[GlobalOpt] Make global SRA offset based

[nikic]

Summary

Currently global SRA uses the GEP structure to determine how to split the global. This patch instead analyses the loads and stores that are performed on the global, and collects which types are used at which offset, and then splits the global according to those.

This is both more general, and works fine with opaque pointers. This is also closer to how ordinary SROA is performed.

[aeubanks]

looks like DebugInfo/Generic/global-sra-array.ll is failing

[nikic]

Update debuginfo test.

[nikic]

Rebase over additional test with [N x i8] type.

[nikic]

Protect against recursive SRA.

[reames]

Hm, your new code looks correct to me.

There's one case in the original code I don't see reflected in the new code though. It doesn't appear to have any tests, and I'm not 100% sure it's reachable at all, but consider the following.

An outer struct with an inner array. The index expression has a non-constant operand for the offset into the array. (So, gep g, 0, v) I believe the existing code will try to strip the struct off, and replace the global with the array. Your code, because of the restriction to entirely constant offsets, will not.

In an opaque pointer world, this example can't arise. If you wanted to add your code without removing the existing, and put the old code under a non-opaque pointer check, I'd be fine with that. Though I'm really hoping you tell me I'm missing something and this case is impossible. :)

[nikic]

In D117223#3243963, @reames wrote:

Hm, your new code looks correct to me.

There's one case in the original code I don't see reflected in the new code though. It doesn't appear to have any tests, and I'm not 100% sure it's reachable at all, but consider the following.

An outer struct with an inner array. The index expression has a non-constant operand for the offset into the array. (So, gep g, 0, v) I believe the existing code will try to strip the struct off, and replace the global with the array. Your code, because of the restriction to entirely constant offsets, will not.

The old code doesn't handle this either, because isSafeSROAGEP() checks that all indices are constant integers. Besides, doing this optimization without constant indices would be illegal, because there may be notional overindexing (this would only be legal with inrange annotations, which is not something the old code tries to handle -- GlobalSplit is the optimization that does).

[reames]

In D117223#3244011, @nikic wrote:

The old code doesn't handle this either, because isSafeSROAGEP() checks that all indices are constant integers.

What had confused me is that we have an early continue in the loop for structs which seemed to allow non-constant operands. I just realized that this is relying on the verifier to disallow non-constant struct indices. Dang that is subtle for no particular reason.

Given that LGTM

[uabelho]

Hello,

We've seen a case where the output from opt is different with/without debug info and ended up bisection at this patch.
After reduction and fiddling it seems like it doesn't actually have to do with debug info at all but something else. Our best reproducer so far is as below.

If we run sroa, early-cse and globalopt like this:

opt -passes='function(sroa,early-cse),invalidate<all>,globalopt' -debug-only=globalopt bbi-65060.ll -o /dev/null

we get one set of printouts and if we run like this where we run the same passes but in two different invocations:

opt -passes='function(sroa,early-cse),invalidate<all>' bbi-65060.ll | opt -passes=globalopt -debug-only=globalopt -o /dev/null

we get something else.

Notably is that with the second way, with the split pipeline we see

PERFORMING GLOBAL SRA ON: @g_93 = internal unnamed_addr global [6 x i16*] undef
MARKING CONSTANT: @g_93.0 = internal unnamed_addr global i16* undef
   *** Marking constant allowed us to simplify all users and delete global!
GLOBAL NEVER LOADED: @g_93.1 = internal unnamed_addr global i16* undef

which we don't see when we run all passes in the same pipe.

This has been very elusive and hard to reproduce and reduce so we wouldn't be surprised if it's connected to some memory issue, iteration order or something like that.

bbi-65060.ll2 KBDownload

[nikic]

@uabelho Thanks for the report. The issue is related to a handling of dead constant users in GlobalStatus, which is why the two opt invocations are relevant (dead constant users will get dropped).

[nikic]

@uabelho Should be fixed by https://github.com/llvm/llvm-project/commit/236fbf571dc6cebcb81ac5187a170c8de6d5bc0e.

[uabelho]

In D117223#3287388, @nikic wrote:

@uabelho Should be fixed by https://github.com/llvm/llvm-project/commit/236fbf571dc6cebcb81ac5187a170c8de6d5bc0e.

Cool! It seems the fix solves the original problem we saw with different result with/without debug info as well.
Thank you!

[uabelho]

In D117223#3287388, @nikic wrote:

@uabelho Should be fixed by https://github.com/llvm/llvm-project/commit/236fbf571dc6cebcb81ac5187a170c8de6d5bc0e.

