Recurse Center, day 1


I left my job two weeks ago. Since then, I’ve been spending less time on the computer, more time with the cats.

But not no time on the computer. Stephen and I put together Fractal Overdrive, rendering fractals incorrectly for aesthetic effect, as a SIGBOVIK submission; and I made it through the easy half of addressing a bug in sccache.

And now: today was my first day at the Recurse Center.

Known unknowns

I’ve got some things I know I want to work on.

You are here

I want to write more. (Yes, that makes this the “sorry this blog has been so quiet” post.)

Along with that, I want to tweak how I build / host / serve this site. But working on tools and infrastructure is too much fun, so I have to give myself a budget: no more time working on serving than on writing. I only get quota once the post goes public– drafts don’t count!

Lock termination checker

In my last job, we dealt with preemption-disabled code regions and locks. We relied on code review to check that the program would eventually reenable preemtion / release the lock – a human check.

I’d like to make something that can verify this property in situ– something lighter than “formally specify the whole thing”, more like LLVM’s thread safety analyzer.

The comment I’ve received more than once about this: “Isn’t that the halting problem?” Which, almost, it’s related. But think about how the same conversation goes with human reasoning:

Alice: Hey, can you take a look at this code? Will this function terminate?

Bob: thinks

Alice: …you’ve been quiet for five minutes. Any luck?

Bob: Well… there is a loop here, and I can’t figure out whether the condition that bounds it will necessarily become true. So I don’t know.

Our checker has the option to say “I don’t know”: “I don’t have enough information”, or “I don’t have enough time”, or “I don’t have an algorithm to solve this”. The feedback here is for humans, who can decide whether “I don’t know” is a problem with the checker or the subject.

I think my first step in this will be going through llvm-tutor; and maybe stepping slightly out of that, to render the control-flow graph in dot. (Or maybe something else, per a tip from my batchmate Zach.)

Then the easy version, “always say I don’t know”; the slightly more complex version, “If the basic blocks form a DAG, say ’terminates’”; and then…

Verification? Symbolic execution? Type systems?

In “watching for that stuff in the background”, I’ve seen some more things I want to learn about.

Statup showdown

I’ve started some experiments, trying to determine which languages take the longest to get to “my code”. I have that chart - then the question is, why do they have these different properties?

The other side of this question: what tools can I use to debug performance, especially where the kernel is involved? What knobs do I have to change the performance, especially when most of the system (language runtime) is outside of my control?

Unknown unknowns

One of the points that’s been reinforced during the RC orientation is being open to opportunities. There’s a lot of interesting projects / areas that folks are or are starting to work on… and I know I’ve got a lot to learn about a lot of things.

So…maybe I’ll port Fractal Overdrive to WGPU? Or start something in the Godot engine? Or… just see what else folks need a partner on.