| 2397 | } |
| 2398 | return Changed; |
| 2399 | } |
| 2400 | |
| 2401 | bool DSEState::eliminateRedundantStoresOfExistingValues() { |
| 2402 | bool MadeChange = false; |
| 2403 | LLVM_DEBUG(dbgs() << "Trying to eliminate MemoryDefs that write the " |
| 2404 | "already existing value\n"); |
| 2405 | for (auto *Def : MemDefs) { |
| 2406 | if (SkipStores.contains(Ptr: Def) || MSSA.isLiveOnEntryDef(MA: Def)) |
| 2407 | continue; |
| 2408 | |
| 2409 | Instruction *DefInst = Def->getMemoryInst(); |
| 2410 | auto MaybeDefLoc = getLocForWrite(I: DefInst); |
| 2411 | if (!MaybeDefLoc || !isRemovable(I: DefInst)) |
| 2412 | continue; |
| 2413 | |
| 2414 | MemoryDef *UpperDef; |
| 2415 | // To conserve compile-time, we avoid walking to the next clobbering def. |
| 2416 | // Instead, we just try to get the optimized access, if it exists. DSE |
| 2417 | // will try to optimize defs during the earlier traversal. |
| 2418 | if (Def->isOptimized()) |
| 2419 | UpperDef = dyn_cast<MemoryDef>(Val: Def->getOptimized()); |
| 2420 | else |
| 2421 | UpperDef = dyn_cast<MemoryDef>(Val: Def->getDefiningAccess()); |
| 2422 | if (!UpperDef || MSSA.isLiveOnEntryDef(MA: UpperDef)) |
| 2423 | continue; |
| 2424 | |
| 2425 | Instruction *UpperInst = UpperDef->getMemoryInst(); |
| 2426 | auto IsRedundantStore = [&]() { |
| 2427 | // We don't care about differences in call attributes here. |
| 2428 | if (DefInst->isIdenticalToWhenDefined(I: UpperInst, |
| 2429 | /*IntersectAttrs=*/true)) |
| 2430 | return true; |
| 2431 | if (auto *MemSetI = dyn_cast<MemSetInst>(Val: UpperInst)) { |
| 2432 | if (auto *SI = dyn_cast<StoreInst>(Val: DefInst)) { |
| 2433 | // MemSetInst must have a write location. |
| 2434 | auto UpperLoc = getLocForWrite(I: UpperInst); |
| 2435 | if (!UpperLoc) |
| 2436 | return false; |
| 2437 | int64_t InstWriteOffset = 0; |
| 2438 | int64_t DepWriteOffset = 0; |
| 2439 | auto OR = isOverwrite(KillingI: UpperInst, DeadI: DefInst, KillingLoc: *UpperLoc, DeadLoc: *MaybeDefLoc, |
| 2440 | KillingOff&: InstWriteOffset, DeadOff&: DepWriteOffset); |
| 2441 | Value *StoredByte = isBytewiseValue(V: SI->getValueOperand(), DL); |
| 2442 | return StoredByte && StoredByte == MemSetI->getOperand(i_nocapture: 1) && |
| 2443 | OR == OW_Complete; |
| 2444 | } |
| 2445 | } |
| 2446 | return false; |
| 2447 | }; |
| 2448 | |
| 2449 | if (!IsRedundantStore() || isReadClobber(DefLoc: *MaybeDefLoc, UseInst: DefInst)) |
| 2450 | continue; |
| 2451 | LLVM_DEBUG(dbgs() << "DSE: Remove No-Op Store:\n DEAD: "<< *DefInst |
| 2452 | << '\n'); |
| 2453 | deleteDeadInstruction(SI: DefInst); |
| 2454 | NumRedundantStores++; |
| 2455 | MadeChange = true; |
| 2456 | } |
| 2457 | return MadeChange; |
| 2458 | } |
| 2459 | |
| 2460 | SmallVector<MemoryLocation, 1> |
| 2461 | DSEState::getInitializesArgMemLoc(const Instruction *I) { |
| 2462 | const CallBase *CB = dyn_cast<CallBase>(Val: I); |
| 2463 | if (!CB) |
| 2464 | return {}; |
| 2465 | |
| 2466 | // Collect aliasing arguments and their initializes ranges. |
| 2467 | SmallMapVector<Value *, SmallVector<ArgumentInitInfo, 2>, 2> Arguments; |
