Hosting WebAssembly in Swift

Today we are going to embed and run WebAssembly (Wasm) modules in a Swift program. Using Wasmer, an embeddable runtime for Wasm, wrapped in a simple Swift API.

Note: 2023-09-23: This is probably outdated by now. API changes in Wasmer.

While it isn’t used that much in production just yet, you likely have heard about WebAssembly (Wasm) before. The technology is most commonly known for running programs written in compiled languages like C, Rust or Swift right within a web browser. Sandboxed, and without any native plugins.

That is not what we are going to do today. Instead of running Wasm programs inside of a web browser, we are going to run them inside of a Swift program.

For the impatient among us, this is what it looks like:

import Wasmer

let wasmData = try Data(contentsOf: URL(fileURLWithPath: "sum.wasm"))
let module   = try WebAssembly.Module(wasmData)
let instance = try WebAssembly.Instance(module)
print(instance.exports.sum(.i32(7), .i32(8)))

We are also not going to look at how to compile Swift itself to WebAssembly. Checkout the SwiftWasm project for that.

So what exactly is Wasm. We’ll look at that in more detail further down below, but essentially a developer can compile a, say Rust, program into a Wasm “binary”. In Rust it looks like this:

$ cargo build --target wasm32-wasi

The result is a .wasm file, for example sum.wasm - a Wasm binary. It doesn’t run natively on your host (it is built for a different platform), but it does run within a web browser, or: using embeddable runtimes like Wasmer.

The wasm32 in the target is like x86 or arm64 - the CPU architecture. The wasi is more like the operating system, what would be win32 or linux in other targets.

But let’s setup a first basic project to get a feel for the technology.

Installing Wasmer

Wasmer is pretty small and can be installed with a little script in like 30 seconds (manual install is fine too, check the docs for instructions):

$ curl https://get.wasmer.io -sSfL | sh

Afterwards you have the wasmer and wapm binaries (they install into ~/.wasmer by default). WAPM is a package manager like Homebrew and can be used to install and run Wasm packages:

$ wapm install -g fortune
[INFO] Installing _/fortune@0.2.0
$ wapm run fortune
The most exciting phrase to hear in science, the one that heralds new discoveries, is not "Eureka!" (I found it!) but "That's funny ..."
    -- Isaac Asimov

A more complex example, a JavaScript engine as a Wasm module: QuickJS:

$ wapm install -g quickjs
[INFO] Installing _/quickjs@0.0.3
Global package installed successfully!

$ wapm run qjs
QuickJS - Type "\h" for help
qjs > console.log("hello")
console.log("hello")
hello
undefined

Looks nice. Let’s compile a small program ourselves.

Compiling a Rust program

Arguably one of the languages which (as of today) support Wasm the best is Rust. We at the ARI think Rust is mostly wrong, but it always pays to watch over the fence. Since a few examples we are going to play with are in Rust, let’s install the toolchain.

Note: Do not install the rust Homebrew package! Instead we are going to use rustup, which can install both Rust compiler/env and the Wasm toolchain we need:

$ brew install rustup
$ rustup-init
$ source $HOME/.cargo/env
$ rustup target add wasm32-wasi

That’s all required to get going with Rust. We are going to compile the cowsay Rust program. Into a Wasm binary. First check out the repository:

$ git clone https://github.com/wapm-packages/cowsay
$ cd cowsay

Then compile it for Wasm:

$ cargo build --target wasm32-wasi --release
...
    Finished dev [unoptimized + debuginfo] target(s) in 35.68s

Just like with Swift Package Manager, this will pull down and compile all the dependencies, then the program itself. The result can be found in the target/wasm32-wasi/release folder:

$ du -sh target/wasm32-wasi/release/cowsay.wasm 
804K	target/wasm32-wasi/release/cowsay.wasm

We can run the module using wasmer:

$ wasmer target/wasm32-wasi/release/cowsay.wasm Swifty Cow!
 _____________
< Swifty Cow! >
 -------------
        \   ^__^
         \  (oo)\_______
            (__)\       )\/\
               ||----w |
                ||     ||

Excellent, we got Wasm cows!

The wasmer tool acts as the runtime for the compiled Wasm program, quite similar to how you invoke Java programs with java, Python programs with python and so on.
Actually it is a lot more similar to invoking a Docker container like docker run -it swift, but we’ll get to that later.

“Very nice” you say, but where is the promised Swift stuff? We ain’t here for the Rost!

