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 "llvm/ADT/APInt.h" 32 #include "llvm/ADT/DenseMap.h" 33 #include "llvm/ADT/MapVector.h" 34 #include "llvm/ADT/PostOrderIterator.h" 35 #include "llvm/ADT/SetVector.h" 36 #include "llvm/ADT/SmallPtrSet.h" 37 #include "llvm/ADT/SmallVector.h" 38 #include "llvm/ADT/Statistic.h" 39 #include "llvm/ADT/StringRef.h" 40 #include "llvm/Analysis/AliasAnalysis.h" 41 #include "llvm/Analysis/CaptureTracking.h" 42 #include "llvm/Analysis/GlobalsModRef.h" 43 #include "llvm/Analysis/LoopInfo.h" 44 #include "llvm/Analysis/MemoryBuiltins.h" 45 #include "llvm/Analysis/MemoryLocation.h" 46 #include "llvm/Analysis/MemorySSA.h" 47 #include "llvm/Analysis/MemorySSAUpdater.h" 48 #include "llvm/Analysis/MustExecute.h" 49 #include "llvm/Analysis/PostDominators.h" 50 #include "llvm/Analysis/TargetLibraryInfo.h" 51 #include "llvm/Analysis/ValueTracking.h" 52 #include "llvm/IR/Argument.h" 53 #include "llvm/IR/BasicBlock.h" 54 #include "llvm/IR/Constant.h" 55 #include "llvm/IR/Constants.h" 56 #include "llvm/IR/DataLayout.h" 57 #include "llvm/IR/DebugInfo.h" 58 #include "llvm/IR/Dominators.h" 59 #include "llvm/IR/Function.h" 60 #include "llvm/IR/IRBuilder.h" 61 #include "llvm/IR/InstIterator.h" 62 #include "llvm/IR/InstrTypes.h" 63 #include "llvm/IR/Instruction.h" 64 #include "llvm/IR/Instructions.h" 65 #include "llvm/IR/IntrinsicInst.h" 66 #include "llvm/IR/Module.h" 67 #include "llvm/IR/PassManager.h" 68 #include "llvm/IR/PatternMatch.h" 69 #include "llvm/IR/Value.h" 70 #include "llvm/Support/Casting.h" 71 #include "llvm/Support/CommandLine.h" 72 #include "llvm/Support/Debug.h" 73 #include "llvm/Support/DebugCounter.h" 74 #include "llvm/Support/ErrorHandling.h" 75 #include "llvm/Support/raw_ostream.h" 76 #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" 77 #include "llvm/Transforms/Utils/BuildLibCalls.h" 78 #include "llvm/Transforms/Utils/Local.h" 79 #include <algorithm> 80 #include <cassert> 81 #include <cstdint> 82 #include <iterator> 83 #include <map> 84 #include <optional> 85 #include <utility> 86 87 using namespace llvm; 88 using namespace PatternMatch; 89 90 #define DEBUG_TYPE "dse" 91 92 STATISTIC(NumRemainingStores, "Number of stores remaining after DSE"); 93 STATISTIC(NumRedundantStores, "Number of redundant stores deleted"); 94 STATISTIC(NumFastStores, "Number of stores deleted"); 95 STATISTIC(NumFastOther, "Number of other instrs removed"); 96 STATISTIC(NumCompletePartials, "Number of stores dead by later partials"); 97 STATISTIC(NumModifiedStores, "Number of stores modified"); 98 STATISTIC(NumCFGChecks, "Number of stores modified"); 99 STATISTIC(NumCFGTries, "Number of stores modified"); 100 STATISTIC(NumCFGSuccess, "Number of stores modified"); 101 STATISTIC(NumGetDomMemoryDefPassed, 102 "Number of times a valid candidate is returned from getDomMemoryDef"); 103 STATISTIC(NumDomMemDefChecks, 104 "Number iterations check for reads in getDomMemoryDef"); 105 106 DEBUG_COUNTER(MemorySSACounter, "dse-memoryssa", 107 "Controls which MemoryDefs are eliminated."); 108 109 static cl::opt<bool> 110 EnablePartialOverwriteTracking("enable-dse-partial-overwrite-tracking", 111 cl::init(true), cl::Hidden, 112 cl::desc("Enable partial-overwrite tracking in DSE")); 113 114 static cl::opt<bool> 115 EnablePartialStoreMerging("enable-dse-partial-store-merging", 116 cl::init(true), cl::Hidden, 117 cl::desc("Enable partial store merging in DSE")); 118 119 static cl::opt<unsigned> 120 MemorySSAScanLimit("dse-memoryssa-scanlimit", cl::init(150), cl::Hidden, 121 cl::desc("The number of memory instructions to scan for " 122 "dead store elimination (default = 150)")); 123 static cl::opt<unsigned> MemorySSAUpwardsStepLimit( 124 "dse-memoryssa-walklimit", cl::init(90), cl::Hidden, 125 cl::desc("The maximum number of steps while walking upwards to find " 126 "MemoryDefs that may be killed (default = 90)")); 127 128 static cl::opt<unsigned> MemorySSAPartialStoreLimit( 129 "dse-memoryssa-partial-store-limit", cl::init(5), cl::Hidden, 130 cl::desc("The maximum number candidates that only partially overwrite the " 131 "killing MemoryDef to consider" 132 " (default = 5)")); 133 134 static cl::opt<unsigned> MemorySSADefsPerBlockLimit( 135 "dse-memoryssa-defs-per-block-limit", cl::init(5000), cl::Hidden, 136 cl::desc("The number of MemoryDefs we consider as candidates to eliminated " 137 "other stores per basic block (default = 5000)")); 138 139 static cl::opt<unsigned> MemorySSASameBBStepCost( 140 "dse-memoryssa-samebb-cost", cl::init(1), cl::Hidden, 141 cl::desc( 142 "The cost of a step in the same basic block as the killing MemoryDef" 143 "(default = 1)")); 144 145 static cl::opt<unsigned> 146 MemorySSAOtherBBStepCost("dse-memoryssa-otherbb-cost", cl::init(5), 147 cl::Hidden, 148 cl::desc("The cost of a step in a different basic " 149 "block than the killing MemoryDef" 150 "(default = 5)")); 151 152 static cl::opt<unsigned> MemorySSAPathCheckLimit( 153 "dse-memoryssa-path-check-limit", cl::init(50), cl::Hidden, 154 cl::desc("The maximum number of blocks to check when trying to prove that " 155 "all paths to an exit go through a killing block (default = 50)")); 156 157 // This flags allows or disallows DSE to optimize MemorySSA during its 158 // traversal. Note that DSE optimizing MemorySSA may impact other passes 159 // downstream of the DSE invocation and can lead to issues not being 160 // reproducible in isolation (i.e. when MemorySSA is built from scratch). In 161 // those cases, the flag can be used to check if DSE's MemorySSA optimizations 162 // impact follow-up passes. 163 static cl::opt<bool> 164 OptimizeMemorySSA("dse-optimize-memoryssa", cl::init(true), cl::Hidden, 165 cl::desc("Allow DSE to optimize memory accesses.")); 166 167 //===----------------------------------------------------------------------===// 168 // Helper functions 169 //===----------------------------------------------------------------------===// 170 using OverlapIntervalsTy = std::map<int64_t, int64_t>; 171 using InstOverlapIntervalsTy = DenseMap<Instruction *, OverlapIntervalsTy>; 172 173 /// Returns true if the end of this instruction can be safely shortened in 174 /// length. 175 static bool isShortenableAtTheEnd(Instruction *I) { 176 // Don't shorten stores for now 177 if (isa<StoreInst>(I)) 178 return false; 179 180 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 181 switch (II->getIntrinsicID()) { 182 default: return false; 183 case Intrinsic::memset: 184 case Intrinsic::memcpy: 185 case Intrinsic::memcpy_element_unordered_atomic: 186 case Intrinsic::memset_element_unordered_atomic: 187 // Do shorten memory intrinsics. 188 // FIXME: Add memmove if it's also safe to transform. 189 return true; 190 } 191 } 192 193 // Don't shorten libcalls calls for now. 194 195 return false; 196 } 197 198 /// Returns true if the beginning of this instruction can be safely shortened 199 /// in length. 200 static bool isShortenableAtTheBeginning(Instruction *I) { 201 // FIXME: Handle only memset for now. Supporting memcpy/memmove should be 202 // easily done by offsetting the source address. 203 return isa<AnyMemSetInst>(I); 204 } 205 206 static std::optional<TypeSize> getPointerSize(const Value *V, 207 const DataLayout &DL, 208 const TargetLibraryInfo &TLI, 209 const Function *F) { 210 uint64_t Size; 211 ObjectSizeOpts Opts; 212 Opts.NullIsUnknownSize = NullPointerIsDefined(F); 213 214 if (getObjectSize(V, Size, DL, &TLI, Opts)) 215 return TypeSize::getFixed(Size); 216 return std::nullopt; 217 } 218 219 namespace { 220 221 enum OverwriteResult { 222 OW_Begin, 223 OW_Complete, 224 OW_End, 225 OW_PartialEarlierWithFullLater, 226 OW_MaybePartial, 227 OW_None, 228 OW_Unknown 229 }; 230 231 } // end anonymous namespace 232 233 /// Check if two instruction are masked stores that completely 234 /// overwrite one another. More specifically, \p KillingI has to 235 /// overwrite \p DeadI. 236 static OverwriteResult isMaskedStoreOverwrite(const Instruction *KillingI, 237 const Instruction *DeadI, 238 BatchAAResults &AA) { 239 const auto *KillingII = dyn_cast<IntrinsicInst>(KillingI); 240 const auto *DeadII = dyn_cast<IntrinsicInst>(DeadI); 241 if (KillingII == nullptr || DeadII == nullptr) 242 return OW_Unknown; 243 if (KillingII->getIntrinsicID() != DeadII->getIntrinsicID()) 244 return OW_Unknown; 245 if (KillingII->getIntrinsicID() == Intrinsic::masked_store) { 246 // Type size. 247 VectorType *KillingTy = 248 cast<VectorType>(KillingII->getArgOperand(0)->getType()); 249 VectorType *DeadTy = cast<VectorType>(DeadII->getArgOperand(0)->getType()); 250 if (KillingTy->getScalarSizeInBits() != DeadTy->getScalarSizeInBits()) 251 return OW_Unknown; 252 // Element count. 253 if (KillingTy->getElementCount() != DeadTy->getElementCount()) 254 return OW_Unknown; 255 // Pointers. 256 Value *KillingPtr = KillingII->getArgOperand(1)->stripPointerCasts(); 257 Value *DeadPtr = DeadII->getArgOperand(1)->stripPointerCasts(); 258 if (KillingPtr != DeadPtr && !AA.isMustAlias(KillingPtr, DeadPtr)) 259 return OW_Unknown; 260 // Masks. 261 // TODO: check that KillingII's mask is a superset of the DeadII's mask. 262 if (KillingII->getArgOperand(3) != DeadII->getArgOperand(3)) 263 return OW_Unknown; 264 return OW_Complete; 265 } 266 return OW_Unknown; 267 } 268 269 /// Return 'OW_Complete' if a store to the 'KillingLoc' location completely 270 /// overwrites a store to the 'DeadLoc' location, 'OW_End' if the end of the 271 /// 'DeadLoc' location is completely overwritten by 'KillingLoc', 'OW_Begin' 272 /// if the beginning of the 'DeadLoc' location is overwritten by 'KillingLoc'. 273 /// 'OW_PartialEarlierWithFullLater' means that a dead (big) store was 274 /// overwritten by a killing (smaller) store which doesn't write outside the big 275 /// store's memory locations. Returns 'OW_Unknown' if nothing can be determined. 276 /// NOTE: This function must only be called if both \p KillingLoc and \p 277 /// DeadLoc belong to the same underlying object with valid \p KillingOff and 278 /// \p DeadOff. 279 static OverwriteResult isPartialOverwrite(const MemoryLocation &KillingLoc, 280 const MemoryLocation &DeadLoc, 281 int64_t KillingOff, int64_t DeadOff, 282 Instruction *DeadI, 283 InstOverlapIntervalsTy &IOL) { 284 const uint64_t KillingSize = KillingLoc.Size.getValue(); 285 const uint64_t DeadSize = DeadLoc.Size.getValue(); 286 // We may now overlap, although the overlap is not complete. There might also 287 // be other incomplete overlaps, and together, they might cover the complete 288 // dead store. 289 // Note: The correctness of this logic depends on the fact that this function 290 // is not even called providing DepWrite when there are any intervening reads. 291 if (EnablePartialOverwriteTracking && 292 KillingOff < int64_t(DeadOff + DeadSize) && 293 int64_t(KillingOff + KillingSize) >= DeadOff) { 294 295 // Insert our part of the overlap into the map. 296 auto &IM = IOL[DeadI]; 297 LLVM_DEBUG(dbgs() << "DSE: Partial overwrite: DeadLoc [" << DeadOff << ", " 298 << int64_t(DeadOff + DeadSize) << ") KillingLoc [" 299 << KillingOff << ", " << int64_t(KillingOff + KillingSize) 300 << ")\n"); 301 302 // Make sure that we only insert non-overlapping intervals and combine 303 // adjacent intervals. The intervals are stored in the map with the ending 304 // offset as the key (in the half-open sense) and the starting offset as 305 // the value. 