xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
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