Modern C++
Programming
12. Templates and
Meta-programming II
Class Templates , Sfinae, and Concepts
Federico Busato
2026-01-06
Class Template
Similarly to function templates, class templates are used to build a family of classes
template<typename T>
struct A { // class template (typename template)
T x = 0;
};
template<int N1>
struct B { // class template (numeric template)
int N = N1;
};
A<int> a1; // a1.x is int x = 0
A<float> a2; // a2.x is float x = 0.0f
B<1> b1; // b1.N is 1
B<2> b2; // b2.N is 2
5/85
Class Template Specialization 1/2
The main difference with template functions is that classes can be partially specialized
Note: Every class specialization (both partial and full) is a completely new class, and it does
not share anything with the generic class
template<typename T, typename R>
struct A {}; // generic class template
template<typename T>
struct A<T, int> {}; // partial specialization
template<>
struct A<float, int> {}; // full specialization
6/85
Class Template Specialization 2/2
template<typename T, typename R>
struct A { // GENERIC class template
T x;
};
template<typename T>
struct A<T, int> { // PARTIAL specialization
T y;
};
A<float, float> a1;
a1.x; // ok, generic template
// a1.y; // compile error
A<float, int> a2;
a2.y; // ok, partial specialization
// a2.x; // compile error
7/85
Example 1: Implement a Simple Type Trait
template<typename T, typename R> // GENERIC template declaration
struct is_same {
static constexpr bool value = false;
};
template<typename T>
struct is_same<T, T> { // PARTIAL template specialization
static constexpr bool value = true;
};
cout << is_same< int, char>::value; // print false, generic template
cout << is_same<float, float>::value; // print true, partial template
8/85
Example 2: Check if a Pointer is const
# include <type_traits>
// std::true_type and std::false_type contain a field "value"
// set to true or false respectively
template<typename T>
struct is_pointer_to_const : std::false_type {}; // GENERIC template declaration
template<typename R> // PARTIAL specialization
struct is_pointer_to_const<const R*> : std::true_type {};
cout << is_pointer_to_const<int*>::value; // print false, generic template
cout << is_pointer_to_const<const int*>::value; // print true, partial template
cout << is_pointer_to_const<int* const>::value; // print false, generic template
9/85
Example 3: Compare Class Templates
# include <type_traits>
template<typename T>
struct A {};
template<typename T, typename R>
struct Compare : std::false_type {}; // GENERIC template declaration
template<typename T, typename R>
struct Compare<A<T>, A<R>> : std::true_type {}; // PARTIAL specialization
cout << Compare<int, float>::value; // false, generic template
cout << Compare<A<int>, A<int>>::value; // true, partial template
cout << Compare<A<int>, A<float>>::value; // true, partial template
10/85
Class Template Constructor
Class template arguments don’t need to be repeated if they are the default ones
template<typename T>
struct A {
A(const A& x); // A(const A<T>& x);
A f(); // A<T> f();
};
11/85
Constructor Template Automatic Deduction (CTAD)
C++17 introduces automatic deduction of class template arguments in constructor
calls
template<typename T, typename R>
struct A {
A(T x, R y) {}
};
A<int, float> a1(3, 4.0f); // < C++17
A a2(3, 4.0f); // C++17
// A<int> a{3, 5}; compile error, "partial" specialization
12/85
CTAD - User-Defined Deduction Guides
Template deduction guide is a mechanism to instruct the compiler how to map
constructor parameter types into class template parameters
template<typename T>
struct MyString {
MyString(T) {}
};
// constructor class instantiation
