Re: Performance optimization of btree binary search

Peter Geoghegan <pg@heroku.com> writes:
> I guess I could write a proper patch to have code setting up a scankey
> also set a flag that indicated that it was acceptable to assume that
> the special built-in comparator would do fine. ...
> I'd be happy with a scheme with only one built-in comparator, and
> allowed a few types to be cataloged such that it was indicated that
> just using the "built-in" comparator was acceptable, knowledge that
> could be passed to _bt_compare via the scankey. I'm thinking of just
> int4, and maybe date and a few other such int4 "covariant" types.

If what you're proposing is that we have a fast path that compares Datums
as Datums, I should think that that would work fine for int2 as well,
*and* for int8 on machines where int8 is pass-by-value.  (Does anyone
still care much about PG's performance on 32-bit hardware?)  We might
have to fool a bit with the fooGetDatum macros in some cases, eg
I think Int16GetDatum isn't careful about sign extension.  Admittedly,
that might introduce an offsetting cost on some hardware, but I think
on most machines sign-extension isn't noticeably more expensive than
zero-extension.
        regards, tom lane