Re: FAIL: gcc.c-torture/execute/950628-1.c execution, -O1

[kenner at vlsi1 dot ultra dot nyu dot edu (Richard Kenner)]

    I don't think so.  It *does* happen to be a more accessible form of
    aliasing information, so that rtx_varies_p and the like can use it.

Hmm.  Of course, the concept of "varies" is a bit dubious.

I'm wondering if *any* of the bits (e.g. scalar and structure) make any
sense as we use alias sets more.

    As for how much we might lose based on reuse, it depends on what the
    lifetime of the object is.  The object has to be live for a large
    enough span that whether or not the value is function invariant
    matters at all.  I suspect, however, that we don't ask this question
    enough to matter much.

Right.

But there are two questions: what to do in the long term and what
to do for GCC 3.0.

I think the former needs more discussion, but I'd say that for 3.0
the *safest* approach is to "allow" multiple stores by changing the
remaining dependence function.  That's both the smallest and safest change.