Efficient coroutines by rewriting bytecode

Mar 29, 2023

One of the projects I work on as part of CC: Tweaked is Cobalt, our Lua runtime. While originally a fork of LuaJ, it’s now very much its own thing, having undergone many large refactors over the years. I’ve written a little bit about Cobalt in the past (see here and here), so do check those posts if you’re interested!

One of Lua’s cooler features is its support for coroutines. Coroutines are a mechanism for writing concurrent programs1 - each coroutine has its own “thread” of execution, and control is passed between them with the coroutine.yield and coroutine.resume functions.

Example code using a coroutine. Feel free to skip if you’re already familiar!
local function counter(name)
  return function()
    for i = 1, 3 do
      print(name, i)
      coroutine.yield() -- Pass control back to the resumer

-- Create two coroutines
local cos = {
  coroutine.create(counter("counter 1")),
  coroutine.create(counter("counter 2")),

while #cos > 0 do
  -- Resume each coroutine
  for i = #cos, 1, -1 do
    if coroutine.status(cos[i]) == "dead" then table.remove(cos, i) end

This example code creates two counters, and then runs the two concurrently, printing the following output:

counter 2	1
counter 1	1
counter 2	2
counter 1	2
counter 2	3
counter 1	3

Implementing coroutines

While coroutines are an incredibly neat feature, they do cause problems when implementing a Lua VM. Unlike languages like C# or Javascript, where asynchronous functions must be explicitly defined with the async keyword, any Lua function is allowed to yield2. This means that almost all of the Lua VM must support yielding3, including:

  1. The Lua interpreter (duh!).
  2. Debug hooks (such as “on call” and “on instruction”).
  3. Any library function (a function written in Java/C rather than Lua) which calls back into Lua. This includes anything which calls a function directly (i.e. table.sort or string.gsub), but also anything which might invoke a metamethod. This includes anything which indexes into a table, so is surprisingly common!

While most of this blog post will be spent on talking about how we implement yielding in library functions, let’s talk about how to implement coroutines in the first place! The JVM doesn’t have built-in support for coroutines4, so we’re going to build everything from scratch. There’s several ways we might go about this:

  1. Create a native thread for each coroutine, with locks to ensure only one thread runs at once.

    This is what LuaJ does. While it’s simple to implement, it can be very resource intensive. Coroutines are expected to be cheap, so programs may chose to run dozens or even hundreds at once. On larger ComputerCraft servers such as SwitchCraft, we saw several thousand threads being used just for coroutines.

  2. Use Continuation Passing Style (CPS).

    If you’re familiar with effect handlers, then I’m sure you’ve seen an implementation of coroutines using shallow handlers. CPS is a common implementation strategy for effects, and so this is a fairly obvious choice.

    For readers who don’t get excited by programming language research, CPS is just a fancy word for callbacks. While NodeJS is the infamous example of using callbacks to implement asynchronous control flow, many other async models such as JS’s Promise and .NET’s Task5 are still heavily callback-based under the hood.

    Unfortunately, the Java runtime is not very amenable callback-heavy code. Each callback must be allocated on the heap, and the lack of tail calls necessitates the use of trampolines.

  1. Throw an exception.

    When a coroutine yields, throw an exception, unwinding the current stack. In this case, any information on the stack is discarded, and so state should be stored on an external object as the function executes.

    This is (mostly)6 what Lua does. The state of Lua’s functions is already stored on the Lua stack, so discarding the native stack doesn’t really matter. However, this doesn’t work very well for C functions which call back into Lua, limiting the places one can yield.

Cobalt uses a slightly odd mixture of 1 and 3. If every function on the call stack supports being resumed (this includes every Lua function, and some library functions), then we throw an exception to unwind the stack. If this is not the case, then we pause the current thread and switch execution over to a new one.

While it is a little cursed, it has also served us very well for the past four years. In some ways, it allows us to get the best of both worlds. Most coroutines won’t have any Java functions on the stack, and so can yield by unwinding. However, if we do have a Java function on the stack, we’re still able to yield as one might expect.

But yeah, it’s also super cursed, so let’s try something different.