MyString(char const*) -> MyString<std::string>; // deduction guide
MyString s{"abc"}; // construct 'MyString<std::string>'
13/85
CTAD - User-Defined Deduction Guides - Aggregate Example
template<typename T>
struct A {
T x, y;
};
template<typename T>
A(T, T) -> A<T>; // deduction guide
// not required in C++20+ for aggregates
A a{1, 3}; // construct 'A<int, int>'
14/85
CTAD - User-Defined Deduction Guides - Independent Argument Example
template<int I>
struct A {
template<typename T>
A(T) {}
};
template<typename T>
A(T) -> A<sizeof(T)>; // deduction guide
A a{1}; // construct 'A<4>', 4 == sizeof(int)
15/85
CTAD - User-Defined Deduction Guides - Universal Reference Example
# include <type_traits> // std::remove_reference_t
template<typename T>
struct A {
template<typename R>
A(R&&) {}
};
template<typename R>
A(R&&) -> A<std::remove_reference_t<R>>; // deduction guide
int x;
A a{x}; // construct 'A<int>' instead of 'A<int&>'
16/85
CTAD - User-Defined Deduction Guides - Iterator Example
# include <type_traits> // std::remove_reference_t
# include <vector> // std::vector
template<typename T>
struct Container {
template<typename Iter>
Container(Iter beg, Iter end) {}
};
template<typename Iter>
Container(Iter b, Iter e) -> // deduction guide
Container<typename std::iterator_traits<Iter>::value_type>;
std::vector v{1, 2, 3};
Container c{v.begin(), v.end()}; // construct 'Container<int>'
17/85
CTAD - User-Defined Deduction Guides - Alias Template
Alias template deduction requires C++20
template<typename T>
struct A {
A(T) {}
};
template<typename T>
A(T) -> A<int>; // deduction guide
template<typename T>
using B = A<T>; // alias template
B c{3.0}; // alias template deduction
// construct 'A<int>'
18/85
CTAD User-Defined Deduction Guides - Limitation
Template deduction guide doesn’t work within the class scope
template<typename T>
struct MyString {
MyString(T) {}
auto f() { return MyString("abc"); } // create 'MyString<const char*>'
}; // NOT 'MyString<std::string>'
MyString(const char*) -> MyString<std::string>; // deduction guide
MyString<const char*> s{"abc"}; // construct 'MyString<const char*>'
The problem can be avoided by using fully-qualifying the class type
auto f() { return ::MyString("abc"); }
19/85
Class + Function - Specialization 1/3
Given a class template and a template member function
template<typename T, typename R>
struct A {
template<typename X, typename Y>
void f();
};
There are two ways to specialize the class/function:
• Generic class + generic function
• Full class specialization + generic/full specialization function
20/85
Class + Function - Specialization 2/3
template<typename T, typename R>
template<typename X, typename Y>
void A<T, R>::f() {}
// ok, A<T, R> and f<X, Y> are not specialized
template<>
template<typename X, typename Y>
void A<int, int>::f() {}
// ok, A<int, int> is full specialized
// ok, f<X, Y> is not specialized
template<>
template<>
void A<int, int>::f<int, int>() {}
// ok, A<int, int> and f<int, int> are full specialized
21/85
Class + Function - Specialization 3/3
template<typename T>
template<typename X, typename Y>
void A<T, int>::f() {}
// error A<T, int> is partially specialized
// (A<T, int> class must be defined before)
template<typename T, typename R>
template<typename X>
void A<T, R>::f<int, X>() {}
// error function members cannot be partially specialized
template<typename T, typename R>
template<>
void A<T, R>::f<int, int>() {}
// error function members of a non-specialized class cannot be specialized
// (requires a binding to a specific template instantiation at compile-time)
22/85
Accessing a Dependent Type - typename Keyword 1/2
Structure templates can have different data members for each specialization.