Unfortunately I've now stumbled upon a case where the compiler starts to hang with that fix:
clang -O2 --target=x86_64-unknown-linux-gnu -o /dev/null bbi-66447.c

[uabelho]

and here's the bbi-66447.c

bbi-66447.c166 KBDownload

[nikic]

@uabelho Thanks for the report, should be fixed by https://github.com/llvm/llvm-project/commit/f8d72100323b127b39fed5f5890b86d81b8a0335.

[uabelho]

In D117223#3337089, @nikic wrote:

@uabelho Thanks for the report, should be fixed by https://github.com/llvm/llvm-project/commit/f8d72100323b127b39fed5f5890b86d81b8a0335.

Yep, thanks!

[fhahn]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

435

I think this limit is more aggressive for struct types than the original code AFAICT, which seems to be causing some code-size regressions. I put up D129525 to make the limit behave more like the original code.

llvm/lib/Transforms/IPO/GlobalOpt.cpp

static bool CleanupConstantGlobalUsers(GlobalVariable *GV,
   }
 
   Changed |=
       RecursivelyDeleteTriviallyDeadInstructionsPermissive(MaybeDeadInsts);
   GV->removeDeadConstantUsers();
   return Changed;
 }
 
-static bool isSafeSROAElementUse(Value *V);
-
-/// Return true if the specified GEP is a safe user of a derived
-/// expression from a global that we want to SROA.
-static bool isSafeSROAGEP(User *U) {
-  // Check to see if this ConstantExpr GEP is SRA'able.  In particular, we
-  // don't like < 3 operand CE's, and we don't like non-constant integer
-  // indices.  This enforces that all uses are 'gep GV, 0, C, ...' for some
-  // value of C.
-  if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
-      !cast<Constant>(U->getOperand(1))->isNullValue())
-    return false;
-
-  gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
-  ++GEPI; // Skip over the pointer index.
-
-  // For all other level we require that the indices are constant and inrange.
-  // In particular, consider: A[0][i].  We cannot know that the user isn't doing
-  // invalid things like allowing i to index an out-of-range subscript that
-  // accesses A[1]. This can also happen between different members of a struct
-  // in llvm IR.
-  for (; GEPI != E; ++GEPI) {
-    if (GEPI.isStruct())
+/// Look at all uses of the global and determine which (offset, type) pairs it
+/// can be split into.
+static bool collectSRATypes(DenseMap<uint64_t, Type *> &Types, GlobalValue *GV,
+                            const DataLayout &DL) {
+  SmallVector<Use *, 16> Worklist;
+  SmallPtrSet<Use *, 16> Visited;
+  auto AppendUses = [&](Value *V) {
+    for (Use &U : V->uses())
+      if (Visited.insert(&U).second)
+        Worklist.push_back(&U);
+  };
+  AppendUses(GV);
+  while (!Worklist.empty()) {
+    Use *U = Worklist.pop_back_val();
+    User *V = U->getUser();
+    if (isa<BitCastOperator>(V) || isa<AddrSpaceCastOperator>(V)) {
+      AppendUses(V);
       continue;
-
-    ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
-    if (!IdxVal || (GEPI.isBoundedSequential() &&
-                    IdxVal->getZExtValue() >= GEPI.getSequentialNumElements()))
-      return false;
     }
 
-  return llvm::all_of(U->users(), isSafeSROAElementUse);
-}
-
-/// Return true if the specified instruction is a safe user of a derived
-/// expression from a global that we want to SROA.
-static bool isSafeSROAElementUse(Value *V) {
-  // We might have a dead and dangling constant hanging off of here.
-  if (Constant *C = dyn_cast<Constant>(V))
-    return isSafeToDestroyConstant(C);
-
-  Instruction *I = dyn_cast<Instruction>(V);
-  if (!I) return false;
-
-  // Loads are ok.
-  if (isa<LoadInst>(I)) return true;
-
-  // Stores *to* the pointer are ok.
-  if (StoreInst *SI = dyn_cast<StoreInst>(I))
-    return SI->getOperand(0) != V;
-
-  // Otherwise, it must be a GEP. Check it and its users are safe to SRA.
-  return isa<GetElementPtrInst>(I) && isSafeSROAGEP(I);
+    if (auto *GEP = dyn_cast<GEPOperator>(V)) {
+      if (!GEP->hasAllConstantIndices())
+        return false;
+      AppendUses(V);
+      continue;
     }
 
