Embind

Embind is used to bind C++ functions and classes to JavaScript, so that the compiled code can be used in a natural way by “normal” JavaScript. Embind also supports calling JavaScript classes from C++.

Embind has support for binding most C++ constructs, including those introduced in C++11 and C++14. Its only significant limitation is that it does not currently support raw pointers with complicated lifetime semantics.

This article shows how to use EMSCRIPTEN_BINDINGS() blocks to create bindings for functions, classes, value types, pointers (including both raw and smart pointers), enums, and constants, and how to create bindings for abstract classes that can be overridden in JavaScript. It also briefly explains how to manage the memory of C++ object handles passed to JavaScript.

Tip

In addition to the code in this article:

Note

Embind was inspired by Boost.Python and uses a very similar approach for defining bindings.

A quick example

The following code uses an EMSCRIPTEN_BINDINGS() block to expose the simple C++ lerp() function() to JavaScript.

// quick_example.cpp
#include <emscripten/bind.h>

using namespace emscripten;

float lerp(float a, float b, float t) {
    return (1 - t) * a + t * b;
}

EMSCRIPTEN_BINDINGS(my_module) {
    function("lerp", &lerp);
}

To compile the above example using embind, we invoke emcc with the bind option:

emcc --bind -o quick_example.js quick_example.cpp

The resulting quick_example.js file can be loaded as a node module or via a <script> tag:

<!doctype html>
<html>
  <script src="quick_example.js"></script>
  <script>
    console.log('lerp result: ' + Module.lerp(1, 2, 0.5));
  </script>
</html>

The code in an EMSCRIPTEN_BINDINGS() block runs when the JavaScript file is initially loaded (at the same time as the global constructors). The function lerp()‘s parameter types and return type are automatically inferred by embind.

All symbols exposed by embind are available on the Emscripten Module object.

Important

Always access objects through the Module object object, as shown above.

While the objects are also available in the global namespace by default, there are cases where they will not be (for example, if you use the closure compiler to minify code or wrap compiled code in a function to avoid polluting the global namespace). You can of course use whatever name you like for the module by assigning it to a new variable: var MyModuleName = Module;.

Classes

Exposing classes to JavaScript requires a more complicated binding statement. For example:

class MyClass {
public:
  MyClass(int x, std::string y)
    : x(x)
    , y(y)
  {}

  void incrementX() {
    ++x;
  }

  int getX() const { return x; }
  void setX(int x_) { x = x_; }

  static std::string getStringFromInstance(const MyClass& instance) {
    return instance.y;
  }

private:
  int x;
  std::string y;
};

// Binding code
EMSCRIPTEN_BINDINGS(my_class_example) {
  class_<MyClass>("MyClass")
    .constructor<int, std::string>()
    .function("incrementX", &MyClass::incrementX)
    .property("x", &MyClass::getX, &MyClass::setX)
    .class_function("getStringFromInstance", &MyClass::getStringFromInstance)
    ;
}

The binding block defines a chain of member function calls on the temporary class_ object (this same style is used in Boost.Python). The functions register the class, its constructor(), member function(), class_function() (static) and property().

Note

This binding block binds the class and all its methods. As a rule you should bind only those items that are actually needed, as each binding increases the code size. For example, it would be rare to bind private or internal methods.

An instance of MyClass can then be created and used in JavaScript as shown below:

var instance = new Module.MyClass(10, "hello");
instance.incrementX();
instance.x; // 12
instance.x = 20; // 20
Module.MyClass.getStringFromInstance(instance); // "hello"
instance.delete();

Memory management

JavaScript, specifically ECMA-262 Edition 5.1, does not support finalizers or weak references with callbacks. Therefore there is no way for Emscripten to automatically call the destructors on C++ objects.

Warning

JavaScript code must explicitly delete any C++ object handles it has received, or the Emscripten heap will grow indefinitely.

var x = new Module.MyClass;
x.method();
x.delete();

var y = Module.myFunctionThatReturnsClassInstance();
y.method();
y.delete();

Value types

Manual memory management for basic types is onerous, so embind provides support for value types. Value arrays are converted to and from JavaScript Arrays and value objects are converted to and from JavaScript Objects.

