Re: BUG #18778: Query planning fails in ExecInitExprRec with unrecognized node type

PG Bug reporting form <noreply(at)postgresql(dot)org> writes:
> The following script:
> CREATE TABLE t (id int, PRIMARY KEY (id)) PARTITION BY RANGE (id);
> CREATE TABLE t0 PARTITION OF t FOR VALUES FROM (0) TO (1);
> CREATE TABLE t1 PARTITION OF t FOR VALUES FROM (1) TO (2);

> SELECT 1 FROM (SELECT EXISTS (SELECT 1 FROM t0 WHERE id = t00.id) AS b FROM
> t0 t00) r, t
> WHERE t.id > CASE WHEN jsonb_build_object(b) IS NULL THEN 1 ELSE 1 END;

> fails with:
> ERROR: XX000: unrecognized node type: 24

Thanks for the report! setrefs.c is supposed to remove
AlternativeSubPlan nodes from the plan, but it's failing to do so
here. Digging, the un-cleaned-up AlternativeSubPlan is inside the
exec_pruning_steps of an Append node's part_prune_info, and I see
that setrefs is totally unaware that those expressions might need
processing. I think it ought to be applying fix_scan_expr to them,
as per the attached. There are a bunch of tidying-up things that
fix_scan_expr does, so I suspect that there may be more bug symptoms
reachable from this oversight. Some of the missed processing may be
redundant --- for example it's likely that
record_plan_function_dependency is duplicative because functions used
here would also be used elsewhere in the query. But it's hard to
believe it all is.

regards, tom lane