Improve usability for "make variable mut" suggestion

[camelid]

With code like

let v = vec![];
// ....
v.push(123);

RA suggests that I make v mut (which I expect!). However, in order to apply the suggestion, I have to navigate to the space between let and v, at which point it's no faster than typing it myself.

When I used to use RLS, I could press Cmd-. at the point of use (v.push) and have it apply the suggestion to the declaration place. Could RA do this as well? It'd be a great quality of life improvement IMO.

Originally posted on Zulip.

[flodiebold]

Related #8449, though this one would be just about providing the rustc suggestion in more places. I think we used to do this, but changed it because it led to some problems?

[jonas-schievink]

The easy solution here is to also provide the quick fix where the main error is reported. The problem is that, for other diagnostics, this will end up with quick fixes that say something like "consider removing this semicolon", while not pointing at a semicolon at all (selecting the quick fix will remove a semicolon somewhere else). This is very confusing and something I'd like to avoid.

Maybe this specific diagnostic could be implemented natively now? That way we'd have full control over how it looks.

[camelid]

Maybe this specific diagnostic could be implemented natively now? That way we'd have full control over how it looks.

Hmm, I guess this particular error doesn't really need the borrow-checker?

[jonas-schievink]

Seems to me that the only tricky part of the diagnostic is implicit borrows, so it has to check not only for &mut <X>, but also for mutable auto-ref in method calls, and also for mutable auto-deref (anything else?).

[camelid]

Yeah, I think as long as RA can see the type signature of the called function (including if it's in a trait), it should be fairly straightforward to determine if there's mutation (famous last words...).

[flodiebold]

We already keep track of autorefs / derefs. I think there's a bit of a tricky part with keeping track where we "need mut" during inference so that everything in &mut foo.bar[baz].qux gets marked as mutable; especially with Index vs. IndexMut and Deref vs. DerefMut. I don't think it's super complicated, but it takes some care and we have to figure out how to represent this information during inference (probably by looking into how rustc does it).

[HKalbasi]

Native diagnostic is implemented in #14232. The suggestion for rustc diagnostic still only works on declaration but I think the native diagnostic is good enough that fixing rustc suggestion doesn't worth the effort.