Consider the example below:

struct Point2f {
    float x;
    float y;
};

struct PersonRecord {
    std::string name;
    int age;
};

PersonRecord findPersonAtLocation(Point2f);

EMSCRIPTEN_BINDINGS(my_value_example) {
    value_array<Point2f>("Point2f")
        .element(&Point2f::x)
        .element(&Point2f::y)
        ;

    value_object<PersonRecord>("PersonRecord")
        .field("name", &PersonRecord::name)
        .field("age", &PersonRecord::age)
        ;

    function("findPersonAtLocation", &findPersonAtLocation);
}

The JavaScript code does not need to worry about lifetime management.

var person = Module.findPersonAtLocation([10.2, 156.5]);
console.log('Found someone! Their name is ' + person.name + ' and they are ' + person.age + ' years old');

Advanced class concepts

Raw pointers

Because raw pointers have unclear lifetime semantics, embind requires their use to be marked with allow_raw_pointers.

For example:

class C {};
C* passThrough(C* ptr) { return ptr; }
EMSCRIPTEN_BINDINGS(raw_pointers) {
    class_<C>("C");
    function("passThrough", &passThrough, allow_raw_pointers());
}

Note

Currently the markup serves only to whitelist smart pointer use, and show that you’ve thought about the use of the raw pointers. Eventually we hope to implement Boost.Python-like raw pointer policies for managing object ownership.

External constructors

There are two ways to specify constructors for a class.

The zero-argument template form invokes the natural constructor with the arguments specified in the template. For example:

class MyClass {
public:
  MyClass(int, float);
  void someFunction();
};

EMSCRIPTEN_BINDINGS(external_constructors) {
  class_<MyClass>("MyClass")
    .constructor<int, float>()
    .function("someFunction", &MyClass::someFunction)
    ;
}

The second form of the constructor takes a function pointer argument, and is used for classes that construct themselves using a factory function. For example:

class MyClass {
  virtual void someFunction() = 0;
};
MyClass* makeMyClass(int, float); //Factory function.

EMSCRIPTEN_BINDINGS(external_constructors) {
  class_<MyClass>("MyClass")
    .constructor(&makeMyClass, allow_raw_pointers())
    .function("someFunction", &MyClass::someFunction)
    ;
}

The two constructors present exactly the same interface for constructing the object in JavaScript. Continuing the example above:

var instance = new MyClass(10, 15.5);
// instance is backed by a raw pointer to a MyClass in the Emscripten heap

Smart pointers

To manage object lifetime with smart pointers, embind must be told about the smart pointer type.

For example, consider managing a class C‘s lifetime with std::shared_ptr<C>. The best way to do this is to use smart_ptr_constructor() to register the smart pointer type:

EMSCRIPTEN_BINDINGS(better_smart_pointers) {
    class_<C>("C")
        .smart_ptr_constructor(&std::make_shared<C>)
        ;
}

When an object of this type is constructed (e.g. using new Module.C()) it returns a std::shared_ptr<C>.

An alternative is to use smart_ptr() in the EMSCRIPTEN_BINDINGS() block:

EMSCRIPTEN_BINDINGS(smart_pointers) {
    class_<C>("C")
        .constructor<>()
        .smart_ptr<std::shared_ptr<C>>()
        ;
}

Using this definition, functions can return std::shared_ptr<C> or take std::shared_ptr<C> as arguments, but new Module.C() would still return a raw pointer.

unique_ptr

embind has built-in support for return values of type std::unique_ptr.

Custom smart pointers

To teach embind about custom smart pointer templates, you must specialize the smart_ptr_trait template.

Non-member-functions on the JavaScript prototype

Methods on the JavaScript class prototype can be non-member functions, as long as the instance handle can be converted to the first argument of the non-member function. The classic example is when the function exposed to JavaScript does not exactly match the behavior of a C++ method.

struct Array10 {
    int& get(size_t index) {
        return data[index];
    }
    int data[10];
};

val Array10_get(Array10& arr, size_t index) {
    if (index < 10) {
        return val(arr.get(index));
    } else {
        return val::undefined();
    }
}

EMSCRIPTEN_BINDINGS(non_member_functions) {
    class_<Array10>("Array10")
        .function("get", &Array10_get)
        ;
}

If JavaScript calls Array10.prototype.get with an invalid index, it will return undefined.

Deriving from C++ classes in JavaScript

If C++ classes have virtual or abstract member functions, it’s possible to override them in JavaScript. Because JavaScript has no knowledge of the C++ vtable, embind needs a bit of glue code to convert C++ virtual function calls into JavaScript calls.

Abstract methods

Let’s begin with a simple case: pure virtual functions that must be implemented in JavaScript.

struct Interface {
    virtual void invoke(const std::string& str) = 0;
};

struct InterfaceWrapper : public wrapper<Interface> {
    EMSCRIPTEN_WRAPPER(InterfaceWrapper);
    void invoke(const std::string& str) {
        return call<void>("invoke", str);
    }
};

EMSCRIPTEN_BINDINGS(interface) {
    class_<Interface>("Interface")
        .function("invoke", &Interface::invoke, pure_virtual())
        .allow_subclass<InterfaceWrapper>("InterfaceWrapper")
        ;
}

allow_subclass() adds two special methods to the Interface binding: extend and implement. extend allows JavaScript to subclass in the style exemplified by Backbone.js. implement is used when you have a JavaScript object, perhaps provided by the browser or some other library, and you want to use it to implement a C++ interface.

