Re: Ping: [Patch] aarch64: Thunderx specific memcpy and memmove

[Siddhesh Poyarekar <siddhesh@gotplt.org>]

On Friday 26 May 2017 02:34 AM, Ramana Radhakrishnan wrote:
> SVE in the ARM world is architectural and not micro-architectural in the context
> of this discussion :) .

Yes, I did not think of SVE as a micro-architecture detail.  I used SVE
as an example to show that multiarch =/=> micro-architectures.

> The difference in the ARM world compared to the x86 world is the number
> of micro-architectures that target the same architectural baseline.
> Pushing in a memcpy
> for every single micro-architecture out there will make the library a
> maintenance
> nightmare !

Nor am I arguing that micro-architectures ==> multiarch.  My comments
have only been pointing out that the IFUNC cost for multiarch is here to
stay and there will likely never be consensus to drop it given the
innovations happening in the ARM server space.

> And we also need to see some numbers which compare the relative performance
> of the routines being put in compared to the generic memcpy otherwise things
> will not improve.  Atleast something like this routine is X % better than the
> generic memcpy.

Yes, and I understand Steve had pointed out the benefits of his
implementation in the original post.  If it turns out that the
implementation is optimal for the general case, by all means merge it
with the generic one but that's not a reason to drop multiarch.

Siddhesh