Beautiful Type Erasure with C++26 Reflection
Beautiful Type Erasure with C++26 Reflection
If you've ever tried using type erasure for something more complicated than std::any or std::function, you've either written 100+ lines of easy-to-mess-up code or reached for a boilerplate-heavy library like Boost.TypeErasure or Folly.Poly.
rjk::duck uses the magic that is C++26 reflection to remove these pain points while preserving all of the customization and performance.
Consider the following basic example:
#include // ...
struct [[=rjk::trait]] Container {
auto size() const -> std::size_t;
auto empty() const -> bool;
auto clear() -> void;
};
rjk::duck c{std::vector{1, 2, 3}};
c.size(); // 3
c = std::string{"hello"}; // swap underlying type at runtime
c.size(); // 5
c = std::map{{1, 2}, {3, 4}};
c.empty(); // false
c.clear();
c.empty(); // true
Simply declare the interface once and let your already-written definitions handle the rest. The library itself is a single header include. It offers owning and non-owning semantics, operators, interface composition, adapters for existing interfaces, extension methods for third-party types, and much more.
We're talking about the bleeding edge here, so right now support is only available for gcc with -std=c++26 -freflection. duck uses reflection in some unique ways that go beyond the simple enum-to-string or JSON serialization examples you may have seen, so we'll spend the rest of the article demystifying the tricks that make this library possible. In particular, we'll walk through tag generation, vtable codegen, overload resolution, and the interconvertibility trick that keeps duck small.
A Brief Intro to C++26 Reflection
You may have noticed the strange bit of syntax from the above example, [[=rjk::trait]]. This is a C++26 annotation, which can be applied to a struct similarly to an attribute. The actual definition for trait is simply:
constexpr inline struct{} trait{};
We can check if a type has the trait annotation like this:
return std::ranges::any_of(
annotations_of(^^MyType),
[](std::meta::info annotation) {
return type_of(annotation) == type_of(^^trait);
}
);
The ^^ operator produces a reflection of something. In this case, we are reflecting both a type (MyType) and a variable (trait). Reflections all have the std::meta::info type and can be queried with a variety of meta-functions, like annotations_of and type_of.
One of the first steps in duck's process is interpreting the members of a trait and converting them to a tag, which is a format that duck uses internally. So for this trait:
struct [[=rjk::trait]] MyTrait {
auto foo() -> void;
auto bar() const -> int;
};
Our goal is to generate a has_fn<void()> and has_fn<int() const> tag. We can implement this pretty easily by inspecting the members of MyTrait and transforming them, like so:
consteval auto members_to_tags(std::meta::info trait) -> std::vector<std::meta::info> {
const auto ctx = std::meta::access_context::unprivileged(); // Only check public members
return members_of(trait, ctx) // Take all members of the trait
| std::views::filter(std::meta::is_user_declared) // Exclude constructors, etc.
| std::views::filter(std::meta::is_function) // Exclude data members
| std::views::filter(std::meta::has_identifier) // Exclude operator functions, etc.
| std::views::transform([](std::meta::info member) {
// identifier_of returns the name as a string_view. fixed_string is a
// custom structural type that can be used as a template argument.
const fixed_string name{identifier_of(member)};
const auto signature = type_of(member); // Returns a function type
// Create has_fn<name, signature>
const auto tag = substitute(^^has_fn, {reflect_constant(name), signature});
return tag;
})
| std::ranges::to<std::vector>();
}
The actual implementation has to handle a variety of other aspects, such as iterating base classes for traits, handling const traits, operators, and more. But the core transformation is as simple as it looks.
Generating a Vtable
The mechanism for code generation we got in C++26 is limited, but still very powerful. Let's walk through each component of how the vtable is generated, starting with creating the struct:
template <typename... Traits>
struct vtable_generator {
struct vtable;
// ...
consteval {
std::vector<std::meta::info> members{ /* typeid, copy, move, destroy... */ };
constexpr static std::array traits{^^Traits...};
template for (constexpr auto trait : traits) {
for (const auto tag : members_to_tags(trait)) {
const auto args = template_arguments_of(tag);
const auto name = extract<fixed_string>(args[0]);
// Remove cvref qualifiers from the function
const auto func_type = remove_fn_qualifiers(args[1]);
const auto signature = add_pointer(prepend_arg(func_type, ^^void*));
const auto member = data_member_spec(signature, {.name = name});
members.push_back(member);
}
}
define_aggregate(^^vtable, members);
}
};
The consteval block is a new feature and is the only context in which we can currently generate code, using define_aggregate. template for is the new syntax for an expansion statement, which lets us iterate over parameter packs without having to rely on fold expressions.
