Hello,

I recently noticed something unexpected: it is possible to construct an
object over constant storage with `std::construct_at`. At least it works in
gcc 13.2.0 and clang 18.
Note that this is not possible with placement new.

I tried playing with the following code:
template <class T>
void swap_const_references(T const& a, T const& b)
{
T c{a};
a.~T();
std::construct_at(&a, b);
b.~T();
std::construct_at(&b, c);
}

And it compiles fine with both compilers. If I replace the
std::construct_at with placement new, the code fails to compile (cannot
convert 'const void*' to 'void*').

This swap_const_references is definitely something we don't want to be able
to write! At least not without relying on UB. I can even make it constexpr
and it is still compiling >< (and working as designed...)

I don't know who is in the wrong here, between the library implementers and
the C++ standard. But here is a case where code with a non-constexpr
placement new is actually safer than the shinny construct_at...

Should we open a defect to fix std::construct_at so it does not accept
constructing over constant storage?

Best regards,
Sébastien BINI