25
votes

Explicit copy constructors disallow something like Foo foo = bar;, and enforce the copy usage as Foo foo(bar);. In addition, explicit copy constructors also disallow returning objects by value from a function. However, I tried replacing the copy initialization with braces, like so

struct Foo
{
    Foo() = default;
    explicit Foo(const Foo&) = default;
};

int main()
{
    Foo bar;
    Foo foo{bar}; // error here
}

and I am getting the error (g++5.2)

error: no matching function for call to 'Foo::Foo(Foo&)'

or (clang++)

error: excess elements in struct initializer

Removing the explicit makes the code compilable under g++, clang++ still fails with the same error (thanks @Steephen). What's going on here? Is the uniform initialization considered as an initializer-list constructor (which trumps all others)? But if that's the case, why does the program compile when the copy constructor is non-explicit?

1
clang shows error even without explicit keyword coliru.stacked-crooked.com/a/1cca94237ee00ea3Steephen
@Steephen indeed, I didn't see that. So it's another funky language-lawyer type question :)vsoftco
I would say that compilers choose to aggregate initialize foo.Jarod42
@Jarod42 You're probably right, it makes sense given the error message.vsoftco
It seems it is the case, as adding int member provide no viable conversion from 'Foo' to 'int' : Demo (and still different behaviour for gcc without the explicit...)Jarod42

1 Answers

24
votes

You've encountered a case that was addressed by the resolution of Core issue 1467 immediately after C++14 was finalized.

Let's first note that class foo is an aggregate. Your code is doing direct-list-initialization for foo. The rules for list-initialization are in [8.5.4p3].

In C++14 (quoting from N4140, the working draft closest to the published standard), the paragraph above started with:

List-initialization of an object or reference of type T is defined as follows:

  • If T is an aggregate, aggregate initialization is performed (8.5.1).

[...]

So, if your class is an aggregate, the compiler tries to do aggregate initialization, which fails.

This was recognized as a problem, and fixed in the working draft. Quoting from the current version, N4527, the above-mentioned paragraph now starts with:

List-initialization of an object or reference of type T is defined as follows:

  • If T is a class type and the initializer list has a single element of type cv U, where U is T or a class derived from T, the object is initialized from that element (by copy-initialization for copy-list-initialization, or by direct-initialization for direct-list-initialization).
  • Otherwise, if T is a character array and the initializer list has a single element that is an appropriately typed string literal (8.5.2), initialization is performed as described in that section.
  • Otherwise, if T is an aggregate, aggregate initialization is performed (8.5.1).

[...]

Your example now falls within the case described by the first bullet point, and foo is direct-list-initialized using the defaulted copy constructor (regardless of whether it's explicit, since it's direct initialization).

That is... if the compiler implements the resolution in the defect report.

  • GCC 5.2.0 (and 6.0.0 trunk) seems to do so, but seems to have a bug related to that explicit.
  • Clang 3.6.0 doesn't, but 3.8.0 trunk does, and does it correctly (explicit doesn't matter).
  • MSVC 14 does, but IntelliSense in the IDE doesn't (squiggles under bar - looks like the EDG compiler used by IntelliSense wasn't updated either).

Update: Since this answer was written, the working draft has been further amended in a couple of ways that are relevant to the example in the question and the explanation above:

  • CWG 2137 indicates that the first bullet in the paragraph quoted above went a bit too far by applying that exception to all class types (the issue notes contain a relevant example). The beginning of the bullet now reads:
    • If T is an aggregate class and [...]
  • The resolution of CWG 1518 contained in paper P0398R0 indicates that classes that declare an explicit constructor (even defaulted) are no longer aggregates.

This doesn't change the fact that, after all the changes are implemented, the example in the question is intended to work, with or without the explicit; it's just worth knowing that the underlying mechanisms that make it work have changed slightly.

Note that all these changes are resolutions of defect reports, so they're supposed to apply when compilers are in C++14 and C++11 modes as well.