Plainly partible

What is partible?

Partible types are specific forms of pseudodata (a generalisation of oracles) whose values satisfy the following properties:

• they are all unique: no two values will ever be the same;

• they are monousal: if it is used, each value can only be used once;

• their splitting is disjoint: the resulting new values are independent.

Why splittable isn't always enough

Further developments

Since its advent, pseudodata (or aspects thereof) have appeared, or can be recognised in other contexts:

• In Simon Peyton Jones's The implementation of functional programming languages (section 9.6 on page 188 of 458), Peter Hancock provides a simpler interface for generating new type variables (of type tvname) for a type checker, using the type name_supply:

|| page 188 of 458

next_name :: name_supply -> tvname

deplete :: name_supply -> name_supply

split :: name_supply -> (name_supply, name_supply)

The crucial point here is the absence of trees - they have been reduced to an implementation detail, oblivious to the users of name_supply values.

• In Efficient and Portable Combined Random Number Generators, Pierre L'Ecuyer suggests the disjoint splitting of random numbers into independent subsequences as needed.

• As previously specified, if pseudodata is used then it remains constant - reusing it doesn't change its value. Lennart Augustsson, Mikael Rittri and Dan Synek take this to an extreme in their functional pearl On generating unique names with their concept implementation for a single-use variant of Hancock's unique-name supply - each one can only be used once, if at all:

abstype uniquesupply

with

new_uniquesupply :: uniquesupply

split_uniquesupply :: uniquesupply -> (uniquesupply, uniquesupply)

get_unique :: uniquesupply -> unique

uniquesupply ::= US

new_uniquesupply = US

split_uniquesupply US = (US, US)

get_unique s = gensym(s)

unique == int

|| Not a regular definition!

gensym :: * -> unique

In contrast to the example by John Launchbury and Simon Peyton Jones in State in Haskell (see pages 39-40 of 51), this monousal strategy completely obviates the need for trees (or other intermediary structured values such as streams).

• Nobuo Yamashita uses a single-use type similar to pseudodata in his IO-alternative oi package: see the Data.OI.Internal module for the details.

A matter of nomenclature

As mentioned earlier, L'Ecuyer suggests the splitting of random numbers be disjoint. But for entities like unique-name supplies, disjoint splitting is an absolute necessity! To avoid having to repeatedly specify it, an alternate terminology is needed - one which clearly indicates that for some types of pseudodata, the disjointedness of its splitting is mandatory, instead of just being very convenient.

The Partible class

Depending on how its corresponding partible type is defined, the disjoint splitting of an unused value can either be a pair or list of new values:

part_uniquesupply :: uniquesupply -> (uniquesupply, uniquesupply)

|| or ||

parts_uniquesupply :: uniquesupply -> [uniquesupply]

As each definition can be defined with the other:

part_uniquesupply u = (u1, u2) where u1:u2:_ = parts_uniquesupply u

|| or ||

parts_uniquesupply u = u1 : parts_uniquesupply u2 where (u1, u2) = part_uniquesupply u

they can both be overloaded in Haskell using default definitions:

class Partible a where
    part :: a -> (a, a)
    parts :: a -> [a]

     -- Minimal complete definition: part or parts
    part u = case parts u of u1:u2:_ -> (u1, u2)
    parts u = case part u of (u1, u2) -> u1 : parts u2

(Of course if it's more efficient to do so, part and parts can both be defined.)

Ideally, each partible type in Haskell should also satisfy the partible laws.

