Re: [PING][PATCHv3 1/2] aarch64: Hoist ZVA check out of the memset function

[Szabolcs Nagy <szabolcs dot nagy at arm dot com>]

On 16/10/17 13:17, Siddhesh Poyarekar wrote:
> On Thursday 12 October 2017 08:11 PM, Siddhesh Poyarekar wrote:
>> Like I said before, the benchmark has been misunderstood and I maintain
>> that the gains will be seen in every benchmark for the input range I
>> pointed out without causing any noticeable change elsewhere.
>>
>> So barring those things I read your response as you thinking that the
>> patch is OK but you're afraid of some unknown unknowns hitting back.  In
>> that case I suggest we commit this now, let CI loops over take over (in
>> Cavium, Linaro, ARM, etc.) and then point out regressions if any.  The
>> valid patch lying around on the mailing list or an arbitrary branch is
>> never going to get tested.
> 
> Szabolcs, have you thought about this some more?
> 

i think the patch can be reduced to have
3 variants instead of 4 (the nozva case does
not seem to buy much over the generic code)

and the alignments have to be checked for the
new fixed zva size cases because they seem to
be wrong.

it's not entirely clear what should happen
for libc internal memset calls, that needs
more thought.

quick test shows that the current benchmark was
not running enough iterations and not testing
enough cases, i'd expect the actual speed up
to be lower than previously reported.

overall this needs more work, as i said string
functions needs to be modified very carefully
since many things depend on their behaviour.