Re: Protected constructor access rules changed?

[Mark Mitchell <mmitchell at usa dot net>]

>>>>> "Rick" == Rick Campbell <campbell@cyberpass.net> writes:

    Rick>     Date: Wed, 4 Feb 1998 08:57:52 GMT From: Mark Mitchell
    Rick> <mmitchell@usa.net>
    
    Rick>     Constructors may not be called by name, except in the
    Rick> initializers portion of a derived class constructor.  I
    Rick> *think* this has been true since the ARM, but I could be
    Rick> wrong.  So, I don't believe there is a bug here.

    Rick> Hmmm.  That may well be, but in a bit over 12 years of C++
    Rick> hacking, I've never heard anyone suggest that before.  While
    Rick> I'm not as familiar with the current spec, I've been back
    Rick> and forth through the ARM enough times that I'm pretty
    Rick> confident that it does not disallow calling constructors by
    Rick> name.

I apologize.  I missed the point of your example, due to an overly
cursory reading.  However, I still believe that egcs is correct, and
here is why:

    [class.protected]

    When a friend or a member function of a derived class references a
    protected nonstatic member of a base class, an access check applies in
    addition to those described earlier in this clause.  Except when
    forming a pointer to a member, the access must be through a pointer
    to, reference to, or object of the derived class itself (or any object
    derived from that class.)... 

    Example:

    class B { protected int i; };

    class D1 : public B { };

    class D2 : public B { void mem(B*, D1*); };

    void D2::mem(B* pb, D1* p1)
    {
      pb->i = 1; // ill-formed
      p1->i = 2; // ill-formed
      this->i = 3; // ok
    }

I aplogoize for wasting your time with my earlier answer.

    Rick> In fact paragraph 2 of 12.1 appears to explicitly allow such
    Rick> a ``call'' -- technically considered a type conversion --
    Rick> and paragraph 13A shows an example of making such a call:

    Rick>     13A functional notation type conversion
    Rick> (_expr.type.conv_) can be used to create new objects of its
    Rick> type.  [Note: The syntax looks like an explicit call of the
    Rick> constructor.  ] [Example: complex zz = complex(1,2.3);
    Rick> cprint( complex(7.8,1.2) ); --end example] An object created
    Rick> in this way is unnamed.  [Note: _class.temporary_ describes
    Rick> the lifetime of temporary objects.  ] [Note: explicit
    Rick> constructor calls do not yield lvalues, see _basic.lval_.  ]

    Rick> At the point, I feel much more confident in saying that
    Rick> egcs's failure to compile the attached code is a true bug.
    Rick> The is purely an access issue.

    Rick> In any case, if what you say is true, that egcs has a
    Rick> different bug in that it permits calling most constructors
    Rick> by name.  For example, in the code that I sent out
    Rick> previously, if you move the first `protected' keyword from
    Rick> just before to just after the the void* constructor, it
    Rick> compiles without incident.

    Rick> 			Rick


-- 
Mark Mitchell		mmitchell@usa.net
Stanford University	http://www.stanford.edu