This creates a variant (a tagged union) that can store either an int or a string.

std::variant< int, std::string > var;

We can store one of either type in it:

var = "hello"s;

And we can access the contents via std::visit:

// Prints "hello\n":
visit( [](auto&& e) {
  std::cout << e << '\n';
}, var );

by passing in a polymorphic lambda or similar function object.

If we are certain we know what type it is, we can get it:

auto str = std::get<std::string>(var);

but this will throw if we get it wrong. get_if:

auto* str  = std::get_if<std::string>(&var);

returns nullptr if you guess wrong.

Variants guarantee no dynamic memory allocation (other than which is allocated by their contained types). Only one of the types in a variant is stored there, and in rare cases (involving exceptions while assigning and no safe way to back out) the variant can become empty.

Variants let you store multiple value types in one variable safely and efficiently. They are basically smart, type-safe unions.

This is an advanced example.

You can use variant for light weight type erasure.

template<class F>
struct pseudo_method {
  F f;
  // enable C++17 class type deduction:
  pseudo_method( F&& fin ):f(std::move(fin)) {}

  // Koenig lookup operator->*, as this is a pseudo-method it is appropriate:
  template<class Variant> // maybe add SFINAE test that LHS is actually a variant.
  friend decltype(auto) operator->*( Variant&& var, pseudo_method const& method ) {
    // var->*method returns a lambda that perfect forwards a function call,
    // behaving like a method pointer basically:
    return [&](auto&&...args)->decltype(auto) {
      // use visit to get the type of the variant:
      return std::visit(
        [&](auto&& self)->decltype(auto) {
          // decltype(x)(x) is perfect forwarding in a lambda:
          return method.f( decltype(self)(self), decltype(args)(args)... );
        },
        std::forward<Var>(var)
      );
    };
  }
};

this creates a type that overloads operator->* with a Variant on the left hand side.

// C++17 class type deduction to find template argument of `print` here.
// a pseudo-method lambda should take `self` as its first argument, then
// the rest of the arguments afterwards, and invoke the action:
pseudo_method print = [](auto&& self, auto&&...args)->decltype(auto) {
  return decltype(self)(self).print( decltype(args)(args)... );
};

Now if we have 2 types each with a print method:

struct A {
  void print( std::ostream& os ) const {
    os << "A";
  }
};
struct B {
  void print( std::ostream& os ) const {
    os << "B";
  }
};

note that they are unrelated types. We can:

std::variant<A,B> var = A{};

(var->*print)(std::cout);

and it will dispatch the call directly to A::print(std::cout) for us. If we instead initialized the var with B{}, it would dispatch to B::print(std::cout).

If we created a new type C:

struct C {};

then:

std::variant<A,B,C> var = A{};
(var->*print)(std::cout);

will fail to compile, because there is no C.print(std::cout) method.

Extending the above would permit free function prints to be detected and used, possibly with use of if constexpr within the print pseudo-method.

live example currently using boost::variant in place of std::variant.

This does not cover allocators.

struct A {};
struct B { B()=default; B(B const&)=default; B(int){}; };
struct C { C()=delete; C(int) {}; C(C const&)=default; };
struct D { D( std::initializer_list<int> ) {}; D(D const&)=default; D()=default; };

std::variant<A,B> var_ab0; // contains a A()
std::variant<A,B> var_ab1 = 7; // contains a B(7)
std::variant<A,B> var_ab2 = var_ab1; // contains a B(7)
std::variant<A,B,C> var_abc0{ std::in_place_type<C>, 7 }; // contains a C(7)
std::variant<C> var_c0; // illegal, no default ctor for C
std::variant<A,D> var_ad0( std::in_place_type<D>, {1,3,3,4} ); // contains D{1,3,3,4}
std::variant<A,D> var_ad1( std::in_place_index<0> ); // contains A{}
std::variant<A,D> var_ad2( std::in_place_index<1>, {1,3,3,4} ); // contains D{1,3,3,4}