X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Fcompiler%2FbasicTypes%2FMkId.lhs;h=84b3546e622f1b1f4421b6bc544caaca89a89b4a;hb=28a464a75e14cece5db40f2765a29348273ff2d2;hp=c8b00b7ca70b21389cc7253c2831468715ec7863;hpb=fa654d6b16ecda7cc8cb780792ca10ec0e227555;p=ghc-hetmet.git diff --git a/ghc/compiler/basicTypes/MkId.lhs b/ghc/compiler/basicTypes/MkId.lhs index c8b00b7..84b3546 100644 --- a/ghc/compiler/basicTypes/MkId.lhs +++ b/ghc/compiler/basicTypes/MkId.lhs @@ -16,7 +16,7 @@ module MkId ( mkDictFunId, mkDefaultMethodId, mkDictSelId, - mkDataConId, mkDataConWrapId, + mkDataConIds, mkRecordSelId, mkPrimOpId, mkFCallId, @@ -30,72 +30,68 @@ module MkId ( mkRuntimeErrorApp, rEC_CON_ERROR_ID, iRREFUT_PAT_ERROR_ID, rUNTIME_ERROR_ID, nON_EXHAUSTIVE_GUARDS_ERROR_ID, nO_METHOD_BINDING_ERROR_ID, - pAT_ERROR_ID + pAT_ERROR_ID, eRROR_ID, + + unsafeCoerceName ) where #include "HsVersions.h" import BasicTypes ( Arity, StrictnessMark(..), isMarkedUnboxed, isMarkedStrict ) -import TysPrim ( openAlphaTyVars, alphaTyVar, alphaTy, betaTyVar, betaTy, - intPrimTy, realWorldStatePrimTy, addrPrimTy +import Rules ( mkSpecInfo ) +import TysPrim ( openAlphaTyVars, alphaTyVar, alphaTy, + realWorldStatePrimTy, addrPrimTy ) import TysWiredIn ( charTy, mkListTy ) import PrelRules ( primOpRules ) -import Rules ( addRule ) -import TcType ( Type, ThetaType, mkDictTy, mkPredTys, mkTyConApp, - mkTyVarTys, mkClassPred, tcEqPred, +import Type ( TyThing(..), mkForAllTy, tyVarsOfTypes ) +import TcType ( Type, ThetaType, mkDictTy, mkPredTys, mkPredTy, + mkTyConApp, mkTyVarTys, mkClassPred, mkFunTys, mkFunTy, mkSigmaTy, tcSplitSigmaTy, isUnLiftedType, mkForAllTys, mkTyVarTy, tyVarsOfType, - tcSplitFunTys, tcSplitForAllTys, mkPredTy + tcSplitFunTys, tcSplitForAllTys, dataConsStupidTheta ) import CoreUtils ( exprType ) -import CoreUnfold ( mkTopUnfolding, mkCompulsoryUnfolding, mkOtherCon ) -import Literal ( Literal(..), nullAddrLit ) -import TyCon ( TyCon, isNewTyCon, tyConTyVars, tyConDataCons, - tyConTheta, isProductTyCon, isDataTyCon, isRecursiveTyCon ) -import Class ( Class, classTyCon, classTyVars, classSelIds ) +import CoreUnfold ( mkTopUnfolding, mkCompulsoryUnfolding ) +import Literal ( nullAddrLit, mkStringLit ) +import TyCon ( TyCon, isNewTyCon, tyConDataCons, FieldLabel, + tyConStupidTheta, isProductTyCon, isDataTyCon, isRecursiveTyCon ) +import Class ( Class, classTyCon, classSelIds ) import Var ( Id, TyVar, Var ) -import VarSet ( isEmptyVarSet ) -import Name ( mkFCallName, Name ) -import PrimOp ( PrimOp(DataToTagOp), primOpSig, mkPrimOpIdName ) +import VarSet ( isEmptyVarSet, subVarSet, varSetElems ) +import Name ( mkFCallName, mkWiredInName, Name, BuiltInSyntax(..) ) +import OccName ( mkOccNameFS, varName ) +import PrimOp ( PrimOp, primOpSig, primOpOcc, primOpTag ) import ForeignCall ( ForeignCall ) -import DataCon ( DataCon, - dataConFieldLabels, dataConRepArity, dataConTyCon, - dataConArgTys, dataConRepType, - dataConOrigArgTys, - dataConName, dataConTheta, - dataConSig, dataConStrictMarks, dataConWorkId, - splitProductType +import DataCon ( DataCon, DataConIds(..), dataConTyVars, + dataConFieldLabels, dataConRepArity, dataConResTys, + dataConRepArgTys, dataConRepType, + dataConSig, dataConStrictMarks, dataConExStricts, + splitProductType, isVanillaDataCon, dataConFieldType, + dataConInstOrigArgTys ) -import Id ( idType, mkGlobalId, mkVanillaGlobal, mkSysLocal, - mkTemplateLocals, mkTemplateLocalsNum, - mkTemplateLocal, idNewStrictness, idName +import Id ( idType, mkGlobalId, mkVanillaGlobal, mkSysLocal, + mkTemplateLocals, mkTemplateLocalsNum, mkExportedLocalId, + mkTemplateLocal, idName ) -import IdInfo ( IdInfo, noCafIdInfo, hasCafIdInfo, - setUnfoldingInfo, +import IdInfo ( IdInfo, noCafIdInfo, setUnfoldingInfo, setArityInfo, setSpecInfo, setCafInfo, - setAllStrictnessInfo, + setAllStrictnessInfo, vanillaIdInfo, GlobalIdDetails(..), CafInfo(..) ) -import NewDemand ( mkStrictSig, strictSigResInfo, DmdResult(..), +import NewDemand ( mkStrictSig, DmdResult(..), mkTopDmdType, topDmd, evalDmd, lazyDmd, retCPR, Demand(..), Demands(..) ) -import FieldLabel ( mkFieldLabel, fieldLabelName, - firstFieldLabelTag, allFieldLabelTags, fieldLabelType - ) import DmdAnal ( dmdAnalTopRhs ) import CoreSyn -import Unique ( mkBuiltinUnique ) +import Unique ( mkBuiltinUnique, mkPrimOpIdUnique ) import Maybes