SwiftyWasmer

Wasmer comes with quite a set of APIs to embed Wasmer into tools written in other programming languages. There is one for Go, one for C/C++, one for JavaScript and one for Rust. Bot none for Swift.

Thanks to Swift’s excellent C integration, we used that and produced: SwiftyWasmer.

To work, the Swift Package Manager requires a pkg-config file. Fortunately wasmer config can generate one for you:

$ wasmer config --pkg-config \
  > /usr/local/lib/pkgconfig/wasmer.pc

Unfortunately the generated file is a little b0rked in 1.0.0. Open up the file in your favorite editor:

$ emacs /usr/local/lib/pkgconfig/wasmer.pc

And adjust two little things:

  1. remove the /wasmer from the Cflags line, it should then read:
    Cflags: -I/Users/helge/.wasmer/include
  2. add -lffi to the Libs line, it should then read:
    Libs: -L/Users/helge/.wasmer/lib -lwasmer -lffi

To link statically, move libwasmer.dylib out of the way:

mv ~/.wasmer/lib/libwasmer.dylib \
   ~/.wasmer/lib/libwasmer.dylib-away

Let’s build something similar to the wasmer CLI tool above, but using Swift. The easiest way to get going it to use swift sh (brew install swift-sh), but feel free to setup an Xcode or SPM tool project:

#!/usr/bin/swift sh
import Wasmer // helje5/SwiftyWasmer

let path     = URL(fileURLWithPath: CommandLine.arguments[1])
let module   = try WebAssembly.Module(contentsOf: path)
let instance = try WebAssembly.Instance(module)

_ = try instance.exports._start()

You can put that into mytool.swift, run chmod +x mytool.swift and then run the tool itself:

$ echo "Hello Swift" | \
  ./mytool.swift target/wasm32-wasi/release/cowsay.wasm 
 _____________
< Hello Swift >
 -------------
        \   ^__^
         \  (oo)\_______
            (__)\       )\/\
               ||----w |
                ||     ||

Note: We cannot pass commandline arguments to cowsay for reasons, but cowsay reads from stdin as a fallback. Which is what the echo pipe does.

The code should be pretty self explanatory. We first build a URL for the file passed in argument[1]. We then create a WebAssembly.Module for that URL:

let module = try WebAssembly.Module(contentsOf: path)

A Module is essentially the compiled Wasm. It can’t be executed on its own, to do that, a WebAssembly.Instance needs to be setup:

let instance = try WebAssembly.Instance(module)

The Instance is the execution environment, the Sandbox. When the Instance is created, you provide the compiled Module and optionally a set of “imports” you want to make available to the module. After the Instance is created, the “exports” are available.
This is quite similar to how dynamic libraries work, they have a set of symbols they “import” and a set of symbols (functions, globals, classes, etc) they “export”.

The default entry point for “tool like” binaries is the _start function, again quite similar to the _main used in C/system executables. This is what we (and the wasmer tool) call to start the Wasm program:

_ = try instance.exports._start()

_start neither takes arguments nor returns values. An Error would be thrown if the module wouldn’t actually export the _start function. For example because it isn’t a commandline tool, but some other module, like a library or plugin.

Excellent, we can run tools compiled for Wasm right from within Swift! An advantage: those Wasm compilations work on all platforms, similar to how you can run a Java program on all platforms (write once, run anywhere). But in a little different way.

Building a Small Rust Lib and Call it from Swift

We are now going to dive a little deeper into what Wasm is and how it works. Let’s start by writing a tinsy Rust library which provides a function to add two numbers.

There is no need to know much about Rust here. What we need to do to setup a library project is similar to SwiftPM. The Rust package manager is called cargo:

$ cargo new --lib sum # create a new lib called `sum`
$ tree sum
sum
├── Cargo.toml
└── src
    └── lib.rs

Add this to the Cargo.toml, to tell Rust that we are creating a library with a “C” interface:

[lib]
crate-type = ["cdylib"]

Then replace the contents of the lib.rs file with:

#[no_mangle]
extern "C" fn sum(a: i32, b: i32) -> i32 {
  let s = a + b;
  println!("From WASM: Sum is: {:?}", s);
  s
}

The #[no_mangle] and extern "C" are similar to Swift’s @_cdecl. All the rest is really similar to Swift (almost like Rust stole all the best ideas from it 😉). We add two integer (32-bit) numbers, print it, and then return the result.

Like with cowsay before, our module can be compiled like this:

$ cargo build --target wasm32-wasi
   Compiling sum v0.1.0 (/tmp/sum)
    Finished dev [unoptimized + debuginfo] target(s) in 0.45s
$ du -sh target/wasm32-wasi/debug/sum.wasm 
1.7M	target/wasm32-wasi/debug/sum.wasm

Note that we didn’t define a _start function, this time we built a library with a single sum function.

