Wingman gets slow on big modules

[isovector]

Even for things like "refine hole" which should be instant. My guess is that waiting for a fresh type-checking result is to blame. We could probably get away with using stale data instead.

[pepeiborra]

GHC is very slow at typechecking holes. Even if Wingman used stale data, the default ghcide code action handler would still be slow and bottleneck the code action response. You may want to play with the settings we use:

haskell-language-server/ghcide/src/Development/IDE/GHC/Compat.hs

Lines 204 to 222 in 9c40dcf

	setUpTypedHoles ::DynFlags -> DynFlags
	setUpTypedHoles df
	= flip gopt_unset Opt_AbstractRefHoleFits -- too spammy
	#if MIN_GHC_API_VERSION(8,8,0)
	$ flip gopt_unset Opt_ShowDocsOfHoleFits -- not used
	#endif
	$ flip gopt_unset Opt_ShowMatchesOfHoleFits -- nice but broken (forgets module qualifiers)
	$ flip gopt_unset Opt_ShowProvOfHoleFits -- not used
	$ flip gopt_unset Opt_ShowTypeAppOfHoleFits -- not used
	$ flip gopt_unset Opt_ShowTypeAppVarsOfHoleFits -- not used
	$ flip gopt_unset Opt_ShowTypeOfHoleFits -- massively simplifies parsing
	$ flip gopt_set Opt_SortBySubsumHoleFits -- very nice and fast enough in most cases
	$ flip gopt_unset Opt_SortValidHoleFits
	$ flip gopt_unset Opt_UnclutterValidHoleFits
	$ df
	{ refLevelHoleFits = Just 1 -- becomes slow at higher levels
	, maxRefHoleFits = Just 10 -- quantity does not impact speed
	, maxValidHoleFits = Nothing -- quantity does not impact speed
	}

[pepeiborra]

Let me insist on playing with the settings we use - Opt_SortBySubsumHoleFits was only "fast enough" for occasional use, and maybe that's no longer good enough

[isovector]

@pepeiborra my understanding is that we currently use the hole fit suggestions to fill in the replace _ with <whatever> code actions, and I'm a little scared to break that.

[pepeiborra]

Some of those options, like Opt_SortBySubsumHoleFits, may not be worth the computational cost. So by all means feel free to tweak them and test the results! At the very least we could expose a config option to enable users to decide

[isovector]

I've had all hole fits turned off globally since 2018 for exactly this performance reason :)

[isovector]

Was playing around with this today, and typechecking is significantly (and noticeably) faster having disabled the hole fit suggestions. Seeing as they mostly suggest things like replace _ with id _, I think the performance gains here more than make up for the slight loss of functionality.

[anka-213]

Was the performance still too bad after disabling Opt_SortBySubsumHoleFits? I would much prefer disabling that and maybe refinement hole fits over disabling type hole fits entirely.

[isovector]

@anka-213 I just tried that this morning, and the behavior is still wicked slow. After case-splitting a HsExpr GhcPs, having hole-fit (even without subsumption holes) suggestions on takes upwards of 1 minute. And this gets invalidated on every keystroke. Comparatively, disabling all hole-fits means the code actions come up in <1 second.

[anka-213]

Ah, that's too bad! Thanks for testing it!

[isovector]

@wz1000 and I were discussing on IRC why this is so likely so slow:

<isovector1> wingman can't (today) use whatever's in scope
<isovector1> so it's not a drop-in replacement :)
<isovector1> i'm happy to just implement that if there's an easy way to see what's in scope
<wz1000> isovector1: I suspect "using whatever is in scope" is where most of the slowdown comes from...
<isovector1> yeah, my guess too?
<wz1000> since you have to typecheck *all* the interfaces
<isovector1> is that stuff not cached?
<wz1000> usually interface typechecking is done lazily, so you only pay for the bits you need
<wz1000> i.e. when you actually use a function

Sounds like this functionality in general is anathema to any sort of good performance.

As an alternative, I'd be happy to implement the same functionality in Wingman, except rather than suggest anything in scope, it suggests only things defined in the current module, plus a whitelist of other functions from anywhere if they're in scope. What do you think, @anka-213 ?