Your June 7 change to expand_expr

[Richard Kenner]

    The other possibility is that Kenner's change is simply incorrect and
    papering over a problem elsewhere -- or that the code that's failing
    for you is doing something unsafe.

I think it's fundamental: there simply is no way to know whether a
REFERENCE_TYPE points into an aggregate or not but yet with the
existing mechanism we *have to know*.

    It's also likely we could use the tri-state to fix 980506-2.c (jfc's
    testcase where cse changes MEM_IN_STRUCT_P in an unsafe manner). 

I think there's more than one such case.

But my concern here is that as more of these cases are found, the bit is
set to the "unknown" value more and more of the time, thus making this
even less useful in practice.  I really think serious thought should
be given to just eliminating it.