From: Daniel James (daniel_at_[hidden])
Date: 2005-03-21 15:20:20
Thorsten Ottosen wrote:
> I would like to thank all who partcipated:
Thank you from me as well, and thanks to Thorsten for managing the review.
> Issues and solutions:
> ---------------------
>
> 1. for those who don't want to include the whole std-lib, individual headers
> can be put in hash/std/XX.hpp (but see 3)
This should probably be a sub-directory of functional, as hash lives at
boost/functional/hash.hpp.
> 4. avoid too many zero hash values by defaulting to something other than 0
I think the preferred solution was to add a constant to hash_combine
(which seems to work well). Although, it might be good to have an
'initial_seed' constant.
> 5. change implementation of hash values for pointers so undefined behavior is
> gone (and consider adding x + (x >> 3) )
I'm going to do that.
> 6. the design aim was predictibility and to minimize collision; for the latter
> part, I
> would like to know if similar values hash to similar buckets or not. If
> very different
> objects end up in the same buskets, wouldn't equality be much faster
> rejected?
> Should this be a concern?
Generally, values that collide should not be similar, (which is why
collisions of sequences of zeros was a problem), although I'm sure that
there are still exceptions to that rule. Especially because it's quite
hard to define what 'similar' means in general.
Daniel