306 int64_t KillingIntStart = KillingOff; 307 int64_t KillingIntEnd = KillingOff + KillingSize; 308 309 // Find any intervals ending at, or after, KillingIntStart which start 310 // before KillingIntEnd. 311 auto ILI = IM.lower_bound(KillingIntStart); 312 if (ILI != IM.end() && ILI->second <= KillingIntEnd) { 313 // This existing interval is overlapped with the current store somewhere 314 // in [KillingIntStart, KillingIntEnd]. Merge them by erasing the existing 315 // intervals and adjusting our start and end. 316 KillingIntStart = std::min(KillingIntStart, ILI->second); 317 KillingIntEnd = std::max(KillingIntEnd, ILI->first); 318 ILI = IM.erase(ILI); 319 320 // Continue erasing and adjusting our end in case other previous 321 // intervals are also overlapped with the current store. 322 // 323 // |--- dead 1 ---| |--- dead 2 ---| 324 // |------- killing---------| 325 // 326 while (ILI != IM.end() && ILI->second <= KillingIntEnd) { 327 assert(ILI->second > KillingIntStart && "Unexpected interval"); 328 KillingIntEnd = std::max(KillingIntEnd, ILI->first); 329 ILI = IM.erase(ILI); 330 } 331 } 332 333 IM[KillingIntEnd] = KillingIntStart; 334 335 ILI = IM.begin(); 336 if (ILI->second <= DeadOff && ILI->first >= int64_t(DeadOff + DeadSize)) { 337 LLVM_DEBUG(dbgs() << "DSE: Full overwrite from partials: DeadLoc [" 338 << DeadOff << ", " << int64_t(DeadOff + DeadSize) 339 << ") Composite KillingLoc [" << ILI->second << ", " 340 << ILI->first << ")\n"); 341 ++NumCompletePartials; 342 return OW_Complete; 343 } 344 } 345 346 // Check for a dead store which writes to all the memory locations that 347 // the killing store writes to. 348 if (EnablePartialStoreMerging && KillingOff >= DeadOff && 349 int64_t(DeadOff + DeadSize) > KillingOff && 350 uint64_t(KillingOff - DeadOff) + KillingSize <= DeadSize) { 351 LLVM_DEBUG(dbgs() << "DSE: Partial overwrite a dead load [" << DeadOff 352 << ", " << int64_t(DeadOff + DeadSize) 353 << ") by a killing store [" << KillingOff << ", " 354 << int64_t(KillingOff + KillingSize) << ")\n"); 355 // TODO: Maybe come up with a better name? 356 return OW_PartialEarlierWithFullLater; 357 } 358 359 // Another interesting case is if the killing store overwrites the end of the 360 // dead store. 361 // 362 // |--dead--| 363 // |-- killing --| 364 // 365 // In this case we may want to trim the size of dead store to avoid 366 // generating stores to addresses which will definitely be overwritten killing 367 // store. 368 if (!EnablePartialOverwriteTracking && 369 (KillingOff > DeadOff && KillingOff < int64_t(DeadOff + DeadSize) && 370 int64_t(KillingOff + KillingSize) >= int64_t(DeadOff + DeadSize))) 371 return OW_End; 372 373 // Finally, we also need to check if the killing store overwrites the 374 // beginning of the dead store. 375 // 376 // |--dead--| 377 // |-- killing --| 378 // 379 // In this case we may want to move the destination address and trim the size 380 // of dead store to avoid generating stores to addresses which will definitely 381 // be overwritten killing store. 382 if (!EnablePartialOverwriteTracking && 383 (KillingOff <= DeadOff && int64_t(KillingOff + KillingSize) > DeadOff)) { 384 assert(int64_t(KillingOff + KillingSize) < int64_t(DeadOff + DeadSize) && 385 "Expect to be handled as OW_Complete"); 386 return OW_Begin; 387 } 388 // Otherwise, they don't completely overlap. 389 return OW_Unknown; 390 } 391 392 /// Returns true if the memory which is accessed by the second instruction is not 393 /// modified between the first and the second instruction. 394 /// Precondition: Second instruction must be dominated by the first 395 /// instruction. 396 static bool 397 memoryIsNotModifiedBetween(Instruction *FirstI, Instruction *SecondI, 398 BatchAAResults &AA, const DataLayout &DL, 399 DominatorTree *DT) { 400 // Do a backwards scan through the CFG from SecondI to FirstI. Look for 401 // instructions which can modify the memory location accessed by SecondI. 402 // 403 // While doing the walk keep track of the address to check. It might be 404 // different in different basic blocks due to PHI translation. 405 using BlockAddressPair = std::pair<BasicBlock *, PHITransAddr>; 406 SmallVector<BlockAddressPair, 16> WorkList; 407 // Keep track of the address we visited each block with. Bail out if we 408 // visit a block with different addresses. 409 DenseMap<BasicBlock *, Value *> Visited; 410 411 BasicBlock::iterator FirstBBI(FirstI); 412 ++FirstBBI; 413 BasicBlock::iterator SecondBBI(SecondI); 414 BasicBlock *FirstBB = FirstI->getParent(); 415 BasicBlock *SecondBB = SecondI->getParent(); 416 MemoryLocation MemLoc; 417 if (auto *MemSet = dyn_cast<MemSetInst>(SecondI)) 418 MemLoc = MemoryLocation::getForDest(MemSet); 419 else 420 MemLoc = MemoryLocation::get(SecondI); 421 422 auto *MemLocPtr = const_cast<Value *>(MemLoc.Ptr); 423 424 // Start checking the SecondBB. 425 WorkList.push_back( 426 std::make_pair(SecondBB, PHITransAddr(MemLocPtr, DL, nullptr))); 427 bool isFirstBlock = true; 428 429 // Check all blocks going backward until we reach the FirstBB. 430 while (!WorkList.empty()) { 431 BlockAddressPair Current = WorkList.pop_back_val(); 432 BasicBlock *B = Current.first; 433 PHITransAddr &Addr = Current.second; 434 Value *Ptr = Addr.getAddr(); 435 436 // Ignore instructions before FirstI if this is the FirstBB. 437 BasicBlock::iterator BI = (B == FirstBB ? FirstBBI : B->begin()); 438 439 BasicBlock::iterator EI; 440 if (isFirstBlock) { 441 // Ignore instructions after SecondI if this is the first visit of SecondBB. 442 assert(B == SecondBB && "first block is not the store block"); 443 EI = SecondBBI; 444 isFirstBlock = false; 445 } else { 446 // It's not SecondBB or (in case of a loop) the second visit of SecondBB. 447 // In this case we also have to look at instructions after SecondI. 448 EI = B->end(); 449 } 450 for (; BI != EI; ++BI) { 451 Instruction *I = &*BI; 452 if (I->mayWriteToMemory() && I != SecondI) 453 if (isModSet(AA.getModRefInfo(I, MemLoc.getWithNewPtr(Ptr)))) 454 return false; 455 } 456 if (B != FirstBB) { 457 assert(B != &FirstBB->getParent()->getEntryBlock() && 458 "Should not hit the entry block because SI must be dominated by LI"); 459 for (BasicBlock *Pred : predecessors(B)) { 460 PHITransAddr PredAddr = Addr; 461 if (PredAddr.needsPHITranslationFromBlock(B)) { 462 if (!PredAddr.isPotentiallyPHITranslatable()) 463 return false; 464 if (!PredAddr.translateValue(B, Pred, DT, false)) 465 return false; 466 } 467 Value *TranslatedPtr = PredAddr.getAddr(); 468 auto Inserted = Visited.insert(std::make_pair(Pred, TranslatedPtr)); 469 if (!Inserted.second) { 470 // We already visited this block before. If it was with a different 471 // address - bail out! 472 if (TranslatedPtr != Inserted.first->second) 473 return false; 474 // ... otherwise just skip it. 475 continue; 476 } 477 WorkList.push_back(std::make_pair(Pred, PredAddr)); 478 } 479 } 480 } 481 return true; 482 } 483 484 static void shortenAssignment(Instruction *Inst, Value *OriginalDest, 485 uint64_t OldOffsetInBits, uint64_t OldSizeInBits, 486 uint64_t NewSizeInBits, bool IsOverwriteEnd) { 487 const DataLayout &DL = Inst->getDataLayout(); 488 uint64_t DeadSliceSizeInBits = OldSizeInBits - NewSizeInBits; 489 uint64_t DeadSliceOffsetInBits = 490 OldOffsetInBits + (IsOverwriteEnd ? NewSizeInBits : 0); 491 auto SetDeadFragExpr = [](auto *Assign, 492 DIExpression::FragmentInfo DeadFragment) { 493 // createFragmentExpression expects an offset relative to the existing 494 // fragment offset if there is one. 495 uint64_t RelativeOffset = DeadFragment.OffsetInBits - 496 Assign->getExpression() 497 ->getFragmentInfo() 498 .value_or(DIExpression::FragmentInfo(0, 0)) 499 .OffsetInBits; 500 if (auto NewExpr = DIExpression::createFragmentExpression( 501 Assign->getExpression(), RelativeOffset, DeadFragment.SizeInBits)) { 502 Assign->setExpression(*NewExpr); 503 return; 504 } 505 // Failed to create a fragment expression for this so discard the value, 506 // making this a kill location. 507 auto *Expr = *DIExpression::createFragmentExpression( 508 DIExpression::get(Assign->getContext(), std::nullopt), 509 DeadFragment.OffsetInBits, DeadFragment.SizeInBits); 510 Assign->setExpression(Expr); 511 Assign->setKillLocation(); 512 }; 513 514 // A DIAssignID to use so that the inserted dbg.assign intrinsics do not 515 // link to any instructions. Created in the loop below (once). 516 DIAssignID *LinkToNothing = nullptr; 517 LLVMContext &Ctx = Inst->getContext(); 518 auto GetDeadLink = [&Ctx, &LinkToNothing]() { 519 if (!LinkToNothing) 520 LinkToNothing = DIAssignID::getDistinct(Ctx); 521 return LinkToNothing; 522 }; 523 524 // Insert an unlinked dbg.assign intrinsic for the dead fragment after each 525 // overlapping dbg.assign intrinsic. The loop invalidates the iterators 526 // returned by getAssignmentMarkers so save a copy of the markers to iterate 527 // over. 528 auto LinkedRange = at::getAssignmentMarkers(Inst); 529 SmallVector<DbgVariableRecord *> LinkedDVRAssigns = 530 at::getDVRAssignmentMarkers(Inst); 531 SmallVector<DbgAssignIntrinsic *> Linked(LinkedRange.begin(), 532 LinkedRange.end()); 533 auto InsertAssignForOverlap = [&](auto *Assign) { 534 std::optional<DIExpression::FragmentInfo> NewFragment; 535 if (!at::calculateFragmentIntersect(DL, OriginalDest, DeadSliceOffsetInBits, 536 DeadSliceSizeInBits, Assign, 537 NewFragment) || 538 !NewFragment) { 539 // We couldn't calculate the intersecting fragment for some reason. Be 540 // cautious and unlink the whole assignment from the store. 541 Assign->setKillAddress(); 542 Assign->setAssignId(GetDeadLink()); 543 return; 544 } 545 // No intersect. 546 if (NewFragment->SizeInBits == 0) 547 return; 548 549 // Fragments overlap: insert a new dbg.assign for this dead part. 550 auto *NewAssign = static_cast<decltype(Assign)>(Assign->clone()); 551 NewAssign->insertAfter(Assign); 552 NewAssign->setAssignId(GetDeadLink()); 553 if (NewFragment) 554 SetDeadFragExpr(NewAssign, *NewFragment); 555 NewAssign->setKillAddress(); 556 }; 557 for_each(Linked, InsertAssignForOverlap); 558 for_each(LinkedDVRAssigns, InsertAssignForOverlap); 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>(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(uint64_t(KillingStart - DeadStart), 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(ToRemoveSize, 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 (auto *AMI = dyn_cast<AtomicMemIntrinsic>(DeadI)) { 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 = AMI->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 Value *DeadWriteLength = DeadIntrinsic->getLength(); 629 Value *TrimmedLength = ConstantInt::get(DeadWriteLength->getType(), NewSize); 630 DeadIntrinsic->setLength(TrimmedLength); 631 DeadIntrinsic->setDestAlignment(PrefAlign); 632 633 Value *OrigDest = DeadIntrinsic->getRawDest(); 634 if (!IsOverwriteEnd) { 635 Value *Indices[1] = { 636 ConstantInt::get(DeadWriteLength->getType(), ToRemoveSize)}; 637 Instruction *NewDestGEP = GetElementPtrInst::CreateInBounds( 638 Type::getInt8Ty(DeadIntrinsic->getContext()), OrigDest, Indices, "", 639 DeadI->getIterator()); 640 NewDestGEP->setDebugLoc(DeadIntrinsic->getDebugLoc()); 641 DeadIntrinsic->setDest(NewDestGEP); 642 } 643 644 // Update attached dbg.assign intrinsics. Assume 8-bit byte. 645 shortenAssignment(DeadI, OrigDest, DeadStart * 8, DeadSize * 8, NewSize * 8, 646 IsOverwriteEnd); 647 648 // Finally update start and size of dead access. 