Update libcxx wait/notify_one/notify_all to use WaitOnAddress/WakeByAddressSingle/WakeByAddressAll on Windows

[johnplatts]

Windows 8 had introduced the WaitOnAddress, WakeByAddressSingle, and WakeByAddressAll functions that can be used to implement __libcpp_platform_wait_on_address and __libcpp_platform_wake_by_address on Windows.

Should a Windows-specific implementation of __libcpp_platform_wait_on_address and __libcpp_platform_wake_by_address that uses the WaitOnAddress/WakeByAddressSingle/WakeByAddressAll APIs be added to libcxx/src/atomic.cpp (although this change requires Windows 8 or later or Windows Server 2012 or later)?

[johnplatts]

@ldionne Should I go ahead and implement __libcpp_platform_wait_on_address using WaitOnAddress on Windows and implement __libcpp_platform_wake_by_address using WakeByAddressSingle/WakeByAddressAll on Windows (this change will require Windows 8 or later along with linking with libsynchronization.a or synchronization.lib)?

Changes also need to be made to clang/lib/Driver/ToolChains/MinGW.cpp to ensure that C++ applications or DLL's that statically link against libc++ also link against libsynchronization.lib.

[RogerSanders]

Any updates/progress here? I ran into this landmine a few days ago, where using std::atomic_flag wait()/notify_all() for thread synchronization resulted in a 2-3x overall program slowdown vs using std::condition_variable, due specifically to the very inefficient implementation here on Windows. The Microsoft STL has this implemented with a platform check, using WaitOnAddress / WakeByAddressSingle / WakeByAddressAll on Windows 8+, with an efficient fallback using some static memory to hold synchronization primitives for earlier platforms. It should be possible to do the same thing here.