emacs - How to keep dir-local variables when switching major modes?

Question

I'm committing to a project where standard indentations and tabs are 3-chars wide, and it's using a mix of HTML, PHP, and JavaScript. Since I use Emacs for everything, and only want the 3-char indentation for this project, I set up a ".dir-locals.el" file at the root of the project to apply to all files/all modes under it:

; Match projets's default indent of 3 spaces per level- and don't add tabs
(
 (nil .
        (
         (tab-width . 3)
         (c-basic-offset . 3)
         (indent-tabs-mode . nil)
         ))
 )

Which works fine when I first open a file. The problem happens when switching major modes- for example to work on a chunk of literal HTML inside of a PHP file. Then I lose all the dir-local variables.

I've also tried explicitly stating all of the modes I use in ".dir-locals.el", and adding to my .emacs file "dir-locals-set-class-variables / dir-locals-set-directory-class". I'm glad to say they all behave consistently, initially setting the dir-local variables, and then losing them as I switch the major mode.

I'm using GNU Emacs 24.3.1.

What's an elegant way of reloading dir-local variables upon switching a buffer's major-mode?

-- edit -- Thanks for the excellent answers and commentary both Aaron and phils! After posting here, I thought it "smelled" like a bug, so entered a report to GNU- will send them a reference to these discussions.

Answer 1

As per comments to Aaron Miller's answer, here is an overview of what happens when a mode function is called (with an explanation of derived modes); how calling a mode manually differs from Emacs calling it automatically; and where after-change-major-mode-hook and hack-local-variables fit into this, in the context of the following suggested code:

(add-hook 'after-change-major-mode-hook 'hack-local-variables)

After visiting a file, Emacs calls normal-mode which "establishes the proper major mode and buffer-local variable bindings" for the buffer. It does this by first calling set-auto-mode, and immediately afterwards calling hack-local-variables, which determines all the directory-local and file-local variables for the buffer, and sets their values accordingly.

For details of how set-auto-mode chooses the mode to call, see C-hig (elisp) Auto Major Mode RET. It actually involves some early local-variable interaction (it needs to check for a mode variable, so there's a specific look-up for that which happens before the mode is set), but the 'proper' local variable processing happens afterwards.

When the selected mode function is actually called, there's a clever sequence of events which is worth detailing. This requires us to understand a little about "derived modes" and "delayed mode hooks"...

Derived modes, and mode hooks

The majority of major modes are defined with the macro define-derived-mode. (Of course there's nothing stopping you from simply writing (defun foo-mode ...) and doing whatever you want; but if you want to ensure that your major mode plays nicely with the rest of Emacs, you'll use the standard macros.)

When you define a derived mode, you must specify the parent mode which it derives from. If the mode has no logical parent, you still use this macro to define it (in order to get all the standard benefits), and you simply specify nil for the parent. Alternatively you could specify fundamental-mode as the parent, as the effect is much the same as for nil, as we shall see momentarily.

define-derived-mode then defines the mode function for you using a standard template, and the very first thing that happens when the mode function is called is:

(delay-mode-hooks
  (PARENT-MODE)
  ,@body
  ...)

or if no parent is set:

(delay-mode-hooks
  (kill-all-local-variables)
  ,@body
  ...)

As fundamental-mode itself calls (kill-all-local-variables) and then immediately returns when called in this situation, the effect of specifying it as the parent is equivalent to if the parent were nil.

Note that kill-all-local-variables runs change-major-mode-hook before doing anything else, so that will be the first hook which is run during this whole sequence (and it happens while the previous major mode is still active, before any of the code for the new mode has been evaluated).

So that's the first thing that happens. The very last thing that the mode function does is to call (run-mode-hooks MODE-HOOK) for its own MODE-HOOK variable (this variable name is literally the mode function's symbol name with a -hook suffix).

