Which set is "Mark's set"? The last one I posted ~100 minutes ago using the signatures mentioned by Mark and including an implementation?
Yeah, when I did the tests your post wasn't up yet, but we have essentially the exact same implementation of Mark's idea.
It's still slower depending on the cost of the move, only for some classes the compiler can optimize that cost away.
I take it you're comparing the return rvalue vs. return rvalue ref? If so, quite possibly yes. I'll see if I can't identify under what practical situations this solution performs worse compared to returning r-value refs.
Anyways, let's explore the best "return rvalues" version. I think the last version I posted should be really close, except if someone can come up with a pass-by-value version and expressions where it works better in practice.
I think that for now we might start with the code I posted and if future compilers are able to support and optimize pass-by-value to allow even better code, we might detect these compilers and use a different implementation.
Sounds good to me. FWIW, I extended this solution to the other cases. All other cases have 4 overloads except for commutative T + U which has 8. It should be a copy/paste replacement that should work with the operators2.hpp test-bed. Link: http://codepad.org/HdII3G0b