Contents

• Lighthouse Benchmarks August 2021
• Demos
• Installation
• The JavaScript/WebAssembly interface
• C as an intermediate language

Other documentation

• Elm data structures
• String encoding
• Garbage Collector

Lighthouse Benchmarks August 2021

Figures noted are ranges across 4 runs of Lighthouse Performance report for Desktop.

Wasm performance is already comparable to JS with --optimize and minification. It is twice as fast as JS without the combination of --optimize and minification.

Until now the focus has been on making the generated code correct, without much serious effort to make it fast. VirtualDom diffing has not been implemented in Wasm yet, which I expect to be a big win Replacing Emscripten with custom code should hugely reduce code size, improving performance

The test application is based on elm-spa-example. The code has been modified so that API endpoints point at local .json files instead of a server, in order to remove most of the network variability from the measurements.

Wasm with -O3

Identical report for all 4 runs

Wasm with -O3 & uglifyjs

JS without --optimize

JS with --optimize

JS with --optimize and uglifyjs

Demos

I have built a few demos here: https://brian-carroll.github.io/elm_c_wasm/

• [Simple example app][demo-app]

  • App compiled from Elm to WebAssembly. Demonstrates the full Elm architecture interacting with kernel code written in C, and with the Elm runtime.

• [Unit tests for core libraries][demo-unit-tests-core]

  • You can see byte-level Elm data structures, arithmetic operations, record updates and accessors, and first-class functions in action.

• [Unit tests for the GC][demo-unit-tests-gc]

  • You can see print-outs of the heap, with byte-level Elm data structures, GC mark bits, etc.

• [Performance micro-benchmarks][demo-benchmark]

  • Reported figures for the Wasm implementations are slower than JS. It's partly due to crossing back and forth across the JS/Wasm boundary on every iteration. Browsers can't optimise that very well yet, though Firefox seems much better at it than Chrome.

  • Despite the limitations, the benchmark is a good way to stress-test the implementation, flush out bugs in the GC, and to find out where the bottlenecks are.

Installation

This project is very much in the early phases of development. It's not even close to being production ready.

The C compilation stage often fails because there are so many kernel modules still not written yet. And the compiler has not really been debugged so it generates bad C code sometimes.

You have been warned!

This is an installation guide for anyone who might want to play with the code and maybe contribute.

• Install Emscripten

  • https://emscripten.org/docs/getting_started/downloads.html

• Clone and build the forked Elm compiler

  • Install Stack (https://www.haskellstack.org)
  • Clone the forked repo (https://github.com/brian-carroll/elm-compiler)
  • cd to the root directory
  • stack init
  • stack build

• Clone this repo

  • In a different directory to the compiler. Perhaps put the two directories beside each other if you like, it doesn't matter.

• Run npm install

  • We need a couple of NPM packages.
    • TypeScript for the wrapper (src/kernel/wrapper/wrapper.ts)
      • This file could eventually be replaced with its compiled JS version
    • http-server to serve assets including the .wasm file
      • If you prefer to use something else, go ahead

• Ensure you're using the forked Elm binary

  • The Makefile invokes elm
  • You can either change the Makefile or change your PATH, whichever you prefer

• Build the demo

  • cd demos/wrapper
  • make

• Run a development server and open the demo

  • npx http-server (still inside demos/wrapper)
  • Open a browser at http://localhost:8080

The JavaScript/WebAssembly interface

Currently Wasm does not have direct support for any Web APIs such as DOM, XmlHttpRequest, etc. You have to call out from WebAssembly to JavaScript to use them. That means that effectful modules like VirtualDom, Browser, and Http, cannot be fully ported to WebAssembly yet. A few effects could be ported, but not enough to justify the effort of porting lots of complex kernel code to C (like Scheduler.js and Platform.js), so I think it's better to leave that for later.

In the meantime, I think a good goal is to compile the pure part of the Elm program to WebAssembly and leave the effectful Kernel code in JavaScript. There would be a "wrapper" Program type that acts as an interface between the Elm runtime and the WebAssembly module.

There are two ways for JavaScript and WebAssembly to talk to each other

• Function calls: The WebAssembly module can import JavaScript functions and call them, and export some of its own functions to make them callable from JavaScript. The main limitation is that the arguments can only be numbers, not data structures.
• Data: For data structures, you need to read and write to the WebAssembly module's memory, which appears in JavaScript as an ArrayBuffer. JavaScript can write bytes into the memory and then call an exported function when the data is ready (perhaps passing the offset and size of the data written). More details here.

The wrapper Program would convert the msg and model to a byte-level representation, write them to the WebAssembly module's memory, and then call an update function exposed by the compiled WebAssembly module.

The update will return a new model and Cmd, which will be decoded from bytes to the JavaScript structures that the Elm runtime kernel code uses.

VirtualDom is a unique effect in Elm in that it has no Cmd. I think the best approach will be to implement the diff algorithm in pure Elm and use ports to apply the patches, as Jakub Hampl demonstrated in Elm 0.18 a couple of years back.

C as an intermediate language

Why use an intermediate language?

It would be possible to compile Elm code directly to WebAssembly. In fact that's what I did in my first attempt. But there are several drawbacks.

Debugging the output of the compiler is not fun if it's WebAssembly. It's a lot easier to spot bugs in generated C code than in generated WebAssembly.

Writing Kernel code in WebAssembly is even worse. I even tried building a little DSL to generate Wasm kernel code from Haskell, but it was practically impossible to debug. I gave up on it. In theory you could compile Elm to WebAssembly but have kernel code in C, but this makes everything very hard.

Having both the compiled code and kernel in C makes everything vastly more do-able. I don't think this project would actually go anywhere if it was direct-to-Wasm.

As for performance, you might expect things to be faster if you skip the C "layer" but I think at this early stage it's more likely the other way around!

C compilers have had a lot of work put into them by hundreds of experts for over 50 years. They have a lot of advanced low-level optimisations built-in. I would have to figure out how to manually do a lot of the basic optimisations before even getting started on Elm-specific ones. And I'm just one person doing a hobby project to punch in some time in the evenings.

In summary I think that using an intermediate language is a massive productivity gain, and that any downsides are more theoretical than practical.

Why C and not Rust?

Rust has great WebAssembly support and I tried to use it early on. But it didn't work out well for me.

By the time the Elm compiler actually gets to the code generation phase, the program has already been validated by Elm's type checker. So I know it's OK but now I have to convince Rust that it's OK, and I found this was very difficult.

Rust is very good for hand-written code that does manual memory management, but not so much for auto-generated code running with a garbage collector. I found it was throwing errors about a lot of things that were actually good ideas in that context.

Also, a language implementation, particularly one with a GC, is going to involve a lot of unsafe Rust. That's quite advanced Rust programming. It would have been too hard for me to know when to go against the normal rules and when not to, on my very first Rust project. Someone who was proficient in Rust may have made a different decision.

For all these reasons, I very reluctantly realised that I had exhausted the alternatives and C was the only tool for this job. It's my first C project in about 10 years.