Version 2 (modified by guest, 6 years ago) (diff)

publish, needs links

DORF -- Implementor's view

The implementation has been prototyped in GHC 7.2.1, see, and SPJ's observations/possible improvements and caveats

The fact that DORF has been 'faked' in existing GHC is good evidence that it's a modest change. Furthermore, we can implement H98-style records/fields using the same mechanism.

DORF is to be enabled by a compiler flag ‑XDeclaredOverloadedRecordFields, which implies flag ‑XNoMonoRecordFields, which in turn implies ‑XDisambiguateRecordFields and -XNamedFieldPuns with ‑XRecordWildCards.

Note we do _not_ assume flag ‑XDotPostfixFuncApply; dot notation is not needed by DORF, it's purely syntactic sugar to suit the taste of the programmer.

DORF is implemented through a class Has' with methods get' and set'. (Very similar in principle to SORF.) There's an instance of Has' for each record/field combination, with the instance generated from the record declaration.

Within each instance, get/set are defined in terms of the record's data constructors, using ‑XDisambiguateRecordFields and friends.

fieldLabel declaration (data dictionary)

There is to be a new declaration type, examples:

    fieldLabel customer_id :: r -> Int
    fieldLabel unitPrice :: (Save r, Num t) => r -> t
    fieldLabel monoField :: Rec -> String   -- equiv to H98 

[fieldLabel is rather long as reserved words go. I'm guessing that field or label would already be heavily used in existing code. Suggestions welcome!]

The fieldLabel declaration desugars to:

    data Proxy_customer_id                  -- phantom, a type 'peg'
    customer_id :: r{ customer_id :: Int } => r -> Int
    customer_id r = get r (undefined :: Proxy_customer_id)

    unit_Price :: (r{ unit_Price :: t}, Save r, Num t) => r -> t

That is: the r{ ... } constraint is added by the desugarer (and will be further desugarred to a Has constraint):

Syntactic sugar for Has

DORF steals from SORF:

    r{ fld :: t }  <===> Has r Proxy_fld t

Using the sugar in the surface syntax (representation) allows for some freedom in the design space 'behind the scenes'.

Should get have a Proxy argument?

I've used a phantom/proxy type (in GHC v 7.2.1) to drive type instancing for Has.

SORF uses a String Kind (which is only partially available with GHC v 7.4.1), with implicit type application (so get does not have a proxy argument). I'll leave it to the implementors to determine which works best.

get is a method of the `Has' class:

    get :: (Has r fld t) => r -> fld -> t

Record declaration

    data Customer_NameAddress = Cust_NA { customer_id :: Int, ... } 

Does _not_ create a field selector function customer_id. Instead it creates a Has instance:

    instance (t ~ Int) 
       => Has Customer_NameAddress Proxy_customer_id t where
               get Cust_NA{customer_id} = customer_id

Note the bare t with type equality constraint. This is unashamedly stolen from SORF's "functional-dependency-like mechanism (but using equalities) for the result type". So type inference binds to this instance based only on the record type and field (type 'peg'), then 'improves' the type of the result. The definition of get uses ‑XDisambiguateRecordFields style (with ‑XNamedFieldPuns).

[It's a wart that in the record declaration, we've had to repeat the type of customer_id when the fieldLabel decl has already stipulated Int. It is legal syntax to omit the type in the record decl, but that currently has a different effect:

    data ... = Cust_NA { customer_id, custName :: String, ... }

currently means customer_id is to be same type as custName.

Opportunity for improvement! ]

Record/field update

Update uses method set from the Has class:

        set :: (Has r fld t) => fld -> t -> r -> r

sets instances are defined using explicit data constructor:

    instance (t ~ String) =>
        Has Customer_NameAddress Proxy_firstName t where
            set _ x (Cust_NA{..}) = Cust_NA{firstName = x, ..}

The definition of set uses ‑XDisambiguateRecordFields style (with ‑XNamedFieldPuns and ‑XRecordWildCards to fill in the un-named fields). Haskell's existing update syntax is desugarred to a call to set:

    myCustNA{ firstName = "John" }
===>  set (undefined :: Proxy_firstName) "John" myCustNA

(That is, with a name or expression preceeding the { ... }. A data constructor prefix continues to use


It is crucial to this proposal that we can implement a polymorphic field update function (set). There's a number of tricky requirements considered below.

In amongst the discussion of dot notation for field selection, there have been aspersions cast on Haskell's current record update syntax.

If we can support update as just a function, there's a chance we can then turn to the syntactic sugar. (For example, the application programmer can develop update idioms to suit their purpose, as just overloaded functions.)

