> The libc's bzero reports only 4MB/s bandwidth, while unrolled
> write reports 150MB/s. I doubt that there might have something
> wrong with libc's assembly version of bzero, even bcopy (pipe has
> low bandwidth, too)

I ran lmbench on a mobilepro 770 (131 MHz cpu) and looking at the the
following:
	"unrolled partial bcopy unaligned
	0.000512 183.06
	0.001024 225.01
	0.002048 120.18
	0.004096 24.72
	0.008192 13.86
	0.016384 13.86
	0.032768 13.87

	"libc bcopy aligned 
	0.000512 13.24
	0.001024 13.11
	0.002048 13.01 
	0.004096 12.91
	0.008192 12.83
	0.016384 12.80
	0.032768 12.81
	0.065536 12.85
	0.131072 12.81
	0.262144 12.81
	0.524288 12.80
	1.05 12.83 
	2.10 12.83
	4.19 12.78

It looks like the unrolled bcopy is managing to somehow benefit from
the cache much more than the libc version.   Once they get out
beyond a couple of K they are really about the same.