Re: Thoughts on Socket (was: New file uploaded to boost)

From: William Kempf (sirwillard_at_[hidden])
Date: 2001-01-03 10:14:50

--- In boost_at_[hidden], "David Abrahams" <abrahams_at_m...> wrote:
> 
> ----- Original Message -----
> From: "Daryle Walker" <darylew_at_m...>
> 
> 
> > I haven't looked at the 'scoped_ptr' stuff, but aren't these 
suggestions
> > overkill?  No other object ('tcpipbuf' is noncopyable) or subclass
> > ('tcpip_data_' is private) is supposed to look at the 
implementation.  The
> > 'tcpipbuf' class could take care of the destruction itself.
> 
> It depends what you mean by "overkill", I guess. The suggestions 
eliminate
> casting, some exception-safety issues, and probably the need to 
write a
> destructor in your class. Sounds like simplification to me. What's 
the
> downside?

Increased compile time dependencies.  You'll have yet another header 
to include and another name brought into the namespace indirectly.

You also have problems with visibility of the destructor for the impl 
class, as discussed in an early thread on this list.

For the "Pimpl Idiom" I prefer to use code like this:

class my_class
{
   ...
private:
   class impl; // scope it in the class to prevent name clashes
   impl* pimpl;
};

This requires code in my_class to manage the lifetime of the impl 
instance, but this is trivial and impacts only the implementation of 
my_class, while making visibility of the destructor a non-issue and 
requires no further compile time dependencies.  This just seems to be 
the safest approach to this pattern.

Bill Kempf