Greg Stark <stark@mit.edu> writes:
> I was originally going to suggest using a crc to combine but iirc we
> changed hash_any() a while back and decided against using crc. I don't know
> if that was wise but wouldn't want to suggest relitigating that.
Nah, CRCs are designed to solve a different problem, ie detecting
single-bit and burst errors with high probability. In particular, I don't
think they make any particular promises with respect to spreading changes
into all bits of the result. That's important for our hash functions
because we usually take just the lowest N bits of the result as being
adequately randomized.
regards, tom lane