wiki:Records/OverloadedRecordFields/Implementation

Version 39 (modified by adamgundry, 8 months ago) (diff)

bug fixing

Overloaded record fields: implementation notes

Here be dragons. This page describes implementation details and progress on the implementation of the overloaded record fields plan. Development of the extension is taking place on forks of the ghc and packages-base repositories (on branch 'overloaded-record-fields'). A prototype implementation is also available.

The basic idea

The Has and Upd classes, and GetResult and SetResult type families, are defined in the module GHC.Records in the base package.

Typechecking a record datatype still generates record selectors, but their names have a $sel prefix and end with the name of their type. Moreover, instances for the classes and type families are generated. For example,

data T = MkT { x :: Int }

generates

$sel_x_T :: T -> Int -- record selector (used to be called `x`)
$sel_x_T (MkT x) = x

$dfHasTx :: forall a . a ~ Int => Has T "x" a -- corresponds to the Has instance decl
$dfHasTx = Has { getField _ = $sel_x_T }

$dfUpdTx :: forall a . a ~ Int => Upd T "x" a -- corresponds to the Upd instance decl
$dfUpdTx = Upd { setField _ s e = s { x = e } }

axiom TFCo:R:GetResult : GetResult T "x" = Int   -- corresponds to the GetResult type family instance
axiom TFCo:R:SetResult : SetResult T "x" Int = T -- corresponds to the SetResult type family instance

The naming of cats

The AvailTC Name [Name] [(OccName, Name)] constructor of AvailInfo represents a type and its pieces that are in scope. Record fields are now stored in a separate list (the third argument), along with their selectors. The IEThingWith name [name] [OccName] constructor of IE, which represents a thing that can be imported or exported, only stores the field labels. SLPJ Whoa! Why should we duplicate this info. My gut feel is that the selector should not appear in the second argument. AMG Does this sound better now? It's helpful if gresFromAvail need not do lookups (it is called by the desugarer).

The Parent type has an extra constructor FldParent Name OccName that stores the parent Name and the field OccName. The GlobalRdrElt (GRE) for a field stores the selector name directly, and uses the FldParent constructor to store the field. Thus a field foo of type T gives rise this entry in the GlobalRdrEnv:

foo |->  GRE $sel_foo_T (FldParent T foo) LocalDef

SLPJ moreover I think we should store the dictionary $dfHasTfoo in the GRE for foo, not the selector. That way we get both getter and setter (via the dictionary) in one go. AMG Now I'm not sure about this. We can't build the dictionary for higher-rank fields, but they have a perfectly good selector. Moreover, with type-changing update there are two dictionaries (one for the getter and one for the setter) and two coercion axioms.

Note that the OccName used when adding a GRE to the environment (greOccName) now depends on the parent field: for FldParent it is the field label rather than the selector name.

The dcFields field of DataCon stores a list of FieldLabel, defined thus:

type FieldLbl a = FieldLabel {
                      flOccName  :: OccName, -- ^ Label of the field
                      flSelector :: a        -- ^ Record selector function
                  }
type FieldLabel = FieldLbl Name

whereas the ifConFields field of IfaceConDecl stores a list of FieldLbl OccName.

Source expressions

The HsExpr type has extra constructors HsOverloadedRecFld OccName and HsSingleRecFld OccName id. When -XOverloadedRecordFields is enabled, and rnExpr encounters HsVar "x" where x refers to multiple GREs that are all record fields, it replaces it with HsOverloadedRecFld "x". When the typechecker sees HsOverloadedRecFld x it emits a wanted constraint Has alpha x beta and returns type alpha -> beta where alpha and beta are fresh unification variables.

When the flag is not enabled, rnExpr turns an unambiguous record field foo into HsSingleRecFld foo $sel_foo_T. The point of this constructor is so we can pretty-print the field name but store the selector name for typechecking.

Where an AST representation type (e.g. HsRecField or ConDeclField) contained an argument of type Located id for a field, it now stores a Located RdrName for the label, and some representation of the selector. The parser uses an error thunk for the selector; it is filled in by the renamer (by rnHsRecFields1 in RnPat, and rnField in RnTypes). The new definition of ConDeclField (used in types) is:

data ConDeclField name
  = ConDeclField { cd_fld_lbl  :: Located RdrName,
                   cd_fld_sel  :: name,
                   cd_fld_type :: LBangType name, 
                   cd_fld_doc  :: Maybe LHsDocString }

The new definition of HsRecField is:

data HsRecField id arg = HsRecField {
        hsRecFieldLbl :: Located RdrName,
        hsRecFieldSel :: Either id [(id, id)],
        hsRecFieldArg :: arg,
        hsRecPun      :: Bool }

The renamer (rnHsRecFields1) supplies Left sel_name for the selector if it is unambiguous, or Right xs if it is ambiguous (because it is for a record update, and there are multiple fields with the correct label in scope). In the latter case, the possibilities xs are represented as a list of (parent name, selector name) pairs. The typechecker (tcExpr) tries three ways to disambiguate the update:

  1. Perhaps only one type has all the fields that are being updated.
  1. Use the type being pushed in, if it is already a TyConApp.
  1. Use the type signature of the record expression, if it exists and is a TyConApp.

