Fix a nasty recursive loop in typechecking interface files

compiler/iface/BuildTyCl.lhs

@@ -115,15 +115,15 @@ mkNewTyConRhs :: Name -> TyCon -> DataCon -> TcRnIf m n AlgTyConRhs
 -- because the latter is part of a knot, whereas the former is not.
 mkNewTyConRhs tycon_name tycon con 
   = do	{ co_tycon_name <- newImplicitBinder tycon_name mkNewTyCoOcc
-	; let co_tycon = mkNewTypeCoercion co_tycon_name tycon etad_rhs
+	; let co_tycon = mkNewTypeCoercion co_tycon_name tycon etad_tvs etad_rhs
               cocon_maybe | all_coercions || isRecursiveTyCon tycon 
 		          = Just co_tycon
                 	  | otherwise              
                 	  = Nothing
 	; return (NewTyCon { data_con    = con, 
 		       	     nt_rhs      = rhs_ty,
-		       	     nt_etad_rhs = etad_rhs,
- 		       	     nt_co = cocon_maybe, 
+		       	     nt_etad_rhs = (etad_tvs, etad_rhs),
+ 		       	     nt_co 	 = cocon_maybe, 
                              -- Coreview looks through newtypes with a Nothing
                              -- for nt_co, or uses explicit coercions otherwise
 		       	     nt_rep = mkNewTyConRep tycon rhs_ty }) }
@@ -137,8 +137,10 @@ mkNewTyConRhs tycon_name tycon con
 	-- Instantiate the data con with the 
 	-- type variables from the tycon
 
-    etad_rhs :: ([TyVar], Type)
-    etad_rhs = eta_reduce (reverse tvs) rhs_ty
+    etad_tvs :: [TyVar]	-- Matched lazily, so that mkNewTypeCoercion can
+    etad_rhs :: Type	-- return a TyCon without pulling on rhs_ty
+			-- See Note [Tricky iface loop] in LoadIface
+    (etad_tvs, etad_rhs) = eta_reduce (reverse tvs) rhs_ty
 
     eta_reduce :: [TyVar]		-- Reversed
 	       -> Type			-- Rhs type
@@ -300,7 +302,7 @@ buildClass class_name tvs sc_theta fds ats sig_stuff tc_isrec
 	; let {	clas_kind = mkArrowKinds (map tyVarKind tvs) liftedTypeKind
 
  	      ; tycon = mkClassTyCon tycon_name clas_kind tvs
-                             rhs rec_clas tc_isrec
+ 	                             rhs rec_clas tc_isrec
 		-- A class can be recursive, and in the case of newtypes 
 		-- this matters.  For example
 		-- 	class C a where { op :: C b => a -> b -> Int }

compiler/iface/LoadIface.lhs

@@ -323,7 +323,27 @@ loadDecl ignore_prags mod (_version, decl)
 	; thing <- forkM doc $ do { bumpDeclStats main_name
 				  ; tcIfaceDecl ignore_prags decl }
 
-	-- Populate the type environment with the implicitTyThings too
+	-- Populate the type environment with the implicitTyThings too.
+	-- 
+	-- Note [Tricky iface loop]
+	-- ~~~~~~~~~~~~~~~~~~~~~~~~
+	-- The delicate point here is that 'mini-env' should be
+	-- buildable from 'thing' without demanding any of the things 'forkM'd 
+	-- by tcIfaceDecl.  For example
+	--	class C a where { data T a; op :: T a -> Int }
+	-- We return the bindings
+	--	[("C", <cls>), ("T", lookup env "T"), ("op", lookup env "op")]
+	-- The call (lookup env "T") must return the tycon T without first demanding
+	-- op; because getting the latter will look up T, hence loop.
+	--
+	-- Of course, there is no reason in principle why (lookup env "T") should demand
+	-- anything do to with op, but take care: 
+	--	(a) implicitTyThings, and 
+	--	(b) getOccName of all the things returned by implicitThings, 
+	-- must not depend on any of the nested type-checks
+	-- 
+	-- All a bit too finely-balanced for my liking.
+
 	; let mini_env = mkOccEnv [(getOccName t, t) | t <- implicitTyThings thing]
 	      lookup n = case lookupOccEnv mini_env (getOccName n) of
 			   Just thing -> thing

compiler/types/Coercion.lhs

@@ -275,8 +275,8 @@ mkUnsafeCoercion ty1 ty2
 
 
 -- See note [Newtype coercions] in TyCon
-mkNewTypeCoercion :: Name -> TyCon -> ([TyVar], Type) -> TyCon
-mkNewTypeCoercion name tycon (tvs, rhs_ty)
+mkNewTypeCoercion :: Name -> TyCon -> [TyVar] -> Type -> TyCon
+mkNewTypeCoercion name tycon tvs rhs_ty
   = mkCoercionTyCon name co_con_arity rule
   where
     co_con_arity = length tvs