It is faster than using a unrolled loop for the use cases the kernel
cares about (cached, sizes typically < 4K)
Signed-off-by: NAndi Kleen <ak@suse.de>