import PrelNames -import Maybe ( isJust ) import Util ( dropList, isSingleton ) import Outputable import FastString -import ListSetOps ( assoc, assocMaybe ) -import UnicodeUtil ( stringToUtf8 ) -import List ( nubBy ) +import ListSetOps ( assoc ) \end{code} %************************************************************************ @@ -138,7 +134,6 @@ ghcPrimIds realWorldPrimId, unsafeCoerceId, nullAddrId, - getTagId, seqId ] \end{code} @@ -149,57 +144,6 @@ ghcPrimIds %* * %************************************************************************ -\begin{code} -mkDataConId :: Name -> DataCon -> Id - -- Makes the *worker* for the data constructor; that is, the function - -- that takes the reprsentation arguments and builds the constructor. -mkDataConId work_name data_con - = mkGlobalId (DataConId data_con) work_name (dataConRepType data_con) info - where - info = noCafIdInfo - `setArityInfo` arity - `setAllStrictnessInfo` Just strict_sig - - arity = dataConRepArity data_con - - strict_sig = mkStrictSig (mkTopDmdType (replicate arity topDmd) cpr_info) - -- Notice that we do *not* say the worker is strict - -- even if the data constructor is declared strict - -- e.g. data T = MkT !(Int,Int) - -- Why? Because the *wrapper* is strict (and its unfolding has case - -- expresssions that do the evals) but the *worker* itself is not. - -- If we pretend it is strict then when we see - -- case x of y -> $wMkT y - -- the simplifier thinks that y is "sure to be evaluated" (because - -- $wMkT is strict) and drops the case. No, $wMkT is not strict. - -- - -- When the simplifer sees a pattern - -- case e of MkT x -> ... - -- it uses the dataConRepStrictness of MkT to mark x as evaluated; - -- but that's fine... dataConRepStrictness comes from the data con - -- not from the worker Id. - - tycon = dataConTyCon data_con - cpr_info | isProductTyCon tycon && - isDataTyCon tycon && - arity > 0 && - arity <= mAX_CPR_SIZE = retCPR - | otherwise = TopRes - -- RetCPR is only true for products that are real data types; - -- that is, not unboxed tuples or [non-recursive] newtypes - -mAX_CPR_SIZE :: Arity -mAX_CPR_SIZE = 10 --- We do not treat very big tuples as CPR-ish: --- a) for a start we get into trouble because there aren't --- "enough" unboxed tuple types (a tiresome restriction, --- but hard to fix), --- b) more importantly, big unboxed tuples get returned mainly --- on the stack, and are often then allocated in the heap --- by the caller. So doing CPR for them may in fact make --- things worse. -\end{code} - The wrapper for a constructor is an ordinary top-level binding that evaluates any strict args, unboxes any args that are going to be flattened, and calls the worker. @@ -237,23 +181,95 @@ Notice that Making an explicit case expression allows the simplifier to eliminate it in the (common) case where the constructor arg is already evaluated. + \begin{code} -mkDataConWrapId data_con - = mkGlobalId (DataConWrapId data_con) (dataConName data_con) wrap_ty info +mkDataConIds :: Name -> Name -> DataCon -> DataConIds + -- Makes the *worker* for the data constructor; that is, the function + -- that takes the reprsentation arguments and builds the constructor. +mkDataConIds wrap_name wkr_name data_con + | isNewTyCon tycon + = NewDC nt_wrap_id + + | any isMarkedStrict all_strict_marks -- Algebraic, needs wrapper + = AlgDC (Just alg_wrap_id) wrk_id + + | otherwise -- Algebraic, no wrapper + = AlgDC Nothing wrk_id where - work_id = dataConWorkId data_con + (tyvars, theta, orig_arg_tys, tycon, res_tys) = dataConSig data_con - info = noCafIdInfo - `setUnfoldingInfo` wrap_unf - -- The NoCaf-ness is set by noCafIdInfo - `setArityInfo` arity - -- It's important to specify the arity, so that partial - -- applications are treated as values - `setAllStrictnessInfo` Just wrap_sig - - wrap_sig = mkStrictSig (mkTopDmdType arg_dmds res_info) - res_info = strictSigResInfo (idNewStrictness work_id) - arg_dmds = map mk_dmd strict_marks + dict_tys = mkPredTys theta + all_arg_tys = dict_tys ++ orig_arg_tys + result_ty = mkTyConApp tycon res_tys + + wrap_ty = mkForAllTys tyvars (mkFunTys all_arg_tys result_ty) + -- We used to include the stupid theta in the wrapper's args + -- but now we don't. Instead the type checker just injects these + -- extra constraints where necessary. + + ----------- Worker (algebraic data types only) -------------- + wrk_id = mkGlobalId (DataConWorkId data_con) wkr_name + (dataConRepType data_con) wkr_info + + wkr_arity = dataConRepArity data_con + wkr_info = noCafIdInfo + `setArityInfo` wkr_arity + `setAllStrictnessInfo` Just wkr_sig + `setUnfoldingInfo` evaldUnfolding -- Record that it's evaluated, + -- even if arity = 0 + + wkr_sig = mkStrictSig (mkTopDmdType (replicate wkr_arity topDmd) cpr_info) + -- Notice that we do *not* say the worker is strict + -- even if the data constructor is declared strict + -- e.g. data T = MkT !