[Git][ghc/ghc][master] Make type-equality on synonyms a bit faster

Marge Bot (@marge-bot) gitlab at gitlab.haskell.org
Fri Jul 12 15:44:51 UTC 2024



Marge Bot pushed to branch master at Glasgow Haskell Compiler / GHC


Commits:
280e4bf5 by Simon Peyton Jones at 2024-07-12T11:43:59-04:00
Make type-equality on synonyms a bit faster

This MR make equality fast for (S tys1 `eqType` S tys2),
where S is a non-forgetful type synonym.

It doesn't affect compile-time allocation much, but then comparison doesn't
allocate anyway.  But it seems like a Good Thing anyway.

See Note [Comparing type synonyms] in GHC.Core.TyCo.Compare
and Note [Forgetful type synonyms] in GHC.Core.TyCon

Addresses #25009.

- - - - -


7 changed files:

- compiler/GHC/Core/Map/Type.hs
- compiler/GHC/Core/TyCo/Compare.hs
- compiler/GHC/Core/TyCon.hs
- compiler/GHC/Core/Type.hs
- compiler/GHC/Core/Unify.hs
- compiler/GHC/Tc/Solver/Equality.hs
- compiler/GHC/Tc/Utils/Unify.hs


Changes:

=====================================
compiler/GHC/Core/Map/Type.hs
=====================================
@@ -38,6 +38,7 @@ import GHC.Prelude
 import GHC.Core.Type
 import GHC.Core.Coercion
 import GHC.Core.TyCo.Rep
+import GHC.Core.TyCon( isForgetfulSynTyCon )
 import GHC.Core.TyCo.Compare( eqForAllVis )
 import GHC.Data.TrieMap
 
@@ -228,10 +229,11 @@ eqDeBruijnType env_t1@(D env1 t1) env_t2@(D env2 t2) =
     andEq TEQX e = hasCast e
     andEq TEQ  e = e
 
-    -- See Note [Comparing nullary type synonyms] in GHC.Core.TyCo.Compare
-    go (D _ (TyConApp tc1 [])) (D _ (TyConApp tc2 []))
-      | tc1 == tc2
-      = TEQ
+    -- See Note [Comparing type synonyms] in GHC.Core.TyCo.Compare
+    go (D env1 (TyConApp tc1 tys1)) (D env2 (TyConApp tc2 tys2))
+      | tc1 == tc2, not (isForgetfulSynTyCon tc1)
+      = gos env1 env2 tys1 tys2
+
     go env_t@(D env t) env_t'@(D env' t')
       | Just new_t  <- coreView t  = go (D env new_t) env_t'
       | Just new_t' <- coreView t' = go env_t (D env' new_t')


=====================================
compiler/GHC/Core/TyCo/Compare.hs
=====================================
@@ -138,35 +138,52 @@ But the left is an AppTy while the right is a TyConApp. The solution is
 to use splitAppTyNoView_maybe to break up the TyConApp into its pieces and
 then continue. Easy to do, but also easy to forget to do.
 
-Note [Comparing nullary type synonyms]
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Note [Comparing type synonyms]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 Consider the task of testing equality between two 'Type's of the form
 
-  TyConApp tc []
+  TyConApp tc tys1  =  TyConApp tc tys2
 
-where @tc@ is a type synonym. A naive way to perform this comparison these
+where `tc` is a type synonym. A naive way to perform this comparison these
 would first expand the synonym and then compare the resulting expansions.
 
-However, this is obviously wasteful and the RHS of @tc@ may be large; it is
-much better to rather compare the TyCons directly. Consequently, before
-expanding type synonyms in type comparisons we first look for a nullary
-TyConApp and simply compare the TyCons if we find one. Of course, if we find
-that the TyCons are *not* equal then we still need to perform the expansion as
-their RHSs may still be equal.
+However, this is obviously wasteful and the RHS of `tc` may be large. We'd
+prefer to compare `tys1 = tys2`.  When is that sound?  Precisely when the
+synonym is not /forgetful/; that is, all its type variables appear in its
+RHS -- see `GHC.Core.TyCon.isForgetfulSynTyCon`.
+
+Of course, if we find that the TyCons are *not* equal then we still need to
+perform the expansion as their RHSs may still be equal.
+
+This works fine for /equality/, but not for /comparison/.  Consider
+   type S a b = (b, a)
+Now consider
+   S Int Bool `compare` S Char Char
+The ordering may depend on whether we expand the synonym or not, and we
+don't want the result to depend on that. So for comparison we stick to
+/nullary/ synonyms only, which is still useful.
 