The compiler needs to know in advance if a symbol within a structure is a type or a
static member when the structure template depends on another template parameter
The keyword typename placed before a structure template solves this ambiguous
template<typename T>
struct A {
using type = int;
};
template<typename R>
void g() {
using X = typename A<R>::type; // "type" is a typename or
} // a data member depending on R
23/85
Accessing a Dependent Type - typename Keyword 2/2
The using keyword can be used to simply the expression to get the structure type
template<typename T>
struct A {
using type = int;
};
template<typename T>
using AType = typename A<T>::type;
template<typename R>
void g() {
using X = AType<R>;
}
24/85
Template Dependent Names - template Keyword
The template keyword tells the compiler that what follows is a template name
(function or class)
note: recent compilers don’t strictly require this keyword in simple cases
template<typename T>
struct A {
template<typename R>
void g() {}
};
template<typename T> // A<T> is a dependent name (from T)
void f(A<T> a) {
// a.g<int>(); // compile error A<T> is dependent on T
// interpreted as: "a.g < int > ();"
// namely: "(a.g < int) > ();"
a.template g<int>(); // ok
}
25/85
Class Template Hierarchy and using
Member of class templates can be used internally in derived class templates by
specifying the particular type of the base class with the keyword using
template<typename T>
struct A {
T x;
void f() {}
};
template<typename T>
struct B : A<T> {
using A<T>::x; // needed (otherwise it could be another specialization)
using A<T>::f; // needed
void g() {
x; // without 'using': this->x
f();
}
};
26/85
virtual Function and Template
Virtual functions cannot have template arguments
• Templates are a compile-time feature
• Virtual functions are a run-time feature
Full story:
The reason for the language disallowing the particular construct is that there are
potentially infinite different types that could be instantiating your template member
function, and that in turn means that the compiler would have to generate code to
dynamically dispatch those many types, which is infeasible
stackoverflow.com/a/79682130
27/85
friend Keyword
template<typename T> struct A {};
template<typename T, typename R> struct B {};
template<typename T> void f() {}
//----------------------------------------------------------------------------------
class C {
friend void f<int>(); // match only f<int>
template<typename T> friend void f(); // match all templates
friend struct A<int>; // match only A<int>
template<typename> friend struct A; // match all A templates
// template<typename T> friend struct B<int, T>;
// partial specialization cannot be declared as a friend
};
28/85
Template Template Arguments
Template template parameters match templates instead of concrete types
template<typename T> struct A {};
template< template<typename> class R >
struct B {
R<int> x;
R<float> y;
};
template< template<typename> class R, typename S >
void f(R<S> x) {} // works with every class with exactly one template parameter
B<A> y;
f( A<int>() );
class and typename keyword are interchangeably in C++17
29/85
Template Meta-Programming
“Metaprogramming is the writing of computer programs with the
ability to treat programs as their data. It means that a program could
be designed to read, generate, analyze or transform other programs, and
even modify itself while running”
“Template meta-programming refers to uses of the C++ template
system to perform computation at compile-time within the code.
