[PATCH] malloc: Optimize small memory zeroing for calloc

[H.J. Lu]

On Wed, Nov 27, 2024, 7:25 PM Wilco Dijkstra <Wilco.Dijkstra@arm.com> wrote:

> Hi H.J.,
>
> If this code is performance critical, it would be useful to optimize the
> generic version first and only consider target-specific code when you
> can beat the generic version by a large margin. For example using
> fixed-size inlined memsets should be faster on most targets and likely
>

I couldn't find a better way which works for all targets.

emit similar code on x86.
>

Very unlikely since the size isn't fixed and x86-64
will use unaligned vector store which isn't suitable
for all targets.


> I think it is fine to extract this into a seperate function, however it
> should
> include the memset so one can change the size of when to call memset -
> it is not obvious to me why the magic 9 is best.
>

Good point.  Will add a macro to specify the unroll size.

Thanks.


> Cheers,
> Wilco
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://sourceware.org/pipermail/libc-alpha/attachments/20241128/e32d6a55/attachment.htm>