649 if (!IsOverwriteEnd) 650 DeadStart += ToRemoveSize; 651 DeadSize = NewSize; 652 653 return true; 654 } 655 656 static bool tryToShortenEnd(Instruction *DeadI, OverlapIntervalsTy &IntervalMap, 657 int64_t &DeadStart, uint64_t &DeadSize) { 658 if (IntervalMap.empty() || !isShortenableAtTheEnd(DeadI)) 659 return false; 660 661 OverlapIntervalsTy::iterator OII = --IntervalMap.end(); 662 int64_t KillingStart = OII->second; 663 uint64_t KillingSize = OII->first - KillingStart; 664 665 assert(OII->first - KillingStart >= 0 && "Size expected to be positive"); 666 667 if (KillingStart > DeadStart && 668 // Note: "KillingStart - KillingStart" is known to be positive due to 669 // preceding check. 670 (uint64_t)(KillingStart - DeadStart) < DeadSize && 671 // Note: "DeadSize - (uint64_t)(KillingStart - DeadStart)" is known to 672 // be non negative due to preceding checks. 673 KillingSize >= DeadSize - (uint64_t)(KillingStart - DeadStart)) { 674 if (tryToShorten(DeadI, DeadStart, DeadSize, KillingStart, KillingSize, 675 true)) { 676 IntervalMap.erase(OII); 677 return true; 678 } 679 } 680 return false; 681 } 682 683 static bool tryToShortenBegin(Instruction *DeadI, 684 OverlapIntervalsTy &IntervalMap, 685 int64_t &DeadStart, uint64_t &DeadSize) { 686 if (IntervalMap.empty() || !isShortenableAtTheBeginning(DeadI)) 687 return false; 688 689 OverlapIntervalsTy::iterator OII = IntervalMap.begin(); 690 int64_t KillingStart = OII->second; 691 uint64_t KillingSize = OII->first - KillingStart; 692 693 assert(OII->first - KillingStart >= 0 && "Size expected to be positive"); 694 695 if (KillingStart <= DeadStart && 696 // Note: "DeadStart - KillingStart" is known to be non negative due to 697 // preceding check. 698 KillingSize > (uint64_t)(DeadStart - KillingStart)) { 699 // Note: "KillingSize - (uint64_t)(DeadStart - DeadStart)" is known to 700 // be positive due to preceding checks. 701 assert(KillingSize - (uint64_t)(DeadStart - KillingStart) < DeadSize && 702 "Should have been handled as OW_Complete"); 703 if (tryToShorten(DeadI, DeadStart, DeadSize, KillingStart, KillingSize, 704 false)) { 705 IntervalMap.erase(OII); 706 return true; 707 } 708 } 709 return false; 710 } 711 712 static Constant * 713 tryToMergePartialOverlappingStores(StoreInst *KillingI, StoreInst *DeadI, 714 int64_t KillingOffset, int64_t DeadOffset, 715 const DataLayout &DL, BatchAAResults &AA, 716 DominatorTree *DT) { 717 718 if (DeadI && isa<ConstantInt>(DeadI->getValueOperand()) && 719 DL.typeSizeEqualsStoreSize(DeadI->getValueOperand()->getType()) && 720 KillingI && isa<ConstantInt>(KillingI->getValueOperand()) && 721 DL.typeSizeEqualsStoreSize(KillingI->getValueOperand()->getType()) && 722 memoryIsNotModifiedBetween(DeadI, KillingI, AA, DL, DT)) { 723 // If the store we find is: 724 // a) partially overwritten by the store to 'Loc' 725 // b) the killing store is fully contained in the dead one and 726 // c) they both have a constant value 727 // d) none of the two stores need padding 728 // Merge the two stores, replacing the dead store's value with a 729 // merge of both values. 730 // TODO: Deal with other constant types (vectors, etc), and probably 731 // some mem intrinsics (if needed) 732 733 APInt DeadValue = cast<ConstantInt>(DeadI->getValueOperand())->getValue(); 734 APInt KillingValue = 735 cast<ConstantInt>(KillingI->getValueOperand())->getValue(); 736 unsigned KillingBits = KillingValue.getBitWidth(); 737 assert(DeadValue.getBitWidth() > KillingValue.getBitWidth()); 738 KillingValue = KillingValue.zext(DeadValue.getBitWidth()); 739 740 // Offset of the smaller store inside the larger store 741 unsigned BitOffsetDiff = (KillingOffset - DeadOffset) * 8; 742 unsigned LShiftAmount = 743 DL.isBigEndian() ? DeadValue.getBitWidth() - BitOffsetDiff - KillingBits 744 : BitOffsetDiff; 745 APInt Mask = APInt::getBitsSet(DeadValue.getBitWidth(), LShiftAmount, 746 LShiftAmount + KillingBits); 747 // Clear the bits we'll be replacing, then OR with the smaller 748 // store, shifted appropriately. 749 APInt Merged = (DeadValue & ~Mask) | (KillingValue << LShiftAmount); 750 LLVM_DEBUG(dbgs() << "DSE: Merge Stores:\n Dead: " << *DeadI 751 << "\n Killing: " << *KillingI 752 << "\n Merged Value: " << Merged << '\n'); 753 return ConstantInt::get(DeadI->getValueOperand()->getType(), Merged); 754 } 755 return nullptr; 756 } 757 758 namespace { 759 // Returns true if \p I is an intrinsic that does not read or write memory. 760 bool isNoopIntrinsic(Instruction *I) { 761 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 762 switch (II->getIntrinsicID()) { 763 case Intrinsic::lifetime_start: 764 case Intrinsic::lifetime_end: 765 case Intrinsic::invariant_end: 766 case Intrinsic::launder_invariant_group: 767 case Intrinsic::assume: 768 return true; 769 case Intrinsic::dbg_declare: 770 case Intrinsic::dbg_label: 771 case Intrinsic::dbg_value: 772 llvm_unreachable("Intrinsic should not be modeled in MemorySSA"); 773 default: 774 return false; 775 } 776 } 777 return false; 778 } 779 780 // Check if we can ignore \p D for DSE. 781 bool canSkipDef(MemoryDef *D, bool DefVisibleToCaller) { 782 Instruction *DI = D->getMemoryInst(); 783 // Calls that only access inaccessible memory cannot read or write any memory 784 // locations we consider for elimination. 785 if (auto *CB = dyn_cast<CallBase>(DI)) 786 if (CB->onlyAccessesInaccessibleMemory()) 787 return true; 788 789 // We can eliminate stores to locations not visible to the caller across 790 // throwing instructions. 791 if (DI->mayThrow() && !DefVisibleToCaller) 792 return true; 793 794 // We can remove the dead stores, irrespective of the fence and its ordering 795 // (release/acquire/seq_cst). Fences only constraints the ordering of 796 // already visible stores, it does not make a store visible to other 797 // threads. So, skipping over a fence does not change a store from being 798 // dead. 799 if (isa<FenceInst>(DI)) 800 return true; 801 802 // Skip intrinsics that do not really read or modify memory. 803 if (isNoopIntrinsic(DI)) 804 return true; 805 806 return false; 807 } 808 809 struct DSEState { 810 Function &F; 811 AliasAnalysis &AA; 812 EarliestEscapeInfo EI; 813 814 /// The single BatchAA instance that is used to cache AA queries. It will 815 /// not be invalidated over the whole run. This is safe, because: 816 /// 1. Only memory writes are removed, so the alias cache for memory 817 /// locations remains valid. 818 /// 2. No new instructions are added (only instructions removed), so cached 819 /// information for a deleted value cannot be accessed by a re-used new 820 /// value pointer. 821 BatchAAResults BatchAA; 822 823 MemorySSA &MSSA; 824 DominatorTree &DT; 825 PostDominatorTree &PDT; 826 const TargetLibraryInfo &TLI; 827 const DataLayout &DL; 828 const LoopInfo &LI; 829 830 // Whether the function contains any irreducible control flow, useful for 831 // being accurately able to detect loops. 832 bool ContainsIrreducibleLoops; 833 834 // All MemoryDefs that potentially could kill other MemDefs. 835 SmallVector<MemoryDef *, 64> MemDefs; 836 // Any that should be skipped as they are already deleted 837 SmallPtrSet<MemoryAccess *, 4> SkipStores; 838 // Keep track whether a given object is captured before return or not. 839 DenseMap<const Value *, bool> CapturedBeforeReturn; 840 // Keep track of all of the objects that are invisible to the caller after 841 // the function returns. 842 DenseMap<const Value *, bool> InvisibleToCallerAfterRet; 843 // Keep track of blocks with throwing instructions not modeled in MemorySSA. 844 SmallPtrSet<BasicBlock *, 16> ThrowingBlocks; 845 // Post-order numbers for each basic block. Used to figure out if memory 846 // accesses are executed before another access. 847 DenseMap<BasicBlock *, unsigned> PostOrderNumbers; 848 849 /// Keep track of instructions (partly) overlapping with killing MemoryDefs per 850 /// basic block. 851 MapVector<BasicBlock *, InstOverlapIntervalsTy> IOLs; 852 // Check if there are root nodes that are terminated by UnreachableInst. 853 // Those roots pessimize post-dominance queries. If there are such roots, 854 // fall back to CFG scan starting from all non-unreachable roots. 855 bool AnyUnreachableExit; 856 857 // Whether or not we should iterate on removing dead stores at the end of the 858 // function due to removing a store causing a previously captured pointer to 859 // no longer be captured. 860 bool ShouldIterateEndOfFunctionDSE; 861 862 /// Dead instructions to be removed at the end of DSE. 863 SmallVector<Instruction *> ToRemove; 864 865 // Class contains self-reference, make sure it's not copied/moved. 866 DSEState(const DSEState &) = delete; 867 DSEState &operator=(const DSEState &) = delete; 868 869 DSEState(Function &F, AliasAnalysis &AA, MemorySSA &MSSA, DominatorTree &DT, 870 PostDominatorTree &PDT, const TargetLibraryInfo &TLI, 871 const LoopInfo &LI) 872 : F(F), AA(AA), EI(DT, &LI), BatchAA(AA, &EI), MSSA(MSSA), DT(DT), 873 PDT(PDT), TLI(TLI), DL(F.getDataLayout()), LI(LI) { 874 // Collect blocks with throwing instructions not modeled in MemorySSA and 875 // alloc-like objects. 876 unsigned PO = 0; 877 for (BasicBlock *BB : post_order(&F)) { 878 PostOrderNumbers[BB] = PO++; 879 for (Instruction &I : *BB) { 880 MemoryAccess *MA = MSSA.getMemoryAccess(&I); 881 if (I.mayThrow() && !MA) 882 ThrowingBlocks.insert(I.getParent()); 883 884 auto *MD = dyn_cast_or_null<MemoryDef>(MA); 885 if (MD && MemDefs.size() < MemorySSADefsPerBlockLimit && 886 (getLocForWrite(&I) || isMemTerminatorInst(&I))) 887 MemDefs.push_back(MD); 888 } 889 } 890 891 // Treat byval or inalloca arguments the same as Allocas, stores to them are 892 // dead at the end of the function. 893 for (Argument &AI : F.args()) 894 if (AI.hasPassPointeeByValueCopyAttr()) 895 InvisibleToCallerAfterRet.insert({&AI, true}); 896 897 // Collect whether there is any irreducible control flow in the function. 898 ContainsIrreducibleLoops = mayContainIrreducibleControl(F, &LI); 899 900 AnyUnreachableExit = any_of(PDT.roots(), [](const BasicBlock *E) { 901 return isa<UnreachableInst>(E->getTerminator()); 902 }); 903 } 904 905 static void pushMemUses(MemoryAccess *Acc, 906 SmallVectorImpl<MemoryAccess *> &WorkList, 907 SmallPtrSetImpl<MemoryAccess *> &Visited) { 908 for (Use &U : Acc->uses()) { 909 auto *MA = cast<MemoryAccess>(U.getUser()); 910 if (Visited.insert(MA).second) 911 WorkList.push_back(MA); 912 } 913 }; 914 915 LocationSize strengthenLocationSize(const Instruction *I, 916 LocationSize Size) const { 917 if (auto *CB = dyn_cast<CallBase>(I)) { 918 LibFunc F; 919 if (TLI.getLibFunc(*CB, F) && TLI.has(F) && 920 (F == LibFunc_memset_chk || F == LibFunc_memcpy_chk)) { 921 // Use the precise location size specified by the 3rd argument 922 // for determining KillingI overwrites DeadLoc if it is a memset_chk 923 // instruction. memset_chk will write either the amount specified as 3rd 924 // argument or the function will immediately abort and exit the program. 