bugfix: a new written tupitem was not visible to other concurrent GPU…

… cores, then it caused NULL-tupitem assertion.

issue reported at heterodb#658
GpuPreAgg(nogroup) write out a result tupitem by the first one.
Once the leader thread gets the role of __insertOneTupleNoGroups(),
1. it acquires the exclusive lock (kds_final->nitems == UINT_MAX)
2. it writes out an empty result tuple.
3. it updated the ROW_INDEX(kds_final) by atomic operation.
4. then, unlock the buffer (kds_final->nitems <-- 1)

However, reader side didn't use volatile read to fetch the resule
tuple-item from ROW_INDEX(kds_final), so it looks here is an empty
tuple on the row_index[0] due to code optimization.

So, we added volatile qualifier to fetch the row_index[0] not to
oversight new values written by the concurrent cores.

src/cuda_gpupreagg.cu

@@ -926,6 +926,7 @@ __insertOneTupleNoGroups(kern_context *kcxt,
 							 kexp_groupby_actions);
 	tupitem->t_len = tupsz;
 	tupitem->rowid = 0;
+	__threadfence();
 	__atomic_write_uint32(KDS_GET_ROWINDEX(kds_final),
 						  __kds_packed((char *)kds_final
 									   + kds_final->length

src/xpu_common.h

@@ -1136,7 +1136,7 @@ KDS_GET_ROWINDEX(const kern_data_store *kds)
 INLINE_FUNCTION(kern_tupitem *)
 KDS_GET_TUPITEM(kern_data_store *kds, uint32_t kds_index)
 {
-	uint32_t	offset = KDS_GET_ROWINDEX(kds)[kds_index];
+	uint32_t	offset = __volatileRead(KDS_GET_ROWINDEX(kds) + kds_index);
 
 	if (!offset)
 		return NULL;
@@ -1145,30 +1145,6 @@ KDS_GET_TUPITEM(kern_data_store *kds, uint32_t kds_index)
 							- __kds_unpack(offset));
 }
 
-/* kern_tupitem by tuple-offset */
-INLINE_FUNCTION(HeapTupleHeaderData *)
-KDS_FETCH_TUPITEM(kern_data_store *kds,
-				  uint32_t tuple_offset,
-				  ItemPointerData *p_self,
-				  uint32_t *p_len)
-{
-	kern_tupitem   *tupitem;
-
-	Assert(kds->format == KDS_FORMAT_ROW ||
-		   kds->format == KDS_FORMAT_HASH);
-	if (tuple_offset == 0)
-		return NULL;
-	Assert(tuple_offset < kds->length);
-	tupitem = (kern_tupitem *)((char *)kds
-							   + kds->length
-							   - __kds_unpack(tuple_offset));
-	if (p_self)
-		*p_self = tupitem->htup.t_ctid;
-	if (p_len)
-		*p_len = tupitem->t_len;
-	return &tupitem->htup;
-}
-
 INLINE_FUNCTION(uint32_t *)
 KDS_GET_HASHSLOT_BASE(const kern_data_store *kds)
 {