RFC: merging MEM_ATTRS

[Richard Kenner]

    For all what exactly?  One alias set to include all variables that didn't
    have their address taken is not useful; most of these accesses are
    already described exactly by their MEM_EXPRs.

Sure it's useful: you can set things up so that the alias set used by
indirect references don't conflict with it.  Essentially for each type,
you have two subset alias sets: one for dereferences and one for things
that aren't addressable.  The two don't conflict.

I haven't worked out all the details of this (or even thought about it
much in the last year or so), but the above is the basic idea.