[Pixman] [PATCH 1/3] mmx: use MMX inline assembly only when USE_MMX

Matt Turner mattst88 at gmail.com
Wed Jul 27 10:19:16 PDT 2011

On Wed, Jul 27, 2011 at 12:59 PM, Soeren Sandmann <sandmann at cs.au.dk> wrote:
> Matt Turner <mattst88 at gmail.com> writes:
>> On Wed, Jul 20, 2011 at 3:29 PM,  <mattst88 at gmail.com> wrote:
>>> +       memcpy(d, s, w);
>> I'm not sure if this change should be made, or whether I should do
>> like the other two hunks and only emit movq instructions for #if
>> defined USE_MMX.
>> memcpy should be optimized in libc, or perhaps the C compiler itself,
>> so.
> Well, if you remove all the MMX instructions from the function, then it
> isn't really MMX specific anymore and could just as well be moved to
> pixman-general.c. I'm not convinced that memcpy() is necessarily
> optimized for MMX in glibc, though.

Right, it doesn't appear to be. It's just as easy to use #ifdef
USE_MMX around the inline assembly, so I'll do that and resend the

> On the other hand, at this point, the main use for MMX on x86 is the
> first version of the OLPC, which had a Geode chip without SSE2 support.
> If you know that the OLPC developers are fine with eliminating the MMX
> support for blt, I'm fine with that too.

I don't see any use in removing working code, so I won't bother with that.


More information about the Pixman mailing list