Hacker News new | ask | show | jobs
by quantummagic 35 days ago
I'm curious, but unable to ascertain, does the same problem exist in the original Zig code? Is this an issue introduced by the translation to Rust? Because if it is a problem that can be replicated in both code bases, it seems a point in Rust's favor, that the issue is easily identifiable with tools that exist in its ecosystem.
3 comments
K0nserv 35 days ago
I'm also curious about that. One thing to keep in mind: the invariants you have to uphold in unsafe blocks are quite stringent. I expect that in some instances the Rust code has new UB due to this.
link
endospore 35 days ago
No. They introduced quite a few aliasing issues that result in immediate UB in Rust but are allowed in Zig. I skimmed over the unsafe blocks with rg for ten minutes and spotted like 3 cases and these wouldn't exist in the original code.

Also the LLM is inventing ways to get around borrow checker limitations with unsafe which is also concerning. But I didn't verify if they are UB or not (either way it's a rejection if I'm to review the code).

link
tuetuopay 34 days ago
As for the specific issue: it does not exist in Zig, because Zig does not have ownership.

In a nutshell, the LLM created abstractions that allow you to write unsound code in safe rust, which is squarely against the language.

To be specific: the abstraction takes a (shared) reference and uses unsafe to wrap it in an owned object, completely erasing le lifetime. In practice, this means users of the abstraction think they own the underlying memory: they choose when to free it. However, it just wraps a pointer that’s owned by someone else (it was a shared reference, remember?), thus it will be freed when you don’t expect it.

So why does it not exist in Zig: it’s a false contract about what it is. The Zig pointer is a pointer with no added lifetime information. You can hold a Zig pointer wrong, but you will hold a lying abstraction wrong. You will misuse it because it doesn’t do what’s written on the tin. You will write bugs with it.

And, LLMs will too. If they do not have the abstraction definition in their context, they also have no way to know the contract is lying.

link