我建议使用Qt。

有一个开源许可证和一个商业许可证。

查看Classdesc http://classdesc.sf.net。它以类“描述符”的形式提供了反射，可以与任何标准c++编译器一起工作(是的，它可以与Visual Studio和GCC一起工作)，并且不需要源代码注释(尽管存在一些处理棘手情况的pragmas)。它已经开发了十多年，并在一些工业规模的项目中使用。

您需要查看您正在尝试做什么，以及RTTI是否满足您的需求。我已经实现了自己的伪反射，用于某些非常特定的目的。例如，我曾经希望能够灵活地配置模拟输出内容。它需要在输出的类中添加一些样板代码:

namespace {
  static bool b2 = Filter::Filterable<const MyObj>::Register("MyObject");
} 

bool MyObj::BuildMap()
{
  Filterable<const OutputDisease>::AddAccess("time", &MyObj::time);
  Filterable<const OutputDisease>::AddAccess("person", &MyObj::id);
  return true;
}

第一个调用将该对象添加到筛选系统，该系统调用BuildMap()方法以确定哪些方法可用。

然后，在配置文件中，你可以这样做:

FILTER-OUTPUT-OBJECT   MyObject
FILTER-OUTPUT-FILENAME file.txt
FILTER-CLAUSE-1        person == 1773
FILTER-CLAUSE-2        time > 2000

通过一些涉及boost的模板魔法，可以在运行时(读取配置文件时)将其转换为一系列方法调用，因此相当高效。我不建议你这样做，除非你真的需要，但是，当你这样做的时候，你可以做一些非常酷的事情。

我想你可能会对Dominic Filion写的“在c++中使用反射模板”这篇文章感兴趣。它在Game Programming Gems 5的1.4部分。不幸的是，我没有带我的副本，但你可以找找看，因为我认为它解释了你想要的东西。

RareCpp库实现了相当简单和直观的反射——所有字段/类型信息都被设计成可以在数组中使用，或者感觉像是数组访问。它是为c++ 17编写的，可与Visual Studios、g++和Clang一起使用。这个库只有头文件，这意味着你只需要将“Reflect.h”复制到你的项目中就可以使用它。

被反射的结构体或类需要REFLECT宏，在该宏中您可以提供所反射的类的名称和字段的名称。

class FuelTank {
    public:
        float capacity;
        float currentLevel;
        float tickMarks[2];

    REFLECT(FuelTank, capacity, currentLevel, tickMarks)
};

这就是全部内容，不需要额外的代码来设置反射。可选地，您可以提供类和字段注释，以便能够遍历超类或向字段添加额外的编译时信息(例如Json::Ignore)。

遍历字段可以简单到…

for ( size_t i=0; i<FuelTank::Class::TotalFields; i++ )
    std::cout << FuelTank::Class::Fields[i].name << std::endl;

您可以通过对象实例循环访问字段值(您可以读取或修改)和字段类型信息……

FuelTank::Class::ForEachField(fuelTank, [&](auto & field, auto & value) {
    using Type = typename std::remove_reference<decltype(value)>::type;
    std::cout << TypeToStr<Type>() << " " << field.name << ": " << value << std::endl;
});

JSON库构建在RandomAccessReflection之上，它可以自动识别适当的JSON输出表示来读写，并且可以递归遍历任何反射字段，以及数组和STL容器。

struct MyOtherObject { int myOtherInt; REFLECT(MyOtherObject, myOtherInt) };
struct MyObject
{
    int myInt;
    std::string myString;
    MyOtherObject myOtherObject;
    std::vector<int> myIntCollection;

    REFLECT(MyObject, myInt, myString, myOtherObject, myIntCollection)
};

int main()
{
    MyObject myObject = {};
    std::cout << "Enter MyObject:" << std::endl;
    std::cin >> Json::in(myObject);
    std::cout << std::endl << std::endl << "You entered:" << std::endl;
    std::cout << Json::pretty(myObject);
}

上面的代码可以这样运行……

Enter MyObject:
{
  "myInt": 1337, "myString": "stringy", "myIntCollection": [2,4,6],
  "myOtherObject": {
    "myOtherInt": 9001
  }
}


You entered:
{
  "myInt": 1337,
  "myString": "stringy",
  "myOtherObject": {
    "myOtherInt": 9001
  },
  "myIntCollection": [ 2, 4, 6 ]
}

参见……

反映文档 反映实现 更多用法示例

I did something like what you're after once, and while it's possible to get some level of reflection and access to higher-level features, the maintenance headache might not be worth it. My system was used to keep the UI classes completely separated from the business logic through delegation akin to Objective-C's concept of message passing and forwarding. The way to do it is to create some base class that is capable of mapping symbols (I used a string pool but you could do it with enums if you prefer speed and compile-time error handling over total flexibility) to function pointers (actually not pure function pointers, but something similar to what Boost has with Boost.Function--which I didn't have access to at the time). You can do the same thing for your member variables as long as you have some common base class capable of representing any value. The entire system was an unabashed ripoff of Key-Value Coding and Delegation, with a few side effects that were perhaps worth the sheer amount of time necessary to get every class that used the system to match all of its methods and members up with legal calls: 1) Any class could call any method on any other class without having to include headers or write fake base classes so the interface could be predefined for the compiler; and 2) The getters and setters of the member variables were easy to make thread-safe because changing or accessing their values was always done through 2 methods in the base class of all objects.

It also led to the possibility of doing some really weird things that otherwise aren't easy in C++. For example I could create an Array object that contained arbitrary items of any type, including itself, and create new arrays dynamically by passing a message to all array items and collecting the return values (similar to map in Lisp). Another was the implementation of key-value observing, whereby I was able to set up the UI to respond immediately to changes in the members of backend classes instead of constantly polling the data or unnecessarily redrawing the display.

也许您更感兴趣的是，您还可以转储为类定义的所有方法和成员，而且是字符串形式。

该系统的缺点可能会让您望而却步:添加所有消息和键值非常繁琐;它比没有反射要慢;你会讨厌看到boost::static_pointer_cast和boost::dynamic_pointer_cast遍布你的代码库;强类型系统的局限性仍然存在，您实际上只是将它们隐藏了一些，所以它不那么明显。字符串中的错别字也不是一个有趣或容易发现的惊喜。

As to how to implement something like this: just use shared and weak pointers to some common base (mine was very imaginatively called "Object") and derive for all the types you want to use. I'd recommend installing Boost.Function instead of doing it the way I did, which was with some custom crap and a ton of ugly macros to wrap the function pointer calls. Since everything is mapped, inspecting objects is just a matter of iterating through all of the keys. Since my classes were essentially as close to a direct ripoff of Cocoa as possible using only C++, if you want something like that then I'd suggest using the Cocoa documentation as a blueprint.