What it means when you convert between different shared_ptrs

Raymond Chen

The C++ shared_ptr manages a reference-counted pointer. Usually, it’s a pointer to an object that will be delete‘d when the last reference expires. But it doesn’t have to be that way.

Recall that a shared_ptr is really two pointers.

• A pointer to a control block which manages the shared and weak reference counts, and which destroys an object (commonly known as the managed object) when the shared reference count drops to zero.
• A pointer to return from the get() method, commonly know as the stored pointer.

Most of the time, the stored pointer points to the managed object, because that what you get when you construct a shared_ptrs from a raw pointer or when you call make_shared. But what is the use case for a shared_ptr where the managed object and the stored pointer are different?

You may want to have a shared_ptr whose get() returns a pointer to a sub-object of another large object. In that case, the managed object is the larger object, and the stored pointer is to the sub-object.

struct Sample
{
    int value1;
    int value2;
};

void consume(std::shared_ptr<int> pint);

std::shared_ptr<Sample> p = std::make_shared<Sample>();
consume(std::shared_ptr<int>(p, &p->value1));

// Or, more tersely
auto p = std::make_shared<Sample>();
consume({ p, &p->value1 });

In the above example, we have a class Sample with two members. We create a shared_ptr to that class and save it in p. But say there’s another function that wants a shared_ptr<int>. No problem, we can convert the std::shared_ptr<Sample> into a std::shared_ptr<int> by reusing the control block (first parameter p) and substituting a new stored pointer (second parameter &p->value1). The consume function can use the shared_ptr<int> to access the value1 member, and the control block of that shared_ptr<int> prevents the Sample from being destroyed, which in turn prevents the value1 from being destroyed.

The general principle is that the lifetime of the stored pointer should be contained with the lifetime of the managed object. It can be a direct containment relationship, like we did with value1, or it could be a more complex chain of lifetime dependencies.

struct Other
{
    int value;
};

struct Sample2
{
    const std::unique_ptr<Other> m_other =
        std::make_unique<Other>();
};

auto p = std::make_shared<Sample2>();
consume({ p, &p->m_other->value });

In this second example, the stored pointer of the shared_ptr<int> we pass to the consume() function points to the value member inside the Other object to which the Sample2 holds a unique pointer. The control block in that shared_ptr<int> controls the lifetime of the Sample2 object, which is acceptable because as long as the Sample2 object remains alive, the value inside the Other will be alive.

Now, the compiler doesn’t check that you have a positive chain of lifetime control from the managed object to the stored pointer. You could do something silly like

struct Sample3
{
    std::unique_ptr<Other> m_other =
        std::make_unique<Other>();
};

auto p = std::make_shared<Sample3>();
consume({ p, &p->m_other->value });
p->m_other = nullptr; // oops, chain is broken

and the shared_ptr<int> will think it’s keeping the value alive, even though you broke the link from the Sample3 to the Other.

Or you can do something even sillier like

int unrelated;
consume({ p, &unrelated });

and the shared_ptr<int> will access unrelated even though its lifetime is unrelated to the Sample2. If unrelated is destroyed, the shared_ptr<int> will have a dangling stored pointer.

These shared_ptr objects in which the managed object is different from the pointed-to object are commonly known as aliasing shared pointers.

Okay, so I showed one way of creating an aliasing shared pointer, namely by constructing a shared_ptr from an existing shared_ptr (which shares the managed object) and providing a different stored pointer. If the new stored pointer points to a base class of the original, then the shared_ptr has a conversion operator that creates an aliasing shared pointer to the base-class subobject.

struct Base
{
};

struct Derived : Base
{
};

std::shared_ptr<Derived> p = std::make_shared<Derived>();
std::shared_ptr<Base> b = p; // auto-conversion
// equivalent to
std::shared_ptr<Base> b(p, p.get());

If you want to do the reverse conversion (from Base to Derived), you can write it out explicitly:

std::shared_ptr<Derived> b(p, static_cast<Derived*>(p.get()));

Of course, this requires that the stored Base pointer really is a pointer to the Base part of a larger Derived object.

The C++ language comes with some helper functions that construct a shared_ptr by casting the stored pointer of another shared_ptr.

Everything looks great until we get to dynamic_pointer_cast, which is not equivalent to a one-liner that uses dynamic_cast!

The reason is that, unlike the other casts, dynamic_cast can change a non-null pointer to a null pointer, which happens if the runtime type does not match. In that case, the dynamic_pointer_case returns an empty shared pointer (rather than a shared pointer with a control block and no stored pointer), because there is nothing whose lifetime needs to be extended.

Now we can finish that table:

This wrinkle about control blocks for null pointers does call out that two boxes in the shared pointer diagram are technically legal though strange.

So far, we’ve been dealing with empty shared pointers (that manage no object and have no stored pointer) and full shared pointers (that manage an object and have a stored pointer). But there are two other boxes, which I’ve named “Phantom” and “Indulgent”. We’ll look at those two weird guys next time.

Raymond Chen

Follow