[FFmpeg-devel] [PATCH] HWAccel infrastructure (take 4)

[Reimar Döffinger]

On Thu, Feb 19, 2009 at 06:07:41PM +0100, Gwenole Beauchesne wrote:
> On Thu, 19 Feb 2009, Reimar D?ffinger wrote:
>
>> On Thu, Feb 19, 2009 at 05:04:27PM +0100, Gwenole Beauchesne wrote:
>>> I don't understand this argument. Actually, ::get_format is a pointer to
>>> function and avcodec_default_get_format() is a function, so IIRC the test
>>> should work. However, that would obviously have failed if ::get_format 
>>> was
>>> a pointer. i.e. sizeof(void *) != sizeof(void (*)(void));
>>
>> As I remember it, taking the address of the function in libavcodec will
>> result a pointer to the actual code, whereas in an external program it
>> will point to the start of the stub which on first call activates the
>> dynamic loader.
>
> Then, there probably is a tool problem. I mean, if you build a shared lib, 
> the address of the function in lavc is also the stub. libavcodec is 
> probably not built correctly on that platform.

When you build a program against a Windows .dll, the compiler generates
two things:
1) An import table where the linker will fill in the addresses of the dll's
function after loading the binary
2) a stub/wrapper that consists of something this:
   jmp         dword ptr [a.dll:test]

Taking the address of the function gives you the address of 2). There is
of course the obvious question of why, but I don't know the details of
that, though I suspect it might have something to do with binary
compatibility.

> The comparison is valid or could you please tell me the chapter/verse in 
> the normative reference saying otherwise? I will check tonight but I am 
> pretty sure this has to work.

Fact is it does not work. Obviously I forgot to say "on Windows" a few
times, but no matter how much it is wrong it still does not work on
Windows (neither will it to my knowledge with -Bsymbolic).
Yes, we might just ignore that, but in my view such inconsistent
behaviour is enough to make such checks worse than no checking at all.
You may of course disagree.

> Anyway, I will probably re-add Ivan's idea to use an extra is_hwaccel bit 
> in PixFmtInfo[] to simplify the check and filter out HW accelerated formats 
> in avcodec_default_get_format(). I am sorry but I believe Michael's 
> suggestion has precedence over yours, unless he tells me otherwise now. ;-)

My suggestion was just to avoid code that I thought should not be
necessary in the first place. Changing avcodec_default_get_format is of
course the more thorough solution, I just considered it overkill.