I think he's trying to get at what happens if an operation becomes non-atomic. In the case of the long long, there could be two register loads, rather than one. Normally, you would expect calc(1) to return m1 at t0, or m1 at t1, dependent on when the change thread hits. However, if your operation requires two register loads (long long on 32), you would have a third value produced, which is neither m1 at t0 or m1 at t1.

This could be considered a non-thread safe operation.