Note

The pure_virtual annotation on the function binding allows JavaScript to throw a helpful error if the JavaScript class does not override invoke(). Otherwise, you may run into confusing errors.

extend example

var DerivedClass = Module.Interface.extend("Interface", {
    // __construct and __destruct are optional.  They are included
    // in this example for illustration purposes.
    // If you override __construct or __destruct, don't forget to
    // call the parent implementation!
    __construct: function() {
        this.__parent.__construct.call(this);
    },
    __destruct: function() {
        this.__parent.__destruct.call(this);
    },
    invoke: function() {
        // your code goes here
    },
});

var instance = new DerivedClass;

implement example

var x = {
    invoke: function(str) {
        console.log('invoking with: ' + str);
    }
};
var interfaceObject = Module.Interface.implement(x);

Now interfaceObject can be passed to any function that takes an Interface pointer or reference.

Non-abstract virtual methods

If a C++ class has a non-pure virtual function, it can be overridden — but does not have to be. This requires a slightly different wrapper implementation:

struct Base {
    virtual void invoke(const std::string& str) {
        // default implementation
    }
};

struct BaseWrapper : public wrapper<Base> {
    EMSCRIPTEN_WRAPPER(BaseWrapper);
    void invoke(const std::string& str) {
        return call<void>("invoke", str);
    }
};

EMSCRIPTEN_BINDINGS(interface) {
    class_<Base>("Base")
        .allow_subclass<BaseWrapper>("BaseWrapper")
        .function("invoke", optional_override([](Base& self, const std::string& str) {
            return self.Base::invoke(str);
        }))
        ;
}

When implementing Base with a JavaScript object, overriding invoke is optional. The special lambda binding for invoke is necessary to avoid infinite mutual recursion between the wrapper and JavaScript.

Base classes

Base class bindings are defined as shown:

EMSCRIPTEN_BINDINGS(base_example) {
    class_<BaseClass>("BaseClass");
    class_<DerivedClass, base<BaseClass>>("DerivedClass");
}

Any member functions defined on BaseClass are then accessible to instances of DerivedClass. In addition, any function that accepts an instance of BaseClass can be given an instance of DerivedClass.

Automatic downcasting

If a C++ class is polymorphic (that is, it has a virtual method), then embind supports automatic downcasting of function return values.

class Base { virtual ~Base() {} }; // the virtual makes Base and Derived polymorphic
class Derived : public Base {};
Base* getDerivedInstance() {
    return new Derived;
}
EMSCRIPTEN_BINDINGS(automatic_downcasting) {
    class_<Base>("Base");
    class_<Derived, base<Base>>("Derived");
    function("getDerivedInstance", &getDerivedInstance, allow_raw_pointers());
}

Calling Module.getDerivedInstance from JavaScript will return a Derived instance handle from which all of Derived‘s methods are available.

Note

Embind must understand the fully-derived type for automatic downcasting to work.

Overloaded functions

Constructors and functions can be overloaded on the number of arguments, but embind does not support overloading based on type. When specifying an overload, use the select_overload() helper function to select the appropriate signature.

struct HasOverloadedMethods {
    void foo();
    void foo(int i);
    void foo(float f) const;
};

EMSCRIPTEN_BINDING(overloads) {
    class_<HasOverloadedMethods>("HasOverloadedMethods")
        .function("foo", select_overload<void()>(&HasOverloadedMethods::foo))
        .function("foo_int", select_overload<void(int)>(&HasOverloadedMethods::foo))
        .function("foo_float", select_overload<void(float)const>(&HasOverloadedMethods::foo))
        ;
}

Enums

Embind‘s enumeration support works with both C++98 enums and C++11 “enum classes”.

enum OldStyle {
    OLD_STYLE_ONE,
    OLD_STYLE_TWO
};

enum class NewStyle {
    ONE,
    TWO
};

EMSCRIPTEN_BINDINGS(my_enum_example) {
    enum_<OldStyle>("OldStyle")
        .value("ONE", OLD_STYLE_ONE)
        .value("TWO", OLD_STYLE_TWO)
        ;
    enum_<NewStyle>("NewStyle")
        .value("ONE", NewStyle::ONE)
        .value("TWO", NewStyle::TWO)
        ;
}

In both cases, JavaScript accesses enumeration values as properties of the type.

