While I’m the author of this blog post, 99% of the work was done by Nicolas B. Pierron.

So far, my role in this project has largely been to play the wise old advisor, nodding and smiling mischeviously whenever Nicolas started exploring new ideas, and emitting cryptic comments in Reverse Jedi Notation.

A few months ago, we published a short (and mysterious) blog post in which we mentioned HolyJIT, an early research project towards a novel approach to writing JITs.

In this blog post, I would like to detail a bit more the ideas behind HolyJIT.

Anatomy of a JIT compiler

If you’re not already familiar with JIT compilers, I suggest you first take a look at Wikipedia before proceeding, because the rest of this blog post won’t be very useful to you.

For performance reasons, real-world JIT compilers such as SpiderMonkey, v8 or HotSpot often contain several implementations of the language. For instance:

The pure interpreter can start running immediately, which is very useful for code that the program is only ever going to run once. The naive compiler will take some time to compile the code fed to it, but will produce faster code once it’s done, so it is useful for code that is used repeatedly. Finally, the optimizing compiler will take much more time to compile the code, but will produce much faster code, so it is useful for performance-critical code.

That’s three implementations of the same language (sometimes more) in the same compiler. In fact, the inline cache strategy often contains a fourth implementation of large subsets of the the same language.

This has two large drawbacks:

Can we do better?

Our objective

The objective with HolyJIT is to see if we can develop a toolbox to:

Some of the ideas in HolyJIT are closely related to multi-stage compilation, as demonstrated for instance in MetaOCaml or Template Haskell. By opposition to such languages, though, the target is explicitly JITs and aims to let JIT developers pick or build their own strategies. By opposition to MetaOCaml, which requires shipping a full OCaml compiler as part of the executable, HolyJIT will not require shipping a full Rust compiler.

I should mention, again, that HolyJIT is early stage research. Some of what will be discussed in this blog post is tested, some is just ideas.

For the rest of this post, I’ll use Host Language to describe the language in which the JIT is implemented (here, Rust) and Target Language to describe the language which we are going to JIT-compile (for instance, JavaScript or regular expressions).

A second look at implementations

Consider the pure interpreter. For the sake of simplicity, we’ll assume that this is an interpreter of bytecode.

It contains:

Actually, assuming that the interpreter is compiled, during compilation, the specification is translated to a Control Flow Graph, which is itself translated to executable format. This executable format may be LLVM IR, machine language, WebAssembly, etc.

Let’s summarize this:

The exact same scheme also describes both the naive native compiler and the optimizing native compiler. Furthermore, since all need to interact in the same executable, they typically share:

In most JITs, the various implementations of the same language within a single JIT differ:

Now, recall that these different specifications as source code are different not because they implement different Target Languages but because the specification contains both the semantics and the strategy used to produce the executable.

Let’s see what we can do about this.

From the interpreter to the compiler

Consider operation + in JavaScript. If we are building a stack machine, the specifications can be translated straightforwardly as the following opcode Op::Add and specifications-as-code:

opcode!(stack, Op::Add => {
    let lval = stack.pop()?;
    let rval = stack.pop()?;
    let lprim = lval.to_primitive()?;
    let rprim = rval.to_primitive()?;
    match (lprim.type_(), rprim.type_()) {
        // `+` may be used for string concatenation.
        (String, _) | (_, String) => {
            let lstring = lprim.to_string()?;
            let rstring = rprim.to_string()?;
            stack.push(JSString::concat(lstring, rstring)?)?
        // `+` may be used for numeric addition.
        _ => {
            let lnum = lprim.to_number()?;
            let rnum = lprim.to_number()?;
            stack.push(JSNumber::add(lnum, rnum)?)?

In the above code, we have used a macro opcode!. Rust macros are powerful enough to perform sophisticated rewriting on the AST and communicate with the rest of the toolchain.

We have also used ? to represent all exceptional cases, including both user-level exceptions (such as Error) and VM-level exceptions (such as out-of-memory). Rust makes it easy to not confuse them.

Once we have thus specified all opcodes, as well as the main loop and the representation of state, we have a fully working interpreter. When encountering Op::Add, this interpreter will execute exactly the code above.

Now, assuming a sufficiently flexible toolchain (such as the existing Rust toolchain), we can extract the CFG for each opcode and use it twice:

  1. convert the CFG to Executable;
  2. keep the CFG as Data.

Build-time tool introduced Conversion of specifications-as-source into CFG as Data.

The first CFG as Executable is exactly what we’re already using for our interpreter. With out-of-the-box Rust, this is the native executable produced by rustc.

Using this CFG as Data, we can now implement the following program, executed at runtime:

  1. Read a chain of bytecode.
  2. Convert chain of bytecode to a vector of CFG-as-Data.
  3. Convert vector of CFG-as-Data to Executable.

