[FFmpeg-devel] [PATCH/RFC] mathops/x86: work around inline asm miscompilation with GCC 4.8.1

[Hendrik Leppkes]

On Sun, Jun 16, 2013 at 8:31 AM, James Almer <jamrial at gmail.com> wrote:
> On 15/06/13 5:48 PM, Hendrik Leppkes wrote:
>> On Sat, Jun 15, 2013 at 10:46 PM, Hendrik Leppkes <h.leppkes at gmail.com> wrote:
>>> The volatile is not required here, and prevents a miscompilation with GCC
>>> 4.8.1 when building on x86 with --cpu=i686
>>
>> I'm not sure if this problem also occurs on 32-bit Linux, if anyone
>> wants to test, that may be useful.
>> A simple build on 32-bit x86 with --cpu=i686 and make fate-vc1-ism
>
> I was able to reproduce it using gcc 4.8.0 on x86 linux, but it seems it's not limited to x86:
> http://fate.ffmpeg.org/report.cgi?time=20130616054303&slot=x86_64-archlinux-gcc-enableshared


Interesting, it doesn't happen on 64-bit for me.

I pin-pointed the specific uses of the function in question that cause
it, but reading the generated asm didn't show anything obvious yet,
but it could be the smallest thing.
Its the 4 mid_pred calls in vc1_pred_b_mv, if anyone else wants to have a look.

- Hendrik