Switch branches/tags
Nothing to show
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
319 lines (259 sloc) 15.6 KB


Java bytecode to C++ transpiler to ahead-of-time compile Java code. This can be useful if you want to target operating systems that don't allow setting the executable flag on memory pages.


Jack tries to provide a reasonable runtime environment for Java applications. The following features should be supported:

  • Java 1.5 language compatibility (no annotations)
  • garbage collection
  • reflection (with limitations, see below)
  • threads
  • exceptions (with limitations, see below)
  • JNI (with limitations, see below)
  • alternative native code interface


  • loading bytecode at runtime
  • full JRE, say good bye to all those EE classes
  • Java memory model for the most part (could potentially honor volatile fields/vars)
  • anything else that's not listed under goals

Code Overview

Jack's code is found in the jack/ folder and is organized as follows.


  • com.badlogic.jack: main package, containing the main entry point, Jack
  • classes that store additional information for classes, methods and fields.
  • com.badlogic.jack.generators: classes that generate parts of the c++ output.


  • is the main entry point. It generates header and implementation files from a directory full of .class files.
  • As a first step metadata (aka info) is collected for all classes, methods and fields. See the package
  • Based on that info, header files are generated for each class, see com.badlogic.jack.generators.HeaderGenerator
  • Next, implementation files are generated for each class, see com.badlogic.jack.generators.ImplementationGenerator. The process goes like this:
    • Methods are emitted, via com.badlogic.jack.generators.MethodGenerator, which in turn uses a series of other generators. During this phase additional information is added to the info classes, such as string literals.
    • <clinit> is emitted, via com.badlogic.jack.generators.ClinitGenerator, which emits static field initialization codes, string literal initialization code and the original <clinit> method body of the class if present.
    • Header and static fields sections are emitted, via com.badlogic.jack.generators.StaticsGenerator.
  • Finally, runtime startup functions and reflection information is generated, via com.badlogic.jack.generators.RuntimeGenerator.

Runtime Library

Jack contains a minimal runtime library in runtime/jack-kernel. It heavily borrows from Avian VM. The runtime/jack-tests folder contains a few basic tests that check language feature implementation and the runtime library. Neither of them links to the normal JRE library since they themselve implement "All The Things".


Currently Jack is tested through compiling and running parts of the runtime library and tests. The simplest way to do this works as follows:

  • Import all three projects (jack/, runtime/jack-kernel/, runtime/jack-tests) into Eclipse. This will put .class files for both library implementations in their respective bin/ folder.
  • Excecute Jack, passing in three parameters specifying:
    • the directory the .class files reside in (e.g. ../runtime/jack-kernel/bin;../runtime/jack-tests/bin, to include both the runtime library and the tests)
    • the directory the .java files reside in (e.g. ../runtime/jack-kernel/src;../runtime/jack-tests/src). This is used to add Java source file lines to the C++ files for easier debugging.
    • the output directory (use native/classes for now).
  • Once Jack is done, fire up Visual C++ 2010 (Express) or Xcode and load the corresponding project (vs10, xcode).
  • Delete all the files in the classes filter/group and reimport everything from native/classes. You'll have to redo this step everytime you add a new Java source file to the projects you translate to C++.
  • Open up native/jack.cpp, mess around with the code to test classes or features, and compile, run and debug it.
  • If you change a runtime library or test Java file, execute Jack again, and wait for Visual Studio/XCode to pick up the changes, then recompile.
  • If you add or remove a file, you'll have to reimport all .cpp files from the native/classes folder.

This process is meant for developing Jack itself. End-users will get a nicer way to transpile and run their stuff.


  • class hierarchy translation
  • method body translation
  • GC (Boehm GC for now)
  • initial classpath implementation based on Avian's classpath (misses JNI parts)
  • reflection: fix instanceof, what was i thinking :p
  • reflection: class descriptors for arrays
  • custom native code bridge (add @DirectNative to class or method, implement missing native methods directly in C++, see jack/native/vm/java_lang.cpp for implementations of Object and Class native methods.
  • partial recompilation, means we only touch .h/.cpp files that actually changed. brings down c++ compile times enormously.