| 2468 | for (unsigned Idx = 0, Count = CB->arg_size(); Idx < Count; ++Idx) { |
| 2469 | Value *CurArg = CB->getArgOperand(i: Idx); |
| 2470 | if (!CurArg->getType()->isPointerTy()) |
| 2471 | continue; |
| 2472 | |
| 2473 | ConstantRangeList Inits; |
| 2474 | Attribute InitializesAttr = CB->getParamAttr(ArgNo: Idx, Kind: Attribute::Initializes); |
| 2475 | // initializes on byval arguments refers to the callee copy, not the |
| 2476 | // original memory the caller passed in. |
| 2477 | if (InitializesAttr.isValid() && !CB->isByValArgument(ArgNo: Idx)) |
| 2478 | Inits = InitializesAttr.getValueAsConstantRangeList(); |
| 2479 | |
| 2480 | // Check whether "CurArg" could alias with global variables. We require |
| 2481 | // either it's function local and isn't captured before or the "CB" only |
| 2482 | // accesses arg or inaccessible mem. |
| 2483 | if (!Inits.empty() && !CB->onlyAccessesInaccessibleMemOrArgMem() && |
| 2484 | !isFuncLocalAndNotCaptured(Arg: CurArg, CB, EA)) |
| 2485 | Inits = ConstantRangeList(); |
| 2486 | |
| 2487 | // We don't perform incorrect DSE on unwind edges in the current function, |
| 2488 | // and use the "initializes" attribute to kill dead stores if: |
| 2489 | // - The call does not throw exceptions, "CB->doesNotThrow()". |
| 2490 | // - Or the callee parameter has "dead_on_unwind" attribute. |
| 2491 | // - Or the argument is invisible to caller on unwind, and there are no |
| 2492 | // unwind edges from this call in the current function (e.g. `CallInst`). |
| 2493 | bool IsDeadOrInvisibleOnUnwind = |
| 2494 | CB->paramHasAttr(ArgNo: Idx, Kind: Attribute::DeadOnUnwind) || |
| 2495 | (isa<CallInst>(Val: CB) && isInvisibleToCallerOnUnwind(V: CurArg)); |
| 2496 | ArgumentInitInfo InitInfo{.Idx: Idx, .IsDeadOrInvisibleOnUnwind: IsDeadOrInvisibleOnUnwind, .Inits: Inits}; |
| 2497 | bool FoundAliasing = false; |
| 2498 | for (auto &[Arg, AliasList] : Arguments) { |
| 2499 | auto AAR = BatchAA.alias(LocA: MemoryLocation::getBeforeOrAfter(Ptr: Arg), |
| 2500 | LocB: MemoryLocation::getBeforeOrAfter(Ptr: CurArg)); |
| 2501 | if (AAR == AliasResult::NoAlias) { |
| 2502 | continue; |
| 2503 | } else if (AAR == AliasResult::MustAlias) { |
| 2504 | FoundAliasing = true; |
| 2505 | AliasList.push_back(Elt: InitInfo); |
| 2506 | } else { |
| 2507 | // For PartialAlias and MayAlias, there is an offset or may be an |
| 2508 | // unknown offset between the arguments and we insert an empty init |
| 2509 | // range to discard the entire initializes info while intersecting. |
| 2510 | FoundAliasing = true; |
| 2511 | AliasList.push_back(Elt: ArgumentInitInfo{.Idx: Idx, .IsDeadOrInvisibleOnUnwind: IsDeadOrInvisibleOnUnwind, |
| 2512 | .Inits: ConstantRangeList()}); |
| 2513 | } |
| 2514 | } |
| 2515 | if (!FoundAliasing) |
| 2516 | Arguments[CurArg] = {InitInfo}; |
| 2517 | } |
| 2518 | |
| 2519 | SmallVector<MemoryLocation, 1> Locations; |
| 2520 | for (const auto &[_, Args] : Arguments) { |
| 2521 | auto IntersectedRanges = |
| 2522 | getIntersectedInitRangeList(Args, CallHasNoUnwindAttr: CB->doesNotThrow()); |
| 2523 | if (IntersectedRanges.empty()) |
| 2524 | continue; |
| 2525 | |
