On Thu, Aug 27, 2026 at 10:45 AM David Rowley <[email protected]> wrote:
>
> On Thu, 27 Aug 2026 at 14:27, Ewan Young <[email protected]> wrote:
> > Rather than pile that cleanup onto this bug-fix thread, I'll send it as a
> > separate follow-up so this fix can be considered on its own.  Will post
> > that shortly and CC you three.
>
> Yes, another thread. Please also include a description of your method
> for verifying the code matches for non-HAVE__BUILTIN_OP_OVERFLOW
> builds.
>
> Also, pg_neg_s32_overflow() was introduced in 2024 by 0a27c3d0f733,
> which is v18+. Depending on Daniel's patience for backpatching two
> different versions, it might be worth just using the == PG_INT32_MIN
> method to fix the bug, then allowing the follow-up patch to switch
> that in v20 only.  That would mean v18 and v19 don't get the most
> optimal fix, but it's hard to imagine that's going to be performance
> critical.

Thanks, that plan makes sense to me -- let's keep the bug fix uniform
across all branches and do the pg_neg_s32_overflow() switch separately.

pg_neg_s32_overflow() only exists from v18 (0a27c3d0f733), while the bug
goes all the way back, so using it in the fix would mean carrying two
versions for the back-branches.  Not worth it for a one-liner.

Attached v4 goes back to the explicit PG_INT32_MIN test:

    if (unlikely(n == PG_INT32_MIN))
        n = PG_INT32_MAX;
    else
        n = -n;

which applies cleanly to every supported branch (I checked REL_14 through
REL_19 and master).  The fix and the regression case are otherwise
unchanged.

I'll take the pg_neg_s32_overflow() cleanup to its own thread, master
only, covering the other open-coded negation checks Ilmari pointed at
(int2/4/8 um/abs/div, lcm, cash, numericvar_to_int64); once this fix
lands I'll fold text_right() into that too.  So the back-branches keep the
plain form and only master gets the helper.

>
> David



-- 
Regards,
Ewan Young

Attachment: v4-0001-Fix-right-with-the-most-negative-integer.patch
Description: Binary data

Reply via email to