refactor queries that don't accept inference variables to use canonicalization

[nikomatsakis]

In #86866 we found various queries that were ICEing when used with inference variables. #86866 worked around these ICEs via simple checks or imprecision, but these queries likely out to be rewritten to use canonicalization or other less hacky techniques.

Discussion with @arora-aman in this Zulip topic, starting about here.

[nikomatsakis]

In the code, some of the places that have to change are marked with FIXME(#86868).

[b-naber]

I could work on this, or do you want to take this, @arora-aman ?

[jackh726]

@b-naber I think you should feel free to claim and work on this

[ouz-a]

If mentorship is still available I would like to work on this

[vincenzopalazzo]

@rustbot claim

[tnuha]

@rustbot claim
It seems like this FIXME is still present in the compiler. Would love some mentorship if still available, otherwise my short notes on this issue (for personal or otherwise) are as follows:

Notes no longer significant due to own research, but would still love any available communication.

#86866 introduced a hack that reduced queries with inference variables to a simple "if this query type has inference, we say it has a significant drop." The "significant drop" bit seems to be related specifically to a clippy lint that's still in nursery, am I right? Or is the goal of this issue more along the lines of checking the queries changed with #86866 and seeing if we can refactor those towards "less hacky" approaches? If it's more this last idea, I wonder again if this already has a tracking issue/something else, and this issue (i.e. #86868) has been inactive for so long for that reason.

Am I right on my current understanding? If it helps any, I've got a Zulip chat for us here: #t-compiler/help > old query refactoring issue, any progress that has been had? @ 💬

[tnuha]

I am currently interpreting the state of this issue as resolving the only remaining FIXME comment:

rust/compiler/rustc_middle/src/ty/util.rs

Lines 1378 to 1384 in 321a89b

	// FIXME(#86868): We should be canonicalizing, or else moving this to a method of inference
	// context, or *something* like that, but for now just avoid passing inference
	// variables to queries that can't cope with them. Instead, conservatively
	// return "true" (may change drop order).
	if query_ty.has_infer() {
	return true;
	}

But will try and start a discussion on Zulip regarding whatever other queries may benefit from canonicalization once I have some significant work there.

[tnuha]

As a quick note, it doesn't immediately break any tests to just remove the block mentioned in my previous comment. In terms of complex refactoring, that's about as simple as it gets so far. I'll double check with some clippy testing first, though.