Other than that, the code is pretty straightforward. We simply collect all of the traits, then collect all of their respective tags, and generate function pointers for each member function in the traits.
One simplification worth acknowledging is that this approach can't handle overloads, since you can't have two members with the same name. The actual code will assign names like slot_0, slot_1, etc. and re-derive them later. We're also just using plain void* here, but in reality we need to also make this const void* based on the function's qualifier.
Creating a static vtable for a type is likewise not too complicated:
// still inside vtable_generator<Traits...>
template <typename T>
consteval static auto make_vtable() -> vtable {
constexpr static auto ctx = std::meta::access_context::unprivileged();
constexpr static auto slots = define_static_array(
nonstatic_data_members_of(^^vtable, ctx)
| std::views::drop(/* copy, move, etc. */)
);
vtable result{};
template for (constexpr auto index : std::views::indices(slots.size())) {
constexpr auto T_member = *std::ranges::find_if(
members_of(^^T, ctx),
[](std::meta::info member) {
return is_function(member)
&& has_identifier(member)
&& identifier_of(member) == identifier_of(slots[index])
&& type_of(member) == type_of(slots[index]);
}
);
result.[: slots[index] :] = convert_to_vtable_func(T_member);
}
return result;
}
template <typename T>
constexpr static auto vtable_for<T> = make_vtable<T>();
Real overload resolution for finding T_member is significantly more involved (see below). We're matching by exact signature here for clarity.
The splice operator ([: expr :]) is another new feature, and it's what brings your code from the land of reflection back into reality. Here, we use it to actually assign to each of the function pointers that were added to the generated vtable.
std::views::indices is a nice library utility we got as well, which is simply equivalent to std::views::iota(0, upper_bound).
The one loose thread remaining is convert_to_vtable_func. Somehow, we need to turn a T_member into an actual auto(*)(void*, Args...) -> Ret that we can store and call through. That conversion is where duck's type erasure actually happens, so it's worth taking apart properly.
From Slot to Call
At its core, this is the same trick any type erasure library uses.
template <typename T, typename Ret, typename... Args, auto Invoker>
struct vtable_fn_maker {
constexpr static auto erased_call(void* self, Args... args) -> Ret {
auto* typed = static_cast<T*>(self);
return std::invoke(Invoker{}, *typed, std::forward<Args>(args)...);
}
};
convert_to_vtable_func (roughly) will end up generating a pointer to this erased_call function, which ultimately gets stored in the vtable.
The Invoker here is suspicious, and doesn't match the T_member we saw from the previous section. Previously, we just looked for an exact function match, but actual overload resolution can't be that simple. As it turns out, duck doesn't try to reimplement it by hand.
Making the Call
The reader accustomed to C++17 might be familiar with this common utility:
template <typename... Callables>
struct overload_set : Callables... {
using Callables::operator()...;
};
Rather than trying to manually reproduce C++ overload resolution, we instead generate a callable overload_set and simply let the language do the work. The actual mechanism relies on two parts.
First, candidate_wrapper:
template <typename Self, auto Member, typename... Args>
struct candidate_wrapper {
constexpr decltype(auto) operator()(Self self, Args... args) const {
return std::invoke(&[:Member:], std::forward<Self>(self), std::forward<Args>(args)...);
}
};
This is a simple wrapper class that takes in some member function from Self, splices it (&[:Member:]) to obtain a member function pointer, and then invokes it with the given type and arguments. The key detail is that this turns any myObj.foo(args...) call into a myWrapper(myObj, args...) call, which can then get substituted into overload_set like so:
consteval auto make_set(std::meta::info type, std::string_view identifier) -> std::meta::info {
const auto members = members_of(type, ctx); // ctx is still a std::meta::access_context
const auto candidates = members
| std::views::filter(std::meta::is_function)
| std::views::filter(std::meta::has_identifier)
| std::views::filter([=](std::meta::info member) {
return identifier_of(member) == identifier;
})
| std::views::transform([=](std::meta::info member) {
// Check the qualifiers and create a const T&, T&&, etc.
