https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62056
--- Comment #5 from Piotr Dziwinski <piotrdz at gmail dot com> --- (In reply to Jonathan Wakely from comment #4) > tr1::tuple doesn't support perfect-forwarding or move semantics > > tr1::tuple doesn't support uses-allocator construction > > tr1::tuple doesn't support 'final' classes > > tr1::tuple doesn't have correct exception specifications > > tr1::tuple doesn't prevent implicit conversions that would use explicit > constructors > > tr1::tuple doesn't support tuple concatenation > > If you can add all those features to the <tr1/tuple> implementation so that > it meets the C++11 requirements and it still compiles faster then I'd be > interested in your analysis of the remaining differences. Otherwise I'm > going to assume the difference is because the <tuple> header contains more > than twice as many lines of code and several additional features. Ok, I understand it now. I was speaking from only somewhat experienced user perspective and I did not realise the deeper implications of standard compliance. Just for the record, I did some testing and found two important factors here are: - dependency to <array> (to enable it in tuple concatenation) - (change in compile time 0.357s -> 0.231s), - allocator constructors - (change 0.231s -> 0.185s). So in the end please ignore my interruption with `std::tr1::tuple`. It seems the recursive version of `std::tuple` is not going to be optimized easily and the new flat implementation is the way to go here.
