Simon Peyton Jones pushed to branch wip/andreask/spec_tyfams at Glasgow Haskell Compiler / GHC Commits: 483a939e by Simon Peyton Jones at 2025-04-26T22:51:58+01:00 Polishing up from Simon - - - - - 7 changed files: - compiler/GHC/Core/Opt/Specialise.hs - compiler/GHC/Core/Predicate.hs - compiler/GHC/Tc/Solver.hs - compiler/GHC/Tc/Solver/Dict.hs - compiler/GHC/Tc/Solver/InertSet.hs - compiler/GHC/Tc/Solver/Monad.hs - compiler/GHC/Tc/Utils/TcType.hs Changes: ===================================== compiler/GHC/Core/Opt/Specialise.hs ===================================== @@ -26,7 +26,7 @@ import GHC.Core.Make ( mkLitRubbish ) import GHC.Core.Unify ( tcMatchTy ) import GHC.Core.Rules import GHC.Core.Utils ( exprIsTrivial, exprIsTopLevelBindable - , mkCast, exprType + , mkCast, exprType, exprIsHNF , stripTicksTop, mkInScopeSetBndrs ) import GHC.Core.FVs import GHC.Core.TyCo.FVs ( tyCoVarsOfTypeList ) @@ -1646,10 +1646,10 @@ specCalls spec_imp env existing_rules calls_for_me fn rhs -- See Note [Inline specialisations] for why we do not -- switch off specialisation for inline functions - = -- pprTrace "specCalls: some" (vcat - -- [ text "function" <+> ppr fn - -- , text "calls:" <+> ppr calls_for_me - -- , text "subst" <+> ppr (se_subst env) ]) $ + = pprTrace "specCalls: some" (vcat + [ text "function" <+> ppr fn + , text "calls:" <+> ppr calls_for_me + , text "subst" <+> ppr (se_subst env) ]) $ foldlM spec_call ([], [], emptyUDs) calls_for_me | otherwise -- No calls or RHS doesn't fit our preconceptions @@ -1705,7 +1705,7 @@ specCalls spec_imp env existing_rules calls_for_me fn rhs , rule_bndrs, rule_lhs_args , spec_bndrs1, dx_binds, spec_args) <- specHeader env rhs_bndrs all_call_args - ; when False $ pprTrace "spec_call" (vcat + ; when True $ pprTrace "spec_call" (vcat [ text "fun: " <+> ppr fn , text "call info: " <+> ppr _ci , text "useful: " <+> ppr useful @@ -3034,7 +3034,7 @@ mkCallUDs' :: SpecEnv -> Id -> [OutExpr] -> UsageDetails mkCallUDs' env f args | wantCallsFor env f -- We want it, and... , not (null ci_key) -- this call site has a useful specialisation - = -- pprTrace "mkCallUDs: keeping" _trace_doc + = pprTrace "mkCallUDs: keeping" _trace_doc singleCall env f ci_key | otherwise -- See also Note [Specialisations already covered] @@ -3042,7 +3042,7 @@ mkCallUDs' env f args emptyUDs where - _trace_doc = vcat [ppr f, ppr args, ppr ci_key] + _trace_doc = vcat [ppr f, ppr args, ppr ci_key, ppr (se_subst env)] pis = fst $ splitPiTys $ idType f constrained_tyvars = tyCoVarsOfTypes $ getTheta pis @@ -3116,22 +3116,64 @@ There really is not much point in specialising f wrt the dictionary d, because the code for the specialised f is not improved at all, because d is lambda-bound. We simply get junk specialisations. -What is "interesting"? Just that it has *some* structure. But what about -variables? We look in the variable's /unfolding/. And that means -that we must be careful to ensure that dictionaries have unfoldings, - -* cloneBndrSM discards non-Stable unfoldings -* specBind updates the unfolding after specialisation - See Note [Update unfolding after specialisation] -* bindAuxiliaryDict adds an unfolding for an aux dict - see Note [Specialisation modulo dictionary selectors] -* specCase adds unfoldings for the new bindings it creates - -We accidentally lost accurate tracking of local variables for a long -time, because cloned variables didn't have unfoldings. But makes a -massive difference in a few cases, eg #5113. For nofib as a -whole it's only a small win: 2.2% improvement in allocation for ansi, -1.2% for bspt, but mostly 0.0! Average 0.1% increase in binary size. +What is "interesting"? Our Main Plan is to use `exprIsConApp_maybe` to see +if the argumeng is a dictionary constructor applied to some arguments, in which +case we can clearly specialise. But there are wrinkles: + +(ID1) Note that we look at the argument /term/, not its /type/. Suppose the + argument is + (% d1, d2 %) |> co + where co :: (% Eq [a], Show [a] %) ~ F Int a, and `F` is a type family. + Then its type (F Int a) looks very un-informative, but the term is super + helpful. See #19747 (where missing this point caused a 70x slow down) + and #7785. + +(ID2) Note that the Main Plan works fine for an argument that is a DFun call, + e.g. $fOrdList $dOrdInt + because `exprIsConApp_maybe` cleverly deals with DFunId applications. Good! + +(ID3) For variables, we look in the variable's /unfolding/. And that means + that we must be careful to ensure that dictionaries /have/ unfoldings: + * cloneBndrSM discards non-Stable unfoldings + * specBind updates the unfolding after specialisation + See Note [Update unfolding after specialisation] + * bindAuxiliaryDict adds an unfolding for an aux dict + see Note [Specialisation modulo dictionary selectors] + * specCase adds unfoldings for the new bindings it creates + + We accidentally lost accurate tracking of local variables for a long + time, because cloned variables didn't have unfoldings. But makes a + massive difference in a few cases, eg #5113. For nofib as a + whole it's only a small win: 2.2% improvement in allocation for ansi, + 1.2% for bspt, but mostly 0.0! Average 0.1% increase in binary size. + +(ID4) We must be very careful not to specialise on a "dictionry" that is, or contains + an implicit parameter, because implicit parameters are emphatically not singleton + types. See #25999: + useImplicit :: (?i :: Int) => Int + useImplicit = ?i + 1 + + foo = let ?i = 1 in (useImplicit, let ?i = 2 in useImplicit) + Both calls to `useImplicit` are at type `?i::Int`, but they pass different values. + We must not specialise on implicit parameters! Hence the call to `couldBeIPLike`. + +(ID5) Suppose the argument is (e |> co). Can we rely on `exprIsConApp_maybe` to deal + with the coercion. No! That only works if (co :: C t1 ~ C t2) with the same type + constructor at the top of both sides. But see the example in (ID1), where that + is not true. For thes same reason, we can't rely on `exprIsConApp_maybe` to look + through unfoldings (because there might be a cast inside), hence dealing with + expandable unfoldings in `interestingDict` directly. + +(ID6) The Main Plan says that it's worth specialising if the argument is an application + of a dictionary contructor. But what if the dictionary has no methods? Then we + gain nothing by specialising, unless the /superclasses/ are interesting. A case + in point is constraint tuples (% d1, .., dn %); a constraint N-tuple is a class + with N superclasses and no methods. + +(ID7) A unary (single-method) class is currently represented by (meth |> co). + We will unwrap the cast (see (ID5)) and then want to reply "yes" if the method + has any struture. We use `exprIsHNF` for this. (We plan a new story for unary + classes, see #23109, and this special case will become irrelevant.) Note [Update unfolding after specialisation] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ @@ -3159,6 +3201,7 @@ Consider (#21848) Now `f` turns into: f @a @b (dd :: D a) (ds :: Show b) a b + = let dc :: D a = %p1 dd -- Superclass selection in meth @a dc .... meth @a dc .... @@ -3174,50 +3217,25 @@ in the NonRec case of specBind. (This is too exotic to trouble with the Rec case.) -} --- interestingDict :: CoreExpr -> Type -> Bool interestingDict :: InScopeEnv -> CoreExpr -> Bool --- A dictionary argument is interesting if it has *some* structure, --- see Note [Interesting dictionary arguments] --- NB: "dictionary" arguments include constraints of all sorts, --- including equality constraints; hence the Coercion case --- To make this work, we need to ensure that dictionaries have --- unfoldings in them. -interestingDict env (Cast arg _) +-- See Note [Interesting dictionary arguments] +interestingDict env (Var v) -- See (ID3) and (ID5) + | Just rhs <- expandUnfolding_maybe (idUnfolding v) + = interestingDict env rhs +interestingDict env (Cast arg _) -- See (ID5) = interestingDict env arg -interestingDict env arg +interestingDict env arg -- Main Plan: use exprIsConApp_maybe | Just (_, _, data_con, _tys, args) <- exprIsConApp_maybe env arg , Just cls <- tyConClass_maybe (dataConTyCon data_con) - , (not . isIPLikePred) (exprType arg) - = if isCTupleClass cls + , (not . couldBeIPLike) (exprType arg) -- See (ID4) + = if null (classMethods cls) -- See (ID6) then any (interestingDict env) args else True + | exprIsHNF arg -- See (ID7) + = True | otherwise = False - --- interestingDict arg _arg_ty --- -- No benefit to specalizing for a ~# b I believe --- -- | (isEqPred arg_ty) = False --- -- | --- -- not (typeDeterminesValue arg_ty) = False -- See Note [Type determines value] --- | otherwise = go arg --- where --- go (Var v) = hasSomeUnfolding (idUnfolding v) --- || isDataConWorkId v --- go (Type _) = False --- go (Coercion _) = False --- go (App fn (Type _)) = go fn --- go (App fn (Coercion _)) = go fn --- go (Tick _ a) = go a --- go (Cast e _) = go e --- go (Lit{}) = True --- go (Case{}) = True --- go (Let{}) = True --- go (App{}) = True --- go (Lam{}) = True - - -- go _ = True - thenUDs :: UsageDetails -> UsageDetails -> UsageDetails thenUDs (MkUD {ud_binds = db1, ud_calls = calls1}) (MkUD {ud_binds = db2, ud_calls = calls2}) ===================================== compiler/GHC/Core/Predicate.hs ===================================== @@ -24,7 +24,7 @@ module GHC.Core.Predicate ( classMethodTy, classMethodInstTy, -- Implicit parameters - isIPLikePred, mentionsIP, isIPTyCon, isIPClass, + couldBeIPLike, mightMentionIP, isIPTyCon, isIPClass, isCallStackTy, isCallStackPred, isCallStackPredTy, isExceptionContextPred, isExceptionContextTy, isIPPred_maybe, @@ -127,7 +127,7 @@ isDictTy ty = isClassPred pred typeDeterminesValue :: Type -> Bool -- See Note [Type determines value] -typeDeterminesValue ty = isDictTy ty && not (isIPLikePred ty) +typeDeterminesValue ty = isDictTy ty && not (couldBeIPLike ty) getClassPredTys :: HasDebugCallStack => PredType -> (Class, [Type]) getClassPredTys ty = case getClassPredTys_maybe ty of @@ -424,44 +424,44 @@ isCallStackTy ty | otherwise = False --- --------------------- isIPLike and mentionsIP -------------------------- +-- --------------------- couldBeIPLike and mightMentionIP -------------------------- -- See Note [Local implicit parameters] -isIPLikePred :: Type -> Bool +couldBeIPLike :: Type -> Bool -- Is `pred`, or any of its superclasses, an implicit parameter? -- See Note [Local implicit parameters] -isIPLikePred pred = - mentions_ip_pred initIPRecTc (const True) (const True) pred - -mentionsIP :: (Type -> Bool) -- ^ predicate on the string - -> (Type -> Bool) -- ^ predicate on the type - -> Class - -> [Type] -> Bool --- ^ @'mentionsIP' str_cond ty_cond cls tys@ returns @True@ if: +couldBeIPLike pred + = might_mention_ip1 initIPRecTc (const True) (const True) pred + +mightMentionIP :: (Type -> Bool) -- ^ predicate on the string + -> (Type -> Bool) -- ^ predicate on the type + -> Class + -> [Type] -> Bool +-- ^ @'mightMentionIP' str_cond ty_cond cls tys@ returns @True@ if: -- -- - @cls tys@ is of the form @IP str ty@, where @str_cond str@ and @ty_cond ty@ -- are both @True@, -- - or any superclass of @cls tys@ has this property. -- -- See Note [Local implicit parameters] -mentionsIP = mentions_ip initIPRecTc +mightMentionIP = might_mention_ip initIPRecTc -mentions_ip :: RecTcChecker -> (Type -> Bool) -> (Type -> Bool) -> Class -> [Type] -> Bool -mentions_ip rec_clss str_cond ty_cond cls tys +might_mention_ip :: RecTcChecker -> (Type -> Bool) -> (Type -> Bool) -> Class -> [Type] -> Bool +might_mention_ip rec_clss str_cond ty_cond cls tys | Just (str_ty, ty) <- isIPPred_maybe cls tys = str_cond str_ty && ty_cond ty | otherwise - = or [ mentions_ip_pred rec_clss str_cond ty_cond (classMethodInstTy sc_sel_id tys) + = or [ might_mention_ip1 rec_clss str_cond ty_cond (classMethodInstTy sc_sel_id tys) | sc_sel_id <- classSCSelIds cls ] -mentions_ip_pred :: RecTcChecker -> (Type -> Bool) -> (Type -> Bool) -> Type -> Bool -mentions_ip_pred rec_clss str_cond ty_cond ty +might_mention_ip1 :: RecTcChecker -> (Type -> Bool) -> (Type -> Bool) -> Type -> Bool +might_mention_ip1 rec_clss str_cond ty_cond ty | Just (cls, tys) <- getClassPredTys_maybe ty , let tc = classTyCon cls , Just rec_clss' <- if isTupleTyCon tc then Just rec_clss else checkRecTc rec_clss tc - = mentions_ip rec_clss' str_cond ty_cond cls tys + = might_mention_ip rec_clss' str_cond ty_cond cls tys | otherwise = False -- Includes things like (D []) where D is -- a Constraint-ranged family; #7785 @@ -474,7 +474,7 @@ initIPRecTc = setRecTcMaxBound 1 initRecTc See also wrinkle (SIP1) in Note [Shadowing of implicit parameters] in GHC.Tc.Solver.Dict. -The function isIPLikePred tells if this predicate, or any of its +The function couldBeIPLike tells if this predicate, or any of its superclasses, is an implicit parameter. Why are implicit parameters special? Unlike normal classes, we can @@ -482,7 +482,7 @@ have local instances for implicit parameters, in the form of let ?x = True in ... So in various places we must be careful not to assume that any value of the right type will do; we must carefully look for the innermost binding. -So isIPLikePred checks whether this is an implicit parameter, or has +So couldBeIPLike checks whether this is an implicit parameter, or has a superclass that is an implicit parameter. Several wrinkles @@ -523,16 +523,16 @@ Small worries (Sept 20): think nothing does. * I'm a little concerned about type variables; such a variable might be instantiated to an implicit parameter. I don't think this - matters in the cases for which isIPLikePred is used, and it's pretty + matters in the cases for which couldBeIPLike is used, and it's pretty obscure anyway. * The superclass hunt stops when it encounters the same class again, but in principle we could have the same class, differently instantiated, and the second time it could have an implicit parameter I'm going to treat these as problems for another day. They are all exotic. -Note [Using typesAreApart when calling mentionsIP] -~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -We call 'mentionsIP' in two situations: +Note [Using typesAreApart when calling mightMentionIP] +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +We call 'mightMentionIP' in two situations: (1) to check that a predicate does not contain any implicit parameters IP str ty, for a fixed literal str and any type ty, ===================================== compiler/GHC/Tc/Solver.hs ===================================== @@ -1902,7 +1902,7 @@ growThetaTyVars theta tcvs | otherwise = transCloVarSet mk_next seed_tcvs where seed_tcvs = tcvs `unionVarSet` tyCoVarsOfTypes ips - (ips, non_ips) = partition isIPLikePred theta + (ips, non_ips) = partition couldBeIPLike theta -- See Note [Inheriting implicit parameters] mk_next :: VarSet -> VarSet -- Maps current set to newly-grown ones ===================================== compiler/GHC/Tc/Solver/Dict.hs ===================================== @@ -749,7 +749,7 @@ shortCutSolver dflags ev_w ev_i -- programs should typecheck regardless of whether we take this step or -- not. See Note [Shortcut solving] - , not (isIPLikePred (ctEvPred ev_w)) -- Not for implicit parameters (#18627) + , not (couldBeIPLike (ctEvPred ev_w)) -- Not for implicit parameters (#18627) , not (xopt LangExt.IncoherentInstances dflags) -- If IncoherentInstances is on then we cannot rely on coherence of proofs ===================================== compiler/GHC/Tc/Solver/InertSet.hs ===================================== @@ -2013,10 +2013,10 @@ solveOneFromTheOther ct_i ct_w is_wsc_orig_w = isWantedSuperclassOrigin orig_w different_level_strategy -- Both Given - | isIPLikePred pred = if lvl_w `strictlyDeeperThan` lvl_i then KeepWork else KeepInert - | otherwise = if lvl_w `strictlyDeeperThan` lvl_i then KeepInert else KeepWork + | couldBeIPLike pred = if lvl_w `strictlyDeeperThan` lvl_i then KeepWork else KeepInert + | otherwise = if