> At four bits, you may be throwing away information and I don't think
> that's cool.  Even if some selected timings are better with more bits
> shifted, all you're really showing is that there is more randomness in
> the upper bits than the lower ones.  But that doesn't mean than the
> lower one contribute nothing at all.

On the contrary, the expected collision rate for a half-full dictionary
is about 21%, whereas I'm getting less than 5%.  I'm taking advantage of
the sequentiality of addresses, just as int and str hashes do for their
values.

However, you're right that it's only one use case.  Although creating a
burst of objects for a throw-away set may itself be common, it's
typically with int or str, and doing it with custom objects is
presumably fairly rare; certainly not a good microbenchmark for the rest
of the interpreter.

Creating a list of 100000 objects, then shuffling and picking a few
increases my collision rate back up to 21%.  That should more accurately
reflect a long-running program using custom objects as keys in a dict.

That said, I still prefer the simplicity of a rotate.  Adding an
arbitrary set of OR, XOR, or add makes me uneasy; I know enough to do
them wrong (reduce entropy), but not enough to do them right.