 We perform this optimisation in a number of places:
 
- * GHC.Core.Types.eqType
- * GHC.Core.Types.nonDetCmpType
- * GHC.Core.Unify.unify_ty
- * GHC.Tc.Solver.Equality.can_eq_nc'
- * TcUnify.uType
+ * GHC.Core.TyCo.Compare.eqType      (works for non-nullary synonyms)
+ * GHC.Core.Map.TYpe.eqDeBruijnType  (works for non-nullary synonyms)
+ * GHC.Core.Types.nonDetCmpType      (nullary only)
 
 This optimisation is especially helpful for the ubiquitous GHC.Types.Type,
 since GHC prefers to use the type synonym over @TYPE 'LiftedRep@ applications
 whenever possible. See Note [Using synonyms to compress types] in
 GHC.Core.Type for details.
 
+Currently-missed opportunity (#25009):
+* In the case of forgetful synonyms, we could still compare the args, pairwise,
+  and then compare the RHS's with a suitably extended RnEnv2.  That would avoid
+  comparing the same arg repeatedly.  e.g.
+      type S a b = (a,a)
+  Compare   S <big> y ~ S <big> y
+  If we expand, we end up compare <big> with itself twice.
+
+  But since forgetful synonyms are rare, we have not tried this.
+
 Note [Type comparisons using object pointer comparisons]
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 Quite often we substitute the type from a definition site into
@@ -341,16 +358,26 @@ inline_generic_eq_type_x syn_flag mult_flag mb_env
   = \ t1 t2 -> t1 `seq` t2 `seq`
     let go = generic_eq_type_x syn_flag mult_flag mb_env
              -- Abbreviation for recursive calls
+
+        gos []       []       = True
+        gos (t1:ts1) (t2:ts2) = go t1 t2 && gos ts1 ts2
+        gos _ _               = False
+
     in case (t1,t2) of
       _ | 1# <- reallyUnsafePtrEquality# t1 t2 -> True
       -- See Note [Type comparisons using object pointer comparisons]
 
-      (TyConApp tc1 [], TyConApp tc2 []) | tc1 == tc2 -> True
-      -- See Note [Comparing nullary type synonyms]
+      (TyConApp tc1 tys1, TyConApp tc2 tys2)
+        | tc1 == tc2, not (isForgetfulSynTyCon tc1)   -- See Note [Comparing type synonyms]
+        -> gos tys1 tys2
 
       _ | ExpandSynonyms <- syn_flag, Just t1' <- coreView t1 -> go t1' t2
         | ExpandSynonyms <- syn_flag, Just t2' <- coreView t2 -> go t1 t2'
 
+      (TyConApp tc1 ts1, TyConApp tc2 ts2)
+        | tc1 == tc2 -> gos ts1 ts2
+        | otherwise  -> False
+
       (TyVarTy tv1, TyVarTy tv2)
         -> case mb_env of
               Nothing  -> tv1 == tv2
@@ -381,14 +408,6 @@ inline_generic_eq_type_x syn_flag mult_flag mb_env
         | Just (s1, t1') <- tcSplitAppTyNoView_maybe t1
         -> go s1 s2 && go t1' t2'
 
-      (TyConApp tc1 ts1, TyConApp tc2 ts2)
-        | tc1 == tc2 -> gos ts1 ts2
-        | otherwise  -> False
-        where
-          gos []       []       = True
-          gos (t1:ts1) (t2:ts2) = go t1 t2 && gos ts1 ts2
-          gos _ _               = False
-
       (ForAllTy (Bndr tv1 vis1) body1, ForAllTy (Bndr tv2 vis2) body2)
         -> case mb_env of
               Nothing -> generic_eq_type_x syn_flag mult_flag
@@ -666,10 +685,11 @@ nonDetCmpTypeX env orig_t1 orig_t2 =
     -- Returns both the resulting ordering relation between
     -- the two types and whether either contains a cast.
     go :: RnEnv2 -> Type -> Type -> TypeOrdering
-    -- See Note [Comparing nullary type synonyms]
+
     go _   (TyConApp tc1 []) (TyConApp tc2 [])
       | tc1 == tc2
-      = TEQ
+      = TEQ    -- See Note [Comparing type synonyms]
+
     go env t1 t2
       | Just t1' <- coreView t1 = go env t1' t2
       | Just t2' <- coreView t2 = go env t1 t2'
@@ -758,8 +778,10 @@ mayLookIdentical orig_ty1 orig_ty2
     orig_env = mkRnEnv2 $ mkInScopeSet $ tyCoVarsOfTypes [orig_ty1, orig_ty2]
 
     go :: RnEnv2 -> Type -> Type -> Bool
-    -- See Note [Comparing nullary type synonyms]
-    go _  (TyConApp tc1 []) (TyConApp tc2 []) | tc1 == tc2 = True
+
+    go env (TyConApp tc1 ts1) (TyConApp tc2 ts2)
+      | tc1 == tc2, not (isForgetfulSynTyCon tc1) -- See Note [Comparing type synonyms]
+      = gos env (tyConBinders tc1) ts1 ts2
 
     go env t1 t2 | Just t1' <- coreView t1 = go env t1' t2
     go env t1 t2 | Just t2' <- coreView t2 = go env t1 t2'


