Modern C++
Programming
9. Templates and
Meta-programming I
Function Templates and Compile-Time Utilities
Federico Busato
2024-03-29
Template Books
C++ Templates: The
Complete Guide (2nd)
D. Vandevoorde, N. M. Josuttis,
D. Gregor, 2017
4/48
Template Overview
Template
A template is a mechanism for generic programming to provide a “schema” (or
placeholders) to represent the structure of an entity
In C++, templates are a compile-time functionality to represent:
• A family of functions
• A family of classes
• A family of variables C++14
5/48
Function Template 1/2
The problem: We want to define a function to handle different types
int add(int a, int b) {
return a + b;
}
float add(float a, float b) { // overloading
return a + b;
}
char add(char a, char b) { ... } // overloading
ClassX add(ClassX a, ClassX b) { ... } // overloading
• Redundant code!!
• How many functions we have to write!?
• If the user introduces a new type we have to write another function!!
6/48
Function Template 2/2
Function Template
A function template is a function schema that operates with generic types
(independent of any particular type) or concrete values
A function template works with multiple types without repeating the entire code for
each of them
template<typename T> // or template<class T>
T add(T a, T b) {
return a + b;
}
int c1 = add(3, 4); // c1 = 7
float c2 = add(3.0f, 4.0f); // c2 = 7.0f
7/48
Templates: Benefits and Drawbacks
Benefits
• Generic Programming: Less code and reusable. Reduce redundancy, better
maintainability and flexibility
• Performance. Computation can be done/optimized at compile-time → faster
Drawbacks
• Readability. “With respect to C++, the syntax and idioms of templates are
esoteric compared to conventional C++ programming, and templates can be very
difficult to understand” [wikipedia] → hard to read, cryptic error messages
• Compile Time/Binary Size. Templates are implicitly instantiated for every
distinct parameters
8/48
Template Instantiation
Template Instantiation
The template instantiation is the substitution of template parameters with concrete
values or types
The compiler automatically generates a function implementation for each template
instantiation
template<typename T>
T add(T a, T b) {
return a + b;
}
add(3, 4); // generates: int add(int, int)
add(3.0f, 4.0f); // generates: float add(float, float)
add(2, 6); // already generated
// other instances are not generated
// e.g. char add(char,char)
9/48
Implicit and Explicit Template Instantiation
Implicit Template Instantiation
Implicit template instantiation occurs when the compiler generates code
depending on the deduced argument types or the explicit template arguments and
only when the definition is needed
Explicit Template Instantiation
Explicit template instantiation occurs when the compiler generates code
depending only on the explicit template arguments specified in the declaration.
Useful when dealing with multiple translation units to reduce the binary size
10/48
Implicit and Explicit Template Instantiation
template<typename T>
void f(T a) {}
void g() {
f(3); // generates: void f(int) → implicit
f<short>(3.0); // generates: void f(short) → implicit
}
template void f<int>(int); // generates: void f(int) → explicit
11/48
Template Parameters
Template Parameters
Template Parameters are the names following the template keyword
template<typename T>
void f() {}
f<int>();
typename T is the template parameter
int is the template argument
A template parameter can be a generic type, i.e. typename , as well as a
non-type template parameters (NTTP), e.g. int , enum , etc.
