[FFmpeg-devel] [PATCH] x86: move XOP emulation code back to x86inc

[James Almer]

On 03/08/15 7:27 AM, Ronald S. Bultje wrote:
> Hi,
> 
> On Mon, Aug 3, 2015 at 2:41 AM, James Almer <jamrial at gmail.com> wrote:
> 
>> Only two functions that use xop multiply-accumulate instructions where the
>> first operand is the same as the fourth actually took advantage of the
>> macros.
>>
>> This further reduces differences with x264's x86inc.
>>
>> Signed-off-by: James Almer <jamrial at gmail.com>
>> ---
>>  libavcodec/x86/flacdsp.asm     |  9 +++++++++
>>  libavutil/x86/x86inc.asm       | 16 ++++++++++++++++
>>  libavutil/x86/x86util.asm      | 19 -------------------
>>  libswresample/x86/resample.asm |  7 ++++++-
>>  4 files changed, 31 insertions(+), 20 deletions(-)
> 
> 
> OK. (I still think if we ever get more users of the macros, they could move
> back to x86util.asm, but not necessary now.)

AMD dropped XOP for Zen, their post-Bulldozer architecture coming next year, and so
far I'm apparently the only one that ever bothered to write XOP versions of existing
functions for ffmpeg.
I really doubt new code that uses these instructions for the specific case where first
and four operand are the same will show up at all, or at least none where planning
the instruction order in the non-XOP versions to reduce register dependency isn't a
better idea.

Pushed then, thanks.