-/// Look at all uses of the global and decide whether it is safe for us to
-/// perform this transformation.
-static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
-  for (User *U : GV->users()) {
-    // The user of the global must be a GEP Inst or a ConstantExpr GEP.
-    if (!isa<GetElementPtrInst>(U) &&
-        (!isa<ConstantExpr>(U) ||
-        cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
+    if (Value *Ptr = getLoadStorePointerOperand(V)) {
+      // This is storing the global address into somewhere, not storing into
+      // the global.
+      if (isa<StoreInst>(V) && U->getOperandNo() == 0)
         return false;
 
-    // Check the gep and it's users are safe to SRA
-    if (!isSafeSROAGEP(U))
+      APInt Offset(DL.getIndexTypeSizeInBits(Ptr->getType()), 0);
+      Ptr = Ptr->stripAndAccumulateConstantOffsets(DL, Offset,
+                                                   /* AllowNonInbounds */ true);
+      if (Ptr != GV || Offset.getActiveBits() >= 64)
         return false;
-  }
 
-  return true;
+      // TODO: We currently require that all accesses at a given offset must
+      // use the same type. This could be relaxed.
+      Type *Ty = getLoadStoreType(V);
+      auto It = Types.try_emplace(Offset.getZExtValue(), Ty).first;
+      if (Ty != It->second)
+        return false;
+      continue;
     }
 
-static bool IsSRASequential(Type *T) {
-  return isa<ArrayType>(T) || isa<VectorType>(T);
-}
-static uint64_t GetSRASequentialNumElements(Type *T) {
-  if (ArrayType *AT = dyn_cast<ArrayType>(T))
-    return AT->getNumElements();
-  return cast<FixedVectorType>(T)->getNumElements();
-}
-static Type *GetSRASequentialElementType(Type *T) {
-  if (ArrayType *AT = dyn_cast<ArrayType>(T))
-    return AT->getElementType();
-  return cast<VectorType>(T)->getElementType();
+    // Ignore dead constant users.
+    if (auto *C = dyn_cast<Constant>(V)) {
+      if (!isSafeToDestroyConstant(C))
+        return false;
+      continue;
     }
-static bool CanDoGlobalSRA(GlobalVariable *GV) {
-  Constant *Init = GV->getInitializer();
 
-  if (isa<StructType>(Init->getType())) {
-    // nothing to check
-  } else if (IsSRASequential(Init->getType())) {
-    if (GetSRASequentialNumElements(Init->getType()) > 16 &&
-        GV->hasNUsesOrMore(16))
-      return false; // It's not worth it.
-  } else
+    // Unknown user.
     return false;
+  }
 
-  return GlobalUsersSafeToSRA(GV);
+  return true;
 }
 
 /// Copy over the debug info for a variable to its SRA replacements.
 static void transferSRADebugInfo(GlobalVariable *GV, GlobalVariable *NGV,
                                  uint64_t FragmentOffsetInBits,
                                  uint64_t FragmentSizeInBits,
                                  uint64_t VarSize) {
   SmallVector<DIGlobalVariableExpression *, 1> GVs;

 }
 
 /// Perform scalar replacement of aggregates on the specified global variable.
 /// This opens the door for other optimizations by exposing the behavior of the
 /// program in a more fine-grained way.  We have determined that this
 /// transformation is safe already.  We return the first global variable we
 /// insert so that the caller can reprocess it.
 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const DataLayout &DL) {
-  // Make sure this global only has simple uses that we can SRA.
-  if (!CanDoGlobalSRA(GV))
+  assert(GV->hasLocalLinkage());
+
+  // Collect types to split into.
+  DenseMap<uint64_t, Type *> Types;
+  if (!collectSRATypes(Types, GV, DL))
     return nullptr;
 
-  assert(GV->hasLocalLinkage());
-  Constant *Init = GV->getInitializer();
-  Type *Ty = Init->getType();
-  uint64_t VarSize = DL.getTypeSizeInBits(Ty);
+  // Don't perform SRA if we would have to split into many globals.
+  if (Types.size() > 16)
+    return nullptr;
 
-  std::map<unsigned, GlobalVariable *> NewGlobals;
+  // Sort by offset.
+  SmallVector<std::pair<uint64_t, Type *>, 16> TypesVector;
+  append_range(TypesVector, Types);
+  sort(TypesVector,
+       [](const auto &A, const auto &B) { return A.first < B.first; });
+
+  // Check that the types are non-overlapping.
+  uint64_t Offset = 0;
+  for (const auto &Pair : TypesVector) {
+    // Overlaps with previous type.
+    if (Pair.first < Offset)
+      return nullptr;
 
-  // Get the alignment of the global, either explicit or target-specific.
-  Align StartAlignment =
-      DL.getValueOrABITypeAlignment(GV->getAlign(), GV->getType());
+    Offset = Pair.first + DL.getTypeAllocSize(Pair.second);
+  }
 