Here is a small Swift tool which can load that module written in Rust and invoke the function:

#!/usr/bin/swift sh
import Wasmer // helje5/SwiftyWasmer

let path     = URL(fileURLWithPath: CommandLine.arguments[1])
let module   = try WebAssembly.Module(contentsOf: path)
let instance = try WebAssembly.Instance(module)

print(try instance.exports.sum(.i32(46), .i32(2)))

Calling it:

$ ./mytool.swift target/wasm32-wasi/debug/sum.wasm 
From WASM: Sum is: 48
[i32(48)]

Note how the Rust module prints the result, and our Swift side also prints the result(s).

An important thing: It doesn’t matter what language was used to produce the Wasm module. As long as it exports a sum function, we can call it using the very same from Swift.

Above we show running things using swift sh. Xcode can be used as well. Just create a new macOS “Tool” project and add this package https://github.com/helje5/SwiftyWasmer as a SwiftPM dependency (e.g. via “File / Swift Packages / Add Dependency”). When compiling in Xcode, make sure you compile for “Your Mac” is the target device (i.e. not an iOS device).

Functions which are no real Functions

Above we’ve seen that the sum call was invoked with two .i32 arguments and that it returns a single .i32 argument. It is a 32bit integer, obviously. Now the interesting part is that Wasm functions only allow for four different datatypes: .i32, .i64, .f32 and .f64. That is it! No strings, no structs, no arrays. Let alone methods or objects.

Disclaimer: we are no experts on this, please feel free to send corrections!

So let’s review what Wasm actually is, it is easy to get wrong, especially if you already have some smattering knowledge about Wasm. With all the talk about “functions” being imported and exported, you might think it is like embedding Python or Java using the JNI. Or maybe like COM or CORBA. That is not the case. It isn’t like JVM or CLR bytecode either.

The WebAssembly website says:

Wasm is a binary instruction format for a stack-based virtual machine.

Wasm itself is really nothing more than that. It specifies the binary machine code a Wasm “CPU” will run. And that machine code is kept very simple. Writing a Wasm program by hand is very much like writing an assembly program for the ARM or Intel CPUs. Besides the instructions to execute, the Wasm runtime also provides a linear memory block to the “machine”.

So what happens if you run a Wasm program is more similar to the process that happens when you run Intel binary code using Rosetta on an M1 Mac. Or some S/390 machine code on an Intel machine using QEmu.

How low level is it? Very, very low level. For example, let’s assume you want to pass a String from the host to the Wasm program. What you essentially have to do is:

  1. copy the string into the memory of the Wasm instance, as bytes, e.g. UTF-8
  2. call a function function with the position of the string in memory, and maybe its length

So consider Wasm like a computer. If you boot it up, it starts executing instructions for its CPU, the Wasm instructions.
Now (unless you are an embedded developer) you very rarely write programs that directly execute on a barebones computer. Instead you’d usually use an operating system, like Windows or Linux. This OS will provide userlevel programs much nicer abstractions for dealing with memory, handling I/O etc.

No different in Wasm “computers”. Currently there are two major Wasm “operating systems”: The older but very capable emscripten and the newer WASI (WebAssembly System Interface). Both provide user level library functionality to Wasm programs. For example in our sum example we used println! to print out a value. That calls into WASI to perform the actual printing (on the host).

This brings us back to “Functions which are no real Functions”. Considering the context, I found it more useful to think of the “Wasm functions” as system calls (i.e. calls going from userspace to kernelspace). If you call, say emscripten write, it’ll call a host provided function with something like _syscall(.i32(4), .i32(2727), .i32(12)). The 4 could be the file descriptor, the 2727 the position of the data in the memory and 12 its length.

Unlike the JVM or the CLR, Wasm has no concept of methods, dynamic dispatch, objects - or any such higher level concept. Both JVM and CLR do act as language bridges somewhat similar to COM (i.e. they allow integrating different languages compiled to their high level OO capable bytecode).

Wasmer is more like Docker

In summary Wasm is less like a scripting language runtime or language integration bridge, but way more like a Docker virtual machine. Or Virtualization.framework. Think in that direction when thinking about additional Wasm usecases. Yes, Wasm can be used to deal with compute intense tasks in web browsers by using compiled code, but it can also be used to host code in isolated environments on the server (or the client!).

