Thread: coredump of 9.3.2

Hi!

Rather simple script (original query was a 500 lines of SQL with >100 Mb of 
gzipped dump. Query is looked strange, but actually it was auto-generated):

CREATE TABLE t (a int, b int);
CREATE TABLE tt (c int);

INSERT INTO t VALUES (1,1), (2,2);
INSERT INTO tt VALUES (2);


SELECT    *
FROM    t
WHERE (    CASE        WHEN a%2 IN (SELECT c FROM tt) THEN a    END IN (SELECT c FROM tt)
);

I suppose, the problem is connected to hashed subplan, but I'm not very familiar 
with executor. And this affects all supported versions of pgsql.


-- 
Teodor Sigaev                                   E-mail: teodor@sigaev.ru
  WWW: http://www.sigaev.ru/
 

On 12/10/2013 09:39 AM, Teodor Sigaev wrote:
> 
> SELECT
>     *
> FROM
>     t
> WHERE (
>     CASE
>         WHEN a%2 IN (SELECT c FROM tt) THEN a
>     END IN (SELECT c FROM tt)
> );

Wow, it wouldn't have occured to me that that was even supported syntax.I'm not suprised that it doesn't work ...


-- 
Josh Berkus
PostgreSQL Experts Inc.
http://pgexperts.com

Teodor Sigaev <teodor@sigaev.ru> writes:
> SELECT
>      *
> FROM
>      t
> WHERE (
>      CASE
>          WHEN a%2 IN (SELECT c FROM tt) THEN a
>      END IN (SELECT c FROM tt)
> );

> I suppose, the problem is connected to hashed subplan, but I'm not very familiar 
> with executor. And this affects all supported versions of pgsql.

It seems to be a planner bug: it's doing the wrong thing with the
PARAM_SUBLINK Params for the nested IN SubLinks.  (They don't look
to be nested textually, but they are, and convert_testexpr is
mistakenly replacing the inner one's Params when it should only
be replacing the outer one's Params.)

I think this has probably been broken since about 2005 :-(.
The comment on convert_testexpr claims it doesn't need to worry
about nested cases; which is true when it's called during
SS_process_sublinks, but not so much when it's called from
convert_ANY_sublink_to_join.  Some digging in the git history
suggests that the latter call existed at the time, meaning the
comment was wrong even when written.
        regards, tom lane