Re: refactor architecture-specific popcount code

On Sat, Jan 31, 2026 at 4:33 AM Nathan Bossart <nathandbossart(at)gmail(dot)com> wrote:
>
> On Fri, Jan 30, 2026 at 03:22:45PM +0700, John Naylor wrote:
> > 0001 - I'm pretty sure this is comparable to HEAD if the optimized
> > function is pg_popcount_sse42(). Has the AVX512 version been tested
> > with 8-byte inputs? That seems to have a lot of pre- and
> > post-processing involved. The inline wrapper only bypasses for 7 or
> > less bytes.
>
> Here [0] is the latest perf data I see for the AVX-512 popcount patch,
> although that's comparing to v16, which IIRC lacks a few other inlining
> tricks. There's a chance the SSE4.2 version is faster at that particular
> length. I'm not sure we need to worry about that, but I can do a bit of
> testing if you'd like.

It might be a good idea to do a little new testing, and I see a use
for a special 8-byte path independent of AVX512: v6 seems to regress a
little for single-words. But, it turns out that when gcc turns
__builtin_popcountl into a single instruction, it's inline, but if it
emits portable bitwise ops, it does so in a function called
__popcountdi2(). That can be avoided by hand-coding in C for normal
builds (and for 32-bit looks cleaner anyway), as in the attached 0005.

My laptop here is really too old to make decisions that are
micro-architecture dependent, but with that caveat, I dusted off the
popcount benchmark and added a test for counting bitmapsets (v7-0004,
applies on top of v6):

select drive_bms_num_members(10000000, 1);

master: 13.2 ticks per call
v6: 15.3
v6+v7-0005 10.8

Again, take this with a grain of salt, but 0005 seems worth looking at.

--
John Naylor
Amazon Web Services