The template argument of a generic type is a built-in or user-declared type, while a
concrete value for a non-type template parameter
12/48
Examples 1/2
int parameter
template<int A, int B>
int add_int() {
return A + B; // sum is computed at compile-time
} // e.g. add_int<3, 4>();
enum parameter
enum class Enum { Left, Right };
template<Enum Z>
int add_enum(int a, int b) {
return (Z == Enum::Left) ? a + b : a;
} // e.g. add_enum<Enum::Left>(3, 4);
13/48
Examples 2/2
• Ceiling division
template<int DIV, typename T>
T ceil_div(T value) {
return (value + DIV - 1) / DIV;
}
// e.g. ceil_div<5>(11); // returns 3
• Rounded division
template<int DIV, typename T>
T round_div(T value) {
return (value + DIV / 2) / DIV;
}
// e.g. round_div<5>(11); // returns 2 (2.2)
Since DIV is known at compile-time, the compiler can heavily optimize the division
(almost for every number, not just for power of two)
14/48
Template Parameter - Default Value 1/3
C++11 Template parameters can have default values
template<int A = 3, int B = 4>
void print1() { cout << A << ", " << B; }
template<int A = 3, int B> // still possible, but little sense
void print2() { cout << A << ", " << B; }
print1<2, 5>(); // print 2, 5
print1<2>(); // print 2, 4 (B: default)
print1<>(); // print 3, 4 (A,B: default)
print1(); // print 3, 4 (A,B: default)
print2<2, 5>(); // print 2, 5
// print2<2>(); compile error
// print2<>(); compile error
// print2(); compile error
15/48
Template Parameters - Default Value 2/3
Template parameters may have no name
void f() {}
template<typename = void>
void g() {}
int main() {
g(); // generated
}
f() is always generated in the final code
g() is generated in the final code only if it is called
16/48
Template Parameters - Default Value 3/3
C++11 Unlike function parameters, template parameters can be initialized by
previous values
template<int A, int B = A + 3>
void f() {
cout << B;
}
template<typename T, int S = sizeof(T)>
void g(T) {
cout << S;
}
f<3>(); // B is 6
g(3); // S is 4
17/48
Function Template Overloading
Template Functions can be overloaded
template<typename T>
T add(T a, T b) {
return a + b;
} // e.g add(3, 4);
template<typename T>
T add(T a, T b, T c) { // different number of parameters
return a + b + c;
} // e.g add(3, 4, 5);
Also, templates themselves can be overloaded
template<int C, typename T>
T add(T a, T b) { // it is not in conflict with
return a + b + C; // T add(T a, T b)
} // "C" is part of the signature
18/48
Template Specialization 1/2
Template Specialization
Template specialization refers to the concrete implementation for a specific
combination of template parameters
The problem:
template<typename T>
bool compare(T a, T b) {
return a < b;
}
The direct comparison between two floating-point values is dangerous due to rounding
errors
19/48
Template Specialization 2/2
Solution: Template specialization
template<>
bool compare<float>(float a, float b) {
return ... // a better floating point implementation
}
Full Specialization: Function templates can be specialized only if ALL template
arguments are specialized
20/48
Template Variable
C++14 allows variables with templates
A template variable can be considered a special case of a class template (see next
lecture)
template<typename T>
constexpr T pi{ 3.1415926535897932385 }; // variable template
template<typename T>
T circular_area(T r) {
return pi<T> * r * r; // pi<T> is a variable template instantiation
}
circular_area(3.3f); // float
circular_area(3.3); // double
// circular_area(3); // compile error, narrowing conversion with "pi"
21/48
Template Parameter Types
Template parameters can be:
• integral type
•
enum , enum class
• floating-point type C++20
• auto placeholder C++17
• class literals and concepts C++20
• generic type typename
and rarely:
• function
• reference/pointer to global static function or object
• pointer to member type
• nullptr t C++14
22/48
Generic Type Notes
Pass multiple values and floating-point types
template<float V> // only in C++20
void print_float() {}
template<typename T>
void print() {
cout << T::x << ", " << T::y;
}
struct Multi {
static const int x = 1;
static constexpr float y = 2.0f;
};
print<Multi>(); // print "1, 2"
23/48
auto Placeholder
C++17 introduces automatic deduction of non-type template parameters with the
auto keyword
template<int X, int Y>
void f() {}
template<typename T1, T1 X, typename T2, T2 Y>
void g1() {} // before C++17
template<auto X, auto Y>
void g2() {}
f<2u, 2u>(); // X: int, Y: int
g1<int, 2, char, 'a'>(); // X: int, Y: char
g2<2, 'a'>(); // X: int, Y: char
24/48
Class Template Parameter Type
C++20 A non-type template parameter of a class literal type:
• A class literal is a class that can be assigned to constexpr variable
• All base classes and non-static data members are public and non-mutable
• All base classes and non-static data members have the same properties
# include <array>
struct A {
int x;
constexpr A(int x1) : x{x1} {}
};
template<A a>
void f() { std::cout << a.x; }
template<std::array array>
void g() { std::cout << array[2]; }
f<A{5}>(); // print '5'
g<std::array{1,2,3}>(); // print '3'
25/48
Template Parameter - Array and Pointer Types
⋆
1/2
Array and pointer
template<int* ptr> // pointer
void g() {
cout << ptr[0];
}
template<int (&array)[3]> // reference
void f() {
cout << array[0];
}
int array[] = {2, 3, 4}; // global
int main() {
f<array>(); // print 2
g<array>(); // print 2
}
Class member
struct A {
int x = 5;
int y[3] = {4, 2, 3};
};
template<int A::*x> // pointer to
void h1() {} // member type
template<int (A::*y)[3]> // pointer to
void h2() {} // member type
int main() {
h1<&A::x>();
h2<&A::y>();
}
26/48
Template Parameter - Function Type
⋆
2/2
Function
template<int (*F)(int, int)> // <-- signature of "f"
int apply1(int a, int b) {
return F(a, b);
}
int f(int a, int b) { return a + b; }
int g(int a, int b) { return a * b; }
template<decltype(f) F> // alternative syntax
int apply2(int a, int b) {
return F(a, b);
}
int main() {
apply1<f>(2, 3); // return 5
apply2<g>(2, 3); // return 6
}
27/48
static assert
C++11 static assert is used to test an assertion at compile-time, e.g.
