the latest patch keeps a timespec: i think that's more correct,
for when we move to nanosecond time nanotime() internally.

The discussion about temporaries and pointer costs reminded me:
shouldn't we  really make nanotime() nanotime be

     struct timespec nanotime __P((void))

so its return value, and anylocal temps, can stay in registers rather
than go via memory, at least on machines with sensible calling conventions.
Would that break anything anywhere?