Templates meta-programming include compile-time constants, data
structures, and complete functions”
30/85
Template Meta-Programming
• Template Meta-Programming is fast (runtime)
Template Metaprogramming is computed at compile-time (nothing is computed at
run-time)
• Template Meta-Programming is Turing Complete
Template Metaprogramming is capable of expressing all tasks that standard
programming language can accomplish
• Template Meta-Programming requires longer compile time
Template recursion heavily slows down the compile time, and requires much more
memory than compiling standard code
• Template Meta-Programming is complex
Everything is expressed recursively. Hard to read, hard to write, and also very hard
to debug
31/85
Example 1: Factorial
template<int N>
struct Factorial { // GENERIC template: Recursive step
static constexpr int value = N * Factorial<N - 1>::value;
};
template<>
struct Factorial<0> { // FULL SPECIALIZATION: Base case
static constexpr int value = 1;
};
constexpr int x = Factorial<5>::value; // 120
// int y = Factorial<-1>::value; // Infinite recursion :)
32/85
Example 1: Factorial (Notes)
The previous example can be easily written as a constexpr in C++14
template<typename T>
constexpr int factorial(T value) {
T tmp = 1;
for (int i = 2; i <= value; i++)
tmp *= i;
return tmp;
};
Advantages:
• Easy to read and write (easy to debug)
• Faster compile time (no recursion)
• Works with different types (typename T)
• Works at run-time and compile-time
33/85
Example 2: Log2
template<int N>
struct Log2 { // GENERIC template: Recursive step
static_assert(N > 0, "N must be greater than zero");
static constexpr int value = 1 + Log2<N / 2>::value;
};
template<>
struct Log2<1> { // FULL SPECIALIZATION: Base case
static constexpr int value = 0;
};
constexpr int x = Log2<20>::value; // 4
34/85
Example 3: Log
template<int A, int B>
struct Max { // utility
static constexpr int value = A > B ? A : B;
};
template<int N, int BASE>
struct Log { // GENERIC template: Recursive step
static_assert(N > 0, "N must be greater than zero");
static_assert(BASE > 0, "BASE must be greater than zero");
// Max is used to avoid Log<0, BASE>
static constexpr int TMP = Max<1, N / BASE>::value;
static constexpr int value = 1 + Log<TMP, BASE>::value;
};
template<int BASE>
struct Log<1, BASE> { // PARTIAL SPECIALIZATION: Base case
static constexpr int value = 0;
};
constexpr int x = Log<20, 2>::value; // 4
35/85
Example 4: Unroll (Compile-time/Run-time Mix)
⋆
template<int NUM_UNROLL, int STEP = 0>
struct Unroll { // GENERIC template: Recursive step
template<typename Op>
static void run(Op op) {
op(STEP);
Unroll<NUM_UNROLL, STEP + 1>::run(op);
}
};
template<int NUM_UNROLL>
struct Unroll<NUM_UNROLL, NUM_UNROLL> { // PARTIAL SPECIALIZATION: Base case
template<typename Op>
static void run(Op) {}
};
auto lambda = [](int step) { cout << step << ", "; };
Unroll<5>::run(lambda); // print "0, 1, 2, 3, 4"
36/85
SFINAE
SFINAE
Substitution Failure Is Not An Error (SFINAE) applies during overload resolution
of function templates. When substituting the deduced type for the template
parameter fails, the specialization is discarded from the overload set instead of
causing a compile error
37/85
The Problem
template<typename T>
T ceil_div(T value, T div);
template<>
unsigned ceil_div<unsigned>(unsigned value, unsigned div) {
return (value + div - 1) / div;
}
template<>
int ceil_div<int>(int value, int div) { // handle negative values
return (value > 0)
∧
(div > 0) ?
