Template (C++)

There are three kinds of templates: function templates, class templates and, since C++14, variable templates. Since C++11, templates may be either variadic or non-variadic; in earlier versions of C++ they are always non-variadic.

C++ Templates are Turing complete.^[2]

A function template behaves like a function except that the template can accept arguments of various types, enabling type-generic behavior (see example). In other words, a function template represents a family of functions. The format for declaring function templates with type parameters is:

template <class Identifier> Declaration;
template <typename Identifier> Declaration;

Both expressions have the same meaning and behave in exactly the same way. The latter form was introduced to avoid confusion,^[3] since a type parameter need not be a class until C++20. (It can be a basic type such as int or double.)

For example, the C++ Standard Library contains the function template max(x, y) which returns the larger of x and y. That function template could be defined like this^[4]:

template <typename T>
[[nodiscard]]
constexpr T& max(const T& a, const T& b) noexcept {
    return a < b ? b : a;
}

This single function definition works with many data types. Specifically, it works with all data types for which < (the less-than operator) is defined and returns a value with a type convertible to bool. The usage of a function template saves space in the source code file in addition to limiting changes to one function description and making the code easier to read.

An instantiated function template usually produces the same object code, though, compared to writing separate functions for all the different data types used in a specific program. For example, if a program uses both an int and a double version of the max() function template above, the compiler will create an object code version of max() that operates on int arguments and another object code version that operates on double arguments.^{[citation needed]} The compiler output will be identical to what would have been produced if the source code had contained two separate non-templated versions of max(), one written to handle int and one written to handle double.

Here is how the function template could be used:

import std;

int main() {
    // This will call max<int> by implicit argument deduction.
    std::println("{}", std::max(3, 7));

    // This will call max<double> by implicit argument deduction.
    std::println("{}", std::max(3.0, 7.0));

    // We need to explicitly specify the type of the arguments; 
    // although std::type_identity could solve this problem...
    std::println("{}", max<double>(3, 7.0));
}

In the first two cases, the template argument T is automatically deduced by the compiler to be int and double, respectively. In the third case automatic deduction of max(3, 7.0) would fail because the type of the parameters must in general match the template arguments exactly. Therefore, we explicitly instantiate the double version with max<double>().

This function template can be instantiated with any copy-constructible type for which the expression y < x is valid. For user-defined types, this implies that the less-than operator (<) must be overloaded in the type.

Since C++20, using auto or concept auto in any of the parameters of a function declaration, that declaration becomes an abbreviated function template declaration.^[5] Such a declaration declares a function template and one invented template parameter for each placeholder is appended to the template parameter list:

void f1(auto); // same as template <class T> void f1(T)
void f2(C1 auto); // same as template <C1 T> void f2(T), if C1 is a concept
void f3(C2 auto...); // same as template <C2... Ts> void f3(Ts...), if C2 is a concept
void f4(C2 auto, ...); // same as template <C2 T> void f4(T...), if C2 is a concept
void f5(const C3 auto*, C4 auto&); // same as template <C3 T, C4 U> void f5(const T*, U&);

Constraining the max() using concepts could look something like this:

using std::totally_ordered;

// in typename declaration:
template <totally_ordered T>
[[nodiscard]]
constexpr T max(T x, T y) noexcept {
    return x < y ? y : x;
}

// in requires clause:
template <typename T>
    requires totally_ordered<T>
[[nodiscard]]
constexpr T max(T x, T y) noexcept {
    return x < y ? y : x;
}

A class template provides a specification for generating classes based on parameters. Class templates are generally used to implement containers. A class template is instantiated by passing a given set of types to it as template arguments.^[6] The C++ Standard Library contains many class templates, in particular the containers adapted from the Standard Template Library, such as vector.

In C++14, templates can be also used for variables, as in the following example:

template <typename T> 
constexpr T PI = T{3.141592653589793238462643383L}; // (Almost) from std::numbers::pi

Although templating on types, as in the examples above, is the most common form of templating in C++, it is also possible to template on values. Thus, for example, a class declared with

template <int K>
class MyClass;

can be instantiated with a specific int.

As a real-world example, the standard library fixed-size array type std::array is templated on both a type (representing the type of object that the array holds) and a number which is of type std::size_t (representing the number of elements the array holds). To create a class Array equivalent to std::array, it can be declared as follows:

template <class T, size_t N> 
struct Array;

and an array of six chars might be declared:

Array<char, 6> myArray;

When a function or class is instantiated from a template, a specialization of that template is created by the compiler for the set of arguments used, and the specialization is referred to as being a generated specialization.

Sometimes, the programmer may decide to implement a special version of a function (or class) for a given set of template type arguments which is called an explicit specialization. In this way certain template types can have a specialized implementation that is optimized for the type or a more meaningful implementation than the generic implementation.

• If a class template is specialized by a subset of its parameters it is called partial template specialization (function templates cannot be partially specialized).
• If all of the parameters are specialized it is a full specialization.

Explicit specialization is used when the behavior of a function or class for particular choices of the template parameters must deviate from the generic behavior: that is, from the code generated by the main template, or templates. For example, the template definition below defines a specific implementation of max() for arguments of type const char*:

import std;

template <> 
[[nodiscard]]
constexpr const char* max(const char* a, const char* b) noexcept {
    // Normally, the result of a direct comparison
    // between two C strings is undefined behaviour;
    // using std::strcmp makes defined.
    return std::strcmp(a, b) > 0 ? a : b;
}

C++11 introduced variadic templates, which can take a variable number of arguments in a manner somewhat similar to variadic functions such as std::printf.

using std::format_string;
using std::ofstream;

enum class Level { ... };

ofstream logFile{"logfile.txt"};

template <typename... Args>
void log(const format_string<Args...>& fmt, Args&&... args) {
    std::println(logFile, fmt, args...);
}

Because only C-style variadic parameters are supported in C++, the only way to get type-safe variadic functions (like in Java is through variadic templates.

