A header-only library providing structure visitors for C++11 and C++14.
In C++ there is no built-in way to iterate over the members of a struct
type.
Oftentimes, an application may contain several small "POD" datatypes, and one would like to be able to easily serialize and deserialize, print them in debugging info, and so on. Usually, the programmer has to write a bunch of boilerplate for each one of these, listing the struct members over and over again.
(This is only the most obvious use of structure visitors.)
Naively one would like to be able to write something like:
for (const auto & member : my_struct) {
std::cerr << member.name << ": " << member.value << std::endl;
}
However, this syntax can never be legal in C++, because when we iterate using a
for loop, the iterator has a fixed static type, and member.value
similarly has
a fixed static type. But the struct member types must be allowed to vary.
The usual way to overcome issues like that (without taking a performance hit) is to use the visitor pattern. For our purposes, a visitor is a generic callable object. Suppose our struct looks like this:
struct my_type {
int a;
float b;
std::string c;
};
and suppose we had a function like this, which calls the visitor v
once for
each member of the struct:
template <typename V>
void visit(const my_type & my_struct, V && v) {
v("a", my_struct.a);
v("b", my_struct.b);
v("c", my_struct.c);
}
(For comparison, see also the function boost::apply_visitor
from the boost::variant
library,
which similarly applies a visitor to the value stored within a variant.)
Then we can "simulate" the for-loop that we wanted to write in a variety of ways. For instance, we can make a template function out of the body of the for-loop and use that as a visitor.
template <typename T>
void log_func(const char * name, const T & value) {
std::cerr << name << ": " << value << std::endl;
}
visit(my_struct, log_func);
Using a template function here means that even though a struct may contain several different types, the compiler figures out which function to call at compile-time, and we don't do any run-time polymorphism -- the whole call can often be inlined.
Basically we are solving the original problem in a very exact way -- there is no longer an explicit iterator, and each time the "loop body" can be instantiated with different types as needed.
If the loop has internal state or "output", we can use a function object (an object which overloads operator()
) as the visitor,
and collect the state in its members. Also in C++14 we have generic lambdas, which sometimes makes all this very terse.
Additionally, while making a visitor is sometimes more verbose than you'd like, it has an added benefit that generic visitors can be used and reused many times. Often, when doing things like logging or serialization, you don't want each struct to get a different implementation or policy, you want to reuse the same code for all of them.
So, if we have a template function visit
for our struct, it may let us simplify code and promote code reuse.
However, that means we still have to actually define visit
for every struct we want to use it
with, and possibly several versions of it, taking const my_type &
, my_type &
, my_type &&
, and so on.
That's also quite a bit of repetitive code, and the whole point of this is to reduce repetition.
Again, ideally we would be able to do something totally generic, like,
template <typename S, typename V>
void for_each(S && s, V && v) {
// Insert magic here...
for (auto && member : s) {
v(member.name, member.value);
}
}
where both the visitor and struct are template parameters, and use this to visit the members of any struct.
Unfortunately, current versions of C++ lack reflection. It's not possible
to programmatically inspect the list of members of a generic class type S
, using templates or
anything else standard, even if S
is a complete type (in which case, the compiler obviously
knows its members). If we're lucky we might get something like this in C++20, but right
now there's no way to actually implement the fully generic for_each
.
This means that any implementation of for_each
requires some help, usually in the form of registration macros
or similar.
This library permits the following syntax in a C++11 program:
struct my_type {
int a;
float b;
std::string c;
};
VISITABLE_STRUCT(my_type, a, b, c);
struct debug_printer {
template <typename T>
void operator()(const char * name, const T & value) {
std::cerr << name << ": " << value << std::endl;
}
};
void debug_print(const my_type & my_struct) {
visit_struct::for_each(my_struct, debug_printer{});
}
Intuitively, you can think that the macro VISITABLE_STRUCT
is defining overloads of visit_struct::for_each
for your structure.
In C++14 this can be made more succinct using a lambda:
void debug_print(const my_type & my_struct) {
visit_struct::for_each(my_struct,
[](const char * name, const auto & value) {
std::cerr << name << ": " << value << std::endl;
});
}
These two things, the macro VISITABLE_STRUCT
and the function visit_struct::for_each
,
represent the most important functionality of the library.
A nice feature of visit_struct
is that for_each
always respects the
C++11 value category of it's arguments.
