Re: [boost] shared_ptr: unsafe, or am I crazy?

From: Luigi Ballabio (ballabio_at_[hidden])
Date: 2002-01-14 15:21:50

>  > At 08:56 AM 1/14/02 -0500, you wrote:
>>  >struct Base { Base(); ~Base(); };
>>  >struct Derived : Base { Derived(); ~Derived(); };
>>  >
>>  >void unsafe()
>>  >{
>>  >     shared_ptr<Derived> d(new Derived);
>>  >     shared_ptr<Base> b(d);
>>  >     d.reset();
>>  >     b.reset(); // equivalent to delete (Base*)(new Derived)!!
>>  >};
>>  >
>>  >Isn't shared_ptr just screaming for the application of John Maddock's new
>>  >is_polymorphic<> trait?
>>
>>  Maybe I'm just about to embarass myself, but isn't Base just screaming for
>>  having a virtual destructor instead, if you want to use it this way?
>
>The point is that non-polymorphic public bases are an important and useful
>construct, and safe as long as you don't explicitly delete a derived through
>the base pointer.

I wholeheartedly agree, and my point was that they should be used 
exactly as such---as non-polymorphic. Maybe I'm just being dense (no 
irony. Maybe I'm just being dense. Really.) but what would be the 
reason for upcasting a derived class to its base, if dynamic dispatch 
is not to be used? If there are no virtual function in Base, casting 
Derived* to Base* or shared_ptr<Derived> to shared_ptr<Base> is 
surely a lot less nasty than slicing Derived to Base when passing by 
value (since no harm is done to the actual object if one doesn't 
delete the Base*). However, it looks suspicious to me, since deleting 
the object through the Base* is not the only harm one can do. What if 
the base class has a setter method, and the derived class overrides 
it to call the base implementation _and_ update some additional state 
specific to Derived? If one calls such method through the 
non-polymorphic Base* or shared_ptr<Base>, isn't the Base 
implementation called, thus leaving the additional Derived state 
invalid? In this case, I'd rather have a visible leak which signals 
me that something is fishy with the cast I let slip in, rather than 
having code that seems to work fine but with a subtle bug which might 
change my results without me noticing. But that's just me (and a 
rather humble me, too).

Bye,
        Luigi

--