KtBind: C++ and Kotlin interoperability

KtBind is a lightweight C++17 header-only library that exposes C++ types to Kotlin and vice versa, mainly to create Kotlin bindings for existing C++ code. The objective of KtBind is to provide an easy-to-use interoperability interface that feels natural from both C++ and Kotlin. KtBind uses C++ compile-time introspection to generate Java Native Interface (JNI) stubs that can be accessed from Kotlin with regular function invocation. The stubs incur minimal or no overhead compared to hand-written JNI code.

This project has been inspired by a similar binding interface between JavaScript and C++ in emscripten, and between Python and C++ in PyBind11 and Boost.Python. Unlike JNA, which one can utilize by means of an intermediary C interface, KtBind offers a direct interface between C++ and Kotlin.

Core features

The following core C++ features can be mapped to Kotlin:

• Functions accepting and returning custom data structures by value or by const reference
• Instance methods and static methods
• Overloaded functions
• Operators
• Instance attributes and static attributes
• Arbitrary exception types
• Callbacks and function objects
• STL data structures

Furthermore, the following core Kotlin features are exposed seamlessly to C++:

• Collection types
• Higher-order functions and lambda expressions
• Arbitrary exception types

Getting started

KtBind is a header-only library, including the interoperability header in your C++ project allows you to create bindings to Kotlin:

#include <ktbind/bind.hpp>

Consider the following C++ class as an example:

struct Sample {
    Sample();
    Sample(const char*);
    Sample(std::string);
    Data get_data();
    void set_data(const Data& data);
};

All Kotlin bindings are registered in the extension module block. In order to expose the member functions of the class Sample, we use native_class and its builder functions constructor and function:

DECLARE_NATIVE_CLASS(Sample, "com.kheiron.example.Sample")

JAVA_EXTENSION_MODULE() {
    using namespace java;
    native_class<Sample>()
        .constructor<Sample()>("create")
        .constructor<Sample(std::string)>("create")
        .function<&Sample::get_data>("get_data")
        .function<&Sample::set_data>("set_data")
    ;
    print_registered_bindings();
}

The type parameter of the template function constructor is a function signature to help choose between multiple available constructors (between constructors that take parameter types const char* and std::string in this case). The non-type template parameter of function is a function pointer, either a member function pointer (as shown above) or a free function pointer.

print_registered_bindings is a utility function that lets you print the Kotlin class definition that corresponds to the registered C++ class definitions. print_registered_bindings prints to Java System.out when you load the compiled shared library (*.so) with Kotlin's System.load(). You would normally use it in the development phase.

The bindings above map to the following class definition in Kotlin:

package com.kheiron.example
import com.kheiron.ktbind.NativeObject

class Sample private constructor() : NativeObject() {
    external override fun close()
    external fun get_data(): Data
    external fun set_data(data: Data)
    companion object {
        @JvmStatic external fun create(): Sample
        @JvmStatic external fun create(str: String): Sample
    }
}

Notice that the Kotlin functions corresponding to the C++ constructors appear in the Kotlin companion object, and the Kotlin class itself has only a private constructor. This highlights an important characteristic of native_class: it serves as a way to expose native objects to Kotlin. The native object lives in the C++ space, and native_class exposes an opaque handle to the object. (This opaque handle is stored in NativeObject.) Whenever a function is called, all parameters are passed by value from Kotlin to C++ by the interoperability layer, and a corresponding function invocation is made on the native object. Let's look at a specific example:

Sample.create().use {
    println(it.get_data())
    it.set_data(Data(true, 42, 65000, 12, 3.1416f, 2.7183, "a Kotlin string", emptyMap()))
}

Here, a Sample object is instantiated in Kotlin but a corresponding native object is immediately created in the C++ space by calling the default constructor. When a function such as get_data() is called, KtBind marshals all parameters, and makes an invocation to Sample::get_data defined in C++, and the return value is passed back to Kotlin.

C++ objects have constructors and destructors but Kotlin (JVM) has garbage collection. In order to ensure that objects are properly reclaimed when they are no longer needed, Sample implements the interface AutoCloseable (via NativeObject). Calling the close method triggers the C++ destructor. (In the example, close() is called automatically at the end of the use block.)