925 // NOTE: AA may determine NoAlias if it can prove that the access size 926 // is larger than the allocation size due to that being UB. To avoid 927 // returning potentially invalid NoAlias results by AA, limit the use of 928 // the precise location size to isOverwrite. 929 if (const auto *Len = dyn_cast<ConstantInt>(CB->getArgOperand(2))) 930 return LocationSize::precise(Len->getZExtValue()); 931 } 932 } 933 return Size; 934 } 935 936 /// Return 'OW_Complete' if a store to the 'KillingLoc' location (by \p 937 /// KillingI instruction) completely overwrites a store to the 'DeadLoc' 938 /// location (by \p DeadI instruction). 939 /// Return OW_MaybePartial if \p KillingI does not completely overwrite 940 /// \p DeadI, but they both write to the same underlying object. In that 941 /// case, use isPartialOverwrite to check if \p KillingI partially overwrites 942 /// \p DeadI. Returns 'OR_None' if \p KillingI is known to not overwrite the 943 /// \p DeadI. Returns 'OW_Unknown' if nothing can be determined. 944 OverwriteResult isOverwrite(const Instruction *KillingI, 945 const Instruction *DeadI, 946 const MemoryLocation &KillingLoc, 947 const MemoryLocation &DeadLoc, 948 int64_t &KillingOff, int64_t &DeadOff) { 949 // AliasAnalysis does not always account for loops. Limit overwrite checks 950 // to dependencies for which we can guarantee they are independent of any 951 // loops they are in. 952 if (!isGuaranteedLoopIndependent(DeadI, KillingI, DeadLoc)) 953 return OW_Unknown; 954 955 LocationSize KillingLocSize = 956 strengthenLocationSize(KillingI, KillingLoc.Size); 957 const Value *DeadPtr = DeadLoc.Ptr->stripPointerCasts(); 958 const Value *KillingPtr = KillingLoc.Ptr->stripPointerCasts(); 959 const Value *DeadUndObj = getUnderlyingObject(DeadPtr); 960 const Value *KillingUndObj = getUnderlyingObject(KillingPtr); 961 962 // Check whether the killing store overwrites the whole object, in which 963 // case the size/offset of the dead store does not matter. 964 if (DeadUndObj == KillingUndObj && KillingLocSize.isPrecise() && 965 isIdentifiedObject(KillingUndObj)) { 966 std::optional<TypeSize> KillingUndObjSize = 967 getPointerSize(KillingUndObj, DL, TLI, &F); 968 if (KillingUndObjSize && *KillingUndObjSize == KillingLocSize.getValue()) 969 return OW_Complete; 970 } 971 972 // FIXME: Vet that this works for size upper-bounds. Seems unlikely that we'll 973 // get imprecise values here, though (except for unknown sizes). 974 if (!KillingLocSize.isPrecise() || !DeadLoc.Size.isPrecise()) { 975 // In case no constant size is known, try to an IR values for the number 976 // of bytes written and check if they match. 977 const auto *KillingMemI = dyn_cast<MemIntrinsic>(KillingI); 978 const auto *DeadMemI = dyn_cast<MemIntrinsic>(DeadI); 979 if (KillingMemI && DeadMemI) { 980 const Value *KillingV = KillingMemI->getLength(); 981 const Value *DeadV = DeadMemI->getLength(); 982 if (KillingV == DeadV && BatchAA.isMustAlias(DeadLoc, KillingLoc)) 983 return OW_Complete; 984 } 985 986 // Masked stores have imprecise locations, but we can reason about them 987 // to some extent. 988 return isMaskedStoreOverwrite(KillingI, DeadI, BatchAA); 989 } 990 991 const TypeSize KillingSize = KillingLocSize.getValue(); 992 const TypeSize DeadSize = DeadLoc.Size.getValue(); 993 // Bail on doing Size comparison which depends on AA for now 994 // TODO: Remove AnyScalable once Alias Analysis deal with scalable vectors 995 const bool AnyScalable = 996 DeadSize.isScalable() || KillingLocSize.isScalable(); 997 998 if (AnyScalable) 999 return OW_Unknown; 1000 // Query the alias information 1001 AliasResult AAR = BatchAA.alias(KillingLoc, DeadLoc); 1002 1003 // If the start pointers are the same, we just have to compare sizes to see if 1004 // the killing store was larger than the dead store. 1005 if (AAR == AliasResult::MustAlias) { 1006 // Make sure that the KillingSize size is >= the DeadSize size. 1007 if (KillingSize >= DeadSize) 1008 return OW_Complete; 1009 } 1010 1011 // If we hit a partial alias we may have a full overwrite 1012 if (AAR == AliasResult::PartialAlias && AAR.hasOffset()) { 1013 int32_t Off = AAR.getOffset(); 1014 if (Off >= 0 && (uint64_t)Off + DeadSize <= KillingSize) 1015 return OW_Complete; 1016 } 1017 1018 // If we can't resolve the same pointers to the same object, then we can't 1019 // analyze them at all. 1020 if (DeadUndObj != KillingUndObj) { 1021 // Non aliasing stores to different objects don't overlap. Note that 1022 // if the killing store is known to overwrite whole object (out of 1023 // bounds access overwrites whole object as well) then it is assumed to 1024 // completely overwrite any store to the same object even if they don't 1025 // actually alias (see next check). 1026 if (AAR == AliasResult::NoAlias) 1027 return OW_None; 1028 return OW_Unknown; 1029 } 1030 1031 // Okay, we have stores to two completely different pointers. Try to 1032 // decompose the pointer into a "base + constant_offset" form. If the base 1033 // pointers are equal, then we can reason about the two stores. 1034 DeadOff = 0; 1035 KillingOff = 0; 1036 const Value *DeadBasePtr = 1037 GetPointerBaseWithConstantOffset(DeadPtr, DeadOff, DL); 1038 const Value *KillingBasePtr = 1039 GetPointerBaseWithConstantOffset(KillingPtr, KillingOff, DL); 1040 1041 // If the base pointers still differ, we have two completely different 1042 // stores. 1043 if (DeadBasePtr != KillingBasePtr) 1044 return OW_Unknown; 1045 1046 // The killing access completely overlaps the dead store if and only if 1047 // both start and end of the dead one is "inside" the killing one: 1048 // |<->|--dead--|<->| 1049 // |-----killing------| 1050 // Accesses may overlap if and only if start of one of them is "inside" 1051 // another one: 1052 // |<->|--dead--|<-------->| 1053 // |-------killing--------| 1054 // OR 1055 // |-------dead-------| 1056 // |<->|---killing---|<----->| 1057 // 1058 // We have to be careful here as *Off is signed while *.Size is unsigned. 1059 1060 // Check if the dead access starts "not before" the killing one. 1061 if (DeadOff >= KillingOff) { 1062 // If the dead access ends "not after" the killing access then the 1063 // dead one is completely overwritten by the killing one. 1064 if (uint64_t(DeadOff - KillingOff) + DeadSize <= KillingSize) 1065 return OW_Complete; 1066 // If start of the dead access is "before" end of the killing access 1067 // then accesses overlap. 1068 else if ((uint64_t)(DeadOff - KillingOff) < KillingSize) 1069 return OW_MaybePartial; 1070 } 1071 // If start of the killing access is "before" end of the dead access then 1072 // accesses overlap. 1073 else if ((uint64_t)(KillingOff - DeadOff) < DeadSize) { 1074 return OW_MaybePartial; 1075 } 1076 1077 // Can reach here only if accesses are known not to overlap. 1078 return OW_None; 1079 } 1080 1081 bool isInvisibleToCallerAfterRet(const Value *V) { 1082 if (isa<AllocaInst>(V)) 1083 return true; 1084 auto I = InvisibleToCallerAfterRet.insert({V, false}); 1085 if (I.second) { 1086 if (!isInvisibleToCallerOnUnwind(V)) { 1087 I.first->second = false; 1088 } else if (isNoAliasCall(V)) { 1089 I.first->second = !PointerMayBeCaptured(V, true, false); 1090 } 1091 } 1092 return I.first->second; 1093 } 1094 1095 bool isInvisibleToCallerOnUnwind(const Value *V) { 1096 bool RequiresNoCaptureBeforeUnwind; 1097 if (!isNotVisibleOnUnwind(V, RequiresNoCaptureBeforeUnwind)) 1098 return false; 1099 if (!RequiresNoCaptureBeforeUnwind) 1100 return true; 1101 1102 auto I = CapturedBeforeReturn.insert({V, true}); 1103 if (I.second) 1104 // NOTE: This could be made more precise by PointerMayBeCapturedBefore 1105 // with the killing MemoryDef. But we refrain from doing so for now to 1106 // limit compile-time and this does not cause any changes to the number 1107 // of stores removed on a large test set in practice. 1108 I.first->second = PointerMayBeCaptured(V, false, true); 1109 return !I.first->second; 1110 } 1111 1112 std::optional<MemoryLocation> getLocForWrite(Instruction *I) const { 1113 if (!I->mayWriteToMemory()) 1114 return std::nullopt; 1115 1116 if (auto *CB = dyn_cast<CallBase>(I)) 1117 return MemoryLocation::getForDest(CB, TLI); 1118 1119 return MemoryLocation::getOrNone(I); 1120 } 1121 1122 /// Assuming this instruction has a dead analyzable write, can we delete 1123 /// this instruction? 1124 bool isRemovable(Instruction *I) { 1125 assert(getLocForWrite(I) && "Must have analyzable write"); 1126 1127 // Don't remove volatile/atomic stores. 1128 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 1129 return SI->isUnordered(); 1130 1131 if (auto *CB = dyn_cast<CallBase>(I)) { 1132 // Don't remove volatile memory intrinsics. 1133 if (auto *MI = dyn_cast<MemIntrinsic>(CB)) 1134 return !MI->isVolatile(); 1135 1136 // Never remove dead lifetime intrinsics, e.g. because they are followed 1137 // by a free. 1138 if (CB->isLifetimeStartOrEnd()) 1139 return false; 1140 1141 return CB->use_empty() && CB->willReturn() && CB->doesNotThrow() && 1142 !CB->isTerminator(); 1143 } 1144 1145 return false; 1146 } 1147 1148 /// Returns true if \p UseInst completely overwrites \p DefLoc 1149 /// (stored by \p DefInst). 1150 bool isCompleteOverwrite(const MemoryLocation &DefLoc, Instruction *DefInst, 1151 Instruction *UseInst) { 1152 // UseInst has a MemoryDef associated in MemorySSA. It's possible for a 1153 // MemoryDef to not write to memory, e.g. a volatile load is modeled as a 1154 // MemoryDef. 1155 if (!UseInst->mayWriteToMemory()) 1156 return false; 1157 1158 if (auto *CB = dyn_cast<CallBase>(UseInst)) 1159 if (CB->onlyAccessesInaccessibleMemory()) 1160 return false; 1161 1162 int64_t InstWriteOffset, DepWriteOffset; 1163 if (auto CC = getLocForWrite(UseInst)) 1164 return isOverwrite(UseInst, DefInst, *CC, DefLoc, InstWriteOffset, 1165 DepWriteOffset) == OW_Complete; 1166 return false; 1167 } 1168 1169 /// Returns true if \p Def is not read before returning from the function. 1170 bool isWriteAtEndOfFunction(MemoryDef *Def, const MemoryLocation &DefLoc) { 1171 LLVM_DEBUG(dbgs() << " Check if def " << *Def << " (" 1172 << *Def->getMemoryInst() 1173 << ") is at the end the function \n"); 1174 SmallVector<MemoryAccess *, 4> WorkList; 1175 SmallPtrSet<MemoryAccess *, 8> Visited; 1176 1177 pushMemUses(Def, WorkList, Visited); 1178 for (unsigned I = 0; I < WorkList.size(); I++) { 1179 if (WorkList.size() >= MemorySSAScanLimit) { 1180 LLVM_DEBUG(dbgs() << " ... hit exploration limit.\n"); 1181 return false; 1182 } 1183 1184 MemoryAccess *UseAccess = WorkList[I]; 1185 if (isa<MemoryPhi>(UseAccess)) { 1186 // AliasAnalysis does not account for loops. Limit elimination to 1187 // candidates for which we can guarantee they always store to the same 1188 // memory location. 1189 if (!isGuaranteedLoopInvariant(DefLoc.Ptr)) 1190 return false; 1191 1192 pushMemUses(cast<MemoryPhi>(UseAccess), WorkList, Visited); 1193 continue; 1194 } 1195 // TODO: Checking for aliasing is expensive. Consider reducing the amount 1196 // of times this is called and/or caching it. 1197 Instruction *UseInst = cast<MemoryUseOrDef>(UseAccess)->getMemoryInst(); 1198 if (isReadClobber(DefLoc, UseInst)) { 1199 LLVM_DEBUG(dbgs() << " ... hit read clobber " << *UseInst << ".