Re: [PATCH v2] x86-64: Optimize bzero

[Noah Goldstein <goldstein.w.n@gmail.com>]

On Fri, Feb 25, 2022 at 7:51 AM Wilco Dijkstra <Wilco.Dijkstra@arm.com> wrote:
>
> Hi Noah,
>
> > What method are you using for getting total function call overhead?
> >
> > Using `perf record` + `pref report` and see a fair amount of variance but much
> > higher memset overheader (counting `_*unaligned_erms` and `*_unaligned`)
> > in `cc1` and `as`.
> >
> > From average of 3 runs compiling file with 1/10/100/1000 functions I get:
>
> For accurate profiles you need to run for longer and increase the sampling rate.

Ran with `-F max`. Would say error bar is ~.5% still (which is [12, 25]% of the
overall overhead so not insignificant.

> The 1000 function version takes less than a second to compile, which is why I
> went for 10000 functions at 20000 samples per second.
>
> I used LD_PRELOAD to use the latest version of __memset_exex_unaligned_erms.
> You can also trivially collect statistics of memset this way. Eg. in this benchmark
> 75% of memset is <= 32 bytes and average memset size 44 bytes, so it has a far
> higher percentage of really small memsets compared to normal compilation.

Does that mean you manually inserted timers to get total function call overhead?

But I see, was just curious if there was some simple audit more
precise than `perf record`
that didn't require building custom infrastructure.

>
> Cheers,
> Wilco