From: Peter Dimov (pdimov_at_[hidden])
Date: 2005-03-16 13:03:41
Dave Harris wrote:
> In-Reply-To: <001001c529b2$4841da80$6501a8c0_at_pdimov2>
> pdimov_at_[hidden] (Peter Dimov) wrote (abridged):
>> Alignment is only going to be a problem if the lower sizeof(size_t) *
>> CHAR_BIT bits of the value produced by reinterpret_cast do not
>> contain enough entropy. I think that in practice the hash function
>> will be good enough.
>
> It partly depends on how the hash_value is used. If the value is
> always a multiple of 8, say, and it is used with a bucket count which
> is a power of 2, then only 1 out of every 8 buckets will get used.
> Entropy in the higher bits doesn't help.
>
> You may be assuming that no-one will use a bucket count which is not a
> prime number. I'd rather not make that assumption if we can reasonably
> avoid it.
This is not an assumption. It is the specification of hash_value that
determines how it is to be used. Currently it does not offer any guarantees
for the entropy of any particular bits, so the client of hash_value is
expected to not do something stupid, like throw away 28 of its 32 bits.