There's a long history of cheating and setting final fields in pseudo-constructors readObject and clone, which is well motivated since Java should really support pseudo-constructors in a better way.. Cheating has used Unsafe or reflection with setAccessible (CopyOnWriteArrayList.resetLock()).
I haven't had any success actually reproducing any data race failures in constructors - on modern machines it would only happen if the JIT reordered the writes. Doug - would CopyOnWriteArrayList.clone() be slightly safer if it had a releaseFence() ?
