[cvoss]

> sure, use an extra atomic to synchronize threads

What? That would make the situation worse. The execution has a weird unintuitive quirk where the actions of thread 3 seem to precede the actions of thread 1, which seem to precede the actions of thread 2, yet thread 2 observes an action of thread 3. Stronger synchronization would forbid such a thing.

The main question of the article is "Is the memory model _weak enough_ to permit the proposed execution?"