[libc++] Use __is_convertible built-in when available

[mcgrathr]

Summary

https://github.com/llvm/llvm-project/issues/62396 reports that
GCC 13 barfs on parsing <type_traits> because of the declarations
of struct __is_convertible. In GCC 13, __is_convertible is a
built-in, but __is_convertible_to is not. Clang has both, so
using either should be fine.

[philnik]

LGTM % comment with a green CI.

libcxx/include/__type_traits/is_convertible.h

27–32

Please add a //TODO: Remove the #else branch once GCC 13 is used in the CI.

[mcgrathr]

libcxx/include/__type_traits/is_convertible.h

27–32

I'm not clear on what the TODO intends to be done later.
Are you saying that you'll remove the fallback for the case of no built-in at all?
I imagine that would break compatibility with many extant GCC versions that don't have either built-in, wouldn't it?

[philnik]

libcxx/include/__type_traits/is_convertible.h

27–32

Yes, that's exactly what I'm saying. We only support the latest released GCC and the last two Clang versions and the one in the same release. People mostly use very recent compiler versions with libc++. Supporting older compilers would be a lot of work for very little benefit.

[mcgrathr]

Added TODO comment

libcxx/include/__type_traits/is_convertible.h

27–32

Understood.

[shafik]

I think this should have a release note.

[philnik]

In D149313#4320325, @shafik wrote:

I think this should have a release note.

Why? We don't support GCC 13 yet.

[shafik]

In D149313#4320424, @philnik wrote:

In D149313#4320325, @shafik wrote:

I think this should have a release note.

Why? We don't support GCC 13 yet.

I must have misunderstood, does this not fix the problem in https://github.com/llvm/llvm-project/issues/62396 even if there is follow-up work todo?

[philnik]

In D149313#4320676, @shafik wrote:

In D149313#4320424, @philnik wrote:

In D149313#4320325, @shafik wrote:

I think this should have a release note.

Why? We don't support GCC 13 yet.

I must have misunderstood, does this not fix the problem in https://github.com/llvm/llvm-project/issues/62396 even if there is follow-up work todo?

Yes, it fixes the problem. But I don't see why we would add a release note. We don't support GCC 13 yet, so it's not unexpected that things break.

libcxx/include/__type_traits/is_convertible.h

 #include <cstddef>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
 _LIBCPP_BEGIN_NAMESPACE_STD
 
-#if __has_builtin(__is_convertible_to) && !defined(_LIBCPP_USE_IS_CONVERTIBLE_FALLBACK)
+#if __has_builtin(__is_convertible) && !defined(_LIBCPP_USE_IS_CONVERTIBLE_FALLBACK)
+
+template <class _T1, class _T2>
+struct _LIBCPP_TEMPLATE_VIS is_convertible : public integral_constant<bool, __is_convertible(_T1, _T2)> {};
+
+#elif __has_builtin(__is_convertible_to) && !defined(_LIBCPP_USE_IS_CONVERTIBLE_FALLBACK)
 
 template <class _T1, class _T2> struct _LIBCPP_TEMPLATE_VIS is_convertible
     : public integral_constant<bool, __is_convertible_to(_T1, _T2)> {};
 
+// TODO: Remove this fallback when GCC < 13 support is no longer required.
+// GCC 13 has the __is_convertible built-in.
 #else  // __has_builtin(__is_convertible_to) && !defined(_LIBCPP_USE_IS_CONVERTIBLE_FALLBACK)
 
 namespace __is_convertible_imp
 {
 template <class _Tp> void  __test_convert(_Tp);
 
 template <class _From, class _To, class = void>
 struct __is_convertible_test : public false_type {};