Automatic instance generation

Has instances are generated, provided the extension is enabled, in tcInstDecls1 (the same time as derived instances (from deriving clauses) are generated). Every record field GRE in scope gives rise to an instance. Such instances are available when typechecking the current module (in tcg_inst_env) but not exported to other modules (via tcg_insts). At the moment, fresh DFunIds are generated for all instances in scope for each module, even though they are exported in interface files. Perhaps this should change.

AMG I wanted to generate each DFunId for Has once, at the field's definition site, but this causes problems for the fields defined in base, as the Has class may not be available. I've reverted to generating fresh DFunIds locally to each module for which -XOverloadedRecordFields is used. Is there a better way to do this?

As well as Has instances, instances of the type family GetResult are generated, and exactly the same question about dfun names applies to their axiom names.

Unused imports

Unused imports and generation of the minimal import list (RnNames.warnUnusedImportDecls) use a map from selector names to labels, in order to print fields correctly. However, fields may currently be reported as unused even if the corresponding Has instance is used. Consider the following:

module A where
  data T = MkT { x,y:Int }

module B where
  data S = MkS { x,y::Bool }

module C where
  import A( T(x) )
  import B( S(x) )

  foo :: T -> Int
  foo r = r.x + 2

Now, do we expect to report the 'x' in S(x) import as unused? Actually the entire 'import B' is unused. Only the typechecker will eventually know that. But I think the type checker does actually record which instances are used, so perhaps we can make use of that info to give accurate unused-import info.

AMG I thought this would be the tcg_ev_binds field of the TcGblEnv, but this seems to be empty by the end of tcRnModule. We could also look at the inert_solved_dicts field of InertSet, but I'm not sure how to propagate the required information out of the TcS monad to the TcM monad where unused names are reported.

GADT record updates

Consider the example

data W a where
    MkW :: a ~ b => { x :: a, y :: b } -> W (a, b)

It would be nice to generate

-- setField :: proxy "x" -> W (a, b) -> a -> W (a, b)
setField _ s e = s { x = e }

but this record update is rejected by the typechecker, even though it is perfectly sensible, because of #2595. The currently implemented workaround is instead to generate the explicit update

setField _ (MkW _ y) x = MkW x y

which is fine, but rather long-winded if there are many constructors or fields. Essentially this is doing the job of the desugarer for record updates.

Note that W does not admit type-changing single update for either field, because of the a ~ b constraint. Without it, though, type-changing update should be allowed.

Type-changing update: phantom arguments

Consider the datatype

data T a = MkT { foo :: Int }

where a is a phantom type argument (it does not occur in the type of foo). The traditional update syntax can change the phantom argument, for example if r :: T Int then r { foo = 3 } :: T Bool typechecks. However, setField cannot do so, because this is illegal:

type instance SetResult (T a) "foo" Int = T b

Note that the result of the type family involves an unbound variable b.

In general, a use of setField can only change type variables that occur in the field type being updated, and do not occur in any of the other fields' types.

Data families

Consider the following:

data family F (a :: *) :: *
data instance F Int  = MkF1 { foo :: Int }
data instance F Bool = MkF2 { foo :: Bool }

This is perfectly sensible, and should give rise to two *different* record selectors foo, and corresponding Has instances:

instance t ~ Int => Has (F Int) "foo" t
instance t ~ Bool => Has (F Bool) "foo" t

However, what can we call the record selectors? They can't both be $sel_foo_F! Ideally we would use the name of the representation tycon, rather than the family tycon, but that isn't introduced until the typechecker (tcDataFamInstDecl in TcInstDcls), and we need to create the selector in the renamer (getLocalNonValBinders in RnNames). We can't just pick an arbitrary unique name, because we need to look up the selector to associate it with its data constructor (extendRecordFieldEnv in RnSource).

Qualified names

Consider the following:

module M where
  data S = MkS { foo :: Int }

module N where
  data T = MkT { foo :: Int }
  data U = MkU { foo :: Int }

module O where
  import M
  import N

  f x = M.foo x
  g x = N.foo x
  h x = foo x

Should there be a difference between f, g and h? It would seem odd if f could turn out to use the foo from T or U even though it explicitly says M.foo. I can see three sensible options:

  • Treat qualified and unqualified fields identically, but issue a warning for qualified fields
  • Forbid referring to overloaded fields with qualified names (so M.foo and N.foo yield errors)
  • Treat a qualified name as a non-overloaded field, generating an ambiguity error if necessary (so M.foo is okay but N.foo is ambiguous)

Of course, it is fine to use a qualified name in a record update.

I've implemented the first option, adding a new warning -fwarn-qualified-overloaded-record-fields which is on by default.

Outstanding bugs

  • typechecker/should_fail/tcfail102 (changed error message)

To do

  • Sort out data families with duplicated fields.
  • Improve error messages from typechecker:
    • Unsolved Accessor p f where p is something silly
  • Consider defaulting Accessor p to p = (->), and defaulting Has r "f" t constraints where there is only one datatype with a field f in scope.
  • Sort out reporting of unused imports.
  • How should dfunids/axioms and instances be propagated?
  • Where should automatic instances be generated for GHCi?
  • How should deprecation work for fields? Not at all?
  • Document the extension, including new warnings.