watever - WebAssembly text bundler

watever is a bundler/transpiler for WebAssembly + JavaScript. We make it easy to write modules in raw WebAssembly text format (WAT) that are consumable across the JavaScript ecosystem.

npm i watever

The watever CLI bundles your WAT to Wasm bytecode and generates a light-weight JS wrapper:

npx watever my-wat-module.wat # creates my-wat-module.wat.js

We aim to solve all the pain points with raw WAT development:

• Statically link multiple WAT modules
• Treeshake the result to produce lean bytecode
• Declare JS imports directly from WAT (instead of in separate JS loader code)
• Pass complex values between WAT and JS, like strings, TypedArrays, objects, etc
• Consume async JS functions from WAT

The goal is to create fully-featured WAT modules without writing any JS glue code. WAT files should stand on their own. After transpilation by watever, they act like nodes in the JS module graph that produce ESM exports & consume ESM imports.

Getting started

Let's say we have a simple WAT module, which exports a function that logs a number:

;; log-number.wat
(module
  (import "js" "console.log" (func $log (param f64)))

  (func $log_number (export "logNumber")
    (call $log (f64.const 3.14159265))
  )
)

The syntax is spec-compliant WAT plus conventions about how import statements are interpreted. In this example, the (import "js" ...) convention allows you to import arbitrary JavaScript objects with inline code.

☝️ If you're not familiar with WAT syntax, we recommend going through this MDN explainer before you continue.

Let's throw the file at watever:

npx watever log-number.wat

This creates a JS file log-number.wat.js which exposes our exported logNumber() as a named JS export, so we can use it like a normal JS module:

import { logNumber } from "./log-number.wat.js";
await logNumber();
// "3.14159265"

The only gotcha here is that our function has become async. That's because behind the scenes we have to unwrap the Promise returned by WebAssembly.instantiate. (We might change this behaviour when top-level await is supported widely enough.)

🎮 Try it yourself: You can find these code snippets under /examples. Clone the repo, transpile with npx watever examples/*.wat and run them like node examples/log-number.js. To run them with deno, use the --deno flag when transpiling: npx watever examples/*.wat --deno.

Hello world

Let's take the last example one step further and make it an actual "hello world" program. To do that, we need to pass a string to console.log, which is not something that plain WebAssembly allows. But watever makes it quite easy!

First, here's how we will use it from JavaScript:

import { helloWorld } from "./hello-world.wat.js";
await helloWorld();
// "Hello, world!"

And here's the code to make that happen:

;; hello-world.wat
(module
  (import "js" "console.log#lift" (func $log (param i32)))
  (import "watever/glue.wat" "lift_raw_string" (func $lift_raw_string (param i32 i32) (result i32)))

  (data (i32.const 0) "Hello, world!") ;; string of length 13

  (func $hello_world (export "helloWorld")
    (call $lift_raw_string
      (i32.const 0) ;; pointer to the string in memory
      (i32.const 13) ;; length
    )
    call $log
  )
)

There is a bit going on here that requires explanation. First, we added an import which points to "watever/glue.wat" and imports a function called "lift_raw_string". This function is used inside $hello_world to put an i32 value on the stack which will be interpreted as "Hello, world!" by the $log function. The import "watever/glue.wat" gets resolved like you're used to: to a file inside node_modules.

There's nothing special about the "watever" library in this regard, you could import from any npm library with the same syntax. Most importantly, you can also import from relative paths to your own WAT files, like (import "./my-helper.wat" ...). All imported WAT is analyzed by watever and the parts that are used are added to the Wasm output. Unused code is eliminated. This is why we call watever a bundler! It enables multi-file WAT development and reusable, shareable code, like any proper programming language.

There is another change we made in the code above, to enable passing a string to JavaScript. We added the instruction #lift at the end of the first import: "console.log#lift". This instruction tells watever to wrap console.log so that it understands the i32 returned by $lift_raw_string and turns it into a JS string. Lifting is our terminology for taking one or more low-level Wasm types (numbers) that somehow describe a high-level JS type, and transforming them into the actual JS type. In the case of $lift_raw_string, a string is described by 1) its starting position in memory and 2) its byte length.

Currently watever supports lifting for integers, floats, booleans, strings, raw bytes arrays (Uint8Array), JS arrays, JS objects, functions and opaque external references.

Passing values, allocating memory

Let's learn some more possibilities. We make another slight change and give our function the following signature:

import { hello } from "./hello-name.wat.js";
let greeting = await hello("Gregor");
console.log(greeting);
// "Hello, Gregor!"

This time, WAT's responsibility is not to log anything but to transform a string. We'll have to create the new string in Wasm memory rather than point to an existing location. The code below shows how to do this.

