I'm trying to think of an optimization for the native types. It should 
be possible to cast any native type element to unsigned char and use the
truncated value for the bytes hash.

Well, for double probably it's required to go from double -> int64_t ->
unsigned char, otherwise the first cast is technically undefined for 
negative values, though it works with gcc.


For non-native types and compound types, Nick's suggestion of
hashing the shape and a couple of values seems to be the best
solution.


Should we do anything about this before 3.3.0?