-  // Loop over all users and create replacement variables for used aggregate
-  // elements.
-  for (User *GEP : GV->users()) {
-    assert(((isa<ConstantExpr>(GEP) && cast<ConstantExpr>(GEP)->getOpcode() ==
-                                           Instruction::GetElementPtr) ||
-            isa<GetElementPtrInst>(GEP)) &&
-           "NonGEP CE's are not SRAable!");
-
-    // Ignore the 1th operand, which has to be zero or else the program is quite
-    // broken (undefined).  Get the 2nd operand, which is the structure or array
-    // index.
-    unsigned ElementIdx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
-    if (NewGlobals.count(ElementIdx) == 1)
-      continue; // we`ve already created replacement variable
-    assert(NewGlobals.count(ElementIdx) == 0);
-
-    Type *ElTy = nullptr;
-    if (StructType *STy = dyn_cast<StructType>(Ty))
-      ElTy = STy->getElementType(ElementIdx);
-    else
-      ElTy = GetSRASequentialElementType(Ty);
-    assert(ElTy);
+  // Some accesses go beyond the end of the global, don't bother.
+  if (Offset > DL.getTypeAllocSize(GV->getValueType()))
+    return nullptr;
+
+  // Collect initializers for new globals.
+  Constant *OrigInit = GV->getInitializer();
+  DenseMap<uint64_t, Constant *> Initializers;
+  for (const auto &Pair : Types) {
+    Constant *NewInit = ConstantFoldLoadFromConst(OrigInit, Pair.second,
+                                                  APInt(64, Pair.first), DL);
+    if (!NewInit) {
+      LLVM_DEBUG(dbgs() << "Global SRA: Failed to evaluate initializer of "
+                        << *GV << " with type " << *Pair.second << " at offset "
+                        << Pair.first << "\n");
+      return nullptr;
+    }
+    Initializers.insert({Pair.first, NewInit});
+  }
 
-    Constant *In = Init->getAggregateElement(ElementIdx);
-    assert(In && "Couldn't get element of initializer?");
+  LLVM_DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV << "\n");
 
+  // Get the alignment of the global, either explicit or target-specific.
+  Align StartAlignment =
+      DL.getValueOrABITypeAlignment(GV->getAlign(), GV->getValueType());
+  uint64_t VarSize = DL.getTypeSizeInBits(GV->getValueType());
+
+  // Create replacement globals.
+  DenseMap<uint64_t, GlobalVariable *> NewGlobals;
+  unsigned NameSuffix = 0;
+  for (auto &Pair : TypesVector) {
+    uint64_t Offset = Pair.first;
+    Type *Ty = Pair.second;
     GlobalVariable *NGV = new GlobalVariable(
-        ElTy, false, GlobalVariable::InternalLinkage, In,
-        GV->getName() + "." + Twine(ElementIdx), GV->getThreadLocalMode(),
-        GV->getType()->getAddressSpace());
+        *GV->getParent(), Ty, false, GlobalVariable::InternalLinkage,
+        Initializers[Offset], GV->getName() + "." + Twine(NameSuffix++), GV,
+        GV->getThreadLocalMode(), GV->getAddressSpace());
     NGV->copyAttributesFrom(GV);
-    NewGlobals.insert(std::make_pair(ElementIdx, NGV));
-
-    if (StructType *STy = dyn_cast<StructType>(Ty)) {
-      const StructLayout &Layout = *DL.getStructLayout(STy);
+    NewGlobals.insert({Offset, NGV});
 
-      // Calculate the known alignment of the field.  If the original aggregate
-      // had 256 byte alignment for example, something might depend on that:
-      // propagate info to each field.
-      uint64_t FieldOffset = Layout.getElementOffset(ElementIdx);
-      Align NewAlign = commonAlignment(StartAlignment, FieldOffset);
-      if (NewAlign > DL.getABITypeAlign(STy->getElementType(ElementIdx)))
-        NGV->setAlignment(NewAlign);
+    Align NewAlign = commonAlignment(StartAlignment, Offset);
+    if (NewAlign > DL.getABITypeAlign(Ty))
 