🤔 This might also give you an idea about whether writing raw WAT sounds fun to you. It's low-level stuff!

;; hello-name.wat
(module
  (import "watever/glue.wat" "lift_string" (func $lift_string (param i32) (result i32)))
  (import "watever/memory.wat" "get_length" (func $get_length (param i32) (result i32)))
  (import "watever/memory.wat" "alloc" (func $alloc (param i32) (result i32)))

  (data (i32.const 0) "Hello, !") ;; all the 8 chars we'll need

  (func $hello (export "hello#lift")
    (param $name i32) (result i32)

    (local $name_length i32)
    (local $greeting_length i32)
    (local $greeting i32)
    (local $i i32)

    ;; get length of input string $name, add 8 to it to calculate length of output string
    (call $get_length (local.get $name))
    local.tee $name_length
    (i32.const 8) (i32.add)
    local.set $greeting_length

    ;; allocate new string of the given length, save pointer
    (call $alloc (local.get $greeting_length))
    local.set $greeting

    ;; write "Hello, " into the new location, byte by byte
    (i32.store8 offset=0 (local.get $greeting) (i32.load8_u offset=0 (i32.const 0))) ;; "H"
    (i32.store8 offset=1 (local.get $greeting) (i32.load8_u offset=1 (i32.const 0))) ;; "e"
    (i32.store8 offset=2 (local.get $greeting) (i32.load8_u offset=2 (i32.const 0))) ;; "l"
    (i32.store8 offset=3 (local.get $greeting) (i32.load8_u offset=3 (i32.const 0))) ;; "l"
    (i32.store8 offset=4 (local.get $greeting) (i32.load8_u offset=4 (i32.const 0))) ;; "o"
    (i32.store8 offset=5 (local.get $greeting) (i32.load8_u offset=5 (i32.const 0))) ;; ","
    (i32.store8 offset=6 (local.get $greeting) (i32.load8_u offset=6 (i32.const 0))) ;; " "

    ;; write $name into the new location, by looping over its bytes
    (local.set $i (i32.const 0))
    loop
      (i32.store8 offset=7
        (i32.add (local.get $greeting) (local.get $i))
        (i32.load8_u
          (i32.add (local.get $name) (local.get $i))
        )
      )
      (br_if 0 (i32.ne (local.get $name_length)
        (local.tee $i (i32.add (local.get $i) (i32.const 1)))
      ))
    end

    ;; write the final "!"
    (i32.store8 offset=7
      (i32.add (local.get $greeting) (local.get $i))
      (i32.load8_u offset=7 (i32.const 0))
    )

    ;; return the new string
    (call $lift_string (local.get $greeting))
  )
)

We'll focus only on the important parts here. First, look at the function signature:

(func $hello (export "hello#lift")
  (param $name i32) (result i32)
  ;; ...

See what happens here?

• The export statement in the function header now reads (export "hello#lift"). This tells watever that the (result i32) returned from this function shall be lifted to a JS value. Lifting is always about going WAT -> JS, so #lift applies to function arguments in an import statement, and to return values in an export statement.

• The string that's passed into hello("Gregor") in JS ends up on the WAT side as a (param $name i32). This is a pointer to the memory address containing the string's bytes (UTF8-encoded by TextEncoder). The number of bytes allocated to a pointer can be obtained with the $get_length helper – this is how we get the length of the input string.

The procedure of transforming a JS string to a WAT number is called lowering. Lowering is about going JS -> WAT, so it happens either when you pass arguments to a WAT function from JS, or return something from a JS function called by WAT.

💡 For lowering, we didn't need an annotation like #lift: it happens automatically – just pass normal JS values to your WAT functions. This is possible because lowering leaves JS numbers and undefined untouched, and thus preserves the behaviour of vanilla (unwrapped) Wasm modules. Apart from that, lowering transforms string, TypedArray and ArrayBuffer to pointers into Wasm memory, and passes all other JS objects as an opaque i32 that represents an external reference (more on that later).

Another notable part is how we create the new string in Wasm memory:

  (call $alloc (local.get $greeting_length)))
  local.set $greeting

To create a new pointer, we call the $alloc function imported from watever/memory.wat. It takes a length (number of bytes), allocates memory of that length and returns a pointer.

The bulk of the function consists of copying characters into the memory addresses starting at $greeting. At the end, we return this string by lifting it:

  (call $lift_string (local.get $greeting))

This $lift_string function is a bit simpler than $lift_raw_string in the last example, because this time, $greeting is a pointer that has its length encoded, so we don't have to manually supply the length.

Some notes about memory:

• You may have wondered whether allocations done by $alloc can interfere with (data ...) sections. The answer is no, because watever/memory.wat knows where the last data section ends, and only allocates at offsets higher than that.

