Under the Hood: Java Bytecode

So, now that we have cleared up how a Java application is executed, I can talk about bytecode manipulation, a technique to modify the compiled classes before execution.
There are two possible times when modification can occur: at build time, modifying the class filers after compilation, or at runtime, using a class loader that modifies the bytes of the class before passing them on and actually defining the class from it. Both have their advantages: Build time simplifies the structure at runtime; your application runs like an ordinary Java application, except that not all of its classes were ordinarily created from Java source code. Runtime gives you the possibility of applying different modifications in different executions, which could be used to, say provide compatibility to different environments. More importantly in my opinion, this allows to create completely new classes based on user input!

Let me give you an example: You make an application that draws the graph of a function the user types into a text field, so basically you take y = f(x) for each coordinate visible in the plot; that's a lot of times and you want to compute your function very quickly. So, instead of parsing the expression, building a tree from it, inserting a new value for x, and then evaluating the function, wouldn't it be nice to parse the expression into a real Java class that is executed on the tested and optimized JVM, with the possibility that your code is even JIT-compiled into machine code for an extra boost?

And guess what - that's not only possible, but even relatively easy, once you've grasped what java byte code looks like, so that's what were going to look at now. It's pretty straight forward that a class consists of fields and methods, and when you think about the fact that inner classes have separate class files, it becomes clear that there must be some information about enclosing class and inner classes, too.

The code of methods is of course where it gets interesting. The JVM is a stack based machine, which means that values are kept in a last-in-first-out data structure, and every operation pops a specific number of values from the stack and pushes its results onto the stack again. Note that this data is not related to local variables or fields; here, we're basically talking about the steps taken in evaluating an expression.
Also, type names have a special form in class files. Where a class name is needed, a string of the form java/lang/Object is used; where it could be a primitive type, too, this is wrapped into L...;, where L somehow was chosen to mean "object type", and ; denotes the end of the type name. As another example, long, which can't use the L, was given J as the identifier, and boolean uses Z; arrays prepend a [ for every dimension. A method, for example int indexOf(char ch) or String substring(int beginIndex, int endIndex) would be described as (C)I or (II)Ljava/lang/String; - note that the name is not part of the signature but stored separately.

And that's all you really need to know, except for the few less than 256 possible commands (opcodes) that the JVM knows. The details of how all the strings are stored in a classfile is hidden from you when you use ASM to generate bytecode.

Now let's look at some bytecode, as this is where it gets interesting. A simple example first:

package Example;

class Example {

    int a;

    void example(int a) {

        this.a = a;

    }

}

...and the bytecode for example():


example(I)V
ALOAD 0
ILOAD 1
PUTFIELD example/Example.a : I
RETURN


MAXSTACK = 2
MAXLOCALS = 2

ALOAD pushes a local variable onto the stack. In this example, there are - surprise - two! Not only is there parameter a, there's also the implicit this of every nonstatic method. When both are on the stack (explaining the MAXSTACK entry above), PUTFIELD can take the second value and store it into the named field of the first reference on the stack, which better be an instance of Example.

Let's take a more complex example:

package Example;

class Example {

    int a;

    Example(int a) {

        this.a = a;

    }

}

...doesn't seem so? Look at the bytecode:

(I)V
ALOAD 0
INVOKESPECIAL java/lang/Object.<init> ()V
ALOAD 0
ILOAD 1
PUTFIELD example/Example.a : I
RETURN

MAXSTACK = 2
MAXLOCALS = 2

First of all, as you see, constructors are internally void methods with special names. The special code here is ironically the one that we didn't write ourselves: the superconstructor invocation.

Okay, that's it for today! Thanks for reading!

Under the Hood: Classloading

I'm kind of a tinkerer... I like exploring new aspects of things I already know, or new things altogether, even if there is no apparent gain in it. And I bet many of you (well, if there were many readers in the first place...) share this trait with me, because it is one of the things that most engineers share.

The "thing" that I explored in the last week is Java itself, and the new aspect is the Java Bytecode, along with a few intricacies of Java that most programmers never have to worry about, namely bytecode manipulation and class loading.

