Which kind of pointer do I use when?

Shared ownership:
The shared_ptr and weak_ptr the standard adopted are pretty much the same as their Boost counterparts. Use them when you need to share a resource and don't know which one will be the last to be alive. Use weak_ptr to observe the shared resource without influencing its lifetime, not to break cycles. Cycles with shared_ptr shouldn't normally happen - two resources can't own each other.

Note that Boost additionally offers shared_array, which might be a suitable alternative to shared_ptr<std::vector<T> const>.

Next, Boost offers intrusive_ptr, which are a lightweight solution if your resource offers reference-counted management already and you want to adopt it to the RAII principle. This one was not adopted by the standard.

Unique ownership:
Boost also has a scoped_ptr, which is not copyable and for which you can not specify a deleter. std::unique_ptr is boost::scoped_ptr on steroids and should be your default choice when you need a smart pointer. It allows you to specify a deleter in its template arguments and is movable, unlike boost::scoped_ptr. It is also fully usable in STL containers as long as you don't use operations that need copyable types (obviously).

Note again, that Boost has an array version: scoped_array, which the standard unified by requiring std::unique_ptr<T[]> partial specialization that will delete[] the pointer instead of deleteing it (with the default_deleter). std::unique_ptr<T[]> also offers operator[] instead of operator* and operator->.

Note that std::auto_ptr is still in the standard, but it is deprecated. §D.10 [depr.auto.ptr]

The class template auto_ptr is deprecated. [ Note: The class template unique_ptr (20.7.1) provides a better solution. —end note ]

No ownership:
Use dumb pointers (raw pointers) or references for non-owning references to resources and when you know that the resource will outlive the referencing object / scope. Prefer references and use raw pointers when you need either nullability or resettability.

If you want a non-owning reference to a resource, but you don't know if the resource will outlive the object that references it, pack the resource in a shared_ptr and use a weak_ptr - you can test if the parent shared_ptr is alive with lock, which will return a shared_ptr that is non-null if the resource still exists. If want to test whether the resource is dead, use expired. The two may sound similar, but are very different in the face of concurrent execution, as expired only guarantees its return value for that single statement. A seemingly innocent test like

if(!wptr.expired())
  something_assuming_the_resource_is_still_alive();

is a potential race condition.

Deciding what smart pointer to use is a question of ownership. When it comes to resource management, object A owns object B if it is in control of the lifetime of object B. For example, member variables are owned by their respective objects because the lifetime of member variables is tied to the lifetime of the object. You choose smart pointers based on how the object is owned.

Note that ownership in a software system is separate from ownership as we would think of it outside of software. For example, a person might "own" their home, but that doesn't necessarily mean that a Person object has control over the lifetime of a House object. Conflating these real world concepts with software concepts is a sure-fire way to program yourself into a hole.

If you have sole ownership of the object, use std::unique_ptr<T>.

If you have shared ownership of the object...
- If there are no cycles in ownership, use std::shared_ptr<T>.
- If there are cycles, define a "direction" and use std::shared_ptr<T> in one direction and std::weak_ptr<T> in the other.

If the object owns you, but there is potential of having no owner, use normal pointers T* (e.g. parent pointers).

If the object owns you (or otherwise has guaranteed existence), use references T&.

Caveat: Be aware of the costs of smart pointers. In memory or performance limited environments, it could be beneficial to just use normal pointers with a more manual scheme for managing memory.

The costs:

• If you have a custom deleter (e.g. you use allocation pools) then this will incur overhead per pointer that may be easily avoided by manual deletion.
• std::shared_ptr has the overhead of a reference count increment on copy, plus a decrement on destruction followed by a 0-count check with deletion of the held object. Depending on the implementation, this can bloat your code and cause performance issues.
• Compile time. As with all templates, smart pointers contribute negatively to compile times.

Examples:

struct BinaryTree
{
    Tree* m_parent;
    std::unique_ptr<BinaryTree> m_children[2]; // or use std::array...
};

A binary tree does not own its parent, but the existence of a tree implies the existence of its parent (or nullptr for root), so that uses a normal pointer. A binary tree (with value semantics) has sole ownership of its children, so those are std::unique_ptr.

struct ListNode
{
    std::shared_ptr<ListNode> m_next;
    std::weak_ptr<ListNode> m_prev;
};

Here, the list node owns its next and previous lists, so we define a direction and use shared_ptr for next and weak_ptr for prev to break the cycle.