[libcxx-commits] [libcxx] [libc++] cv-qualified types in atomic and atomic_ref (P3323R1) (PR #121414)

[Louis Dionne via libcxx-commits]

================
@@ -251,17 +284,24 @@ struct atomic_ref : public __atomic_ref_base<_Tp> {
 
   _LIBCPP_HIDE_FROM_ABI atomic_ref(const atomic_ref&) noexcept = default;
 
-  _LIBCPP_HIDE_FROM_ABI _Tp operator=(_Tp __desired) const noexcept { return __base::operator=(__desired); }
+  _LIBCPP_HIDE_FROM_ABI __base::value_type operator=(__base::value_type __desired) const noexcept
+    requires(!is_const_v<_Tp>)
+  {
+    return __base::operator=(__desired);
+  }
 
   atomic_ref& operator=(const atomic_ref&) = delete;
 };
 
 template <class _Tp>
-  requires(std::integral<_Tp> && !std::same_as<bool, _Tp>)
+  requires(integral<_Tp> && !same_as<bool, __remove_cv_t<_Tp>>)
 struct atomic_ref<_Tp> : public __atomic_ref_base<_Tp> {
   using __base _LIBCPP_NODEBUG = __atomic_ref_base<_Tp>;
 
+  static_assert(__base::is_always_lock_free || !is_volatile_v<_Tp>);
+
   using difference_type = __base::value_type;
+  using value_type      = __base::value_type;
----------------
ldionne wrote:

I would get rid of this one and the one for `floating_point` since this is already provided in `__atomic_ref_impl`. This makes it clearer that we're only providing the typedefs that we actually need to fulfill the public API, which in this case is `difference_type`.

If you have trouble referring to `value_type` (I think you might need something like `typename atomic_ref::value_type` which is annoying), you could always do

```
using typename __base::value_type;
```

IIRC that's legal and that makes it clear that we're only bringing the name that already exists into the current scope instead of introducing a new name.

https://github.com/llvm/llvm-project/pull/121414