boost::shared_ptr and multiple inheritance

[EnsGabe]

Suppose you have a class heirarchy as such:


class Base{
....
};


class Mid1 : public Base{
....
};

class Mid2 : public Base{
....
};


class Derived: public Mid1, Mid2{


};

Mid1, Mid2, and Base are ABC.

What is an effective way to manage the lifetime of this object?
Currently, I leak Derived (and anything similar to Derived) all over
the place by not reclaiming them at all. That works for now, in as
much as when I am done with them I am done with execution of my
program, but is not a solution that will work for the future.

My use at the moment is as such:

class ObjectHolder {
public:
....
void addMid1(Mid1 *p) { assert(p); mid1s.push_back(p); };
void addMid2(Mid2 *p) { assert(p); mid2s.push_back(p); };
private:
std::vector<Mid1*> mid1s;
std::vector<Mid2*> mid2s;
....
};

There will be many different classes that are equivalent to Derived
(in the sense that they are deriving from Mid1 and Mid2.) There will
also be classes that derive from one or the other of Mid1 and Mid2.
All of them will be registered exactly once in mid1s and/or mid2s,
whichever they are derived from. References to the objects will not
exist outside of ObjectHolder. Objects may contain references to
other objects; the directed graph that they form will be acyclic.

My first instinct was to delete every element of mid1s and mid2s in
the destructor for ObjectHolder, but that will not work when a Derived
has been inserted into both vectors. My next idea was to use
boost::shared_ptr and use that to automatically manage the lifetime of
the object, but there was nothing in the documentation that I saw that
led me to believe that one could use a shared_ptr across a class
heirarchy. Since the 'this' pointer for an object is different
depending on the type it is being used as, I'm (possibly erroneously)
assuming that a shared_ptr will mismanage the reference count for the
pointer since they are using different pointers. Am I wrong in my
understanding of the semantics of boost::shared_ptr?

(PS: It's obvious that what would solve my problem is a GC- I'm not
ruling it out at this point, but I am exploring my options)

 

[Kai-Uwe Bux]

Suppose you have a class heirarchy as such:


class Base{
...
};


class Mid1 : public Base{
...
};

class Mid2 : public Base{
...
};


class Derived: public Mid1, Mid2{


};

Mid1, Mid2, and Base are ABC.

What is an effective way to manage the lifetime of this object?
Currently, I leak Derived (and anything similar to Derived) all over
the place by not reclaiming them at all. That works for now, in as
much as when I am done with them I am done with execution of my
program, but is not a solution that will work for the future.

My use at the moment is as such:

class ObjectHolder {
public:
...
void addMid1(Mid1 *p) { assert(p); mid1s.push_back(p); };
void addMid2(Mid2 *p) { assert(p); mid2s.push_back(p); };
private:
std::vector<Mid1*> mid1s;
std::vector<Mid2*> mid2s;
...
};

There will be many different classes that are equivalent to Derived
(in the sense that they are deriving from Mid1 and Mid2.) There will
also be classes that derive from one or the other of Mid1 and Mid2.
All of them will be registered exactly once in mid1s and/or mid2s,
whichever they are derived from. References to the objects will not
exist outside of ObjectHolder. Objects may contain references to
other objects; the directed graph that they form will be acyclic.

My first instinct was to delete every element of mid1s and mid2s in
the destructor for ObjectHolder, but that will not work when a Derived
has been inserted into both vectors.

The following is based on the idea that the objects should unregister
themselves upon destruction.

#include <set>
#include <iostream>

class Owner;

class BaseA {

Owner * owner_ptr;

public:

virtual
~BaseA ( void );

void register_owner ( Owner * p );

};

class BaseB {

Owner * owner_ptr;

public:

virtual
~BaseB ( void );

void register_owner ( Owner * p );

};

class Owner {

std::set< BaseA * > all_a;
std::set< BaseB * > all_b;

public:

void register_a ( BaseA * p ) {
all_a.insert( p );
p->register_owner( this );
}

void register_b ( BaseB * p ) {
all_b.insert( p );
p->register_owner( this );
}

void unregister_a ( BaseA * p ) {
all_a.erase( p );
}

void unregister_b ( BaseB * p ) {
all_b.erase( p );
}

~Owner ( void ) {
while ( ! all_a.empty() ) {
std::set< BaseA * >::iterator first = all_a.begin();
BaseA * a_ptr = *first;
all_a.erase( first );
delete a_ptr;
}
while ( ! all_b.empty() ) {
std::set< BaseB * >::iterator first = all_b.begin();
BaseB * b_ptr = *first;
all_b.erase( first );
delete b_ptr;
}
}

};


