Monday, November 05, 2007

Ruby Community Seeks Autotranslator

As many of you know, Ruby was created in Japan by Yukihiro Matsumoto, and most of the core development team is still Japanese to this day. This has posed a serious problem for the Ruby community, since the language barrier between the Japanese core team and community and the English-speaking community is extremely high. Only a few members of the core team can speak English comfortably, so discussions about the future of Ruby, bug fixes, and new features happens almost entirely on the Japanese ruby-dev mailing list. That leaves those of us English speakers on the ruby-core mailing list out in the cold.

We need a two-way autotranslator.

Yes, we all know that automated translation technology is not perfect, and that for East Asian languages it's often barely readable. But even having partial, confusing translations of the Japanese emails would be better than having nothing at all, since we'd know that certain topics are being discussed. And English to JP translators do a bit better than the reverse direction, so core team members interested in ruby-core emails would get the same benefit.

I imagine this is also part of the reason Rails has not taken off as quickly in Japan as it has in the English-speaking world: the Rails core team is peopled primarily by English speakers, and the main Rails lists are all in English. Presumably, an autotranslating gateway would be useful for many such communities.

But here's the problem: I know of no such service.

There are multiple translation services, for free and for pay, that can handle Japanese to some level. Google Translate and Babelfish are the two I use regularly. But these only support translating a block of text or a URL entered into a web form. There also does not appear to be a Google API for Translate, so screen-scraping would be the only option at present.

The odd thing about this is that autotranslators are good enough now that there could easily be a generic translation service for dozens of languages. Enter in source and target languages, source and target mailing lists, and it would busily chew through mail. For closely-related European languages, autotranslators do an extremely good job. And just last night I translated a Chinese blog post using Google Translate that ended up reading as almost perfect English. The time is ripe for such a service, and making it freely available could knock down some huge barriers between international communities.

So, who's going to set it up first and grab the brass ring (or is there a service I've overlooked)?

Updated Alioth Numbers for JRuby 1.1b1

Oh yes, you all know you love the Alioth Shootout.

Isaac Gouy has updated the JRuby numbers, and modified the default comparison to be with Ruby 1.8.6 rather than with Groovy as it was before. And true to form, JRuby is faster than Ruby on 14 out of 18 benchmarks.

There are reasons for all four benchmarks that are slower:

  • pidigits is simply too short for JRuby to hit its full stride. Alioth runs it with n = 2500, which on my system doing a simple "time" results in JRuby taking 11 seconds and Ruby taking 5. If I bump that up to 5000, JRuby takes 27 seconds to Ruby's 31.
  • regex-dna and recursive-complement are both hitting the Regexp performance problem we have in JRuby 1.0.x and in the 1.1 beta. We expect to have that resolved for 1.1 final, and Ola Bini and Marcin Mielczynski are each developing separate Regexp engines to that end.
  • startup, beyond being a touch unfair for a JVM-based language right now, is actually about half our fault and half the JVM's. The JVM half is the unpleasantly high cost of classloading, and specifically the cost of generating many small temporary classes into their own classloaders, as we have to do in JRuby to avoid leaking classes as we JIT and bind methods at runtime. The JRuby half is the fact that we're loading and generating so many classes, most of them too far ahead of time or that will never be used. So there's blame to go around, but we'll never have Ruby's time for this.
Standard disclaimer applies about reading too much into these kinds of benchmarks, but it's certainly a nice set of numbers to wake up to.

Closures Prototype Applied to JRuby Compiler

A bit ago, I was catching up on my feeds and noticed that Neal Gafter had announced the first prototype of Java closures. I've been a fan of the BGGR proposal, so I thought I'd catch up on the current status and try applying it to a pain point in the JRuby source: the compiler.

The current compiler is made up of two halves: the AST walker and the bytecode emitter. The AST walker recursively walks the AST, calling appropriate methods on a set of interfaces into the bytecode emitter. The bytecode emitter, in turn, spits out appropriate bytecodes and calls back to the AST walker. Back and forth, the AST is traversed and all nested structures are assembled appropriately into a functional Java method.

This back and forth is key to the structure and relative simplicity of the compiler. Take for example the following method in ASTCompiler.java, which compiles a Ruby "next" keyword (similar to Java's "continue"):

public static void compileNext(Node node, MethodCompiler context) {
context.lineNumber(node.getPosition());

final NextNode nextNode = (NextNode) node;

ClosureCallback valueCallback = new ClosureCallback() {
public void compile(MethodCompiler context) {
if (nextNode.getValueNode() != null) {
ASTCompiler.compile(nextNode.getValueNode(), context);
} else {
context.loadNil();
}
}
};

context.pollThreadEvents();
context.issueNextEvent(valueCallback);
}

First, the "lineNumber" operation is called on the MethodCompiler, my interface for primary bytecode emitter. This emits bytecode for line number information based on the parsed position in the Ruby AST.