(Int,Int) + -- Why? Because the *wrapper* is strict (and its unfolding has case + -- expresssions that do the evals) but the *worker* itself is not. + -- If we pretend it is strict then when we see + -- case x of y -> $wMkT y + -- the simplifier thinks that y is "sure to be evaluated" (because + -- $wMkT is strict) and drops the case. No, $wMkT is not strict. + -- + -- When the simplifer sees a pattern + -- case e of MkT x -> ... + -- it uses the dataConRepStrictness of MkT to mark x as evaluated; + -- but that's fine... dataConRepStrictness comes from the data con + -- not from the worker Id. + + cpr_info | isProductTyCon tycon && + isDataTyCon tycon && + wkr_arity > 0 && + wkr_arity <= mAX_CPR_SIZE = retCPR + | otherwise = TopRes + -- RetCPR is only true for products that are real data types; + -- that is, not unboxed tuples or [non-recursive] newtypes + + ----------- Wrappers for newtypes -------------- + nt_wrap_id = mkGlobalId (DataConWrapId data_con) wrap_name wrap_ty nt_wrap_info + nt_wrap_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo + `setArityInfo` 1 -- Arity 1 + `setUnfoldingInfo` newtype_unf + newtype_unf = ASSERT( isVanillaDataCon data_con && + isSingleton orig_arg_tys ) + -- No existentials on a newtype, but it can have a context + -- e.g. newtype Eq a => T a = MkT (...) + mkTopUnfolding $ Note InlineMe $ + mkLams tyvars $ Lam id_arg1 $ + mkNewTypeBody tycon result_ty (Var id_arg1) + + id_arg1 = mkTemplateLocal 1 (head orig_arg_tys) + + ----------- Wrappers for algebraic data types -------------- + alg_wrap_id = mkGlobalId (DataConWrapId data_con) wrap_name wrap_ty alg_wrap_info + alg_wrap_info = noCafIdInfo -- The NoCaf-ness is set by noCafIdInfo + `setArityInfo` alg_arity + -- It's important to specify the arity, so that partial + -- applications are treated as values + `setUnfoldingInfo` alg_unf + `setAllStrictnessInfo` Just wrap_sig + + all_strict_marks = dataConExStricts data_con ++ dataConStrictMarks data_con + wrap_sig = mkStrictSig (mkTopDmdType arg_dmds cpr_info) + arg_dmds = map mk_dmd all_strict_marks mk_dmd str | isMarkedStrict str = evalDmd | otherwise = lazyDmd -- The Cpr info can be important inside INLINE rhss, where the @@ -265,65 +281,19 @@ mkDataConWrapId data_con -- ...(let w = C x in ...(w p q)...)... -- we want to see that w is strict in its two arguments - wrap_unf | isNewTyCon tycon - = ASSERT( null ex_tyvars && null ex_dict_args && isSingleton orig_arg_tys ) - -- No existentials on a newtype, but it can have a context - -- e.g. newtype Eq a => T a = MkT (...) - mkTopUnfolding $ Note InlineMe $ - mkLams tyvars $ Lam id_arg1 $ - mkNewTypeBody tycon result_ty (Var id_arg1) - - | not (any isMarkedStrict strict_marks) - = mkCompulsoryUnfolding (Var work_id) - -- The common case. Not only is this efficient, - -- but it also ensures that the wrapper is replaced - -- by the worker even when there are no args. - -- f (:) x - -- becomes - -- f $w: x - -- This is really important in rule matching, - -- (We could match on the wrappers, - -- but that makes it less likely that rules will match - -- when we bring bits of unfoldings together.) - -- - -- NB: because of this special case, (map (:) ys) turns into - -- (map $w: ys). The top-level defn for (:) is never used. - -- This is somewhat of a bore, but I'm currently leaving it - -- as is, so that there still is a top level curried (:) for - -- the interpreter to call. - - | otherwise - = mkTopUnfolding $ Note InlineMe $ - mkLams all_tyvars $ - mkLams ex_dict_args $ mkLams id_args $ - foldr mk_case con_app - (zip (ex_dict_args++id_args) strict_marks) i3 [] - - con_app i rep_ids = mkApps (Var work_id) - (map varToCoreExpr (all_tyvars ++ reverse rep_ids)) - - (tyvars, _, ex_tyvars, ex_theta, orig_arg_tys, tycon) = dataConSig data_con - all_tyvars = tyvars ++ ex_tyvars - - ex_dict_tys = mkPredTys ex_theta - all_arg_tys = ex_dict_tys ++ orig_arg_tys - result_ty = mkTyConApp tycon (mkTyVarTys tyvars) - - wrap_ty = mkForAllTys all_tyvars (mkFunTys all_arg_tys result_ty) - -- We used to include the stupid theta in the wrapper's args - -- but now we don't. Instead the type checker just injects these - -- extra constraints where necessary. - - mkLocals i tys = (zipWith mkTemplateLocal [i..i+n-1] tys, i+n) - where - n = length tys + alg_unf = mkTopUnfolding $ Note InlineMe $ + mkLams tyvars $ + mkLams dict_args $ mkLams id_args $ + foldr mk_case con_app + (zip (dict_args ++ id_args) all_strict_marks) + i3 [] - (ex_dict_args,i2) = mkLocals 1 ex_dict_tys - (id_args,i3) = mkLocals i2 orig_arg_tys - arity = i3-1 - (id_arg1:_) = id_args -- Used for newtype only + con_app i rep_ids = mkApps (Var wrk_id) + (map varToCoreExpr (tyvars ++ reverse rep_ids)) - strict_marks = dataConStrictMarks data_con + (dict_args,i2) = mkLocals 1 dict_tys + (id_args,i3) = mkLocals i2 orig_arg_tys + alg_arity = i3-1 mk_case :: (Id, StrictnessMark) -- Arg, strictness @@ -337,15 +307,32 @@ mkDataConWrapId data_con MarkedStrict | isUnLiftedType (idType arg) -> body i (arg:rep_args) | otherwise -> - Case (Var arg) arg [(DEFAULT,[], body i (arg:rep_args))] + Case (Var arg) arg result_ty [(DEFAULT,[], body i (arg:rep_args))] MarkedUnboxed -> case splitProductType "do_unbox" (idType arg) of (tycon, tycon_args, con, tys) -> - Case (Var arg) arg [(DataAlt con, con_args, - body i' (reverse con_args ++ rep_args))] + Case (Var arg) arg result_ty + [(DataAlt con, + con_args, + body i' (reverse con_args ++ rep_args))] where (con_args, i') = mkLocals i tys + +mAX_CPR_SIZE :: Arity +mAX_CPR_SIZE = 10 +-- We do not treat very big tuples as CPR-ish: +-- a) for a start we get into trouble because there aren't +-- "enough" unboxed tuple types (a tiresome restriction, +-- but hard to fix), +-- b) more importantly, big unboxed tuples get returned mainly +-- on the stack, and are often then allocated in the heap +-- by the caller. So doing CPR for them may in fact make +-- things worse. + +mkLocals i tys = (zipWith mkTemplateLocal [i..i+n-1] tys, i+n) + where + n = length tys \end{code} @@ -393,41 +380,84 @@ Similarly for (recursive) newtypes unN :: forall b. N -> b -> b unN = /\b -> \n:N -> (coerce (forall a. a->a) n) + +Note [Naughty record selectors] +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +A "naughty" field is one for which we can't define a record +selector, because an existential type variable would escape. For example: + data T = forall a. MkT { x,y::a } +We obviously can't define + x (MkT v _) = v +Nevertheless we *do* put a RecordSelId into the type environment +so that if the user tries to use 'x' as a selector we can bleat +helpfully, rather than saying unhelpfully that 'x' is not in scope. +Hence the sel_naughty flag, to identify record selcectors that don't really exist. + +In general, a field is naughty if its type mentions a type variable that +isn't in the result type of the constructor. + +For GADTs, we require that all constructors with a common field 'f' have the same +result type (modulo alpha conversion). [Checked in TcTyClsDecls.checkValidTyCon] +E.g. + data T where + T1 { f :: a } :: T [a] + T2 { f :: a, y :: b } :: T [a] +and now the selector takes that type as its argument: + f :: forall a. T [a] -> a + f t = case t of + T1 { f = v } -> v + T2 { f = v } -> v +Note the forall'd tyvars of the selector are just the free tyvars +of the result type; there may be other tyvars in the constructor's +type (e.g. 'b' in T2). + \begin{code} + +-- XXX - autrijus - +-- Plan: 1. Determine naughtiness by comparing field type vs result type +-- 2. Install naughty ones with selector_ty of type _|_ and fill in mzero for info +-- 3. If it's not naughty, do the normal plan. + +mkRecordSelId :: TyCon -> FieldLabel -> Id mkRecordSelId tycon field_label -- Assumes that all fields with the same field label have the same type - -- - -- Annoyingly, we have to pass in the unpackCString# Id, because - -- we can't conjure it up out of thin air - = sel_id + | is_naughty = naughty_id + | otherwise = sel_id where - sel_id = mkGlobalId (RecordSelId field_label) (fieldLabelName field_label) selector_ty info - field_ty = fieldLabelType field_label - data_cons = tyConDataCons