Metaprogramming API

TeaVM was designed to create web applications. One important requirement for web applications is their size. Nobody will use a compiler which produces tens of megabytes of JS. That’s why tools like TeaVM and GWT perform advanced optimizations to reduce code size. Unfortunately, it’s impossible to make these optimizations reflection-friendly. Even if TeaVM implemented reflection, any attempt to use it could lead to “dependency explosion”, i.e. huge JS file. That’s why TeaVM does not provide reflection support. Instead, it comes with its own replacement, called metaprogramming API.


Unlike Java reflection, TeaVM metaprogramming does not work in runtime. It allows you to write a code that runs in the compile-time. TeaVM can query from compiler much about classes, methods and fields and generate JS code that will be executed at run-time.

One, familiar with GWT can find this approach very close to deferred binding (aka generators). However, TeaVM’s approach is slightly more powerful, as you can see below.

Quick start

Let’s start by writing a method which reads field called foo or returns null if such field is not available.

    public static Object getFoo(Object obj) {
        return getFoo(obj.getClass(), obj);

    private static native Object getFooImpl(Class<?> cls, Object obj);

    private static void getFooImpl(ReflectClass<Object> cls, Value<Object> obj) {
        ReflectField field = cls.getField("a");
        if (field != null) {
            exit(() -> field.get(obj));
        } else {
            exit(() -> null);

Here, our method simply delegates all work to getFooImpl, which is a native method marked with @Meta annotation. This annotation does all magic. It tells compiler to generate body of getFooImpl by invoking another getFooImpl method with slightly different signature. It’s quite easy to find the difference: returning value must be void, the ReflectClass argument must correspond to Class and Value argument must correspond to any other argument. Note that only one argument can be of ReflectClass type.

GWT’s entry point to generators are special classes created by GWT.create method, which requires single class literal argument (class variable are not allowed). TeaVM’s approach looks similar, however, it does not restrict developer to the literal argument. You can get class from anywhere, for example, by calling Object.getClass().

The second getFooImpl which has actual Java code is executed in run-time. It provides access to classes available to compiler via API which resembles Java reflection, with Reflect prefix added to each class. We use it to find a field called “a”.

To get field’s value and return it from method, we use Metaprogramming.exit (or just exit, since there’s a convention to always statically import entire Metaprogramming class). This method takes lambda which will be again executed in run-time and causes original getFoo method to return evaluated expression.

We can now test this method:

class A {
    private String foo;
    A(String foo) { = foo; }
class B {

public static void main(String[] args) {
    System.out.println(getFoo(new A("barbaz"));
    System.out.println(getFoo(new B()));

So, we can freely switch between compile-time and run-time. To switch from runtime to compile-time, we declare a pair of methods with equal names and similar signatures, mark first one with @Meta annotation. To switch back, we call special method like exit or emit.

GWT requires you to generate Java code in compile-time which it further compiles to JavaScript. TeaVM works with bytecode so it would be hard to write the code that generates byte-code. Mataprogramming API does byte-code generation for you using lambdas as templates.

@Meta annotation

@Meta annotation must be put on a static native method. Another static non-abstract non-native method with the same name must exist in the same class. Its signature must correspond to methods of the original method:

  • It should be void no matter which type is returned by the original method.
  • Any Class argument can be mapped to ReflectClass argument.
  • Any other argument must be mapped to Value argument.
  • There can be at most one ReflectClass argument.

@Meta annotation causes the first method’s body to be generated by the second method which runs in compile-time.

Emitting bytecode

There are several methods which can be called from compile-time to generate code. All they take lambda as an argument, and simply write the lambda’s body to the generated code. Unlike normal lambdas, template lambdas have some restrictions over variables they can capture.

  • It’s allowed to capture primitive values (numbers, strings).
  • It’s allowed to capture Value, ReflectClass, ReflectField, ReflectMethod.
  • Any other value is disallowed.

These captured variables act as template parameters.

The main (and the simplest) one is Metaprogramming.emit. It simply writes template as-is. exit method writes template and additional return statement which returns expression evaluated by lambda. lazy does not write template immediately. Instead, it produces Value which is written as soon as accessed by another lambda.

Passing data between template lambdas

In the real world templates can’t be isolated. Value produced by one template may be needed to another template. TeaVM uses Value for this purpose. Methods like emit and lazy produce Value. Value can’t be read at compile-time. The only way to read Value is to capture it by another lambda and call get method there. For example:

Value a = emit(() -> 2);
Value b = emit(() -> a.get() + 3);
exit(() -> b.get());

Reflection restrictions

You can use metaprogramming reflection much like usual Java reflection. However, it imposes several restrictions:

  • You can’t search and enumerate fields and methods of a class in template lambdas.
  • You can’t get and set value of ReflectField in compile-time.
  • You can’t do ReflectMethod.invoke methods in compile-time.


Of course, there is an alternative to reflection proxies. You can call Metaprogramming.proxy method which accepts interface and InvocationHandler.

Improve this page