Just in case you are not aware of it, I'll first describe the lifecycle of a HelloWorld program from writing it, up to its execution. It all starts, of course, with a source file. In the standard case, that's HelloWorld.java, but there are other languages, like Groovy, Scala and JRuby, that all run on the JVM.
This leads us directly to our next step, compilation. The JVM is, as the name suggests, a machine, just like any computer (except it's "virtual"), and has an instruction set it can execute. The Java compiler javac translates the more or less human readable source code into a .class file that contains code executable by the JVM, as do other compilers like gcj or scalac, except that in case of scalac, the source is written in a different language of course.

One aside about gcj: there are a few Java compiler vendors around, but by far not as many as for other languages like C. The answer is, again, the JVM: while a C compiler creates code for a specific architecture and OS, and therefore each new platform needs a new compiler (although much logic can be reused, of course), Java targets only a single platform, the JVM. Therefore, there's no inherent need for numerous compilers and Java doesn't have the problem of incompatible dialects. Still, there is some competition; for example, gcj is an open source alternative to javac. More importantly, the part of Java that is platform specific is the JVM. In Java's early days, Microsoft had its own Java implementation, which was kind of crappy compared to Sun's and was thankfully soon discontinued. Additionally, there are VMs implemented in Java, or ones that target platforms where Oracle does not provide support, or ones with smaller memory requirements, etc.

So, now we have a class file consisting of Java bytecode, independent of the language we started with. Using the java command, we can launch the main method of that class, but this involves more steps than are apparent!
Every java application has a system class loader. A ClassLoader is responsible for finding class files, reading them, and loading the classes into memory. In the most basic case, the classloader has a list of URLs (the classpath) that contains locations where classes are found. By default, this classpath contains the JRE classes (plus some more) and those found in the current directory, e.g. HelloWorld.class. Other classloaders delegate to a parent classloader, but may add new search locations or other means of getting the bytes that make up a class.
Now, the classloader loads the main class of the application, and the main thread starts, executing the main method. This method may in turn require other classes which are then loaded by the same classloader that also loaded the current class. This means by extension that normally, all classes are loaded by the same classloader that also loaded the main class. However, it's also possible to create and invoke classloaders directly, creating a class that was loaded by a different classloader than the current class. This is important because of two things. Firstly, as I said, the new classloader can add new ways of loading classes (the main reason for using a separate classloader), and secondly because it can introduce subtle problems when multiple classloaders define classes with the same name: they are not compatible, and you get a ClassCastException when you try to mix them. Don't worry, there's an easy solution: define an interface that is loaded by a common parent classloader, and only cast to the interface.

And finally, our main method can execute. And since this post was already long enough, I'll tell the story about bytecode manipulation with ASM next time ;)

Entities and values

As so often in programming, one big question arose when I worked on my Undo/Replication framework: What is an entity, and what is a value?

I use the word entity to make clear that I mean something different from objects. An object is something identified by an address pointer, whereas for a primitive value, its bytes in memory directly represent the value instead of an address. What I mean with entity/value is best explained by an example from databases:

Suppose you have a database table storing the data of a company's employees. There is a surname, a prename, a birth date, and so on. Besides others, there's an address. The simplest approach is to handle each of these items as a table column: Surname and prename are strings, the birthdate uses a special date column type. The address can also be simply taken as a string, but there are other ways:

An address is composed of a street name, a number, probably also an appartment number, country, state etc. It could be beneficial to store all these data in a table on its own, give it an ID (the database way of saying address), and only reference this ID in the employee table.

Don't forget the third way! You could use all those columns, but put them directly into the employee table, so there would be prename, surname, birth date, street, stree no., appartment no., etc.

What's the best? Of course, it depends! Variant one is easy to use, but lacks intrinsic validity checks: When the database doesn't know that something should be an (appartment) number, how should it check it's even a number? Variant two can handle a very specific case "better": two workers have the same address. In this case, the same ID can be referenced multiple times, so that both employees not only have equal addresses but also have the same one. In this particular case, the benefit is not too big, but there are use cases where it's important that something can be referenced. About the third, it really combines the downsides of the both above... I personally see no benefit in using it. It there are, I'm sorry I didn't see them.

