Program Reduction: A Win for Recursion Schemes

I first encountered the “recursion schemes” idea in Haskell while attending an excellent presentation given by Tim Williams in 2013. His ideas in that talk made a deep impression on me, and I resolved to find a way to prove the potential he promised in my own work.

Now, the basic idea of recursion schemes goes back at least as far as the 1991 paper entitled Functional Programming with Bananas, Lenses, Envelopes and Barbed Wire, by Meijer, Hughes, Fokkinga and Paterson. In the years since, many excellent blog articles have sprung up that delve neatly into the theory and application of recursion schemes, as well as recorded talks and presentations, and even a library entitled recursion-schemes by the esteemable Edward Kmett.

However, despite my enchantment with the beautiful theory – and especially it’s connections with category theory – it took a while before the promise of these abstraction really paid off. And pay off it did. In fact, what happened in hnix last night was such a revelation, I felt the time had come to tell the story here.

Note that I’m not going to explain recursion schemes here, or how they work. Instead, I recommend you start with an introduction like this one, and then come back to this article once the basics are familiar.

Starting the experiment: hnix

The hnix project, which aims to provide another, complete implementation of the Nix language, began in 2014. Since Nix is itself a pure, functional, lazily evaluated language, it seemed only natural to implement it using Haskell, which also shares these properties.

The first step was to write a parser, to read the Nix source code into an abstract syntax tree (AST) for the purposes of pretty printing and evaluation. Since the language is highly recursive (that is, expressions may contain expressions), what better time to see whether recursion schemes would fit the bill.

This decision was made purely because I liked the idea, and not for any immediate practical advantage. In fact, there is a slight cost when you start using recursion schemes: the management of Fix type wrapper layers, which appear everywhere throughout the resulting tree (though, see the end of this article for a way to avoid this).

For the sake of exposition, here’s what the core expression type for hnix looks like, minus a few of its supporting types. As you can see, there are many places in the structure where recursion is possible, each identified by the type variable r:

The use of Fix here to express recursion allows separating the information stored at each layer of the AST from the recursive structure of that tree. This separation gives us some freedoms that will become clearer in a moment, but essentially the trick is that we now use a generic function cata (for catamorphism) to recurse over syntax trees, passing it a function that defines how each “data layer” separately reduces:

Annotated expression trees

The first moment when this recursion scheme went from an interesting idea to a good design decision was when someone created an issue on GitHub, asking to add location information to the expression tree. Ordinarily, this would have required extending the type of the tree to include such information, but with the scheme above we were able to simply define a new fixed-point type:

Then we had to enrich the parser to create these annotated expressions, by extending what we’d written before for regular ones:

And finally, we wrap each parser for plain expressions so they become parsers for annotated expressions:

Note that the parsing code is the only module that needs to be modified invasively, since it is expressed using direct recursion. This requires us to inject the new data layer (the one containing the annotation information) at each step. However, for algorithms that are not directly recursive, such as the reduction steps passed to cata, there is a better way…

Abstract Definitional Interpreters

While at ICFP 2017, I attended a very interesting lecture titled Abstracting Definition Interpreters (ADI) by David Darais. It showed how recursive code, such as traditional evaluators, could be rewritten to inject new behavior at each layer of the recursion – even changing the logic of the recursion in arbitrary ways.

But wait… aren’t recursion schemes all about sifting data apart from recursion so we can manage the layers in a general way too? Only ADI does it for behavior, while recursion schemes apply the idea to data. Surely there’s a duality to be explored here.

While pondering this, I sat down with Conal Elliott during the break, and together we found a way to unify the two ideas: to apply the ADI approach to catamorphism-based evaluators operating on fixed-point data structures. The result was the following pair of functions:

These functions can used as replacements for cata and cataM, while modify the algorithm they’re passed in f using the transformation given in g. We’ll see how this was put to use in each of the following sections.

Error reporting with context

In the first version of the hnix evaluator, error messages were reported to the user with nice location information, using the trifecta library to show the source context of the error. We later moved to megaparsec to resolve some performance issues, but still the errors were flat in this respect. To capture a sense of where things really go wrong during evaluation, we need to see a trace of all the evaluation steps leading up to that point, similar to what Nix itself offers when you supply the --show-trace option.

This is the first place I thought the adi function could help, and indeed we were able to implement full contextual error reporting in just two lines of code, by extending the original evaluator to append “stack frames” within a MonadReader context around each evaluation step (some of this code has been simplified to omit extraneous detail):

The framedEvalExpr function takes monadic f-algebras for the original expression functor – those including a monad m, so we can use Framed e m to require a MonadReader instance providing a list of frames – and transforms it into an f-algebra for location-annotated expression trees that reports all the locations leading up to an error whenever an exception is thrown.

The rest of the logic happens in throwError, which queries the MonadReader instance for the current list of frames, and reports all the positions to the user including whatever we knew about the expression trees at those points in time. All without adding a single line to the original algorithm, or teaching the existing evaluator anything more about location-annotated trees.

