Re: Incorrect bitfield aliasing with Tree SSA

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

    > The "alias set of the field" would be that of its type, for example
    > 'int'.  Your patch will cause any indirect memory access through a
    > pointer to 'int' to be deemed conflicting with any access to the
    > field, which is precisely what DECL_NONADDRESSABLE_P is meant to
    > avoid.

    You are not answering my question of how the alias set of the field is
    ever used in RTL if the purpose of the flag is *not* to use it (but
    use its parent).

I read all the way to the end of this thread in the hope that things would be
clarified by the time I got back online.  It's a little clearer, but there
was still no direct answer to that question and I think it's important that
there be.

The *whole point* here is that the "alias set of the field" is *not* used
in that case: it's not supposed to be.

I use the term in quotes because I don't really like to think of fields as
*having* an alias set.  *Expressions*, more precisely the subset of those
that reference memory, are what have alias sets.  get_alias_set does it
precisely right and the relevant function is component_uses_parent_alias_set.

If I have
	struct foo {int i; float f;} bar;

and the first field is DECL_NONADDRESSABLE_P, what it means is that *for
every aliasing purpose*, a reference to that field is to be treated the
same way you'd treat a reference to the entire structure, meaning it
should have the alias set of BAR, which is exactly what get_alias_set does.