Now, what should have triggered your attention was the word "ID". As soon as something has an identity, it's an entity. Note how the prename is a String, which is an Object type in Java, but only a value. On the other hand, the address in case 2 has an identifier and is an entity.

And now we're back: What I do is to assign IDs to the objects I want to replicate. The objects have values which can be changed, and I transmit changes like "property XY of object 01 changed from 'ab' to 'cd'."

And now the big question: what about collections? Should it be "element 'ab' was added to list 01" (i.e. they are entities) or "element 'ab' was added to list XY of object 01?"

Replicating Game States - JGroups

So now I've got it. My goals in implementing Replication was to keep it out of the core functionality while still making it easy to use without worrying about replication yourself. How do you do that? By providing a general interface to other software (in my case, that's a listener) and implementing a specific one for the task at hand.

So my History class, which stores all changes to the managed objects, has a listener for two possible events: A new modification is executed; and the "current" state is moved somewhere else (e.g. undoing something.) My replication listener does not support undo right now, because I haven't figured out how to best identify a state (which is where the program is going) yet, but that should be easy. The other thing, executing modifications, does work. That means, whenever a modification is executed locally, it is sent over the network so that the partner(s) can execute it too.

The only downside is that the partners also "execute it locally", so I had to use a trick to suppress resending these received modifications. And it works pretty well:

//initialization code
final History h = createHistory(createKey("test-history"));
final JChannel channel = new JGroupsReplicationListener(h) {
    @Override
    public void receive(Message msg) {
        Modification m = (Modification) deserialize(msg.getBuffer());
        if(m instanceof Creation) id = ((Creation) m).getId();
        super.receive(msg);
    }
}.getChannel();

//executing the actions (creating an object)
h.pushHistoryForThread();
try {
    TestBean t = (TestBean) h.getObjectStore().get(Creation.create(new TestBean()));
    t.setA(1);
    t.setB(1);
} finally {
    h.popHistoryForThread();
}


//printing the results
System.out.printf("current state: %s%n", h.getObjectStore().get(id));


I made this test a GUI application so that I can control the timing on multiple virtual machines. The public void receive(Message msg)seen above is only needed for this Test, again. It's not necessary to do something like that to get the replication. The code in the middle looks exactly the same as the example last time, except I haven't hidden some of the details in a setTest() method.

Now what's probably the most interesting is how the networking works; let me tell you, I didn't write any! As hinted by me before, and in the title of course, is that I have tried JGroups to do this. In JGroups, you create Channels and let them join into a cluster; all channels in the cluster can receive and send messages, either to only only one recipient, or to all at the same time. JGroups provides a configurable protocol stack that takes care of reliability, synchronity and such things. For my tests, I have used a UDP-based protocol stack. For wide-area connections, a TCP-based approach is probably easier.

Replicating Game States - Working code!

Before you get too excited - no, there's no multiplayer yet (well, who would have guessed that...). There is, however, working code for undoing and redoing actions (even with multiple, branching histories), and even for replicating the state on multiple virtual machines.

The key to make such a library is that it looks like ordinary Java code from the outside, so that others can use your code without wondering why it looks so weird. Let's take a look:

//Test t = new Test(); is an attribute

//replaced angle with square brackets because of HTML markup
List[StateStamp] states = new ArrayList[StateStamp]();
History h = createHistory(createKey("test"));
h.pushHistoryForThread();
try {
    states.add(h.getCurrentState()); // 0
    print();

    setTest(t);
    states.add(h.getCurrentState()); // 1
    print();

    getTest().setA(1);
    states.add(h.getCurrentState()); // 2
    print();

    getTest().setB(1);
    states.add(h.getCurrentState()); // 3
    print();

    Deletion.delete(getTest());
    states.add(h.getCurrentState()); // 4
    print();

    h.goToState(states.get(3));
    print();
    h.goToState(states.get(2));
    print();
    h.goToState(states.get(1));
    print();
    h.goToState(states.get(0));
    print();
} finally {
    h.popHistoryForThread();
}


It's probably easy to to guess that print(); prints the current state of the test object. Besides that, the only code that looks different from normal operations is marked in bold. Most of it is only necessary for this test case: storing the previous states so that you can go back. Really mandatory is only the colored code (and the actual functional code of course, but that doesn't count.)

