Re: [PATCH 1/3] aarch64: Optimized memset specific to AmpereComputing emag

[Siddhesh Poyarekar <siddhesh at gotplt dot org>]

On 18/12/18 3:33 PM, Feng Xue wrote:
> This version uses general register based memory store instead of
> vector register based, for the former is faster than the latter
> in emag.

Barring a couple of instances that show a 5%+ difference from 
__memset_generic (which also seems sporadic, maybe due to noise?), 
everything else seems to be in the 1-2% range.  Is that a significant 
enough difference to warrant a new variant?

It may not be worth adding another variant for a mere 1-2% overall gain 
for string routines but maybe I've misread the results and you have a 
better justification for this.  Please let me know if you do.

This is not a strong opposition though; if Szabolcs, Wilco, etc. don't 
mind maintaining another variant (by calling it zva64_nosimd) despite 
the seemingly low relative performance gain then I won't stand in the way.

Siddhesh