Examples

• Yamashita's single-use type can be made partible:

data OI                 = OI OI#

instance Partible OI where
    part = partOI

partOI                  :: OI -> (OI, OI)
partOI (OI h)           = case part# h of (# h1, h2 #) -> (OI h1, OI h2)

runOI                   :: (OI -> a) -> IO a
runOI g                 = IO $ \s -> case dispense# s of
                                       (# s', h #) -> seq# (g (OI h)) s'

invokes                 :: Monomo a => String -> IO a -> OI -> a
(name `invokes` IO act) (OI h)
                        = (name `invokes#` act) h

class Monomo a

 -- local definitions --
 --
type OI#                = String -> State# RealWorld

part#                   :: OI# -> (# OI#, OI# #)
part# h                 = case h "partOI" of
                            s -> case dispense# s of
                                   (# s', h1 #) ->
                                     case dispense# s' of
                                       (# _, h2 #) -> (# h1, h2 #)

dispense#               :: IO# OI#
dispense# s             = case newMutVar# () s of
                            (# s', r #) -> (# s', expire# s' r #)

expire#                 :: State# s -> MutVar# s () -> String -> State# s
expire# s r name        = case atomicModifyMutVar# r use s of
                            (# s', () #) -> s'
                          where
                              use x   =  (error nowUsed, x)
                              nowUsed =  name' ++ ": already expired"
                              name'   =  if all isSpace name then "(unknown)"
                                          else name

invokes#                :: Monomo a => String -> IO# a -> OI# -> a
(name `invokes#` act) h = case act (noDuplicate# (h name)) of (# _, t #) -> t

type IO# a              = State# RealWorld -> (# State# RealWorld, a #)

Some notes:

• the elementary reuse-error reporting is optional;
• the use of the Monomo in invokes leverage Haskell's type system to provide an extra measure of safety, by restricting any type-polymorphism in the result: for more information, look into the history of Standard ML.

• Using the reformulated OI type, an unique-name supply is easily defined:

data Fresh a = Fresh (OI -> a) OI

instance Partible (Fresh a) where
    parts (Fresh g u) = [ Fresh g v | v <- parts u ]

afresh :: (OI -> a) -> OI -> Fresh a
afresh g u = Fresh g u

fresh :: Fresh a -> [a]
fresh (Fresh g u) = [ g v | v <- parts u ]

instance Functor Fresh where
    fmap f (Fresh g u) = Fresh (f . g) u

• Instances for various standard Haskell types are also a simple matter:

instance (Ix a, Partible b) => Partible (Array a b) where
    part arr = case unzip (map part' (assocs arr)) of
                 (al1, al2) -> (new al1, new al2)
               where
                   new          = array (bounds arr)
                   part' (i, u) = case part u of
                                    (u1, u2) -> ((i, u1), (i, u2))

instance (Partible a, Partible b) => Partible (Either a b) where
    parts (Left u)  = map Left (parts u)
    parts (Right v) = map Right (parts v)

instance (Partible a, Partible b) => Partible (a, b) where
    parts (u, v) = zip (parts u) (parts v)

instance (Partible a, Partible b, Partible c) => Partible (a, b, c) where
    parts (u, v, w) = zip3 (parts u) (parts v) (parts w)

instance (Partible a, Partible b, Partible c, Partible d) => Partible (a, b, c, d) where
    parts (u, v, w, x) = zip4 (parts u) (parts v) (parts w) (parts x)

instance (Partible a, Partible b, Partible c, Partible d, Partible e) => Partible (a, b, c, d, e) where
    parts (u, v, w, x, y) = zip5 (parts u) (parts v) (parts w) (parts x) (parts y)

 -- etc.

No list or Maybe instances

The unit type () is clearly not partible, because of its single value:

-- instance Partible () where part () = ((), ())  {- WRONG! -}

Therefore, because of their void values:

• [] :: [a]
  

• Nothing :: Maybe a
  

instances for for the list or Maybe types are at best dubious:

instance Partible a => Partible [a] where
    part [] = ([], [])  -- !?
            .
            .
            .

instance Partible a => Partible (Maybe a) where
    part Nothing = (Nothing, Nothing)  -- ?!
            .
            .
            .

An alternative is to repurpose their non-void values to form a new type e.g:

data Some a = Only a | More a (Some a)

instance Partible a => Partible (Some a) where
    parts (Only u)    = map Only (parts u)
    parts (More u us) = zipWith More (parts u) (parts us)