{-# LANGUAGE DeriveDataTypeable #-} {-# LANGUAGE DerivingStrategies #-} -- | Renamer-level information about 'Name's. -- -- Renamer equivalent of 'TyThing'. module GHC.Types.GREInfo where import GHC.Prelude import GHC.Types.Basic import GHC.Types.FieldLabel import GHC.Types.Name import GHC.Types.Unique import GHC.Types.Unique.Set import GHC.Utils.Outputable import GHC.Utils.Panic import Control.DeepSeq ( NFData(..), deepseq ) import Data.Data ( Data ) import Data.List.NonEmpty ( NonEmpty ) import qualified Data.List.NonEmpty as NonEmpty {-********************************************************************** * * GREInfo * * ************************************************************************ Note [GREInfo] ~~~~~~~~~~~~~~ In the renamer, we sometimes need a bit more information about a 'Name', e.g. whether it is a type constructor, class, data constructor, record field, etc. For example, when typechecking record construction, the renamer needs to look up the fields of the data constructor being used (see e.g. GHC.Rename.Pat.rnHsRecFields). Extra information also allows us to provide better error messages when a fatal error occurs in the renamer, as it allows us to distinguish classes, type families, type synonyms, etc. For imported Names, we have access to the full type information in the form of a TyThing (although see Note [Retrieving the GREInfo from interfaces]). However, for Names in the module currently being renamed, we don't yet have full information. Instead of using TyThing, we use the GREInfo type, and this information gets affixed to each element in the GlobalRdrEnv. This allows us to treat imported and local Names in a consistent manner: always look at the GREInfo. Note [Retrieving the GREInfo from interfaces] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ If we have a TyThing, we can easily compute the corresponding GREInfo: this is done in GHC.Types.TyThing.tyThingGREInfo. However, one often needs to produce GlobalRdrElts (and thus their GREInfos) directly after loading interface files, before they are typechecked. For example: - GHC.Tc.Module.tcRnModuleTcRnM first calls tcRnImports, which starts off calling rnImports which transitively calls filterImports. That function is responsible for coughing up GlobalRdrElts (and their GREInfos) obtained from interfaces, but we will only typecheck the interfaces after we have finished processing the imports (see e.g. the logic at the start of tcRnImports which sets eps_is_boot, which decides whether we should look in the boot or non-boot interface for any particular module). - GHC.Tc.Utils.Backpack.mergeSignatures first loads the relevant signature interfaces to merge them, but only later on does it typecheck them. In both of these examples, what's important is that we **lazily** produce the GREInfo: it should only be consulted once the interfaces have been typechecked, which will add the necessary information to the type-level environment. In particular, the respective functions 'filterImports' and 'mergeSignatures' should NOT force the gre_info field. We delay the loading of interfaces by making the gre_info field of 'GlobalRdrElt' a thunk which, when forced, loads the interface, looks up the 'Name' in the type environment to get its associated TyThing, and computes the GREInfo from that. See 'GHC.Rename.Env.lookupGREInfo'. A possible alternative design would be to change the AvailInfo datatype to also store GREInfo. We currently don't do that, as this would mean that every time an interface re-exports something it has to also provide its GREInfo, which could lead to bloat. Note [Forcing GREInfo] ~~~~~~~~~~~~~~~~~~~~~~ The GREInfo field of a GlobalRdrElt needs to be lazy, as explained in Note [Retrieving the GREInfo from interfaces]. For imported things, this field is usually a thunk which looks up the GREInfo in a type environment (see GHC.Rename.Env.lookupGREInfo). We thus need to be careful not to introduce space leaks: such thunks could end up retaining old type environments, which would violate invariant (5) of Note [GHC Heap Invariants] in GHC.Driver.Make. This can happen, for example, when reloading in GHCi (see e.g. test T15369, which can trigger the ghci leak check if we're not careful). A naive approach is to simply deeply force the whole GlobalRdrEnv. However, forcing the GREInfo thunks can force the loading of interface files which we otherwise might not need to load, so it leads to wasted work. Instead, whenever we are about to store the GlobalRdrEnv somewhere (such as in ModDetails), we dehydrate it by stripping away the GREInfo field, turning it into (). See 'forceGlobalRdrEnv' and its cousin 'hydrateGlobalRdrEnv', as well as Note [IfGlobalRdrEnv] in GHC.Types.Name.Reader. Search for references to this note in the code for illustration. -} -- | Information about a 'Name' that is pertinent to the renamer. -- -- See Note [GREInfo] data GREInfo -- | No particular information... e.g. a function = Vanilla -- | An unbound GRE... could be anything | UnboundGRE -- | 'TyCon' | IAmTyCon !