(value / div) : (value + div - 1) / div;
}
What about long long int , long long unsigned , short , unsigned short ,
etc.?
38/85
std::enable_if Type Trait
The common way to adopt SFINAE is using the
std::enable_if/std::enable_if_t type traits
std::enable_if allows a function template or a class template specialization to
include or exclude itself from a set of matching functions/classes
template<bool Condition, typename T = void>
struct enable_if {
// "type" is not defined if "Condition == false"
};
template<typename T>
struct enable_if<true, T> {
using type = T;
};
helper alias: std::enable_if_t<T> instead of typename std::enable_if<T>::type
39/85
Function SFINAE - Return type 1/5
# include <type_traits> // std::is_signed_v, std::enable_if_t
template<typename T>
std::enable_if_t<std::is_signed_v<T>>
f(T) {
cout << "signed";
}
template<typename T>
std::enable_if_t<!std::is_signed_v<T>>
f(T) {
cout << "unsigned";
}
f(1); // print "signed"
f(1u); // print "unsigned"
40/85
Function SFINAE - Parameter 2/5
# include <type_traits> compiler-explorer
template<typename T>
void f(std::enable_if_t<std::is_signed_v<T>, T>) {
cout << "signed";
}
template<typename T>
void f(std::enable_if_t<!std::is_signed_v<T>, T>) {
cout << "unsigned";
}
// NOTE: explicit SFINAE on parameter prevents argument deduction
f<int>(1); // print "signed"
f<unsigned>(1u); // print "unsigned"
// f(1); // compile error
// f(1u); // compile error
41/85
Function SFINAE - Hidden Parameter 3/5
# include <type_traits>
template<typename T>
void f(T,
std::enable_if_t<std::is_signed_v<T>, int> = 0) {
cout << "signed";
}
template<typename T>
void f(T,
std::enable_if_t<!std::is_signed_v<T>, int> = 0) {
cout << "unsigned";
}
f(1); // print "signed"
f(1u); // print "unsigned"
42/85
Function SFINAE - Hidden Template Parameter 4/5
# include <type_traits>
template<typename T,
std::enable_if_t<std::is_signed_v<T>, int> = 0>
void f(T) {}
template<typename T,
std::enable_if_t<!std::is_signed_v<T>, int> = 0>
void f(T) {}
f(4);
f(4u);
43/85
Function SFINAE - decltype + return type 5/5
# include <type_traits>
template<typename T, typename R> // (1)
decltype(T{} + R{}) add(T a, R b) { // T{} + R{} is not possible with 'A'
return a + b;
}
template<typename T, typename R> // (2)
std::enable_if_t<std::is_class_v<T>, T> // 'int' is not a class
add(T a, R b) {
return a;
}
struct A {};
add(1, 2u); // call (1)
add(A{}, A{}); // call (2)
// if 'A' supports operator+, then we have a conflict
44/85
Function SFINAE Example - Array vs. Pointer
# include <type_traits> compiler-explorer
template<typename T, int Size>
void f(T (&array)[Size]) {} // (1)
// template<typename T>
// void f(T array) {} // (2)
template<typename T>
std::enable_if_t<std::is_pointer_v<T>>
f(T ptr) {} // (3)
// void f(int* pointer) {} // (4) has the highest priority among (1), (2), and (3)
int array[3];
f(array); // It is not possible to call (1) if (2) is present
// The reason is that 'array' decays to a pointer
// Now with (3), the code calls (1)
45/85
Function SFINAE Notes
The wrong way to achieve SFINAE
template<typename T, typename = std::enable_if_t<std::is_signed_v<T>>>
void f(T) {}
// template<typename T, typename = std::enable_if_t<!std::is_signed_v<T>>>
// void f(T) {}
// compile error redefinition of the second template parameter
Using std::enable_if_t for the return type prevents auto deduction
// template<typename T>
// std::enable_if_t<std::is_signed_v<T>, auto> f(T) {}
// compile error auto is not allowed here
46/85
Class SFINAE
# include <type_traits>
template<typename T, typename Enable = void>
struct A;
template<typename T>
struct A<T, std::enable_if_t<std::is_signed_v<T>>> {
};
template<typename T>
struct A<T, std::enable_if_t<!std::is_signed_v<T>>> {
};
A<int> a1;
A<unsigned> a2;
47/85
Check Struct Member
⋆
1/3
SFINAE can be also used to check if a structure has a specific data member or type
Let consider the following structures:
struct A {
static int x;
int y;
using type = int;
};
struct B {};
48/85
Check Struct Member - Variable
⋆
2/3
# include <type_traits>
template<typename T, typename = void>
struct has_x : std::false_type {};
template<typename T>
struct has_x<T, decltype((void) T::x)> : std::true_type {};
template<typename T, typename = void>
struct has_y : std::false_type {};
template<typename T>
struct has_y<T, decltype((void) std::declval<T>().y)> : std::true_type {};
has_x< A >::value; // returns true
has_x< B >::value; // returns false
has_y< A >::value; // returns true
has_y< B >::value; // returns false
49/85
Check Struct Member - Type
⋆
3/3
template<typename...>
using void_t = void; // included in C++17 <utility>
template<typename T, typename = void>
struct has_type : std::false_type {};
template<typename T>
struct has_type<T,
std::void_t<typename T::type> > : std::true_type {};
has_type< A >::value; // returns true
has_type< B >::value; // returns false
50/85
Support Trait for Stream Operator
⋆
template<typename T>
using EnableP = decltype( std::declval<std::ostream&>() <<
std::declval<T>() );
template<typename T, typename = void>
struct is_stream_supported : std::false_type {};
template<typename T>
struct is_stream_supported<T, EnableP<T>> : std::true_type {};
struct A {};
is_stream_supported<int>::value; // returns true
is_stream_supported<A>::value; // returns false
51/85
Variadic Template 1/2
Variadic template (C++11)
A variadic template captures a parameter pack of arguments, which hold an
arbitrary number of values or types
template<typename... TArgs> // Variadic typename -> parameter pack: ... TArgs
void f(TArgs... args) {} // pack expansion -> pattern: TArgs
A parameter pack is introduced by an identifier TArgs prefixed by an ellipsis
... TArgs . Once captured, a parameter pack can later be used in a pattern
expanded by an ellipsis ...
A pack expansion is equivalent to a comma-separated list of instances of the pattern
A pattern is a set of tokens containing the identifiers of one or more parameter packs.
