There are two timing issues in the JVM package export handling on non-TSO systems.

I believe we have issues with:
1) boolean is_exported_unqualified relative to is_exported_all_Unnamed
and
2) boolean is_exported_unqualified relative to _qualified_exports

Reflection::verify_class_access is an example of a lock-free reader which calls:
    is_unqual_exported:
       assert(!(_qualified_exports != NULL && _is_exported_unqualified)
    assert(!(_is_exported_allUnnamed && _is_exported_unqualified)

But another thread could at the same time under a lock run:
    set_unqual_exported()
       _ is_exported_unqualified = true
       _ is_exported_allUnnamed = false
       _ qualified_exports = NULL

With an out-of-order write, you can provoke either of the asserts in is_unqual_exported.