DEV Community

Cover image for In And Out Of Rabbit Holes
Ben Lovy
Ben Lovy

Posted on • Updated on


In And Out Of Rabbit Holes

More Fun With WebAssembly

In the past week, I've fallen into the following rabbit holes wholly outside of the necessary requirements while getting my new side project off the ground. Here's how I've gotten out.

Dynamic dispatch in Rust

Rust binaries can bloat pretty quickly, and one way to mitigate this is to prefer dynamic dispatch (i.e. Box<dyn Trait>) to monomorphization (fn<T: Trait>(obj: T) {}). In brief, the latter generic function syntax will create a completely separate function implementation in your binary for each type it's used with, and will select the properly typed version at compile time. The former will not bind this function call at compile time, and instead merely ensure that the trait indicated is implemented. It will then dynamically bind the method call at runtime for the type it's ultimately called from.

I rewrote all my generic functions to take trait objects and my bundle got smaller, YMMV.

As it turns out, this is not an insignificant rewrite. One thing to note is that while a Box<dyn Trait> is Sized (because a Box is sized), a trait object (the dyn Trait part) by definition cannot be. The type is unknown, all we do know is that it implements this interface, and the method calls will be dispatched dynamically at runtime. This means you'll need to do some manual plumbing if you want to automate trait-to-trait machinery, and be prepared to get up close and personal with the borrow checker. I'm still not quite sure I've got it right, but it does compile.

As an example, I've got two related traits, Drawable and Widget. A Drawable is a type that knows how to paint itself to the canvas, and a Widget is an organizational component that contains a 2d grid of child Widgets.

Some widgets are also Drawable, of course, so that eventually something gets drawn to the screen. The eventual idea is to provide a set of Widgets generic enough that games built on top of this library (or whatever) never need to manually implement Drawable, they can just compose Widgets like Text and Button and Area which handle all the details, and I'm using the game I'm building on top of it to drive what widgets need writing.

The easiest way I could figure out how to make this work is to have implementors of this trait write a function that returns some other type, MountedWidget:

/// Trait representing things that can be drawn to the canvas
pub trait Drawable {
    /// Draw this game element with the given top left corner
    fn draw_at(&self, top_left: Point, w: WindowPtr) -> Result<Point>;
    /// Get the Region of the bounding box of this drawable
    fn get_region(&self, top_left: Point, w: WindowPtr) -> Result<Region>;

/// Trait representing sets of 0 or more Drawables
/// Each one can have variable number rows and elements in each row
pub trait Widget {
    /// Make this object into a Widget
    fn mount_widget(&self) -> MountedWidget;
Enter fullscreen mode Exit fullscreen mode

The MountedWidget provides its own Drawable implementation that knows how to methodically draw its way through a grid of children, and optionally contain a raw Drawable itself:

/// A container struct for a widget
pub struct MountedWidget {
    children: Vec<Vec<Box<dyn Widget>>>,
    drawable: Option<Box<dyn Drawable>>,

Enter fullscreen mode Exit fullscreen mode

Part of me thinks I should be able to streamline this even further and avoid allocating an intermediary struct, but this setup got me something working. The unfortunate part is that as written every widget gets re-created and dropped for every frame - clearly the way to go is to mount it all first and adjust as needed but it's a start, at least.


I've said it before and I'll say it again: cargo is the crème de la crème of package managers. Everyone else is missing out.

One way I hoped to leverage it was by pulling out my Canvas mounting and drawing stuff as its own crate, and letting incremental compilation cache the build separately. As it turns out, this was really really easy. Here's what a standard library with three modules might look like, directory-wise:

$ tree
├── Cargo.toml
└── src

1 directory, 6 files

Enter fullscreen mode Exit fullscreen mode

To turn your "drawing" module into its own crate, make it look like this:

├── Cargo.toml
└── src
    ├── drawing
    │   ├── Cargo.toml
    │   ├── LICENSE
    │   ├──
    │   └── src
    │       └──

Enter fullscreen mode Exit fullscreen mode

That's all! In Cargo.toml for the parent crate, just add the dependency:

path = "src/drawing"
Enter fullscreen mode Exit fullscreen mode

It could not be easier, and the efficiency gained makes a real difference especially with these hefty WASM builds. Feel free to grab that directory and plop it anywhere you like (i.e. hosted in a git repo), you can point your Cargo.toml where you need.

Size optimization

This link in the RustWasm book has some good tips. You need to install the Binaryen toolkit to get the full benefit - it can run further speed and size optimizations on your compiled WASM output beyond what LLVM will do via rustc. You'll need to have cmake installed, which is available in all major repositories (apt,homebrew,chocolatey, etc.)

