Re: optimize lookups in snapshot [sub]xip arrays

On Wed, Aug 10, 2022 at 5:00 AM Nathan Bossart <nathandbossart(at)gmail(dot)com> wrote:
>
> On Tue, Aug 09, 2022 at 01:21:41PM +0700, John Naylor wrote:
> > I decided I wasn't quite comfortable changing snapshot handling
> > without further guarantees. To this end, 0002 in the attached v11 is
> > an addendum that adds assert checking (also pgindent and some
> > comment-smithing). As I suspected, make check-world passes even with
> > purposefully screwed-up coding. 0003 uses pg_lfind32 in syscache.c and
> > I verified that sticking in the wrong answer will lead to a crash in
> > assert-enabled builds in short order. I'd kind of like to throw this
> > (or something else suitable) at the build farm first for that reason.
> > It's simpler than the qsort/qunique/binary search that was there
> > before, so that's nice, but I've not tried to test performance.
>
> Your adjustments in 0002 seem reasonable to me. I think it makes sense to
> ensure there is test coverage for pg_lfind32(), but I don't know if that
> syscache code is the right choice. For non-USE_SSE2 builds, it might make
> these lookups more expensive.

I think that for non-USE_SSE2 builds, there is no additional overhead
as all assertion-related code in pg_lfind32 depends on USE_SSE2.

> I'll look around to see if there are any
> other suitable candidates.

As you proposed, having a test module for that seems to be a good
idea. We can add test codes for future optimizations that utilize SIMD
operations.

Regards,

--
Masahiko Sawada
EDB: https://www.enterprisedb.com/