recur

Created: 2012-03-24 00:56
Updated: 2018-09-14 20:42

recur-defun and recur-let

TL;DR

This is an extension to elisp which lets you write limited recursion without blowing the stack. Eg:

``````(recur-defun* my-prod (list &optional (acc 1))
(if list (recur (cdr list) (* acc (car list)))
acc))
(my-prod (list 1 2 3 4))
``````

Can calculate the product of a list of elements without ever blowing the stack. These forms expand to loops. The recursion is limited to self-recursion.

It is completely stand alone (although it requires `cl.el`), which is the major improvement over the last version, which required all of my trashy utils.

Context

When I first encountered clojure's limited tail call optimization, in the form of "recur" forms inside functions and loops, I was a bit disappointed. Even though (obviously) I write most of my lisp code in hoary old emacs lisp, I think I lean towards the scheme side of things in a lot of respects, and I like writing functions recursively when possible. Lots of people disagree, but I find this is a clear, concise and error avoiding way of specifying lots of algorithms. I find loops, on the other hand, error prone and ugly for many kinds of tasks, although the Common Lisp Loop Macro is pretty nice for lots of simple cases.

However, if you read the above blog post, you'll notice some nuance to my opinion which eventually made me warm to the `recur` form in clojure. Since the stack is a limited resource in most programming systems for reasons of optimization, and since tail calls don't grow the stack, and are therefore something like "meta-semantically" different than non-tail calls, I rather like that Clojure forces you to explicitly mark an attempted tail call, and barfs at you if you place it in a non-tail location. This seems like a good idea, even if you lose a bit of elegance in not using the name of the function itself to recur. Reflection on this theme eventually reveals that the real limitation of `recur` in Clojure is that it only allows self-recursion. A conforming scheme optimizes all tail calls, regardless of their location. In practice, lots of algorithms depend on self-recursion only (but see the Ackerman numbers for a standard counter example of a sort - they are usually computed with several mutually recursive functions), and since the JVM doesn't support complete tail call optimization, `recur` represents an elegant compromise between slow, trampoline based full optimization, and relying entirely on loops or other ad-hoc iteration techniques.

I promise I am getting somewhere.

Once again I am faced with a comparison between Clojure and Emacs Lisp, because both are Lisps which find themselves living in somewhat primitive run-times. Emacs Lisp is somewhat constrained by legacy considerations (although I don't rightly know why the EL engine doesn't optimize tail calls - if anyone does, please contact me) and frankly, somewhat outmoded design. Clojure, similarly saddled with the JVM, shows considerable insight and care in its design, providing modern, powerful features to a difficult environment. `Recur` is one such feature.

I wanted to add this capability to Emacs Lisp because I kept running into places in other projects where I was writing big, gross loops where a recursive implementation would have been easier to write and read. Technically, I've already done this since I've ported all of Clojure's `defn` and `loop\let` destructuring bind forms into Elisp, and these forms provide for the use of the `recur` keyword without growing the stack. Read about those features here. However, the "Clojure-like emacs" project was one of my first very large forays into Emacs Lisp and Programming Language Implementation, and even though I frequently use it, it is a little slow (unless one byte-compiles it) and frightening to use. The implementation is crufty and needs an overhaul.

So I decided I'd re-implement the features I wanted in a more Emacs Lisp flavored way. This library provides those features in two forms. One is a `recur-let` form which allows recursive calls on a let's binding forms, similar to Scheme's `named-let` feature. The other is a `recur-defun*` form, which allows you to declare a function (with a common lisp-like lambda-list) which can `recur` to itself using the recur keyword. Neither forms grow the stack during their pseudo-recursion. It is quite possible to write infinite loops, for instance.

Examples Are In Order

Here is an example of `recur-let`

``````(require 'recur)
(recur-let ((x 0))
(if (< x 10) (recur (+ x 1)) x))
``````

This evaluates to 10.

Here is an example of `recur-defun*`

``````(recur-defun* my-prod (list &optional (acc 1))
(if list (recur (cdr list) (* acc (car list)))
acc))
(my-prod (list 1 2 3 4))
``````

The second form evaluates, as it should, to 24. (Thanks to Joseph Gay for the correction).

Conclusions

This implementation is much, much simpler than the implementation associated with `defn` and its kin from my clojure library. Although this is a very early version, I think the library is stable enough to use for real things.

Most of the complexity here was not even related to the implementation of recursion itself, but with parsing lambda-lists, for which emacs seems to have no native functions. I am sure I could have used some piece of `cl.el` code, but a cursory search didn't reveal anything.

Most of the code is commented. I hope someone out there finds this useful. If you are interested in code-walking macros, this might be a good example to get a sense for how they work and what they do. The critical part of this implementation is a codewalker which searches for `recur` and expands it appropriately or throws an error if we aren't in tail position.