aggregate alias anomaly

[Jeff Sturm <jsturm at one-point dot com>]

Given:

struct X {int x;};
struct Y {int y;};

int f(struct X *x, struct Y *y) {
        x->x = 0;
        y->y = 1;
        return x->x;
}

gcc -O2 -fomit-frame-pointer yields:

f:
        movl    4(%esp), %eax
        movl    8(%esp), %edx
        movl    $0, (%eax)
        movl    $1, (%edx)
        movl    (%eax), %eax
        ret

for every released compiler I tried, plus mainline.  Note the useless load
of x->x, given that x and y cannot possibly have conflicting alias sets.

Currently, tree-ssa fares better:

f:
        movl    4(%esp), %eax
        movl    $0, (%eax)
        movl    8(%esp), %eax
        movl    $1, (%eax)
        xorl    %eax, %eax
        ret

In both cases the alias sets appear correct, but mainline tests the alias
sets of `struct X' vs. `int' which must conflict, since the former is a
superset of the latter.

Since I'm no expert in RTL I'll ask: is this a known problem?  Is it hard
to fix for some reason?  (For tree-ssa I read the description of
may_alias_p, and it makes sense to me, whereas I'm lost in flow.c.)

As far as I can see, alias analysis for Java is useless on mainline.

Jeff