[FFmpeg-devel] [PATCH] get_cabac_inline_x86: Don't inline if 32-bit Windows

[Martin Storsjö]

On Mon, 2 Jan 2023, Christopher Degawa wrote:

> previouslly, it only was an issue with 32-bit clang from msys2's
> mingw32 repo, however, at some point with an update to gcc 12.2.0,
> the same issue popped up. Tested with a clean clone of ffmpeg, and even
> tested with n5.0, but the issue persists, so I presume it's a compiler
> issue.
>
> Related: https://trac.ffmpeg.org/ticket/8903
>
> Signed-off-by: Christopher Degawa <ccom at randomderp.com>
> ---
> libavcodec/x86/cabac.h | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/libavcodec/x86/cabac.h b/libavcodec/x86/cabac.h
> index b046a56a6b..70f990db8d 100644
> --- a/libavcodec/x86/cabac.h
> +++ b/libavcodec/x86/cabac.h
> @@ -178,7 +178,7 @@
> #if HAVE_7REGS && !BROKEN_COMPILER
> #define get_cabac_inline get_cabac_inline_x86
> static
> -#if defined(_WIN32) && !defined(_WIN64) && defined(__clang__)
> +#if defined(_WIN32) && !defined(_WIN64)
> av_noinline
> #else
> av_always_inline
> -- 
> 2.39.0

I'm ok with this change (although I'd reword the commit message a bit); 
the inline assembly here is brittle and it's easy to trigger this failure.

It's very hard to say whether this is a compiler bug, or just our assembly 
crossing the line for what we can demand that the compiler accommodates, 
in combination with deep inlineing - the patch avoids inlineing the 
function which makes it much more possible to handle the inline assembly 
constraints.

I and Hendrik discussed the issue on irc; I reproduced the issue with 
MSYS2's 32 bit GCC, and Hendrik also separately reproduced it with a 
separate build of GCC, by adding the options "-march=pentium4 
-mtune=generic", which are MSYS2's current defaults, which trigger the 
failure (at least in mingw builds, with GCC 12.2.0).

So if there's no reasonable opposition, I'd go ahead and push this patch 
with a slightly reworded commit message.

// Martin