Re: define pg_structiszero(addr, s, r)

Hi,

On Thu, Nov 07, 2024 at 09:44:32AM +0900, Michael Paquier wrote:
> On Thu, Nov 07, 2024 at 08:10:17AM +1300, David Rowley wrote:
> > Did you try with a size where there's a decent remainder, say 124
> > bytes? FWIW, one of the cases has 112 bytes, and I think that is
> > aligned memory meaning we'll do the first 64 in the SIMD loop and have
> > to do 48 bytes in the byte-at-a-time loop. If you had the loop Michael
> > mentioned, that would instead be 6 loops of size_t-at-a-time.
> 
> See the attached allzeros.c, based on the previous versions exchanged.
> And now just imagine a structure like that:
> #define BLCKSZ 48
> typedef union AlignedBlock
> {
>     char        data[BLCKSZ];
>     double      force_align_d;
>     int64_t     force_align_i64;
> } AlignedBlock;
> 
> The allzero check is used for pgstat entries, and it could be possible
> that some out-of-core code needs to rely on such small-ish sizes, or
> even something else when a patch author feels like it.  So let's make
> that optimized as much as we think we can: that's what this discussion
> is about.

Yeah, fully agree. My initial testing was not "good" enough and so was not showing
as much improvement as your's and David's ones.

Please find v8 attached.

Regards,

-- 
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com