State machines the boring way

The main reason we need all this extra threading machinery is that most library functions don’t support being unwound and resumed. If we want to remove the need for extra threads, we’ll need to find a way to make library functions resumable.

Let’s take a simple function, table.unpack. This takes a table as an argument (i.e. { 1, 2, 3 }), and “unpacks” them, returning them as multiple values. This is implemented in Java a little bit like this:

public static Varargs unpack(LuaState state, Varargs args) {
  LuaTable table = args.arg(1).checkTable();

  int length = Operations.length(state, table);
  LuaValue[] result = new LuaValue[length];
  for (int i = 0; i < length; i++) {
    LuaValue item = Operations.get(state, table, i + 1);
    result[i] = value;

  return varargsOf(result);

There’s two places where this function may yield:

To work out how we’re going to make this resumable, let’s start by sketching out the control flow graph (CFG) of the function. Slightly curiously, we’re also going to create special nodes in the graph for each place the function may yield.

G init LuaTable table = ...; len int length = Operations.length(...); init->len arr LuaValue[] result = new LuaValue[length]; len->arr loop for (...) { arr->loop 2 LuaValue item = Operations.get(...); loop->2 Loop condition is true 5 return varargsOf(result); loop->5 Loop condition is false 3 result[i] = item; i++; 2->3 3->loop Next loop iteration

Now that we’ve drawn out our graph of what the function is doing, we’re going to convert it into a state machine. Instead of writing our code “normally”, each block in the graph becomes a separate state in the state machine. This is then implemented in Java by a big ol’ loop and switch statement:

LuaTable table = ...;
int state = 0;
int length, i; LuaValue[] result; LuaValue item;
while (true) {
  switch(state) {
    case 0:
      length = Operations.length(table);
    case 1:
      result = new LuaValue[length];
    case 2: // for(...) {
      if (i >= length) return varargsOf(result);
    case 3:
      item = Operations.get(state, table, i + 1);
    case 4:
      result[i] = item;
      state = 2; // Jump back to state #2, our for loop.
      throw IllegalStateException(); // Literally!

Now, this isn’t very helpful on its own: it’s got the same functionality as the original code, it’s just harder to read! However, it offers one key benefit: we can now easily jump to any bit of the function, just by setting state to a different value.

Let’s see how we can use that to implement a resumable function. Firstly, instead of setting state to 0, we’ll read it from an object:

// ...
int state = resume.state;
int length = resume.length, i = resume.i;
LuaValue[] result = resume.result;
LuaValue item = resume.item;
// ...

Secondly, whenever one of our blocks yields, we’ll save everything back to that object:

switch {
  // ...
  case 3:
    try {
      item = Operations.get(state, table, i + 1);
    } catch (UnwindThrowable e) {
      resume.state = 3;
      resume.length = length;
      resume.i = i;
      resume.result = result;
      throw e;
  // ...

Hopefully you can now see where we’re going here. The first time we execute our function, we’ll create our resume object, save it on the Lua stack, and then pass that off to the main body. When the function resumes, we read the same object from the stack, and call the main body again, which will continue execution where it left off!

This is pretty amazing: we’ve now got a library call which can be unwound and resumed, meaning we can yield without having to create a new thread. We’ve won!, but at what cost? The resulting code is pretty ugly7, and while writing it is fairly mechanical, it’s also very error prone.

Hang on a moment. Mechanical and easy to make mistakes? That sounds like a job which should be automated.

State machines the Fun™ way

When you compile a Java file, it is not converted directly to machine code. Instead, it is converted to Java bytecode, a special format understood by the JVM. One of the nice properties of Java bytecode is that it is incredibly easy to work with8 (that, in part, is why Minecraft modding took off!), meaning it’s quite simple to either generate bytecode, or even rewrite existing code.

You can probably see where the rest of this post is going now. Instead of writing the state machine ourselves, we’ll just write our unwindable functions normally. Then, after the code is compiled, we just automatically apply the process we’ve already discussed; splitting a function into basic blocks and converting it into a state machine9.