Module.OldStyle.ONE;
Module.NewStyle.TWO;

Constants

To expose a C++ constant() to JavaScript, simply write:

EMSCRIPTEN_BINDINGS(my_constant_example) {
    constant("SOME_CONSTANT", SOME_CONSTANT);
}

SOME_CONSTANT can have any type known to embind.

Memory views

In some cases it is valuable to expose raw binary data directly to JavaScript code as a typed array, allowing it to be used without copying. This is useful for instance for uploading large WebGL textures directly from the heap.

Memory views should be treated like raw pointers; lifetime and validity are not managed by the runtime and it’s easy to corrupt data if the underlying object is modified or deallocated.

#include <emscripten/bind.h>
#include <emscripten/val.h>

using namespace emscripten;

unsigned char *byteBuffer = /* ... */;
size_t bufferLength = /* ... */;

val getBytes() {
    return val(typed_memory_view(bufferLength, byteBuffer));
}

EMSCRIPTEN_BINDINGS(memory_view_example) {
    function("getBytes", &getBytes);
}

The calling JavaScript code will receive a typed array view into the emscripten heap:

var myUint8Array = Module.getBytes()
var xhr = new XMLHttpRequest();
xhr.open('POST', /* ... */);
xhr.send(myUint8Array);

The typed array view will be of the appropriate matching type, such as Uint8Array for an unsigned char array or pointer.

Using val to transliterate JavaScript to C++

Embind provides a C++ class, emscripten::val, which you can use to transliterate JavaScript code to C++. Using val you can call JavaScript objects from your C++, read and write their properties, or coerce them to C++ values like a bool, int, or std::string.

The example below shows how you can use val to call the JavaScript Web Audio API from C++:

Note

This example is based on the excellent Web Audio tutorial: Making sine, square, sawtooth and triangle waves (stuartmemo.com). There is an even simpler example in the emscripten::val documentation.

First consider the JavaScript below, which shows how to use the API:

// Get web audio api context
var AudioContext = window.AudioContext || window.webkitAudioContext;

// Got an AudioContext: Create context and OscillatorNode
var context = new AudioContext();
var oscillator = context.createOscillator();

// Configuring oscillator: set OscillatorNode type and frequency
oscillator.type = 'triangle';
oscillator.frequency.value = 261.63; // value in hertz - middle C

// Playing
oscillator.connect(context.destination);
oscillator.start();

// All done!

The code can be transliterated to C++ using val, as shown below:

#include <emscripten/val.h>
#include <stdio.h>
#include <math.h>

using namespace emscripten;

int main() {
  val AudioContext = val::global("AudioContext");
  if (!AudioContext.as<bool>()) {
    printf("No global AudioContext, trying webkitAudioContext\n");
    AudioContext = val::global("webkitAudioContext");
  }

  printf("Got an AudioContext\n");
  val context = AudioContext.new_();
  val oscillator = context.call<val>("createOscillator");

  printf("Configuring oscillator\n");
  oscillator.set("type", val("triangle"));
  oscillator["frequency"].set("value", val(261.63)); // Middle C

  printf("Playing\n");
  oscillator.call<void>("connect", context["destination"]);
  oscillator.call<void>("start", 0);

  printf("All done!\n");
}

First we use global() to get the symbol for the global AudioContext object (or webkitAudioContext if that does not exist). We then use new_() to create the context, and from this context we can create an oscillator, set() its properties (again using val) and then play the tone.

The example can be compiled on the Linux/Mac OS X terminal with:

./emcc -O2 -Wall -Werror --bind -o oscillator.html oscillator.cpp

Built-in type conversions

Out of the box, embind provides converters for many standard C++ types:

C++ type JavaScript type
void undefined
bool true or false
char Number
signed char Number
unsigned char Number
short Number
unsigned short Number
int Number
unsigned int Number
long Number
unsigned long Number
float Number
double Number
std::string ArrayBuffer, Uint8Array, Uint8ClampedArray, Int8Array, or String
std::wstring String (UTF-16 code units)
emscripten::val anything

For convenience, embind provides factory functions to register std::vector<T> (register_vector()) and std::map<K, V> (register_map()) types:

EMSCRIPTEN_BINDINGS(stl_wrappers) {
    register_vector<int>("VectorInt");
    register_map<int,int>("MapIntInt");
}

Performance

At time of writing there has been no comprehensive embind performance testing, either against standard benchmarks, or relative to WebIDL Binder.

The call overhead for simple functions has been measured at about 200 ns. While there is room for further optimisation, so far its performance in real-world applications has proved to be more than acceptable.