Then we get a reference to the NextNode passed in.

Now here's where it gets a little tricky. The "next" operation can be compiled in one of two ways. If it occurs within a normal loop, and the compiler has an appropriate jump location, it will compile as a normal Java GOTO operation. If, on the other hand, the "next" occurs within a closure (and not within an immediately-enclosing loop), we must initiate a non-local branch operation. In short, we must throw a NextJump.

In Ruby, unlike in Java, "next" can take an optional value. In the simple case, where "next" is within a normal loop, this value is ignored. When a "next" occurs within a closure, the given value becomes the local return from that invocation of the closure. The idea is that you might write code like this, where you want to do an explicit local return from a closure rather than let the return value "fall off the end":
def foo
puts "still going" while yield
end

a = 0
foo {next false if a > 4; a += 4; true}

...which simply prints "still going" four times.

The straightforward way to compile this non-local "next" would be to evaluate the argument, construct a NextJump object, swap the two so we can call the NextJump(IRubyObject value) constructor with the given value, and then raise the exception. But that requires us to juggle values around all the time. This simple case doesn't seem like such a problem, but imagine the hundreds or thousands of nodes the compiler will handle for a given method, all spending at least part of their time juggling stack values around. It would be a miserable waste.

So the compiler constructs a poor-man's closure: an anonymous inner class. The inner class implements our "ClosureCallback" interface which has a single method "compile" accepting a single MethodCompiler parameter "context". This allows the non-local "next" bytecode emitter to first construct the NextJump, then ask the AST compiler to continue processing AST nodes. The compiler walks the "value" node for the "next" operation, again causing appropriate bytecode emitter calls to be made, and finally we have our value on the stack, exactly where we want it. We continue constructing the NextJump and happily toss it into the ether.

The final line of the compileNext method initiates this process.

So what would this look like with the closure specification in play? We'll simplify it with a function object.
public static void compileNext(Node node, MethodCompiler context) {
context.lineNumber(node.getPosition());

final NextNode nextNode = (NextNode) node;

ClosureCallback valueCallback = { MethodCompiler => context
if (nextNode.getValueNode() != null) {
ASTCompiler.compile(nextNode.getValueNode(), context);
} else {
context.loadNil();
}
};

context.pollThreadEvents();
context.issueNextEvent(valueCallback);
}

That's starting to look a little cleaner. Gone is the explicit "new"ing of a ClosureCallback anonymous class, along with the superfluous "compiler" method declaration. We're also seeing a bit of magic outside the function type: closure conversion. Our little closure that accepts a MethodCompiler parameter is being coerced into the appropriate interface type for the "valueCallback" variable.

How about a more complicated example? Here's a much longer method from JRuby that handles "operator assignment", or any code that looks like a += b:
public static void compileOpAsgn(Node node, MethodCompiler context) {
context.lineNumber(node.getPosition());

// FIXME: This is a little more complicated than it needs to be;
// do we see now why closures would be nice in Java?

final OpAsgnNode opAsgnNode = (OpAsgnNode) node;

final ClosureCallback receiverCallback = new ClosureCallback() {
public void compile(MethodCompiler context) {
ASTCompiler.compile(opAsgnNode.getReceiverNode(), context); // [recv]
context.duplicateCurrentValue(); // [recv, recv]
}
};

BranchCallback doneBranch = new BranchCallback() {
public void branch(MethodCompiler context) {
// get rid of extra receiver, leave the variable result present
context.swapValues();
context.consumeCurrentValue();
}
};

// Just evaluate the value and stuff it in an argument array
final ArrayCallback justEvalValue = new ArrayCallback() {
public void nextValue(MethodCompiler context, Object sourceArray,
int index) {
compile(((Node[]) sourceArray)[index], context);
}
};

BranchCallback assignBranch = new BranchCallback() {
public void branch(MethodCompiler context) {
// eliminate extra value, eval new one and assign
context.consumeCurrentValue();
context.createObjectArray(new Node[]{opAsgnNode.getValueNode()}, justEvalValue);
context.getInvocationCompiler().invokeAttrAssign(opAsgnNode.getVariableNameAsgn());
}
};

