[DAGCombiner] Make merging consecutive stores aggressive

[junbuml]

Summary

Instead of looking at mergeable stores only from the end of store sequences,
we can start from a store which is not in the middle of store sequence of the same type.
This should fix PR27816 : https://llvm.org/bugs/show_bug.cgi?id=27816

[niravd]

This change is obviated by D14834 which is waiting on a final okay from the AMDGPU folks.

[junbuml]

This change is obviated by D14834 which is waiting on a final okay from the AMDGPU folks.

Yes, looks like D14834 cover what I want to do here.
Thanks Nirav !

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
   // Don't merge vectors into wider vectors if the source data comes from loads.
   // TODO: This restriction can be lifted by using logic similar to the
   // ExtractVecSrc case.
   if (MemVT.isVector() && IsLoadSrc)
     return false;
 
   // Only look at ends of store sequences.
   SDValue Chain = SDValue(St, 0);
-  if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
+  if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE) {
+    StoreSDNode *StUser = cast<StoreSDNode>(*Chain->use_begin());
+    if (St->getMemoryVT() == StUser->getMemoryVT())
       return false;
+  }
 
   // Save the LoadSDNodes that we find in the chain.
   // We need to make sure that these nodes do not interfere with
   // any of the store nodes.
   SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
 
   // Save the StoreSDNodes that we find in the chain.
   SmallVector<MemOpLink, 8> StoreNodes;

test/CodeGen/AArch64/merge-store.ll

while.end.i:
   %call.i = tail call i8* @foo()
   store i8* %call.i, i8** bitcast (%struct1*** @gv1 to i8**), align 8
   br label %exit
 
 exit:
   ret void
 }
 
+; PR27816 - Continue merging consecutive stores when the next store has different type in store sequences
+
+; CHECK-LABEL: merge_const_store
+; CHECK: str x{{[0-9]+}}, {{.*}}
+; CHECK: str wzr
+
+%struct.A = type { i8, i8, i8, i8, i8, i8, i8, i8, i32}
+
+define void @merge_const_store(i32 %count, %struct.A* nocapture %p, i8 %val8)  {
+  %1 = icmp sgt i32 %count, 0
+  br i1 %1, label %.lr.ph, label %._crit_edge
+.lr.ph:
+  %i.02 = phi i32 [ %add, %.lr.ph ], [ 0, %0 ]
+  %.01 = phi %struct.A* [ %addr, %.lr.ph ], [ %p, %0 ]
+  %a0 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 0
+  store i8 1, i8* %a0, align 1
+  %a1 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 1
+  store i8 2, i8* %a1, align 1
+  %a2 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 2
+  store i8 3, i8* %a2, align 1
+  %a3 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 3
+  store i8 4, i8* %a3, align 1
+  %a4 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 4
+  store i8 5, i8* %a4, align 1
+  %a5 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 5
+  store i8 6, i8* %a5, align 1
+  %a6 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 6
+  store i8 7, i8* %a6, align 1
+  %a7 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 7
+  store i8 8, i8* %a7, align 1
+  %addr_last = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 8
+  store i32 0, i32* %addr_last, align 4
+  %add = add nsw i32 %i.02, 1
+  %addr = getelementptr inbounds %struct.A, %struct.A* %.01, i64 1
+  %exitcond = icmp eq i32 %add, %count
+  br i1 %exitcond, label %._crit_edge, label %.lr.ph
+._crit_edge:
+  ret void
+}
+
+
 declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1)
 declare i32 @fcntl(i32, i32, ...)
 declare noalias i8* @foo()