What =delete means

In my C++ training classes, I often explain the meaning of =delete as

The library author is saying, “I know what you’re trying to do, and what you’re trying to do is wrong.”

In some situations there’s not much difference between =delete’ing a function and not-=delete’ing it. In fact, the most common way that people first learn about =delete is in a context where it doesn’t really matter — which is why they start out a little bit confused!

struct MoveOnly {
    MoveOnly& operator=(MoveOnly&&);

    MoveOnly(const MoveOnly&) = delete;  // redundant
    MoveOnly& operator=(const MoveOnly&) = delete;  // redundant

The question of whether to explicitly =delete here, or just leave the copy members undeclared, is purely stylistic. Me personally, I’d leave them undeclared and save two lines of code. Someone else might err on the side of being explicit.

Where =delete really matters is when it’s the best match in an overload set with at least one other candidate. My go-to example for this is std::cref, which takes a const T& and turns it into a reference_wrapper<const T>. Recall that const T& is happy to bind to both lvalues and rvalues. So, if the library designer provides only one overload of std::cref, then it’ll accept both lvalues and rvalues:

template<class T>
auto cref(const T&) -> std::reference_wrapper<const T>;

int i = 42;
auto r1 = std::cref(i);   // OK
auto r2 = std::cref(42);  // OK but dangling!

When the STL’s designers added a deleted overload for const T&&, they were saying, “I know what you’re trying to do [namely, call std::cref on an rvalue], and what you’re trying to do is wrong.”

template<class T>
auto cref(const T&) -> std::reference_wrapper<const T>;

template<class T>
auto cref(const T&&) = delete;

auto r2 = std::cref(42);  // Error, best match is deleted

This is not the same as the author saying “I don’t know what you’re trying to do.” It’s not even the same as the author saying “I know what you’re trying to do, but I don’t do it; maybe somebody else might.” In those cases, the author might leave the function undeclared, or have it SFINAE away. With =delete, the library author is specifically stating that for the thing you’re trying to do, the buck stops with them — and guess what? They don’t want you to do it.

There are very few uses of =delete in the STL. They tend to get used in three main situations:

First, of course, there’s deleting the copy operations of types you want to make immobile, such as lock_guard and std::pmr::monotonic_buffer_resource.

Second, there’s the ones isomorphic to std::cref — the STL loves to use “rvalueness” as a proxy for “short-livedness.” Sometimes, as with std::cref and std::addressof, I think it’s a defensible design choice, good enough to use as an example in class. Sometimes, as with the constructor of regex_iterator, I think it’s just silly: I don’t use that as an example. For more on this topic, see “Value category is not lifetime” (2019-03-11).

A similar situation arises with nullptr; for example, std::string_view deletes its constructor and assignment operator from nullptr_t:

std::string_view sv1 = nullptr;   // Error, best match is deleted
std::string_view sv2 = (char*)0;  // OK, UB at runtime

Perhaps the most interesting and defensible case is C++20’s deleted overloads of << for “character types” that aren’t char. Godbolt:

std::cout << u8"hello world!" << '\n';  // const char8_t*
std::cout << L'h' << '\n';              // wchar_t

Using C++17’s overload set, which had no overloads for const char8_t* nor wchar_t, this example would just choose the best-matching candidates: ostream::operator<<(const void*) and ostream::operator<<(int) respectively. In C++20, ostream provides specifically =delete’d overloads for these types, so you get a nice compiler error instead of gibberish at runtime. This is the library designer saying, “I understand what you’re trying to do [print a wchar_t to a narrow-character terminal], and what you’re trying to do is wrong.”

Now, maybe the designer actually does want you to be able to do it, someday! But they know that if you try it this year, you’ll just end up with gibberish and tears. So it’s best, instead of allowing the less-good const void* or int overloads to be picked, to shut you down at compile time. =delete provides the mechanism to do just that.

The third and most interesting use of =delete in the STL is to disable specific template arguments. For example, the STL provides bodies for std::make_unique<Widget> and std::make_unique<Widget[]>, but if you try std::make_unique<Widget[10]> you’ll hit a deleted overload. “I know what you’re trying to do, and you’re wrong.” (You probably wanted std::make_unique<Widget[]>(10) instead.)

Finally, although the STL doesn’t do this anywhere yet, there’s one more place you might consider using =delete: If your library provides a function that’s meant to be called through ADL, and then you deprecate and remove that function. Godbolt:

namespace mine {
    struct Widget {};
    void oldapi(Widget);
    void newapi(Widget);
} // namespace mine

struct Sink { Sink(auto) {} };
int oldapi(Sink);

int main() {
    mine::Widget w;

We can mark oldapi(Widget) as [[deprecated("please switch to newapi")]] for a few releases, but at some point we’ll want to get rid of it. We could just remove the declaration altogether; if we do that, then any code (such as main) that’s still calling oldapi via ADL might silently switch to some other oldapi, such as oldapi(Sink). So, if we’re really worried about this possibility, then we might change the declaration of our oldapi from [[deprecated]] to =delete’d. Again, this is us saying, “I know what you’re trying to do, main, and you’re (now) wrong to want to do it.”

Notice that this logic doesn’t apply if we expect the caller to always call mine::oldapi via its qualified name. The compiler is perfectly capable of telling the programmer that the qualified name mine::oldapi doesn’t exist (anymore). The programmer’s intent in that case is very clear; the compiler doesn’t need any help from the library designer. Where =delete is useful is where the library designer sees a specific pattern of (potential) misuse: in the same way that we craft positive overload sets to say “I know what this caller wants, and I’m implementing the best way to do that thing,” =delete allows us to craft “negative” overload sets to say “I know what this caller wants, and I don’t want them to be doing that.

See also:

Posted 2021-10-17