✅ Deploy Preview for salsa-rs canceled.

CodSpeed Performance Report

Merging #969 will improve performances by ×3.4

_{Comparing MichaReiser:store-tracked-struct-ids-only-once (6a63bff) with master (a2cd1b8)}

Summary

⚡ 1 improvements
✅ 11 untouched benchmarks

Benchmarks breakdown

This looks promising :)

The results on ty's benchmarks aren't as impressive but still a solid 1-3% performance win and a reduction in memory usage.

The memory usage, for posterity: astral-sh/ruff#19843 (comment)

Uhhm, why?

I don't get it. It seems that now there's a huge regression after I rebased past the persistence changes but they shouldn't matter?

I can't reproduce this regression locally or in ty. I'm not sure what's up here with codspeed I'm sure something changed but it isn't clear what and I wonder if it's just because it now sees the LocalKey frame. I also noticed that the numbers differ by a lot if I make a change which the following commit undoes again.

Hmm, I think this introduced a subtle regression regarding fixpoint iteration.

Before: Salsa seeded the tracked_struct_ids in every iteration with the values from the previous revision and the values from the previous iteration.

Now: Salsa seeds the tracked_struct_ids initially with the values from the previous iteration. Later iterations seed tracked_struct_ids with the values from the previous iteration but any tracked struct id that wasn't created in the first iteration isn't copied over anymore because the new QueryRevisions removes unused tracked struct ids eagerly and not only when diff_outputs is called (which we only ran between the memo from the last revision with the memo from the new revision).

Now, this is sort of a pre-existing issue for queries that aren't cycle heads because they already had the behavior that we now observe for cycle heads too.