[FFmpeg-cvslog] r16488 - in trunk: configure libavutil/mem.c

[ramiro]

  wrote:
> Michael Niedermayer wrote:
>> On Fri, Jan 09, 2009 at 03:35:52AM -0200, ramiro wrote:
>>> Michael Niedermayer wrote:
>>>> On Thu, Jan 08, 2009 at 07:40:18PM -0200, ramiro wrote:
>>>>> Michael Niedermayer wrote:
>>>>>> On Thu, Jan 08, 2009 at 09:34:06PM +0100, Vitor Sessak wrote:
>>>>>>> ramiro wrote:
>>>>>>>> Author: ramiro
>>>>>>>> Date: Thu Jan  8 00:36:34 2009
>>>>>>>> New Revision: 16488
>>>>>>>>
>>>>>>>> Log:
>>>>>>>> Use posix_memalign() if available.
>>>>>>> By looking at FATE, it looks like this change broke H264 decoding on
>>> gcc
>>>>>>> 2.95.3. See for example
>>>>>>> http://fate.multimedia.cx/index.php?test_result=6753660 .

And now it randomagically works.

>>>>>> as far as i understand it fate didnt build 3 revissions and this one
>>>>>> does not look like the most likely cause, 16490 looks more likely
>>>>> I couldn't get a hold of gcc 2.95 to test, but this patch should fix it.
>>>> i seriously doubt this patch will fix it. The problem IIRC was that
>>>> gcc 2.95 had some problem with alignment somewhere. gcc 2.95 should not
>>>> have any problem with sse as such.
>>> I finally got gcc 2.95 working. This patch really changes nothing.
>>>
>>> So, what could be done here?
>>> - check gcc version in configure and have sse depend on it
>>> or
>>> - #if defined(HAVE_SSE) && (!defined(__GNUC__) || __GNUC__ > 2)
>>> or
>>> - disable HAVE_SSE in some internal header based on __GNUC__
>>>
>>> I'm slightly more in favor of configure, because then we can keep
>>> HAVE_SSE as the only basis to disable building SSE code in the future.
>>> This patch only disabled selecting it at run-time. The functions are
>>> still being built.
>> fine though you need approval from the configure maintainers for this
> 
> Are there any binutils versions still in use that don't support sse?
> 
>> also strictly the correct solution would be to leave HAVE_SSE enabled
>> as gcc 2.95 supports it and just disable functions where actual alignment
>> problems exists.
> 
> Disabling working code out of laziness is unacceptable IMO.

I don't understand what you suggest here.

What about leaving things as is and adding a section in the 
documentation about gcc 2 support suggesting to --disable-sse?

Ramiro Polla