Layout compatible classes involving [[no_unique_address]]

[leni536]

Consider the following example:

#include <type_traits>
#include <cstddef>

class A1 {
    private:
    int i;
    char ch;
};

class B1 {
    public:
    [[no_unique_address]] A1 a;
    char ch;
};

static_assert(std::is_standard_layout_v<B1>);

class A2 {
    public:
    int i;
    char ch;
};

class B2 {
    public:
    [[no_unique_address]] A2 b2;
    char ch;
};

static_assert(std::is_standard_layout_v<B1>);
static_assert(offsetof(B1, ch) != offsetof(B2, ch));

Godbolt: https://godbolt.org/z/1o91bE

Following the standard wording A1 and A2 are layout compatible [1]. Similarly B1 and B2 are layout compatible. However the ch member in B1 has different offset than the ch member in B2.

A1 is not POD for the purpose of layout while A2 is, this affects the offset of the ch member.

Two possible directions to fix this:

1. Modify the standard to disallow [[no_unique_address]] non-empty non-static data members to be part of a common initial sequence (or maybe only allow as the last element of a common initial sequence).
2. Modify the ABI to disable overlapping of [[no_unique_address]] non-empty data members in standard-layout classes.

Personally I would prefer the first option of these two.

[1] https://eel.is/c++draft/class#mem.general-22

[zygoloid]

I've forwarded this to the C++ committee for consideration.

[frederick-vs-ja]

A2 is a POD type in C++98/03 while A1 is not (although it's a POD type since C++11). In C++98 without the resolution of CWG 43, a subobject of POD type is effectively never potentially-overlapping, which is the reason for "POD for the purpose of layout" IMO.

I believe that "POD for the purpose of layout" should be removed from Itanium C++ ABI in v2 for consistency and complexity of implementation, since it exists almost only for a defect in C++98.

[leni536]

A somewhat similar issue with [[no_unique_address]] and common initial sequences:

#include <type_traits>
#include <cstddef>
 
struct A {};
struct B {};
struct C : A { char ch; };

struct D {
    [[no_unique_address]] A a;
    C c;
};

static_assert(std::is_standard_layout_v<D>);

struct E {
    [[no_unique_address]] B b;
    C c;
};

static_assert(std::is_standard_layout_v<E>);

#ifdef CHECK_LAYOUT_COMPATIBLE
static_assert(!std::is_layout_compatible_v<D, E>);
#endif
static_assert(offsetof(D, c) != offsetof(E,c));

https://godbolt.org/z/7M93vr4zb

D and E are layout compatible according to standard wording. Here the issue doesn't involve "POD for the purpose of layout".

I've forwarded this to the C++ committee for consideration.

@zygoloid : did this go anywhere? I'm not aware of a defect report in this area.

[frederick-vs-ja]

As CWG2759 has been resolved for a long while, can we consider the original issue resolved?