Provide a method to attach a database even if it's different from the current attached one

[ChayimFriedman2]

rust-analyzer needs this.

[netlify]

✅ Deploy Preview for salsa-rs canceled.

Name	Link
🔨 Latest commit	678a650
🔍 Latest deploy log	https://app.netlify.com/projects/salsa-rs/deploys/68cbc75f28484c0008459e98

[codspeed-hq]

CodSpeed Performance Report

Merging #992 will not alter performance

_{Comparing ChayimFriedman2:attach-allow-change (678a650) with master (3713cd7)}

Summary

✅ 12 untouched

[ChayimFriedman2]

The Clippy failure seems unrelated?

[MichaReiser]

Can you say more why RA needs this? Isn't this technically unsound because it means that the outer query won't capture all dependencies?

[Veykril]

We currently create sub databases to speculatively do highlighting in doc strings and the like. We already take care of not intermixing queries between these, but we are now in a situation where we create these temporary sub databases within query calls (we still don't mix things though) and that now breaks due to the checks salsa does.

I do think we should mark this unsafe or give it a dummy param to mark it as problematic if misused (if it is not unsound, unsure if this is a soundness problem or not given you can already smuggle IDs between databases).

[MichaReiser]

Funny, Niko and I just talked about some changes to fixpoint iteration that might also allow speculative execution

[ChayimFriedman2]

I don't think this is unsound? First, you cannot avoid recording query deps with this, as you can't call queries on the old db (they will attach it again), and second, even if you could I don't think not tracking some inputs is unsound.

[ChayimFriedman2]

Funny, Niko and I just talked about some changes to fixpoint iteration that might also allow speculative execution

We do like speculative execution, but the speculative execution that @Veykril talked about is of another kind and unrelated to Salsa.

[ChayimFriedman2]

What's up with the test?

[MichaReiser]

The output is different based on compiler version. You need to gate this test to only run for Rust version >= whatever version beta is

[ChayimFriedman2]

Locally it outputs what I pushed now. But then it errors at CI. And if I update to the CI version it errors again.

[Veykril]

The compile fail tests are gated by the persistence feature which is disabled by default, so they don't run with a plain cargo test