The double-checked locking implemented in com.sun.tools.javac.code.Types.arraySuperType is broken.

Although it could be fixed (by for instance making arraySuperType volatile) the question is if the synchronization is necessary to begin with:

    A) Other parts of javac isn't thread-safe
    B) Instances of Type shouldn't be compared referentially, so returning different objects for different invocations shouldn't be a problem to begin with.