On 03/01/2008, Thorsten Ottosen <thorsten.ottosen@dezide.com> wrote:
Can boost::compressed_pair not be used to remove the overhead of stateless functors and hashers?
I used to do that but I found that in practice the difference was so slight it didn't seem worth it. If you have 4 zero-sized function objects (2 hash functions, 2 equality functions), it's the difference between 1 or 2 bytes and 4 bytes. Which often doesn't make a difference because of padding. The pointer to the current pair of objects (typically 4 bytes) is where the real cost lies. I could possibly use a different implementation for empty objects so that the pointer isn't required, but I'm not sure about the technicalities or if it would be worth the extra complexity. I could also use a bool instead of a pointer but that has costs elsewhere (I haven't measured this, so I'm not entirely sure how expensive it would be). It might be worth using compressed_pair (or something similar) for allocators, but I never got round to doing that. I've generally not been that concerned about the size of the container object, as it tends to be small compared to the memory required for the buckets and nodes. Daniel