clang rejects regular constexpr member function in struct with virtual base class

[Rush10233]

First, the following code is rejected by clang but accepted by other compilers like GCC, MSVC, and ICC:

    struct base { };
    struct derived : virtual base {
        constexpr int tst(){return 1;}
    };

https://godbolt.org/z/rj63nKP4v

Clang reports the following message:

<source>:3:23: error: constexpr member function not allowed in struct with virtual base class
    3 |         constexpr int tst(){return 1;}
      |                       ^
<source>:2:22: note: virtual base class declared here
    2 |     struct derived : virtual base {
      |                      ^~~~~~~~~~~~
1 error generated.

The diagnostic seems not such necessary because member function tst doesn't involve any override and can be specified during compile stage.

It's more interesting that the overloaded copy assignment operator is also accepted by other compilers, which definitely contains the instantiation of the virtual base class:

    struct base { };
    struct derived : virtual base {
        //reject
        //constexpr derived(const derived &)=default;
        //reject
        //constexpr derived(derived &&)=default;
        //accept
        constexpr derived& operator=(derived const&) =default;
    };

https://godbolt.org/z/rYPEYTex8

Copy ctors and move ctors are always rejected though.

[MitalAshok]

For the first one (normal member function)

llvm-project/clang/lib/Sema/SemaDeclCXX.cpp

Lines 1826 to 1827 in 1b8ab2f

	// FIXME: This only applies to constructors and destructors, not arbitrary
	// member functions.

In C++20 (and before), this matches all of the requirements https://timsong-cpp.github.io/cppwp/n4861/dcl.constexpr#3

Notably C++20 had a requirement for constructors/destructors only:

• if the function is a constructor or destructor, its class shall not have any virtual base classes;

Which is being applied to all member functions by Clang. Stuff like this needs to be accepted: https://godbolt.org/z/76WzG3jv6

---

For the second one (defaulted copy-assign operator): In C++23 after p2448r2, it should be accepted.

The constructors not being allowed to be constexpr is still not allowed in C++23. https://eel.is/c++draft/dcl.constexpr#3.2

A function is constexpr-suitable if:

• [...]
• if the function is a constructor or destructor, its class does not have any virtual base classes.
  Except for instantiated constexpr functions, non-templated constexpr functions shall be constexpr-suitable.

So constexpr derived(const derived &)=default; is always ill-formed.

I believe gcc shouldn't accept it before C++23. https://timsong-cpp.github.io/cppwp/n4861/dcl.fct.def.default#3.sentence-1 (C++20)

An explicitly-defaulted function that is not defined as deleted may be declared constexpr or consteval only if it is constexpr-compatible ([special]).

[tbaederr]

There is even a FIXME comment about this:

llvm-project/clang/lib/Sema/SemaDeclCXX.cpp

Lines 1820 to 1841 in 208a08c

	if (MD && MD->isInstance()) {
	// C++11 [dcl.constexpr]p4:
	// The definition of a constexpr constructor shall satisfy the following
	// constraints:
	// - the class shall not have any virtual base classes;
	//
	// FIXME: This only applies to constructors and destructors, not arbitrary
	// member functions.
	const CXXRecordDecl *RD = MD->getParent();
	if (RD->getNumVBases()) {
	if (Kind == CheckConstexprKind::CheckValid)
	return false;
	Diag(NewFD->getLocation(), diag::err_constexpr_virtual_base)
	<< isa<CXXConstructorDecl>(NewFD)
	<< getRecordDiagFromTagKind(RD->getTagKind()) << RD->getNumVBases();
	for (const auto &I : RD->vbases())
	Diag(I.getBeginLoc(), diag::note_constexpr_virtual_base_here)
	<< I.getSourceRange();
	return false;
	}
	}

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: None (Rush10233)

First, the following code is rejected by clang but accepted by other compilers like GCC, MSVC, and ICC:

    struct base { };
    struct derived : virtual base {
        constexpr int tst(){return 1;}
    };

https://godbolt.org/z/rj63nKP4v

Clang reports the following message:

<source>:3:23: error: constexpr member function not allowed in struct with virtual base class
    3 |         constexpr int tst(){return 1;}
      |                       ^
<source>:2:22: note: virtual base class declared here
    2 |     struct derived : virtual base {
      |                      ^~~~~~~~~~~~
1 error generated.

The diagnostic seems not such necessary because member function tst doesn't involve any override and can be specified during compile stage.

It's more interesting that the overloaded copy assignment operator is also accepted by other compilers, which definitely contains the instantiation of the virtual base class:

    struct base { };
    struct derived : virtual base {
        //reject
        //constexpr derived(const derived &)=default;
        //reject
        //constexpr derived(derived &&)=default;
        //accept
        constexpr derived& operator=(derived const&) =default;
    };

https://godbolt.org/z/rYPEYTex8

Copy ctors and move ctors are always rejected though.

[frederick-vs-ja]

In C++20 (and before), this matches all of the requirements https://timsong-cpp.github.io/cppwp/n4861/dcl.constexpr#3

However, in this example it seems that tst can never produce a constant (sub)expression (even in C++23).

As derived is not an implicit-lifetime type and can't have any constexpr constructor due to the virtual base, it is impossible to start lifetime of a derived object in constant evaluation. As a result, a call to the non-static member function tst is not allowed in constant evaluation as any attempt would cause UB, and it's ill-formed, no diagnostic required to mark tst constexpr until C++23.

In C++23, it should be well-formed to make tst constexpr.

[MitalAshok]

@frederick-vs-ja You don't need an object to start its lifetime inside the constant expression to call a member function on it (https://stackoverflow.com/q/32803708)

struct base { };
struct derived : virtual base {
    constexpr int tst() {return 1;}
};
extern derived d;
static_assert(d.tst() == 1);  // Well formed in C++20 (C++11 even I think)

[frederick-vs-ja]

@frederick-vs-ja You don't need an object to start its lifetime inside the constant expression to call a member function on it (https://stackoverflow.com/q/32803708)

struct base { };
struct derived : virtual base {
    constexpr int tst() {return 1;}
};
extern derived d;
static_assert(d.tst() == 1);  // Well formed in C++20 (C++11 even I think)

Yeah, you're right (thanks to P2280R4). Not sure whether #95474 should also address this.

[MitalAshok]

It was valid even before because there are no unknown references (d is a constant glvalue, and there is nothing with reference or pointer type), so this is unrelated. See the second example (auto rando() here: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p2280r4.html#other-examples which is only invalid because of the reference type, and would be valid if the parameter was by-value.

[zygoloid]

Wow. Looks like I got this wrong 13 years ago. Oops!