| 2526 | for (const auto &Arg : Args) { |
| 2527 | for (const auto &Range : IntersectedRanges) { |
| 2528 | int64_t Start = Range.getLower().getSExtValue(); |
| 2529 | int64_t End = Range.getUpper().getSExtValue(); |
| 2530 | // For now, we only handle locations starting at offset 0. |
| 2531 | if (Start == 0) |
| 2532 | Locations.push_back(Elt: MemoryLocation(CB->getArgOperand(i: Arg.Idx), |
| 2533 | LocationSize::precise(Value: End - Start), |
| 2534 | CB->getAAMetadata())); |
| 2535 | } |
| 2536 | } |
| 2537 | } |
| 2538 | return Locations; |
| 2539 | } |
| 2540 | |
| 2541 | std::pair<bool, bool> |
| 2542 | DSEState::eliminateDeadDefs(const MemoryLocationWrapper &KillingLocWrapper) { |
| 2543 | bool Changed = false; |
| 2544 | bool DeletedKillingLoc = false; |
| 2545 | unsigned ScanLimit = Opts.dse_memoryssa_scanlimit; |
| 2546 | unsigned WalkerStepLimit = Opts.dse_memoryssa_walklimit; |
| 2547 | unsigned PartialLimit = Opts.dse_memoryssa_partial_store_limit; |
| 2548 | // Worklist of MemoryAccesses that may be killed by |
| 2549 | // "KillingLocWrapper.MemDef". |
| 2550 | SmallSetVector<MemoryAccess *, 8> ToCheck; |
| 2551 | // Track MemoryAccesses that have been deleted in the loop below, so we can |
| 2552 | // skip them. Don't use SkipStores for this, which may contain reused |
| 2553 | // MemoryAccess addresses. |
| 2554 | SmallPtrSet<MemoryAccess *, 8> Deleted; |
| 2555 | [[maybe_unused]] unsigned OrigNumSkipStores = SkipStores.size(); |
| 2556 | ToCheck.insert(X: KillingLocWrapper.MemDef->getDefiningAccess()); |
| 2557 | |
| 2558 | // Check if MemoryAccesses in the worklist are killed by |
| 2559 | // "KillingLocWrapper.MemDef". |
| 2560 | for (unsigned I = 0; I < ToCheck.size(); I++) { |
| 2561 | MemoryAccess *Current = ToCheck[I]; |
| 2562 | if (Deleted.contains(Ptr: Current)) |
| 2563 | continue; |
| 2564 | std::optional<MemoryAccess *> MaybeDeadAccess = getDomMemoryDef( |
| 2565 | KillingDef: KillingLocWrapper.MemDef, StartAccess: Current, KillingLoc: KillingLocWrapper.MemLoc, |
| 2566 | KillingUndObj: KillingLocWrapper.UnderlyingObject, ScanLimit, WalkerStepLimit, |
| 2567 | IsMemTerm: isMemTerminatorInst(I: KillingLocWrapper.DefInst), PartialLimit, |
| 2568 | IsInitializesAttrMemLoc: KillingLocWrapper.DefByInitializesAttr); |
| 2569 | |
| 2570 | if (!MaybeDeadAccess) { |
| 2571 | LLVM_DEBUG(dbgs() << " finished walk\n"); |
| 2572 | continue; |
| 2573 | } |
| 2574 | MemoryAccess *DeadAccess = *MaybeDeadAccess; |
| 2575 | LLVM_DEBUG(dbgs() << " Checking if we can kill "<< *DeadAccess); |
| 2576 | if (isa<MemoryPhi>(Val: DeadAccess)) { |
| 2577 | LLVM_DEBUG(dbgs() << "\n ... adding incoming values to worklist\n"); |
| 2578 | for (Value *V : cast<MemoryPhi>(Val: DeadAccess)->incoming_values()) { |
| 2579 | MemoryAccess *IncomingAccess = cast<MemoryAccess>(Val: V); |
| 2580 | BasicBlock *IncomingBlock = IncomingAccess->getBlock(); |
| 2581 | BasicBlock *PhiBlock = DeadAccess->getBlock(); |
| 2582 | |
| 2583 | // We only consider incoming MemoryAccesses that come before the |
| 2584 | // MemoryPhi. Otherwise we could discover candidates that do not |