That is, if my_struct
is a const l-value reference, non-const l-value reference, or r-value
reference, then for_each
will pass each of the fields to the visitor correspondingly,
and the visitor is also forwarded properly.
It should be noted that there are already libraries that permit iterating over a structure like
this, such as boost::fusion
, which does this and much more. Or boost::hana
, which is like
a modern successor to boost::fusion
which takes advantage of C++14.
However, our library can be used as a single-header, header-only library with no external dependencies.
The core visit_struct.hpp
is in total about four hundred lines of code, depending on how you count,
and is fully functional on its own. For some applications, visit_struct
is all that you need.
Additionally, the syntax for doing these kind of visitations is (IMO) a little nicer than in fusion
or hana
. And visit_struct
has much better compiler support right now than hana
. hana
requires
a high level of conformance to C++14. It only supports gcc-6
and up for instance, and doesn't work with
any versions of MSVC. (Its support on clang
is quite good.) visit_struct
can be used with
many "first generation C++11 compilers" that are now quite old, like gcc-4.8
and MSVC 2013.
Note: The macro VISITABLE_STRUCT
must be used at filescope, an error will occur if it is
used within a namespace. You can simply include the namespaces as part of the type, e.g.
VISITABLE_STRUCT(foo::bar::baz, a, b, c);
visit_struct also has support code so that it can be used with "fusion-adapted structures".
That is, any structure that boost::fusion
knows about, can also be used with visit_struct::for_each
,
if you include the extra header.
#include <visit_struct/visit_struct_boost_fusion.hpp>
This compatability header means that you don't have to register a struct once with fusion
and once with visit_struct
.
It may help if you are migrating from one library to the other.
visit_struct also has a similar compatibility header for boost::hana
.
#include <visit_struct/visit_struct_boost_hana.hpp>
A drawback of the basic syntax is that you have to repeat the field member names.
This introduces a maintenance burden: What if someone adds a field member and doesn't update the list?
- It is possible to write a static assertion that all of the members are registered, by comparing sizeof the struct with what it should be given the known registered members. (See test_fully_visitable.cpp )
- It may be useful to register only a subset of the field members for serialization.
- It may be a requirement for you that you cannot change the header where the struct is defined, and you still want to visit it, so the first syntax may be pretty much the only option for you.
However, none of these changes the fact that with the first syntax, you have to write the names twice.
If visit_struct were e.g. a clang plugin instead of a header-only library, then perhaps we could make the syntax look like this:
struct my_type {
__attribute__("visitable") int a;
__attribute__("visitable") float b;
__attribute__("visitable") std::string c;
};
void debug_print(const my_type & my_struct) {
__builtin_visit_struct(my_struct,
[](const char * name, const auto & member) {
std::cout << name << ": " << member << std::endl;
});
}
We don't offer a clang plugin like this, but we do offer an additional header,
visit_struct_intrusive.hpp
which uses macros to get pretty close to this syntax, and which is portable:
struct my_type {
BEGIN_VISITABLES(my_type);
VISITABLE(int, a);
VISITABLE(float, b);
VISITABLE(std::string, c);
END_VISITABLES;
};
This declares a structure which is essentially the same as
struct my_type {
int a;
float b;
std::string c;
};
There are no additional data members defined within the type, although there are some "secret" static declarations which are occurring. (Basically, a bunch of typedef's.) That's why it's "intrusive". There is still no run-time overhead.
Each line above expands to a separate series of declarations within the body of my_type
, and arbitrary other C++
declarations may appear between them.
struct my_type {
int not_visitable;
double not_visitable_either;
BEGIN_VISITABLES(my_type);
VISITABLE(int, a);
VISITABLE(float, b);
typedef std::pair<std::string, std::string> spair;
VISITABLE(spair, p);
void do_nothing() const { }
VISITABLE(std::string, c);
END_VISITABLES;
};
When visit_struct::for_each
is used, each member declared with VISITABLE
will be visited, in the order that they are declared.
The benefits of this version are that, you don't need to type all the member names twice, and you don't need to jump out of your namespaces back to filescope in order to register a struct. The main drawbacks are that this is still somewhat verbose, the implementation is a bit more complicated, and this one may not be useful in some cases, like if the struct you want to visit belongs to some other project and you can't change its definition.
visit_struct also supports visiting two instances of the same struct type at once.
