Hacker News Reader

Comment by aw1621107 | Hacker News Reader

Thanks for the link! That's an interesting approach indeed. If I'm reading the paper [0] correctly the concepts involved look pretty similar to what Rust uses. I think they turn the interprocedural analysis into an intraprocedural one as well by effectively deriving borrowing information from function implementations and attaching that to signatures for use by callers. It's a compatibility hazard, which is why Rust chose to not do the same, but in principle it makes it easier to derive more precise borrowing information.

[0]: https://dl.acm.org/doi/epdf/10.1145/3814943.3816182

View on HN

They globally derive the borrowing information for all pointers in the whole program by classifying the role of each pointer. If there is a way to do this in a way that makes the program safe, this is computed automatically using a SAT solver.

> They globally derive the borrowing information for all pointers in the whole program by classifying the role of each pointer.

Right; by my reading this is done function-by-function by the constraint generator before passing that information to the solver for verification.

The classification can not be done on the function level. There are two ways, either this is done using AI or using a SAT solver.

The paper explicitly says that if you use an LLM it is prompted function-by-function.

For the SAT solver I think classification can be done function-by-function if you manually drive the solver in the single-function mode, but I think you're more right in that the whole-program mode doesn't actually do that.

I think my misunderstanding stemmed from interpreting the start of section 5.4 and figure 3 as showing that constraint generation occurred strictly before the verifier was invoked and glossing over the bit at the start of section 6 that said that the process basically works the other way around when using the solver to build the p-model.

Perhaps the LLM can infer enough context from one function.

Possibly? Though given its outputs apparently aren't completely out there I'd hazard a guess that the inferred context usually shouldn't be too far off of what can be deduced using more traditional methods.

Reply on news.ycombinator.com