Example:

Thread 1 does A.lock, A.unlock, B.lock, B.unlock

Thread 2 does B.lock, B.unlock, A.lock, A.unlock

That's fine, but swap the middle pair of unlock followed by lock on both threads and you have a standard deadlock based on different lock nestings.

And to reiterate from my previous comment, I do think that transform is or should be forbidden. That's not for synchronization reasons per se, but because the transform can change the set of externally visible side effects of the code. This argument requires that a store with release semantics is considered to be an externally visible side effect. That makes intuitive sense to me, though I don't remember what e.g. the C++ standard actually says about that.