This paper is a concise introduction to the static reflection facilities proposed in P0194. See also P0385 for examples, in-depth design rational, and future extensions. A previous version of the functionality described here has been implemented in clang and is available for experimentation.
reflexpr was preferred to $reflect. The get_base_name operation was changed to get_name which only works for types that have a name (e.g., const and & qualified types do not have a name).C++ provides many facilities that allow for the compile-time inspection and manipulation of types and values. Indeed an entire field, template metaprogramming, arose to take advantage of these features. While the feats accomplished have been both surprising and impressive, C++ still lacks some fundamental building blocks in this area. There is no way, for example, to discover an arbitrary class's name or query its member variables. This paper describes an attempt to extend the language with the most crucial of these building blocks.
Lets look at a simple example illustrating our proposed design:
template <typename T>
T min(const T& a, const T& b) {
log() << "min<"
<< get_display_name_v<reflexpr(T)>
<< ">(" << a << ", " << b << ") = ";
T result = a < b ? a : b;
log() << result << std::endl;
return result;
}Here we define a min function that closely resembles, aside from the logging, the semantics of std::min. Assuming several calls, the output of this function might look something like this:
min<int>(1, 33)
min<std::string>(hello, world)
...Note that the type argument as well as the value arguments are printed out. Our proposed reflection syntax and library is what makes this possible.
The key expression is get_display_name_v<reflexpr(T)>, which consists of two parts. The first is reflexpr(T). This expression produces a special type that contains meta-information concerning T (e.g. name, member variable, and inheritance information). We call these types "meta objects" following industry practice. reflexpr's argument doesn't necessarily have to be a type though; many kinds of syntax are recognized in the general case. The second part is the call to get_display_name_v. This is where we extract a piece of information from the meta object, in this case the "display name". Most uses of the reflection facilities we provide are variations of this simple theme: reflect to produce a metaobject and query that meta object for information.
The design space for reflection is vast. It is tempting to set complete coverage as a goal even though the value proposition is dubious at best. Instead, we attempted a pragmatic approach where we add a minimal set of features that are tied to concrete use cases and compliment the existing C++ feature set. You will not see, for instance, a replacement or alternative for std::is_const or other well-established pre-existing reflection capabilities.
We also are not, in this iteration, proposing a high-level reflection API. There is still a lot of discussion as to what such an API should look like. More experimentation is required. Instead we provide low-level interfaces that can be built upon.
Our vision for a complete reflection software stack is as follows
At the bottom are low-level reflection facilities. That is what we are proposing here. On top of that is a higher-level reflection library that, building on the lower-level facilities, provides an easy and convenient way to write libraries requiring reflection. Higher yet are domain specific libraries. This could be a serialization library or an automatic database schema generator. Finally, on the top, is user code which makes use of the domain specific libraries.
The danger of designing facilities instead of a high-level API is that the former ends up inadequate for the latter. Fortunately, we've been developing in parallel several variations of high-level APIs that make use of our facilities. See the following snippet making use of the mirror library, a Boost.MPL style reflection library built on our facilities:
template <typename T>
bool generic_equal(const T& a, const T& b)
{
using metaT = reflexpr(T);
bool result = true;
mirror::for_each<mirror::get_data_members_t<metaT>>(
compare_data_members<T>{a, b, result}
);
return result;
}At the 2016 Issaquah meeting, Louis Dionne presented a Boost.Hana-styled reflection library built on the facilities described here along with an impressive JSON serialization library built on top of that. We find this result encouraging and suggestive that our facilities are as general and API-agnostic as we had hoped.
As mentioned above, we're aiming for a minimal set of functionality that still satisfies a number of use cases. Here's a basic summary of what's included and what's not.
Included:
Data members. e.g. walking through the data members of a class
Member types. e.g. walking through nested types or typedefs in a class
Enumerators. The ability to, for example, make one-line serialization routines for enums
Template instantiations. The ability to reflect on instantiated templates, such as std::vector<int>
Alias support. The ability to distinguish between a typedef and its underlying type
Not Included:
Namespace member sequences. We're not providing the ability to walk through all the declarations of a namespace. Efficient implementations of such a feature seem unlikely and use cases are wanting.
Functions. We think we know what we want and how we can do it, but this chapter is complex. This is comming in a separate proposal, P0670.
Class templates. It is unclear, as of yet, what reflecting on an uninstantiated template, like vector, would look like.
Building new datatypes. We have plans to extend this proposal at some point with the ability to build new datatypes with an identifier generation facility. See the design document for more details.
Reflection facilities already in C++. As mentioned before, we're in the business of cooperating with existing facilities instead of replacing what already exists.
Anonymous functions. An extension to this paper which supports anonymous functions will be forthcoming.
Attributes. We feel that reflecting on attributes would be a highly valuable extension to this proposal. However, this feature would cause a significant change to how we use attributes today. We intend to propose this in a future paper.
The primary consideration at the language level was what to call the reflexpr operator. There was a bikeshedding at the 2016 Kona meeting and reflexpr had the most consensus when compared to several other options.
