Thank you again for the PR.

I am closing in favor of #7010, which merges your change, and another, to v2.4.x. We are not going to be using the next branch in the future, so this PR does not need to be merged to next.

If you are interested, see the changes in the second commit of #7010. We need NOINLINE to stop the compiler from optimizing access to thread local storage in this circumstance:

auto x = TLS_get_foo();
on_thread_t th(other);
auto y = TLS_get_foo();

The compiler (or link time optimizer) does not know that on_thread_t th(other) will switch the coroutine to a different POSIX thread. Thus it may assume that both calls to TLS_get_foo(), after inlining, both access a thread_local, which on_thread_t's constructor does not modify, thus they return the same value. Using NOINLINE, we prevent this optimization.

I'll add, because thread_local is a C++11 feature, I've decided to not worry about platforms which do not support it (which might be CentOS 6 with an old system gcc 4.8 compiler).