So I was doing some profiling of memcpy() calls to try and tune
the UltraSPARC-III memcpy a little more, and you wouldn't believe
what was at the top of the list during a kernel build :-)

It's bh_lru_install.  On a 64-bit system every time the local
cpu's lru->bh[0] doesn't match 'bh' we do a 64-byte memcpy()
from the stack into the lru->bh[] array.

It shouldn't be too hard to make the code just work without
an on-stack copy, shuffling the lru->bh[] array entries
directly.

For the curious, attached is profiling output after a full kernel
build on this box.  I basically rigged it such that the per-cpu
timer profiling stuff didn't touch prof_buffer, and every
memcpy() call recorded __builtin_return_address(0).  It also
keeps alignment and length distribution information accessible
via proc.  Example dumps and that debugging patch are attached
too.