• $alloc is also directly used by our JS wrapper when it copies lowered values to Wasm memory (like $name above). All pointers created by $alloc will be automatically deallocated at the end of each function call into WAT. This means that most of the time, you don't have to think about freeing memory. (In case you do need to keep memory around, there is an escape hatch – see the next example.)

• $alloc handles dynamically growing the memory for you.

To enable these conveniences, the module watever/memory.wat (which declares and exports a Wasm memory) is imported implicitly in every watever bundle, so that it can be exported and used by our JS wrappers. Because of that, as you may have noticed, none of our previous code examples had to declare memory. The downside is that currently, watever is not compatible with Wasm modules that declare their own memory, e.g. to implement custom garbage collection schemes. We expect that this can be resolved when multiple memories become available.

Importing JS and handling async

Finally, we will add one more feature to our WAT module and interact with an async JS API. The module will take a github username and respond with the following message, which contains the user's number of published github repos:

import { hello } from "./hello-github-user.wat.js";
let greeting = await hello("mitschabaude");
console.log(greeting);
// "Hello, mitschabaude! I see you have 19 github repositories."

To achieve this, we're going to import the following JS module, which can fetch the number of repos from Github's REST API:

// github.js
export async function numberOfRepos(username) {
  let res = await fetch(`https://api.github.com/users/${username}/repos`);
  let json = await res.json();
  return `${json.length}`;
}

Here' the WAT in all its glory:

;; hello-github-user.wat
(module
  (import "./github.js" "numberOfRepos#lift" (func $number_of_repos (param i32) (result i32)))

  (import "watever/glue.wat" "lift_string" (func $lift_string (param i32) (result i32)))
  (import "watever/glue.wat" "lift_extern" (func $lift_extern (param i32) (result i32)))
  (import "watever/memory.wat" "get_length" (func $get_length (param i32) (result i32)))
  (import "watever/memory.wat" "alloc" (func $alloc (param i32) (result i32)))
  (import "watever/memory.wat" "keep" (func $keep (param i32)))
  (import "watever/memory.wat" "free" (func $free (param i32)))
  (import "watever/promise.wat" "then" (func $then (param i32 i32) (result i32)))

  (data (i32.const 0) "Hello, ! I see you have  github repositories.")

  (table 1 funcref)
  (export "table" (table 0))
  (elem (i32.const 0) $create_greeting)

  (global $username (mut i32) (i32.const 0))

  (func $hello_github_user (export "hello#lift")
    (param $username i32) (result i32)
    (local $promise i32)

    ;; store the $username pointer as a global and use $keep to keep it in memory
    (global.set $username (local.get $username))
    (call $keep (local.get $username))

    ;; call the async JS function, with the $username as argument
    (call $number_of_repos (call $lift_string (local.get $username)))
    local.set $promise

    ;; chain the function at table index 0 after the $promise
    (call $then
      (local.get $promise)
      (i32.const 0)
    )
    ;; this produces a new promise which we take from the stack...
    local.set $promise

    ;; ...and return (and tell JS to read it as an external reference)
    (call $lift_extern (local.get $promise))
  )

  (func $create_greeting
    (param $reponumber i32) (result i32)
    (local $greeting i32)

    ;; put length of new string on stack
    (call $get_length (global.get $username))
    (call $get_length (local.get $reponumber))
    i32.const 45
    (i32.add) (i32.add)

    ;; allocate new string $greeting
    call $alloc
    local.set $greeting

    ;; copy string fragments into $greeting
    local.get $greeting
    i32.const 0 i32.const 7
    call $copy_string ;; "Hello, "
    global.get $username (call $get_length (global.get $username))
    call $copy_string ;; ${username}
    i32.const 7 i32.const 17
    call $copy_string ;; "! I see you have "
    local.get $reponumber (call $get_length (local.get $reponumber))
    call $copy_string ;; ${repos}
    i32.const 24 i32.const 21
    call $copy_string ;; " github repositories."
    drop

    ;; free the string kept in memory before
    (call $free (global.get $username))

    ;; return the new, combined string
    (call $lift_string (local.get $greeting))
  )

  (func $copy_string
    (param $target i32) (param $source i32) (param $length i32)
    (result i32)
    (local $i i32)
    (local.set $i (i32.const 0))
    loop
      (i32.store8
        (i32.add (local.get $target) (local.get $i))
        (i32.load8_u (i32.add (local.get $source) (local.get $i)))
      )
      (br_if 0 (i32.ne (local.get $length)
        (local.tee $i (i32.add (local.get $i) (i32.const 1)))
      ))
    end
    (i32.add (local.get $target) (local.get $i))
  )
)

This may be too much to digest, so let's focus on the interesting parts: The exported function $hello_github_user and the code preceding it. First, we create a table containing a single function reference (to $create_greeting):

(table 1 funcref)
(export "table" (table 0))
(elem (i32.const 0) $create_greeting)

The idea is this: To handle a promise, we basically do .then(callback) where the callback is another Wasm function. However, because function references are not (yet) first-class values, we actually have to pass the callback as a table index. Behind the scenes, the JS wrapper will then assume that our module exports a table called "table", and get the callback with

let callback = instance.exports.table.get(index);

This is admittedly a bit messy. But it will soon get very nice, when reference types land, because then we'll just create the function reference from the function directly (no table needed).

