Re: C++ PATCH: PR 11493 and PR 11495

[Mark Mitchell <mark at codesourcery dot com>]

On Wed, 2003-07-16 at 05:55, Gerald Pfeifer wrote:
> (First, late me state that I really appreciate the huge number of fixes
> and improvements you've recently made!)
> 
> Unfortunately, I found the current failure mode for this patch not to be
> sufficiently helpful:
> 
>   template <class T>
>   struct D : T {
>     void dosomething();
>   };
> 
>   template <class T>
>   void D<T>::dosomething() {
>     f();         // problem
>     this->g();   // fix variant 1
>     T::h();      // fix variant 2
>   }
> 
>   x.cc: In member function `void D<T>::dosomething()':
>   x.cc:8: error: there are no arguments to `f' that depend on a template
>     parameter, so a declaration of `f' must be available
>   x.cc:8: error: (if you use `-fpermissive', G++ will accept your code, but
>    allowing the use of an undeclared name is deprecated)
> 
> Without following gcc-patches and your changes to libstdc++, I certainly
> would not have known that I need to add "this->" to qualify such
> functions.

FWIW, the EDG front end, when running in it's strict mode, just says:

  "test.C", line 8: error: identifier "f" is undefined

The rules for name lookup in templates are pretty arcane.  Knowing when
you need to say "this->" or "typename" is tricky.

One reason I didn't suggest the "this->" trick in the message was that
it may not be the right answer.  If the function is static, 
that will work -- but many people would prefer to write "D<T>::f()". 

Also, the user might be trying to call a namespace-scope function.  For
example, V3 had erroneous code like this:

  template <typename T>
  void D<T>::f(const char* s1, T t) { strlen (s1); }

Here, the solution is "#include <cstring>".

> Alos, I'm not a language lawyer, but is "this->g()" indeed a declaration??

No -- but the use of "this->" postpones lookup because the type of
"this" is dependent on a template parameter.

> I'm a bit afraid I may start sounding like Joseph, but for changes like
> this, which surely will break tons of C++ programs out there, we really
> need to provide more information for our users, for example by providing
> a clear explanation as part of the documentation for -fpermissive and as
> part of the release notes.
> 
> Most importantly, this should include how to rewrite programs affected
> by this.

Do you think the bits that Wolfgang pointed at our sufficient?  Do you
have ideas about how to make them more accessible to people?

Thanks,

-- 
Mark Mitchell <mark@codesourcery.com>
CodeSourcery, LLC