ClosureCallback receiver2Callback = new ClosureCallback() {
public void compile(MethodCompiler context) {
context.getInvocationCompiler().invokeDynamic(
opAsgnNode.getVariableName(), receiverCallback, null,
CallType.FUNCTIONAL, null, false);
}
};

if (opAsgnNode.getOperatorName() == "||") {
// if lhs is true, don't eval rhs and assign
receiver2Callback.compile(context);
context.duplicateCurrentValue();
context.performBooleanBranch(doneBranch, assignBranch);
} else if (opAsgnNode.getOperatorName() == "&&") {
// if lhs is true, eval rhs and assign
receiver2Callback.compile(context);
context.duplicateCurrentValue();
context.performBooleanBranch(assignBranch, doneBranch);
} else {
// eval new value, call operator on old value, and assign
ClosureCallback argsCallback = new ClosureCallback() {
public void compile(MethodCompiler context) {
context.createObjectArray(new Node[]{opAsgnNode.getValueNode()}, justEvalValue);
}
};
context.getInvocationCompiler().invokeDynamic(
opAsgnNode.getOperatorName(), receiver2Callback, argsCallback,
CallType.FUNCTIONAL, null, false);
context.createObjectArray(1);
context.getInvocationCompiler().invokeAttrAssign(opAsgnNode.getVariableNameAsgn());
}

context.pollThreadEvents();
}

Gods, what a monster. And notice my snarky comment at the top about how nice closures would be (it's really there in the source, see for yourself). This method obviously needs to be refactored, but there's a key goal here that isn't addressed easily by currently-available Java syntax: the caller and the callee must cooperate to produce the final result. And in this case that means numerous closures.

I will spare you the walkthrough on this, and I will also spare you the one or two other methods in the ASTCompiler class that are even worse. Instead, we'll jump to the endgame:
public static void compileOpAsgn(Node node, MethodCompiler context) {
context.lineNumber(node.getPosition());

// FIXME: This is a little more complicated than it needs to be;
// do we see now why closures would be nice in Java?

final OpAsgnNode opAsgnNode = (OpAsgnNode) node;

ClosureCallback receiverCallback = { MethodCompiler context =>
ASTCompiler.compile(opAsgnNode.getReceiverNode(), context); // [recv]
context.duplicateCurrentValue(); // [recv, recv]
};

BranchCallback doneBranch = { MethodCompiler context =>
// get rid of extra receiver, leave the variable result present
context.swapValues();
context.consumeCurrentValue();
};

// Just evaluate the value and stuff it in an argument array
ArrayCallback justEvalValue = { MethodCompiler context, Object sourceArray, int index =>
compile(((Node[]) sourceArray)[index], context);
};

BranchCallback assignBranch = { MethodCompiler context =>
// eliminate extra value, eval new one and assign
context.consumeCurrentValue();
context.createObjectArray(new Node[]{opAsgnNode.getValueNode()}, justEvalValue);
context.getInvocationCompiler().invokeAttrAssign(opAsgnNode.getVariableNameAsgn());
};

ClosureCallback receiver2Callback = { MethodCompiler context =>
context.getInvocationCompiler().invokeDynamic(
opAsgnNode.getVariableName(), receiverCallback, null,
CallType.FUNCTIONAL, null, false);
};

// eval new value, call operator on old value, and assign
ClosureCallback argsCallback = { MethodCompiler context =>
context.createObjectArray(new Node[]{opAsgnNode.getValueNode()}, justEvalValue);
};

if (opAsgnNode.getOperatorName() == "||") {
// if lhs is true, don't eval rhs and assign
receiver2Callback.compile(context);
context.duplicateCurrentValue();
context.performBooleanBranch(doneBranch, assignBranch);
} else if (opAsgnNode.getOperatorName() == "&&") {
// if lhs is true, eval rhs and assign
receiver2Callback.compile(context);
context.duplicateCurrentValue();
context.performBooleanBranch(assignBranch, doneBranch);
} else {
context.getInvocationCompiler().invokeDynamic(
opAsgnNode.getOperatorName(), receiver2Callback, argsCallback,
CallType.FUNCTIONAL, null, false);
context.createObjectArray(1);
context.getInvocationCompiler().invokeAttrAssign(
opAsgnNode.getVariableNameAsgn());
}

context.pollThreadEvents();
}

There's two things I'd like you to notice here. First, it's a bit shorter as a result of the literal function objects and closure conversion. It's also a bit DRYer, which naturally plays into code reduction. Second, there's far less noise to contend with. Rather than having a minimum of five verbose lines to define a one-line closure (for example), we now have three terse ones. We've managed to tighten the focus to the lines of code we're actually interested in: the bodies of the closures.

Of course this quick tour doesn't get into the much wider range of features that the closures proposal contains, such as non-local returns. It also doesn't show closures being invoked because with closure conversion many existing interfaces can be represented as function objects automatically.

I'll be looking at the closure proposal a bit more closely, and time permitting I'll try to get a simple JRuby prototype compiler wired up using the techniques above. I'd recommend you give it a try too, and offer Neal your feedback.