Store tracked struct ids as ThinVec on Revisions

[MichaReiser]

We don't need a map. All we use it for is to seed new ids. This helps to reduce the QueryRevisionsExtra size (which will make up for the regression that I introduce in #882)

[netlify]

✅ Deploy Preview for salsa-rs canceled.

Name	Link
🔨 Latest commit	a6231f3
🔍 Latest deploy log	https://app.netlify.com/projects/salsa-rs/deploys/683d96378d48450008a91360

[codspeed-hq]

CodSpeed Performance Report

Merging #892 will not alter performance

_{Comparing MichaReiser:revisions-identity-vec (a6231f3) with master (14318b7)}

Summary

✅ 12 untouched benchmarks

[Veykril]

Can't this be a Box<[T]> even? I don't think we ever push into it?

[MichaReiser]

Can't this be a Box<[T]> even? I don't think we ever push into it?

Is it possible to delete from a box slice? We use retain. We could convert to a Vec and then back to a Boxed slice but that could be heavy on allocations

https://github.com/MichaReiser/salsa/blob/80655d09e691fb76fc0173cceeb9a08ed278b306/src/function/diff_outputs.rs#L53-L59

[MichaReiser]

I guess, we could use a ThinVec

[ibraheemdev]

This looks good to me, I was trying to refactor the code to avoid the retain call and instead filter the tracked struct IDs when we convert the map to a list. That way we could use Box<[_]> instead, but I'm not sure if that's possible.

[MichaReiser]

This looks good to me, I was trying to refactor the code to avoid the retain call and instead filter the tracked struct IDs when we convert the map to a list. That way we could use Box<[_]> instead, but I'm not sure if that's possible.

That would be nice! It might still be worth it from a perf perspective to avoid the linear traversal