Emacs Lisp Readable Closures

I’ve stated before that one of the unique features of Emacs Lisp is that its closures are readable. Closures can be serialized by the printer and read back in with the reader. I am unaware of any other programming language that has this feature. In fact it’s essential for Elisp byte-code compilation because byte-compiled Elisp files are merely s-expressions of byte-code dumped out as source.

Lisp Printing

The Lisp family of languages are homoiconic. Lisp source code is written in the syntax of its own data structures, s-expressions. Since a compiler/interpreter is usually provided at run-time, a consequence of this is that reading and printing are a fundamental feature of Lisps. A value can be handed to the printer, which will serialize the value into an s-expression as a sequence of characters. Later on the reader can parse the s-expression back into an equal value.

To compare, JavaScript originally had half of this in place. JavaScript has convenient object syntax for defining an associative array, known today as JSON. The eval function could (dangerously) be used as a reader for parsing a string containing JSON-encoded data into a value. But until JSON.stringify() became standard, developers had to write their own printer. Lisp s-expression syntax is much more powerful (and complicated) than JSON, maintaining both identity and cycles (e.g. *print-circle*).

Not all values can be read. They’ll still print (when *print-readably* is nil) but will do so using special syntax that will signal an error in the reader: #<. For example, in Emacs Lisp buffers cannot be serialized so they print using this syntax.

(prin1-to-string (current-buffer))
;; => "#<buffer *scratch*>"

It doesn’t matter what’s between the angle brackets, or even that there’s a closing angle bracket. The reader will signal an error as soon as it hits a #<.

Almost Everything Prints Readably

Elisp has a small set of primitive data types. All of these primitive types print readably:

Here are all the non-readable types. Each one has a good reason for not being serializable.

And that’s it. Every other value in Elisp is constructed from one or more of these primitives, including keymaps, functions, macros, syntax tables, defstruct structs, and EIEIO objects. This means that as long as these values don’t refer to an unreadable value, they themselves can be printed.

An interesting note here is that, unlike the Common Lisp Object System (CLOS), EIEIO objects are readable by default. To Elisp they’re just vectors, so of course they print. CLOS objects are unreadable without manually defining a print method per class.

Elisp Closures

Elisp got lexical scoping in Emacs 24, released in June 2012. It’s now one of the relatively few languages to have both dynamic and lexical scope. Like Common Lisp, variables declared with defvar (and family) continue to have dynamic scope. For backwards compatibility with old Lisp code, lexical scope is disabled by default. It’s enabled for a specific file or buffer by setting lexical-binding to non-nil.

With lexical scope, anonymous functions become closures, a powerful functional programming primitive: a function plus a captured lexical environment. It also provides some performance benefits. In my own tests, compiled Elisp with lexical scope enabled is about 10% to 15% faster than with the default dynamic scope.

What do closures look like in Emacs Lisp? It takes on two forms depending on whether the closure is compiled or not. For example, consider this function, foo, that takes two arguments and returns a closure that returns the first argument.

;; -*- lexical-binding: t; -*-
(defun foo (x y)
  (lambda () x))

(foo :bar :ignored)
;; => (closure ((y . :ignored) (x . :bar) t) () x)

An uncompiled closure is a list beginning with the symbol closure. The second element is the lexical environment, the third is the argument list (lambda list), and the rest is the body of the function. Here we can see that both x and y have been “closed over.” This is a little bit sloppy because the function never makes use of y. Capturing it has a few problems.

Fortunately the compiler is smart enough to see this and will avoid capturing unused variables. To prove this, I’ve now compiled foo so that it returns a compiled closure.

(foo :bar :ignored)
;; => #[0 "\300\207" [:bar] 1]

What’s returned here is a byte-code function object, with the #[...] syntax. It has these elements:

  1. The function’s lambda list (zero arguments)
  2. Byte-codes stored in a unibyte string
  3. Constants vector
  4. Maximum stack space needed by this function

Notice that the lexical environment has been captured in the constants vector, specifically noting the lack of :ignored in this vector. The compiler didn’t capture it.

For those curious about the byte-code here’s an explanation. The string syntax shown is in octal, representing a string containing two bytes: 192 and 135. The Elisp byte-code interpreter is stack-based. The 192 (constant 0) says to push the first constant onto the stack. The 135 (return) says to pop the top element from the stack and return it.

(coerce "\300\207" 'list)
;; => (192 135)

The Readable Closures Catch

Since closures are byte-code function objects, they print readably. You can capture an environment in a closure, serialize it, read it back in, and evaluate it. That’s pretty cool! This means closures can be transmitted to other Emacs instances in a multi-processing setup (i.e. Elnode, Async)

The catch is that it’s easy to accidentally capture an unreadable value, especially buffers. Consider this function bar which uses a temporary buffer as an efficient string builder. It returns a closure that returns the result. (Weird, but stick with me here!)

(defun bar (n)
  (with-temp-buffer
    (let ((standard-output (current-buffer)))
      (loop for i from 0 to n do (princ i))
      (let ((string (buffer-string)))
        (lambda () string)))))

The compiled form looks fine,

(bar 3)
;; => #[0 "\300\207" ["0123"] 1]

But the interpreted form of the closure has a problem. The with-temp-buffer macro silently introduced a new binding — an abstraction leak.

(bar 3)
;; => (closure ((string . "0123")
;;              (temp-buffer . #<killed buffer>)
;;              (n . 3) t)
;;      () string)

The temporary buffer is mistakenly captured in the closure making it unreadable, but only in its uncompiled form. This creates the awkward situation where compiled and uncompiled code has different behavior.

Have a comment on this article? Start a discussion in my public inbox by sending an email to ~skeeto/public-inbox@lists.sr.ht [mailing list etiquette] , or see existing discussions.

null program

Chris Wellons

wellons@nullprogram.com (PGP)
~skeeto/public-inbox@lists.sr.ht (view)