For instance, the function call
visit_struct::for_each(s1, s2, v);
is similar to
v("a", s1.a, s2.a);
v("b", s1.b, s2.b);
v("c", s1.c, s2.c);
This is useful for implementing generic equality and comparison operators for visitable
structures, for instance. Here's an example of a generic function struct_eq
which
compares any two visitable structures for equality using operator ==
on each field,
and which short-circuits properly.
struct eq_visitor {
bool result = true;
template <typename T>
void operator()(const char *, const T & t1, const T & t2) {
result = result && (t1 == t2);
}
};
template <typename T>
bool struct_eq(const T & t1, const T & t2) {
eq_visitor vis;
visit_struct::for_each(t1, t2, vis);
return vis.result;
}
On clang 3.5 with a simple example, this compiles the same assembly as a hand-rolled equality operator. See it on godbolt compiler explorer.
Besides iteration over an instance of a registered struct, visit_struct also supports visiting the definition of the struct. In this case, instead of passing you the field name and the field value within some instance, it passes you the field name and the pointer to member corresponding to that field.
Suppose that you are serializing many structs in your program as json. You might also want to be able to emit the json schema associated to each struct that your program is expecting, especially to produce good diagnostics if loading the data fails. When you visit without an instance, you can get all the type information for the struct, but you don't have to actually instantiate it, which might be complicated or expensive.
The function call
visit_struct::visit_pointers<my_type>(v);
is similar to
v("a", &my_type::a);
v("b", &my_type::b);
v("c", &my_type::c);
These may be especially useful when you have a C++14 compiler which has proper constexpr
support.
In that case, these visitations are constexpr
also, so you can use this
for some nifty metaprogramming purposes. (For an example, check out test_fully_visitable.cpp.)
There are two alternate versions of this visitation.
This function call
visit_struct::visit_types<my_type>(v);
is similar to
v("a", visit_struct::type_c<a>());
v("b", visit_struct::type_c<b>());
v("c", visit_struct::type_c<c>());
Here, type_c
is just a tag, so that your visitor can take appropriate action using tag dispatch.
This syntax is a little simpler than the pointer to member syntax.
In the third version, you get passed an "accessor", that is, a function object that implements the function computed by the pointer-to-member.
This call
visit_struct::visit_accessors<my_type>(v);
is roughly similar to
v("a", [](auto s) { return s.a; });
v("b", [](auto s) { return s.b; });
v("c", [](auto s) { return s.c; });
Accessors are convenient because they can be used easily with other standard algorithms that require function objects, they avoid the syntax of member pointers, and because they are well-supported by hana and fusion.
Much thanks to Jarod42 for this patch and subsequent suggestions.
Note: The compatibility headers for boost::fusion
and boost::hana
don't
currently support visit_pointers
. They only support visit_types
, and visit_accessors
.
To my knowledge, there is no way to get the pointers-to-members from boost::fusion
or boost::hana
.
That is, there is no publicly exposed interface to get them.
If you really want or need to be able to get the pointers to members, that's a pretty good reason to use visit_struct
honestly.
If you think you need the fusion or hana compatibility, then you should probably avoid anything to do with member pointers here, and stick to accessors instead.
for_each
is quite powerful, and by crafting special visitors, there is a lot that you can do with it.
However, one thing that you cannot easily do is implement std::tuple
methods, like std::get<i>
to get the i
'th member of the struct.
Most if not all libraries that support struct-field reflection support this in some way.
So, we decided that we should support this also.
We didn't change our implementation of for_each
, which works well on all targets.
But we have added new functions which allow indexed access to structures, and to the metadata.
visit_struct::get<i>(s);
Gets (a reference to) the i
'th visitable member of the struct s
. Index is 0-based. Analogous to std::get
.
visit_struct::get_name<i, S>();
visit_struct::get_name<i>(s);
Gets a string constant representing the name of the i
'th member of the struct type S
. The struct type may be passed as a second template parameter.