Well, mostly. The nice thing about doing this on the bytecode level is that we don’t need that while/switch loop any more, and can can get away with good old labels and goto. This means our original unpack definition is (automatically!) converted to something like this:

public static Varargs unpack(LuaState state, Varargs args, UnpackState reusme) {
  if (resume != null) {
    switch(resume.state) {
      case 0: table, length = ...; goto state_0;
      case 1: table, result, length, i, item = ...; goto state_1;
      default: throw new IllegalStateException();

  LuaTable table = args.arg(1).checkTable();

  int length = try { Operations.length(state, table); } catch (UnwindThrowable e) { ... }
  LuaValue[] result = new LuaValue[length];
  for (int i = 0; i < length; i++) {
    LuaValue item = try { Operations.get(state, table, i + 1); } catch(UnwindThrowable e) { ... }
    result[i] = value;

  return varargsOf(result);

Ugly? Very! But at least we don’t have to write it by hand any more :).

Bonus: load-bearing state machines

One thing to note here is that for the past four years, all of Cobalt has actually been using the “boring” state machine method, with the exception of one function: load.

This function loads a block of Lua code, but instead of taking a string, it takes an arbitrary “reader” function:

local fn, err = load(function()
  return file:read("*l") -- Read a line from a file at a time.

Of course, this callback function may yield. And suddenly we’ve opened a whole can of worms, as now the entire Lua parser and compiler needs to support being suspended and resumed at arbitrary points. Using a hand-written state machine clearly isn’t practical!

Before I decided to write my own bytecode transform, I did look at other options. Nothing I’ve done here is especially novel: there’s several other tools out there for applying similar transforms. The most popular of these is Kotlin, a new programming language for the JVM with built-in support for coroutines. It’s fairly easy to convert our Lua compiler to use Kotlin, but the results are disappointing.

The real nastiness with load is that the yielding nature of the code permeates everywhere, even into the inner lexer loop. But despite that, most of the time none of these functions will yield! Most existing transforms understandably aren’t designed for this use case, and so do not perform very well. The Kotlin version of our parser10 is twice as slow as the original, and performs 3 times as many allocations.

Cobalt’s bytecode transform fortunately does not have this issue. Any unwind/resume code stays out of the hot-path, and allocations are only performed when a yield occurs, meaning that the transformed parser performs within a few percentage points of the original.

I think I’d call that a success :).

  1. Here I’m meaning concurrency to have the ability to have overlapped exeuction of multiple tasks. This is different to parallelism, where those tasks are executed simultaneously (i.e. on multiple cores).↩︎

  2. It’s probably worth mentioning the classic What Color is your Function? blog post. I personally don’t agree with its conclusions - after all, function colours are just effect types right?! - but it is an interesting read!↩︎

  3. Well it doesn’t have to. PUC Lua doesn’t allow yielding from within many library functions, throwing the somewhat infamous “attempt to yield across a C-call boundary” error.

    However, one of the requirements of Cobalt is that we should be able to pause (and then resume) the execution of the Lua runtime at any point. But there’s little difference between pausing the runtime and pausing just one coroutine, and so it makes sense to support both.↩︎

  4. Well, yet. Project Loom (which adds support for green threads) is on the horizon. However, eventually we’d be like to persist and restore the current Lua VM, and so we’re better off rolling our own mechanism here anyway.↩︎

  5. .NET is a little more complex then that. I can heartily recommend this blog post, which goes into the history behind their current implementation, and some neat tricks they use to avoid allocations. Unfortunately for us, Java doesn’t support value types, so there’s very little we can use there :(.↩︎

  6. The public API is implemented in terms of callbacks, see the Lua manual.↩︎

  7. For the sake of transparency, the actual implementation of table.unpack is quite different to the one displayed here. There we prefer duplicating code (which is bad for other reasons) and encode the current state implicitly via what fields are set.

    We do use a more normal state machine elsewhere (see our implementation of table.sort), but those are much more complex functions.↩︎

  8. Aside from a few gnarly bits, the Java bytecode specification is pretty simple. There’s also a lot of very good libraries for working with bytecode (notably OW2’s ASM).↩︎

  9. If you’re interested, this is the commit which implements this. The actual bytecode transformation sits at about 1.3k LOC, which is pretty good going.↩︎

  10. Remember, every function is marked suspend. Normal Kotlin code should have the same performance as Java.↩︎