-      // Copy over the debug info for the variable.
-      uint64_t Size = DL.getTypeAllocSizeInBits(NGV->getValueType());
-      uint64_t FragmentOffsetInBits = Layout.getElementOffsetInBits(ElementIdx);
-      transferSRADebugInfo(GV, NGV, FragmentOffsetInBits, Size, VarSize);
-    } else {
-      uint64_t EltSize = DL.getTypeAllocSize(ElTy);
-      Align EltAlign = DL.getABITypeAlign(ElTy);
-      uint64_t FragmentSizeInBits = DL.getTypeAllocSizeInBits(ElTy);
-
-      // Calculate the known alignment of the field.  If the original aggregate
-      // had 256 byte alignment for example, something might depend on that:
-      // propagate info to each field.
-      Align NewAlign = commonAlignment(StartAlignment, EltSize * ElementIdx);
-      if (NewAlign > EltAlign)
-        NGV->setAlignment(NewAlign);
-      transferSRADebugInfo(GV, NGV, FragmentSizeInBits * ElementIdx,
-                           FragmentSizeInBits, VarSize);
-    }
+    transferSRADebugInfo(GV, NGV, Offset * 8, DL.getTypeAllocSizeInBits(Ty),
+                         VarSize);
   }
 
-  if (NewGlobals.empty())
-    return nullptr;
-
-  Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
-  for (auto NewGlobalVar : NewGlobals)
-    Globals.push_back(NewGlobalVar.second);
-
-  LLVM_DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV << "\n");
-
-  Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
+  // Replace uses of the original global with uses of the new global.
+  SmallVector<Value *, 16> Worklist;
+  SmallPtrSet<Value *, 16> Visited;
+  SmallVector<WeakTrackingVH, 16> DeadInsts;
+  auto AppendUsers = [&](Value *V) {
+    for (User *U : V->users())
+      if (Visited.insert(U).second)
+        Worklist.push_back(U);
+  };
+  AppendUsers(GV);
+  while (!Worklist.empty()) {
+    Value *V = Worklist.pop_back_val();
+    if (isa<BitCastOperator>(V) || isa<AddrSpaceCastOperator>(V) ||
+        isa<GEPOperator>(V)) {
+      AppendUsers(V);
+      if (isa<Instruction>(V))
+        DeadInsts.push_back(V);
+      continue;
+    }
 
-  // Loop over all of the uses of the global, replacing the constantexpr geps,
-  // with smaller constantexpr geps or direct references.
-  while (!GV->use_empty()) {
-    User *GEP = GV->user_back();
-    assert(((isa<ConstantExpr>(GEP) &&
-             cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
-            isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
-
-    // Ignore the 1th operand, which has to be zero or else the program is quite
-    // broken (undefined).  Get the 2nd operand, which is the structure or array
-    // index.
-    unsigned ElementIdx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
-    assert(NewGlobals.count(ElementIdx) == 1);
-
-    Value *NewPtr = NewGlobals[ElementIdx];
-    Type *NewTy = NewGlobals[ElementIdx]->getValueType();
-
-    // Form a shorter GEP if needed.
-    if (GEP->getNumOperands() > 3) {
-      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
-        SmallVector<Constant*, 8> Idxs;
-        Idxs.push_back(NullInt);
-        for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
-          Idxs.push_back(CE->getOperand(i));
-        NewPtr =
-            ConstantExpr::getGetElementPtr(NewTy, cast<Constant>(NewPtr), Idxs);
+    if (Value *Ptr = getLoadStorePointerOperand(V)) {
+      APInt Offset(DL.getIndexTypeSizeInBits(Ptr->getType()), 0);
+      Ptr = Ptr->stripAndAccumulateConstantOffsets(DL, Offset,
+                                                   /* AllowNonInbounds */ true);
+      assert(Ptr == GV && "Load/store must be from/to global");
+      GlobalVariable *NGV = NewGlobals[Offset.getZExtValue()];
+      assert(NGV && "Must have replacement global for this offset");
+
+      // Update the pointer operand and recalculate alignment.
+      Align PrefAlign = DL.getPrefTypeAlign(getLoadStoreType(V));
+      Align NewAlign =
+          getOrEnforceKnownAlignment(NGV, PrefAlign, DL, cast<Instruction>(V));
+
+      if (auto *LI = dyn_cast<LoadInst>(V)) {
+        LI->setOperand(0, NGV);
+        LI->setAlignment(NewAlign);
       } else {
-        GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
-        SmallVector<Value*, 8> Idxs;
-        Idxs.push_back(NullInt);
-        for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
-          Idxs.push_back(GEPI->getOperand(i));
-        NewPtr = GetElementPtrInst::Create(
-            NewTy, NewPtr, Idxs, GEPI->getName() + "." + Twine(ElementIdx),
-            GEPI);
-      }
-    }
-    GEP->replaceAllUsesWith(NewPtr);
-
-    // We changed the pointer of any memory access user. Recalculate alignments.
-    for (User *U : NewPtr->users()) {
-      if (auto *Load = dyn_cast<LoadInst>(U)) {
-        Align PrefAlign = DL.getPrefTypeAlign(Load->getType());
-        Align NewAlign = getOrEnforceKnownAlignment(Load->getPointerOperand(),
-                                                    PrefAlign, DL, Load);
-        Load->setAlignment(NewAlign);
-      }
-      if (auto *Store = dyn_cast<StoreInst>(U)) {
-        Align PrefAlign =
-            DL.getPrefTypeAlign(Store->getValueOperand()->getType());
-        Align NewAlign = getOrEnforceKnownAlignment(Store->getPointerOperand(),
-                                                    PrefAlign, DL, Store);
-        Store->setAlignment(NewAlign);
+        auto *SI = cast<StoreInst>(V);
+        SI->setOperand(1, NGV);
+        SI->setAlignment(NewAlign);
       }
+      continue;
     }
 
