`<complex>`: Structured binding for `complex` is accidentally enabled before C++26, in a C++26-incompatible way

[frederick-vs-ja]

When attempting to implement WG21-P2819R2 "Add The Tuple Protocol To complex", I found that the following thing is accepted by MSVC STL (Godbolt link).

#include <complex>

int main() {
    auto [xy] = std::complex<double>{};
}

Currently, MSVC STL's complex satisfies the conditions in [dcl.struct.bind]/8 - it has only one data member of type value_type[2] which is public and inherited from its base class.

Since C++26, the aforementioned example will certainly stop working. Do we want to reject it in old modes by making something private? Given complex doesn't seem able to be passed or returned in registers, perhaps such change is ABI-compatible.

[StephanTLavavej]

I'm somewhat worried about finding novel ways to break ABI (I was scarred by defaulting tuple<>'s copy ctor, the one-line comment there marks a severe regression), but I think we could do manual inspection of what codegen is emitted when passing complex around.

[vmichal]

I would maybe consider looking into it. I have a very rough idea, however, if you could provide me with some examples of previous work, I would be grateful. All decisions whether something is or is not an ABI break seem to be made purely by optical inspection.