[FFmpeg-trac] #4148(avcodec:open): Crash in ff_add_bytes_l2_sse2 when decoding attached APNG file

[FFmpeg]

#4148: Crash in ff_add_bytes_l2_sse2 when decoding attached APNG file
-------------------------------------+-------------------------------------
             Reporter:  benoit       |                    Owner:
                 Type:  defect       |                   Status:  open
             Priority:  important    |                Component:  avcodec
              Version:  git-master   |               Resolution:
             Keywords:  png crash    |               Blocked By:
  SIGSEGV regression                 |  Reproduced by developer:  1
             Blocking:               |
Analyzed by developer:  0            |
-------------------------------------+-------------------------------------

Comment (by kurosu):

 Crap, my comment was not posted, so I'll make it briefer:

 0) Are you sure src1 is always aligned? I think not
 1) Can the caller be fixed to use aligned addresses without copying around
 data? If yes, that should be done to have as few unaligned loads as
 possible
 2) Speed generally matters, so deciding in case of unaligned addresses to
 use the slow, scalar version depends on how often and how large the areas
 it applies to are
 3) If speed matters, a SIMD handling unaligned addresses should be used
 4) Compare its speed to the aligned version for aligned addresses. Larger
 areas, older PCs should show larger differences.
 5) If the speed difference is large enough, add a branch for unaligned
 addresses, otherwise just replace the aligned code
 6) If you're right about src1:
 {{{
 test dstq, 15
 jnz .handle_unaligned
 }}}
 If you're not:
 {{{
 mov iq, dstq
 or   iq, src1q
 test iq, 15
 jnz .handle_unaligned
 }}}
 then xor iq on each path.

--
Ticket URL: <https://trac.ffmpeg.org/ticket/4148#comment:6>
FFmpeg <https://ffmpeg.org>
FFmpeg issue tracker