Compiler optimization breaks multi-threaded code

[stephan]

> atomicOp uses a CAS loop for the RMW operations.
Ignore my comment. I should have looked at the code in core.atomic 
before commenting. I just had one test case with atomicOp!("+=") that 
worked, and assumed that atomicOp!("+=") was implemented with "lock xadd".

> I'm thinking of exposing atomicStore and atomicLoad in core.atomic so folks have something to use until the compiler work is taken care of.
This would be a good solution. Not only does it solve this ordering 
issue, but it is actually what I want to do with a shared variable while 
I am already within a critical section anyways.