> Yeah, the patch in its current form is wrong, because it will give wrong
> answers if the operator being used in a SAOP is non-strict. I modified
> the patch to consider operator strictness before doing anything with nulls.
I tried to review this patch without any familiarity to the code.
arrayconst_next_fn():
> + /* skip nulls if ok to do so */
> + if (state->opisstrict)
> + {
> + while (state->elem_nulls[state->next_elem])
> + state->next_elem++;
> + }
Shouldn't we check if we consumed all elements (state->next_elem >=
state->num_elems) inside the while loop?
arrayexpr_next_fn():
> + /* skip nulls if ok to do so */
> + if (state->opisstrict)
> + {
> + Node *node = (Node *) lfirst(state->next);
> +
> + while (IsA(node, Const) && ((Const *) node)->constisnull)
> + state->next = lnext(state->next);
> + }
I cannot find a way to test this change. Can you imagine a query to
exercise it on the regression tests?