SIMD vs no SIMD on i686 (was: [Bug 3975] executable stack flag prevents use of libx264.so.148 with selinux in enforcing mode)

[Nicolas Chauvet]

2016-07-19 10:31 GMT+02:00 Simon Farnsworth <simon at farnz.org.uk>:
>
>> On 19 Jul 2016, at 02:12, Sérgio Basto <sergio at serjux.com> wrote:
>>
>> On Ter, 2016-07-19 at 02:55 +0200, dominik at greysector.net wrote:
>>> Hello, Sérgio.
>>>
>>> On Tuesday, 19 July 2016 at 02:35, RPM Fusion Bugzilla wrote:
>>>>
>>>> https://bugzilla.rpmfusion.org/show_bug.cgi?id=3975
>>>>
>>>> --- Comment #4 from Sérgio Basto <sergio at serjux.com> 2016-07-19
>>>> 02:35:19 CEST ---
>>>> (In reply to comment #3)
>>>>>
>>>>> I'll take a look at this. Could you try running execstack -c on
>>>>> the installed
>>>>> library in the meantime?
>>>> Be my guest , in  meantime I read your thread on packaging Mailing
>>>> list about
>>>> sse3 , I'd like understand if we need 2 builds for i686 ... one
>>>> with sse2 other
>>>> without it, can you give us your opinion ?
>>> The only real concern here are applications linked against libx264,
>>> which someone might want to run on low-end hardware, because I don't
>>> think anyone would want to encode anything to H.264 on non-SSE2
>>> capable
>>> CPU (i.e. Pentium 3 or Athlon XP and older). Considering last non-
>>> SSE2
>>> CPUs went out of production about 8 years ago, I think it's fairly
>>> safe
>>> to assume that the impact of doing SSE2-only builds would be
>>> negligible,
>>> if any.
>>
>> From x264.spec we can read "#i686 is disabled on purpose - re-enabled
>> with sse2 build"
>>
>> It is safe to say enable sse2 build is enable asm code ? if yes %ifarch
>> i686 we can/should build with enable asm code ?  and only one time
>> isn't it ? I tried that, but kwizart didn't agree ...
>>
>
> It may not be worth the effort (as there's upstream effort here, not just packaging, if I understand correctly), but https://lwn.net/Articles/691932/ documents GCC 6 function multi-versioning, which enables you to build both SSE2 and non-SSE versions of a given functions, and have GCC generate code to choose the right version at run time - as it's part of the compiler, the optimiser should also be aware, and able to avoid dispatch code when calling from the SSE2 version to another SSE2 function.

Indeed, that's clearly the way to go for upstream, even if before that
there is hwcaps


-- 
-

Nicolas (kwizart)