Hi,

On Fri, 18 Sept 2026 at 06:22, Richard Guo <[email protected]> wrote:
>
> Further fuzzing with Claude on the join alias found this bug.  The
> following queries fail in various ways on master and all supported
> branches:
>
> create table t (a int, b int);
>
> select 1 from ((select (select 1) as x) s cross join t) j
>   where (select 1 where j is null) is null;
> ERROR:  cannot handle unplanned sub-select
>
> select 1 from ((select (select 1) as x) s cross join t) j
>   where (1, 1) in (select (j is null)::int, count(*) from t);
> TRAP: failed Assert("!IsA(node, SubLink)"), File: "prepagg.c"
>
> select 1 from ((select (select 1) as x) s cross join t) j
>   where exists (select 1 from t tablesample system ((j is null)::int * 100));
> ERROR:  unrecognized node type: 22
>
> Once subquery s is flattened, the joinaliasvars entry for j.x is no
> longer a Var but the SubLink (select 1), so expanding a reference to j
> inside a sub-select inserts a SubLink into that sub-select.
>
> But flatten_join_alias_vars_mutator fails to notice this and thus does
> not set the sub-select's hasSubLinks flag.  So preprocess_expression
> skips SS_process_sublinks, and the SubLink survives into code that
> can't cope with one.
>
> The fix is to make the same checkExprHasSubLink() test in the
> whole-row path.  See attached.

This looks reasonable to me. I can see how the recursive call might
look sufficient here, but once the alias entry already contains a
SubLink, there needn't be another join Var to expand and trigger the
check. Handling it like the single-column case seems a good fit.

Patch LGTM.

Regards,
Ayush


Reply via email to