| 2585 | // strictly dominate our starting def. |
| 2586 | if (PostOrderNumbers[IncomingBlock] > PostOrderNumbers[PhiBlock]) |
| 2587 | ToCheck.insert(X: IncomingAccess); |
| 2588 | } |
| 2589 | continue; |
| 2590 | } |
| 2591 | // We cannot apply the initializes attribute to DeadAccess/DeadDef. |
| 2592 | // It would incorrectly consider a call instruction as redundant store |
| 2593 | // and remove this call instruction. |
| 2594 | // TODO: this conflates the existence of a MemoryLocation with being able |
| 2595 | // to delete the instruction. Fix isRemovable() to consider calls with |
| 2596 | // side effects that cannot be removed, e.g. calls with the initializes |
| 2597 | // attribute, and remove getLocForInst(ConsiderInitializesAttr = false). |
| 2598 | MemoryDefWrapper DeadDefWrapper( |
| 2599 | cast<MemoryDef>(Val: DeadAccess), |
| 2600 | getLocForInst(I: cast<MemoryDef>(Val: DeadAccess)->getMemoryInst(), |
| 2601 | /*ConsiderInitializesAttr=*/false)); |
| 2602 | assert(DeadDefWrapper.DefinedLocations.size() == 1); |
| 2603 | MemoryLocationWrapper &DeadLocWrapper = |
| 2604 | DeadDefWrapper.DefinedLocations.front(); |
| 2605 | LLVM_DEBUG(dbgs() << " ("<< *DeadLocWrapper.DefInst << ")\n"); |
| 2606 | ToCheck.insert(X: DeadLocWrapper.MemDef->getDefiningAccess()); |
| 2607 | NumGetDomMemoryDefPassed++; |
| 2608 | |
| 2609 | if (!DebugCounter::shouldExecute(Counter&: MemorySSACounter)) |
| 2610 | continue; |
| 2611 | if (isMemTerminatorInst(I: KillingLocWrapper.DefInst)) { |
| 2612 | if (KillingLocWrapper.UnderlyingObject != DeadLocWrapper.UnderlyingObject) |
| 2613 | continue; |
| 2614 | LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: " |
| 2615 | << *DeadLocWrapper.DefInst << "\n KILLER: " |
| 2616 | << *KillingLocWrapper.DefInst << '\n'); |
| 2617 | deleteDeadInstruction(SI: DeadLocWrapper.DefInst, Deleted: &Deleted); |
| 2618 | ++NumFastStores; |
| 2619 | Changed = true; |
| 2620 | } else { |
| 2621 | // Check if DeadI overwrites KillingI. |
| 2622 | int64_t KillingOffset = 0; |
| 2623 | int64_t DeadOffset = 0; |
| 2624 | OverwriteResult OR = |
| 2625 | isOverwrite(KillingI: KillingLocWrapper.DefInst, DeadI: DeadLocWrapper.DefInst, |
| 2626 | KillingLoc: KillingLocWrapper.MemLoc, DeadLoc: DeadLocWrapper.MemLoc, |
| 2627 | KillingOff&: KillingOffset, DeadOff&: DeadOffset); |
| 2628 | if (OR == OW_MaybePartial) { |
| 2629 | auto &IOL = IOLs[DeadLocWrapper.DefInst->getParent()]; |
| 2630 | OR = isPartialOverwrite(Opts, KillingLoc: KillingLocWrapper.MemLoc, |
| 2631 | DeadLoc: DeadLocWrapper.MemLoc, KillingOff: KillingOffset, |
| 2632 | DeadOff: DeadOffset, DeadI: DeadLocWrapper.DefInst, IOL); |
| 2633 | } |
| 2634 | if (Opts.enable_dse_partial_store_merging && |
| 2635 | OR == OW_PartialEarlierWithFullLater) { |
| 2636 | auto *DeadSI = dyn_cast<StoreInst>(Val: DeadLocWrapper.DefInst); |
| 2637 | auto *KillingSI = dyn_cast<StoreInst>(Val: KillingLocWrapper.DefInst); |
| 2638 | // We are re-using tryToMergePartialOverlappingStores, which requires |
| 2639 | // DeadSI to dominate KillingSI. |
| 2640 | // TODO: implement tryToMergeParialOverlappingStores using MemorySSA. |
