MLton 20070826 FunctionalRecordUpdate
Home  Index  
Functional record update is the copying of a record while replacing the values of some of the fields. Standard ML does not have explicit syntax for functional record update. We will show below how to implement functional record update in SML, with a little boilerplate code.

As an example, the functional update of the record

{a = 13, b = 14, c = 15} 
with c = 16 yields a new record
{a = 13, b = 14, c = 16}
Functional record update also makes sense with multiple simultaneous updates. For example, the functional update of the record above with a = 18, c = 19 yields a new record
{a = 18, b = 14, c = 19}

One could easily imagine an extension of the SML that supports functional record update. For example

e with {a = 16, b = 17}
would create a copy of the record denoted by e with field a replaced with 16 and b replaced with 17.

Since there is no such syntax in SML, we now show how to implement functional record update directly. We first give a simple implementation that has a number of problems. We then give an advanced implementation, that, while complex underneath, is a reusable library that admits simple use.

Simple implementation

To support functional record update on the record type

{a: 'a, b: 'b, c: 'c} 
first, define an update function for each component.
fun withA ({a = _, b, c}, a) = {a = a, b = b, c = c}
fun withB ({a, b = _, c}, b) = {a = a, b = b, c = c}
fun withC ({a, b, c = _}, c) = {a = a, b = b, c = c}
Then, one can express e with {a = 16, b = 17}  as
withB (withA (e, 16), 17)
With infix notation
infix withA withB withC
the syntax is almost as concise as a language extension.
e withA 16 withB 17

This approach suffers from the fact that the amount of boilerplate code is quadratic in the number of record fields. Furthermore, changing, adding, or deleting a field requires time proportional to the number of fields (because each with function must be changed). It is also annoying to have to define a with function, possibly with a fixity declaration, for each field.

Fortunately, there is a solution to these problems.

Advanced implementation

Using Fold one can define a family of makeUpdate<N> functions and single update operator U so that one can define a functional record update function for any record type simply by specifying a (trivial) isomorphism between that type and a product type. For example, suppose that we would like to do functional record update on records with fields a and b. Then one defines a function updateAB as follows.

val updateAB =
   fn z =>
   let
      fun p2r (v1 & v2) = {a = v1, b = v2}
      fun r2p {a = v1, b = v2} = (v1 & v2)
   in
      makeUpdate2 (p2r, p2r, r2p)
   end
   z

The functions p2r (think product to record) and r2p (think record to product) specify an isomorphism between a,b records and binary products. There is a second use of p2r to work around the lack of first-class polymorphism in SML.

With the definition of updateAB in place, the following expressions are valid.

updateAB {a = 13, b = "hello"} (U#b "goodbye") $
updateAB {a = 13.5, b = true} (U#b false) (U#a 12.5) $

As another example, suppose that we would like to do functional record update on records with fields b, c, and d. Then one defines a function updateBCD as follows.

val updateBCD =
   fn z =>
   let
      fun p2r (v1 & v2 & v3) = {b = v1, c = v2, d = v3}
      fun r2p {b = v1, c = v2, d = v3} = (v1 & v2 & v3)
   in
      makeUpdate3 (p2r, p2r, r2p)
   end
   z

With the definition of updateBCD in place, the following expression is valid.

updateBCD {b = 1, c = 2, d = 3} (U#c 4) (U#c 5) $

Note that not all fields need be updated and that the same field may be updated multiple times. Further note that the same U operator is used for all update functions (in the above, for both updateAB and updateBCD).

In general, to define a functional-record-update function on records with fields f1, f2, ..., fN, use the following template.

val update =
   fn z =>
   let 
      fun p2r (v1 & v2 & ... & vn) = {f1 = v1, f2 = v2, ..., fn = vn}
      fun r2p {f1 = v1, f2 = v2, ..., fn = vn} = (v1 & v2 & ... & vn)
   in
      makeUpdateN (p2r, p2r, r2p)
   end
   z

With this, one can update a record as follows.

update {f1 = v1, ..., fn = vn} (U#fi1 vi1) ... (U#fim vim) $

If makeUpdateN is not already defined for the desired N, a generic makeUpdate function and special value, A, is defined so that one can use the following for makeUpdateN, where A is repeated N times.

makeUpdate A ... A $

The FunctionalRecordUpdate structure

Here is the implementation of functional record update.

structure FunctionalRecordUpdate =
   struct
      datatype ('x, 'y) u = X of 'x | Y of 'y

      val makeUpdate =
         fn z =>
         Fold.fold
         (((), (),
           fn f => f o X,
           fn (a, u) => case u of X x => x | _ => a),
          fn (p, up, _, _) => fn (p2r, p2r', r2p) => fn r =>
          Fold.fold ((p2r' (p id), up, r2p r),
                     fn (_, _, p) => p2r p))
         z
         
      val A =
         fn z =>
         Fold.step0
         (fn (_, _, p, up) =>
          (p, up, fn f => p (f o X) & (f o Y),
           fn (a & b, u) =>
           (case u of X x => up (a, x) | _ => a)
           & (case u of Y y => y | _ => b)))
         z

      fun makeUpdate2 z = makeUpdate A A $ z
      fun makeUpdate3 z = makeUpdate A A A $ z
      fun makeUpdate4 z = makeUpdate A A A A $ z

      fun U s v = Fold.step0 (fn (r, up, p) => (r, up, up (p, s r v)))
   end

The idea of makeUpdate is to inductively build the update function for n-ary product types. Each A supplied to makeUpdate adds one more level to the product. When finished with its arguments, makeUpdate begins a second fold, this time to process a variable number of U steps. The second fold begins by converting the supplied record to a product, using the supplied isomorphism (p2r'). Each step works by selecting a "path", s r v), from the inductively constructed product, reformatted by the supplied isomorphism to look like a record. Then, the inductively constructed update function is applied to the record-as-product and the path up (p, s r v) to yield a new record-as-product. Finally, at the end of the fold, the product is converted back to a record using the supplied isomorphism (p2r).

Efficiency

With MLton, the efficiency of this approach is as good as one would expect with the special syntax. Namely a sequence of updates will be optimized into a single record construction that copies the unchanged fields and fills in the changed fields with their new values.


Last edited on 2007-08-15 22:06:30 by MatthewFluet.