-    if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
-      GEPI->eraseFromParent();
-    else
-      cast<ConstantExpr>(GEP)->destroyConstant();
+    assert(isa<Constant>(V) && isSafeToDestroyConstant(cast<Constant>(V)) &&
+           "Other users can only be dead constants");
   }
 
-  // Delete the old global, now that it is dead.
-  Globals.erase(GV);
+  // Delete old instructions and global.
+  RecursivelyDeleteTriviallyDeadInstructions(DeadInsts);
+  GV->removeDeadConstantUsers();
+  GV->eraseFromParent();
   ++NumSRA;
 
   assert(NewGlobals.size() > 0);
   return NewGlobals.begin()->second;
 }
 
 /// Return true if all users of the specified value will trap if the value is
 /// dynamically null.  PHIs keeps track of any phi nodes we've seen to avoid

llvm/test/DebugInfo/Generic/global-sra-array.ll

 ;   return array[0].a + array[1].a;
 ; }
 
 %struct.anon = type { i32, i8 }
 
 ; This array is first split into two struct, which are then split into their
 ; elements, of which only .a survives.
 @array = internal global [2 x %struct.anon] zeroinitializer, align 16, !dbg !0
-; CHECK: @array.0.0 = internal unnamed_addr global i32 0, align 16, !dbg ![[EL0:.*]]
-; CHECK: @array.1.0 = internal unnamed_addr global i32 0, align 8, !dbg ![[EL1:.*]]
-;
+; CHECK: @array.0 = internal unnamed_addr global i32 0, align 16, !dbg ![[EL0:.*]]
+; CHECK: @array.1 = internal unnamed_addr global i32 0, align 8, !dbg ![[EL1:.*]]
+
 ; CHECK: ![[EL0]] = !DIGlobalVariableExpression(var: ![[VAR:.*]], expr: !DIExpression(DW_OP_LLVM_fragment, 0, 32))
 ; CHECK: ![[VAR]] = distinct !DIGlobalVariable(name: "array"
 ; CHECK: ![[EL1]] = !DIGlobalVariableExpression(var: ![[VAR]], expr: !DIExpression(DW_OP_LLVM_fragment, 64, 32))
 
 
 ; Function Attrs: nounwind optsize ssp uwtable
 define void @foo(i32 %in) #0 {
 entry:

llvm/test/Transforms/GlobalOpt/externally-initialized-global-ctr.ll

 }
 
 declare void @test(i8*)
 
 ; The preferred alignment is available.
 
 define void @print() {
 ; CHECK-LABEL: @print(
-; CHECK-NEXT:    [[TMP1:%.*]] = load i8*, i8** @_ZL14buttonInitData.0.0, align 16
+; CHECK-NEXT:    [[TMP1:%.*]] = load i8*, i8** @_ZL14buttonInitData.0, align 16
 ; CHECK-NEXT:    call void @test(i8* [[TMP1]])
 ; CHECK-NEXT:    ret void
 ;
   %1 = load i8*, i8** getelementptr inbounds ([1 x %struct.ButtonInitData], [1 x %struct.ButtonInitData]* @_ZL14buttonInitData, i32 0, i32 0, i32 0), align 4
   call void @test(i8* %1)
   ret void
 }
 
 !2009 = !{}

llvm/test/Transforms/GlobalOpt/globalsra-align.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals
 ; RUN: opt < %s -passes=globalopt -S | FileCheck %s
 
 target datalayout = "p:16:32:64" ; 16-bit pointers with 32-bit ABI alignment and 64-bit preferred alignmentt
 
 @a = internal externally_initialized global [3 x [7 x i32*]] zeroinitializer, align 16
 
 ; PR50253
 ; The alignments are correct initially, but they should be updated
 ; after transforming the global. The stored global pointer array retains
 ; its original "align 16", so access to element N into the new array
 ; should be offset by the ABI alignment of N pointers.
 ; Loaded globals are split into individual pointers and use the
 ; preferred alignment from the datalayout.
 
