Anonymous types and [basic.link] for nonlocal variables

[Joe Buck]

On Tue, Mar 09, 2004 at 02:29:43PM -0800, Matt Austern wrote:
> Geoff checked in a patch a few days ago, 
> http://gcc.gnu.org/ml/gcc-patches/2004-03/msg00346.html
> 
> Among other things, this patch enforces the rule that variables with 
> linkage may not be defined in terms of names with no linkage.  
> Previously we enforced that rule selectively: we made a special 
> exception for anonymous types.  That is, we used to allow things like
>    enum { foo } e;
> or
>    struct { int x; int y; } pairs[] = { { 1,2 }, { 5,8 } };
> 
> I didn't notice this patch when it went by a few days ago.  Now I'm 
> wondering if it's a good idea.  Allowing variables to be defined in 
> terms of anonymous types was a deliberate decision (there was a comment 
> in the code pointing it out), because forms like this are legal in C 
> and common, albeit illegal, in C++.  If we're going to reverse a 
> deliberate decision, we should make sure to give it a little thought.
> 
> This is one of those annoying problems where in one sense this is 
> clearly a correct change: it brings us into closer conformance with the 
> language standard that we claim to implement.  On the other hand, it 
> breaks the QT build.  I bet QT isn't alone.

In my view, anything that breaks QT is a regression, no matter how
many standards documents we can point to as justification.  Certainly
-pedantic should object to such usage, and perhaps future QT versions
should avoid such usage if possible, but we need to build QT.