SRA problem with uninitialzed fields

[Richard Kenner]

    A proper implementation that keeps the types those that accord with
    the language semantics would be for the GIMPLE generated to include
    explicit mask operations for assignments to bit-fields.  If it doesn't
    represent that assignment to a bit-field from the same type may change
    the value, then there are likely to be problems.  At one point there
    were, but I don't see such problems at present in some quick tests.

Well, look at the test case I sent. A mask is clearly required on the
comparison (at least, I think so!), but the only mechanism to generate it
is the hook, which is false for C++.