[PATCH v2] x86-64: Optimize bzero

[Noah Goldstein]

On Thu, Feb 10, 2022 at 7:02 AM Wilco Dijkstra <Wilco.Dijkstra@arm.com> wrote:
>
> Hi,
>
> >> The saving is in the lane-cross broadcast which is on the critical
> >> path for memsets in [VEC_SIZE, 2 * VEC_SIZE] (think 32-64).
>
> What is the speedup in eg. bench-memset? Generally the OoO engine will
> be able to hide a small increase in latency, so I'd be surprised it shows up
> as a significant gain.

Well comparing the previous sse2 bzero against avx2/evex/avx512 versions
there is obviously speedup. Comparing memset-${version} vs bzero-${version}
it's ambiguous if there is any benefits.

>
> If you can show a good speedup in an important application (or benchmark
> like SPEC2017) then it may be worth pursuing. However there are other
> optimization opportunities that may be easier or give a larger benefit.
>
> >> Agreed it's not clear if it's worth it to start replacing memset calls with
> >> bzero calls, but at the very least this will improve existing code that
> >> uses bzero.
>
> No code uses bzero, no compiler emits bzero. It died 2 decades ago...
>
> > My point is this is a lot of code and infrastructure for a symbol marked
> > as legacy for POSIX.1-2001 and removed on POSIX.1-2008 for the sake of
> > marginal gains in specific cases.
>
> Indeed, what we really should discuss is how to remove the last traces of
> bcopy and bcmp from GLIBC. Do we need to keep a compatibility symbol
> or could we just get rid of it altogether?
>
> Cheers,
> Wilco