在智能指针容器中添加项目

Question

Several ways to add an item in a container of smart pointers. I am wondering which way you will go for.

class MyContainer
{
private:
    std::vector<std::unique_ptr<Item>> mItems;

public:
    bool Add(Item* item);
    // This is Way 1
    //
    // Advantages: 
    // - Easy to add derived items, such as Add(new DerivedItem);
    // - No interface change if smart pointer type changes to such as shared_ptr;
    //
    // Disadvantages:
    // - Don't explicitly show the item to add must be allocated on heap;
    // - If failed to add, user has to delete the item.

    bool Add(std::unique_ptr<Item> item);
    // This is Way 2
    // Disadvantages and advantages are reversed from Way 1.
    // Such as to add derived item, Add(std::unique_ptr<Item>(new DerivedItem));
    //                                                    |
    //                               easy to write DerivedItem here for an error

    bool Add(std::unique_ptr<Item>& item);
    // This is Way 3
    // Similar to Way 2, but when failed to add, item still exist if it is a 
    // reference of outer unique_ptr<Item>

};

I personally go for Way 1. Any more advantages for Way 2 and 3 or disadvantages of Way 1 that I should go for 2 or 3?

sftrabbit gives many good points. In the following common case. How to use Way 2 or 3 to do it easily? User uses a dialog to generate a new derived item. It is put on std::unique_ptr<DerivedItem> item . When click 'OK' button, it is added to the container. If failed to add, go back to the dialog for an edit.

Answer 1

I vote for:

bool Add(std::unique_ptr<Item> item);

Reasons:

1. It's clear from the function signature that the client needs to pass ownership of the object to the MyContainer . If you choose option 1 instead, it's still not clear if the client should delete the object themselves or not, or even if they should be passing a dynamically allocated object.

2. Client is forced to explicitly transfer ownership with std::move if they already have the object managed by a named std::unique_ptr . They won't accidentally lose ownership. Option 3 does not clearly express that it's going to take ownership.

3. When we have std::make_unique (N3588) , the method for adding an element will be:

    container.Add(std::make_unique<Item>()); 

   This avoids using new and improves exception safety in some situations.

4. The issue you've given for derived objects is not really a problem. You'll get a compile-time error if you do it incorrectly.

5. If the interface changes to use a different type of smart pointer, the client will want to know. They don't want to continue passing objects thinking that they're passing ownership if in reality they're sharing it. They'll especially want to know if the opposite happens.

Answer 2

Unfortunately the first way seriously compromises type safety – you've pointed that our yourself in the disadvantages. I think those concerns are overriding any advantages this method may have.

In particular, the potential error with the second method when using a derived object is caught at compile time so it's annoying, but safe !

I agree about your assessment that this usage leaks implementation details but in my experience such leakage is unavoidable – I agree with sfrabbit that this is actually a detail that the user of the class needs to know about.

Answer 3

Consider this as another tool in the toolbox:

bool Add(std::unique_ptr<Item>&& item);

This combines the advantages of Way 2 and Way 3. Ie it will only accept rvalue unique_ptr s (like 2), but if there is some failure in adding it to the container, it can retain ownership like 3. Could be used something like:

void
foo(MyContainer& c)
{
    std::unique_ptr<Item> p = get_Item();
    try
    {
        c.Add(std::move(p));
    }
    catch (...)
    {
        log_error_wtih(*p);
    }
}