| 2641 | if (DeadSI && KillingSI && DT.dominates(Def: DeadSI, User: KillingSI)) { |
| 2642 | if (Constant *Merged = tryToMergePartialOverlappingStores( |
| 2643 | KillingI: KillingSI, DeadI: DeadSI, KillingOffset, DeadOffset, DL, AA&: BatchAA, |
| 2644 | DT: &DT)) { |
| 2645 | |
| 2646 | // Update stored value of earlier store to merged constant. |
| 2647 | DeadSI->setOperand(i_nocapture: 0, Val_nocapture: Merged); |
| 2648 | ++NumModifiedStores; |
| 2649 | Changed = true; |
| 2650 | DeletedKillingLoc = true; |
| 2651 | |
| 2652 | // Remove killing store and remove any outstanding overlap |
| 2653 | // intervals for the updated store. |
| 2654 | deleteDeadInstruction(SI: KillingSI, Deleted: &Deleted); |
| 2655 | auto I = IOLs.find(Key: DeadSI->getParent()); |
| 2656 | if (I != IOLs.end()) |
| 2657 | I->second.erase(Key: DeadSI); |
| 2658 | break; |
| 2659 | } |
| 2660 | } |
| 2661 | } |
| 2662 | if (OR == OW_Complete) { |
| 2663 | LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: " |
| 2664 | << *DeadLocWrapper.DefInst << "\n KILLER: " |
| 2665 | << *KillingLocWrapper.DefInst << '\n'); |
| 2666 | deleteDeadInstruction(SI: DeadLocWrapper.DefInst, Deleted: &Deleted); |
| 2667 | ++NumFastStores; |
| 2668 | Changed = true; |
| 2669 | } |
| 2670 | } |
| 2671 | } |
| 2672 | |
| 2673 | assert(SkipStores.size() - OrigNumSkipStores == Deleted.size() && |
| 2674 | "SkipStores and Deleted out of sync?"); |
| 2675 | |
| 2676 | return {Changed, DeletedKillingLoc}; |
| 2677 | } |
| 2678 | |
| 2679 | bool DSEState::eliminateDeadDefs(const MemoryDefWrapper &KillingDefWrapper) { |
| 2680 | if (KillingDefWrapper.DefinedLocations.empty()) { |
| 2681 | LLVM_DEBUG(dbgs() << "Failed to find analyzable write location for " |
| 2682 | << *KillingDefWrapper.DefInst << "\n"); |
| 2683 | return false; |
| 2684 | } |
| 2685 | |
| 2686 | bool MadeChange = false; |
| 2687 | for (auto &KillingLocWrapper : KillingDefWrapper.DefinedLocations) { |
| 2688 | LLVM_DEBUG(dbgs() << "Trying to eliminate MemoryDefs killed by " |
| 2689 | << *KillingLocWrapper.MemDef << " (" |
| 2690 | << *KillingLocWrapper.DefInst << ")\n"); |
| 2691 | auto [Changed, DeletedKillingLoc] = eliminateDeadDefs(KillingLocWrapper); |
| 2692 | MadeChange |= Changed; |
| 2693 | |
| 2694 | // Check if the store is a no-op. |
| 2695 | if (!DeletedKillingLoc && storeIsNoop(Def: KillingLocWrapper.MemDef, |
| 2696 | DefUO: KillingLocWrapper.UnderlyingObject)) { |
| 2697 | LLVM_DEBUG(dbgs() << "DSE: Remove No-Op Store:\n DEAD: " |
| 2698 | << *KillingLocWrapper.DefInst << '\n'); |
| 2699 | deleteDeadInstruction(SI: KillingLocWrapper.DefInst); |
| 2700 | NumRedundantStores++; |
| 2701 | MadeChange = true; |
| 2702 | continue; |
| 2703 | } |
| 2704 | // Can we form a calloc from a memset/malloc pair? |
| 2705 | if (!DeletedKillingLoc && |
| 2706 | tryFoldIntoCalloc(Def: KillingLocWrapper.MemDef, |
| 2707 | DefUO: KillingLocWrapper.UnderlyingObject)) { |
| 2708 | LLVM_DEBUG(dbgs() << "DSE: Remove memset after forming calloc:\n" |
| 2709 | << " DEAD: "<< *KillingLocWrapper.DefInst << '\n'); |
| 2710 | deleteDeadInstruction(SI: KillingLocWrapper.DefInst); |