$ git clone
$ cmake . && make
Enter fullscreen mode Exit fullscreen mode

It will take a little while. There's several frontends we won't be using, see the readme for usage. I just symlinked wasm-opt to my user's path:

$ ln -s /home/ben/code/extern/binaryen/bin/wasm-opt /home/ben/.local/bin/
Enter fullscreen mode Exit fullscreen mode

I then wrote a script to handle the wasm-opt call:


function wasm_size {
    wc -c $1

function echo_size {
    echo "$(eval wasm_size $1)"

function extract_size {
    wasm_size $1 | sed 's/^\([0-9]\+\).*/\1/'

# $1 = target $2 = focus $3 = level
function shrink {
    if [ "$2" = "size" ]; then
        if [ "$3" = "aggro" ]; then
        if [ "$3" = "aggro" ]; then
    COMMAND="wasm-opt $ARG -o $1 $WASM"
    echo $COMMAND
    eval $COMMAND

function choose_smaller {
    shrink $NORMAL_TARGET $2 $3
    NORMAL_SIZE="$(eval extract_size $NORMAL_TARGET)"
    shrink $AGGRO_TARGET $2 $3
    AGGRO_SIZE="$(eval extract_size $AGGRO_TARGET)"
    if [ $NORMAL_SIZE -lt $AGGRO_SIZE ]; then
        echo "Normal settings smaller, saving..."; mv $NORMAL_TARGET $WASM; rm $AGGRO_TARGET;
        echo "Aggressive settings smaller, saving..."; mv $AGGRO_TARGET $WASM; rm $NORMAL_TARGET;

# parse args
for i in "$@"
case $i in
    # unknown option
# last line is target, non-opt, no equals sign
if [ -n $1 ]; then

echo_size $WASM
if [ -z $FOCUS ]; then
echo "Shrinking, optimizing for ${FOCUS_STR}."
if [ "$LEVEL" = "aggro" ]; then
    echo "Using aggressive optimizations."
if [ "$FOCUS" = "size" ]; then
    choose_smaller $1
    shrink $WASM $FOCUS $LEVEL
echo_size $WASM

Enter fullscreen mode Exit fullscreen mode

There's also a Makefile to call it for me:

.PHONY: all clean help

RUSTCLEAN=cargo clean
RUST=wasm-pack build
OPT=./ -f=speed -l=aggro

all: $(PKGDIR)/$(EXEC)



    @echo "Usage: make {all|clean|help}" 1>&2 && false
Enter fullscreen mode Exit fullscreen mode

The script passes the proper args to wasm-opt for either size or speed, and aggressive or normal. If you choose size, it runs it both with aggressive or not and saves the smaller of the two. To tweak it, set the options in the OPT line of the makefile. Seems like I should be able to get some mileage out of this setup for now.

This project had a rabbit-hole-in-a-rabbit-hole writing the extract_size function in the bash script. The sed call was my first solution. Then I decided for some odd reason I wanted to try to do it without a call or subshell, using string substitutions like the argument matching or something. What a waste of a morning. I'm sure there's a simple solution staring at me in the face, but I didn't find it and even if I had it would have made no difference. Why do we do this to ourselves?


This is less of a rabbit hole so much as a way around them.

Use console_error_panic_hook. With it, when your module panics you get actual useful error output to the browser console instead of just "Unreachable executed". This is obviously an improvement.

Also, wasm-pack build runs a release build by default, with no debug symbols. When debugging, use wasm-pack build --debug or add debug = true to your Cargo.toml. Now your errors will actually have the name of the Rust function that tripped instead of webassembly[37] or some nonsense. I didn't realize this for too long and thought debugging WASM apps was just like that. It doesn't have to be like that.

Stay tuned

Even with all this time....alternatively spent, the thing does the thing and my little grid of test widgets is accurately painted to the canvas, so I chalk it up as a successful week. Up next, Ben attempts to provide a procedural macro-style DSL! This should be a mess, you won't want to miss it.

Photo by Gary Bendig on Unsplash

Top comments (2)

jeikabu profile image

I use workspaces in a lot of my rust projects. But now that I'm looking at how you structured this I'm wondering why I bother...

deciduously profile image
Ben Lovy

Yeah, that was what I reached for first, rembering this chapter, but it didn't seem to solve the right problem. For this particular case, I don't want to share a Cargo.lock - the eventual goal is the split this crate out entirely into a separate project. Maybe I'll revisit as the project grows but for now I don't feel a need for structure beyond modules and crates.