Of course, a history for the modifications must be created. Then, the history is set for the thread: Code running in the thread (that is, everything in this method, and nothing more) can access the history without passing it as a parameter (which would make using the library awkward) or declaring it as a static variable somewhere (which is ugly from an architecture standpoint.) Everything after this is wrapped in try/finally, so that the last method is not skipped in case of an exception: this removes the association of the history with the current Thread.

Of course, implementing setTest(), setA() etc. is different from a usual simple setter. But the point here is that using the code is as simple as shown here. The result is this:

Test@732A54F9[a=0, b=0] not in store
Test@7A6D084B[a=0, b=0] in store
Test@7A6D084B[a=1, b=0] in store
Test@7A6D084B[a=1, b=1] in store
Test@7A6D084B[a=1, b=1] not in store
Test@15301ED8[a=1, b=1] in store
Test@15301ED8[a=1, b=0] in store
Test@15301ED8[a=0, b=0] in store
Test@15301ED8[a=0, b=0] not in store


You see here how the test object runs through different states up to its deletion, then goes back to its initial state. What's not shown here, but is working, is adding another branch to this history. Of course, the actual objects have only one state at a time, but the history can be switched arbitrarily between multiple histories:

This is the actual test case I ran: I went back to the state after creating the test object, changed its a value, and then switched over to the state before, and then after, deleting the test again.

I have code for replication of states going, but it's not yet as clean as for undo: a lot of handling is necessary for the network, which should be transparent to the user. I'll show it once I'm finished, but in the meantime you can look here.

Quantum Mechanics and Magic AI

Okay, the title is kind of a stretch, but Quantum Mechanics sounds so much cooler than probability...

So what do I want to talk about? One thought that I often have is that the computer has the advantage of perfect memory. If it once views your hand, it can perfectly memorize all the cards in it. It follows that it can say that as long as a card doesn't leave your hand, you still have it in hand. But when the game progresses, things get more vague: You shuffle cards into your library without revealing them before. Here, probability comes into play.

I assume that the computer knows my deck list for simplicity. (It could even use heuristics to guess what cards might be in your deck, but that only complicates the matter.) At the beginning of the game, before even drawing your opening hand, each card in the library has an equal chance of being one of the, say, 60 cards in your deck. For example, in your mono green deck, a card has a 24 in 60 chance of being a forest. These chances don't change as you draw seven cards, at least from the computer's point of view. Even so, the computer can say that the probability of you having a Giant Growth in hand is (# Giant Growth in deck)/(# cards in deck)*(# cards in hand), and it can have "fear" that you might play that card during combat. The greater the probability, the greater the fear.

Now comes the "collapse of the wave function": the computer observes the cards in your hand. (You see, I can even use QM terminology here ;)) Suddenly, the probability of every of the cards in your hand becomes 100% for the card the AI has observed. Technically, as the hand is not a sorted zone, the AI should not remember which card is which. Let's say you have 4 different cards in hand, then the AI can assume that every of the cards has a 25% probability of being any of these cards.

When you now shuffle one card back into your library, nothing really changes, except that there's only 3 cards for a total of 75% per previously observed card.

I hope that it's clear what I'm saying. I have the feeling to make too many words about a simple concept, yet at the same time I feel that all this seems abstract and not very understandable... well, I should have made some images, but I'm too lazy...

Let me end with this: Magic is a game of uncertainty, and luckily the computer has the capabilities to process these. When an AI can make decisions based on what it sees, why not on what it doesn't see? Assigning these possibilities is pretty simple; every card in the game has a total of 100% of being some card from the deck list, and every card in the deck list is represented to 100% among all cards in the game.
The problem is to design the AI to use that information; it's often hard enough to process the known information, so even more the unknown. But in principle, there's no difference. And even if it is too hard, there are some shortcuts: If you have only one Morph card and play a Morph card, the AI knows which it is, even though it's face down. Such probability collapses can happen all the time, and it would be a waste to let them go unconsidered.

