Miri reports Undefined Behavior in tests about thunderdome

[CosmicHorrorDev]

Looks like you can. I opened #42

[CosmicHorrorDev]

Went ahead and pushed changes to consolidate a lot of the branches

[LPGhatguy]

Hey, thanks for checking this out! MIRI is serious business, we should fix these issues and add it to our CI pipeline.

[parasyte]

Similar to orlp/slotmap#92 (comment), using MIRIFLAGS='-Zmiri-tree-borrows' does not report any issues!

[CosmicHorrorDev]

Noticed this because of the linked issue

It looks like this crate is very similar to slab which implements .get2_mut() by using .split_at_mut() on the underlying Vec to get two mutable slices that each contain one of the desired elements. Then you can call .get_mut() on each of the slices to get a mutable reference

https://docs.rs/slab/latest/src/slab/lib.rs.html#751-773

Do you think the same technique would work here? It allows you to avoid using unsafe

[LPGhatguy]

Thanks for the investigation and PR!

In general, I'm happy to reduce unsafety in crates when there's an easy replacement. In this case though, I'm a little concerned — this safe alternative has a lot more code! It's hard to make sure that all of these branches are tested and work correctly when compared with the original code.

Based on orlp/slotmap#92 (comment), I'm not sure we should change the implementation of this method. It seems like there might be a compelling argument that MIRI is raising a false alarm here. I am not familiar enough with the current status of UB in Rust to know for sure.

[CosmicHorrorDev]

The main part that has a lot of explicit branches is when the two indices occupy the same slot. I was mostly trying to avoid duplicate checks, but they could probably just use .get_mut() and the compiler should be smart enough to remove the duplicate checks. That should reduce a lot of the near-duplicate logic

I was mostly just making this change to reduce unsafe usage, not because I'm worried about aliasing issues here (which the compiler doesn't have a firm stance on yet). slab seems to be largely the same as this library aside from the generational indices and manages to avoid any unsafe aside from a couple of unsafe methods they expose for comparison