The lack of a decent compile-time string representation forced us to choose between several undesirable options. For consistency with the other metafunctions we decided to use integral_constant<const char (&)[N], STR> where STR is the name of a static, null-terminated byte string of length N.
We assume WG21 will incorporate proper compile-time strings at some point and consider this a placeholder implementation.
ObjectSequenceBecause there isn't a native type-list implementation in the standard library, we included a placeholder implementation in the reflection proposal. It provides the ability to query size and get an element by index.
template <ObjectSequence T>
constexpr auto get_size_v = get_size<T>::value;
template <size_t I, ObjectSequence S>
using get_element_t = typename get<I, S>::type;Additionally, the unpack_sequence_v metafunction was provided that enables the convenient conversion of a ObjectSequence into another type-list representation, such as a std::tuple.
template <template <class...> class Tpl, ObjectSequence S>
constexpr auto unpack_sequence_v = unpack_sequence<Tpl, S>::value;We assume WG21 will incorporate proper compile-time type lists at some point and consider this a placeholder implementation as well.
All types containing metainformation satisfy the reflect::Object concept. Beyond that, there are several other concepts that provide more specialized information. Generally a metaobject will satisfy several of the concepts below.
Class-like things:
Record. A union or a ClassClass. A class or a structBase. The public B part of class A : public B {};, for exampleRecordMember. Data members and member typesScopes:
Namespace. A namespaceScope. A namespace, class, or enumeration scopeScopeMember. Something in a scopeGlobalScope. The global namespace, ::Enums:
Enum. An enum.Enumerator. An enumerator.Types:
Typed. Something with a type, such as a member variable.Type. A type.Expressions:
Variable. A variable.Constant. A constant expression, like an enumerator.Other:
Named. Something with a name.Alias. An alias, such as a typedef.ObjectSequence. Our stand-in for type lists.In the following sections we'll go into more detail for some of the more important concepts and operations.
The reflexpr operation always produces a type that satisfies the Object concept. Objects provide the ability to query source location and the reflects_same method determines whether or not two objects reflect the same underlying entity.
template <Object T> struct get_source_line;
template <Object T> struct get_source_column;
template <Object T> struct get_source_file_name;
template <Object T1, Object T2>
struct reflects_same;Additionally, get_source_location is provided for compatibility with the std::source_location datatype in the library fundamentals TS.
template <Object T>
struct get_source_location;
// return a std::source_location objectRecordA record is a union, class, or struct (i.e. a class type).
The most general way to query members is through use of get_data_members and get_member_types. These provide lists of all the public, private, and protected members. Because the access of private members can cause abstraction leaks, two other variants are provided.
The get_public_* metafunctions return only the public members of a record. This operation can be used safely on third party code. The get_accessible_* metafunctions, on the other hand, will include private members as well if the reflexpr operation's surrounding context allows it (e.g. it is found in a member function or friend class). Encapsulation cannot be broken with either of these two variants.
template <Record T> struct get_data_members;
template <Record T> struct get_public_data_members;
template <Record T> struct get_accessible_data_members;
template <Record T> struct get_member_types;
template <Record T> struct get_public_member_types;
template <Record T> struct get_accessible_member_types;Most entities that can be reflected upon have a name of some sort and the Named concept supports this. There are two primary string-returning operations, get_display_name and get_name.
template <Named T> struct get_display_name;
template <Named T> struct get_name;While the semantics of both of these functions is implementation defined, they have a clear difference in intent as described below.
get_nameget_name returns the name of the underlying reflected entity if it has one. Note that abbreviations are elaborated and instantiated template classes return the name of the template class itself. Decorated types, such as those with const and volatile do not have a name and the get_name operation on these returns "".
get_name_v<reflexpr(unsigned)>
// "unsigned int"
using foo = int;
get_name_v<reflexpr(foo)>
// "foo"
get_name_v<reflexpr(std::vector<int>)>
// "vector"
get_name_v<reflexpr(volatile std::size_t* [10])>
// ""get_display_nameget_display_name provides a way for compilers to provide a non-portable, but human readable, representation of the underlying entity. The intent is for this to hook into the technology already used in compilers to provide human readable diagnostics.
get_display_name_v<reflexpr(unsigned)>
// "unsigned"
using foo = int;
get_display_name_v<reflexpr(foo)>
// "foo"
get_display_name_v<reflexpr(std::vector<int>)>
// "std::vector<int>"
get_display_name_v<reflexpr(volatile std::size_t* [10])>
// "volatile std::size_t *[10]"AliasAliases (viz. typedefs, etc.) provide a get_aliased operation which returns the underlying entity being reflected.
template <Alias T> struct get_aliased;For example:
using MyInt = int;
get_name_v<reflexpr(MyInt)> // "MyInt"
get_name_v<get_aliased_t<reflexpr(MyInt)>> // "int"Note that get_aliased_t always returns the true underlying type. Walking through typedefs of typedefs is not supported as a concession to compiler implementers.
We've overviewed a proposal for adding static reflection to C++. The feature set provided here goes a long way towards filling the holes metaprogrammers face today and will be extended to support even more ambitious reflection capabilities in the future.