=====================================
compiler/GHC/Core/TyCon.hs
=====================================
@@ -841,7 +841,8 @@ data TyConDetails =
                                  --          are fine), again after expanding any
                                  --          nested synonyms
 
-        synIsForgetful :: Bool,  -- True <=  at least one argument is not mentioned
+        synIsForgetful :: Bool,  -- See Note [Forgetful type synonyms]
+                                 -- True <=  at least one argument is not mentioned
                                  --          in the RHS (or is mentioned only under
                                  --          forgetful synonyms)
                                  -- Test is conservative, so True does not guarantee
@@ -2121,11 +2122,43 @@ isFamFreeTyCon (TyCon { tyConDetails = details })
 -- True. Thus, False means that all bound variables appear on the RHS;
 -- True may not mean anything, as the test to set this flag is
 -- conservative.
+--
+-- See Note [Forgetful type synonyms]
 isForgetfulSynTyCon :: TyCon -> Bool
 isForgetfulSynTyCon (TyCon { tyConDetails = details })
   | SynonymTyCon { synIsForgetful = forget } <- details = forget
   | otherwise                                           = False
 
+{- Note [Forgetful type synonyms]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+A type synonyms is /forgetful/ if its RHS fails to mention one (or more) of its bound variables.
+
+Forgetfulness is conservative:
+  * A non-forgetful synonym /guarantees/ to mention all its bound variables in its RHS.
+  * It is always safe to classify a synonym as forgetful.
+
+Examples:
+    type R = Int             -- Not forgetful
+    type S a = Int           -- Forgetful
+    type T1 a = Int -> S a   -- Forgetful
+    type T2 a = a -> S a     -- Not forgetful
+    type T3 a = Int -> F a   -- Not forgetful
+      where type family F a
+
+* R shows that nullary synonyms are not forgetful.
+
+* T2 shows that forgetfulness needs to account for uses of forgetful
+  synonyms. `a` appears on the RHS, but only under a forgetful S
+
+* T3 shows that non-forgetfulness is not the same as injectivity. T3 mentions its
+  bound variable on its RHS, but under a type family.  So it is entirely possible
+  that    T3 Int ~ T3 Bool
+
+* Since type synonyms are non-recursive, we don't need a fixpoint analysis to
+  determine forgetfulness.  It's rather easy -- see `GHC.Core.Type.buildSynTyCon`,
+  which is a bit over-conservative for over-saturated synonyms.
+-}
+
 -- As for newtypes, it is in some contexts important to distinguish between
 -- closed synonyms and synonym families, as synonym families have no unique
 -- right hand side to which a synonym family application can expand.


=====================================
compiler/GHC/Core/Type.hs
=====================================
@@ -2327,7 +2327,7 @@ buildSynTyCon name binders res_kind roles rhs
     is_forgetful = not (all ((`elemVarSet` rhs_tyvars) . binderVar) binders) ||
                    uniqSetAny isForgetfulSynTyCon rhs_tycons
          -- NB: is_forgetful is allowed to be conservative, returning True more often
-         -- than it should. See comments on GHC.Core.TyCon.isForgetfulSynTyCon
+         -- than it should. See Note [Forgetful type synonyms] in GHC.Core.TyCon
 
     rhs_tycons = tyConsOfType   rhs
     rhs_tyvars = tyCoVarsOfType rhs
@@ -3275,8 +3275,9 @@ efficient. Specifically, we strive to
 Goal (b) is particularly useful as it makes traversals (e.g. free variable
 traversal, substitution, and comparison) more efficient.
 Comparison in particular takes special advantage of nullary type synonym
-applications (e.g. things like @TyConApp typeTyCon []@), Note [Comparing
-nullary type synonyms] in "GHC.Core.Type".
+applications (e.g. things like @TyConApp typeTyCon []@). See
+* Note [Comparing type synonyms] in "GHC.Core.TyCo.Compare"
+* Note [Unifying type synonyms] in "GHC.Core.Unify"
 
 To accomplish these we use a number of tricks, implemented by mkTyConApp.
 


=====================================
compiler/GHC/Core/Unify.hs
=====================================
@@ -1089,6 +1089,47 @@ of arity n:
   If we couldn't decompose in the previous step, we return SurelyApart.
 
