try to molify nvc++ and gcc at the same time

Author: ericniebler

libcudacxx/include/cuda/__utility/__basic_any/basic_any_fwd.h

@@ -113,7 +113,7 @@ enum class __vtable_kind : uint8_t
   __rtti,
 };
 
-inline constexpr uint8_t __basic_any_version = 0;
+inline constexpr uint8_t __basic_any_version = 1;
 
 template <class _Interface>
 extern _Interface __remove_ireference_v; // specialized in interfaces.cuh

libcudacxx/include/cuda/__utility/__basic_any/iset.h

@@ -115,19 +115,19 @@ struct __iset_vptr : __base_vptr
   __iset_vptr() = default;
 
   _CCCL_API constexpr __iset_vptr(__iset_vtable const* __vptr) noexcept
-      : __base_vptr(__vptr)
+      : __base_vptr{__vptr}
   {}
 
   _CCCL_API constexpr __iset_vptr(__base_vptr __vptr) noexcept
-      : __base_vptr(__vptr)
+      : __base_vptr{__vptr}
   {}
 
   // Permit narrowing conversions from a super-set __vptr. Warning: we can't
   // simply constrain this because then the ctor from __base_vptr would be
   // selected instead, giving the wrong result.
   template <class... _Others>
   _CCCL_API __iset_vptr(__iset_vptr<_Others...> __vptr) noexcept
-      : __base_vptr(__vptr->__query_interface(__iunknown()))
+      : __base_vptr{__vptr->__query_interface(__iunknown())}
   {
     static_assert(::cuda::std::__type_set_contains_v<::cuda::std::__make_type_set<_Others...>, _Interfaces...>, "");
     _CCCL_ASSERT(__vptr_->__kind_ == __vtable_kind::__rtti && __vptr_->__cookie_ == 0xDEADBEEF,

libcudacxx/include/cuda/__utility/__basic_any/rtti.h

@@ -94,8 +94,25 @@ static_assert(sizeof(__rtti_base) == sizeof(uint64_t) + sizeof(void*));
 // Used to map an interface typeid to a pointer to the vtable for that interface.
 struct __base_info
 {
+  using __cast_fn_t = auto(__rtti const*) noexcept -> __base_vptr;
+
   ::cuda::std::__type_info_ptr __typeid_;
-  __base_vptr __vptr_;
+  union
+  {
+    __cast_fn_t* __cast_fn_; // used when __basic_any_version >= 1,
+    __base_vptr __vptr_v0_; // used when __basic_any_version == 0
+  };
+
+  [[nodiscard]] _CCCL_API auto __get_vptr(__rtti const* __rtti_ptr, uint8_t __version) const noexcept -> __base_vptr
+  {
+    return __version >= 1 ? __cast_fn_(__rtti_ptr) : __vptr_v0_;
+  }
+
+  template <class _VTable, class _Interface>
+  [[nodiscard]] _CCCL_API static auto __cast_fn_impl(__rtti const* __rtti_ptr) noexcept -> __base_vptr
+  {
+    return {static_cast<__vptr_for<_Interface>>(static_cast<_VTable const*>(__rtti_ptr))};
+  }
 };
 
 inline constexpr size_t __half_size_t_bits = sizeof(size_t) * CHAR_BIT / 2;
@@ -164,35 +181,34 @@ struct __rtti : __rtti_base
     {
       if (&__id == __base_vptr_map_[__i].__typeid_)
       {
-        return static_cast<__vptr_for<_Interface>>(__base_vptr_map_[__i].__vptr_);
+        return static_cast<__vptr_for<_Interface>>(__base_vptr_map_[__i].__get_vptr(this, __version_));
       }
     }
 
     for (size_t __i = 0; __i < __nbr_interfaces_; ++__i)
     {
       if (__id == *__base_vptr_map_[__i].__typeid_)
       {
-        return static_cast<__vptr_for<_Interface>>(__base_vptr_map_[__i].__vptr_);
+        return static_cast<__vptr_for<_Interface>>(__base_vptr_map_[__i].__get_vptr(this, __version_));
       }
     }
 
     return nullptr;
   }
 
-  void (*__dtor_)(void*, bool) noexcept;
-  __object_metadata const* __object_info_;
+  void (*__dtor_)(void*, bool) noexcept            = nullptr;
+  __object_metadata const* __object_info_          = nullptr;
   ::cuda::std::__type_info_ptr __interface_typeid_ = nullptr;
-  __base_info const* __base_vptr_map_;
+  __base_info const* __base_vptr_map_              = nullptr;
 };
 
 template <size_t _NbrInterfaces>
 struct __rtti_ex : __rtti
 {
-  template <class _Tp, class _Super, class... _Interfaces, class _VPtr>
-  _CCCL_API constexpr __rtti_ex(__tag<_Tp, _Super> __type, __tag<_Interfaces...> __ibases, _VPtr) noexcept
+  template <class _Tp, class _Super, class... _Interfaces, class _VTable>
+  _CCCL_API constexpr __rtti_ex(__tag<_Tp, _Super> __type, __tag<_Interfaces...> __ibases, _VTable const*) noexcept
       : __rtti{__type, __ibases, __base_vptr_array}
-      , __base_vptr_array{
-          {&_CCCL_TYPEID(_Interfaces), static_cast<__vptr_for<_Interfaces>>(static_cast<_VPtr>(this))}...}
+      , __base_vptr_array{{&_CCCL_TYPEID(_Interfaces), {&__base_info::__cast_fn_impl<_VTable, _Interfaces>}}...}
   {}
 
   __base_info __base_vptr_array[_NbrInterfaces];
@@ -225,8 +241,8 @@ template <class _SrcInterface, class _DstInterface>
   else
   {
     //! Slow down-casts and cross-casts:
-    __rtti const* rtti = __src_vptr->__query_interface(__iunknown());
-    return rtti->__query_interface(_DstInterface());
+    __rtti const* __rtti_ptr = __src_vptr->__query_interface(__iunknown());
+    return __rtti_ptr->__query_interface(_DstInterface());
   }
 }
 

libcudacxx/include/cuda/__utility/__basic_any/virtual_ptrs.h

@@ -32,12 +32,6 @@ _CCCL_BEGIN_NAMESPACE_CUDA
 
 struct __base_vptr
 {
-  __base_vptr() = default;
-
-  _CCCL_NODEBUG_API constexpr __base_vptr(__rtti_base const* __vptr) noexcept
-      : __vptr_(__vptr)
-  {}
-
   template <class _VTable>
   [[nodiscard]] _CCCL_NODEBUG_API explicit constexpr operator _VTable const*() const noexcept
   {