This sort of abstraction also supports a better division of labor: The core evaluator, where most of the work is done, works on the simplest type: trees without any annotations, just Fix ExprF. All of the additional enrichments for things like error reporting happen outside of this logic, reducing the overall complexity by localizing each bit of functionality to its own module.

Adding tracing to any evaluator

The second application of this technique used adiM to introduce both Reader and IO effects, to produce a tracing report during evaluation showing which parts of the tree we’re working on as we go. Again, without modifying the original evaluator, or even knowing anything about which evaluator we end up extending:

Reducing programs to test cases

Finally we come to the motivation for this article. Although the above expressiveness and flexibility was enough to convince me of the potential in the recursion schemes approach, I still hadn’t found its “killer app”: something that recursion schemes is able to make so much easier that it’s well worth whatever boilerplate the technique induces. But yesterday I think I found that example.

A problem with evaluating lazy functional languages, like nix or Haskell, is that it allows for self-referential structures by way of tying the knot. This is great for writing compact and elegant code, but extremely tricky if you happen to get the evaluation semantics wrong, which was the case in hnix yesterday. It ended up that somewhere deep within the evaluation of nixpkgs, I ended up forcing a thunk that I was already in the process of forcing. This means that somewhere in the evaluation hnix was either being too eager, or I’d gotten the scoping wrong and a self-reference was occurring where it shouldn’t have.

Either way, it ended up proving very difficult to delve deep into the thousands of lines of highly recursive, lazy, higher-order code. How was I going to find the root cause of the problem?

It occurred to me that even though the expression tree involved were massive, I’d only evaluated a fraction of it before encountering the bug. So why can’t I just output that fragment along with the failure, to make it easier to see what was actually involved in producing the problem? The algorithm seemed simple enough:

  1. Start with an unadorned expression tree.

  2. Annotate it with mutable booleans at every layer, to indicate whether we’ve forced the evaluation of that layer or not (i.e., whether it contributed to the final outcome).

  3. Strip away from the tree anything that was never referenced.

  4. Further compact the tree based on the logical consequences of step 3. For example, if we force the true branch of an if, but never the false, we can replace the if statement expression with just the true branch.

The tree that results from this winnowing should yield exactly the same behavior, but be potentially much smaller and simpler. After all, there are over 1.2 million lines of code already in nixpkgs, and it’s hard to know due to the lazy nature of Nix how much of it we actually touched during evaluation. There has to be a better way!

At first I thought this would be a typical hard problem: That is, easy to imagine a solution for, but many long hours of elbow grease to make it happen. I logged a bug in the tracker describing the idea, wondering how many days it would take to realize, and how much the code would have to change to make it possible.

Two hours later, it was working: thanks to both recursion schemes and abstract definitional interpreters.

This was accomplished by first defining the type of boolean-flagged trees, which extend whatever other kind of tree (given by the functor f) we might be working with:

Then a function in IO that takes a given expression tree, and enriches it with all the booleans set to False, to mean unvisited:

Of course we also need a way to strip away the annotations later. cata makes this one really easy:

Now using adiM we can fold in the IO monad, allowing us to toggle these IORef boolean as we evaluate. For these function we need to know the types of the trees involved, so that we can reduce this appropriately in pruneTree.

That I can insert this whole function here in a blog post is, to me, a testament to the power of the abstractions involved. This is all we needed to color the tree with the locations where we actually performed evaluation!

The last step is to cull the tree of its dead wood, by applying logical transforms wherever lack of evaluation implies a reduction in the size of the tree. This function is the only genuinely complex part and is a bit too long to include here. But it’s also a pure function, making it easier to verify and test. Here’s an excerpt of what it looks like:

And that’s pretty much it. Now I can input a program like this:

And automatically reduce it to a smaller program producing the same result:

It even works for my huge thunk recursion problem in nixpkgs, reducing that large amount of code spread across many files, to a single file containing less than 10k lines of code with very few imports remaining (those that could not be statically determined).

Conclusion

In the four years I’ve been using recursion schemes in this project and elsewhere, I haven’t really encountered many downsides, while reaping many benefits. The worst aspect is definitely the extra boilerplate needed to define and work with the types (see my parsec-free library for a rather extreme case of this).

However, using modern GHC with bidirectional pattern synonyms, you can easily hide this boilerplate from your users. For example, if you’re using the Free type instead of Fix – because your functor doesn’t have its own base case – you would have code like that follows this general pattern:

In conclusion: I highly recommend the recursion schemes approach. Even if you start using it just because it sounds cool, it’s a sound design decision to include at the beginning of a project that could potentially lead to big wins down the line. Had the hnix project started with traditional recursive types, it would have much harder at this point to retrofit it with fixed-points of functors; but since that decision was made from the start, we’re able to keep extending its functionality in ways like those I’ve described above, but where each new layer of functionality is almost entirely self-contained and separately testable and reviewable.