> When hacking on the V-Kernel, I found a measurable win in eliminating
   > all FP in the kernel, and using a fixed-point load average.  I don't
   > know if Core would buy that, because the floating-point load averages
   > were, last time I looked, a machine-independent part of the kernel.

   It seems like if somebody did the work, it would be worth putting in.
   Ideally there shouldn't be any floating point code in the kernel,
   other than to support the FPU's use by user processes.

Um, the load average *is* kept in fixed point.  See sys/resource.h and
vm/vm_meter.c.  Indeed, I just grepped through the kernel and found no
obvious uses of floating point (outside of the netiso code, anyway).