tycon - tyvars = tyConTyVars tycon -- These scope over the types in - -- the FieldLabels of constructors of this type - data_ty = mkTyConApp tycon tyvar_tys - tyvar_tys = mkTyVarTys tyvars - - -- Very tiresomely, the selectors are (unnecessarily!) overloaded over + is_naughty = not (tyVarsOfType field_ty `subVarSet` tyvar_set) + sel_id_details = RecordSelId tycon field_label is_naughty + + -- Escapist case here for naughty construcotrs + -- We give it no IdInfo, and a type of forall a.a (never looked at) + naughty_id = mkGlobalId sel_id_details field_label forall_a_a noCafIdInfo + forall_a_a = mkForAllTy alphaTyVar (mkTyVarTy alphaTyVar) + + -- Normal case starts here + sel_id = mkGlobalId sel_id_details field_label selector_ty info + data_cons = tyConDataCons tycon + data_cons_w_field = filter has_field data_cons -- Can't be empty! + has_field con = field_label `elem` dataConFieldLabels con + + con1 = head data_cons_w_field + res_tys = dataConResTys con1 + tyvar_set = tyVarsOfTypes res_tys + tyvars = varSetElems tyvar_set + data_ty = mkTyConApp tycon res_tys + field_ty = dataConFieldType con1 field_label + + -- *Very* tiresomely, the selectors are (unnecessarily!) overloaded over -- just the dictionaries in the types of the constructors that contain -- the relevant field. [The Report says that pattern matching on a -- constructor gives the same constraints as applying it.] Urgh. -- -- However, not all data cons have all constraints (because of - -- TcTyDecls.thinContext). So we need to find all the data cons + -- BuildTyCl.mkDataConStupidTheta). So we need to find all the data cons -- involved in the pattern match and take the union of their constraints. - -- - -- NB: this code relies on the fact that DataCons are quantified over - -- the identical type variables as their parent TyCon - tycon_theta = tyConTheta tycon -- The context on the data decl - -- eg data (Eq a, Ord b) => T a b = ... - needed_preds = [pred | (DataAlt dc, _, _) <- the_alts, pred <- dataConTheta dc] - dict_tys = map mkPredTy (nubBy tcEqPred needed_preds) - n_dict_tys = length dict_tys + stupid_dict_tys = mkPredTys (dataConsStupidTheta data_cons_w_field) + n_stupid_dicts = length stupid_dict_tys (field_tyvars,field_theta,field_tau) = tcSplitSigmaTy field_ty - field_dict_tys = map mkPredTy field_theta + field_dict_tys = mkPredTys field_theta n_field_dict_tys = length field_dict_tys -- If the field has a universally quantified type we have to -- be a bit careful. Suppose we have @@ -446,10 +476,10 @@ mkRecordSelId tycon field_label selector_ty :: Type selector_ty = mkForAllTys tyvars $ mkForAllTys field_tyvars $ - mkFunTys dict_tys $ mkFunTys field_dict_tys $ + mkFunTys stupid_dict_tys $ mkFunTys field_dict_tys $ mkFunTy data_ty field_tau - arity = 1 + n_dict_tys + n_field_dict_tys + arity = 1 + n_stupid_dicts + n_field_dict_tys (strict_sig, rhs_w_str) = dmdAnalTopRhs sel_rhs -- Use the demand analyser to work out strictness. @@ -462,21 +492,22 @@ mkRecordSelId tycon field_label `setAllStrictnessInfo` Just strict_sig -- Allocate Ids. We do it a funny way round because field_dict_tys is - -- almost always empty. Also note that we use length_tycon_theta + -- almost always empty. Also note that we use max_dict_tys -- rather than n_dict_tys, because the latter gives an infinite loop: -- n_dict tys depends on the_alts, which depens on arg_ids, which depends -- on arity, which depends on n_dict tys. Sigh! Mega sigh! - field_dict_base = length tycon_theta + 1 - dict_id_base = field_dict_base + n_field_dict_tys - field_base = dict_id_base + 1 - dict_ids = mkTemplateLocalsNum 1 dict_tys - field_dict_ids = mkTemplateLocalsNum field_dict_base field_dict_tys - data_id = mkTemplateLocal dict_id_base data_ty - - alts = map mk_maybe_alt data_cons - the_alts = catMaybes alts + stupid_dict_ids = mkTemplateLocalsNum 1 stupid_dict_tys + max_stupid_dicts = length (tyConStupidTheta tycon) + field_dict_base = max_stupid_dicts + 1 + field_dict_ids = mkTemplateLocalsNum field_dict_base field_dict_tys + dict_id_base = field_dict_base + n_field_dict_tys + data_id = mkTemplateLocal dict_id_base data_ty + arg_base = dict_id_base + 1 + + the_alts :: [CoreAlt] + the_alts = map mk_alt data_cons_w_field -- Already sorted by data-con + no_default = length data_cons == length data_cons_w_field -- No default needed - no_default = all isJust