So if we consider a mode named child-mode which is derived from parent-mode which is derived from grandparent-mode, the whole chain of events when we call (child-mode) looks something like this:

(delay-mode-hooks
  (delay-mode-hooks
    (delay-mode-hooks
      (kill-all-local-variables) ;; runs change-major-mode-hook
      ,@grandparent-body)
    (run-mode-hooks 'grandparent-mode-hook)
    ,@parent-body)
  (run-mode-hooks 'parent-mode-hook)
  ,@child-body)
(run-mode-hooks 'child-mode-hook)

What does delay-mode-hooks do? It simply binds the variable delay-mode-hooks, which is checked by run-mode-hooks. When this variable is non-nil, run-mode-hooks just pushes its argument onto a list of hooks to be run at some future time, and returns immediately.

Only when delay-mode-hooks is nil will run-mode-hooks actually run the hooks. In the above example, this is not until (run-mode-hooks 'child-mode-hook) is called.

For the general case of (run-mode-hooks HOOKS), the following hooks run in sequence:

• change-major-mode-after-body-hook
• delayed-mode-hooks (in the sequence in which they would otherwise have run)
• HOOKS (being the argument to run-mode-hooks)
• after-change-major-mode-hook

So when we call (child-mode), the full sequence is:

(run-hooks 'change-major-mode-hook) ;; actually the first thing done by
(kill-all-local-variables)          ;; <-- this function
,@grandparent-body
,@parent-body
,@child-body
(run-hooks 'change-major-mode-after-body-hook)
(run-hooks 'grandparent-mode-hook)
(run-hooks 'parent-mode-hook)
(run-hooks 'child-mode-hook)
(run-hooks 'after-change-major-mode-hook)

Back to local variables...

Which brings us back to after-change-major-mode-hook and using it to call hack-local-variables:

(add-hook 'after-change-major-mode-hook 'hack-local-variables)

We can now see clearly that if we do this, there are two possible sequences of note:

1. We manually change to foo-mode:

   (foo-mode)
    => (kill-all-local-variables)
    => [...]
    => (run-hooks 'after-change-major-mode-hook)
        => (hack-local-variables)
   

2. We visit a file for which foo-mode is the automatic choice:

   (normal-mode)
    => (set-auto-mode)
        => (foo-mode)
            => (kill-all-local-variables)
            => [...]
            => (run-hooks 'after-change-major-mode-hook)
                => (hack-local-variables)
    => (hack-local-variables)
   

Is it a problem that hack-local-variables runs twice? Maybe, maybe not. At minimum it's slightly inefficient, but that's probably not a significant concern for most people. For me, the main thing is that I wouldn't want to rely upon this arrangement always being fine in all situations, as it's certainly not the expected behaviour.

(Personally I do actually cause this to happen in certain specific cases, and it works just fine; but of course those cases are easily tested -- whereas doing this as standard means that all cases are affected, and testing is impractical.)

So I would propose a small tweak to the technique, so that our additional call to the function does not happen if normal-mode is executing:

(defvar my-hack-local-variables-after-major-mode-change t
  "Whether to process local variables after a major mode change.
Disabled by advice if the mode change is triggered by `normal-mode',
as local variables are processed automatically in that instance.")

(defadvice normal-mode (around my-do-not-hack-local-variables-twice)
  "Prevents `after-change-major-mode-hook' from processing local variables.
See `my-after-change-major-mode-hack-local-variables'."
  (let ((my-hack-local-variables-after-major-mode-change nil))
    ad-do-it))
(ad-activate 'normal-mode)

(add-hook 'after-change-major-mode-hook 
          'my-after-change-major-mode-hack-local-variables)

(defun my-after-change-major-mode-hack-local-variables ()
  "Callback function for `after-change-major-mode-hook'."
  (when my-hack-local-variables-after-major-mode-change
    (hack-local-variables)))

Disadvantages to this?

The major one is that you can no longer change the mode of a buffer which sets its major mode using a local variable. Or rather, it will be changed back immediately as a result of the local variable processing.

That's not impossible to overcome, but I'm going to call it out of scope for the moment :)