Replicating Game States - revised

The last time I talked about my Undo system, which should also allow multiplayer through its structure of storing actions that can be transmitted over the network, was a long time ago. I don't want to directly talk about this today; you'll see.

Especially one thought has gone through my mind in the last weeks: Replicating a state across the network properly is nothing specific to Magic - so why do I implement it in the Core of Laterna Magica? I thought about it, and there doesn't seem to be a specific reason for doing so. Another result of that thought was that there are probably complete libraries out there which do exactly what I want. Well, that's unfortunately not the case, so I'm stuck/blessed with implementing it myself.

I haven't really gotten into development yet. I have a rough skeleton, but no really fancy functionality yet. But I'm having my thoughts. I think stepping away from the application's needs and looking at the general requirements of such a library help in developing a better piece of software. I haven't done this for my current undo system, which I find kind of disappointing. Normally, I don't have problems with spotting such generalizations.

Okay, back to topic: what does a library for state replication need?

• There needs to a notion of what has happened on a higher, application level: Transactions
• Transactions consist of atomic modifications, which are the core of the state
• Optionally, they may also consist of other sub-transactions
• Transactions and modifications must be transmittable over the network. If the whole state is represented as a tree of subtransactions under a root-transaction, then there's of course the need to transmit unfinished transactions. Otherwise, it might be enough to transmit finished transactions. I think that the first is generally more flexible. It might be a good idea to give the user a chance to mark transactions as atomic so that it's only transmitted when it is finished.
• Probably the most interesting requirement is that, of course, after transmitting the state, both/all states have to be the same!
• The system should be able to detect that there are state conflicts and that a transmitted transaction can't be carried out locally.
• When we're already speaking of it, it's probably nice (but also probably impossible, so see that as an utopic remark) to resolve/merge such conflicts.

Now guess what I have an example for! Of course for the "most interesting" one in my list! It might seem at first that by defining modification, and making them granular enough to catch all actions in a system, there is no way that states may get out of sync. That's unfortunately not true. Let's take a simple case. This case actually only has a theoretical problem, but it's easier to illustrate and serves  the purpose. Our state consists of two variables, foo and bar. The user may only set foo, and when he does so, the program automatically changes bar. Now consider this:

When the state replication system transmits these two modifications, the receiver does its own set operation on bar, because foo has changed. Then, the original set operation is transmitted, setting bar another time (fortunately to the same value).

Until now, I thought that this particular case is just me wanting to be clean. Now that I wrote this, I even see more problems, real ones: If we're able to detect conflicts, then this shouldn't even run through, because the last modification should be the second but really is the third. And even if we don't: this second modification in State 2 needs to be transmitted to State 1. Even if that doesn't result in errors, which is likely, it will waste network bandwidth at least.

Now for the real problem I spotted here before: Let's replace "bar = 1" with "put a token onto the battlefield". Setting a number multiple times is redundant, but creating an Object is not. There is a very definite difference between one and two tokens, you know!

Now, let's solve these problems. There are basically two ways that can lead to this problem: a complex setter, and events.

In the first case, the setter for foo includes code that calls the setter of bar. The problem here is that our replication library can't know that setting foo doesn't really do this: it actually sets foo and bar. So, refactor foo's setter so that it does only what it says, and not more. Hide this from the user by adding another method doSomething() that does what he previously did by setting foo.

Second is events. In this case, the change to foo fires an event that leads to bar being changed as well. In this case, the simple answer that is not so simple to implement is: don't fire events while replicating the state. The more complex answer that is simpler to implement is that this is similar to the first case: Don't write events that fire on state changes and perform state changes. Instead, fire a new Event in doSomething() that triggers the change. As the state replication won't call doSomething(), everything is fine.

Both solutions shown here require changes in the application, not in the library. This might seem bad, but it's actually logical: The library may sanely assume that variables change and objects are created, and that it may do that by itself as well. If the state does not fulfill the contract that it just contains values, but instead employs complex logic that the library can't predict, the task is impossible. The library needs some way to change values without triggering application logic.

Thanks for staying wiht me, and a happy new year to all of you!