i386 memcpy and memset

I've committed improved memcpy and memset routines to the i386
libkern source tree.

These are 'rep xxx' based routines but are significantly faster
than the old routines for short transfers, and slightly faster for
long transfers [1].

The x86_64 routines would almost certainly benefit from the same
changes.

Also I haven't put the changes into libc - I would rather the
same source file got built for both environments.

	David

[1] except for memset to an 8n+4 boundary on P2/P3/P4 when the new
code is about three times faster than the old code for long transfers!

-- 
David Laight: [email protected]