https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104928
--- Comment #10 from Benjamin Buch <benni.buch at gmail dot com> --- Okay, I understand this. https://gcc.gnu.org/onlinedocs/libstdc++/manual/status.html#status.iso.2020 defines P1135R6 is implemented since libstdc++ 11.1 and the feature test macro __cpp_lib_semaphore >= 201907L is defined accordingly while the implementation is known to produce wrong code for nearly 2 years and three major releases now. It brings us into a situation again, where users have to detect compiler and compiler version to check whether they can use std::counting_semaphore instead of just checking the feature test macro. In my option this is a very bad and hard to acceptable situation, even in an experimental mode. Especially since C++20 already has a wide adoption these days. So I ask you to undefine the feature test macro and set the implementation state to partial. Also I propose to introduce -std=c++XXpreview flags for all (new) experimental language modes instead of -std=c++XX which misleadingly indicates that the mode is ready to use. The documentation might name the modes as experimental, but I think its much more user-friendly to make the flag itself expressive. (MSVC also starts doing this with the upcoming release 2022 17.13 for C++23. There will be a c++23preview mode.) I understand that this might break existing code. Nevertheless in my opinion its better to loudly fail at compile time that silently doing the wrong thing at runtime. We all know that software contains bug. That's normal. To actively inform the users is important and user-friendly. Sometimes even more important than fixing the bug itself. ;-)
