On 14 March 2013 16:31, Neil Groves <neil@grovescomputing.com> wrote:
Virtual functions calls add very appreciable overhead relative to the numerous small functions that would need runtime polymorphism for the idiom to work. The overhead is particularly severe with super-scalar processor architectures. To some this won't be important of course.
Of course, if performance it isn't important, then just use std::set and be done with it. Is there any criterion besides performance that people choose between set and unordered_set where generalization would make any sense?
The key question is: who gets to make that decision? ``my_algorithm`` may be happy to use std::set, but it's really the caller's choice. Perhaps the information ``my_algorithm`` needs already lives somewhere in a boost::multi_index_container. Does it really need to be copied just to call ``my_algorithm``? Perhaps I would like to compile ``my_algorithm`` once and stick it in a library file. Does it really need to be recompiled for each concrete container type? -Andy