(TyConFlavour Name) -- | 'ConLike' | IAmConLike !ConInfo -- ^ The constructor fields. -- See Note [Local constructor info in the renamer]. -- | Record field | IAmRecField !RecFieldInfo deriving Data instance NFData GREInfo where rnf Vanilla = () rnf UnboundGRE = () rnf (IAmTyCon tc) = rnf tc rnf (IAmConLike info) = rnf info rnf (IAmRecField info) = rnf info plusGREInfo :: GREInfo -> GREInfo -> GREInfo plusGREInfo Vanilla Vanilla = Vanilla plusGREInfo UnboundGRE UnboundGRE = UnboundGRE plusGREInfo (IAmTyCon {}) info2@(IAmTyCon {}) = info2 plusGREInfo (IAmConLike {}) info2@(IAmConLike {}) = info2 plusGREInfo (IAmRecField {}) info2@(IAmRecField {}) = info2 plusGREInfo info1 info2 = pprPanic "plusInfo" $ vcat [ text "info1:" <+> ppr info1 , text "info2:" <+> ppr info2 ] instance Outputable GREInfo where ppr Vanilla = text "Vanilla" ppr UnboundGRE = text "UnboundGRE" ppr (IAmTyCon flav) = text "TyCon" <+> ppr flav ppr (IAmConLike info) = text "ConLike" <+> ppr info ppr (IAmRecField info) = text "RecField" <+> ppr info {-********************************************************************** * * Constructor info * * ************************************************************************ Note [Local constructor info in the renamer] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ As explained in Note [GREInfo], information pertinent to the renamer is stored using the GREInfo datatype. What information do we need about constructors? Consider the following example: data T = T1 { x, y :: Int } | T2 { x :: Int } | T3 | T4 Int Bool We need to know: * The fields of the data constructor, so that - We can complain if you say `T1 { v = 3 }`, where `v` is not a field of `T1` See the following call stack * GHC.Rename.Expr.rnExpr (RecordCon case) * GHC.Rename.Pat.rnHsRecFields * GHC.Rename.Env.lookupRecFieldOcc - Ditto if you pattern match on `T1 { v = x }`. See the following call stack * GHC.Rename.Pat.rnHsRecPatsAndThen * GHC.Rename.Pat.rnHsRecFields * GHC.Rename.Env.lookupRecFieldOcc - We can fill in the dots if you say `T1 {..}` in construction or pattern matching See GHC.Rename.Pat.rnHsRecFields.rn_dotdot * Whether the contructor is nullary. We need to know this to accept `T2 {..}`, and `T3 {..}`, but reject `T4 {..}`, in both construction and pattern matching. See GHC.Rename.Pat.rnHsRecFields.rn_dotdot and Note [Nullary constructors and empty record wildcards] Note [Nullary constructors and empty record wildcards] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ A nullary constructor is one with no arguments. For example, both `data T = MkT` and `data T = MkT {}` are nullary. For consistency and TH convenience, it was agreed that a `{..}` match or usage on nullary constructors would be accepted. This is done as as per https://github.com/ghc-proposals/ghc-proposals/blob/master/proposals/0496-empty-record-wildcards.rst -} -- | Information about the record fields of a constructor. -- -- See Note [Local constructor info in the renamer] data ConInfo = ConHasRecordFields (NonEmpty FieldLabel) | ConHasPositionalArgs | ConIsNullary deriving stock Eq deriving Data instance NFData ConInfo where rnf ConIsNullary = () rnf ConHasPositionalArgs = () rnf (ConHasRecordFields flds) = rnf flds mkConInfo :: Arity -> [FieldLabel] -> ConInfo mkConInfo 0 _ = ConIsNullary mkConInfo _ fields = maybe ConHasPositionalArgs ConHasRecordFields $ NonEmpty.nonEmpty fields conInfoFields :: ConInfo -> [FieldLabel] conInfoFields (ConHasRecordFields fields) = NonEmpty.toList fields conInfoFields ConHasPositionalArgs = [] conInfoFields ConIsNullary = [] instance Outputable ConInfo where ppr ConIsNullary = text "ConIsNullary" ppr ConHasPositionalArgs = text "ConHasPositionalArgs" ppr (ConHasRecordFields fieldLabels) = text "ConHasRecordFields" <+> braces (ppr fieldLabels) -- | The 'Name' of a 'ConLike'. -- -- Useful when we are in the renamer and don't yet have a full 'DataCon' or -- 'PatSyn' to hand. data ConLikeName = DataConName { conLikeName_Name :: !Name } | PatSynName { conLikeName_Name :: !Name } deriving (Eq, Data) instance Outputable ConLikeName where ppr = ppr . conLikeName_Name instance Uniquable ConLikeName where getUnique = getUnique . conLikeName_Name instance NFData ConLikeName where rnf = rnf . conLikeName_Name {-********************************************************************** * * Record field info * * **********************************************************************-} data RecFieldInfo = RecFieldInfo { recFieldLabel :: !FieldLabel , recFieldCons :: !(UniqSet ConLikeName) -- ^ The constructors which have this field label. -- Always non-empty. -- -- NB: these constructors will always share a single parent, -- as the field label disambiguates between parents in the presence -- of duplicate record fields. } deriving (Eq, Data) instance NFData RecFieldInfo where rnf (RecFieldInfo lbl cons) = rnf lbl `seq` nonDetStrictFoldUniqSet deepseq () cons instance Outputable RecFieldInfo where ppr (RecFieldInfo { recFieldLabel = fl, recFieldCons = cons }) = text "RecFieldInfo" <+> braces (text "recFieldLabel:" <+> ppr fl <> comma <+> text "recFieldCons:" <+> pprWithCommas ppr (nonDetEltsUniqSet cons))