[libc++] Fix tuple_cat for element with unconstrained constructor

libcxx/include/tuple

@@ -1338,12 +1338,25 @@ struct __tuple_cat<tuple<_Types...>, __tuple_indices<_I0...>, __tuple_indices<_J
   }
 };
 
+template <class _TupleDst, class _TupleSrc, size_t... _Indices>
+inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 _TupleDst
+__tuple_cat_select_element_wise(_TupleSrc&& __src, __tuple_indices<_Indices...>) {
+  static_assert(tuple_size<_TupleDst>::value == tuple_size<_TupleSrc>::value,
+                "misuse of __tuple_cat_select_element_wise with tuples of different sizes");
+  return _TupleDst(std::get<_Indices>(std::forward<_TupleSrc>(__src))...);
+}
+
 template <class _Tuple0, class... _Tuples>
 inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 typename __tuple_cat_return<_Tuple0, _Tuples...>::type
 tuple_cat(_Tuple0&& __t0, _Tuples&&... __tpls) {
-  using _T0 _LIBCPP_NODEBUG = __libcpp_remove_reference_t<_Tuple0>;
-  return __tuple_cat<tuple<>, __tuple_indices<>, typename __make_tuple_indices<tuple_size<_T0>::value>::type>()(
-      tuple<>(), std::forward<_Tuple0>(__t0), std::forward<_Tuples>(__tpls)...);
+  using _T0 _LIBCPP_NODEBUG          = __libcpp_remove_reference_t<_Tuple0>;
+  using _TRet _LIBCPP_NODEBUG        = typename __tuple_cat_return<_Tuple0, _Tuples...>::type;
+  using _T0Indices _LIBCPP_NODEBUG   = typename __make_tuple_indices<tuple_size<_T0>::value>::type;
+  using _TRetIndices _LIBCPP_NODEBUG = typename __make_tuple_indices<tuple_size<_TRet>::value>::type;
+  return std::__tuple_cat_select_element_wise<_TRet>(
+      __tuple_cat<tuple<>, __tuple_indices<>, _T0Indices>()(
+          tuple<>(), std::forward<_Tuple0>(__t0), std::forward<_Tuples>(__tpls)...),
+      _TRetIndices());
 }
 
 template <class... _Tp, class _Alloc>

libcxx/test/std/utilities/tuple/tuple.tuple/tuple.creation/tuple_cat.pass.cpp

@@ -31,6 +31,48 @@ template<typename ...Ts>
 void forward_as_tuple(Ts...) = delete;
 }
 
+// https://github.com/llvm/llvm-project/issues/41034
+struct Unconstrained {
+  int data;
+  template <typename Arg>
+  TEST_CONSTEXPR_CXX14 Unconstrained(Arg arg) : data(arg) {}
+};
+
+TEST_CONSTEXPR_CXX14 bool test_tuple_cat_with_unconstrained_constructor() {
+  {
+    auto tup_src = std::tuple<Unconstrained>(Unconstrained(5));
+    auto tup     = std::tuple_cat(tup_src);
+    assert(std::get<0>(tup).data == 5);
+  }
+  {
+    auto tup = std::tuple_cat(std::tuple<Unconstrained>(Unconstrained(6)));
+    assert(std::get<0>(tup).data == 6);
+  }
+  {
+    auto tup = std::tuple_cat(std::tuple<Unconstrained>(Unconstrained(7)), std::tuple<>());
+    assert(std::get<0>(tup).data == 7);
+  }
+#if TEST_STD_VER >= 17
+  {
+    auto tup_src = std::tuple(Unconstrained(8));
+    auto tup     = std::tuple_cat(tup_src);
+    ASSERT_SAME_TYPE(decltype(tup), std::tuple<Unconstrained>);
+    assert(std::get<0>(tup).data == 8);
+  }
+  {
+    auto tup = std::tuple_cat(std::tuple(Unconstrained(9)));
+    ASSERT_SAME_TYPE(decltype(tup), std::tuple<Unconstrained>);
+    assert(std::get<0>(tup).data == 9);
+  }
+  {
+    auto tup = std::tuple_cat(std::tuple(Unconstrained(10)), std::tuple());
+    ASSERT_SAME_TYPE(decltype(tup), std::tuple<Unconstrained>);
+    assert(std::get<0>(tup).data == 10);
+  }
+#endif
+  return true;
+}
+
 int main(int, char**)
 {
     {
@@ -270,5 +312,13 @@ int main(int, char**)
         assert(std::get<0>(t).i == 1);
         assert(std::get<0>(t2).i == 1);
     }
-  return 0;
+    // See https://github.com/llvm/llvm-project/issues/41034
+    {
+      test_tuple_cat_with_unconstrained_constructor();
+#if TEST_STD_VER >= 14
+      static_assert(test_tuple_cat_with_unconstrained_constructor(), "");
+#endif
+    }
+
+    return 0;
 }