 ;.
-; CHECK: @[[A_1:[a-zA-Z0-9_$"\\.-]+]] = internal unnamed_addr externally_initialized global [7 x i32*] zeroinitializer, align 16
-; CHECK: @[[A_2_0:[a-zA-Z0-9_$"\\.-]+]] = internal unnamed_addr externally_initialized global i32* null, align 8
-; CHECK: @[[A_2_1:[a-zA-Z0-9_$"\\.-]+]] = internal unnamed_addr externally_initialized global i32* null, align 8
-; CHECK: @[[A_2_2:[a-zA-Z0-9_$"\\.-]+]] = internal unnamed_addr externally_initialized global i32* null, align 8
-; CHECK: @[[A_2_3:[a-zA-Z0-9_$"\\.-]+]] = internal unnamed_addr externally_initialized global i32* null, align 8
+; CHECK: @[[A_4:[a-zA-Z0-9_$"\\.-]+]] = internal unnamed_addr externally_initialized global i32* null, align 8
+; CHECK: @[[A_5:[a-zA-Z0-9_$"\\.-]+]] = internal unnamed_addr externally_initialized global i32* null, align 16
+; CHECK: @[[A_6:[a-zA-Z0-9_$"\\.-]+]] = internal unnamed_addr externally_initialized global i32* null, align 16
+; CHECK: @[[A_7:[a-zA-Z0-9_$"\\.-]+]] = internal unnamed_addr externally_initialized global i32* null, align 16
 ;.
 define i32* @reduce_align_0() {
 ; CHECK-LABEL: @reduce_align_0(
-; CHECK-NEXT:    [[X:%.*]] = load i32*, i32** @a.2.0, align 8
-; CHECK-NEXT:    store i32* null, i32** getelementptr inbounds ([7 x i32*], [7 x i32*]* @a.1, i32 0, i64 0), align 16
+; CHECK-NEXT:    [[X:%.*]] = load i32*, i32** @a.4, align 8
 ; CHECK-NEXT:    ret i32* [[X]]
 ;
   %x = load i32*, i32** getelementptr inbounds ([3 x [7 x i32*]], [3 x [7 x i32*]]* @a, i64 0, i64 2, i64 0), align 8
   store i32* null, i32** getelementptr inbounds ([3 x [7 x i32*]], [3 x [7 x i32*]]* @a, i64 0, i64 1, i64 0), align 4
   ret i32* %x
 }
 
 define i32* @reduce_align_1() {
 ; CHECK-LABEL: @reduce_align_1(
-; CHECK-NEXT:    [[X:%.*]] = load i32*, i32** @a.2.1, align 8
-; CHECK-NEXT:    store i32* null, i32** getelementptr inbounds ([7 x i32*], [7 x i32*]* @a.1, i32 0, i64 1), align 4
+; CHECK-NEXT:    [[X:%.*]] = load i32*, i32** @a.5, align 16
 ; CHECK-NEXT:    ret i32* [[X]]
 ;
   %x = load i32*, i32** getelementptr inbounds ([3 x [7 x i32*]], [3 x [7 x i32*]]* @a, i64 0, i64 2, i64 1), align 4
   store i32* null, i32** getelementptr inbounds ([3 x [7 x i32*]], [3 x [7 x i32*]]* @a, i64 0, i64 1, i64 1), align 16
   ret i32* %x
 }
 
 define i32* @reduce_align_2() {
 ; CHECK-LABEL: @reduce_align_2(
-; CHECK-NEXT:    [[X:%.*]] = load i32*, i32** @a.2.2, align 8
-; CHECK-NEXT:    store i32* null, i32** getelementptr inbounds ([7 x i32*], [7 x i32*]* @a.1, i32 0, i64 2), align 8
+; CHECK-NEXT:    [[X:%.*]] = load i32*, i32** @a.6, align 16
 ; CHECK-NEXT:    ret i32* [[X]]
 ;
   %x = load i32*, i32** getelementptr inbounds ([3 x [7 x i32*]], [3 x [7 x i32*]]* @a, i64 0, i64 2, i64 2), align 16
   store i32* null, i32** getelementptr inbounds ([3 x [7 x i32*]], [3 x [7 x i32*]]* @a, i64 0, i64 1, i64 2), align 4
   ret i32* %x
 }
 