| 2711 | MadeChange = true; |
| 2712 | continue; |
| 2713 | } |
| 2714 | } |
| 2715 | return MadeChange; |
| 2716 | } |
| 2717 | |
| 2718 | static bool eliminateDeadStores(Function &F, AliasAnalysis &AA, MemorySSA &MSSA, |
| 2719 | DominatorTree &DT, PostDominatorTree &PDT, |
| 2720 | const TargetLibraryInfo &TLI, |
| 2721 | const CycleInfo &CI) { |
| 2722 | const ScalarOptions &Opts = |
| 2723 | F.getContext().getLibraryOptions().get<ScalarOptions>(); |
| 2724 | bool MadeChange = false; |
| 2725 | DSEState State(Opts, F, AA, MSSA, DT, PDT, TLI, CI); |
| 2726 | // For each store: |
| 2727 | for (unsigned I = 0; I < State.MemDefs.size(); I++) { |
| 2728 | MemoryDef *KillingDef = State.MemDefs[I]; |
| 2729 | if (State.SkipStores.count(Ptr: KillingDef)) |
| 2730 | continue; |
| 2731 | |
| 2732 | MemoryDefWrapper KillingDefWrapper( |
| 2733 | KillingDef, |
| 2734 | State.getLocForInst(I: KillingDef->getMemoryInst(), |
| 2735 | ConsiderInitializesAttr: Opts.enable_dse_initializes_attr_improvement)); |
| 2736 | MadeChange |= State.eliminateDeadDefs(KillingDefWrapper); |
| 2737 | } |
| 2738 | |
| 2739 | if (Opts.enable_dse_partial_overwrite_tracking) |
| 2740 | for (auto &KV : State.IOLs) |
| 2741 | MadeChange |= State.removePartiallyOverlappedStores(IOL&: KV.second); |
| 2742 | |
| 2743 | MadeChange |= State.eliminateRedundantStoresOfExistingValues(); |
| 2744 | MadeChange |= State.eliminateDeadWritesAtEndOfFunction(); |
| 2745 | MadeChange |= State.eliminateRedundantStoresViaDominatingConditions(); |
| 2746 | |
| 2747 | while (!State.ToRemove.empty()) { |
| 2748 | Instruction *DeadInst = State.ToRemove.pop_back_val(); |
| 2749 | DeadInst->eraseFromParent(); |
| 2750 | } |
| 2751 | |
| 2752 | return MadeChange; |
| 2753 | } |
| 2754 | |
| 2755 | //===----------------------------------------------------------------------===// |
| 2756 | // DSE Pass |
| 2757 | //===----------------------------------------------------------------------===// |
| 2758 | PreservedAnalyses DSEPass::run(Function &F, FunctionAnalysisManager &AM) { |
| 2759 | AliasAnalysis &AA = AM.getResult<AAManager>(IR&: F); |
| 2760 | const TargetLibraryInfo &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F); |
| 2761 | DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 2762 | MemorySSA &MSSA = AM.getResult<MemorySSAAnalysis>(IR&: F).getMSSA(); |
| 2763 | PostDominatorTree &PDT = AM.getResult<PostDominatorTreeAnalysis>(IR&: F); |
| 2764 | CycleInfo &CI = AM.getResult<CycleAnalysis>(IR&: F); |
| 2765 | |
| 2766 | bool Changed = eliminateDeadStores(F, AA, MSSA, DT, PDT, TLI, CI); |
| 2767 | |
| 2768 | #ifdef LLVM_ENABLE_STATS |
| 2769 | if (AreStatisticsEnabled()) |
| 2770 | for (auto &I : instructions(F)) |
| 2771 | NumRemainingStores += isa<StoreInst>(Val: &I); |
| 2772 | #endif |
| 2773 | |
| 2774 | if (!Changed) |
| 2775 | return PreservedAnalyses::all(); |
| 2776 | |
| 2777 | PreservedAnalyses PA; |
| 2778 | PA.preserveSet<CFGAnalyses>(); |
| 2779 | PA.preserve<MemorySSAAnalysis>(); |
| 2780 | return PA; |
| 2781 | } |
| 2782 | |
| 2783 | namespace { |
| 2784 | |
