MSVC Compatibility: Implement intrinsics for P0466R5 - Layout-compatibility and Pointer-interconvertibility Traits

[MahmoudGSaleh]

Bugzilla Link	48860
Version	11.0
OS	Windows NT
CC	@CaseyCarter,@zygoloid

Extended Description

MSVC recently implemented the following intrinsics needed for P0466R5 Layout-compatibility and Pointer-interconvertibility Traits:

• __builtin_is_layout_compatible
• __builtin_is_pointer_interconvertible_base_of
• __builtin_is_pointer_interconvertible_with_class
• __builtin_is_corresponding_member

MSVC STL is planning to use them in <type_traits> as follows (PR: microsoft/STL#1575):

// STRUCT TEMPLATE is_layout_compatible
template <class _Ty1, class _Ty2>
struct is_layout_compatible : bool_constant<__builtin_is_layout_compatible(_Ty1, _Ty2)> {};

template <class _Ty1, class _Ty2>
inline constexpr bool is_layout_compatible_v = __builtin_is_layout_compatible(_Ty1, _Ty2);

// STRUCT TEMPLATE is_pointer_interconvertible_base_of
template <class _Base, class _Derived>
struct is_pointer_interconvertible_base_of
    : bool_constant<__builtin_is_pointer_interconvertible_base_of(_Base, _Derived)> {};

template <class _Base, class _Derived>
inline constexpr bool is_pointer_interconvertible_base_of_v = __builtin_is_pointer_interconvertible_base_of(
    _Base, _Derived);

// FUNCTION TEMPLATE is_pointer_interconvertible_with_class
template <class _ClassTy, class _MemberTy>
_NODISCARD constexpr bool is_pointer_interconvertible_with_class(_MemberTy _ClassTy::*_Pm) noexcept {
    return __builtin_is_pointer_interconvertible_with_class(_ClassTy, _Pm);
}

// FUNCTION TEMPLATE is_corresponding_member
template <class _ClassTy1, class _ClassTy2, class _MemberTy1, class _MemberTy2>
_NODISCARD constexpr bool is_corresponding_member(_MemberTy1 _ClassTy1::*_Pm1, _MemberTy2 _ClassTy2::*_Pm2) noexcept {
    return __builtin_is_corresponding_member(_ClassTy1, _ClassTy2, _Pm1, _Pm2);

[zygoloid]

Do we know why MSVC has deviated from the longstanding rule that compiler intrinsic forms of type traits are named by prefixing simply __ to the name? (Eg, __is_layout_compatible, __is_pointer_interconvertible_base_of, __is_pointer_interconvertible_with_class, __is_corresponding_member.)

In particular, if MSVC is avoiding name collisions with something, will we need to avoid adding the normal name in addition to the __builtin_ name?

[MahmoudGSaleh]

Do we know why MSVC has deviated from the longstanding rule that compiler
intrinsic forms of type traits are named by prefixing simply __ to the name?
(Eg, __is_layout_compatible, __is_pointer_interconvertible_base_of,
__is_pointer_interconvertible_with_class, __is_corresponding_member.)

In particular, if MSVC is avoiding name collisions with something, will we
need to avoid adding the normal name in addition to the _builtin name?

Agree and thanks for catching this. In the past we've used the __builtin* form; e.g. __builtin_zero_non_value_bits, however not in <type_traits>. We will rename the intrinsics to follow the __is_* form instead. The change should appear in Visual Studio 2019 version 16.10 preview 1.
I will update this bug report once those renames are shipped.

[MahmoudGSaleh]

Update: The C++ toolset update renaming the intrinsics (removing the __builtin prefix) has shipped with VS 2019 preview 1.

MSVC STL moved to the new names as well: https://github.com/microsoft/STL/pull/1711/files

[MahmoudGSaleh]

With the latest MSVC update, the intrinsic names are:

• __is_layout_compatible
• __is_pointer_interconvertible_base_of
• __is_pointer_interconvertible_with_class
• __is_corresponding_member

[ldionne]

GCC also does this. It would be great to implement this in Clang so that libc++ can start using the traits.

[cor3ntin]

@MahmoudGSaleh how should msvc::no_unique_address behave for the purpose of layout compatibility, in the context of https://cplusplus.github.io/CWG/issues/2759.html ?