 Afterwards, the rest of the code doesn't have to worry about type families.
+
+Note [Unifying type synonyms]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Consider the task of unifying two 'Type's of the form
+
+  TyConApp tc [] ~ TyConApp tc []
+
+where `tc` is a type synonym. A naive way to perform this comparison these
+would first expand the synonym and then compare the resulting expansions.
+
+However, this is obviously wasteful and the RHS of `tc` may be large; it is
+much better to rather compare the TyCons directly. Consequently, before
+expanding type synonyms in type comparisons we first look for a nullary
+TyConApp and simply compare the TyCons if we find one.
+
+Of course, if we find that the TyCons are *not* equal then we still need to
+perform the expansion as their RHSs may still be unifiable.  E.g
+    type T = S (a->a)
+    type S a = [a]
+and consider
+    T Int ~ S (Int -> Int)
+
+We can't decompose non-nullary synonyms.  E.g.
+    type R a = F a    -- Where F is a type family
+and consider
+    R (a->a) ~ R Int
+We can't conclude that  (a->) ~ Int.  (There is a currently-missed opportunity
+here; if we knew that R was /injective/, perhaps we could decompose.)
+
+We perform the nullary-type-synonym optimisation in a number of places:
+
+ * GHC.Core.Unify.unify_ty
+ * GHC.Tc.Solver.Equality.can_eq_nc'
+ * GHC.Tc.Utils.Unify.uType
+
+This optimisation is especially helpful for the ubiquitous GHC.Types.Type,
+since GHC prefers to use the type synonym over @TYPE 'LiftedRep@ applications
+whenever possible. See Note [Using synonyms to compress types] in
+GHC.Core.Type for details.
+
+c.f. Note [Comparing type synonyms] in GHC.Core.TyCo.Compare
 -}
 
 -------------- unify_ty: the main workhorse -----------
@@ -1107,7 +1148,7 @@ unify_ty :: UMEnv
 -- Respects newtypes, PredTypes
 -- See Note [Computing equality on types] in GHC.Core.Type
 unify_ty _env (TyConApp tc1 []) (TyConApp tc2 []) _kco
-  -- See Note [Comparing nullary type synonyms] in GHC.Core.TyCo.Compare
+  -- See Note [Unifying type synonyms]
   | tc1 == tc2
   = return ()
 


=====================================
compiler/GHC/Tc/Solver/Equality.hs
=====================================
@@ -305,7 +305,7 @@ can_eq_nc
    -> Type -> Type    -- RHS, after and before type-synonym expansion, resp
    -> TcS (StopOrContinue (Either IrredCt EqCt))
 
--- See Note [Comparing nullary type synonyms] in GHC.Core.TyCo.Compare
+-- See Note [Unifying type synonyms] in GHC.Core.Unify
 can_eq_nc _flat _rdr_env _envs ev eq_rel ty1@(TyConApp tc1 []) _ps_ty1 (TyConApp tc2 []) _ps_ty2
   | tc1 == tc2
   = canEqReflexive ev eq_rel ty1


=====================================
compiler/GHC/Tc/Utils/Unify.hs
=====================================
@@ -2187,11 +2187,12 @@ uType env@(UE { u_role = role }) orig_ty1 orig_ty2
                               ; uType env orig_ty1 ty2 }
                Nothing -> uUnfilledVar env IsSwapped tv2 ty1 }
 
-      -- See Note [Expanding synonyms during unification]
+      -- See Note [Unifying type synonyms] in GHC.Core.Unify
     go ty1@(TyConApp tc1 []) (TyConApp tc2 [])
       | tc1 == tc2
       = return $ mkReflCo role ty1
 
+        -- Now expand synonyms
         -- See Note [Expanding synonyms during unification]
         --
         -- Also NB that we recurse to 'go' so that we don't push a
@@ -2349,7 +2350,7 @@ We expand synonyms during unification, but:
  * The problem case immediately above can happen only with arguments
    to the tycon. So we check for nullary tycons *before* expanding.
    This is particularly helpful when checking (* ~ *), because * is
-   now a type synonym.
+   now a type synonym.  See Note [Unifying type synonyms] in GHC.Core.Unify.
 
 Note [Deferred unification]
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~



View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/280e4bf5ca62ec51eaeedd04b0db78b085257ab8

-- 
This project does not include diff previews in email notifications.
View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/280e4bf5ca62ec51eaeedd04b0db78b085257ab8
You're receiving this email because of your account on gitlab.haskell.org.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-commits/attachments/20240712/0e461528/attachment-0001.html>


More information about the ghc-commits mailing list