Tomas Vondra <tomas.von...@enterprisedb.com> writes:
> IMO just bumping up the constants from ~65k to 1M is a net loss, for
> most users. We add this to bitmapsets, which means we're using ~8kB with
> the current values, but this jumps to 128kB with this higher value. This
> also means bms_next_member etc. have to walk much more memory, which is
> bound to have some performance impact for everyone.
Hmm, do we really have any places that include OUTER_VAR etc in
bitmapsets? They shouldn't appear in relid sets, for sure.
I agree though that if they did, this would have bad performance
consequences.
I still think the negative-special-values approach is better.
If there are any places that that would break, we'd find out about
it in short order, rather than having a silent performance lossage.
regards, tom lane
