[Ffmpeg-devel] [PATCH] Snow mmx+sse2 asm optimizations
Wed Mar 8 02:39:43 CET 2006
On Wed, 2006-03-08 at 01:31 +0100, Michael Niedermayer wrote:
> On Tue, Mar 07, 2006 at 07:07:06PM -0500, Robert Edele wrote:
> > On Tue, 2006-03-07 at 15:34 -0800, Loren Merritt wrote:
> > > On Tue, 7 Mar 2006, Robert Edele wrote:
> > > > On Mon, 2006-03-06 at 02:06 +0100, Michael Niedermayer wrote:
> > > >> On Sun, Mar 05, 2006 at 06:09:09PM -0500, Robert Edele wrote:
> > > >>> + ::
> > > >>> + "m"(b0),"m"(b1),"m"(b2),"m"(b3),"m"(b4),"m"(b5),"d"(end_w2):
> > > >>> + "%"REG_a"","%"REG_b"","%"REG_c"");
> > > >>
> > > >> this code is not valid, REG_d is changed but neither output nor on the clobber list
> > > >
> > > > REG_d is on the input list, so GCC recognizes it as clobbered? GCC
> > > > also refuses that I put it REG_d into the clobber list. I believe the
> > > > code is good as is?
> > >
> > > If it's both input and clobbered, put it on the output list with "+d".
> > Putting "+d"(end_w2) pr "=d"(end_w2) in the output list causes the
> > program to crash.
> well, you will have to debug it i fear if you want to see the code in cvs
> > The variable is not used after the asm block.
> maybe not at the C level but at the asm level gcc is free to use the register
> and unless you mark it as changed u MUST gurantee that it didnt change
> these are the rules of gcc asm, you can look it up in the gcc manual
> page or some introduction to gcc asm (try google)
I found two bugs, one of which was just as you said - a clobber list
issue. The other bug doesn't effect the result but could effect cache
performance. Both have been fixed and the code now runs with REG_d in
the clobber list.
The new and hopefully correct patch is attached.
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 93624 bytes
Desc: not available
More information about the ffmpeg-devel