BaseA::~BaseA ( void ) {
owner_ptr->unregister_a( this );
std::cout << "a";
}

void BaseA::register_owner ( Owner * p ) {
owner_ptr = p;
}


BaseB::~BaseB ( void ) {
owner_ptr->unregister_b( this );
}

void BaseB::register_owner ( Owner * p ) {
owner_ptr = p;
std::cout << "b";
}

class X : public BaseA, public BaseB {

virtual
~X ( void ) {
std::cout << "x";
}

};

int main ( void ) {
Owner o;
for ( unsigned int i = 0; i < 10; ++i ) {
X * x = new X;
o.register_a( x );
o.register_b( x );
BaseA * a = new BaseA;
o.register_a( a );
BaseB * b = new BaseB;
o.register_b( b );
}
}

My next idea was to use
boost::shared_ptr and use that to automatically manage the lifetime of
the object, but there was nothing in the documentation that I saw that
led me to believe that one could use a shared_ptr across a class
heirarchy. Since the 'this' pointer for an object is different
depending on the type it is being used as, I'm (possibly erroneously)
assuming that a shared_ptr will mismanage the reference count for the
pointer since they are using different pointers. Am I wrong in my
understanding of the semantics of boost::shared_ptr?

[snip]

If D is derived from T, then tr1::shared_ptr<T> allows assignment and copy
construction from tr1::shared_ptr<D>. So, if you do:

std::vector< shared_ptr<Mid1> > mid1s;
std::vector< shared_ptr<Mid2> > mid2s;

and

void addMid1 ( shared_ptr<Mid1> p );
void addMid2 ( shared_ptr<Mid2> p );

you can use it as follows:

shared_ptr< Derived > d_ptr = new Derived ( ... );
addMid1( d_ptr );
addMid2( d_ptr );

and it should work just fine. Have you tried?



Best

Kai-Uwe Bux

 

[EnsGabe]

The following is based on the idea that the objects should unregister
themselves upon destruction.

#include <set>
#include <iostream>

class Owner;

class BaseA {

  Owner * owner_ptr;

public:

  virtual
  ~BaseA ( void );

  void register_owner ( Owner * p );

};

class BaseB {

  Owner * owner_ptr;

public:

  virtual
  ~BaseB ( void );

  void register_owner ( Owner * p );

};

class Owner {

  std::set< BaseA * > all_a;
  std::set< BaseB * > all_b;

public:

  void register_a ( BaseA * p ) {
    all_a.insert( p );
    p->register_owner( this );
  }

  void register_b ( BaseB * p ) {
    all_b.insert( p );
    p->register_owner( this );
  }

  void unregister_a ( BaseA * p ) {
    all_a.erase( p );
  }

  void unregister_b ( BaseB * p ) {
    all_b.erase( p );
  }

  ~Owner ( void ) {
    while ( ! all_a.empty() ) {
      std::set< BaseA * >::iterator first = all_a.begin();
      BaseA * a_ptr = *first;
      all_a.erase( first );
      delete a_ptr;
    }
    while ( ! all_b.empty() ) {
      std::set< BaseB * >::iterator first = all_b.begin();
      BaseB * b_ptr = *first;
      all_b.erase( first );
      delete b_ptr;
    }
  }

};

BaseA::~BaseA ( void ) {
  owner_ptr->unregister_a( this );
  std::cout << "a";

}

void BaseA::register_owner ( Owner * p ) {
  owner_ptr = p;

}

BaseB::~BaseB ( void ) {
  owner_ptr->unregister_b( this );

}

void BaseB::register_owner ( Owner * p ) {
  owner_ptr = p;
  std::cout << "b";

}

class X : public BaseA, public BaseB {

  virtual
  ~X ( void ) {
    std::cout << "x";
  }

};

int main ( void ) {
  Owner o;
  for ( unsigned int i = 0; i < 10; ++i ) {
    X * x = new X;
    o.register_a( x );
    o.register_b( x );
    BaseA * a = new BaseA;
    o.register_a( a );
    BaseB * b = new BaseB;
    o.register_b( b );
  }

}

Interesting. This couples the classes together, though. I'd like to
avoid that.

If D is derived from T, then tr1::shared_ptr<T> allows assignment and copy
construction from tr1::shared_ptr<D>. So, if you do:

  std::vector< shared_ptr<Mid1> > mid1s;
  std::vector< shared_ptr<Mid2> > mid2s;

and

  void addMid1 ( shared_ptr<Mid1> p );
  void addMid2 ( shared_ptr<Mid2> p );

you can use it as follows:

  shared_ptr< Derived > d_ptr = new Derived ( ... );
  addMid1( d_ptr );
  addMid2( d_ptr );

and it should work just fine. Have you tried?