 define i32* @reduce_align_3() {
 ; CHECK-LABEL: @reduce_align_3(
-; CHECK-NEXT:    [[X:%.*]] = load i32*, i32** @a.2.3, align 8
-; CHECK-NEXT:    store i32* null, i32** getelementptr inbounds ([7 x i32*], [7 x i32*]* @a.1, i32 0, i64 3), align 4
+; CHECK-NEXT:    [[X:%.*]] = load i32*, i32** @a.7, align 16
 ; CHECK-NEXT:    ret i32* [[X]]
 ;
   %x = load i32*, i32** getelementptr inbounds ([3 x [7 x i32*]], [3 x [7 x i32*]]* @a, i64 0, i64 2, i64 3), align 4
   store i32* null, i32** getelementptr inbounds ([3 x [7 x i32*]], [3 x [7 x i32*]]* @a, i64 0, i64 1, i64 3), align 8
   ret i32* %x
 }

llvm/test/Transforms/GlobalOpt/globalsra-generic-type.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -S -globalopt < %s | FileCheck %s
 
 @g = internal global [8 x i8] undef
 
 define void @test() {
 ; CHECK-LABEL: @test(
-; CHECK-NEXT:    store i32 1, i32* bitcast ([8 x i8]* @g to i32*), align 4
-; CHECK-NEXT:    store i32 2, i32* getelementptr (i32, i32* bitcast ([8 x i8]* @g to i32*), i64 1), align 4
 ; CHECK-NEXT:    ret void
 ;
   store i32 1, i32* bitcast ([8 x i8]* @g to i32*)
   store i32 2, i32* getelementptr (i32, i32* bitcast ([8 x i8]* @g to i32*), i64 1)
   ret void
 }
 
 define i32 @load1() {
 ; CHECK-LABEL: @load1(
-; CHECK-NEXT:    [[V:%.*]] = load i32, i32* bitcast ([8 x i8]* @g to i32*), align 4
-; CHECK-NEXT:    ret i32 [[V]]
+; CHECK-NEXT:    ret i32 1
 ;
   %v = load i32, i32* bitcast ([8 x i8]* @g to i32*)
   ret i32 %v
 }
 
 define i32 @load2() {
 ; CHECK-LABEL: @load2(
-; CHECK-NEXT:    [[V:%.*]] = load i32, i32* getelementptr (i32, i32* bitcast ([8 x i8]* @g to i32*), i64 1), align 4
-; CHECK-NEXT:    ret i32 [[V]]
+; CHECK-NEXT:    ret i32 2
 ;
   %v = load i32, i32* getelementptr (i32, i32* bitcast ([8 x i8]* @g to i32*), i64 1)
   ret i32 %v
 }

llvm/test/Transforms/GlobalOpt/globalsra-opaque-ptr.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals
+; RUN: opt -S -globalopt -opaque-pointers < %s | FileCheck %s
+
+; Global SRA should not be performed here (or at least not naively), as
+; offset 4 is accessed as both i32 and i64.
+
+%T = type { i32, i32, i32, i32 }
+@g = internal global %T zeroinitializer
+
+;.
+; CHECK: @[[G:[a-zA-Z0-9_$"\\.-]+]] = internal unnamed_addr global [[T:%.*]] zeroinitializer
+;.
+define void @test1() {
+; CHECK-LABEL: @test1(
+; CHECK-NEXT:    store i32 1, ptr getelementptr inbounds ([[T:%.*]], ptr @g, i64 0, i32 1), align 4
+; CHECK-NEXT:    store i32 2, ptr getelementptr inbounds ([[T]], ptr @g, i64 0, i32 2), align 4
+; CHECK-NEXT:    ret void
+;
+  store i32 1, ptr getelementptr (%T, ptr @g, i64 0, i32 1)
+  store i32 2, ptr getelementptr (%T, ptr @g, i64 0, i32 2)
+  ret void
+}
+
+define i32 @load1() {
+; CHECK-LABEL: @load1(
+; CHECK-NEXT:    [[V:%.*]] = load i32, ptr getelementptr inbounds ([[T:%.*]], ptr @g, i64 0, i32 1), align 4
+; CHECK-NEXT:    ret i32 [[V]]
+;
+  %v = load i32, ptr getelementptr (%T, ptr @g, i64 0, i32 1)
+  ret i32 %v
+}
+
+define i64 @load2() {
+; CHECK-LABEL: @load2(
+; CHECK-NEXT:    [[V:%.*]] = load i64, ptr getelementptr inbounds ([[T:%.*]], ptr @g, i64 0, i32 2), align 4
+; CHECK-NEXT:    ret i64 [[V]]
+;
+  %v = load i64, ptr getelementptr (%T, ptr @g, i64 0, i32 2)
+  ret i64 %v
+}