Direct initialization

Initializes an object from explicit set of constructor arguments.


T object ( arg );

T object ( arg1, arg2, ... );

T object { arg }; (2) (since C++11)
T ( other )

T ( arg1, arg2, ... ).

static_cast< T >( other ) (4)
new T( args, ... ) (5)
Class::Class() : member( args, ... ) { ... } (6)
[arg]() { ... } (7) (since C++11)


Direct initialization is performed in the following situations:

1) Initialization with a nonempty parenthesized list of expressions or braced-init-lists (since C++11).
2) Initialization of an object of non-class type with a single brace-enclosed initializer (note: for class types and other uses of braced-init-list, see list-initialization) (since C++11).
3) Initialization of a prvalue temporary (until C++17)the result object of a prvalue (since C++17) by function-style cast or with a parenthesized expression list.
4) Initialization of a prvalue temporary (until C++17)the result object of a prvalue (since C++17) by a static_cast expression.
5) Initialization of an object with dynamic storage duration by a new-expression with a non-empty initializer.
6) Initialization of a base or a non-static member by constructor initializer list.
7) Initialization of closure object members from the variables caught by copy in a lambda-expression.

The effects of direct initialization are:

  • If T is an array type,
  • the program is ill-formed.
(until C++20)
struct A
    explicit A(int i = 0) {}
A a[2](A(1)); // OK: initializes a[0] with A(1) and a[1] with A()
A b[2]{A(1)}; // error: implicit copy-list-initialization of b[1]
              //        from {} selected explicit constructor
(since C++20)
  • If T is a class type,
  • if the initializer is a prvalue expression whose type is the same class as T (ignoring cv-qualification), the initializer expression itself, rather than a temporary materialized from it, is used to initialize the destination object.
    (Before C++17, the compiler may elide the construction from the prvalue temporary in this case, but the appropriate constructor must still be accessible: see copy elision)
(since C++17)
  • the constructors of T are examined and the best match is selected by overload resolution. The constructor is then called to initialize the object.
  • otherwise, if the destination type is a (possibly cv-qualified) aggregate class, it is initialized as described in aggregate initialization except that narrowing conversions are permitted, designated initializers are not allowed, a temporary bound to a reference does not have its lifetime extended, there is no brace elision, and any elements without an initializer are value-initialized.
struct B
    int a;
    int&& r;
int f();
int n = 10;
B b1{1, f()};            // OK, lifetime is extended
B b2(1, f());            // well-formed, but dangling reference
B b3{1.0, 1};            // error: narrowing conversion
B b4(1.0, 1);            // well-formed, but dangling reference
B b5(1.0, std::move(n)); // OK
(since C++20)
  • Otherwise, if T is a non-class type but the source type is a class type, the conversion functions of the source type and its base classes, if any, are examined and the best match is selected by overload resolution. The selected user-defined conversion is then used to convert the initializer expression into the object being initialized.
  • Otherwise, if T is bool and the source type is std::nullptr_t, the value of the initialized object is false.
  • Otherwise, standard conversions are used, if necessary, to convert the value of other to the cv-unqualified version of T, and the initial value of the object being initialized is the (possibly converted) value.


Direct-initialization is more permissive than copy-initialization: copy-initialization only considers non-explicit constructors and non-explicit user-defined conversion functions, while direct-initialization considers all constructors and all user-defined conversion functions.

In case of ambiguity between a variable declaration using the direct-initialization syntax (1) (with round parentheses) and a function declaration, the compiler always chooses function declaration. This disambiguation rule is sometimes counter-intuitive and has been called the most vexing parse.

#include <fstream>
#include <iterator>
#include <string>
int main()
    std::ifstream file("data.txt");
    // The following is a function declaration:
    std::string foo1(std::istreambuf_iterator<char>(file),
    // It declares a function called foo1, whose return type is std::string,
    // first parameter has type std::istreambuf_iterator<char> and the name "file",
    // second parameter has no name and has type std::istreambuf_iterator<char>(),
    // which is rewritten to function pointer type std::istreambuf_iterator<char>(*)()
    // Pre-C++11 fix (to declare a variable) - add extra parentheses around one
    // of the arguments:
    std::string str1((std::istreambuf_iterator<char>(file)),
    // Post-C++11 fix (to declare a variable) - use list-initialization for any
    // of the arguments:
    std::string str2(std::istreambuf_iterator<char>{file}, {});


#include <iostream>
#include <memory>
#include <string>
struct Foo
    int mem;
    explicit Foo(int n) : mem(n) {}
int main()
    std::string s1("test"); // constructor from const char*
    std::string s2(10, 'a');
    std::unique_ptr<int> p(new int(1));  // OK: explicit constructors allowed
//  std::unique_ptr<int> p = new int(1); // error: constructor is explicit
    Foo f(2); // f is direct-initialized:
              // constructor parameter n is copy-initialized from the rvalue 2
              // f.mem is direct-initialized from the parameter n
//  Foo f2 = 2; // error: constructor is explicit
    std::cout << s1 << ' ' << s2 << ' ' << *p << ' ' << f.mem  << '\n';


test aaaaaaaaaa 1 2

See also

© cppreference.com
Licensed under the Creative Commons Attribution-ShareAlike Unported License v3.0.