Lisp: Icing or Cake?

June 01, 2024

The Spring Lisp Game Jam 2024 ended one week ago. 48 games were submitted, a new record for the jam! This past week has been a time for participants to play and rate each other’s games. As I explored the entries, I noticed two distinct meta-patterns in how people approached building games with Lisp. I think these patterns apply more broadly to all applications of Lisp. Let’s talk about these patterns in some detail, with examples.

But first!

Here’s the breakdown of the jam submissions by language:

lang       entries    % (rounded)
----       -------    -----------
guile      15         31
fennel     10         21
clojure    5          10
cl         5          10
racket     4          8
elisp      4          8
s7         3          6
kawa       1          2
owl        1          2

I haven’t rolled up the various Schemes (Guile, Racket, S7, Kawa) into a general scheme category because Scheme is so minimally specified and they are all very distinct implementations for different purposes, not to mention that Racket has a lot more going on than just Scheme.

For the first time ever, Guile came out on top with the most submissions! There’s a very specific reason for this outcome. 11 out of the 15 Guile games were built for the web with Hoot, a Scheme-to-WebAssembly compiler that I work on at the Spritely Institute. 2 of those 11 were official Spritely projects. We put out a call for people to try making games with Hoot before the jam started, and a lot of people took us up on it! Very cool!

The next most popular language, which is typically the most popular language in these jams, is Fennel. Fennel is a Lisp that compiles to Lua. It’s very cool, too!

Also of note, at least to me as a Schemer, is that three games used S7. Hmm, there might be something relevant to this post going on there.

The patterns I’m about to talk about could sort of be framed as “The Guile Way vs. The Fennel Way”, but I don’t want to do that. It's not an “us vs. them” thing. It’s wonderful that there are so many flavors of Lisp these days that anyone can find a great implementation that suits their preferences. Not only that, but many of these implementations can be used to make games that anyone can easily play in their web browser! That was not the case several years ago. Incredible!

I want to preface the rest of this post by saying that both patterns are valid, and while I prefer one over the other, that is not to say that the other is inferior. I'll also show how these patterns can be thought of as two ends of a spectrum and how, in the end, compromises must be made. Okay, let’s get into it!

Lisp as icing

The icing pattern is using Lisp as a “scripting” language on top of a cake that is made from C, Rust, and other static languages. The typical way to do this is by embedding a Lisp interpreter into the larger program. If you’re most interested in writing the high-level parts of an application in Lisp then this pattern is the fastest way to get there. All you need is a suitable interpreter/compiler and a way to add the necessary hooks into your application. Since the program is mainly C/Rust/whatever, you can then use emscripten to compile it to WebAssembly and deploy to the web. Instant gratification, but strongly tied to static languages and their toolchains.

S7 is an example of an embeddable Scheme. Guile is also used for extending C programs, though typically that involves dynamically linking to libguile rather than embedding the interpreter into the program’s executable. Fennel takes a different approach, recognizing that there are many existing applications that are already extensible through Lua, and provides a lispy language that compiles to Lua.

Lisp as cake

The cake pattern is using Lisp to implement as much of the software stack as possible. It’s Lisp all the way down... sorta. Rather than embedding Lisp into a non-Lisp program, the cake pattern does the inverse: the majority of the program is written in Lisp. When necessary, shared libraries can be called via a foreign function interface, but this should be kept to a minimum. This approach takes longer to yield results. Time is spent implementing missing libraries for your Lisp of choice and writing wrappers around the C shared libraries you can’t avoid using. Web deployment gets trickier, too, since the project is not so easily emscriptenable.

(You may recognize this as the classic embed vs. extend debate. You’re correct! I'm just adding my own thoughts and applying it specifically to some real-world Lisp projects.)

I mentioned Guile as an option for icing, but Guile really shines best as cake. The initial vision for Guile was to Emacsify other programs by adding a Scheme interpreter to them. These days, the best practice is to write your program in Scheme to begin with. Common Lisp is probably the best example, though. Implementations like SBCL have good C FFIs and can compile efficient native executables, minimizing the desire to use some C for performance reasons.

Case studies

Let’s take a look at some of the languages and libraries used for the Lisp Game Jam and evaluate their icing/cake-ness.

Fennel + love2d

love2d has been a popular choice for solo or small team game development for many years. It is a C++ program that embeds a Lua interpreter, which means it’s a perfect target for Fennel. Most Linux distributions package love2d, so it’s easy to run .love files natively. Additionally, thanks to emscripten, love2d games can be deployed to the web. Thus most Fennel games use love2d. ./soko.bin and Gnomic Vengeance are two games that use this stack.

Fennel + love2d is a perfect example of Lisp as icing. Fennel sits at the very top of the stack, but there’s not really a path to spread Lisp into the layers below. It is also the most successful Lisp game development stack to date.

S7 + raylib

This stack is new to me, but two games used it this time around: GhostHop and Life Predictor. (You really gotta play GhostHop, btw. It’s a great little puzzle game and it is playable on mobile devices.) Raylib is a C library with bindings for many higher-level languages that has become quite popular in recent years. S7 is also implemented in C and is easily embeddable. This makes the combination easy to deploy on the web with emscripten.

S7 + raylib is another example of Lisp as icing. I’m curious to see if this stack becomes more popular in future jams.

Guile + Chickadee

This is the stack that I helped build. Chickadee is a game library for Guile that implements almost all of the interesting parts in Scheme, including rendering. Two games were built with Chickadee in the most recent jam: Turbo Racer 3000 and Bloatrunner.