lvl_w `strictlyDeeperThan` lvl_i then KeepInert else KeepWork -- See Note [Replacement vs keeping] part (1) - -- For the isIPLikePred case see Note [Shadowing of implicit parameters] + -- For the couldBeIPLike case see Note [Shadowing of implicit parameters] -- in GHC.Tc.Solver.Dict same_level_strategy -- Both Given ===================================== compiler/GHC/Tc/Solver/Monad.hs ===================================== @@ -396,8 +396,8 @@ updInertDicts dict_ct@(DictCt { di_cls = cls, di_ev = ev, di_tys = tys }) -- an implicit parameter (?str :: ty) for the given 'str' and any type 'ty'? does_not_mention_ip_for :: Type -> DictCt -> Bool does_not_mention_ip_for str_ty (DictCt { di_cls = cls, di_tys = tys }) - = not $ mentionsIP (not . typesAreApart str_ty) (const True) cls tys - -- See Note [Using typesAreApart when calling mentionsIP] + = not $ mightMentionIP (not . typesAreApart str_ty) (const True) cls tys + -- See Note [Using typesAreApart when calling mightMmentionIP] -- in GHC.Core.Predicate updInertIrreds :: IrredCt -> TcS () @@ -533,7 +533,7 @@ updSolvedDicts what dict_ct@(DictCt { di_cls = cls, di_tys = tys, di_ev = ev }) = do { is_callstack <- is_tyConTy isCallStackTy callStackTyConName ; is_exceptionCtx <- is_tyConTy isExceptionContextTy exceptionContextTyConName ; let contains_callstack_or_exceptionCtx = - mentionsIP + mightMentionIP (const True) -- NB: the name of the call-stack IP is irrelevant -- e.g (?foo :: CallStack) counts! @@ -551,9 +551,9 @@ updSolvedDicts what dict_ct@(DictCt { di_cls = cls, di_tys = tys, di_ev = ev }) -- Return a predicate that decides whether a type is CallStack -- or ExceptionContext, accounting for e.g. type family reduction, as - -- per Note [Using typesAreApart when calling mentionsIP]. + -- per Note [Using typesAreApart when calling mightMentionIP]. -- - -- See Note [Using isCallStackTy in mentionsIP]. + -- See Note [Using isCallStackTy in mightMentionIP]. is_tyConTy :: (Type -> Bool) -> Name -> TcS (Type -> Bool) is_tyConTy is_eq tc_name = do { (mb_tc, _) <- wrapTcS $ TcM.tryTc $ TcM.tcLookupTyCon tc_name @@ -581,14 +581,14 @@ in a different context! See also Note [Shadowing of implicit parameters], which deals with a similar problem with Given implicit parameter constraints. -Note [Using isCallStackTy in mentionsIP] -~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +Note [Using isCallStackTy in mightMentionIP] +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ To implement Note [Don't add HasCallStack constraints to the solved set], we need to check whether a constraint contains a HasCallStack or HasExceptionContext constraint. We do this using the 'mentionsIP' function, but as per -Note [Using typesAreApart when calling mentionsIP] we don't want to simply do: +Note [Using typesAreApart when calling mightMentions] we don't want to simply do: - mentionsIP + mightMentionIP (const True) -- (ignore the implicit parameter string) (isCallStackTy <||> isExceptionContextTy) ===================================== compiler/GHC/Tc/Utils/TcType.hs ===================================== @@ -155,7 +155,7 @@ module GHC.Tc.Utils.TcType ( mkTyConTy, mkTyVarTy, mkTyVarTys, mkTyCoVarTy, mkTyCoVarTys, - isClassPred, isEqPred, isIPLikePred, isEqClassPred, + isClassPred, isEqPred, couldBeIPLike, isEqClassPred, isEqualityClass, mkClassPred, tcSplitQuantPredTy, tcSplitDFunTy, tcSplitDFunHead, tcSplitMethodTy, isRuntimeRepVar, isFixedRuntimeRepKind, @@ -1819,7 +1819,7 @@ pickCapturedPreds pickCapturedPreds qtvs theta = filter captured theta where - captured pred = isIPLikePred pred || (tyCoVarsOfType pred `intersectsVarSet` qtvs) + captured pred = couldBeIPLike pred || (tyCoVarsOfType pred `intersectsVarSet` qtvs) -- Superclasses View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/483a939e118d1472ce32ac95b43e6d5f... -- View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/483a939e118d1472ce32ac95b43e6d5f... You're receiving this email because of your account on gitlab.haskell.org.