Note [Compatibility]

GHC/Core/FamInstEnv.hs:494 compiler

Two patterns are /compatible/ if either of the following conditions hold:
1) The patterns are apart.
2) The patterns unify with a substitution S, and their right hand sides
equal under that substitution.

For open type families, only compatible instances are allowed. For closed
type families, the story is slightly more complicated. Consider the following:

type family F a where
  F Int = Bool
  F a   = Int

g :: Show a => a -> F a
g x = length (show x)

Should that type-check? No. We need to allow for the possibility that 'a'
might be Int and therefore 'F a' should be Bool. We can simplify 'F a' to Int
only when we can be sure that 'a' is not Int.

To achieve this, after finding a possible match within the equations, we have to
go back to all previous equations and check that, under the substitution induced
by the match, other branches are surely apart, using `tcUnifyTysFG`. (See
Note [Apartness and type families] in GHC.Core.Unify.) This is similar to what
happens with class instance selection, when we need to guarantee that there is
only a match and no unifiers. The exact algorithm is different here because the
potentially-overlapping group is closed.

As another example, consider this:

type family G x where
  G Int = Bool
  G a   = Double

type family H y
no instances

Now, we want to simplify (G (H Char)). We can't, because (H Char) might later
simplify to be Int. So, (G (H Char)) is stuck, for now.

While everything above is quite sound, it isn't as expressive as we'd like.
Consider this:

type family J a where
  J Int = Int
  J a   = a

Can we simplify (J b) to b? Sure we can. Yes, the first equation matches if
b is instantiated with Int, but the RHSs coincide there, so it's all OK.

So, the rule is this: when looking up a branch in a closed type family, we
find a branch that matches the target, but then we make sure that the target
is apart from every previous *incompatible* branch. We don't check the
branches that are compatible with the matching branch, because they are either
irrelevant (clause 1 of compatible) or benign (clause 2 of compatible).

References 1

Referenced by 3