[Haskell-cafe] PROPOSAL: Record field type inference

[John Meacham]

This is also available as html at

http://repetae.net/computer/jhc/record_inference.html

Record Type Inference
=====================

An extension to the named field mechanism that will greatly enhance the
utility of them when combined with the existing `DisambiguateRecordFields`,
`RecordPuns`, and `RecordWildCards`.

The proposal is to allow the types of record fields to be inferred by the
normal type inference engine. It would look like

~~~~ {.haskell}
data Rec = Rec {fieldA,fieldB,fieldC}

f Rec { .. } = Rec { .. } where
    fieldA = True
    fieldB = 4
~~~~

This would infer the types `Bool`, `Int`, and `forall a . a` for the fields of
the record constructor and `f :: Rec -> Rec` for f. For the purposes of type
checking the fields are treated as monomorphic and not generalized but
defaulted like normal after typechecking the module. Other than infering the
types of the record fields, the records have the normal syntax. The extensions
`RecordPuns`, `RecordWildCards` and `DisambiguateRecordFields` will be enabled
when record field inference is enabled.

Selector functions will not be created for infered records, as in, the names
are field labels and not functions. This means they do not share a namespace
with functions and do not conflict with each other. Multiple records may have
the same field names in the same module. This means the following is fine.

~~~~ {.haskell}
data Rec1 = Rec1 {input, withFoo, withoutFoo }
data Rec2 = Rec2 {input, withBar, withoutBar }

f Rec1 { .. } = case input of
    [] -> Rec1 { .. }
    (x:xs) -> if hasFoo x
        then Rec1 { withFoo = x:withFoo, .. }
        else Rec1 { withoutFoo = x:withoutFoo, .. }
~~~~

Possible extensions
-------------------

### as-pattern disambiguation

In order to make the disambiguation of record fields more useful without
relying on the type checker for disambiguation, We can declare that variables
explicitly bound to a constsructor in a pattern match use that constructor to
disambiguate fields for operations on the variable. This is a purely syntactic
transformation that can happen before typechecking. It can be used as follows.

~~~~ {.haskell}
-- use the input bound by a Rec1 to update the input bound by a Rec2
f r1 at Rec1 { input } r2 at Rec2 {} = case input of
    xs | any hasBar xs = f r1 { input = [] } r2 { input }
~~~~

### Field label inference

It is concievable that we may want to infer the fields themselves of a record,
as in:

~~~~ {.haskell}
-- infer that R has the field labels bob and susan
data R = R { ..}
f x at R {bob} = R {susan = bob}
~~~~

In order to implement this, a pass through the file will collect every field
label that is used with an explicit R constructor and treat the record as if
it were declared with those names as infered fields.