Re: [PATCH v3 2/2] powerpc: Add optimized stpncpy for POWER9

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 30/09/2020 11:21, Raphael M Zinsly wrote:
> Hi Adhemerval,
> 
> On 30/09/2020 10:42, Adhemerval Zanella wrote:
>>
>>
>> On 29/09/2020 12:21, Raphael Moreira Zinsly via Libc-alpha wrote:
>>> Add stpncpy support into the POWER9 strncpy.
>>
>> The benchmark numbers you provided [1] seems to show it is slight worse than
>> the generic_strncpy which uses the same strategy as string/strncpy.c
>> (which would use VSX instruction through memset/memcpy).
> 
> My implementation is always better than the generic_strncpy, almost three times better in average. And it calls memset as well.
> 
> Are you talking about __strncpy_ppc? For some reason it is using strnlen_ppc instead of the strnlen_power8, but I didn't touch it.
> 
>> Did you compare this
>> optimization against an implementation that just call power8/9 memset/memcpy
>> instead?
>>
> 
> Not sure if I understand, isn't that generic_strncpy and strncpy_ppc?


Right, I misread the benchmark.  And I tested my own suggestion on the power9
from gcc farm and it seems that although it is slight faster than power7
variant it does not really beat power8 (as expected since it calls strnlen and
then memcpy/memset and access the input twice).

I do not really oppose it and it is up to the arch maintainer, but I still think
these micro-optimizations tends to just add extra maintainability and icache
pressure where the microbenchmark does not really catch.

> 
> 
>> It should resulting a smaller implementation which reduces i-cache size and
>> the code is much more simpler and maintainable.  The same applies for stpncpy.
>>
>> I tried to dissuade Intel developers that such micro-optimization are not
>> really a real gain and instead we should optimize only a handful of string
>> operations (memcpy/memset/etc.) and use composable implementation instead
>> (as generic strncpy).  It still resulted on 1a153e47fcc, but I think we
>> might do better for powerpc.
>>
>> [1] https://sourceware.org/pipermail/libc-alpha/2020-September/118049.html
>>
> 
> Best Regards,