On xf86 and xf86_64 architectures it means We dont know if a or b is written to that memory location(as a last action), because load/store operations of 32 (for both) or 64 bit (xf86_64 only) memory aligned datatypes are atomic.
On other architectures usually We dont even know what is written there (a garbage) is a valid answer - for sure on RISC architectures, I currently don't know on GPU's.
Note that The fact the code works doesn't imply that it is correct and in the 99% of the times it's the source of sentences like "there's a compiler bug, the code was working until the previous version" or "the code works on the development machine. The server selected for production is broken" :)
EDIT:
On NVidia GPUs we have weakly-ordered memory model. In the description on the Cuda C Programming guide it's not explicitly stated that store operations are atomic. The write operations come from the same thread, so it does not mean that load/store operations are atomic.