When a pattern contains more than one parameter pack, all packs must have the same length
53/85
Variadic Template 2/2
template<typename... TArgs>
void f(TArgs... args) { // Typename expansion
int values[] = {args...}; // Arguments expansion
}
f(1, 2, 3);
The pack TArgs expands in a template-argument-list, i.e. list of template arguments
The pack args expands in an initializer-list, i.e. list of values
The number of variadic arguments can be retrieved with the sizeof... operator
sizeof...(args) // e.g. 3
Note: variadic arguments must be the last one in the declaration
C++20 idioms for parameter packs
54/85
Example 1
// BASE CASE
template<typename T, typename R>
auto add(T a, R b) {
return a + b;
}
// RECURSIVE CASE
template<typename T, typename... TArgs> // Variadic typename
auto add(T a, TArgs... args) { // Typename expansion
return a + add(args...); // Arguments expansion
}
add(2, 3.0); // 5
add(2, 3.0, 4); // 9
add(2, 3.0, 4, 5); // 14
// add(2); // compile error the base case accepts only two arguments
55/85
Example 2 - Function Unpack
template<typename T, typename... TArgs>
auto add(T a, TArgs... args); // see previous slides
struct A {
int v;
int f() { return v; }
};
template<typename... TArgs>
int f(TArgs... args) {
return add(args.f()...); // equivalent to: 'A{1}.f(), A{2}.f(), A{3}.f()'
}
f(A{1}, A{2}, A{3}); // return 6
56/85
Example 3 - Function Application
template<typename T, typename... TArgs>
auto add(T a, TArgs... args); // see previous slides
template<typename T>
T square(T value) { return value * value; }
//-----------------------------------------------------------
template<typename... TArgs>
auto add_square(TArgs... args) {
return add(square(args)...); // square() is applied to each
} // variadic argument
add_square(2, 2, 3.0f); // returns 17.0f
57/85
Example 4 - Type Expansion
template<typename... TArgs>
int g(TArgs... args) {}
template<typename... TArgs>
int f(TArgs... args) {
g<std::make_unsigned_t<TArgs>...>(args...);
}
f(1, 2, 3);
58/85
Function Initializer List Types
template<typename... TArgs>
void f(TArgs... args) {} // pass by-value
template<typename... TArgs>
void g(const TArgs&... args) {} // pass by-const reference
template<typename... TArgs>
void h(TArgs*... args) {} // pass by-pointer
template<int... Sizes>
void l(int (&...arrays)[Sizes]) {} // pass a list of array references
int a[] = {1, 2};
int b[] = {1, 2, 3};
f(1, 2.0);
h(a, b);
l(a, b); // same as g()
59/85
Homogeneous Variadic Template Parameters
Parameter pack can be also used to create a homogeneous variadic template
parameters
template<int... IntSeq> // sequence of integers
void f() {}
f<1, 2, 3>();
template<int... IntSeq> // sequence of integers
class A {};
A<1, 2, 3> a{};
60/85
Other Usages
Variadic templates can be also applied to lambdas with generic parameters (C++14)
and concepts (C++20)
auto lambda = [](auto... args) {};
lambda(1, 2u, 3.0f, 1ull);
void f(std::floating_point auto... args) {}
f(1.0, 2.0f); // ok
// f(1u, 2.0f); // compile error
61/85
Advanced Usages
⋆
Besides initializer-lists, template-argument-list, parameter pack can be used in:
capture list, constructor initializer-list, using declaration
template<typename... BaseClasses>
struct A : BaseClasses... { // : BaseClass_1, BaseClass_2, ...
A(int v) : BaseClasses...{v} {} // BaseClass_1{v}, BaseClass_2{v}, ...
using BaseClasses::f;
// equivalent to:
// using BaseClass_1::f;
// using BaseClass_2::f;
// ...
};
void f(auto... args) {
auto lambda = [arg&...](){}; // capture by-reference
}
62/85
Folding Expression 1/2
C++17 Folding expressions perform a fold of a template parameter pack over any
binary operator in C++ ( + , * , , , += , && , <= etc.)
