Am Sonntag, den 23.11.2014, 14:39 -0500 schrieb Rich Felker:
> On Sun, Nov 23, 2014 at 07:01:24PM +0100, Jens Gustedt wrote:
> > No, I don't think it is evaluated, 6.5.3.4 says:
> > 
> >   If the type of the operand is a variable length array type, the
> >   operand is evaluated;
> > 
> > observe that it says VLA, and not VM type.
> > 
> >   otherwise, the operand is not evaluated and the result is an integer
> >   constant.
> > 
> > There are no exception for other VM types. To deduce the size for our
> > case here, an evaluation isn't necessary.
> > 
> > (Hoping that the __typeof__ stuff follows the same rules.)
> 
> OK, thanks for clarifying this. In that case I think we're safe as
> long as __typeof__ follows the same rules.

Bad news, it seems that typeof does in fact evaluate VM types in typeof:

https://gcc.gnu.org/ml/gcc-patches/2013-11/msg01378.html

I am not much convinced that it is necessary to have that rule, if it
is not a VLA. For any value expression, the necessary type information
should be available without evaluation of the expression. (Type
expressions are different, but we don't need that, here.) But anyhow,
we can't change that.

So the only way to get away with it to support all allowed atomic
types would be to use the __auto_type extension that they introduce in
that patch. That would only allow us to provide support for recent
gcc.

For older gcc, say from the point where they have the __sync builtins,
we could still

 - support all types with sizes that fit to the natively supported
   instruction set, usually 1, 2, 4 and 8, by juggling with sizeof
   expressions.
 - Only allow the atomic type specifier version "_Atomic(typename)"
   and not the qualifier version "_Atomic typename"
 - Only have the atomic generic macros to use these types, no
   operators allowed.
 - Don't support atomic struct

One way to do this would be

#define _Atomic(T) __typeof__(__typeof__(T) volatile[1])

e.g the result for int would be equivalent to

typedef int volatile atomic_int[1];

This would still have it binary compatible with future version that
really have _Atomic as a keyword. On the other hand, it would not
allow to assign such a variable.

Jens

-- 
:: INRIA Nancy Grand Est ::: AlGorille ::: ICube/ICPS :::
:: ::::::::::::::: office Strasbourg : +33 368854536   ::
:: :::::::::::::::::::::: gsm France : +33 651400183   ::
:: ::::::::::::::: gsm international : +49 15737185122 ::
:: http://icube-icps.unistra.fr/index.php/Jens_Gustedt ::