(in order, search for FIXME in the code)

  • array covariance, e.g. Object[] arr = new String[10]. Damn it.
  • exceptions: set signal handlers, use setjmp, finally is thankfully handled by javac
  • reflection: rest of class/constructor/method/field descriptors
  • reflection: method#invoke, newInstance, newArray, might get away with not using libffi
  • jni: minimally viable product :p
  • threads
  • add the rest of Avian's classpath + JNI implementations.
  • add unit tests, get some from Avian, see if OpenJDK has anything useful.


The following describes how the Java bytecode is translated to C++.

Class hierarchies

Java class hierarchies are directly mapped to C++ class hierarchies. That way i don't have to implement my own vtable abstraction. Some funkiness is involved to cope with bridge methods and covariant return types, which usually turn up if you have a concrete generic class, e.g. List<String>#get(0). Covariance itself is supported by C++, however, the covariant return type needs to be complete (fully defined). Forward declarations do not work, instead the full class declaration of a covariant return type is included in the header the class that has the return type covariant method. This might lead to cyclic dependencies in some ill-conceived cases.

The full Java class hierarchy is retained in the C++ class hierarchy.

Fields and methods are added as is, with prefixes (f_ for fields, m_ for methods).


We use the String implementation by Avian, which includes UTF8 handling. Strings are hence stored as 16-bit character sequences. String literals are constructed in a funky way, you don't want to know how...

Class initialization

The JVM usually calls MyClass#<clinit> the first time that class is referenced at runtime (see the Java language specs and JVM specs for more info). In an ahead-of-time compiled environment this would be to expensive as tons of methods would have to be augmented with checks whether <clinit> was already called.

Jack initializes all classes at start up, in an order that guarantees that guarantees that all classes a class depends on are initialized before the class itself is initialized.


All java primitive types are mapped to equivalent C++ types, see vm/types.h (to be extended for more platforms, looking at you long long). Reference types are always represented as pointers, e.g. a field of type java.lang.Object would translate to java_lang_Object*.

The array type is special and implemented in C++, subclassing java.lang.Object (see vm/array.h). Multidimensional arrays are just nested C++ Array instances. The transpiler will replace array access with calls to equivalent methods/operators of the C++ Array class.


Reflection data is created before class initialization. It includes class, constructor, method and field descriptors. The descriptors are actually modelled via Java class instances, see classpath/src/java/lang/reflect and classpath/src/java/lang/Class.

Only a minimal set of reflection capabilities will be exposed. These are modelled after Avian VMs classpath (but don't use Avian's implementation, just the APIs which are a subset of the JRE reflection classes). Using Java class instances has the benefit that all the pesky reflection code can be written in Java instead of C++, not including creating new class/array instances and invoking methods.

These three operations are performed in native code. The reflection classes have native methods which implement that functionality.

Special care needs to be taken for array class descriptors. The strategy would be to collect all array types in the entire program and generate class descriptors for all of those. A concrete Array implementation would then refer to the precompiled class descriptor. This might blow up with arrays created via reflection if the array is never used (e.g. assigned to a variable with an array type), but instead treated and stored as an object. Very unlikely, but could happen.

Primitive types also have class descriptors, those need to be generated and stored somewhere as well. I guess it makes sense to keep them at the same place as the array class descriptors.

Method body translation

Jack uses Soot, a Java bytecode analysis framework. It parses Java bytecode and translates it into 3-address code representation called Jimple which is like an infinite register machine representation. This has benefits over the Java bytecode stack machine representation. Stack machines are load/store heavy. 3-address code as generated by Soot is also load/store heavy, but can be expressed with a series of local variables. Those can in turn be easily optimized to use registers most of the time by any capable C++ compiler. It also makes it easier for me to directly pass arguments to methods. On top of that, many loophole and more complex optimizations can be implemented directly with Soot, e.g. elimination of many load/stores (which the c++ compiler already does for us most of the time).

Another benefit of using Soot and it's 3-address code representation is the fact that there are only 15 different statement types (plus a few dozen value types). That's a lot more managable than the 200 something JVM instructions.

The 3-address code can be translated to C++ almost directly. See all the nasty Compiler#translateXXX methods.

