Proposed resolution:

This wording is relative to N4981.

1. In [util.smartptr.shared.const]/9, specify that `shared_ptr(nullptr_t p, D d);` checks whether d(static_cast<T*>(nullptr)) is well-formed. This requires expressing the the constraints for the `Y*` constructors and the `nullptr_t` constructors separately, which is mostly editorial:

   
   template<class Y, class D> shared_ptr(Y* p, D d);
   template<class Y, class D, class A> shared_ptr(Y* p, D d, A a);
   template<class D> shared_ptr(nullptr_t p, D d);
   template<class D, class A> shared_ptr(nullptr_t p, D d, A a);
   

   -9- Constraints: is_move_constructible_v<D> is `true`, and `d(p)` is a well-formed expression. For the first two overloads:

   1. (9.1) If `T` is an array type, then either `T` is `U[N]` and `Y(*)[N]` is convertible to `T*`, or `T` is `U[]` and `Y(*)[]` is convertible to `T*`.
   2. (9.2) If `T` is not an array type, then `Y*` is convertible to `T*`.

   
   template<class D> shared_ptr(nullptr_t p, D d);
   template<class D, class A> shared_ptr(nullptr_t p, D d, A a);
   

   -?- Constraints: is_move_constructible_v<D> is `true`, and d(static_cast<T*>(p)) is a well-formed expression.

The following code doesn't compile on conforming implementations:

    #include <memory>

    void f() {
        std::shared_ptr<int>(new int, [](auto pointer) { delete pointer; });
    }

This is caused by the constraint on `shared_ptr(nullptr_t p, D d);` being that `d(p)` is valid ([util.smartptr.shared.const] p9), which leads to a hard error inside the lambda since it is called with a `nullptr_t`. This seems unintended.

See LLVM issue 93071 comment for additional context.