Unary/Binary folding
template<typename... Args>
auto add_unary(Args... args) { // Unary folding
return (... + args); // unfold: 1 + 2.0f + 3ull
}
template<typename... Args>
auto add_binary(Args... args) { // Binary folding
return (1 + ... + args); // unfold: 1 + 1 + 2.0f + 3ull
} // the value on the left of the ellipsis is typically the identity
add_unary(1, 2.0f, 3ll); // returns 6.0f (float)
add_binary(1, 2.0f, 3ll); // returns 7.0f (float)
63/85
Example 1 - Extract The Last Argument
template<typename... TArgs>
int f(TArgs... args) {
return (args, ...); // the comma operator discards left values
} // same as (..., args)
f(1, 2, 3); // return 3
64/85
Example 2 - Function Application
Same example of “Variadic Template - Function Application” ... but shorter
template<typename T>
T square(T value) { return value * value; }
template<typename... TArgs>
auto add_square(TArgs... args) {
return (square(args) + ...); // square() is applied to each
} // variadic argument
add_square(2, 2, 3.0f); // returns 17.0f
Note: binary unfolding (1 + ... + args) cannot be combined with functions
65/85
Example 3 - Homogeneous Variadic Parameter Type
Parameter pack can be constrained to obtain a homogeneous variadic parameter
type
template <typename ... TArgs>
std::enable_if_t<(std::is_same_v<TArgs, int> && ... && true)>
f(const TArgs ... args) {}
f(1, 2, 3); // ok
// f(1u, 2, 3); // compile error
66/85
Variadic Template and Classes
template<int... NArgs>
struct Add; // data structure declaration
template<int N1, int N2>
struct Add<N1, N2> { // BASE case
static constexpr int value = N1 + N2;
};
template<int N1, int... NArgs>
struct Add<N1, NArgs...> { // RECURSIVE case
static constexpr int value = N1 + Add<NArgs...>::value;
};
Add<2, 3, 4>::value; // returns 9
// Add<>; // compile error no match
// Add<2>::value; // compile error
// call Add<N1, NArgs...>, then Add<>
67/85
Variadic Class Template
⋆
Variadic Template can be used to build recursive data structures
template<typename... TArgs>
struct Tuple; // data structure declaration
template<typename T>
struct Tuple<T> { // base case
T value; // specialization with one parameter
};
template<typename T, typename... TArgs>
struct Tuple<T, TArgs...> { // recursive case
T value; // specialization with more
Tuple<TArgs...> tail; // than one parameter
};
Tuple<int, float, char> t1 { 2, 2.0, 'a' };
t1.value; // 2
t1.tail.value; // 2.0
t1.tail.tail.value; // 'a'
68/85
Variadic Template and Class Specialization
⋆
1/3
Get function arity at compile-time:
template <typename T>
struct GetArity;
// generic function pointer
template<typename R, typename... Args>
struct GetArity<R(*)(Args...)> {
static constexpr int value = sizeof...(Args);
};
// generic function reference
template<typename R, typename... Args>
struct GetArity<R(&)(Args...)> {
static constexpr int value = sizeof...(Args);
};
// generic function object
template<typename R, typename... Args>
struct GetArity<R(Args...)> {
static constexpr int value = sizeof...(Args);
};
69/85
Variadic Template and Class Specialization
⋆
2/3
void f(int, char, double) {}
int main() {
// function object
GetArity<decltype(f)>::value;
auto& g = f;
// function reference
GetArity<decltype(g)>::value;
// function reference
GetArity<decltype((f))>::value;
auto* h = f;
// function pointer
GetArity<decltype(h)>::value;
}
Get function arity from template parameter
70/85
Variadic Template and Class Specialization
⋆
3/3
Get operator() (and lambda) arity at compile-time:
template <typename T>
struct GetArity;
template<typename R, typename C, typename... Args>
struct GetArity<R(C::*)(Args...)> { // class member
static constexpr int value = sizeof...(Args);
};
template<typename R, typename C, typename... Args>
struct GetArity<R(C::*)(Args...) const> { // "const" class member
static constexpr int value = sizeof...(Args);
};
struct A {
void operator()(char, char) {}
void operator()(char, char) const {}
};
GetArity<A>::value; // call GetArity<R(C::*)(Args...)>
GetArity<const A>::value; // call GetArity<R(C::*)(Args...) const>
71/85
C++20 Concepts Overview 1/2
C++20 introduces concepts as an extension for templates to enforce constraints,
which specifies the requirements on template arguments
Concepts allows performing compile-time validation of template arguments
Advantages compared to SFINAE ( std::enable_if_t ):
• Better express the intention
• Concepts are easier to read and write
• Clear compile-time messages for debugging
• Faster compile time
• The concept behind C++ concepts
• Concepts vs type traits
• Constraints and concepts
• What are C++20 concepts and constraints? How to use them?
