Replace checkHomeUnitsClosed with a faster implementation

[wz1000]

GHC gives a default "main" home unit, and it is difficult to avoid this as
each ghc session must have an active home unit at all times, but in a multiple
component session, there is no other good choice for a default unit. We could
pick one arbitrarily, but this is ugly and complicates the code a lot.

We never use this default "main" unit for anything, as the GHC sessions
corresponding to any file/component have the active unit set to the correct one
for that component.

When checking for home unit closure, we must make sure to include only the
actual units in the project, not the bogus "main" unit that GHC forces us to
have. Including the main unit seems to make the checkHomeUnitsClosed function loop
forever for some reason.

Use a faster implementation of checkHomeUnitsClosed
GHC had an implementation of this function, but it was horribly inefficient
We should move back to the GHC implementation on compilers where
https://gitlab.haskell.org/ghc/ghc/-/merge_requests/12162 is included

Fixes #4046

[michaelpj]

Hooray!

Please put the reasoning into the code! Can we have a test case? It would be nice to know what its going wrong with the checkHomeUnitsClosed function - it would be much better for it to throw or fail somehow rather than looping!

[wz1000]

I'm investigating that currently, it seems like checkHomeUnitsClosed is just extremely slow - it traverses all paths in the unit dependency graph, which seems to work when your dependency closure is small but quickly grows to be massive.

[michaelpj]

I'm investigating that currently, it seems like checkHomeUnitsClosed is just extremely slow - it traverses all paths in the unit dependency graph, which seems to work when your dependency closure is small but quickly grows to be massive.

That definitely seems bad. I would have thought this was a standard graph problem of some kind 🤔

[wz1000]

I've put up a patch here, I will copy the implementation to ghcide so we can use it on released GHCs https://gitlab.haskell.org/ghc/ghc/-/merge_requests/12162

[fendor]

Idea looks great, I do have a couple of questions, though.

[michaelpj]

I'm assuming we're going to wait for the GHC MR to land so we can incorporate any comments from upstream.

[ch1bo]

Seems like https://gitlab.haskell.org/ghc/ghc/-/merge_requests/12162 will get merged / backported. Can we get this proper workaround for < 9.10 compilers now? :)

[michaelpj]

Yes, we should be able to copy in the final code now.