C++11 introduced template aliases, which act like parameterized typedefs.

The following code shows renaming std::map to TreeMap and std::unordered_map to HashMap, as well as creating an alias StringHashMap for std::unordered_map<K, std::string>. This allows, for example, StringHashMap<int> to be used as shorthand for std::unordered_map<int, std::string>.

using String = std::string;

// allowing optional specialization of hash functions, allocators, etc.
template <
    typename K, 
    typename V, 
    typename Compare = std::less<K>, 
    typename Alloc = std::allocator<std::pair<const K, T>>
>
using TreeMap = std::map<K, V, Compare, Alloc>;

template <
    typename K, 
    typename V, 
    typename HashFn = std::hash<K>, 
    typename KeyEq = std::equal_to<K>, 
    typename Alloc = std::allocator<std::pair<const K, T>>
> 
using HashMap = std::unordered_map<K, V, HashFn, KeyEq, Alloc>;

// or, only allowing K and V to be specialized:
template <typename K, typename V>
using TreeMap = std::map<K, V>;

template <typename K, typename V> 
using HashMap = std::unordered_map<K, V>;

// Defining StringHashMap<K> = HashMap<K, String>
template <typename K>
using StringHashMap = HashMap<K, String>;

StringHashMap<int> myMap = /* something here... */;

Since C++20, templates can be constrained similarly to generics wildcards in Java or C# and Rust where clauses. This is done using concepts, which represent a set of boolean predicates evaluated at compile time.

For example, this code defines a concept representing an upper bound on inheritance. A class satisfies this concept if it inherits from Player, and classes that do not cannot be used as the template parameter in processListOfPlayers().

import std;

using std::is_base_of_v;
using std::vector;

class Player {
    // ...
};

template <typename T>
concept ExtendsPlayer = is_base_of_v<Player, T>;

// T is required to be a type whose inheritance upper bound is Player,
// blocking any type that does not inherit from Player
template <ExtendsPlayer T>
void processListOfPlayers(vector<T> players) {
    // ...
}

In C++03, "exported templates" were added to C++.^[7] These were later removed in C++11, due to very few compilers actually supporting the feature.^[8] The only compiler known to support exported templates was Comeau C/C++. Among the cited reasons for removal were:

• Expensive or complex to implement
• Little benefit to most users as well as little interest
• Difficult to use
• Changes to meanings of existing language features
• Restricting the future development of C++

An "exported template" is essentially a class template whose static data members and non-inline methods are exported. It must be marked by the keyword export. What distinguishes an "exported template" is the fact that it does not need to be defined in a translation unit that uses the template.^[9][10] For example (in C++03):

File1.cpp:

#include <iostream>

static void trace() {
    std::cout << "File 1" << std::endl;
}

export template <typename T>
T min(const T& x, const T& y);

int main() {
    trace();
    std::cout << min(2, 3) << std::endl;
}

File2.cpp:

#include <iostream>

static void trace() {
    std::cout << "File 2" << std::endl;
}

export template <typename T>
T min(const T& x, const T& y) {
    trace();
    return a < b ? a : b;
}

With the introduction of modules in C++20, the keyword export was re-added to C++. This re-allowed declarations like this:

import std;

using std::is_base_of_v;

export class Atom {
    // ...
};

export template <typename T>
concept ExtendsAtom = is_base_of_v<Atom, T>;

export template <ExtendsAtom Instance, typename... Bases>
class Cluster: public Bases... {
private:
    Instance x;
public:
    explicit Cluster(Instance x, Bases&&... bases):
        Bases(bases)..., x{x} {}

    // ...
};

The compilation speed benefits intended to be offered by exported templates are offered by modules anyway, making the feature essentially obsolete and superseded by modules.

Initially, the concept of templates was not included in some languages, such as Java and C# 1.0. Java's adoption of generics mimics the behavior of templates, but is technically different. C# added generics (parameterized types) in .NET 2.0. The generics in Ada predate C++ templates.

Although C++ templates, Java generics, and .NET generics are often considered similar, generics only mimic the basic behavior of C++ templates.^[11] Some of the advanced template features utilized by libraries such as Boost and STLSoft, and implementations of the STL, for template metaprogramming (explicit or partial specialization, default template arguments, template non-type arguments, template template arguments, ...) are unavailable with generics.

In C++ templates, compile-time cases were historically performed by pattern matching over the template arguments. For example, the template base class in the Factorial example below is implemented by matching 0 rather than with an inequality test, which was previously unavailable. However, the arrival in C++11 of standard library features such as std::conditional has provided another, more flexible way to handle conditional template instantiation.

// Induction
template <unsigned int N> 
struct Factorial {
    static constexpr unsigned int value = N * Factorial<N - 1>::value;
};

// Base case via template specialization:
template <> 
struct Factorial<0> {
    static constexpr unsigned int value = 1;
};

With these definitions, one can compute, say 6! at compile time using the expression Factorial<6>::value.

Alternatively, constexpr in C++11 / if constexpr in C++17 can be used to calculate such values directly using a function at compile-time:

template <unsigned int N>
[[nodiscard]]
constexpr unsigned int factorial() noexcept {
    if constexpr (N <= 1) {
        return 1;
    } else {
        return N * factorial<N - 1>();
    }
}

Because of this, template meta-programming is now mostly used to do operations on types.

• Template metaprogramming
• Metaprogramming
• Monomorphization
• Generic programming
• Header-only
• Substitution failure is not an error
• Curiously recurring template pattern
• List of C++ template libraries