This allows a more efficient representation for 64-bit hosts.
It should be about the same for 32-bit hosts, as we can still
access the individual pieces of the double.
Signed-off-by: NRichard Henderson <rth@twiddle.net>