It was possible for a thread to read invalid memory after a conflict
when multiple threads were synchronising.

I haven't been successful in constructing a test case that triggers
this, but we have some internal code that ran into it.