Notice that we have used Data, a type we have yet to define. Its C++ definition looks as follows:

struct Data {
    bool b = true;
    short s = 82;
    int i = 1024;
    long l = 111000111000;
    float f = M_PI;
    double d = M_E;
    std::string str = "sample string in struct";
    std::map<std::string, std::vector<std::string>> map = {
        {"one", {"a","b","c"}},
        {"two", {"x","y","z"}},
        {"three", {}},
        {"four", {"l"}}
    };
};

The corresponding Kotlin definition is as shown below:

data class Data(
        val b: Boolean,
        val s: Short,
        val i: Int,
        val l: Long,
        val f: Float,
        val d: Double,
        val str: String,
        val map: Map<String, List<String>>
)

As shown in the example, KtBind supports fundamental types, object types, generic types, and composite types, up to arbitrary levels of nesting. Data classes are always passed by value, and registered in C++ with data_class:

// ...
DECLARE_DATA_CLASS(Data, "com.kheiron.example.Data")

JAVA_EXTENSION_MODULE() {
    // ...
    data_class<Data>()
        .field<&Data::b>("b")
        .field<&Data::s>("s")
        .field<&Data::i>("i")
        .field<&Data::l>("l")
        .field<&Data::f>("f")
        .field<&Data::d>("d")
        .field<&Data::str>("str")
        .field<&Data::map>("map")
    ;
}

Both the C++ and the Kotlin definition of a data class might have fields not registered in the binding but the values of these fields will not be transferred across the language boundary, and will always take their initial values.

Type mapping

KtBind recognizes several widely-used types and marshals them automatically between C++ and Kotlin without explicit user-defined type specification:

Java boxing an unboxing for types is performed automatically.

Exceptions

Exceptions thrown in C++ automatically trigger a Java exception when crossing the language boundary. The interoperability layer catches all exceptions that inherit from std::exception, and throws a java.lang.Exception before passing control back to the JVM.

Exceptions originating from Java/Kotlin are automatically wrapped in a C++ type called JavaException, which derives from std::exception. The function what() in JavaException retrieves the Java exception message. C++ code can catch JavaException and take appropriate action, which causes the exception to be cleared in Java.

Callbacks

Passing a callback or lambda from Kotlin to C++ is fully supported. The callback or lambda is wrapped in a std::function<R(Args...) that allows C++ code to trigger the lambda at any time (even from a different thread). Take the following class definition in Kotlin:

class LambdaSample private constructor() : NativeObject() {
    external override fun close()
    companion object {
       @JvmStatic external fun pass_callback_arguments(str: String, callback: (String, Short, Int, Long) -> String): String
    }
}

and the way it is used:

val result = LambdaSample.pass_callback_arguments("callback") { str, short, int, long -> "($str, $short, $int, $long)" }

The corresponding C++ function definition looks as follows:

std::string pass_callback_arguments(std::string str, std::function<std::string(std::string, short, int, long)> fun) {
    return fun(str, 4, 82, 112);
}

In the example above, result evaluates to "(callback, 4, 82, 112)".

Binding registration

The macro JAVA_EXTENSION_MODULE in KtBind expands into a pair of function definitions:

JNIEXPORT jint JNI_OnLoad(JavaVM* vm, void* reserved) { ... }
JNIEXPORT void JNI_OnUnload(JavaVM *vm, void *reserved) { ... }

These definitions, in turn, iterate over the function and field bindings registered with native_class and data_class.

Each function binding generates a function pointer at compile time, which are passed to the JNI function RegisterNatives. Each of these function pointers points at a static member function of a template class, where the template parameters capture the type information extracted from the function signature. When the function is invoked through the pointer, the function makes the appropriate type conversions to cast Java types into C++ types and back. For example, the C++ function signature

bool func(const std::string& str, const std::vector<int>& vec, double d);

causes the function adapter template to be instantiated with parameter types std::string, std::vector<int> and double and return type bool. When Java calls the pointer through JNI, the adapter transforms the types std::string and std::vector<int>. (double and bool need no transformation.) For each transformed type, a temporary object is created, all of which are then used in invoking the original function func.

Internally, field bindings utilize JNI accessor functions like GetObjectField and SetObjectField to extract and populate Java objects. Like with function bindinds, KtBind uses C++ type information to make the appropriate JNI function call. For instance, setting a field with type double entails a call to GetDoubleField (from Java to C++) or SetDoubleField (from C++ to Java). If the type is a composite type, such as a std::vector<T>, then a Java object is constructed recursively, and then set with SetObjectField. For example,

• Setting a field of type std::vector<int> first creates a java.util.ArrayList with JNI's NewObject, then sets elements with the add method (invoked using JNI's CallBooleanMethod), performing boxing for the primitive type int with valueOf, and finally uses SetObjectField with the newly created java.util.ArrayList instance.
• Setting an std::vector<std::string> field involves creating a java.util.ArrayList with JNI's NewObject, and a call to JNI's NewStringUTF for each string element. The strings are then added to the java.util.ArrayList instance with add, and finally to the field with SetObjectField.