Re: [C++ Patch] PR 35331

[Mark Mitchell <mark at codesourcery dot com>]

Paolo Carlini wrote:
> Hi Mark, all,
> > This is the sort of situation where I think we should be much more 
> > brutal.
> > > + template<typename...> struct A;
> > > + + template<typename...T> struct A<T*> // { dg-error "not expanded|T" }
> > > + {                                   // { dg-error "not expanded|T" }
> > > +   friend void foo();
> > > + };
> > Rather than trying to make add_friend -- and every other function that 
> > might be called as we try to build up members of a erroneous class -- 
> > we should short-circuit the whole process.  I would suggest we just 
> > skip everything inside the erroneous class -- look for the matching 
> > closing brace and be done with it.  That will no doubt eliminate a 
> > bunch of crashses, making the compiler more robust.  The downside, of 
> > course, is that the user may not get errors about things within the 
> > erroneous class until they fix the class -- but with a class that 
> > makes no sense, it's not clear that there's a whole lot in there we 
> > can be intelligent about.
> Below you will find my second try, which tries to implement your 
> suggestions. Tested x86_64-linux.

It looks better -- but can we go even further?  Can we arrange to have 
begin_class_definition return error_mark_node in this case?  I'd like to 
avoid having these half-formed bits of tree structure hanging about; 
they're just so darned crash-prone.

Thanks,

--
Mark Mitchell
CodeSourcery
mark@codesourcery.com
(650) 331-3385 x713