Following some discussion with Roland, and patterned after the style
anointed by Linus last week, here is a new version of the 32-bit MMIO
copy routine needed by our InfiniPath device.

The name of the routine has changed from memcpy_toio32 to
__raw_memcpy_toio32.  This reflects the basic nature of the routine;
it dodes not guarantee the order in which writes are performed, nor does
it perform a memory barrier after it is done.

The reason for this is that our chip treats the first and last writes
to some MMIO regions specially; our driver performs those directly using
writel, and uses __raw_memcpy_toio32 for the bits in between.

Regarding the specialised x86_64 implementation, Andi Kleen asked me
to perform some measurements of its performance impact.  It makes a
difference of about 5% in performance on moderately large copies over
the HyperTransport bus, compared to the generic implementation.

	<b