[FFmpeg-devel] [PATCH] avcodec/vp9mvs: fix misaligned access when clearing VP9mv

[Ronald S. Bultje]

Hi,

On Sun, Jun 2, 2024 at 9:12 AM James Almer <jamrial at gmail.com> wrote:

> On 6/2/2024 10:06 AM, James Almer wrote:
> > On 6/2/2024 9:14 AM, Kacper Michajłow wrote:
> >> Fixes runtime error: member access within misaligned address
> >> <addr> for type 'av_alias64', which requires 8 byte alignment.
> >>
> >> VP9mv is aligned to 4 bytes, so instead doing 8 bytes clear, let's do
> >> 2 times 4 bytes.
> >>
> >> Signed-off-by: Kacper Michajłow <kasper93 at gmail.com>
> >> ---
> >>   libavcodec/vp9mvs.c | 3 ++-
> >>   1 file changed, 2 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/libavcodec/vp9mvs.c b/libavcodec/vp9mvs.c
> >> index b706d1660f..790cf629a6 100644
> >> --- a/libavcodec/vp9mvs.c
> >> +++ b/libavcodec/vp9mvs.c
> >> @@ -294,7 +294,8 @@ void ff_vp9_fill_mv(VP9TileData *td, VP9mv *mv,
> >> int mode, int sb)
> >>       VP9Block *b = td->b;
> >>       if (mode == ZEROMV) {
> >> -        AV_ZERO64(mv);
> >> +        AV_ZERO32(&mv[0]);
> >> +        AV_ZERO32(&mv[1]);
> >>       } else {
> >>           int hp;
> >
> > IMO just move mv in VP9Block to the top of the struct. That will make
> > sure it's aligned to at the very least 16 byte (Since it's av_malloc'd).
>
> Actually nevermind, VP9mv has two int16_t and given what's passed to
> ff_vp9_fill_mv() it's not enough.
>

Do compilers on relevant platforms convert this to a single 64bit
(unaligned) zero-move? Otherwise, we may want an unaligned AV_ZERO64() so
as to not slow down platforms supporting unaligned writes.

Ronald