Updating multiple fields

The syntax already supports updating multiple fields:

    myCustNA { firstName = "Fred", lastName = "Dagg" }

The simplest implementation is to turn this into two (nested) updates, but that makes it inefficient generating then discarding the interim result. It may also block updates of linked fields that share a type parametric in the record type. The prototype for this proposal has explored updating with a type instance pairing the two fields:

    instance (t ~ (String, String)) =>
        Has Customer_NameAddress
              (Proxy_firstName, Proxy_lastName) t     where ...

but in general, this would need instances for each perm/comb of fields.

Type-changing update

Haskell 98's record update can change the type of the record, by changing the type of a field that is parametric in the record's type.

There has been some feedback that there are legitimate use-cases for type-changing update -- for example traversing a record structure applying some type-changing transformation.

This proposal does support type changing, but at cost of considerable extra complexity.

So the earlier definitions of Has/get/set have been "economical with the truth". Instead:

    class Has r fld t	where
        get  :: r -> fld -> GetResult r fld t
        set  :: (Has (SetResult r fld t) fld t) =>
                   fld -> t -> r -> SetResult r fld t

The type functions are to handle the possibly-changing types:

    type family GetResult  r fld t   :: *  -- result from get
    type family SetResult  r fld t   :: *  -- result from set

For monomorphic (non-changing) fields, GetResult returns t and SetResult returns r, so this amounts to the simpler definitions for Has/get/set given earlier.

These are type families, not associated types, because in many cases, the result from get depends only on fld, and the result from set depends only on the record type r. In a few cases, the type function must be sensitive to the combination of field type and record type.

The extra Has constraint on set's result is to 'improve' t by gathering constraints from the type of set's resulting record type.

Note that the field value's type t is the type to-be in the result, _not_ the type as-was in the record being updated. So the result from set has that type `inserted'.

Example, based on field unit_Price:

    data Customer_Price a = Num a => Cust_Price {
                                       unit_Price  :: a,
                                       ... }
    type instance GetResult (Customer_Price a) Proxy_unit_Price t
           = a           -- type as is
    type instance SetResult (Customer_Price _a) Proxy_unit_Price t
           = Customer_Price t      -- set record type per arg to set
    instance (Num t) =>
        Has (Customer_Price a) Proxy_unitPrice t        where
            get Cust_Price{unit_Price} _ = unit_Price
            set _ x Cust_Price{..} = Cust_Price{ unit_Price = x, .. }

(The method definitions are 'uninteresting', compared to the drama to get the types right.)

The extra complexity to support changing type could be somewhat reduced using a separate Set class with four type parameters, including both as-was and resulting record types, and equality constraints to improve them -- rather than type family SetResult.

This would mean, though, that the type sugar for Has constraints would not be adequate. Since that sugar is to be visible but the instance definitions are to be 'internal', this proposal prefers to support the sugar.

Selecting polymorphic/higher-ranked fields

Note that initialising records with polymorphic fields (using record constructor syntax) is not affected. This proposal implements selecting/applying those fields in polymorphic contexts. This includes fields with class-constrained types 'sealed' within the record.

To support higher-ranked fields, this proposal follows SORF's approach (with three parameters to Has) to obtain a polymorphic type:

    data HR	= HR {rev :: forall a_. [a_] -> [a_]}   -- per SPJ
    fieldLabel rev :: r -> (forall a_. [a_] -> [a_])
    data Proxy_rev
    rev :: Has r Proxy_rev t => r -> t
    rev r = get r (undefined :: Proxy_rev)
    type instance GetResult r Proxy_rev t = t     -- plain t
                   -- field's type is whatever's there (it's opaque)
                   -- improved by the instance constraint
    type instance SetResult HR Proxy_rev t = HR
                   -- the H-R type is hidded inside HR
    instance (t ~ ([a_] -> [a_])) =>              -- same as SORF
        Has HR Proxy_rev t    where
            get HR{rev} _ = rev

Updating polymorphic/higher-ranked fields

The prototype for this proposal does include a method of updating Higher-ranked fields. SPJ has quickly reviewed the prototype:

"Your trick with SetTy to support update of polymorphic fields is, I belive, an (ingenious) hack that does not scale. I think it works only for fields that are quantified over one type variable with no constraints.
So, I think that update of polymorphic fields remains problematic. "

Note that the "(ingenious)" and unscalable "hack" appears only in compiler-generated code.

Is it a requirement to be able to update polymorphic fields? Is it sufficient to be able to update other (monomorphic) fields in records that also contain poly fields?

Representation hiding/import/export

See the discussion under <Application Programmer's view> and <No Mono Record Fields>. When import/exporting do we need to also export the Proxy_type? If not exported, update syntax cannot be desuggarred to use it.)

Should application programmers declare instances for `Has'/set?

Nothing so far suggests they should. (And there's obvious dangers in allowing it.)

But updating through virtual fields might need it. See <DORF -- comparison to SORF>#<Virtual record selectors>.

Attachments (2)

Download all attachments as: .zip