[FFmpeg-devel] [PATCH 0/2] Fix strict aliasing violations in AAC and AC3 parsers

[Måns Rullgård]

M?ns Rullg?rd <mans at mansr.com> writes:

> "Jason Garrett-Glaser" <darkshikari at gmail.com> writes:
>
>>> am i interperting this correctly, that the code is hugely faster with
>>> -fno-strict-aliasing
>>> ?
>>> and iam of course in favor of the faster variant for h264.c
>>
>> Yes, but it makes absolutely no sense why it would be faster--the
>> oprofile says even the asm functions are faster which of course makes
>> no sense at all.
>>
>> Others in #ffmpeg-devel have reported small speed increases or
>> decreases, but nothing like what I posted.
>
> Changes in decoding time after adding -fno-strict-aliasing on ARM:
>
> H.264   -2.4%
> MPEG4   +0.14%
> MPEG2   -2.3%
> Vorbis  +0.08%
> Flac    +3.6%
> MP3     +2.6%
> AC3     +0.83%
> DTS     +0.81%

BTW, that was with codesourcery gcc 4.2.1.

> I will test with Jason's sample on Core2 and see if I get anything
> resembling his results.

Tested on Core2, and I get no such numbers:

vanilla              203.7867 s
-fno-strict-aliasing 204.2257 s

This was with gcc 4.2.4.

-- 
M?ns Rullg?rd
mans at mansr.com