Glad I could be of help.
I found this through code analysis. I've been working on a custom PG
checker, adapting the Clang Static Checker for my bachelor thesis.
Always nice to see, when academic work has real world benefits :)
Bernd
On 8/29/25 3:02 PM, David Rowley wrote:
> On Fri, 29 Aug 2025 at 23:45, Bernd Reiß <bd_reiss@gmx.at> wrote:
>> Thanks for the quick response and the review.
> Thanks for the report, investigation and patch.
>
> I've pushed and backpatched this to 15. v14 doesn't have the
> RelOptInfo.live_parts field, so it didn't suffer from the issue.
> Technically, 15 isn't broken either as the bms_del_member() function
> in that version wouldn't pfree the set. I decided to patch 15 anyway
> to keep the code the same and to avoid assuming it's ok to ignore the
> return value of bms_del_member().
>
>> This is admittedly a pretty remote edge case, but still, better safe
>> than sorry.
> Did you find it through code analysis or from a crash?
>
> It would just have been a matter of time before someone hit this.
>
> David
>
>