\n"); 1200 return false; 1201 } 1202 1203 if (MemoryDef *UseDef = dyn_cast<MemoryDef>(UseAccess)) 1204 pushMemUses(UseDef, WorkList, Visited); 1205 } 1206 return true; 1207 } 1208 1209 /// If \p I is a memory terminator like llvm.lifetime.end or free, return a 1210 /// pair with the MemoryLocation terminated by \p I and a boolean flag 1211 /// indicating whether \p I is a free-like call. 1212 std::optional<std::pair<MemoryLocation, bool>> 1213 getLocForTerminator(Instruction *I) const { 1214 uint64_t Len; 1215 Value *Ptr; 1216 if (match(I, m_Intrinsic<Intrinsic::lifetime_end>(m_ConstantInt(Len), 1217 m_Value(Ptr)))) 1218 return {std::make_pair(MemoryLocation(Ptr, Len), false)}; 1219 1220 if (auto *CB = dyn_cast<CallBase>(I)) { 1221 if (Value *FreedOp = getFreedOperand(CB, &TLI)) 1222 return {std::make_pair(MemoryLocation::getAfter(FreedOp), true)}; 1223 } 1224 1225 return std::nullopt; 1226 } 1227 1228 /// Returns true if \p I is a memory terminator instruction like 1229 /// llvm.lifetime.end or free. 1230 bool isMemTerminatorInst(Instruction *I) const { 1231 auto *CB = dyn_cast<CallBase>(I); 1232 return CB && (CB->getIntrinsicID() == Intrinsic::lifetime_end || 1233 getFreedOperand(CB, &TLI) != nullptr); 1234 } 1235 1236 /// Returns true if \p MaybeTerm is a memory terminator for \p Loc from 1237 /// instruction \p AccessI. 1238 bool isMemTerminator(const MemoryLocation &Loc, Instruction *AccessI, 1239 Instruction *MaybeTerm) { 1240 std::optional<std::pair<MemoryLocation, bool>> MaybeTermLoc = 1241 getLocForTerminator(MaybeTerm); 1242 1243 if (!MaybeTermLoc) 1244 return false; 1245 1246 // If the terminator is a free-like call, all accesses to the underlying 1247 // object can be considered terminated. 1248 if (getUnderlyingObject(Loc.Ptr) != 1249 getUnderlyingObject(MaybeTermLoc->first.Ptr)) 1250 return false; 1251 1252 auto TermLoc = MaybeTermLoc->first; 1253 if (MaybeTermLoc->second) { 1254 const Value *LocUO = getUnderlyingObject(Loc.Ptr); 1255 return BatchAA.isMustAlias(TermLoc.Ptr, LocUO); 1256 } 1257 int64_t InstWriteOffset = 0; 1258 int64_t DepWriteOffset = 0; 1259 return isOverwrite(MaybeTerm, AccessI, TermLoc, Loc, InstWriteOffset, 1260 DepWriteOffset) == OW_Complete; 1261 } 1262 1263 // Returns true if \p Use may read from \p DefLoc. 1264 bool isReadClobber(const MemoryLocation &DefLoc, Instruction *UseInst) { 1265 if (isNoopIntrinsic(UseInst)) 1266 return false; 1267 1268 // Monotonic or weaker atomic stores can be re-ordered and do not need to be 1269 // treated as read clobber. 1270 if (auto SI = dyn_cast<StoreInst>(UseInst)) 1271 return isStrongerThan(SI->getOrdering(), AtomicOrdering::Monotonic); 1272 1273 if (!UseInst->mayReadFromMemory()) 1274 return false; 1275 1276 if (auto *CB = dyn_cast<CallBase>(UseInst)) 1277 if (CB->onlyAccessesInaccessibleMemory()) 1278 return false; 1279 1280 return isRefSet(BatchAA.getModRefInfo(UseInst, DefLoc)); 1281 } 1282 1283 /// Returns true if a dependency between \p Current and \p KillingDef is 1284 /// guaranteed to be loop invariant for the loops that they are in. Either 1285 /// because they are known to be in the same block, in the same loop level or 1286 /// by guaranteeing that \p CurrentLoc only references a single MemoryLocation 1287 /// during execution of the containing function. 1288 bool isGuaranteedLoopIndependent(const Instruction *Current, 1289 const Instruction *KillingDef, 1290 const MemoryLocation &CurrentLoc) { 1291 // If the dependency is within the same block or loop level (being careful 1292 // of irreducible loops), we know that AA will return a valid result for the 1293 // memory dependency. (Both at the function level, outside of any loop, 1294 // would also be valid but we currently disable that to limit compile time). 1295 if (Current->getParent() == KillingDef->getParent()) 1296 return true; 1297 const Loop *CurrentLI = LI.getLoopFor(Current->getParent()); 1298 if (!ContainsIrreducibleLoops && CurrentLI && 1299 CurrentLI == LI.getLoopFor(KillingDef->getParent())) 1300 return true; 1301 // Otherwise check the memory location is invariant to any loops. 1302 return isGuaranteedLoopInvariant(CurrentLoc.Ptr); 1303 } 1304 1305 /// Returns true if \p Ptr is guaranteed to be loop invariant for any possible 1306 /// loop. In particular, this guarantees that it only references a single 1307 /// MemoryLocation during execution of the containing function. 1308 bool isGuaranteedLoopInvariant(const Value *Ptr) { 1309 Ptr = Ptr->stripPointerCasts(); 1310 if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) 1311 if (GEP->hasAllConstantIndices()) 1312 Ptr = GEP->getPointerOperand()->stripPointerCasts(); 1313 1314 if (auto *I = dyn_cast<Instruction>(Ptr)) { 1315 return I->getParent()->isEntryBlock() || 1316 (!ContainsIrreducibleLoops && !LI.getLoopFor(I->getParent())); 1317 } 1318 return true; 1319 } 1320 1321 // Find a MemoryDef writing to \p KillingLoc and dominating \p StartAccess, 1322 // with no read access between them or on any other path to a function exit 1323 // block if \p KillingLoc is not accessible after the function returns. If 1324 // there is no such MemoryDef, return std::nullopt. The returned value may not 1325 // (completely) overwrite \p KillingLoc. Currently we bail out when we 1326 // encounter an aliasing MemoryUse (read). 1327 std::optional<MemoryAccess *> 1328 getDomMemoryDef(MemoryDef *KillingDef, MemoryAccess *StartAccess, 1329 const MemoryLocation &KillingLoc, const Value *KillingUndObj, 1330 unsigned &ScanLimit, unsigned &WalkerStepLimit, 1331 bool IsMemTerm, unsigned &PartialLimit) { 1332 if (ScanLimit == 0 || WalkerStepLimit == 0) { 1333 LLVM_DEBUG(dbgs() << "\n ... hit scan limit\n"); 1334 return std::nullopt; 1335 } 1336 1337 MemoryAccess *Current = StartAccess; 1338 Instruction *KillingI = KillingDef->getMemoryInst(); 1339 LLVM_DEBUG(dbgs() << " trying to get dominating access\n"); 1340 1341 // Only optimize defining access of KillingDef when directly starting at its 1342 // defining access. The defining access also must only access KillingLoc. At 1343 // the moment we only support instructions with a single write location, so 1344 // it should be sufficient to disable optimizations for instructions that 1345 // also read from memory. 1346 bool CanOptimize = OptimizeMemorySSA && 1347 KillingDef->getDefiningAccess() == StartAccess && 1348 !KillingI->mayReadFromMemory(); 1349 1350 // Find the next clobbering Mod access for DefLoc, starting at StartAccess. 1351 std::optional<MemoryLocation> CurrentLoc; 1352 for (;; Current = cast<MemoryDef>(Current)->getDefiningAccess()) { 1353 LLVM_DEBUG({ 1354 dbgs() << " visiting " << *Current; 1355 if (!MSSA.isLiveOnEntryDef(Current) && isa<MemoryUseOrDef>(Current)) 1356 dbgs() << " (" << *cast<MemoryUseOrDef>(Current)->getMemoryInst() 1357 << ")"; 1358 dbgs() << "\n"; 1359 }); 1360 1361 // Reached TOP. 1362 if (MSSA.isLiveOnEntryDef(Current)) { 1363 LLVM_DEBUG(dbgs() << " ... found LiveOnEntryDef\n"); 1364 if (CanOptimize && Current != KillingDef->getDefiningAccess()) 1365 // The first clobbering def is... none. 1366 KillingDef->setOptimized(Current); 1367 return std::nullopt; 1368 } 1369 1370 // Cost of a step. Accesses in the same block are more likely to be valid 1371 // candidates for elimination, hence consider them cheaper. 1372 unsigned StepCost = KillingDef->getBlock() == Current->getBlock() 1373 ? MemorySSASameBBStepCost 1374 : MemorySSAOtherBBStepCost; 1375 if (WalkerStepLimit <= StepCost) { 1376 LLVM_DEBUG(dbgs() << " ... hit walker step limit\n"); 1377 return std::nullopt; 1378 } 1379 WalkerStepLimit -= StepCost; 1380 1381 // Return for MemoryPhis. They cannot be eliminated directly and the 1382 // caller is responsible for traversing them. 1383 if (isa<MemoryPhi>(Current)) { 1384 LLVM_DEBUG(dbgs() << " ... found MemoryPhi\n"); 1385 return Current; 1386 } 1387 1388 // Below, check if CurrentDef is a valid candidate to be eliminated by 1389 // KillingDef. If it is not, check the next candidate. 1390 MemoryDef *CurrentDef = cast<MemoryDef>(Current); 1391 Instruction *CurrentI = CurrentDef->getMemoryInst(); 1392 1393 if (canSkipDef(CurrentDef, !isInvisibleToCallerOnUnwind(KillingUndObj))) { 1394 CanOptimize = false; 1395 continue; 1396 } 1397 1398 // Before we try to remove anything, check for any extra throwing 1399 // instructions that block us from DSEing 1400 if (mayThrowBetween(KillingI, CurrentI, KillingUndObj)) { 1401 LLVM_DEBUG(dbgs() << " ... skip, may throw!\n"); 1402 return std::nullopt; 1403 } 1404 1405 // Check for anything that looks like it will be a barrier to further 1406 // removal 1407 if (isDSEBarrier(KillingUndObj, CurrentI)) { 1408 LLVM_DEBUG(dbgs() << " ... skip, barrier\n"); 1409 return std::nullopt; 1410 } 1411 1412 // If Current is known to be on path that reads DefLoc or is a read 1413 // clobber, bail out, as the path is not profitable. We skip this check 1414 // for intrinsic calls, because the code knows how to handle memcpy 1415 // intrinsics. 1416 if (!isa<IntrinsicInst>(CurrentI) && isReadClobber(KillingLoc, CurrentI)) 1417 return std::nullopt; 1418 1419 // Quick check if there are direct uses that are read-clobbers. 1420 if (any_of(Current->uses(), [this, &KillingLoc, StartAccess](Use &U) { 1421 if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(U.getUser())) 1422 return !MSSA.dominates(StartAccess, UseOrDef) && 1423 isReadClobber(KillingLoc, UseOrDef->getMemoryInst()); 1424 return false; 1425 })) { 1426 LLVM_DEBUG(dbgs() << " ... found a read clobber\n"); 1427 return std::nullopt; 1428 } 1429 1430 // If Current does not have an analyzable write location or is not 1431 // removable, skip it. 1432 CurrentLoc = getLocForWrite(CurrentI); 1433 if (!CurrentLoc || !isRemovable(CurrentI)) { 1434 CanOptimize = false; 1435 continue; 1436 } 1437 1438 // AliasAnalysis does not account for loops. Limit elimination to 1439 // candidates for which we can guarantee they always store to the same 1440 // memory location and not located in different loops. 1441 if (!isGuaranteedLoopIndependent(CurrentI, KillingI, *CurrentLoc)) { 1442 LLVM_DEBUG(dbgs() << " ... not guaranteed loop independent\n"); 1443 CanOptimize = false; 1444 continue; 1445 } 1446 1447 if (IsMemTerm) { 1448 // If the killing def is a memory terminator (e.g. lifetime.end), check 1449 // the next candidate if the current Current does not write the same 1450 // underlying object as the terminator. 1451 if (!isMemTerminator(*CurrentLoc, CurrentI, KillingI)) { 1452 CanOptimize = false; 1453 continue; 1454 } 1455 } else { 1456 int64_t KillingOffset = 0; 1457 int64_t DeadOffset = 0; 1458 auto OR = isOverwrite(KillingI, CurrentI, KillingLoc, *CurrentLoc, 1459 KillingOffset, DeadOffset); 1460 if (CanOptimize) { 1461 // CurrentDef is the earliest write clobber of KillingDef. Use it as 1462 // optimized access. Do not optimize if CurrentDef is already the 1463 // defining access of KillingDef. 1464 if (CurrentDef != KillingDef->getDefiningAccess() && 1465 (OR == OW_Complete || OR == OW_MaybePartial)) 1466 KillingDef->setOptimized(CurrentDef); 1467 1468 // Once a may-aliasing def is encountered do not set an optimized 1469 // access. 