Esa Ilari Vuokko wrote:
My "use case" is in Win32-package, where we simply call a FFI imported call and on failures call GetLastError to get extended error code. GetLastError of course only works on same OS thread and contains sensible results only before next Windows API call. It's much like C/posix errno.
So, we *must* call GetLastError in the same OS thread, and there must not be any Windows API calls in between, including those caused by, say, rts' memory allocation or calls to lock mutexes. Unless the error code is saved via Get/SetLastError.
Also, I'd like not to run the code in a another OS thread (and that won't fix the rts-caused Windows API calls), because the calls are supposed to be fast, and OS-level context switch isn't fast.
Is this problem real, or did I miss some magic feature of rts?
This is most definitely a real problem. On Unix we have some hacks to make errno thread-local: basically we store it in the thread state object for the Haskell thread, and make sure we save/restore it around any RTS operations that might affect it. We need to do the same tricks on Windows using GetLastError/SetLastError, it shouldn't be too difficult. Would you like to submit a ticket? Cheers, Simon