alts -- No default needed default_alt | no_default = [] | otherwise = [(DEFAULT, [], error_expr)] @@ -485,11 +516,11 @@ mkRecordSelId tycon field_label | otherwise = MayHaveCafRefs sel_rhs = mkLams tyvars $ mkLams field_tyvars $ - mkLams dict_ids $ mkLams field_dict_ids $ + mkLams stupid_dict_ids $ mkLams field_dict_ids $ Lam data_id $ sel_body sel_body | isNewTyCon tycon = mk_result (mkNewTypeBody tycon field_ty (Var data_id)) - | otherwise = Case (Var data_id) data_id (default_alt ++ the_alts) + | otherwise = Case (Var data_id) data_id field_tau (default_alt ++ the_alts) mk_result poly_result = mkVarApps (mkVarApps poly_result field_tyvars) field_dict_ids -- We pull the field lambdas to the top, so we need to @@ -499,22 +530,28 @@ mkRecordSelId tycon field_label -- foo :: forall a. T -> a -> a -- foo = /\a. \t:T. case t of { MkT f -> f a } - mk_maybe_alt data_con - = case maybe_the_arg_id of - Nothing -> Nothing - Just the_arg_id -> Just (mkReboxingAlt uniqs data_con arg_ids body) - where - body = mk_result (Var the_arg_id) - where - arg_ids = mkTemplateLocalsNum field_base (dataConOrigArgTys data_con) - -- No need to instantiate; same tyvars in datacon as tycon - - unpack_base = field_base + length arg_ids - uniqs = map mkBuiltinUnique [unpack_base..] - - -- arity+1 avoids all shadowing - maybe_the_arg_id = assocMaybe (field_lbls `zip` arg_ids) field_label - field_lbls = dataConFieldLabels data_con + mk_alt data_con + = -- In the non-vanilla case, the pattern must bind type variables and + -- the context stuff; hence the arg_prefix binding below + mkReboxingAlt uniqs data_con (arg_prefix ++ arg_ids) + (mk_result (Var the_arg_id)) + where + (arg_prefix, arg_ids) + | isVanillaDataCon data_con -- Instantiate from commmon base + = ([], mkTemplateLocalsNum arg_base (dataConInstOrigArgTys data_con res_tys)) + | otherwise -- The case pattern binds type variables, which are used + -- in the types of the arguments of the pattern + = (dc_tyvars ++ mkTemplateLocalsNum arg_base (mkPredTys dc_theta), + mkTemplateLocalsNum arg_base' dc_arg_tys) + + (dc_tyvars, dc_theta, dc_arg_tys, _, _) = dataConSig data_con + arg_base' = arg_base + length dc_theta + + unpack_base = arg_base' + length dc_arg_tys + uniqs = map mkBuiltinUnique [unpack_base..] + + the_arg_id = assoc "mkRecordSelId:mk_alt" (field_lbls `zip` arg_ids) field_label + field_lbls = dataConFieldLabels data_con error_expr = mkRuntimeErrorApp rEC_SEL_ERROR_ID field_tau full_msg full_msg = showSDoc (sep [text "No match in record selector", ppr sel_id]) @@ -536,7 +573,7 @@ mkRecordSelId tycon field_label mkReboxingAlt :: [Unique] -- Uniques for the new Ids -> DataCon - -> [Var] -- Source-level args + -> [Var] -- Source-level args, including existential dicts -> CoreExpr -- RHS -> CoreAlt @@ -551,7 +588,7 @@ mkReboxingAlt us con args rhs (DataAlt con, args', mkLets binds rhs) where - stricts = dataConStrictMarks con + stricts = dataConExStricts con ++ dataConStrictMarks con go [] stricts us = ([], []) @@ -604,12 +641,10 @@ This is unlike ordinary record selectors, which have all the for-alls at the outside. When dealing with classes it's very convenient to recover the original type signature from the class op selector. -ToDo: unify with mkRecordSelId? - \begin{code} mkDictSelId :: Name -> Class -> Id mkDictSelId name clas - = mkGlobalId (RecordSelId field_lbl) name sel_ty info + = mkGlobalId (ClassOpId clas) name sel_ty info where sel_ty = mkForAllTys tyvars (mkFunTy (idType dict_id) (idType the_arg_id)) -- We can't just say (exprType rhs), because that would give a type @@ -618,10 +653,7 @@ mkDictSelId name clas -- But it's type must expose the representation of the dictionary -- to gat (say) C a -> (a -> a) - field_lbl = mkFieldLabel name tycon sel_ty tag - tag = assoc "MkId.mkDictSelId" (map idName (classSelIds clas) `zip` allFieldLabelTags) name - - info = noCafIdInfo + info = noCafIdInfo `setArityInfo` 1 `setUnfoldingInfo` mkTopUnfolding rhs `setAllStrictnessInfo` Just strict_sig @@ -638,21 +670,19 @@ mkDictSelId name clas | otherwise = Eval (Prod [ if the_arg_id == id then evalDmd else Abs | id <- arg_ids ]) - tyvars = classTyVars clas - tycon = classTyCon clas [data_con] = tyConDataCons tycon - tyvar_tys = mkTyVarTys tyvars - arg_tys = dataConArgTys data_con tyvar_tys - the_arg_id = arg_ids !! (tag - firstFieldLabelTag) + tyvars = dataConTyVars data_con + arg_tys = dataConRepArgTys data_con + the_arg_id = assoc "MkId.mkDictSelId" (map idName (classSelIds clas) `zip` arg_ids) name - pred = mkClassPred clas tyvar_tys + pred = mkClassPred clas (mkTyVarTys tyvars) (dict_id:arg_ids) = mkTemplateLocals (mkPredTy pred : arg_tys) rhs | isNewTyCon tycon = mkLams tyvars $ Lam dict_id $ mkNewTypeBody tycon (head arg_tys) (Var dict_id) | otherwise = mkLams tyvars $ Lam dict_id $ - Case (Var dict_id) dict_id + Case (Var dict_id) dict_id (idType the_arg_id) [(DataAlt data_con, arg_ids, Var the_arg_id)] mkNewTypeBody tycon result_ty result_expr @@ -678,17 +708,16 @@ mkPrimOpId prim_op where (tyvars,arg_tys,res_ty, arity, strict_sig) = primOpSig prim_op ty = mkForAllTys tyvars (mkFunTys arg_tys res_ty) - name = mkPrimOpIdName prim_op + name = mkWiredInName gHC_PRIM (primOpOcc prim_op) + (mkPrimOpIdUnique (primOpTag prim_op)) + Nothing (AnId id) UserSyntax id = mkGlobalId (PrimOpId prim_op) name ty info info = noCafIdInfo - `setSpecInfo` rules - `setArityInfo` arity + `setSpecInfo` mkSpecInfo (primOpRules prim_op name) + `setArityInfo` arity `setAllStrictnessInfo` Just strict_sig - rules = foldl (addRule id) emptyCoreRules (primOpRules prim_op) - - -- For each ccall we manufacture a separate CCallOpId, giving it -- a fresh unique, a type that is correct for this particular ccall, -- and a CCall structure that gives the correct details about calling @@ -734,11 +763,9 @@ Dict funs and default methods are *not* ImplicitIds. Their definition involves user-written code, so we can't figure out their strictness etc based on fixed info, as we can for constructors and record selectors (say). -We build them as GlobalIds, but when in the module where they are -bound, we turn the Id at the *binding site* into an exported LocalId. -This ensures that they are taken to account by free-variable finding -and dependency analysis (e.g. CoreFVs.exprFreeVars). The simplifier -will propagate the LocalId to all occurrence sites. +We build them as LocalIds, but with External Names. This ensures that +they are taken to account by free-variable finding and dependency +analysis (e.g. CoreFVs.exprFreeVars). Why shouldn't they be bound as GlobalIds? Because, in particular, if they are globals, the specialiser floats dict uses above their defns, @@ -751,11 +778,11 @@ It's OK for dfuns to be LocalIds, because we form the instance-env to pass on to the next module (md_insts) in CoreTidy, afer tidying and globalising the top-level Ids. -BUT make sure they are *exported* LocalIds (setIdLocalExported) so +BUT make sure they are *exported* LocalIds (mkExportedLocalId) so that they aren't discarded by the occurrence analyser. \begin{code} -mkDefaultMethodId dm_name ty = mkVanillaGlobal dm_name ty noCafIdInfo +mkDefaultMethodId dm_name ty = mkExportedLocalId dm_name ty mkDictFunId :: Name -- Name to use for the dict fun; -> [TyVar] @@ -765,7 +792,7 @@ mkDictFunId :: Name -- Name to use for the dict fun; -> Id mkDictFunId dfun_name inst_tyvars dfun_theta clas inst_tys - = mkVanillaGlobal dfun_name dfun_ty noCafIdInfo + = mkExportedLocalId dfun_name dfun_ty where dfun_ty = mkSigmaTy inst_tyvars dfun_theta (mkDictTy clas inst_tys) @@ -775,7 +802,7 @@ mkDictFunId dfun_name inst_tyvars dfun_theta clas inst_tys (class_tyvars, sc_theta, _, _) = classBigSig clas not_const (clas, tys) = not (isEmptyVarSet (tyVarsOfTypes tys)) - sc_theta' = substClasses (mkTopTyVarSubst class_tyvars inst_tys) sc_theta + sc_theta' = substClasses (zipTopTvSubst class_tyvars inst_tys) sc_theta dfun_theta = case inst_decl_theta of [] -> [] -- If inst_decl_theta is empty, then we don't -- want to have any dict arguments, so that we can @@ -821,6 +848,29 @@ they can unify with both unlifted and lifted types. Hence we provide another gun with which to shoot yourself in the foot. \begin{code} +mkWiredInIdName mod fs uniq id + = mkWiredInName mod (mkOccNameFS varName fs) uniq Nothing (AnId id) UserSyntax + +unsafeCoerceName = mkWiredInIdName gHC_PRIM FSLIT("unsafeCoerce#") unsafeCoerceIdKey unsafeCoerceId +nullAddrName = mkWiredInIdName gHC_PRIM FSLIT("nullAddr#") nullAddrIdKey nullAddrId +seqName = mkWiredInIdName gHC_PRIM FSLIT("seq") seqIdKey seqId +realWorldName = mkWiredInIdName gHC_PRIM FSLIT("realWorld#") realWorldPrimIdKey realWorldPrimId +lazyIdName = mkWiredInIdName