| 2785 | /// A legacy pass for the legacy pass manager that wraps \c DSEPass. |
| 2786 | class DSELegacyPass : public FunctionPass { |
| 2787 | public: |
| 2788 | static char ID; // Pass identification, replacement for typeid |
| 2789 | |
| 2790 | DSELegacyPass() : FunctionPass(ID) { |
| 2791 | initializeDSELegacyPassPass(*PassRegistry::getPassRegistry()); |
| 2792 | } |
| 2793 | |
| 2794 | bool runOnFunction(Function &F) override { |
| 2795 | if (skipFunction(F)) |
| 2796 | return false; |
| 2797 | |
| 2798 | AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); |
| 2799 | DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
| 2800 | const TargetLibraryInfo &TLI = |
| 2801 | getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); |
| 2802 | MemorySSA &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA(); |
| 2803 | PostDominatorTree &PDT = |
| 2804 | getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree(); |
| 2805 | CycleInfo &CI = getAnalysis<CycleInfoWrapperPass>().getResult(); |
| 2806 | |
| 2807 | bool Changed = eliminateDeadStores(F, AA, MSSA, DT, PDT, TLI, CI); |
| 2808 | |
| 2809 | #ifdef LLVM_ENABLE_STATS |
| 2810 | if (AreStatisticsEnabled()) |
| 2811 | for (auto &I : instructions(F)) |
| 2812 | NumRemainingStores += isa<StoreInst>(Val: &I); |
| 2813 | #endif |
| 2814 | |
| 2815 | return Changed; |
| 2816 | } |
| 2817 | |
| 2818 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 2819 | AU.setPreservesCFG(); |
| 2820 | AU.addRequired<AAResultsWrapperPass>(); |
| 2821 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
| 2822 | AU.addPreserved<GlobalsAAWrapperPass>(); |
| 2823 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 2824 | AU.addRequired<PostDominatorTreeWrapperPass>(); |
| 2825 | AU.addRequired<MemorySSAWrapperPass>(); |
| 2826 | AU.addPreserved<MemorySSAWrapperPass>(); |
| 2827 | AU.addRequired<CycleInfoWrapperPass>(); |
| 2828 | AU.addRequired<AssumptionCacheTracker>(); |
| 2829 | } |
| 2830 | }; |
| 2831 | |
| 2832 | } // end anonymous namespace |
| 2833 | |
| 2834 | char DSELegacyPass::ID = 0; |
| 2835 | |
| 2836 | INITIALIZE_PASS_BEGIN(DSELegacyPass, "dse", "Dead Store Elimination", false, |
| 2837 | false) |
| 2838 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
| 2839 | INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass) |
| 2840 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
| 2841 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) |
| 2842 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) |
| 2843 | INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass) |
| 2844 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
| 2845 | INITIALIZE_PASS_DEPENDENCY(CycleInfoWrapperPass) |
| 2846 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) |
| 2847 | INITIALIZE_PASS_END(DSELegacyPass, "dse", "Dead Store Elimination", false, |
| 2848 | false) |
| 2849 | |
| 2850 | LLVM_ABI FunctionPass *llvm::createDeadStoreEliminationPass() { |
| 2851 | return new DSELegacyPass(); |
| 2852 | } |
| 2853 |
Generated on 2026-Oct-20 from project llvm_projects revision 7be7bc433
Powered by 
Code Browser 2.1
Generator usage only permitted with license.