Guile + Chickadee is an example of Lisp as cake. Chickadee wraps some C libraries for low-level tasks such as loading images, audio, and fonts, but it is written in pure Scheme. All the matrix and vector math is in Scheme. Chickadee comes with a set of rendering primitives comparable to love2d and raylib but they’re all implemented in Scheme. I’ve even made progress on rendering vector graphics with Scheme, whereas most other Lisp game libraries use a C library such as nanosvg. Chickadee has pushed the limits of Guile’s compiler and virtual machine, and Guile has been improved as a result. But it’s the long road. Chickadee is mostly developed by me, alone, in my very limited spare time. It is taking a long time to reach feature parity with more popular game development libraries, but it works quite well for what it is.

Hoot + HTML5 canvas

I also helped build this one. Hoot is a Scheme-to-WebAssembly compiler. Rather than compile the Guile VM (written in C) to Wasm using emscripten, Hoot implements a complete Wasm toolchain and a new backend for Guile’s compiler that emits Wasm directly. Hoot is written entirely in Scheme. Unlike C programs compiled with emscripten that target Wasm 1.0 with linear memory, Hoot targets Wasm 2.0 with GC managed heap types. This gives Hoot a significant advantage: Hoot binaries do not ship a garbage collector and thus are much smaller than Lisp runtimes compiled via emscripten. The Wasm binary for my game weighs in at < 2MiB whereas the love2d game I checked had a nearly 6MiB love.wasm. Hoot programs can also easily interoperate with JavaScript. Scheme objects can easily be passed to JavaScript, and vice versa, as they are managed in the same heap. With all of the browser APIs just a Wasm import away, an obvious choice for games was the built-in HTML5 canvas API for easy 2D rendering.

11 games used Hoot in the jam, including (shameless plug) Cirkoban and Lambda Dungeon.

Hoot + HTML5 canvas is mostly dense cake with a bit of icing. On one hand, it took a year and significant funding to boot Hoot. We said “no” to emscripten, built our own toolchain, and extended Guile’s compiler. It's Lisp all the way until you hit the browser runtime! We even have a Wasm interpreter that runs on the Guile VM! Hoot rules! It was a risk but it paid off. On the other hand, the canvas API is very high-level. The more cake thing to do would be to use Hoot’s JS FFI to call WebGL and/or WebGPU. Indeed, this is the plan for the future! Wasm GC needs some improvements to make this feasible, but my personal goal is to get Chickadee ported to Hoot. I want Chickadee games to be easy to play natively and in browsers, just like love2d games.

The cake/icing spectrum

I must acknowledge the limitations of the cake approach. We’re not living in a world of Lisp machines, but a world of glorified PDP-11s. Even the tallest of Lisp cakes sits atop an even larger cake made mostly of C. All modern Lisp systems bottom out at some point. Emacs rests on a C core. Guile’s VM is written in C. Hoot runs on mammoth JavaScript engines written in C++ like V8. Games on Hoot currently render with HTML5 canvas rather than WebGL/WebGPU. Good luck using OpenGL without libGL; Chickadee uses guile-opengl which uses the C FFI to call into libGL. Then there’s libpng, FreeType, and more. Who the heck wants to rewrite all this in Lisp? Who even has the resources? Does spending all this time taking the scenic route matter at all, or are we just deluding ourselves because we have fun writing Lisp code?

I think it does matter. Every piece of the stack that can be reclaimed from the likes of C is a small victory. The parts written in Lisp are much easier to hack on, and some of those things become live hackable while our programs are running. They are also memory safe, typically, thanks to GC managed runtimes. Less FFI calls means less overhead from traversing the Lisp/C boundary and more safety. As more of the stack becomes Lisp, it starts looking less like icing and more like cake.

Moving beyond games, we can look to the Guix project as a great example of just how tasty the cake can get. Guix took the functional packaging model from the Nix project and made a fresh implementation, replacing the Nix language with Guile. Why? For code staging, code sharing, and improved hackability. Guix also uses an init system written in Guile rather than systemd. Why? For code staging, code sharing, and improved hackability. These are real advantages that make the trade-off of not using the industry-standard thing worth it.

I’ve been using Guix since the early days, and back then it was easy to make the argument that Guix was just reinventing wheels for no reason. But now, over 10 years later, the insistence on maximizing the usage of Lisp has been key to the success of the project. As a user, once you learn the Guix idioms and a bit of Guile, you unlock extraordinary power to craft your OS to your liking. It’s the closest thing you can get to a Lisp machine on modern hardware. The cake approach paid off for Guix, and it could pay off for other projects, too.

If Common Lisp is more your thing, and even if it isn’t, you’ll be amazed by the Trial game engine and how much of it is implemented in Common Lisp rather than wrapping C libraries.

There’s also projects like Pre-Scheme that give me hope that one day the layers below the managed GC runtime can be implemented in Lisp. Pre-Scheme was developed and successfully used for Scheme 48 and I am looking forward to a modern revival of it thanks to an NLnet grant.

I'm a cake boy

That’s right, I said it: I’m a cake boy. I want to see projects continue to push the boundaries of what Lisp can do. When it comes to the Lisp Game Jam, what excites me most are not the games themselves, but the small advances made to reclaim another little slice of the cake from stale, dry C. I intend to keep pushing the limits for Guile game development with my Chickadee project.

It’s not a piece of cake to bake a lispy cake, and the way is often hazy, but I know we can’t be lazy and just do the cooking by the book. Rewrite it in Rust? No way! Rewrite it in Lisp!