How the MCF thread model works

Overview

The now committed patch implements compiler, linker and standard library support for the MCF thread model. The CRT part has not been committed to mingw-w64.

If GCC is configured with --enable-threads=mcf:

1. It has __USING_MCFGTHREAD__ as a pre-defined macro, which enables special code paths in libstdc++ and mingw-w64 CRT.
2. It has -lmcfgthread -lntdll as default libraries. (NTDLL is only needed for static linking.)
3. It installs 'gthr-mcf.h' as 'bits/gthr-default.h', which is just a wrapper for 'mcfgthread/gthr.h'.

In libgcc, most use of threading infrastructure is internal. In libstdc++ there are some public functions that are to be replaced with MCF implementations:

1. __cxa_thread_atexit(): On Linux it is provided in libstdc++ and jumps to __cxa_thread_atexit_impl() in libc. Recent mingw-w64 provides this function, so it is opted out to prefer the mingw-w64 one. (There's a bug about it: https://github.com/msys2/MINGW-packages/issues/2519)
2. __cxa_guard_acquire(), __cxa_guard_release() and __cxa_guard_abort(): On Linux they are implemented with futexes. On most other platforms they are implemented with a single static mutex. They call __MCF_cxa_guard_acquire(), __MCF_cxa_guard_release() and __MCF_cxa_guard_abort(), respectively.