Best

Kai-Uwe Bux

I've tried it, and it compiles, but I learned long ago that is a far
cry from saying that something is working. Can you point me in the
direction of documentation for this?

 

[Kai-Uwe Bux]

EnsGabe wrote:

[snip]

I've tried it, and it compiles, but I learned long ago that is a far
cry from saying that something is working. Can you point me in the
direction of documentation for this?

I am looking at the draft n2521 for C++0X. There, you will find what you are
looking for in clause [20.6.6.2.1/19-22]:

shared_ptr(shared_ptr const& r);
template<class Y> shared_ptr(shared_ptr<Y> const& r);
Requires: For the second constructor Y* shall be convertible to T*.
Effects: If r is empty, constructs an empty shared_ptr object; otherwise,
constructs a shared_ptr object that shares ownership with r.
Postconditions: get() == r.get() && use_count() == r.use_count().
Throws: nothing.
...

Note the postconditions.

Assignment is dealt with in [20.6.6.2.3]. Whichever version of shared_ptr<>
you are using, look up the documentation of copy-construction and
assignment.



Best

Kai-Uwe Bux

 

[James Kanze]

Suppose you have a class heirarchy as such:

class Base{
...
};

class Mid1 : public Base{
...
};

class Mid2 : public Base{
...
};

class Derived: public Mid1, Mid2{
};

Mid1, Mid2, and Base are ABC.

Are you sure that you want several instances of Base in the
final object? Or do you want virtual inheritance.

What is an effective way to manage the lifetime of this object?

In what way? How you manage lifetime of an object depends on
the semantics of the object.

Currently, I leak Derived (and anything similar to Derived)
all over the place by not reclaiming them at all.

You mean you never call delete on them. (If the objects don't
have a deterministic lifetime, requiring a explicit call to
delete, then garbage collection is the obvious answer.)

That works for now, in as much as when I am done with them I
am done with execution of my program, but is not a solution
that will work for the future.

My use at the moment is as such:

class ObjectHolder {
public:
...
void addMid1(Mid1 *p) { assert(p); mid1s.push_back(p); };
void addMid2(Mid2 *p) { assert(p); mid2s.push_back(p); };
private:
std::vector<Mid1*> mid1s;
std::vector<Mid2*> mid2s;
...
};

There will be many different classes that are equivalent to
Derived (in the sense that they are deriving from Mid1 and
Mid2.) There will also be classes that derive from one or the
other of Mid1 and Mid2. All of them will be registered
exactly once in mid1s and/or mid2s, whichever they are derived
from. References to the objects will not exist outside of
ObjectHolder. Objects may contain references to other
objects; the directed graph that they form will be acyclic.

In which case, reference counted pointers can be used as a
substitute for garbage collection. It's more invasive, and
usually somewhat slower, but installing boost::shared_ptr is a
lot easier than installing the Boehm collector.

My first instinct was to delete every element of mid1s and
mid2s in the destructor for ObjectHolder, but that will not
work when a Derived has been inserted into both vectors. My
next idea was to use boost::shared_ptr and use that to
automatically manage the lifetime of the object, but there was
nothing in the documentation that I saw that led me to believe
that one could use a shared_ptr across a class heirarchy.

You can, provided that all of the instances of shared_ptr derive
from the same initial shared_ptr. (What you can't do is create
an initial shared_ptr twice from a raw pointer. Basically,
you should probably create the initial shared_ptr as a
shared_ptr< Derived >, at the site of the new, and use nothing
but shared_ptr after that---shared_ptr supports all of the usual
pointer conversions, or it should.)

Since the 'this' pointer for an object is different
depending on the type it is being used as, I'm (possibly erroneously)
assuming that a shared_ptr will mismanage the reference count for the
pointer since they are using different pointers. Am I wrong in my
understanding of the semantics of boost::shared_ptr?

Sort of. Globally, there are two large families of reference
counted pointers: invasive and non-invasive. Since you're
dealing with a known hierarchy, you can easily use either. The
invasive pointers have the advantage that you can create new
smart pointers from the raw pointer anywhere you want; you
could, for example, have the only smart pointers in the two
containers, and that would work. On the other hand, when
multiple inheritance is involved, the base class (which derived
from the RefCntObj, or whatever) *must* be virtual, period.
Non-invasive pointers have the advantage that the pointed to
object doesn't need to be aware that reference counted pointers
are being used---you can even create a shared_ptr<int> (not that
there would ever be any reason to). On the other hand, every
time you create an instance of the smart pointer from a raw
pointer, you get a new counter.

Which is more appropriate in your case depends on the way the
objects are allocated and who manages the containers.