At least for the docs stuff, I wasn't intending to use MonadFail (in fact I think it's a huge antipattern), so unless there's a particular reason we want to update to GHC 8.6.x very soon, I'd rather refactor so that these sorts of changes won't be necessary and then upgrade; does that sound okay?

@hdgarrood For sure! I don't intend on making changes in that area anymore and you've got push permissions for this branch, so knock yourself out :)

Ah I see: most of the MonadFail constraints are actually arising from inside the type checker, and they're just propagating out to everywhere else.

Since most of them seem to be due to pattern matches like

  expr@(TypedValue _ value ty) <- infer ...

how do we feel about having infer return something like

data InferredValue = InferredValue
  { inferredCheck :: Bool
  , inferredValue :: Expr
  , inferredType :: Type
  , inferredExpr :: Expr
  }

where the inferredExpr field is always a TypedValue whose fields are the same as the other three fields? Then we can replace the above with

  InferredValue _ value ty expr <- infer ...

and we can recover the current infer with fmap inferredExpr . infer.

Actually how about we just set -XNoMonadFailDesugaring for now (which is the same as what we have now), and refactor so that we aren't using fail in a later PR? If we want, we could put it off until we upgrade to GHC 8.8 (at the latest).

I guess I'd be fine with setting that flag. We could also just make a function that does that pattern match and throws an internal error when it fails (as that's essentially the behaviour we've got right now).

Yep, that should be fine too. I think I prefer the InferredValue approach though, since the compiler can help you avoid making mistakes that way.

Looks good to me! Sorry for not getting to this myself.

No worries at all :)