Re: [PATCH] Tree SRA and atomicity/volatility

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 1/8/07, Eric Botcazou <ebotcazou@adacore.com> wrote:
> > > Exactly.  Gigi makes sure that a full-word load is always possible, but
> > > it relies on the side-effectness for keeping it together.
> >
> > I don't see where the tree-ssa part of the middle-end guarantees this and
> > also expand does not seem to honor this as then we'd get a full load
> > for every component load.  It looks to me that for a component write
> > you then need a read-modify-write cycle as we get for targets that
> > don't support sub-word loads/stores and/or strict-alignment targets.
>
> I'm not sure I understand, maybe my wording was ambiguous.  Gigi has all the
> magic to support atomicity (type wrapping, alignment tweaking, countermeasure
> for problematic targets, e.g. Alpha, and so on) and generates the full-word
> access on its own.  The only requirement is for the middle-end to fully honor
> the TREE_THIS_VOLATILE and TREE_SIDE_EFFECTS flags.

Oh, so the only problem you see are for structure copies that are
decomposed by SRA?  Still, middle-end support for this kind of stuff
could be useful - maybe using volatile bitfields instead of a volatile
structure type.

struct val {
 volatile unsigned val1 : 8;
 volatile unsigned val2 : 8;
};

(at the moment accesses to the above are also decomposed by expand).

Richard.