Re: [PATCH 3/4] S390: Do not call memcpy, memcmp, memset within libc.so via ifunc-plt.

[Adhemerval Zanella <adhemerval dot zanella at linaro dot org>]

On 11/05/2016 10:40, Stefan Liebler wrote:
> On 05/04/2016 05:22 PM, Adhemerval Zanella wrote:
>>
>>
>> On 27/04/2016 05:14, Stefan Liebler wrote:
>>> On 04/26/2016 03:35 PM, Adhemerval Zanella wrote:
>>>>
>>>>
>>>> On 26/04/2016 09:07, Stefan Liebler wrote:
>>>>> On s390, the memcpy, memcmp, memset functions are IFUNC symbols,
>>>>> which are created with s390_libc_ifunc-macro.
>>>>> This macro creates a __GI_ symbol which is set to the
>>>>> ifunced symbol. Thus calls within libc.so to e.g. memcpy
>>>>> result in a call to *ABS*+0x954c0@plt stub and afterwards
>>>>> to the resolved memcpy-ifunc-variant.
>>>>>
>>>>> This patch sets the __GI_ symbol to the default-ifunc-variant
>>>>> to avoid the plt call. The __GI_ symbols are now created at the
>>>>> default variant of ifunced function.
>>>>
>>>> Is the internal ifunc plt usage leading to a failure in s390/s390x
>>>> (as for powerpc32 and i686) or is it an optimization fix?
>>>
>>> No it does not lead to a failure on s390.
>>> It is an optimization to avoid the extra call to the plt-stub if called within libc.so.
>>>
>>
>> Right because on 19c4bec0f43599eecc2f32de96ae179cd7d64053 I did the
>> exact opposite because for POWER it shows that the gains of using
>> optimized version over default one shows a good improvement in
>> algorithms that use these symbol internally (like regex).
>>
>> It might be the case where the default s390 version is good enough
>> and shows no performance difference.
>>
> 
> Hi Adhemerval,
> 
> I've retested the patchset in context of regex with these functions called with/without ifunc-plt.
> The version without ifunc-plt is slightly faster.
> 
> But you are right. This decision has to be redetermined with further variants or newer machines.
> 
> In case of the string-functions, calling them via ifunc within libc.so could gain improvement.
> 
> Bye
> Stefan
> 

Fair enough them.  I would expect without ifunc-plt to be slower in newer
architectures, but it could be case where the performance differences
between the base implementation and the arch optimized one is that large
(for instance powerpc64 memcpy the base one only uses 8-bytes store/loads,
which show large differences depending of the workload compared to power7
one).