Curiously recurring template pattern

The curiously recurring template pattern (CRTP) is an idiom in C++ in which a class X derives from a class template instantiation using X itself as template argument.^[1] More generally it is known as F-bound polymorphism, and it is a form of F-bounded quantification.

History

The technique was formalized in the 1980s as "F-bounded quantification."^[2] The name "CRTP" was independently coined by Jim Coplien in 1995,^[3] who had observed it in some of the earliest C++ template code as well as in code examples that Timothy Budd created in his multiparadigm language Leda.^[4] It is sometimes called "Upside-Down Inheritance"^[5][6] due to the way it allows class hierarchies to be extended by substituting different base classes.

General form

// The Curiously Recurring Template Pattern (CRTP)
template<class T>
class Base
{
    // methods within Base can use template to access members of Derived
};
class Derived : public Base<Derived>
{
    // ...
};

Some use cases for this pattern are static polymorphism and other metaprogramming techniques such as those described by Andrei Alexandrescu in Modern C++ Design.^[7] It also figures prominently in the C++ implementation of the Data, Context and Interaction paradigm.^[8]

Static polymorphism

Typically, the base class template will take advantage of the fact that member function bodies (definitions) are not instantiated until long after their declarations, and will use members of the derived class within its own member functions, via the use of a cast; e.g.:

template <class T> 
struct Base
{
    void interface()
    {
        // ...
        static_cast<T*>(this)->implementation();
        // ...
    }

    static void static_func()
    {
        // ...
        T::static_sub_func();
        // ...
    }
};

struct Derived : public Base<Derived>
{
    void implementation();
    static void static_sub_func();
};

In the above example, note in particular that the function Base<Derived>::implementation(), though declared before the existence of the struct Derived is known by the compiler (i.e., before Derived is declared), is not actually instantiated by the compiler until it is actually called by some later code which occurs after the declaration of Derived (not shown in the above example), so that at the time the function "implementation" is instantiated, the declaration of Derived::implementation() is known.

This technique achieves a similar effect to the use of virtual functions, without the costs (and some flexibility) of dynamic polymorphism. This particular use of the CRTP has been called "simulated dynamic binding" by some.^[9] This pattern is used extensively in the Windows ATL and WTL libraries.

To elaborate on the above example, consider a base class with no virtual functions. Whenever the base class calls another member function, it will always call its own base class functions. When we derive a class from this base class, we inherit all the member variables and member functions that weren't overridden (no constructors or destructors). If the derived class calls an inherited function which then calls another member function, that function will never call any derived or overridden member functions in the derived class.

However, if base class member functions use CRTP for all member function calls, the overridden functions in the derived class will be selected at compile time. This effectively emulates the virtual function call system at compile time without the costs in size or function call overhead (VTBL structures, and method lookups, multiple-inheritance VTBL machinery) at the disadvantage of not being able to make this choice at runtime.

Object counter

The main purpose of an object counter is retrieving statistics of object creation and destruction for a given class.^[10] This can be easily solved using CRTP:

template <typename T>
struct counter
{
    static int objects_created;
    static int objects_alive;

    counter()
    {
        ++objects_created;
        ++objects_alive;
    }
    
    counter(const counter&)
    {
        ++objects_created;
        ++objects_alive;
    }
protected:
    ~counter() // objects should never be removed through pointers of this type
    {
        --objects_alive;
    }
};
template <typename T> int counter<T>::objects_created( 0 );
template <typename T> int counter<T>::objects_alive( 0 );

class X : counter<X>
{
    // ...
};

class Y : counter<Y>
{
    // ...
};

Each time an object of class X is created, the constructor of counter<X> is called, incrementing both the created and alive count. Each time an object of class X is destroyed, the alive count is decremented. It is important to note that counter<X> and counter<Y> are two separate classes and this is why they will keep separate counts of X's and Y's. In this example of CRTP, this distinction of classes is the only use of the template parameter (T in counter<T>) and the reason why we cannot use a simple un-templated base class.

Polymorphic copy construction

When using polymorphism, one sometimes needs to create copies of objects by the base class pointer. A commonly used idiom for this is adding a virtual clone function that is defined in every derived class. The CRTP can be used to avoid having to duplicate that function or other similar functions in every derived class.

// Base class has a pure virtual function for cloning
class Shape {
public:
    virtual ~Shape() {};
    virtual Shape *clone() const = 0;
};
// This CRTP class implements clone() for Derived
template <typename Derived>
class Shape_CRTP : public Shape {
public:
    virtual Shape *clone() const {
        return new Derived(static_cast<Derived const&>(*this));
    }
};

// Nice macro which ensures correct CRTP usage
#define Derive_Shape_CRTP(Type) class Type: public Shape_CRTP<Type>

// Every derived class inherits from Shape_CRTP instead of Shape
Derive_Shape_CRTP(Square) {};
Derive_Shape_CRTP(Circle) {};

This allows obtaining copies of squares, circles or any other shapes by shapePtr->clone().

Pitfalls

One issue with static polymorphism is that without using a general base class like "Shape" from the above example, derived classes cannot be stored homogeneously as each CRTP base class is a unique type. For this reason, it is more common to inherit from a shared base class with a virtual destructor, like the example above.

In other languages

In Java and the .NET Framework, it is common for a class to extend generic superclasses and interfaces having that same class as a type parameter. Usually it is done to make sure that method parameters and return values are correctly typed as their declaring class. For example, the following Java class implements the generic interface Comparable from the standard library:

public class Item implements Comparable<Item> {
	private String name;
	
	@Override
	public int compareTo(final Item other) {
		return name.compareTo(other.name);
	}
}

This ensures that no attempt is made at compile time to compare an Item to something that is not an Item, such as a String. Unlike C++, a Java class not only cannot extend the same superclass twice with different type parameters (a specific case of the more general lack of multiple inheritance in Java), but it also cannot implement the same interface twice with different type parameters (due to the fact that Java uses type erasure to handle generics). For example, the class Item above cannot implement both Comparable<Item> and Comparable<Object>. See Generics in Java § Problems with type erasure for more information.

See also

• Barton–Nackman trick
• F-bounded quantification

References