On Tue, 5 Dec 2023, Jonathan Wakely wrote:
> On Wed, 22 Nov 2023 at 14:50, Jonathan Wakely <jwak...@redhat.com> wrote:
> >
> > On Mon, 20 Nov 2023 at 02:56, Jason Merrill wrote:
> > >
> > > Tested x86_64-pc-linux-gnu. Are the library bits OK? Any comments
> > > before I
> > > push this?
> >
> > The library parts are OK.
> >
> > The variable template is_trivially_copyable_v just uses
> > __is_trivially_copyable so should be just as efficient, and the change
> > to <bit> is fine.
> >
> > The variable template is_trivially_destructible_v instantiates the
> > is_trivially_destructible type trait, which instantiates
> > __is_destructible_safe and __is_destructible_impl, which is probably
> > why we used the built-in directly in <variant>. But that's an
> > acceptable overhead to avoid using the built-in in a mangled context,
> > and it would be good to optimize the variable template anyway, as a
> > separate change.
>
> This actually causes a regression:
>
> FAIL: 20_util/variant/87619.cc -std=gnu++20 (test for excess errors)
> FAIL: 20_util/variant/87619.cc -std=gnu++23 (test for excess errors)
> FAIL: 20_util/variant/87619.cc -std=gnu++26 (test for excess errors)
>
> It's OK for C++17 because the changed code is only used for C++20 and later.
>
> That test instantiates a very large variant to check that we don't hit
> our template instantiation depth limit. Using the variable template
> (which uses the class template) instead of the built-in causes it to
> fail now.
Could we pass down __trivially_destructible from _Variadic_storage to
_Variadic_union and use that as the dtor's constraint instead of
recursively re-computing it? This reduces the maximum template
instantiation depth for 87619.cc to ~270 from ~780 so that the depth is
roughly #variants rather than 4 * #variants.
diff --git a/libstdc++-v3/include/std/variant b/libstdc++-v3/include/std/variant
index 20a76c8aa87..4b9002e0917 100644
--- a/libstdc++-v3/include/std/variant
+++ b/libstdc++-v3/include/std/variant
@@ -392,7 +392,7 @@ namespace __variant
};
// Defines members and ctors.
- template<typename... _Types>
+ template<bool __trivially_destructible, typename... _Types>
union _Variadic_union
{
_Variadic_union() = default;
@@ -401,8 +401,8 @@ namespace __variant
_Variadic_union(in_place_index_t<_Np>, _Args&&...) = delete;
};
- template<typename _First, typename... _Rest>
- union _Variadic_union<_First, _Rest...>
+ template<bool __trivially_destructible, typename _First, typename... _Rest>
+ union _Variadic_union<__trivially_destructible, _First, _Rest...>
{
constexpr _Variadic_union() : _M_rest() { }
@@ -427,13 +427,12 @@ namespace __variant
~_Variadic_union() = default;
constexpr ~_Variadic_union()
- requires (!is_trivially_destructible_v<_First>)
- || (!is_trivially_destructible_v<_Variadic_union<_Rest...>>)
+ requires (!__trivially_destructible)
{ }
#endif
_Uninitialized<_First> _M_first;
- _Variadic_union<_Rest...> _M_rest;
+ _Variadic_union<__trivially_destructible, _Rest...> _M_rest;
};
// _Never_valueless_alt is true for variant alternatives that can
@@ -514,7 +513,7 @@ namespace __variant
return this->_M_index != __index_type(variant_npos);
}
- _Variadic_union<_Types...> _M_u;
+ _Variadic_union<false, _Types...> _M_u;
using __index_type = __select_index<_Types...>;
__index_type _M_index;
};
@@ -552,7 +551,7 @@ namespace __variant
return this->_M_index != static_cast<__index_type>(variant_npos);
}
- _Variadic_union<_Types...> _M_u;
+ _Variadic_union<true, _Types...> _M_u;
using __index_type = __select_index<_Types...>;
__index_type _M_index;
};
>
> So optimizing the variable template is now a priority.
>
>