If the original Java source is available, Jack will add it as comments to the C++ source code on a per statement basis. This makes following and debugging the C++ translation a bit easier.


Are a strange beast. Two strategies are available:

  • setjmp/longjmp: for each try/catch block of a method, a jump address is pushed onto a thread local stack of jump addresses. When an exception is thrown, the top of the jump address stack is poped, and longjmp is called to jump to this address. Execution will continue in the catch handler which checks if the exception type can be handled. if it can, the exception handler code is executed. Otherwise the next jump address is popped from the stack. If no fitting exception handler can be found, the app will terminate. Local variables of a method containing a try/catch block must be declared as volatile for this to work, which can potentially hurt performance.
  • C++ exceptions: originally not considered as they lack finally blocks. Have to reinvestigate due to a funky finding, see next paragraph.

In both cases we don't have to care about finally blocks. Javac will inline those in the exception handlers as well as before any return statements of a method. This makes our life considerably easier.

C++ automatic stack unwinding would be a nice to have instead of the confusing mess that is setjmp/longjmp. Setjmp/longjmp seem to have better performance characteristics (not counting the volatile locals).

I guess i'll try both :)

Threading & Monitors

Threading itself should be straight forward, using whatever non-standard Windows supplies as well as Posix on any sane OS. The GC and exceptions have to be considered here. It looks like Boehm GC is up for the task, and making jump buffers thread local is trivial should i chose to use setjmp/longjmp.

Thread interruption will be interesting, especially with regards to interruptable i/o methods. This is an area where more research is needed, Avian should give me some hints.

Monitors are a different beast. Every class instance in Java is potentially a monitor. The mutex has to be stored somewhere, potentially bloating objects. Here's a nice starting point

Finally, locks need to be released in case of exceptions. Javac makes our live a bit easier again by automatically generating try/catch blocks around synchronized statements/methods:

public void test() {
   synchronized(this) {

And the corresponding Jimple representation:

   (soot.jimple.JimpleBody)     public void test()
        jack.Synchronized r0, r2;
        java.lang.Throwable $r3;

        r0 := @this: jack.Synchronized;
        r2 = r0;
        entermonitor r0;

        virtualinvoke r0.<jack.Synchronized: void test2()>();
        exitmonitor r2;

        goto label5;

        $r3 := @caughtexception;

        exitmonitor r2;

        throw $r3;


        catch java.lang.Throwable from label0 to label1 with label2;
        catch java.lang.Throwable from label3 to label4 with label2;


Given full reflection data and a non-moving GC, it should be almost easy to implement at least the most useful parts of the JNI (lock primitive arrays, get direct buffer addresses, fetch method ids and invoke them).

Native Java methods need a way to find and invoke their C/JNI counter part. Here's how that could be done naively:

  • for each native method of a Java class, add a corresponding function pointer field in the C++ class.
  • for each native method, generate a C++ body in the transpiler that:
    • makes sure any locks are obtained if the method is synchronized
    • checks if the function pointer field is set
      • if no, get the address via libdl or whatever Win32 provides
      • if yes, fetch the address from the field
    • invoke the JNI function

The arguments to the JNI function can be the direct pointers to any Java class instance. We just hide them behind suitable definitions of jstring, jarray, jobject and so on. JNIEnv has to be implemented at least to some degree so we can pass it in as well. A Java class that implements most of JNIEnv might actually work using the reflection data from other classes and one or two native methods.

Care has to be taken if exceptions are thrown in the native code, or if Java code is called from the JNI code that itself throws an exception. Research this.

As long as the GC is non-moving we should be mostly safe when passing Java objects directly. Threading might be a fun cause for heisenbugs in that case. Research this.

Besides a minimal JNI implementation, it would be nice to interface with C/C++ more directly. Figure out a nice way to do that.

Garbage Collection

Out of laziness Jack uses the Boehm GC for now. Mono uses/used it, so it's good enough for this as well. All classes derrive from gc to override the new operator. Primitive arrays are instantiated via GC_MALLOC_ATOMIC, anything else is allocated via GC_MALLOC. Since the Boehm GC is non-moving we can be a bit lazy on the JNI side of things.