Unlike Docker Wasm doesn’t need a Linux kernel to run. Or images for a specific instruction set. “Wasm images” can be run on any platform which have a runtime available. Yes, even in the browser if that is desirable.

The environment provided by WASI is also very much like the Docker environment. You can remap files, you can provide the files the sandbox can even access (by default none), etc.

As an example, this is how you run nginx in Wasmer (uses emscripten):

$ wapm run nginx -p example -c nginx.conf
2015/10/21 07:28:00 [notice] 73097#0: nginx/1.15.3
2015/10/21 07:28:00 [notice] 73097#0: built by clang 6.0.1  (emscripten 1.38.12 : 1.38.12)
2015/10/21 07:28:00 [notice] 73097#0: OS: Darwin

In the future you might able to run a Macro.swift server alongside your nginx frontend proxy, while connecting to some database written in Rust. Unlike with Docker, you don’t have to wrap each in a full Linux environment. We’d need something like compose for that, and SwiftyWasmer might be used to write such tooling 😉

We could also see that the tech might be used to offer very lightweight AWS Lambda like functions, without all the overhead required to boot up a Linux kernel. Yet still giving the user the choice what language to write the functions in.

Compiling Swift to Wasm

The original article didn’t talk about this, but we just gave it a try and it worked nicely: Compiling Swift itself to Wasm. And then running that Wasm Swift binary from within Swift ∞

To get going, one needs to download a Swift toolchain from the SwiftWasm project, for example: Swift Wasm 5.3.1. Install the package, and you’ll find the Swift Wasm toolchain in: /Library/Developer/Toolchains/.

Add it to your path when playing w/ SwiftWasm:

$ export PATH=/Library/Developer/Toolchains/swift-wasm-5.3.1-RELEASE.xctoolchain/usr/bin:$PATH

Let’s pull down a great Swift package, cows, and build it for Wasm:

$ git clone https://github.com/AlwaysRightInstitute/cows
$ cd cows
$ swift build --triple wasm32-unknown-wasi
[9/9] Linking vaca.wasm
$ du -sh .build/debug/vaca.wasm 
 25M	.build/debug/vaca.wasm

And yay, you can then run this in Wasmer:

$ wasmer .build/debug/vaca.wasm compiler
          (__)
        /  .\/.     ______
       |  /\_|     |      \
       |  |___     |       |
       |   ---@    |_______|
    *  |  |   ----   |    |
     \ |  |_____
      \|________|
CompuCow Discovers Bug in Compiler

Or in Swift (e.g. using the swasi-run tool included in SwiftyWasmer):

$ swift run swasi-run vaca.wasm 
        o
        | [---]
        |   |
        |   |                              |------========|
   /----|---|\                             | **** |=======|
  /___/___\___\                         o  | **** |=======|
  |            |                     ___|  |==============|
  |           |                ___  {(__)} |==============|
  \-----------/             [](   )={(oo)} |==============|
   \  \   /  /             /---===--{ \/ } |
-----------------         / | NASA  |====  |
|               |        *  ||------||-----^
-----------------           ||      |      |
  /    /  \   \             ^^      ^      |
 /     ----    \
  ^^         ^^           This cow jumped over the Moon

Wasm Swift running within a Swift host.

Closing Notes

All that technology, while in development for years, still seems very early. It is quite interesting and - if anything - a fun toy to play with!

It is definitely worth watching where this technology is going.

When some VC friend asked us what we think of the idea of running server side code using Wasm, we gave him like a set of reasons why this is utter nonsense. Except maybe for edgy edge cases. But we also pointed out that the JVM is big on the server, despite being invented for set-top boxes and phones.
So I guess we’ll see whether this is the next big thing after Docker 😉

Pro tip: To troll Wasm evangelists/fanboys, always use “WASM” (all uppercase) and “Web Assembly” (w/ a space) when referring to the technology. That’s always a winner.

What’s Missing in SwiftyWasmer

Quite a few things, an imcomplete list:

  • Import objects do not seem to fully work in the 1.0.0 C API yet, e.g. you can’t configure the WASI environment yet (commandline, env vars, file mappings).
  • The 1.0.0 C API also seems to have issues with executing different WASI versions, though we may be just holding it wrong
  • There is no neat way in SwiftyWasmer yet to export functions to Wasmer (the ABI is a little unfortunate the integrate other languages),
  • Many other things :-)

Contact

Feedback is warmly welcome: @helje5, me@helgehess.eu.

Want to support my work? Buy an app! You don’t have to use it! 😀

Written on January 10, 2021