Joel Jacobson <joel@trustly.com> writes:
> I've seen a performance trick in other hash functions [1]
> to instead read multiple bytes in each iteration,
> and then handle the remaining bytes after the loop.
> [1] https://github.com/wangyi-fudan/wyhash/blob/master/wyhash.h#L29
I can't get very excited about this, seeing that we're only going to
be hashing short strings. I don't really believe your 30% number
for short strings; and even if I did, there's no evidence that the
hash functions are worth any further optimization in terms of our
overall performance.
Also, as best I can tell, the approach you propose would result
in an endianness dependence, meaning we'd have to have separate
lookup tables for BE and LE machines. That's not a dealbreaker
perhaps, but it is certainly another point on the "it's not worth it"
side of the argument.
regards, tom lane