[FFmpeg-devel] [PATCH] Fix configure to disable MMX for Pentium Pro

[Reimar Döffinger]

On Thu, Aug 26, 2010 at 10:38:53PM +0200, Michael Niedermayer wrote:
> On Thu, Aug 26, 2010 at 10:22:57PM +0200, Reimar D?ffinger wrote:
> > On Thu, Aug 26, 2010 at 09:53:44PM +0200, Michael Niedermayer wrote:
> > > On Thu, Aug 26, 2010 at 08:52:24PM +0200, Reimar D?ffinger wrote:
> > > > On Thu, Aug 26, 2010 at 01:37:06PM -0400, Alex Converse wrote:
> > > > > > We should revert r24946: ?Disable MMX for i686 and pentiumpro
> > > > > 
> > > > > If something should be reverted it is dropping runtime detection for MMX
> > > > 
> > > > At the very least I think it would be nice to be able to have SSE,
> > > > does someone know out of their head why we have
> > > > sse_deps="mmx"
> > > > ? I always thought that at least in principle you could have and
> > > > use SSE without MMX and that also would avoid the emms issue...
> > > 
> > > there is no cpu that supports SSE without MMX and if the features no
> > > longer match the cpu then alot of assumtations of what is faster and
> > > what cpus code will run on become wrong
> > > its btw silly to disable mmx on a mmx cpu unless its for debuging or other
> > > uncommon things
> > 
> > How about "emms causes all kinds of issues, I'd like a build without
> > those issues but still ok speed?"
> 
> wrong assumtion thus following theorems are flawed

How about, "while nearly impossible, I can imagine MMX being deprecated
while SSE being kept similarly to 3dnow, as well as I can imagine
global optimization might actually cause real issues with our too-late
emms and thus I'd like to be able to do some tests of SSE-without-MMX
just in case if I don't miss an obvious problem".

> besides i object to allowing (distro) people to build slow versions of
> ffmpeg like SSE without MMX this would only lead to a big disadvantage to
> our users

Maybe, I can't estimate the impact out of my head.

> also some SSE instructions access MMX registers,

I thought I remembered something like this, but I wasn't sure.

> this whole idea of SSE
> withouzt MMX is completly silly and sounds like trolling to me even if its
> likely not intended to be

It's curiosity first of all, secondly a bit of a "long term research"-kind
of thing and not so much meant as a solution to the current issue.

> thats besides there are code paths that lack SSE* optimizations thus
> it wont be ok speed at all for some common codecs

I'd find it interesting to find that out.