volatile correctness: combine vs. target.md

[Richard Guenther]

On Fri, Dec 2, 2011 at 1:59 PM, Andrew Haley <aph@redhat.com> wrote:
> On Fri, Dec 2, 2011 at 5:46 AM, <Paul_Koning@dell.com> wrote:
>> >
>> > ...
>>>> > >> It's never correct to exchange volatile accesses.
>>> > >
>>> > >That's not true.  volatile accesses to different memory locations
>>> > >have no special dependence.  If it happens that GCC doesn't
>>> > >do this kind of things then this is only because most passes
>>> > >don't thouch volatile stmts at all (thus the reports for sub-optimal
>>> > >code with volatile - we don't even try to do legal transformations).
>> >
>> > I'm confused.  Do you mean that in
>> >        Volatile int *a, *b;
>> >        Int j, k;
>> >        J = *a; k = *b;
>> > it is ok to fetch *b before *a?  I can't think of any device driver writer who would expect that.
>
> On 12/02/2011 12:27 PM, Richard Guenther wrote:
>> It's ok in terms of GCC internals.  Whether it's ok in terms
>> of the C language specification I don't know (but I think it is ok).
>
> It's not.
>
> 5.1.2.3 Program execution
>
> Accessing a volatile object, modifying an object, modifying a file, or
> calling a function that does any of those operations are all side
> effects, which are changes in the state of the execution
> environment. Evaluation of an expression may produce side effects. At
> certain specified points in the execution sequence called sequence
> points, all side effects of previous evaluations shall be complete and
> no side effects of subsequent evaluations shall have taken place. (A
> summary of the sequence points is given in annex C.)
>
> ...
>
> The least requirements on a conforming implementation are: — At
> sequence points, volatile objects are stable in the sense that
> previous accesses are complete and subsequent accesses have not yet
> occurred.
>
> So, in here:
>
>       Volatile int *a, *b;
>       Int j, k;
>       J = *a; k = *b;
>
> the fetch from *a must complete before the sequence point that is the
> semicolon.

I see.  As we do not explicitely model this dependency we probably
get lucky by the if (gimple_has_volatile_ops ()) bail-out; most
passes do.

Richard.

> Andrew.