[PATCH v2] x86-64: Optimize bzero

[Adhemerval Zanella]

On 09/02/2022 19:14, Noah Goldstein wrote:
> On Wed, Feb 9, 2022 at 5:41 AM Adhemerval Zanella via Libc-alpha
> <libc-alpha@sourceware.org> wrote:
>>
>>
>>
>> On 08/02/2022 19:43, H.J. Lu via Libc-alpha wrote:
>>> Rebase against the current master branch.
>>>
>>> --
>>> memset with zero as the value to set is by far the majority value (99%+
>>> for Python3 and GCC).
>>>
>>> bzero can be slightly more optimized for this case by using a zero-idiom
>>> xor for broadcasting the set value to a register (vector or GPR).
>>>
>>> Co-developed-by: Noah Goldstein <goldstein.w.n@gmail.com>
>>
>> Is it really worth to ressurerect bzero with this multiple ifunc variants?
>> Would Python3/GCC or any programs start to replace memset with bzero for
>> the sake of this optimization?
>>
>> I agree with Wilco where the gain are marginal in this case.
> 
> The cost is only 1 cache line and it doesn't interfere with memset at all
> so it's unlikely to cause any problems.
> 
> 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).
> 
> As well for EVEX + AVX512, because it uses predicate execution for
> [0, VEC_SIZE] there is a slight benefit there (although only in throughput
> because the critical path in mask construction is longer than the lane
> VEC setup).
> 
> 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.
> 

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.