The Limits of Emacs Advice
Today at work I was using impatient-mode to share some code with Brian. It makes for a really handy live pastebin. To limit the buffer to the relevant code, I narrowed it down with
narrow-to-region. However, the browser wouldn’t update to show only the narrowed region until I made an edit. This makes sense because impatient-mode hooks
after-change-functions. Narrowing the buffer doesn’t change anything in the buffer, so, as expected, this hook is not called.
The solution would be to also join whatever hook is called when the buffer restriction changes. Unfortunately, no such hook exists. I thought I could create this hook with some advice, but this turns out to be currently impossible.
What’s advice? It’s a handy feature of Emacs lisp that allows users to modify the behavior of almost any function without having to redefine it. It works a little bit like methods in the Common Lisp Object System (CLOS): advice is code than can be evaluated before, after, or around a function.
Advice is defined with
defadvice. Duh. For example, say we wanted to be silly and have Emacs say “Ouch!” when a line is killed with
kill-line. We can advise this function to display a message.
(defadvice kill-line (after say-ouch activate) (message "Ouch!"))
This says we want to advise the function
kill-line, we want this advise to execute after
kill-line has run, our advice is named ”
say-ouch”, and we want to immediately activate this advice so it gets used right away. The rest is the body of the advice, like the body of a function. After evaluating this
defadvice, every time I hit
C-k Emacs says “Ouch!” in the minibuffer. Cool!
narrow-to-region and widen
A hook is a variable that holds a list of functions. (Or maybe hooks are the functions in this list? Emacs’ documentation calls both of these things hooks.) These functions are called, usually without arguments, when some specific event occurs. For example, every mode has its own mode hook which is called when the mode is activated in a buffer. This allows users to extend or modify the mode – like by enabling additional minor modes – without editing the mode’s source code directly.
To make our hook work we need to advise
widen to run the hook after they’ve done their work. These are the primitive narrowing functions which all the other narrowing functions eventually call, like
narrow-to-page, and any other mode-specific narrowing. Advising these two functions will cover all buffer narrowing. It should be this simple.
(defvar change-restriction-hook ()) (defadvice narrow-to-region (after hook activate) (run-hooks 'change-restriction-hook)) (defadvice widen (after hook activate) (run-hooks 'change-restriction-hook))
At first this seems to work. I can add a test hook see them activate when I use
M-x narrow-to-region and
M-x widen. However, when I use other narrowing functions, like
narrow-to-defun, my hook functions aren’t called.
Is there a narrowing primitive I missed? I check the source code. Nope, these are lisp functions which ultimately call
narrow-to-region. Is the advice not getting used when called indirectly? I test that out.
(defun foo () (interactive) (narrow-to-region 1 2))
This works fine. Hmmm, these other functions are byte-compiled, maybe that’s the problem.
Bingo. The advice has stopped working. It has something to do with byte-compilation.
Let’s take a look at the bytecode for
(symbol-function 'foo) ;; => #[nil "\300\301}\207" [1 2] 2 nil nil]
I don’t know too much about Emacs’ byte code, but here’s the gist of it. A compiled function is a special type of vector (hence the
# form). This is a legal s-expression which you can use directly in regular Elisp code just like it was a function. The only reason you’d do so is for obfuscation, so it would look very suspicious.
The first element of this function vector is the parameter list – empty in this case. The second is a string containing the actual bytecodes. The rest holds the various constants from the function body. This includes the symbols of other functions called by this function. It’s important to note that
narrow-to-region does not appear in this list!
Curious. Let’s take a closer look at the bytecode.
(coerce (aref (symbol-function 'foo2) 1) 'list) ;; => (192 193 125 135)
bytecomp.el from the Emacs distribution I can see that codes 192 and 193 are used for accessing constants. This pushes my constants 1 and 2 onto a stack for use as function arguments. Next up is 125, which corresponds to
It turns out
narrow-to-region is so special – probably because it’s used very frequently – that it gets its own bytecode. The primitive function call is being compiled away into a single instruction. This means my advice will not be considered in byte-compiled code. Darnit. The same is true for
widen (code 126).
Where to go now?
Since it’s not possible to hook or advise the buffer-narrowing primitives, impatient-mode would need to hook some other event that tends to happen at the same time. Perhaps any time a command is executed in the current buffer it could check for changes to the buffer restriction and, if so, update any attached web clients. I’ll figure something out.blog comments powered by Disqus