24 Days of GHC Extensions: Recursive Do

Ordinarily, when we’re programming computations under a monad, we’re limited to forming bindings in the order they appear in a do block. However, there are a class of monads that support extra value recursion through the type class MonadFix. This extra functionality has special syntax support in GHC, and can be useful in a variety of situations.

MonadFix and Recursive Bindings

First of all, we’ll take a look at how MonadFix can be used to easily create recursive data structures. Before we get to that, lets look at a little program that works with rose trees. We’ll begin with the definition of a rose tree:

{-# LANGUAGE RecursiveDo #-}
import Control.Monad.Fix
data RoseTree a = RoseTree a [RoseTree a]
  deriving (Show)

A rose tree is a node paired up with a list of child trees. Thus this data structure is a simple model of a multi-way tree. One thing we’d like to be able to do with our rose trees is to pair each element in the tree with the largest element in the entire tree. For example, if we have the following:

exampleTree :: RoseTree Int
exampleTree = RoseTree 5 [RoseTree 4 [], RoseTree 6 []]

then we can pair each item in the tree with the biggest element - 6:

.> pureMax exampleTree
RoseTree (5, 6) [RoseTree (4, 6) [], RoseTree (4, 6) []]

On the surface, we might think that this operation should take two traversals of the tree - one to find out what the largest element is, and then another pass through the tree to change all the elements. Interestingly, this can actually be done in a single pass, if we exploit laziness! Here’s how:

pureMax :: Ord a => RoseTree a -> RoseTree (a, a)
pureMax tree =
  let (t, largest) = go largest tree
  in t
 where
  go :: Ord a => a -> RoseTree a -> (RoseTree (a, a), a)
  go biggest (RoseTree x []) = (RoseTree (x, biggest) [], x)
  go biggest (RoseTree x xs) =
      let sub = map (go biggest) xs
          (xs', largests) = unzip sub
      in (RoseTree (x, biggest) xs', max x (maximum largests))

If you’ve not seen this before - this may seem a little cryptic! Let’s take it slowly. We walk through our tree with the go function. go takes two paremeters - one is the tree that we’re working on, while the other is the known largest element. But wait… don’t we need to calculate the largest element? You’re right - and that’s exactly what we do in go.

Notice that go returns two values - the relabelled rose tree, along with the known largest element in that tree. The magic happens right at the start of pureMax - notice that we pattern match to bind the variable largest, though we also feed in largest to relabel the trees.

What’s happening is that we’re exploiting lazy evaluation - what we actually do is relabel the tree with a thunk at every node. Once we’ve finished recursing the entire tree, we’ve got enough information, should we need to force that node. This technique is known as circular programming, and Richard Bird has a lovely write up on a similar problem - the repmin problem.

So far, we’ve seen a pure solution to the problem, but what happens if we need to use effectful computations to determine the largest value? For example, let’s work with an exclusive secret santa. In this exclusive secret santa, people can invite others into the game - forming a tree of invites. Furthermore, we also have a database lookup function that will tell us what their budget is. We’d like to be able to relabel the invite tree with the minimum budget for fairness. If we had a tree of budgets, that would be easy - we could just use a variation of pureMax above. However, determining the budget requires a database lookup.

It looks like we’re stuck, but what we can do is make an effectful variation of pureMax:

impureMin :: (MonadFix m, Ord b) => (a -> m b) -> RoseTree a -> m (RoseTree (a, b))
impureMin f tree = do
  rec (t, smallest) <- go smallest tree
  return t
 where
  go smallest (RoseTree x []) = do
    b <- f x
    return (RoseTree (x, smallest) [], b)

  go smallest (RoseTree x xs) = do
    sub <- mapM (go smallest) xs
    b <- f x
    let (xs', bs) = unzip sub
    return (RoseTree (x, smallest) xs', min b (minimum bs))

If you compare this to pureMax you should notice that the programs are very similar. Infact, all we’ve had to do is replace the pure let x = y bindings with effectful x <- y bindings, and call out to our effectful function. Finally, the magic sauce at the top is to use a new piece of syntax rec. rec comes from the RecursiveDo extension, and while the details are beyond the scope of this post (there’s a whole thesis on it!), we can see here that it serves the same purpose as forming recursive bindings, as we did with let.

To wrap up this example, let’s work with some test data and see how it all plays out:

budget :: String -> IO Int
budget "Ada"      = return 10 -- A struggling startup programmer
budget "Curry"    = return 50 -- A big-earner in finance
budget "Dijkstra" = return 20 -- Teaching is the real reward
budget "Howard"   = return 5  -- An frugile undergraduate!
inviteTree = RoseTree "Ada" [ RoseTree "Dijkstra" []
                            , RoseTree "Curry" [ RoseTree "Howard" []]
                            ]

Now we can ask our system what budget we should use

.> impureMin budget inviteTree
RoseTree ("Ada",5) [RoseTree ("Dijkstra",5) [],RoseTree ("Curry",5) [RoseTree ("Howard",5) []]]

Alright, 5 gold coins it is!

In this post I’ve shown just one of many things that can be done with RecursiveDo syntax. Another common usage of this extension is in reactive-banana - here it’s possible for reactive behaviors to depend on events, where the events sample the same behavior. A more extreme example (if you really want a headache!) is the Tardis monad - a monad where data can travel forwards and backwards in time!

Edits:


This post is part of 24 Days of GHC Extensions - for more posts like this, check out the calendar.


You can contact me via email at ollie@ocharles.org.uk or tweet to me @acid2. I share almost all of my work at GitHub. This post is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Unported License.