Re: [boost] Re: [boost.tr1?] request for a hash<> implementation

From: Peter Dimov (pdimov_at_[hidden])
Date: 2005-02-07 10:55:13

John Maddock wrote:
>
> If we want a library of string hash functions though, we'd better
> give them different names - there are after all many different ways
> of hashing a string.

There aren't _that_ many good ways to hash a sequence. If you deviate from 
*5/*31, hashpjw, and the shift/xor family there is substantial risk that the 
new and improved hash function leads to pathological cases where a hash_map 
loses to a map with a constant factor of three or more.

For contiguous 8 bit characters there are fast alternatives that process 
four characters at a time but they don't generalize to arbitrary sequences.

Cryptographic hash functions are obviously best from a quality standpoint, 
but they are too slow for a hash_map.

All IMO, of course. :-)