pREL_BASE FSLIT("lazy") lazyIdKey lazyId + +errorName = mkWiredInIdName pREL_ERR FSLIT("error") errorIdKey eRROR_ID +recSelErrorName = mkWiredInIdName pREL_ERR FSLIT("recSelError") recSelErrorIdKey rEC_SEL_ERROR_ID +runtimeErrorName = mkWiredInIdName pREL_ERR FSLIT("runtimeError") runtimeErrorIdKey rUNTIME_ERROR_ID +irrefutPatErrorName = mkWiredInIdName pREL_ERR FSLIT("irrefutPatError") irrefutPatErrorIdKey iRREFUT_PAT_ERROR_ID +recConErrorName = mkWiredInIdName pREL_ERR FSLIT("recConError") recConErrorIdKey rEC_CON_ERROR_ID +patErrorName = mkWiredInIdName pREL_ERR FSLIT("patError") patErrorIdKey pAT_ERROR_ID +noMethodBindingErrorName = mkWiredInIdName pREL_ERR FSLIT("noMethodBindingError") + noMethodBindingErrorIdKey nO_METHOD_BINDING_ERROR_ID +nonExhaustiveGuardsErrorName + = mkWiredInIdName pREL_ERR FSLIT("nonExhaustiveGuardsError") + nonExhaustiveGuardsErrorIdKey nON_EXHAUSTIVE_GUARDS_ERROR_ID +\end{code} + +\begin{code} -- unsafeCoerce# :: forall a b. a -> b unsafeCoerceId = pcMiscPrelId unsafeCoerceName ty info @@ -849,10 +899,11 @@ seqId info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs - ty = mkForAllTys [alphaTyVar,betaTyVar] - (mkFunTy alphaTy (mkFunTy betaTy betaTy)) - [x,y] = mkTemplateLocals [alphaTy, betaTy] - rhs = mkLams [alphaTyVar,betaTyVar,x,y] (Case (Var x) x [(DEFAULT, [], Var y)]) + ty = mkForAllTys [alphaTyVar,openBetaTyVar] + (mkFunTy alphaTy (mkFunTy openBetaTy openBetaTy)) + [x,y] = mkTemplateLocals [alphaTy, openBetaTy] +-- gaw 2004 + rhs = mkLams [alphaTyVar,openBetaTyVar,x,y] (Case (Var x) x openBetaTy [(DEFAULT, [], Var y)]) -- lazy :: forall a?. a? -> a? (i.e. works for unboxed types too) -- Used to lazify pseq: pseq a b = a `seq` lazy b @@ -872,24 +923,6 @@ lazyIdUnfolding = mkLams [openAlphaTyVar,x] (Var x) [x] = mkTemplateLocals [openAlphaTy] \end{code} -@getTag#@ is another function which can't be defined in Haskell. It needs to -evaluate its argument and call the dataToTag# primitive. - -\begin{code} -getTagId - = pcMiscPrelId getTagName ty info - where - info = noCafIdInfo `setUnfoldingInfo` mkCompulsoryUnfolding rhs - -- We don't provide a defn for this; you must inline it - - ty = mkForAllTys [alphaTyVar] (mkFunTy alphaTy intPrimTy) - [x,y] = mkTemplateLocals [alphaTy,alphaTy] - rhs = mkLams [alphaTyVar,x] $ - Case (Var x) y [ (DEFAULT, [], mkApps (Var dataToTagId) [Type alphaTy, Var y]) ] - -dataToTagId = mkPrimOpId DataToTagOp -\end{code} - @realWorld#@ used to be a magic literal, \tr{void#}. If things get nasty as-is, change it back to a literal (@Literal@). @@ -903,8 +936,8 @@ This comes up in strictness analysis \begin{code} realWorldPrimId -- :: State# RealWorld = pcMiscPrelId realWorldName realWorldStatePrimTy - (noCafIdInfo `setUnfoldingInfo` mkOtherCon []) - -- The mkOtherCon makes it look that realWorld# is evaluated + (noCafIdInfo `setUnfoldingInfo` evaldUnfolding) + -- The evaldUnfolding makes it look that realWorld# is evaluated -- which in turn makes Simplify.interestingArg return True, -- which in turn makes INLINE things applied to realWorld# likely -- to be inlined @@ -946,15 +979,15 @@ mkRuntimeErrorApp mkRuntimeErrorApp err_id res_ty err_msg = mkApps (Var err_id) [Type res_ty, err_string] where - err_string = Lit (MachStr (mkFastString (stringToUtf8 err_msg))) + err_string = Lit (mkStringLit err_msg) rEC_SEL_ERROR_ID = mkRuntimeErrorId recSelErrorName rUNTIME_ERROR_ID = mkRuntimeErrorId runtimeErrorName iRREFUT_PAT_ERROR_ID = mkRuntimeErrorId irrefutPatErrorName rEC_CON_ERROR_ID = mkRuntimeErrorId recConErrorName -nON_EXHAUSTIVE_GUARDS_ERROR_ID = mkRuntimeErrorId nonExhaustiveGuardsErrorName pAT_ERROR_ID = mkRuntimeErrorId patErrorName nO_METHOD_BINDING_ERROR_ID = mkRuntimeErrorId noMethodBindingErrorName +nON_EXHAUSTIVE_GUARDS_ERROR_ID = mkRuntimeErrorId nonExhaustiveGuardsErrorName -- The runtime error Ids take a UTF8-encoded string as argument mkRuntimeErrorId name = pc_bottoming_Id name runtimeErrorTy @@ -991,7 +1024,7 @@ pcMiscPrelId name ty info pc_bottoming_Id name ty = pcMiscPrelId name ty bottoming_info where - bottoming_info = hasCafIdInfo `setAllStrictnessInfo` Just strict_sig + bottoming_info = vanillaIdInfo `setAllStrictnessInfo` Just strict_sig -- Do *not* mark them as NoCafRefs, because they can indeed have -- CAF refs. For example, pAT_ERROR_ID calls GHC.Err.untangle, -- which has some CAFs