[boost] Re: move_ptr

From: Howard Hinnant (hinnant_at_[hidden])
Date: 2004-01-23 22:28:40

On Jan 23, 2004, at 9:24 PM, Reece Dunn wrote:

> I'm not sure of how is_convertable is implemented, but can't 
> noncopyable be added as a special case, like generating a compile-time 
> error when used as a parameter to it, e.g.
>   error: cannot instantiate disallowed_conversion< noncopyable >
>   from is_convertable< noncopyable, T >::type
> or something similar.
>
> This should be a case of adding some MPL code to detect is_same< 
> noncopyable, T >. The only problem I forsee would be in compile time, 
> or am I missing something?
>
> e.g. (pseudocode style):
>
> is_convertable_ex< T1, T2 >::type =
>   if_
>   <
>      or_< is_same< T1, noncopyable >, is_same< T2, noncopyable > >,
>      disallowed_conversion< noncopyable >,
>      is_convertable< T1, T2 >
>   >::type

The big picture is how does is_convertible handle types with private 
copy ctors (as opposed to just deriving from noncopyable).  And it is a 
bigger problem than just saying:  is_convertible should be a compiler 
intrinsic and it should just do the right thing.

C++ is currently set up so that given a class:

class foo
{
public:
private:
        void bar();
};

If you change the access of foo::bar()  from private to protected or 
public, the behavior of your working program is guaranteed not to 
change.  This is a useful invariant in the language.  This invariant 
translates to compiler behavior as:  Name lookup occurs without regard 
to access.  Access is checked after name lookup has identified a unique 
best overload.

So the suitability (lookup) of copy constructors and conversion 
constructors is performed without regard to access (and that is how 
is_convertible works).  If we make is_convertible smarter than that 
(via language changes), then a major cornerstone of the language is 
changed:  changing access can silently effect runtime behavior.  I do 
not currently possess the foresight to know if such a change would be 
wise or not.  I'm nervous enough to guess that it would be a bad 
change.  But perhaps others have better guesses.  See D&E section 2.10 
for a more informed discussion.

-Howard