Signature help suppresses arity error in editor

[andrewbranch]

Bug Report

🔎 Search Terms

signature help missing errors

🕗 Version & Regression Information

Saw this for the first time a few months ago, but could be older.

⏯ Playground Link

Playground link with relevant code

💻 Code

declare function f(x: string, y: number): any;

// 1. Put cursor inside call parens
// 2. Request signature help with cmd+shift+space
// 3. Type a quote mark
// 4. Observed behavior: no squigglies
f()

🙁 Actual behavior

No error after following instructions, until making another edit elsewhere / with signature help closed

🙂 Expected behavior

Arity error on the call

[RyanCavanaugh]

I'm pretty sure we have an incremental parser bug. If I move atomically (with pasting) from

fun('a', 3, 3, 'a', 3)

to

fun('a', 3, 3, 'a', 3, 3)

then the error is always presented.

If I insert the comma, then the argument, with delays, then it's a reliable repro that the error doesn't appear. More tellingingly, atomic edits from that point forward don't present an error unless I remove the line and put it back again

[andrewbranch]

Could be, but I’m pretty sure signature help is also related somehow. I have automatic signature help turned off in my editor but frequently invoke it manually. By default, signature help appears when you type a comma, so that creates another difference vs. pasting. But in my editor, I can’t repro without invoking signature help.

[DanielRosenwasser]

Fairly certain it's not related to incremental parsing. If you dismiss signature help and start typing, errors show up again. Manually triggering and typing makes it disappear.

signatureHelpPoisoningTypeCheckerCache.webm

Here's my theory from #52369:

What are the mechanics here that're making this happen? If you're using VS Code, that initial , will start a signature help session. Every subsequent character will likey be used as a retrigger for signature help prior to requesting diagnostics. And so signature help (which I recall is checked in a specific way to handle overloads that differ slightly in arity) keeps getting requested for a specific expression before the rest of the type-checker gets to walk over it to report errors.

I believe that because signature help gets checked in a special way, it is likely poisoning the type-checker state we use for reporting errors as well.