1470 if (OR != OW_None) 1471 CanOptimize = false; 1472 } 1473 1474 // If Current does not write to the same object as KillingDef, check 1475 // the next candidate. 1476 if (OR == OW_Unknown || OR == OW_None) 1477 continue; 1478 else if (OR == OW_MaybePartial) { 1479 // If KillingDef only partially overwrites Current, check the next 1480 // candidate if the partial step limit is exceeded. This aggressively 1481 // limits the number of candidates for partial store elimination, 1482 // which are less likely to be removable in the end. 1483 if (PartialLimit <= 1) { 1484 WalkerStepLimit -= 1; 1485 LLVM_DEBUG(dbgs() << " ... reached partial limit ... continue with next access\n"); 1486 continue; 1487 } 1488 PartialLimit -= 1; 1489 } 1490 } 1491 break; 1492 }; 1493 1494 // Accesses to objects accessible after the function returns can only be 1495 // eliminated if the access is dead along all paths to the exit. Collect 1496 // the blocks with killing (=completely overwriting MemoryDefs) and check if 1497 // they cover all paths from MaybeDeadAccess to any function exit. 1498 SmallPtrSet<Instruction *, 16> KillingDefs; 1499 KillingDefs.insert(KillingDef->getMemoryInst()); 1500 MemoryAccess *MaybeDeadAccess = Current; 1501 MemoryLocation MaybeDeadLoc = *CurrentLoc; 1502 Instruction *MaybeDeadI = cast<MemoryDef>(MaybeDeadAccess)->getMemoryInst(); 1503 LLVM_DEBUG(dbgs() << " Checking for reads of " << *MaybeDeadAccess << " (" 1504 << *MaybeDeadI << ")\n"); 1505 1506 SmallVector<MemoryAccess *, 32> WorkList; 1507 SmallPtrSet<MemoryAccess *, 32> Visited; 1508 pushMemUses(MaybeDeadAccess, WorkList, Visited); 1509 1510 // Check if DeadDef may be read. 1511 for (unsigned I = 0; I < WorkList.size(); I++) { 1512 MemoryAccess *UseAccess = WorkList[I]; 1513 1514 LLVM_DEBUG(dbgs() << " " << *UseAccess); 1515 // Bail out if the number of accesses to check exceeds the scan limit. 1516 if (ScanLimit < (WorkList.size() - I)) { 1517 LLVM_DEBUG(dbgs() << "\n ... hit scan limit\n"); 1518 return std::nullopt; 1519 } 1520 --ScanLimit; 1521 NumDomMemDefChecks++; 1522 1523 if (isa<MemoryPhi>(UseAccess)) { 1524 if (any_of(KillingDefs, [this, UseAccess](Instruction *KI) { 1525 return DT.properlyDominates(KI->getParent(), 1526 UseAccess->getBlock()); 1527 })) { 1528 LLVM_DEBUG(dbgs() << " ... skipping, dominated by killing block\n"); 1529 continue; 1530 } 1531 LLVM_DEBUG(dbgs() << "\n ... adding PHI uses\n"); 1532 pushMemUses(UseAccess, WorkList, Visited); 1533 continue; 1534 } 1535 1536 Instruction *UseInst = cast<MemoryUseOrDef>(UseAccess)->getMemoryInst(); 1537 LLVM_DEBUG(dbgs() << " (" << *UseInst << ")\n"); 1538 1539 if (any_of(KillingDefs, [this, UseInst](Instruction *KI) { 1540 return DT.dominates(KI, UseInst); 1541 })) { 1542 LLVM_DEBUG(dbgs() << " ... skipping, dominated by killing def\n"); 1543 continue; 1544 } 1545 1546 // A memory terminator kills all preceeding MemoryDefs and all succeeding 1547 // MemoryAccesses. We do not have to check it's users. 1548 if (isMemTerminator(MaybeDeadLoc, MaybeDeadI, UseInst)) { 1549 LLVM_DEBUG( 1550 dbgs() 1551 << " ... skipping, memterminator invalidates following accesses\n"); 1552 continue; 1553 } 1554 1555 if (isNoopIntrinsic(cast<MemoryUseOrDef>(UseAccess)->getMemoryInst())) { 1556 LLVM_DEBUG(dbgs() << " ... adding uses of intrinsic\n"); 1557 pushMemUses(UseAccess, WorkList, Visited); 1558 continue; 1559 } 1560 1561 if (UseInst->mayThrow() && !isInvisibleToCallerOnUnwind(KillingUndObj)) { 1562 LLVM_DEBUG(dbgs() << " ... found throwing instruction\n"); 1563 return std::nullopt; 1564 } 1565 1566 // Uses which may read the original MemoryDef mean we cannot eliminate the 1567 // original MD. Stop walk. 1568 if (isReadClobber(MaybeDeadLoc, UseInst)) { 1569 LLVM_DEBUG(dbgs() << " ... found read clobber\n"); 1570 return std::nullopt; 1571 } 1572 1573 // If this worklist walks back to the original memory access (and the 1574 // pointer is not guarenteed loop invariant) then we cannot assume that a 1575 // store kills itself. 1576 if (MaybeDeadAccess == UseAccess && 1577 !isGuaranteedLoopInvariant(MaybeDeadLoc.Ptr)) { 1578 LLVM_DEBUG(dbgs() << " ... found not loop invariant self access\n"); 1579 return std::nullopt; 1580 } 1581 // Otherwise, for the KillingDef and MaybeDeadAccess we only have to check 1582 // if it reads the memory location. 1583 // TODO: It would probably be better to check for self-reads before 1584 // calling the function. 1585 if (KillingDef == UseAccess || MaybeDeadAccess == UseAccess) { 1586 LLVM_DEBUG(dbgs() << " ... skipping killing def/dom access\n"); 1587 continue; 1588 } 1589 1590 // Check all uses for MemoryDefs, except for defs completely overwriting 1591 // the original location. Otherwise we have to check uses of *all* 1592 // MemoryDefs we discover, including non-aliasing ones. Otherwise we might 1593 // miss cases like the following 1594 // 1 = Def(LoE) ; <----- DeadDef stores [0,1] 1595 // 2 = Def(1) ; (2, 1) = NoAlias, stores [2,3] 1596 // Use(2) ; MayAlias 2 *and* 1, loads [0, 3]. 1597 // (The Use points to the *first* Def it may alias) 1598 // 3 = Def(1) ; <---- Current (3, 2) = NoAlias, (3,1) = MayAlias, 1599 // stores [0,1] 1600 if (MemoryDef *UseDef = dyn_cast<MemoryDef>(UseAccess)) { 1601 if (isCompleteOverwrite(MaybeDeadLoc, MaybeDeadI, UseInst)) { 1602 BasicBlock *MaybeKillingBlock = UseInst->getParent(); 1603 if (PostOrderNumbers.find(MaybeKillingBlock)->second < 1604 PostOrderNumbers.find(MaybeDeadAccess->getBlock())->second) { 1605 if (!isInvisibleToCallerAfterRet(KillingUndObj)) { 1606 LLVM_DEBUG(dbgs() 1607 << " ... found killing def " << *UseInst << "\n"); 1608 KillingDefs.insert(UseInst); 1609 } 1610 } else { 1611 LLVM_DEBUG(dbgs() 1612 << " ... found preceeding def " << *UseInst << "\n"); 1613 return std::nullopt; 1614 } 1615 } else 1616 pushMemUses(UseDef, WorkList, Visited); 1617 } 1618 } 1619 1620 // For accesses to locations visible after the function returns, make sure 1621 // that the location is dead (=overwritten) along all paths from 1622 // MaybeDeadAccess to the exit. 1623 if (!isInvisibleToCallerAfterRet(KillingUndObj)) { 1624 SmallPtrSet<BasicBlock *, 16> KillingBlocks; 1625 for (Instruction *KD : KillingDefs) 1626 KillingBlocks.insert(KD->getParent()); 1627 assert(!KillingBlocks.empty() && 1628 "Expected at least a single killing block"); 1629 1630 // Find the common post-dominator of all killing blocks. 1631 BasicBlock *CommonPred = *KillingBlocks.begin(); 1632 for (BasicBlock *BB : llvm::drop_begin(KillingBlocks)) { 1633 if (!CommonPred) 1634 break; 1635 CommonPred = PDT.findNearestCommonDominator(CommonPred, BB); 1636 } 1637 1638 // If the common post-dominator does not post-dominate MaybeDeadAccess, 1639 // there is a path from MaybeDeadAccess to an exit not going through a 1640 // killing block. 1641 if (!PDT.dominates(CommonPred, MaybeDeadAccess->getBlock())) { 1642 if (!AnyUnreachableExit) 1643 return std::nullopt; 1644 1645 // Fall back to CFG scan starting at all non-unreachable roots if not 1646 // all paths to the exit go through CommonPred. 1647 CommonPred = nullptr; 1648 } 1649 1650 // If CommonPred itself is in the set of killing blocks, we're done. 1651 if (KillingBlocks.count(CommonPred)) 1652 return {MaybeDeadAccess}; 1653 1654 SetVector<BasicBlock *> WorkList; 1655 // If CommonPred is null, there are multiple exits from the function. 1656 // They all have to be added to the worklist. 1657 if (CommonPred) 1658 WorkList.insert(CommonPred); 1659 else 1660 for (BasicBlock *R : PDT.roots()) { 1661 if (!isa<UnreachableInst>(R->getTerminator())) 1662 WorkList.insert(R); 1663 } 1664 1665 NumCFGTries++; 1666 // Check if all paths starting from an exit node go through one of the 1667 // killing blocks before reaching MaybeDeadAccess. 1668 for (unsigned I = 0; I < WorkList.size(); I++) { 1669 NumCFGChecks++; 1670 BasicBlock *Current = WorkList[I]; 1671 if (KillingBlocks.count(Current)) 1672 continue; 1673 if (Current == MaybeDeadAccess->getBlock()) 1674 return std::nullopt; 1675 1676 // MaybeDeadAccess is reachable from the entry, so we don't have to 1677 // explore unreachable blocks further. 1678 if (!DT.isReachableFromEntry(Current)) 1679 continue; 1680 1681 for (BasicBlock *Pred : predecessors(Current)) 1682 WorkList.insert(Pred); 1683 1684 if (WorkList.size() >= MemorySSAPathCheckLimit) 1685 return std::nullopt; 1686 } 1687 NumCFGSuccess++; 1688 } 1689 1690 // No aliasing MemoryUses of MaybeDeadAccess found, MaybeDeadAccess is 1691 // potentially dead. 1692 return {MaybeDeadAccess}; 1693 } 1694 1695 /// Delete dead memory defs and recursively add their operands to ToRemove if 1696 /// they became dead. 1697 void 1698 deleteDeadInstruction(Instruction *SI, 1699 SmallPtrSetImpl<MemoryAccess *> *Deleted = nullptr) { 1700 MemorySSAUpdater Updater(&MSSA); 1701 SmallVector<Instruction *, 32> NowDeadInsts; 1702 NowDeadInsts.push_back(SI); 1703 --NumFastOther; 1704 1705 while (!NowDeadInsts.empty()) { 1706 Instruction *DeadInst = NowDeadInsts.pop_back_val(); 1707 ++NumFastOther; 1708 1709 // Try to preserve debug information attached to the dead instruction. 1710 salvageDebugInfo(*DeadInst); 1711 salvageKnowledge(DeadInst); 1712 1713 // Remove the Instruction from MSSA. 1714 MemoryAccess *MA = MSSA.getMemoryAccess(DeadInst); 1715 bool IsMemDef = MA && isa<MemoryDef>(MA); 1716 if (MA) { 1717 if (IsMemDef) { 1718 auto *MD = cast<MemoryDef>(MA); 1719 SkipStores.insert(MD); 1720 if (Deleted) 1721 Deleted->insert(MD); 1722 if (auto *SI = dyn_cast<StoreInst>(MD->getMemoryInst())) { 1723 if (SI->getValueOperand()->getType()->isPointerTy()) { 1724 const Value *UO = getUnderlyingObject(SI->getValueOperand()); 1725 if (CapturedBeforeReturn.erase(UO)) 1726 ShouldIterateEndOfFunctionDSE = true; 1727 InvisibleToCallerAfterRet.erase(UO); 1728 } 1729 } 1730 } 1731 1732 Updater.removeMemoryAccess(MA); 1733 } 1734 1735 auto I = IOLs.find(DeadInst->getParent()); 1736 if (I != IOLs.end()) 1737 I->second.erase(DeadInst); 1738 // Remove its operands 1739 for (Use &O : DeadInst->operands()) 1740 if (Instruction *OpI = dyn_cast<Instruction>(O)) { 1741 O.set(PoisonValue::get(O->getType())); 1742 if (isInstructionTriviallyDead(OpI, &TLI)) 1743 NowDeadInsts.push_back(OpI); 1744 } 1745 1746 EI.removeInstruction(DeadInst); 1747 // Remove memory defs directly if they don't produce results, but only 1748 // queue other dead instructions for later removal. They may have been 1749 // used as memory locations that have been cached by BatchAA. Removing 1750 // them here may lead to newly created instructions to be allocated at the 1751 // same address, yielding stale cache entries. 1752 if (IsMemDef && DeadInst->getType()->isVoidTy()) 1753 DeadInst->eraseFromParent(); 1754 else 1755 ToRemove.push_back(DeadInst); 1756 } 1757 } 1758 1759 // Check for any extra throws between \p KillingI and \p DeadI that block 1760 // DSE. This only checks extra maythrows (those that aren't MemoryDef's). 1761 // MemoryDef that may throw are handled during the walk from one def to the 1762 // next. 1763 bool mayThrowBetween(Instruction *KillingI, Instruction *DeadI, 1764 const Value *KillingUndObj) { 1765 // First see if we can ignore it by using the fact that KillingI is an 1766 // alloca/alloca like object that is not visible to the caller during 1767 // execution of the function. 