[Ffmpeg-devel] DECLARE_ALIGNED_16 not honoured correctly in MinGW

[Benjamin Larsson]

Steve Lhomme skrev:
> Hi,
>
> We are encountering a crash when calling DrFFMPEG from DrDivX on 
> Windows. The crash occurs when decoding WMA content. After some 
> investigation it appears that this declaration in wmadec.c doesn't 
> produce what it's supposed to do:
>
> DECLARE_ALIGNED_16(FFTSample, output[BLOCK_MAX_SIZE * 2]);
>
> When the DLL (could be libavcodec) is called from MSVC code 'output' 
> is not aligned to 16 bits (0x01F36D28 in this case for example). And 
> it crashes the SSE code used later. But it works fine when everything 
> is built with MinGW. It appears that gcc assumes the stack is already 
> aligned a certain way when it does the alignment of this variable in 
> the stack. And this is not true when the stack is given by MSVC code.
>
> Now I understand there's no support for MSVC here, but on the other 
> hand libavcodec is a library, no matter how it's being called. Plus 
> the bug is not in MSVC (who has the right to do whatever they want 
> with the stack) but in gcc.
>
> So my question is: does anybody know a way to make gcc behave 
> correctly in such case ?
>
> The other possibility I see to avoid the crash in my case is to add a 
> test in ff_imdct_calc_sse() and vector_fmul_add_add_sse() to use the C 
> code if one of the parameter passed is not aligned as it should. (I 
> could make the patch if the solution is OK)
>
> Thanks
> Steve

I'm just confirming this. There has been several reported instances of 
this on irc. Both with mingw and with gcc on BeOS. And I would reject 
the proposed imdct hack and use an aligned malloc instead (assuming that 
atleast mallocs can be aligned properly), that should work just as well 
if I remember the code correctly.

MvH
Benjamin Larsson