sizeof , literals, templates, constexpr
If the static assertion fails, the program does not compile
static_assert(2 + 2 == 4, "test1"); // ok, it compiles
static_assert(2 + 2 == 5, "test2"); // compile error, print "test2"
C++17: assertions without messages
template<typename T, typename R>
void f() { static_assert(sizeof(T) == sizeof(R)); }
f<int, unsigned>(); // ok, it compiles
// f<int, char>(); // compile error
C++26: assertions with text formatting
static_assert(sizeof(T) != 4, std::format("test1 with sizeof(T)={}", sizeof(T)));
28/48
using Keyword 1/2
using keyword (C++11)
The using keyword introduces an alias-declaration or alias-template
• using is an enhanced version of typedef with a more readable syntax
• using can be combined with templates, as opposite to typedef
• using is useful to simplify complex template expression
• using allows introducing new names for partial and full specializations
typedef int distance_t; // equal to:
using distance_t = int;
typedef void (*function)(int, float); // equal to:
using function = void (*)(int, float);
29/48
using Keyword 2/2
Full/Partial specialization alias:
template<typename T, int Size>
struct Vector {}; // see next lecture for further details
// on class template
template<int Size>
using Bitset = Vector<bool, Size>; // partial specialization alias
using IntV4 = Vector<int, 4>; // full specialization alias
Accessing a type within a structure:
struct A {
using type = int;
};
using Alias = A::type;
30/48
decltype Keyword (value) 1/3
C++11 decltype keyword captures the type of entity or an expression
• decltype never executes, it is always evaluated at compile-type
int x = 3;
int& y = x;
const int z = 4;
int array[2];
void f(int, float);
decltype(x) d1; // int
decltype(2 + 3.0) d2; // double
decltype(y) d3; // int&
decltype(z) d4; // const int
decltype(array) d5; // int[2]
decltype(f(1, 2.0f)) d6; // void
using function = decltype(f);
31/48
decltype Keyword ((expression))
⋆
2/3
bool f(int) { return true; }
struct A {
int x;
};
int x = 3;
const A a{4};
decltype(x) d1; // int
decltype((x)) d2 = x; // int&
decltype(f) d3; // bool (int)
decltype((f)) d4 = f; // bool (&)(int)
decltype(a.x) d5; // int
decltype((a.x)) d6 = x; // const int
www.ibm.com/support/knowledgecenter
32/48
decltype Keyword + Function templates 3/3
C++11
template<typename T, typename R>
decltype(T{} + R{}) add(T x, R y) {
return x + y;
}
unsigned v1 = add(1, 2u);
double v2 = add(1.5, 2u);
C++14
template<typename T, typename R>
auto add(T x, R y) {
return x + y;
}
33/48
Type Traits 1/4
Introspection
Introspection is the ability to inspect a type and query its properties
Reflection
Reflection is the ability of a computer program to examine, introspect, and modify
its own structure and behavior
C++ provides compile-time reflection and introspection capabilities through
type traits
34/48
Type Traits 2/4
Type traits (C++11)
Type traits define a compile-time interface to query or modify the properties of
types
The problem:
template<typename T>
T integral_div(T a, T b) {
return a / b;
}
integral_div(7, 2); // returns 3 (int)
integral_div(7l, 2l); // returns 3 (long int)
integral_div(7.0, 3.0); // !!! a floating-point value is not an integral type
Two alternatives: (1) Specialize (2) Type Traits + static assert
35/48
Type Traits 3/4
If we want to prevent floating-point/other objects division at compile-time, a
first solution consists in specialize for all integral types
template<typename T>
T integral_div(T a, T b); // declaration (error for other types)
template<>
char integral_div<char>(char a, char b) { // specialization
return a / b;
}
template<>
int integral_div<int>(int a, int b) { // specialization
return a / b;
}
...unsigned char
...short
...
Very redundant!!
36/48
Type Traits 4/4
The best solution is to use type traits
# include <type_traits> // <-- std type traits library
template<typename T>
T integral_div(T a, T b) {
static_assert(std::is_integral<T>::value,
"integral_div accepts only integral types");
return a / b;
}
std::is integral<T> is a struct with a static constexpr boolean field value
value is true if T is bool , char , short , int , long , long long , false otherwise
C++17 provides utilities to improve the readability of type traits
std::is_integral_v<T>; // std::is_integral<T>::value
37/48
Type Traits Library - Query Fundamental and Scalar Types 1/3
• is integral checks for an integral type ( bool , char , unsigned char ,
short , int , long , etc.)