1768 if (KillingUndObj && isInvisibleToCallerOnUnwind(KillingUndObj)) 1769 return false; 1770 1771 if (KillingI->getParent() == DeadI->getParent()) 1772 return ThrowingBlocks.count(KillingI->getParent()); 1773 return !ThrowingBlocks.empty(); 1774 } 1775 1776 // Check if \p DeadI acts as a DSE barrier for \p KillingI. The following 1777 // instructions act as barriers: 1778 // * A memory instruction that may throw and \p KillingI accesses a non-stack 1779 // object. 1780 // * Atomic stores stronger that monotonic. 1781 bool isDSEBarrier(const Value *KillingUndObj, Instruction *DeadI) { 1782 // If DeadI may throw it acts as a barrier, unless we are to an 1783 // alloca/alloca like object that does not escape. 1784 if (DeadI->mayThrow() && !isInvisibleToCallerOnUnwind(KillingUndObj)) 1785 return true; 1786 1787 // If DeadI is an atomic load/store stronger than monotonic, do not try to 1788 // eliminate/reorder it. 1789 if (DeadI->isAtomic()) { 1790 if (auto *LI = dyn_cast<LoadInst>(DeadI)) 1791 return isStrongerThanMonotonic(LI->getOrdering()); 1792 if (auto *SI = dyn_cast<StoreInst>(DeadI)) 1793 return isStrongerThanMonotonic(SI->getOrdering()); 1794 if (auto *ARMW = dyn_cast<AtomicRMWInst>(DeadI)) 1795 return isStrongerThanMonotonic(ARMW->getOrdering()); 1796 if (auto *CmpXchg = dyn_cast<AtomicCmpXchgInst>(DeadI)) 1797 return isStrongerThanMonotonic(CmpXchg->getSuccessOrdering()) || 1798 isStrongerThanMonotonic(CmpXchg->getFailureOrdering()); 1799 llvm_unreachable("other instructions should be skipped in MemorySSA"); 1800 } 1801 return false; 1802 } 1803 1804 /// Eliminate writes to objects that are not visible in the caller and are not 1805 /// accessed before returning from the function. 1806 bool eliminateDeadWritesAtEndOfFunction() { 1807 bool MadeChange = false; 1808 LLVM_DEBUG( 1809 dbgs() 1810 << "Trying to eliminate MemoryDefs at the end of the function\n"); 1811 do { 1812 ShouldIterateEndOfFunctionDSE = false; 1813 for (MemoryDef *Def : llvm::reverse(MemDefs)) { 1814 if (SkipStores.contains(Def)) 1815 continue; 1816 1817 Instruction *DefI = Def->getMemoryInst(); 1818 auto DefLoc = getLocForWrite(DefI); 1819 if (!DefLoc || !isRemovable(DefI)) { 1820 LLVM_DEBUG(dbgs() << " ... could not get location for write or " 1821 "instruction not removable.\n"); 1822 continue; 1823 } 1824 1825 // NOTE: Currently eliminating writes at the end of a function is 1826 // limited to MemoryDefs with a single underlying object, to save 1827 // compile-time. In practice it appears the case with multiple 1828 // underlying objects is very uncommon. If it turns out to be important, 1829 // we can use getUnderlyingObjects here instead. 1830 const Value *UO = getUnderlyingObject(DefLoc->Ptr); 1831 if (!isInvisibleToCallerAfterRet(UO)) 1832 continue; 1833 1834 if (isWriteAtEndOfFunction(Def, *DefLoc)) { 1835 // See through pointer-to-pointer bitcasts 1836 LLVM_DEBUG(dbgs() << " ... MemoryDef is not accessed until the end " 1837 "of the function\n"); 1838 deleteDeadInstruction(DefI); 1839 ++NumFastStores; 1840 MadeChange = true; 1841 } 1842 } 1843 } while (ShouldIterateEndOfFunctionDSE); 1844 return MadeChange; 1845 } 1846 1847 /// If we have a zero initializing memset following a call to malloc, 1848 /// try folding it into a call to calloc. 1849 bool tryFoldIntoCalloc(MemoryDef *Def, const Value *DefUO) { 1850 Instruction *DefI = Def->getMemoryInst(); 1851 MemSetInst *MemSet = dyn_cast<MemSetInst>(DefI); 1852 if (!MemSet) 1853 // TODO: Could handle zero store to small allocation as well. 1854 return false; 1855 Constant *StoredConstant = dyn_cast<Constant>(MemSet->getValue()); 1856 if (!StoredConstant || !StoredConstant->isNullValue()) 1857 return false; 1858 1859 if (!isRemovable(DefI)) 1860 // The memset might be volatile.. 1861 return false; 1862 1863 if (F.hasFnAttribute(Attribute::SanitizeMemory) || 1864 F.hasFnAttribute(Attribute::SanitizeAddress) || 1865 F.hasFnAttribute(Attribute::SanitizeHWAddress) || 1866 F.getName() == "calloc") 1867 return false; 1868 auto *Malloc = const_cast<CallInst *>(dyn_cast<CallInst>(DefUO)); 1869 if (!Malloc) 1870 return false; 1871 auto *InnerCallee = Malloc->getCalledFunction(); 1872 if (!InnerCallee) 1873 return false; 1874 LibFunc Func; 1875 if (!TLI.getLibFunc(*InnerCallee, Func) || !TLI.has(Func) || 1876 Func != LibFunc_malloc) 1877 return false; 1878 // Gracefully handle malloc with unexpected memory attributes. 1879 auto *MallocDef = dyn_cast_or_null<MemoryDef>(MSSA.getMemoryAccess(Malloc)); 1880 if (!MallocDef) 1881 return false; 1882 1883 auto shouldCreateCalloc = [](CallInst *Malloc, CallInst *Memset) { 1884 // Check for br(icmp ptr, null), truebb, falsebb) pattern at the end 1885 // of malloc block 1886 auto *MallocBB = Malloc->getParent(), 1887 *MemsetBB = Memset->getParent(); 1888 if (MallocBB == MemsetBB) 1889 return true; 1890 auto *Ptr = Memset->getArgOperand(0); 1891 auto *TI = MallocBB->getTerminator(); 1892 ICmpInst::Predicate Pred; 1893 BasicBlock *TrueBB, *FalseBB; 1894 if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Ptr), m_Zero()), TrueBB, 1895 FalseBB))) 1896 return false; 1897 if (Pred != ICmpInst::ICMP_EQ || MemsetBB != FalseBB) 1898 return false; 1899 return true; 1900 }; 1901 1902 if (Malloc->getOperand(0) != MemSet->getLength()) 1903 return false; 1904 if (!shouldCreateCalloc(Malloc, MemSet) || 1905 !DT.dominates(Malloc, MemSet) || 1906 !memoryIsNotModifiedBetween(Malloc, MemSet, BatchAA, DL, &DT)) 1907 return false; 1908 IRBuilder<> IRB(Malloc); 1909 Type *SizeTTy = Malloc->getArgOperand(0)->getType(); 1910 auto *Calloc = emitCalloc(ConstantInt::get(SizeTTy, 1), 1911 Malloc->getArgOperand(0), IRB, TLI); 1912 if (!Calloc) 1913 return false; 1914 1915 MemorySSAUpdater Updater(&MSSA); 1916 auto *NewAccess = 1917 Updater.createMemoryAccessAfter(cast<Instruction>(Calloc), nullptr, 1918 MallocDef); 1919 auto *NewAccessMD = cast<MemoryDef>(NewAccess); 1920 Updater.insertDef(NewAccessMD, /*RenameUses=*/true); 1921 Malloc->replaceAllUsesWith(Calloc); 1922 deleteDeadInstruction(Malloc); 1923 return true; 1924 } 1925 1926 // Check if there is a dominating condition, that implies that the value 1927 // being stored in a ptr is already present in the ptr. 1928 bool dominatingConditionImpliesValue(MemoryDef *Def) { 1929 auto *StoreI = cast<StoreInst>(Def->getMemoryInst()); 1930 BasicBlock *StoreBB = StoreI->getParent(); 1931 Value *StorePtr = StoreI->getPointerOperand(); 1932 Value *StoreVal = StoreI->getValueOperand(); 1933 1934 DomTreeNode *IDom = DT.getNode(StoreBB)->getIDom(); 1935 if (!IDom) 1936 return false; 1937 1938 auto *BI = dyn_cast<BranchInst>(IDom->getBlock()->getTerminator()); 1939 if (!BI || !BI->isConditional()) 1940 return false; 1941 1942 // In case both blocks are the same, it is not possible to determine 1943 // if optimization is possible. (We would not want to optimize a store 1944 // in the FalseBB if condition is true and vice versa.) 1945 if (BI->getSuccessor(0) == BI->getSuccessor(1)) 1946 return false; 1947 1948 Instruction *ICmpL; 1949 ICmpInst::Predicate Pred; 1950 if (!match(BI->getCondition(), 1951 m_c_ICmp(Pred, 1952 m_CombineAnd(m_Load(m_Specific(StorePtr)), 1953 m_Instruction(ICmpL)), 1954 m_Specific(StoreVal))) || 1955 !ICmpInst::isEquality(Pred)) 1956 return false; 1957 1958 // In case the else blocks also branches to the if block or the other way 1959 // around it is not possible to determine if the optimization is possible. 1960 if (Pred == ICmpInst::ICMP_EQ && 1961 !DT.dominates(BasicBlockEdge(BI->getParent(), BI->getSuccessor(0)), 1962 StoreBB)) 1963 return false; 1964 1965 if (Pred == ICmpInst::ICMP_NE && 1966 !DT.dominates(BasicBlockEdge(BI->getParent(), BI->getSuccessor(1)), 1967 StoreBB)) 1968 return false; 1969 1970 MemoryAccess *LoadAcc = MSSA.getMemoryAccess(ICmpL); 1971 MemoryAccess *ClobAcc = 1972 MSSA.getSkipSelfWalker()->getClobberingMemoryAccess(Def, BatchAA); 1973 1974 return MSSA.dominates(ClobAcc, LoadAcc); 1975 } 1976 1977 /// \returns true if \p Def is a no-op store, either because it 1978 /// directly stores back a loaded value or stores zero to a calloced object. 1979 bool storeIsNoop(MemoryDef *Def, const Value *DefUO) { 1980 Instruction *DefI = Def->getMemoryInst(); 1981 StoreInst *Store = dyn_cast<StoreInst>(DefI); 1982 MemSetInst *MemSet = dyn_cast<MemSetInst>(DefI); 1983 Constant *StoredConstant = nullptr; 1984 if (Store) 1985 StoredConstant = dyn_cast<Constant>(Store->getOperand(0)); 1986 else if (MemSet) 1987 StoredConstant = dyn_cast<Constant>(MemSet->getValue()); 1988 else 1989 return false; 1990 1991 if (!isRemovable(DefI)) 1992 return false; 1993 1994 if (StoredConstant) { 1995 Constant *InitC = 1996 getInitialValueOfAllocation(DefUO, &TLI, StoredConstant->getType()); 1997 // If the clobbering access is LiveOnEntry, no instructions between them 1998 // can modify the memory location. 1999 if (InitC && InitC == StoredConstant) 2000 return MSSA.isLiveOnEntryDef( 2001 MSSA.getSkipSelfWalker()->getClobberingMemoryAccess(Def, BatchAA)); 2002 } 2003 2004 if (!Store) 2005 return false; 2006 2007 if (dominatingConditionImpliesValue(Def)) 2008 return true; 2009 2010 if (auto *LoadI = dyn_cast<LoadInst>(Store->getOperand(0))) { 2011 if (LoadI->getPointerOperand() == Store->getOperand(1)) { 2012 // Get the defining access for the load. 2013 auto *LoadAccess = MSSA.getMemoryAccess(LoadI)->getDefiningAccess(); 2014 // Fast path: the defining accesses are the same. 2015 if (LoadAccess == Def->getDefiningAccess()) 2016 return true; 2017 2018 // Look through phi accesses. Recursively scan all phi accesses by 2019 // adding them to a worklist. Bail when we run into a memory def that 2020 // does not match LoadAccess. 2021 SetVector<MemoryAccess *> ToCheck; 2022 MemoryAccess *Current = 2023 MSSA.getWalker()->getClobberingMemoryAccess(Def, BatchAA); 2024 // We don't want to bail when we run into the store memory def. But, 2025 // the phi access may point to it. So, pretend like we've already 2026 // checked it. 2027 ToCheck.insert(Def); 2028 ToCheck.insert(Current); 2029 // Start at current (1) to simulate already having checked Def. 2030 for (unsigned I = 1; I < ToCheck.size(); ++I) { 2031 Current = ToCheck[I]; 2032 if (auto PhiAccess = dyn_cast<MemoryPhi>(Current)) { 2033 // Check all the operands. 2034 for (auto &Use : PhiAccess->incoming_values()) 2035 ToCheck.insert(cast<MemoryAccess>(&Use)); 2036 continue; 2037 } 2038 2039 // If we found a memory def, bail. This happens when we have an 2040 // unrelated write in between an otherwise noop store. 2041 assert(isa<MemoryDef>(Current) && 2042 "Only MemoryDefs should reach here."); 2043 // TODO: Skip no alias MemoryDefs that have no aliasing reads. 2044 // We are searching for the definition of the store's destination. 