72/85
C++20 Concepts Overview 2/2
Keyword introduced:
concept Constrain
requires Constrain list/Requirements, clause and expression
73/85
The Problem
Goal: define a function to sum only arithmetic types
template<typename T>
T add(T valueA, T valueB) {
return valueA + valueB;
}
struct A {};
add(3, 4); // ok
// add(A{}, A{}); // not supported
SFINAE solution (ugly, verbose):
template<typename T>
std::enable_if_t<std::is_arithmetic_v<T>, T>
add(T valueA, T valueB) {
return valueA + valueB;
}
74/85
concept Keyword
[template arguments]
concept [name] = [compile-time boolean expression];
Example: arithmetic type concept
template<typename T>
concept Arithmetic = std::is_arithmetic_v<T>;
• Template argument constrain
template<Arithmetic T>
T add(T valueA, T valueB) {
return valueA + valueB;
}
• auto deduction constrain (constrained auto )
auto add(Arithmetic auto valueA, Arithmetic auto valueB) {
return valueA + valueB;
}
75/85
requires Clause
requires [compile-time boolean expression or Concept]
it acts like SFINAE
• After template parameter list
template<typename T>
requires Arithmetic<T>
T add(T valueA, T valueB) {
return valueA + valueB;
}
• After function declaration
template<typename T>
T add(T valueA, T valueB) requires (sizeof(T) == 4) {
return valueA + valueB;
}
76/85
requires Clause and concept Notes
Concepts and requirements can have multiple statements. It must be a primary
expression, e.g. constexpr value (not a constexpr function) or a sequence of
primary expressions joined with the operator && or ||
template<typename T>
concept Arithmetic2 = std::is_arithmetic_v<T> && sizeof(T) >= 4;
Concepts and requirements can be used together
template<Arithmetic T>
requires (sizeof(T) >= 4)
T add(T valueA, T valueB) {
77/85
requires Expression 1/2
A requires expression is a compile-time expression of type bool that defines the
constraints on template arguments
requires [(arguments)] {
[SFINAE contrain]; // or
requires [predicate];
} -> bool
template<typename T>
concept MyConcept = requires (T a, T b) { // First case: SFINAE constrains
a + b; // Req. 1 - support add operator
a[0]; // Req. 2 - support subscript operator
a.x; // Req. 3 - has "x" data member
a.f(); // Req. 4 - has "f" function member
typename T::type; // Req. 5 - has "type" field
};
78/85
requires Expression 2/2
Concept library
# include <concepts>
template<typename T>
concept MyConcept2 = requires (T a, T b) {
{*a + 1} -> std::convertible_to<float>; // Req. 6 - can be deferred and the sum
// with an integer is convertible
// to float
{a * a} -> std::same_as<int>; // Req. 7 - "a * a" must be valid and
// the result type is "int"
};
79/85
requires Expression + Clause
requires expression can be combined with requires clause
(see requires definition, second case) to compute a boolean value starting from
SFINAE expressions
template<typename T>
concept Arithmetic = requires { // expression -> bool (zero args)
T::value; // clause -> direct SFINAE
requires std::is_arithmetic_v<T>; // clause -> SFINAE from boolean
};
template<typename T>
concept MyConcept = requires (T value) { // expression -> bool (one arg)
requires sizeof(value) >= 4; // clause -> SFINAE from boolean
requires std::is_floating_point_v<T>; // clause -> SFINAE from boolean
};
80/85
requires Clause + Expression
requires clause can be combined with requires expression to apply SFINAE
(functions, structures) starting from a compile-time boolean expressions
template<typename T>
void f(T a) requires requires { T::value; }
// clause -> SFINAE followed by
// expression -> bool (zero args)
{}
template<typename T>
T increment(T a) requires requires (T x) { x + 1; }
// clause -> SFINAE followed by
// expression -> bool (one arg)
{
return a + 1;
}
81/85
requires and constexpr
Some examples:
• constexpr bool has_member_x = requires(T v){ v.x; };
• if constexpr (MyConcept<T>)
• static_assert(requires(T v){ ++v; }, "no increment");
• template<typename Iter>
constexpr bool is_iterator() {
return requires(Iter it) { *it++; };
}
82/85
Template Debugging
• -ftemplate-backtrace-limit=<N> Maximum number of template
instantiation notes for a single warning/error to N , default 10
N=1 is useful when looking only at the lasted instantiation (much less verbose
output). N=100 (or higher) if you are looking at all template instantiations (rare)
• -ftemplate-depth=<N> Set the maximum instantiation depth for template
classes to N , default 900
• -Wfatal-errors Abort compilation on the first error occurred rather than
trying to keep going and printing further error messages
84/85
Template Debugging
• -fdiagnostics-show-template-tree Display the templates as an indented
text tree
map<
[...],
map<
[float != double],
[...]>>
85/85
