[BUG]: libcudacxx clashes w/ libc++: ambiguous overload resolution for `__swallow`

[Artem-B]

Is this a duplicate?

• [X] I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Compile-time Error

Component

libcu++

Describe the bug

https://godbolt.org/z/47xKeqKnM

Compilation fails because of an ambiguous overload resolution caused by both libcudacxx and libc++ providing overloads for __swallow:

/opt/compiler-explorer/clang-trunk-20240430/bin/../include/c++/v1/tuple:439:58: note: candidate function [with _Tp = <cuda::std::__tuple_leaf<0, std::__wrap_iter<int *>> &, cuda::std::__tuple_leaf<1, std::__wrap_iter<int *>> &>]
  439 | _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 void __swallow(_Tp&&...) _NOEXCEPT {}
      |                                                          ^
/opt/compiler-explorer/libs/cccl/trunk/libcudacxx/include/cuda/std/detail/libcxx/include/tuple:427:36: note: candidate function [with _Tp = <cuda::std::__tuple_leaf<0, std::__wrap_iter<int *>> &, cuda::std::__tuple_leaf<1, std::__wrap_iter<int *>> &>]
  427 | _LIBCUDACXX_INLINE_VISIBILITY void __swallow(_Tp &&...) noexcept {}
      |                                    ^

How to Reproduce

Reproducer on godbolt: https://godbolt.org/z/47xKeqKnM

Expected behavior

No conflicts with libc++

Reproduction link

https://godbolt.org/z/47xKeqKnM

Operating System

Debian/testing

nvidia-smi output

N/a

NVCC version

NVCC does not support libc++.

The issue affects clang + libc++ only.

[Artem-B]

Looks like libcudacxx is missing this fix from libc++: https://reviews.llvm.org/D50106

Basically calls to __swallow need to be done with fully qualified name. I guess in libcudacxx it means _CUDA_VSTD::. The diff below appears to fix this issue.

diff --git a/libcudacxx/include/cuda/std/detail/libcxx/include/tuple b/libcudacxx/include/cuda/std/detail/libcxx/include/tuple
index 22cb1d51f..cc2682931 100644
--- a/libcudacxx/include/cuda/std/detail/libcxx/include/tuple
+++ b/libcudacxx/include/cuda/std/detail/libcxx/include/tuple
@@ -494,7 +494,7 @@ struct _LIBCUDACXX_DECLSPEC_EMPTY_BASES __tuple_impl<__tuple_indices<_Indx...>,
   _LIBCUDACXX_INLINE_VISIBILITY __tuple_impl& operator=(_Tuple&& __t) noexcept(
     (__all<_LIBCUDACXX_TRAIT(is_nothrow_assignable, _Tp&, __tuple_elem_at<_Tuple, _Indx>)...>::value))
   {
-    __swallow(__tuple_leaf<_Indx, _Tp>::operator=(
+    _CUDA_VSTD::__swallow(__tuple_leaf<_Indx, _Tp>::operator=(
       _CUDA_VSTD::forward<__tuple_elem_at<_Tuple, _Indx>>(_CUDA_VSTD::get<_Indx>(__t)))...);
     return *this;
   }
@@ -505,14 +505,14 @@ struct _LIBCUDACXX_DECLSPEC_EMPTY_BASES __tuple_impl<__tuple_indices<_Indx...>,
   _LIBCUDACXX_INLINE_VISIBILITY __tuple_impl&
   operator=(const __tuple_impl& __t) noexcept((__all<_LIBCUDACXX_TRAIT(is_nothrow_copy_assignable, _Tp)...>::value))
   {
-    __swallow(__tuple_leaf<_Indx, _Tp>::operator=(static_cast<const __tuple_leaf<_Indx, _Tp>&>(__t).get())...);
+    _CUDA_VSTD::__swallow(__tuple_leaf<_Indx, _Tp>::operator=(static_cast<const __tuple_leaf<_Indx, _Tp>&>(__t).get())...);
     return *this;
   }

   _LIBCUDACXX_INLINE_VISIBILITY __tuple_impl&
   operator=(__tuple_impl&& __t) noexcept((__all<_LIBCUDACXX_TRAIT(is_nothrow_move_assignable, _Tp)...>::value))
   {
-    __swallow(__tuple_leaf<_Indx, _Tp>::operator=(
+    _CUDA_VSTD::__swallow(__tuple_leaf<_Indx, _Tp>::operator=(
       _CUDA_VSTD::forward<_Tp>(static_cast<__tuple_leaf<_Indx, _Tp>&>(__t).get()))...);
     return *this;
   }
@@ -520,7 +520,7 @@ struct _LIBCUDACXX_DECLSPEC_EMPTY_BASES __tuple_impl<__tuple_indices<_Indx...>,
   _LIBCUDACXX_INLINE_VISIBILITY void
   swap(__tuple_impl& __t) noexcept(__all<__is_nothrow_swappable<_Tp>::value...>::value)
   {
-    __swallow(__tuple_leaf<_Indx, _Tp>::swap(static_cast<__tuple_leaf<_Indx, _Tp>&>(__t))...);
+    _CUDA_VSTD::__swallow(__tuple_leaf<_Indx, _Tp>::swap(static_cast<__tuple_leaf<_Indx, _Tp>&>(__t))...);
   }
 };

[Artem-B]

@jrhemstad, @miscco ^^^

[miscco]

I adopted this in #1685