2045 // So, if that is the same definition as the load, then this is a 2046 // noop. Otherwise, fail. 2047 if (LoadAccess != Current) 2048 return false; 2049 } 2050 return true; 2051 } 2052 } 2053 2054 return false; 2055 } 2056 2057 bool removePartiallyOverlappedStores(InstOverlapIntervalsTy &IOL) { 2058 bool Changed = false; 2059 for (auto OI : IOL) { 2060 Instruction *DeadI = OI.first; 2061 MemoryLocation Loc = *getLocForWrite(DeadI); 2062 assert(isRemovable(DeadI) && "Expect only removable instruction"); 2063 2064 const Value *Ptr = Loc.Ptr->stripPointerCasts(); 2065 int64_t DeadStart = 0; 2066 uint64_t DeadSize = Loc.Size.getValue(); 2067 GetPointerBaseWithConstantOffset(Ptr, DeadStart, DL); 2068 OverlapIntervalsTy &IntervalMap = OI.second; 2069 Changed |= tryToShortenEnd(DeadI, IntervalMap, DeadStart, DeadSize); 2070 if (IntervalMap.empty()) 2071 continue; 2072 Changed |= tryToShortenBegin(DeadI, IntervalMap, DeadStart, DeadSize); 2073 } 2074 return Changed; 2075 } 2076 2077 /// Eliminates writes to locations where the value that is being written 2078 /// is already stored at the same location. 2079 bool eliminateRedundantStoresOfExistingValues() { 2080 bool MadeChange = false; 2081 LLVM_DEBUG(dbgs() << "Trying to eliminate MemoryDefs that write the " 2082 "already existing value\n"); 2083 for (auto *Def : MemDefs) { 2084 if (SkipStores.contains(Def) || MSSA.isLiveOnEntryDef(Def)) 2085 continue; 2086 2087 Instruction *DefInst = Def->getMemoryInst(); 2088 auto MaybeDefLoc = getLocForWrite(DefInst); 2089 if (!MaybeDefLoc || !isRemovable(DefInst)) 2090 continue; 2091 2092 MemoryDef *UpperDef; 2093 // To conserve compile-time, we avoid walking to the next clobbering def. 2094 // Instead, we just try to get the optimized access, if it exists. DSE 2095 // will try to optimize defs during the earlier traversal. 2096 if (Def->isOptimized()) 2097 UpperDef = dyn_cast<MemoryDef>(Def->getOptimized()); 2098 else 2099 UpperDef = dyn_cast<MemoryDef>(Def->getDefiningAccess()); 2100 if (!UpperDef || MSSA.isLiveOnEntryDef(UpperDef)) 2101 continue; 2102 2103 Instruction *UpperInst = UpperDef->getMemoryInst(); 2104 auto IsRedundantStore = [&]() { 2105 if (DefInst->isIdenticalTo(UpperInst)) 2106 return true; 2107 if (auto *MemSetI = dyn_cast<MemSetInst>(UpperInst)) { 2108 if (auto *SI = dyn_cast<StoreInst>(DefInst)) { 2109 // MemSetInst must have a write location. 2110 auto UpperLoc = getLocForWrite(UpperInst); 2111 if (!UpperLoc) 2112 return false; 2113 int64_t InstWriteOffset = 0; 2114 int64_t DepWriteOffset = 0; 2115 auto OR = isOverwrite(UpperInst, DefInst, *UpperLoc, *MaybeDefLoc, 2116 InstWriteOffset, DepWriteOffset); 2117 Value *StoredByte = isBytewiseValue(SI->getValueOperand(), DL); 2118 return StoredByte && StoredByte == MemSetI->getOperand(1) && 2119 OR == OW_Complete; 2120 } 2121 } 2122 return false; 2123 }; 2124 2125 if (!IsRedundantStore() || isReadClobber(*MaybeDefLoc, DefInst)) 2126 continue; 2127 LLVM_DEBUG(dbgs() << "DSE: Remove No-Op Store:\n DEAD: " << *DefInst 2128 << '\n'); 2129 deleteDeadInstruction(DefInst); 2130 NumRedundantStores++; 2131 MadeChange = true; 2132 } 2133 return MadeChange; 2134 } 2135 }; 2136 2137 static bool eliminateDeadStores(Function &F, AliasAnalysis &AA, MemorySSA &MSSA, 2138 DominatorTree &DT, PostDominatorTree &PDT, 2139 const TargetLibraryInfo &TLI, 2140 const LoopInfo &LI) { 2141 bool MadeChange = false; 2142 2143 DSEState State(F, AA, MSSA, DT, PDT, TLI, LI); 2144 // For each store: 2145 for (unsigned I = 0; I < State.MemDefs.size(); I++) { 2146 MemoryDef *KillingDef = State.MemDefs[I]; 2147 if (State.SkipStores.count(KillingDef)) 2148 continue; 2149 Instruction *KillingI = KillingDef->getMemoryInst(); 2150 2151 std::optional<MemoryLocation> MaybeKillingLoc; 2152 if (State.isMemTerminatorInst(KillingI)) { 2153 if (auto KillingLoc = State.getLocForTerminator(KillingI)) 2154 MaybeKillingLoc = KillingLoc->first; 2155 } else { 2156 MaybeKillingLoc = State.getLocForWrite(KillingI); 2157 } 2158 2159 if (!MaybeKillingLoc) { 2160 LLVM_DEBUG(dbgs() << "Failed to find analyzable write location for " 2161 << *KillingI << "\n"); 2162 continue; 2163 } 2164 MemoryLocation KillingLoc = *MaybeKillingLoc; 2165 assert(KillingLoc.Ptr && "KillingLoc should not be null"); 2166 const Value *KillingUndObj = getUnderlyingObject(KillingLoc.Ptr); 2167 LLVM_DEBUG(dbgs() << "Trying to eliminate MemoryDefs killed by " 2168 << *KillingDef << " (" << *KillingI << ")\n"); 2169 2170 unsigned ScanLimit = MemorySSAScanLimit; 2171 unsigned WalkerStepLimit = MemorySSAUpwardsStepLimit; 2172 unsigned PartialLimit = MemorySSAPartialStoreLimit; 2173 // Worklist of MemoryAccesses that may be killed by KillingDef. 2174 SmallSetVector<MemoryAccess *, 8> ToCheck; 2175 // Track MemoryAccesses that have been deleted in the loop below, so we can 2176 // skip them. Don't use SkipStores for this, which may contain reused 2177 // MemoryAccess addresses. 2178 SmallPtrSet<MemoryAccess *, 8> Deleted; 2179 [[maybe_unused]] unsigned OrigNumSkipStores = State.SkipStores.size(); 2180 ToCheck.insert(KillingDef->getDefiningAccess()); 2181 2182 bool Shortend = false; 2183 bool IsMemTerm = State.isMemTerminatorInst(KillingI); 2184 // Check if MemoryAccesses in the worklist are killed by KillingDef. 2185 for (unsigned I = 0; I < ToCheck.size(); I++) { 2186 MemoryAccess *Current = ToCheck[I]; 2187 if (Deleted.contains(Current)) 2188 continue; 2189 2190 std::optional<MemoryAccess *> MaybeDeadAccess = State.getDomMemoryDef( 2191 KillingDef, Current, KillingLoc, KillingUndObj, ScanLimit, 2192 WalkerStepLimit, IsMemTerm, PartialLimit); 2193 2194 if (!MaybeDeadAccess) { 2195 LLVM_DEBUG(dbgs() << " finished walk\n"); 2196 continue; 2197 } 2198 2199 MemoryAccess *DeadAccess = *MaybeDeadAccess; 2200 LLVM_DEBUG(dbgs() << " Checking if we can kill " << *DeadAccess); 2201 if (isa<MemoryPhi>(DeadAccess)) { 2202 LLVM_DEBUG(dbgs() << "\n ... adding incoming values to worklist\n"); 2203 for (Value *V : cast<MemoryPhi>(DeadAccess)->incoming_values()) { 2204 MemoryAccess *IncomingAccess = cast<MemoryAccess>(V); 2205 BasicBlock *IncomingBlock = IncomingAccess->getBlock(); 2206 BasicBlock *PhiBlock = DeadAccess->getBlock(); 2207 2208 // We only consider incoming MemoryAccesses that come before the 2209 // MemoryPhi. Otherwise we could discover candidates that do not 2210 // strictly dominate our starting def. 2211 if (State.PostOrderNumbers[IncomingBlock] > 2212 State.PostOrderNumbers[PhiBlock]) 2213 ToCheck.insert(IncomingAccess); 2214 } 2215 continue; 2216 } 2217 auto *DeadDefAccess = cast<MemoryDef>(DeadAccess); 2218 Instruction *DeadI = DeadDefAccess->getMemoryInst(); 2219 LLVM_DEBUG(dbgs() << " (" << *DeadI << ")\n"); 2220 ToCheck.insert(DeadDefAccess->getDefiningAccess()); 2221 NumGetDomMemoryDefPassed++; 2222 2223 if (!DebugCounter::shouldExecute(MemorySSACounter)) 2224 continue; 2225 2226 MemoryLocation DeadLoc = *State.getLocForWrite(DeadI); 2227 2228 if (IsMemTerm) { 2229 const Value *DeadUndObj = getUnderlyingObject(DeadLoc.Ptr); 2230 if (KillingUndObj != DeadUndObj) 2231 continue; 2232 LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *DeadI 2233 << "\n KILLER: " << *KillingI << '\n'); 2234 State.deleteDeadInstruction(DeadI, &Deleted); 2235 ++NumFastStores; 2236 MadeChange = true; 2237 } else { 2238 // Check if DeadI overwrites KillingI. 2239 int64_t KillingOffset = 0; 2240 int64_t DeadOffset = 0; 2241 OverwriteResult OR = State.isOverwrite( 2242 KillingI, DeadI, KillingLoc, DeadLoc, KillingOffset, DeadOffset); 2243 if (OR == OW_MaybePartial) { 2244 auto Iter = State.IOLs.insert( 2245 std::make_pair<BasicBlock *, InstOverlapIntervalsTy>( 2246 DeadI->getParent(), InstOverlapIntervalsTy())); 2247 auto &IOL = Iter.first->second; 2248 OR = isPartialOverwrite(KillingLoc, DeadLoc, KillingOffset, 2249 DeadOffset, DeadI, IOL); 2250 } 2251 2252 if (EnablePartialStoreMerging && OR == OW_PartialEarlierWithFullLater) { 2253 auto *DeadSI = dyn_cast<StoreInst>(DeadI); 2254 auto *KillingSI = dyn_cast<StoreInst>(KillingI); 2255 // We are re-using tryToMergePartialOverlappingStores, which requires 2256 // DeadSI to dominate KillingSI. 2257 // TODO: implement tryToMergeParialOverlappingStores using MemorySSA. 2258 if (DeadSI && KillingSI && DT.dominates(DeadSI, KillingSI)) { 2259 if (Constant *Merged = tryToMergePartialOverlappingStores( 2260 KillingSI, DeadSI, KillingOffset, DeadOffset, State.DL, 2261 State.BatchAA, &DT)) { 2262 2263 // Update stored value of earlier store to merged constant. 2264 DeadSI->setOperand(0, Merged); 2265 ++NumModifiedStores; 2266 MadeChange = true; 2267 2268 Shortend = true; 2269 // Remove killing store and remove any outstanding overlap 2270 // intervals for the updated store. 2271 State.deleteDeadInstruction(KillingSI, &Deleted); 2272 auto I = State.IOLs.find(DeadSI->getParent()); 2273 if (I != State.IOLs.end()) 2274 I->second.erase(DeadSI); 2275 break; 2276 } 2277 } 2278 } 2279 2280 if (OR == OW_Complete) { 2281 LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *DeadI 2282 << "\n KILLER: " << *KillingI << '\n'); 2283 State.deleteDeadInstruction(DeadI, &Deleted); 2284 ++NumFastStores; 2285 MadeChange = true; 2286 } 2287 } 2288 } 2289 2290 assert(State.SkipStores.size() - OrigNumSkipStores == Deleted.size() && 2291 "SkipStores and Deleted out of sync?"); 2292 2293 // Check if the store is a no-op. 2294 if (!Shortend && State.storeIsNoop(KillingDef, KillingUndObj)) { 2295 LLVM_DEBUG(dbgs() << "DSE: Remove No-Op Store:\n DEAD: " << *KillingI 2296 << '\n'); 2297 State.deleteDeadInstruction(KillingI); 2298 NumRedundantStores++; 2299 MadeChange = true; 2300 continue; 2301 } 2302 2303 // Can we form a calloc from a memset/malloc pair? 2304 if (!Shortend && State.tryFoldIntoCalloc(KillingDef, KillingUndObj)) { 2305 LLVM_DEBUG(dbgs() << "DSE: Remove memset after forming calloc:\n" 2306 << " DEAD: " << *KillingI << '\n'); 2307 State.deleteDeadInstruction(KillingI); 2308 MadeChange = true; 2309 continue; 2310 } 2311 } 2312 2313 if (EnablePartialOverwriteTracking) 2314 for (auto &KV : State.IOLs) 2315 MadeChange |= State.removePartiallyOverlappedStores(KV.second); 2316 2317 MadeChange |= State.eliminateRedundantStoresOfExistingValues(); 2318 MadeChange |= State.eliminateDeadWritesAtEndOfFunction(); 2319 2320 while (!State.ToRemove.empty()) { 2321 Instruction *DeadInst = State.ToRemove.pop_back_val(); 2322 DeadInst->eraseFromParent(); 2323 } 2324 2325 return MadeChange; 2326 } 2327 } // end anonymous namespace 2328 2329 //===----------------------------------------------------------------------===// 2330 // DSE Pass 2331 //===----------------------------------------------------------------------===// 2332 PreservedAnalyses DSEPass::run(Function &F, FunctionAnalysisManager &AM) { 2333 AliasAnalysis &AA = AM.getResult<AAManager>(F); 2334 const TargetLibraryInfo &TLI = AM.getResult<TargetLibraryAnalysis>(F); 2335 DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(F); 2336 MemorySSA &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA(); 2337 PostDominatorTree &PDT = AM.getResult<PostDominatorTreeAnalysis>(F); 2338 LoopInfo &LI = AM.getResult<LoopAnalysis>(F); 2339 2340 bool Changed = eliminateDeadStores(F, AA, MSSA, DT, PDT, TLI, LI); 2341 2342 #ifdef LLVM_ENABLE_STATS 2343 if (AreStatisticsEnabled()) 2344 for (auto &I : instructions(F)) 2345 NumRemainingStores += isa<StoreInst>(&I); 2346 #endif 2347 2348 if (!Changed) 2349 return PreservedAnalyses::all(); 2350 2351 PreservedAnalyses PA; 2352 PA.preserveSet<CFGAnalyses>(); 2353 PA.preserve<MemorySSAAnalysis>(); 2354 PA.preserve<LoopAnalysis>(); 2355 return PA; 2356 } 2357