[C++] Implement constexpr placement new (P2747R2)

[cor3ntin]

https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2024/p2747r2.html

[cor3ntin]

I started work on this

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: cor3ntin (cor3ntin)

https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2024/p2747r2.html

[tbaederr]

I'm not much of a spec person - does this include placing into an array of different type, e.g.:

char B[sizeof(int));
new (B) int{12};

?

Or is it only supposed to work if the types match, e.g.

int a;
new (&a) int{12};

[frederick-vs-ja]

I'm not much of a spec person - does this include placing into an array of different type, e.g.:

char B[sizeof(int));
new (B) int{12};

?

Or is it only supposed to work if the types match, e.g.

int a;
new (&a) int{12};

I believe the former is not yet covered by P2747R2. Construction is still required to be performed on "properly typed" storage in constant evaluation.

[cor3ntin]

The interconvertibility requirement allows things like

struct S {
  int i;
};

S s;
new (&s) int(0);

because the pointer to s can be converted to a pointer to int (ie we have to chase various bases, subobjects and union members, but it is implementable.

(Without that restriction all existing c++ implementations would have to be thrown away afaik)

[zygoloid]

How can we choose a union member if multiple union members have the same type? That case hits angelic nondeterminism, which we can't really support here.

[cor3ntin]

@zygoloid Haha, we were having that conversation at the same moment. Brian started a thread with CWG https://lists.isocpp.org/core/2024/07/16009.php (in part because the complexity introduced by interconvertibility may not be justified in the cases that would be implementable)