This, along with the main loop, is a naive native compiler. We have obtained this native compiler by introducing the following tools in our toolchain:

Runtime tool introduced Convert CFG-as-Data to Executable.

Runtime tool introduced Combine blocks of CFG-as-Data.

These tools are entirely independent from the target language.

In effect, we have extracted the compiler from the interpreter.

While this result is not new (that’s pretty much the point of Multi-Stage Compilation), the technique we have employed is, as far as we can tell, both novel and more flexible than alternative approaches.

For one thing, with a single specification as source, we have both the interpreter and the compiler. This is a good thing, since we need both of them to produce a high-performance JIT.

In particular, we can go further.

From the compiler to the tracing compiler

Caveat At this stage of the project, this section is mostly hypothetical.

Consider, again, the implementation of + in JavaScript. Let’s say that we wish to implement a tracing JIT compiler as an inline cache generator.

I assume it’s a well-known result, but experience indicates that we can turn an interpreter into a naive tracing compiler by performing a simple (and automated) rewrite of the CFG.

Let’s inform our toolchain that we want to do this:

opcode!(stack, Op::Add => {
    let lval = stack.pop()?;
    let rval = stack.pop()?;
        let lprim = lval.to_primitive()?;
        let rprim = rval.to_primitive()?;
        match (lprim.type_(), rprim.type_()) {
            // `+` may be used for string concatenation.
            (String, _) | (_, String) => {
                let lstring = lprim.to_string()?;
                let rstring = rprim.to_string()?;
                stack.push(JSString::concat(lstring, rstring)?)
            // `+` may be used for numeric addition.
            _ => {
                let lnum = lprim.to_number()?;
                let rnum = lprim.to_number()?;
                stack.push(JSNumber::add(lnum, rnum)?)

So far, HolyJIT was extracting a single CFG from the specification as source. Macro trace! instructs it to also extract a second CFG (we’ll call it a Tracing CFG), obtained by transforming the Interpreter CFG to emit a trace of CFG-as-data.

For this, we need three tools:

Build-time tool introduced Conversion of CFG into tracing CFG.

Build-time tool introduced Introducing the CFG-as-data for inline caching.

Runtime tool introduced Letting the CFG assemble CFG-as-data.

This new Tracing CFG, which was, once again, extracted from the specification as source, may now be stored as CFG-as-data, and executed much as the naive compiler.

In other words, with minimal annotations, we can extract a tracing compiler from the interpreter.


What about security?

It is still too early to be sure. The ability to combine CFGs dynamically and convert them to executable is basicaly the same level of power as what current JITs do, so we don’t expect to make a difference here.

We hope that the ability to mostly decouple the compilation strategy from the target language will make it easier to fix security issues, though, and to experiment with new security strategies.

What about performance?

Our initial experiments lead us to believe that we can end up with essentially the same generated code as existing JITs, with the bonus that we can more easily experiment with additional optimization strategies.

How many ways do you need to translate CFG to Executable?

There are several possible implementation strategies. One possible way is to generate a portable assembly format such as WebAssembly. We could therefore translate CFG to WebAssembly both at build-time (for the interpreter) and at runtime (for the native compilers).

Of course, we may decide to generate directly optimized Executables at build-time (for the interpreter), effectively reusing the existing Rust infrastructure, and a different target at runtime.

Isn’t the CFG just a new bytecode format?

Somewhat, yes. Most JITs use at least one intermediate format between the bytecode and the executable. The CFG here is such a format. So, in effect, some kind of very high-level bytecode.

Wait, what about actually optimizing the CFG?

Yes, the discussion above didn’t mention constant propagation, dead code elimination, constant subexpression elimination, …

Of course, they need to be part of the result. At this stage, there are several places at which they may be added, but the most reasonable seems to be while generating Executable from CFG. The main question is whether we always want to apply these optimizations or whether they should take place only in some strategies. The former would be more regular and easier to trust, while the latter would let Target Language developers further fine-tune their JIT strategies.

Do you actually need a bytecode for JavaScript?

Not necessarily. This seems to be the simplest way to do things, but it’s not written in stone.

How do you switch from the interpreter to the naive compiler to the optimizing compiler?

Not determined yet. If we wish our toolkit to be as generic as possible, decision might be left in the hands of the implementor of the target language.

Why Rust and not X?

Many other languages would work too. Rust combines the ability to define low-level data structures and algorithms with strong type-safety, no dependency on a specific garbage-collector, and a very flexible toolchain.

Plus, we like it.

How close are you to release?

In a galaxy far, far away.

Where can I find the code?

A proof of concept is available on github.