[MPlayer-dev-eng] [PATCH] configure cmov detection

[Ivan Kalvachev]

2006/10/26, Ivan Kalvachev <ikalvachev at gmail.com>:
> 2006/10/26, Guillaume Poirier <gpoirier at mplayerhq.hu>:
> > Hi,
> >
> > Diego Biurrun wrote:
> > > On Sun, Oct 15, 2006 at 08:14:45PM +0200, Diego Biurrun wrote:
> > >
> > >>On Sun, Oct 15, 2006 at 09:08:28PM +0300, Ivan Kalvachev wrote:
> >
> > [..]
> >
> > >>OK, then it should just be a matter of adding the proper ifdefs to
> > >>libavcdodec/cabac.h.  Since we already have a precedent for that it
> > >>should suffice.
> > >
> > >
> > > This has been solved differently in FFmpeg, HAVE_CMOV is now defined
> > > there, so we should do the same.  We can either apply Ivan's patch to
> > > configure or port the stuff that Guillaume put into FFmpeg's configure.
> >
> >
> > What I have written of ffmpeg was quite simple and straightforward:
> > the user passes the minimum cpu requirement to the configure script,
> > and based on it we activate or not CMOV usage.
> >
> > On MPlayer, things are different, and could be implemented in a more
> > intelligent manner.
> >
> > We are supposed to be parse /pro/cpuinfo, or the output of this tool,
> > so we can detect CMOV support on host cpu that way (note that FFmpeg
> > doesn't do that AFAIK).
> >
> > We are also supposed to detect the CPU type that fits -mcpu, -march,
> > -mtune flags of gcc, so we can do the same trick as in FFMPEG (i.e.
> > enable CMOV depending on the cpu name we are about to pass to -march).
> >
> > We can also do as Ivan did, which is to test if an asm sample compiles
> > and executes correctly.
> >
> >
> >
> > I think that using /proc/cpuinfo (or the program TOOLS/cpuinfo) is the
> > safest and easiest way to do it BUT that means we need to disable CMOV
> > in case we cross-compile or in case we compile with runtime cpu
> > detection. Not a big deal maybe, but Geexbox folks may not like this.
> >
> > The safest solution is do do as I did on FFmpeg. It requires more
> > maintainance at first look. GeexBox folks will be happier because we
> > should be able to make do with cross-compilation.
> >
> > I don't have any opinion on Ivan's solution, it just doesn't look to
> > be superior to either of the two solutions above.
> >
> > Thoughts?
>
> We can do both.
> As Zuxy pointed out, mmx do both.
> I don't see problem doing the same for cmov.
>
> I did follow the mmx detection as sample, but skipped the "early cut"
> based on specific platform test, as /proc/cpuinfo test are linux
> centric. While
> the asm detection is the only 100% sure.
>
> I'm going to commit my patch, so Guillaume could implement proper
> platform specific tests.
> I guess these are required for cross-compiling.
>
> The only change I'm gonna do, is the line
> "+$_def_cmov    // only define if you have CMOV (i686+, without VIA C3)"
>
> (that's according to what Rathann told me on irc).


Committed