I'll try out your new approach by replacing some of the ATOMIC_BLOCK macros I'm currently using a number of programs. I also recently ported the ATOMIC_BLOCK macros over to G++ for an ARM Cortex3 architecture (Leaflabs Maple board) for my own use -- I'll try to port your approach as well to test portability.
I also agree it would be nice for portability to get an implementation that doesn't rely on a relatively exotic feature of the Gnu compilers.
Also, since the original ATOMIC_BLOCK macros are actually based on a "for" loop under the hood to specify the code block, it would be theoretically possible to inadvertantly place a "break" or "continue" statement in such a block, and the compiler would not pick it up as out of place. Not a biggie in practice, but a bug yours doesn't suffer from.