• is floating point checks for a floating-point type ( float , double )
• is arithmetic checks for a integral or floating-point type
• is signed checks for a signed type ( float , int , etc.)
• is unsigned checks for an unsigned type ( unsigned , bool , etc.)
• is enum checks for an enumerator type ( enum , enum class )
• is void checks for ( void )
• is pointer checks for a pointer ( T* )
•
is null pointer checks for a ( nullptr ) C++14
38/48
Type Traits Library - Query References, Functions, Objects 2/3
Entity type queries:
• is reference checks for a reference ( T& )
• is array checks for an array ( T (&)[N] )
• is function checks for a function type
Class queries:
• is class checks for a class type ( struct , class )
• is abstract checks for a class with at least one pure virtual function
• is polymorphic checks for a class with at least one virtual function
39/48
Type Traits Library - Query Type Relation 3/3
Type property queries:
• is const checks if a type is const
Type relation:
• is same<T, R> checks if T and R are the same type
• is base of<T, R> checks if T is base of R
• is convertible<T, R> checks if T can be converted to R
Full list: en.cppreference.com/w/cpp/header/type_traits
40/48
Example - const Deduction
# include <type_traits>
template<typename T>
void f(T x) { cout << std::is_const_v<T>; }
template<typename T>
void g(T& x) { cout << std::is_const_v<T>; }
template<typename T>
void h(T& x) {
cout << std::is_const_v<T>;
x = nullptr; // ok, it compiles for T: (const int)*
}
const int a = 3;
f(a); // print false, "const" drop in pass by-value
g(a); // print true
const int* b = new int;
h(b); // print false!! T: (const int)*
41/48
Example - Type Relation
# include <type_traits>
template<typename T, typename R>
T add(T a, R b) {
static_assert(std::is_same_v<T, R>, "T and R must have the same type");
return a + b;
}
add(1, 2); // ok
// add(1, 2.0); // compile error, "T and R must have the same type"
# include <type_traits>
struct A {};
struct B : A {};
std::is_base_of_v<A, B>; // true
std::is_convertible_v<int, float>; // true
42/48
Type Manipulation
Type traits allow also to manipulate types by using the type field
Example: produce unsigned from int
# include <type_traits>
using U = typename std::make unsigned<int>::type; // see next lecture to understand
// why 'typename' is needed here
U y = 5; // unsigned
C++14 provides utilities to improve the readability of type traits
std::make_unsigned_t<T>; // instead of 'typename std::make_unsigned<T>::type'
43/48
Type Traits Library - Type Manipulation 1/2
Signed and Unsigned types:
• make signed makes a signed type
• make unsigned makes an unsigned type
Pointers and References:
• remove pointer remove pointer ( T* → T )
• remove reference remove reference ( T& → T )
• add pointer add pointer ( T → T* )
• add lvalue reference add reference ( T → T& )
44/48
Type Traits Library - Type Transformation 2/2
const specifiers:
• remove const remove const ( const T → T )
• add const add const
Other type transformation:
• common type<T, R> returns the common type between T and R
• conditional<pred, T, R> returns T if pred is true , R otherwise
• decay<T> returns the same type as a function parameter passed by-value
45/48
Type Manipulation Example
# include <type_traits>
template<typename T>
void f(T ptr) {
using R = std::remove_pointer_t<T>;
R x = ptr[0]; // char
}
template<typename T>
void g(T x) {
using R = std::add_const_t<T>;
R y = 3;
// y = 4; // compile error
}
char a[] = "abc";
f(a); // T: char*
g(3); // T: int
46/48
std::common type Example
# include <type_traits>
template<typename T, typename R>
std::common_type_t<R, T> // <-- return type
add(T a, R b) {
return a + b;
}
// we can also use decltype to derive the result type
using result_t = decltype(add(3, 4.0f));
result_t x = add(3, 4.0f);
47/48
std::conditional Example
# include <type_traits>
template<typename T, typename R>
auto f(T a, R b) {
constexpr bool pred = sizeof(T) > sizeof(R);
using S = std::conditional_t<pred, T, R>;
return static_cast<S>(a) + static_cast<S>(b);
}
f( 2, 'a'); // return 'int'
f( 2, 2ull); // return 'unsigned long long'
f(2.0f, 2ull); // return 'unsigned long long'
48/48
