...Blast. it turns out that, due to a glitch with the mips ->
{mipsel,mipseb}, all the profiling of rcons blitting i'd done had used
the non-tuned C-coded kernel memmove().

I will rebuild a kernel and rerun the tests as soon as I get done with
an attempt to do a `make world'.  UTSL'ing, using a tuned bcopy()
should make an immediate 4x difference on write bandwidth; i dunno
about read bandwidth.