If an instance is available, it may be passed as an argument, and the struct type will be deduced (the argument will not be accessed).
visit_struct::get_pointer<i, S>();
visit_struct::get_pointer<i>(s);
Gets the pointer-to-member for the i
'th visitable element of the struct type S
.
visit_struct::get_accessor<i, S>();
visit_struct::get_accessor<i>(s);
Gets the accessor corresponding to the i
'th visitable element of the struct type S
.
visit_struct::type_at<i, S>
This alias template gives the declared type of the i
'th member of S
.
visit_struct::field_count<S>();
visit_struct::field_count(s);
Gets a size_t
which tells how many visitable fields there are.
visit_struct::get_name<S>();
visit_struct::get_name(s);
Gets a string constant representing the name of the structure. The string here is the token that you passed to the visit_struct
macro in order to register the structure.
This could be useful for error messages. E.g. "Failed to match json input with struct of type 'foo', layout: ..."
There are other ways to get a name for the type, such as typeid
, but it has implementation-defined behavior and sometimes gives a mangled name. However, the visit_struct
name might not
always be acceptable either -- it might contain namespaces, or not, depending on if you use standard or intrusive syntax, for instance.
Since the programmer is already taking the trouble of passing this name into a macro to register the struct, we think we might as well give programmatic access to that string if they want it.
Note that there is no equivalent feature in fusion
or hana
to the best of my knowledge, so there's no support for this in the compatibility headers.
visit_struct::apply_visitor(v, s);
visit_struct::apply_visitor(v, s1, s2);
This is an alternate syntax for for_each
. The only difference is that the visitor comes first rather than last.
Historically, apply_visitor
is a much older part of visit_struct
than for_each
. Its syntax is similar to boost::apply_visitor
from the boost::variant
library.
For a long time, apply_visitor
was the only function in the library.
However, experience has shown that for_each
is little nicer syntax than apply_visitor
. It reads more like a for loop -- the bounds of the loop come first, which are the structure, then the body of the loop, which is repeated.
Additionally, in C++14 one may often use generic lambdas. Then the code is a little more readable if the lambda comes last, since it may span several lines of code.
(I won't say I wasn't influenced by ldionne's opinion. He makes this same point in the boost::hana
docs here.)
So, nowadays I prefer and recommend for_each
. The original apply_visitor
syntax isn't going to be deprecated or broken though.
visit_struct::traits::is_visitable<S>::value
This type trait can be used to check if a structure is visitable. The above expression should resolve to boolean true or false. I consider it part of the forward-facing interface, you can use it in SFINAE to easily select types that visit_struct
knows how to use.
When using VISITABLE_STRUCT
, the maximum number of members which can be registered
is visit_struct::max_visitable_members
, which is by default 69.
When using the intrusive syntax, the maximum number of members is visit_struct::max_visitable_members_intrusive
,
which is by default 100.
These limits can both be increased, see the source comments and also IMPLEMENTATION_NOTES.md.
visit_struct targets C++11 -- you need to have r-value references at least, and for the intrusive syntax, you need variadic templates also.
visit_struct is known to work with versions of gcc >= 4.8.2
and versions of clang >= 3.5
.
The appveyor build tests against MSVC 2013, 2015, 2017.
MSVC 2015 is believed to be fully supported.
For MSVC 2013, the basic syntax is supported, the intrusive syntax doesn't work there and now isn't tested. Again, patches welcome.
Much thanks again to Jarod42 for significant patches related to MSVC support.
visit_struct
attempts to target three different levels of constexpr
support.
- No support
- C++11 support
- C++14 extended support
This is controlled by two macros VISIT_STRUCT_CONSTEXPR
and VISIT_STRUCT_CXX14_CONSTEXPR
. We use these tokens where we would use the constexpr
keyword.
In the visit_struct.hpp
header, these macros are defined to either constexpr
or nothing.
We attempt to guess the appropriate setting by inspecting the preprocessor symbols __cplusplus
and _MSC_VER
.
If it doesn't work on your compiler, please open a github issue, especially if you know how to fix it :)
In the meantime, if you don't want to tweak the headers for your project, you can override the behavior by defining these macros yourself, before including visit_struct.hpp
.
If the header sees that you have defined them it won't touch them and will defer to your settings. In most cases this should not be necessary.
On gcc and clang, we assume at least C++11 constexpr support. If you enabled a later standard using -std=...
, we turn on the full constexpr
.
On MSVC currently the settings are:
- VS2013: no support
- VS2015: C++11 support
- VS2017: C++14 extended support
visit_struct is available under the boost software license.
- map-macro from swansontec
- boost-hana from ldionne
- pod flat reflection from apolukhin
- self-aware structs from jckarter