const auto self_type = self_type_of(member);
const auto params = parameters_of(member);
const std::array member_and_self{reflect_constant(member), self_type};
const auto args = std::views::concat(
member_and_self,
params | std::views::transform(std::meta::type_of)
);
return substitute(^^candidate_wrapper, args);
});
return substitute(^^overload_set, candidates);
}
There are a couple of subtleties that come with getting this right:
self_type_ofwill query the qualifiers of the member function and determine the qualifiers to apply to theSelftemplate parameterstd::views::concatis a new utility that lets us cheaply concatenate two ranges, without having to materialize them into an actualstd::vectorparamsgets a small transformation to convert the parameter reflections into type reflections
But the payoff is that overload resolution now works cleanly:
struct [[=rjk::trait]] MyTrait {
auto foo(int) -> int;
};
struct MyStruct {
auto foo(double) -> int;
auto foo(int) const -> int;
};
make_set(^^MyStruct, "foo") will collect both of the overloads, and then we will attempt to call them on a MyStruct& and MyStruct&& with an int as an argument. This will match the auto foo(int) const -> int signature. So even though MyStruct doesn't literally define a mutable foo member function that accepts an int, it can still match MyTrait.
The T_member search from make_vtable was the simplification. In reality, each vtable slot is filled with a vtable_fn_maker whose Invoker is our compile-time generated overload_set. Resolution happens right inside of erased_call, handled entirely by the compiler.
All of this gets us a fully populated vtable_for<T>, but it doesn't answer the question of how we can get the clean myDuck.foo(5) syntax we're looking for. Somehow, duck needs to grow a callable member, one for each trait member function.
Growing an Interface
Reflection doesn't let us inject member functions yet, so the best we can do is starting with a simple wrapper type which our duck class can then inherit from. So for the following simple trait:
struct [[=rjk::trait]] MyTrait {
auto foo(int) -> int;
auto bar() -> double;
};
We can fill in duck like this (skipping some pieces for brevity, e.g. we need to forward declare duck):
// Wraps a raw call to the vtable from before cleanly (no void* exposed)
template <typename Ret, typename... Args, auto VtableMember>
class vtable_function;
template <typename Ret, typename... Args, auto VtableMember>
class vtable_function<Ret(Args...), VtableMember> {
public:
constexpr vtable_function(duck<Traits...>& owner)
: m_owner(owner) {}
constexpr auto operator()(Args... args) -> Ret {
return m_owner->get_vtable()->[:VtableMember:](
m_owner->get_underlying(),
std::forward<Args>(args)...
);
}
private:
duck<Traits...>* m_owner;
};
// generated (somewhere) by define_aggregate
struct vtable_wrapper {
vtable_function<int(int), /* foo slot */> foo;
vtable_function<double(), /* bar slot */> bar;
};
class duck : public vtable_wrapper {
private:
auto get_vtable() const -> const vtable* { ... }
auto get_underlying() -> void* { ... }
public:
template <typename Ret, typename... Args, auto VtableMember>
friend class vtable_function;
// ...
};
duck inherits from vtable_wrapper and thereby takes in all of the vtable_function callable objects with the appropriate names, so we get the myDuck.bar() syntax we want. We also have to go through each of the vtable functions in the constructor and set their owner back-pointer to look at the enclosing duck.
But this approach has a problem, and it's a big one. That owner back pointer is not cheap: sizeof(duck) now scales linearly with the number of traits we use, since each one needs a back pointer. Combining that with the vtable pointer and the data pointer that duck already stores, this could make what should be a lean type incredibly large.
Realistically, there's no reason these vtable_function objects shouldn't know about duck. After all, they're each simulating member functions and aren't used in any other context, so there should be some way for them to recover the this pointer without having to explicitly store it.
Pointer-Interconvertibility
To understand the solution, we must first take a detour to an oft-forgot feature of the C++ standard. Consider the following example:
struct SomeType {};
Comments
No comments yet. Start the discussion.