Autocomplete often fails on new version

[sanbox-irl]

Hello!

Right now, having just upgraded the RA -- autocomplete often appears to fail to successfully return results while typing normally. If I hit escape, and then resume typing, it will find autocomplete solutions, though it appears slightly laggy.

I am running on a relatively slow, laptop processor, runnings VSCode, and the most recent version of the RA (the one released just a few hours ago).

I can take a video if that would help?

[SomeoneToIgnore]

Hello.
A video/gif would be helpful, it looks like we're dealing with multiple issues here.
Also a glance at the project with the issue might help.

I've created #6614 that helps the issue a bit.
We might consider adding a setting toggle to disable this particular feature, since there might always be a slow laptop with RA somewhere 🙂

It should be straightforward, here's the completion config:
https://github.com/rust-analyzer/rust-analyzer/blob/156f7d6963fb6b47570bfa457fdf51733a182054/crates/completion/src/config.rs#L10

and its field usages show the rest of the places to fix.

[georgewfraser]

Same issue, and I am on a pretty fast laptop. This is extremely annoying---is there a way to revert to the old version?

[georgewfraser]

(notice how if you type two characters in rapid succession, autocomplete fails and VSCode reverts to string-based autocomplete)

[kjeremy]

This may be related to #6565

…

On Tue, Nov 24, 2020, 3:25 PM George Fraser ***@***.***> wrote: (notice how if you type two characters in rapid succession, autocomplete fails and VSCode reverts to string-based autocomplete) — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#6612 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABBACRHI5ZETHLADXPO7U4LSRQJEXANCNFSM4UAALQVA> .

[SomeoneToIgnore]

We've added a feature toggle to the latest nightly: #6631
Will be in the next stable release on Monday.

It's on by default, since there is not that much feedback on the completion issues, I've decided to leave it on by default.

[sanbox-irl]

Great news! Thank you, I appreciate that a lot!

@georgewfraser in the meantime, until monday (unless you want to go nightly) -- you can revert the client in VSC trivially. go to the extensions panel and then right click on Rust Analyzer, and then you can change version back to last monday.

However, you'll probably (not sure on the backend here) need to replace the rust-analyzer rust binary. To replace that, you'll probably need to go to the releases tag here in this repo, and download the exe (or compile it yourself) and then give the path to it in your preferences in VSC as an override for the server location. It sounds more complicated than it is; it's really just replacing a file.

[sanbox-irl]

Going along with this idea here -- why do I, and the other person in this thread, sometimes get no results when we're typing? How does that RA handle that? Is there effectively some amount of time x that it will give to answer the query and then will fail?

I was imagining that auto-imports, because it's necessarily a "fuzzy" operation, aren't as important as current imports. Would it be possible to suggest what RA suggested last week (the ones already in scope), and then, if there's enough time/performance is good enough, also gather the auto-imports?

Apologies if this is already what you do, or this has already been thought. I'm only a user and not a contributor, so I'm sorry if I'm overstepping a bit

[SomeoneToIgnore]

sometimes get no results when we're typing

That can be caused by multiple issues, a few ideas from me:

• before the autocompletions feature changes, VSCode behaved so that after the initial set of completion proposals is received, it starts to filter out this set on every new symbol added up intil there are no results that contain all symbols you've entered, and only then it queries for more.
  After the completion changes, we've switched VSCode to be more eager on requesting the new items (even if the fuzzy completions are off) that is supposed to help with the situation.

• there can be various issues with the way RA derives the completions: it needs good type inference (which sometimes can fail for complex cases, procmacro cases and similar) and it needs good fallback mechanisms which sometimes we lack.

I don't have a real-world example now, but something like this (where <|> is your caret and you're typing more to get the completions)

fn test() -> Result<(), ()> {
    let a = ha<|>
    Ok(())
}

can confuse RA, since it considers Ok(()) to be related to what you type and it might cause various side effects up to zero completion proposed.

I'm afraid I cannot get more specific without any MRE that does not work for you.

suggest what RA suggested last week (the ones already in scope), and then, if there's enough time/performance is good enough, also gather the auto-imports?

We do something similar with fuzzy imports already: first (and always) we fill in the "regular" completions, the ones that were before the new feature, then we add the fuzzy ones to the same suggestion set.
We don't check for "if there's enough time/performance" conditions and I'm not sure if we should: it's hard to formalise and might be tough to implement.
IMO it's more beneficial to invest into the performance improvements instead and work on the completion prioritisation to show better results.

[georgewfraser]

@SomeoneToIgnore does rust-analyzer have performance benchmarks to detect regressions? If not, I might be interested in setting that up...I've contributed before, in the syntax colorization area.