Zonk the returned kind in tcFamTyPats

The motivation is given in Note [tcFamTyPats: zonking the result kind].

Fixes #19250  --  the fix is easy.

1 files changed, 31 insertions, 0 deletions

diff --git a/compiler/GHC/Tc/Gen/HsType.hs b/compiler/GHC/Tc/Gen/HsType.hs
index 0a188d9020..39b18e063c 100644
--- a/compiler/GHC/Tc/Gen/HsType.hs
+++ b/compiler/GHC/Tc/Gen/HsType.hs
@@ -680,6 +680,9 @@ tcFamTyPats fam_tc hs_pats
        ; let fun_ty = mkTyConApp fam_tc []
        ; (fam_app, res_kind) <- tcInferTyApps mode lhs_fun fun_ty hs_pats
 
+       -- Hack alert: see Note [tcFamTyPats: zonking the result kind]
+       ; res_kind <- zonkTcType res_kind
+
        ; traceTc "End tcFamTyPats }" $
          vcat [ ppr fam_tc, text "res_kind:" <+> ppr res_kind ]
 
@@ -689,6 +692,34 @@ tcFamTyPats fam_tc hs_pats
     fam_arity = tyConArity fam_tc
     lhs_fun   = noLoc (HsTyVar noExtField NotPromoted (noLoc fam_name))
 
+{- Note [tcFamTyPats: zonking the result kind]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Consider (#19250)
+    F :: forall k. k -> k
+    type instance F (x :: Constraint) = ()
+
+The tricky point is this:
+  is that () an empty type tuple (() :: Type), or
+  an empty constraint tuple (() :: Constraint)?
+We work this out in a hacky way, by looking at the expected kind:
+see Note [Inferring tuple kinds].
+
+In this case, we kind-check the RHS using the kind gotten from the LHS:
+see the call to tcCheckLHsType in tcTyFamInstEqnGuts in GHC.Tc.Tycl.
+
+But we want the kind from the LHS to be /zonked/, so that when
+kind-checking the RHS (tcCheckLHsType) we can "see" what we learned
+from kind-checking the LHS (tcFamTyPats).  In our example above, the
+type of the LHS is just `kappa` (by instantiating the forall k), but
+then we learn (from x::Constraint) that kappa ~ Constraint.  We want
+that info when kind-checking the RHS.
+
+Easy solution: just zonk that return kind.  Of course this won't help
+if there is lots of type-family reduction to do, but it works fine in
+common cases.
+-}
+
+
 {-
 ************************************************************************
 *                                                                      *