To see WAT handling promises in action, check out the code of $hello_github_user. A couple of things to note:

• We import the JS function with (import "./github.js" "numberOfRepos" ...), another twist on import syntax convention.
• At the beginning in $hello_github_user, we make sure that we can access the input $username later, when the JS promise resolves. For that, we store the pointer in a mutable global. We also invoke $keep from watever/memory.wat, which causes the pointer to stay alive even after this function returns.
• We then call the JS function $number_of_repos, which returns a promise. Automatic lowering turns that promise into an i32 which represents an opaque external reference to the promise in JS land. In the future, this will be Wasm-native behaviour, and use the externref type, but this is so useful that watever sort-of polyfills it.
• We then call another helper, $then from watever/promises.wat. This is where we pass our promise and the table index 0 to chain another WAT function to the promise.
• We get another promise (again, opaque) and return it.

In the callback $create_greeting, the main thing of interest is that we use $free to remove the lock on the $username pointer that was created by $keep. This just means it gets deallocated like all other pointers at the end of the function call.

Under the hood

To get a glimpse at how all this works, let's look at the autogenerated JS file from our last example. I added comments and formatting for clarity.

// hello-github-user.wat.js
import { wrap } from "watever/wrap.js";
import { numberOfRepos } from "./github.js";
import { addLock, removeLock } from "watever/wrap.js";

// Wasm bytecode (base64)
let wasm = "AGFzbQEAAAABGgVgAX8Bf2ACf38Bf2AAAGABfwBgA...";

let { table, hello, memory, alloc, reset } = wrap(
  wasm,
  // exports:
  ["table", "hello#lift", "memory", "alloc", "reset"],
  // imports:
  {
    "./github.js": { "numberOfRepos#lift": numberOfRepos },
    "watever/wrap.js": { "addLock": addLock, "removeLock": removeLock },
    "js": { "(p, c) => p.then(x => c(x))#lift": "(p, c) => p.then(x => c(x))" },
  }
);

export { table, hello, memory, alloc, reset };

Some remarks:

• All JS wrapper code is contained in the wrap() function that is imported from watever, so its size impact when using multiple WAT modules in one project stays small (~1.6 kB minzipped).
• As you can see, our (import "./github.js" "numberOfRepos" ...) is converted into a JS import statement and numberOfRepos is passed to wrap() (which will pass it to WebAssembly.instantiate), together with other JS imports. On the other hand, imported WAT got bundled directly into the Wasm bytecode.
• Wasm bytecode is inlined as a base64 string. I think this is a far better default than fetching a .wasm file, for the same reason that we bundle JS modules by default: because, when deployed on the web, it saves a roundtrip to the server that is only initiated after the importing file has loaded. Code-splitting is a good thing, but should be applied with purpose and not as a random by-product of using Wasm in JS.

FAQ

Are there any non-trivial projects written in raw WAT?

Yes! As a proof of concept that watever is usable & useful, I've created watsign, a library that implements fast cryptographic signatures (ed25519) in WAT. The API and algorithms are the same as those of popular crypto library tweetnacl-js, with our version being much faster.

Another nice example of raw WAT power (unrelated to watever) is binji/raw-wasm which contains awesome demos.

How am I supposed to integrate this into my project?

You can use JS produced by the watever CLI directly in node (as ES module) or deno. To use it on the web, you'll need a second build step after watever that performs JS bundling, like webpack.

In the future, I plan to create plugins to integrate watever in popular JS build pipelines, to only have one build step. This will also enable optimizations like tree-shaking WAT modules based on what is imported from them on the JS side. In fact, this project originated from esbuild-plugin-wat which quickly got out of hand as I wanted to add more and more features 😅

TODOs

• create comprehensive API docs of watever/*.wat modules
• watch mode for the CLI
• enable post-MVP syntax that's still missing from @webassemblyjs
• merge together start sections from imported modules
• think about this: if we would go all in on externref and do lowering inside functions, we could avoid the need for implicitly exported memory and offer true compatibility with the wider wasm world