The thesis that underlies my project to translate the Emacs C code to Common Lisp is that Emacs Lisp is close enough to Common Lisp that the parts of the Emacs C code that implement Lisp can be dropped in favor of the generally superior CL implementation.  This is generally true, but there are a few difficult bits.

Symbols

The primary problem is the translation of symbols when used as variable references.  Consider this code:

(defvar global 73)
(defun function (argument)
  (let ((local (something-else))
    (+ local argument global)))

More is going on here than meets the eye.

First, Emacs Lisp uses dynamic binding by default (optional lexical binding is a new feature in Emacs 24).  This applies to function arguments as well as other bindings.  So, you might think you could translate this straightforwardly to:

(defvar global 73)
(declare (special global))
(defun function (argument)
  (declare (special argument))
  (let ((local (something-else))
    (declare (special local))
    (+ local argument global)))

This was the approach taken by elisp.lisp; it defined macros for let and let* (but forgot defun) to do the dirty work:

(defmacro el::let* ((&rest vars) &rest forms)
  "Emacs-Lisp version of `let*' (everything special)."
  `(let* ,vars (declare (special ,@(mapcar #'from-list vars))) ,@forms))

But not so fast!  Emacs also has buffer-local variables.  These are variables where the value is associated with the current buffer; switching buffers makes a different binding visible to Lisp.  These require no special syntax, and a variable can be made buffer-local at any time.  So, we can break the above translation simply by evaluating:

(make-local-variable 'global)
(setq global 0)

Whoops!  Now the function will return the wrong result — the translation will have no way to know that is should refer to the buffer-local value.  (Well, ok, pretend that the setq magically worked somehow…)

My idea for implementing this is pretty convoluted.  Actually I have two ideas, one “user” and one “kernel”:

User

I think it is possible to use define-symbol-macro on all symbols that come from Elisp, so that we can tell the CL compiler about the real implementation.  However, a symbol can either be treated as a variable, or it can be treated as a symbol-macro — not both at the same time.  So, we will need a second location of some kind to store the real value.  Right now I’m thinking a symbol in another package, but maybe a cons or some other object would work better. In either case, we’d need a macro, a setf method for its expansion, and some extra-tricky redefinitions of let and defun to account for this change.

This would look something like:

(define-symbol-macro global (elisp:get-elisp-value 'global))
(defsetf elisp:get-elisp-value elisp:set-elisp-value))
;; Details left as an exercise for the reader.

This solution then has to be applied to buffer-, keyboard-, and frame-local variables.

Kernel

The kernel method is a lot simpler to explain: hack a Common Lisp implementation to directly know about buffer-locals.  SMOP!  But on the whole I think this approach is to be less preferred.

Other Problems

Emacs Lisp also freely extends other typical data types with custom attributes.  I consider this part of the genius of Emacs; a more ordinary program would work within the strictures of some defined, external language, but Emacs is not so cautious or constrained.  (Emacs is sort of a case study in breaking generally accepted rules of programming; which makes one wonder whether those rules are any good at all.)

So, for example, strings in Emacs have properties as a built-in component.  The solution here is simple — we will just translate the Emacs string data type as a whole, something we probably have to do anyway, because Emacs also has its own idiosyncratic approach to different encodings.

In elisp, aref can be used to access elements of other vector-like objects, not just arrays; there are some other odd little cases like this.  This is also easily handled; but it left me wondering why things like aref aren’t generic methods in CL. It often seems to me that a simpler, more orthogonal language lies inside of CL, struggling to get free. I try not to think these thoughts, though, as that way lies Scheme and the ridiculous fragmentation that has left Lisp unpopular.

This is a followup to my earlier post on converting the Emacs C code into Common Lisp.  This one is a bit more technical, diving into some specifics of the conversion process.

Basics

One important fact is that we do not need to convert an arbitrary C program to Common Lisp.  This might or might not be efficiently possible — but we do not care.  We only need to convert Emacs.  This is simpler for two reasons.  First, we can just ignore any C construct that Emacs does not use.  If the translator barfs after some new update, we can fix it then.  Second, Emacs itself is already written in a relatively Lispy style, being a Lisp implementation itself.  We further exploit this by allowing the translator to know some details about Emacs.  As a trivial example, all the Smumble globals created by the DEFUN marco need not be translated into Common Lisp as structure constants — they are an artifact of the implementation, and will show up directly in the generated defuns instead.

What to ignore

A good portion of Emacs is simply redundant in the CL world.  There are a few types (cons, vector, integers, functions) that are shareable — in fact, sharing these is part of the goal of this effort.  There are also a number of functions which are effectively identical.  There are also entire redundant modules, like the garbage collector, or the bytecode interpreter.

The question is how to have the translator differentiate between what is useful and what is not, without breaking builds of future versions of Emacs.

I don’t currently think there is a high road to solving this problem.  For modules like the GC, I plan to have ad hoc translator rules for the particular source files.  For functions and data types, I’m adding new GCC attributes that I can use to mark the ignorable definitions.

Types

There are two type-related issues that arise when translating the source.

First, how should Emacs-specific types be represented?  Primarily these types are structures, like struct buffer or struct string (we cannot use the CL string type, because Emacs adds properties directly to the string, and Emacs has its own idiosyncratic character handling).  My answer here is to just straightforwardly translate them to defstruct.

The other question is when translating a C function, what do we do with the types of local variables?  For the most part I am pretending that they don’t exist.  This works fine except for local arrays and structures, but these are easily handled by initializing variables properly. My rationale is that while this is slower, it lets me get something working more quickly, and we can always update the translator to emit CL type declarations later on.

This simple approach doesn’t actually cover all the needed cases.  For example, there is code in Emacs that takes the address of a local variable and passes it somewhere.  This is easy to deal with; much of the remaining work is just digging through the code looking for special cases to clean up.

I’m similarly omitting type declarations from the generated structures.  One possible nice side effect of this approach is that it will make it easier to lift Emacs’ file-size restrictions, because there will no longer be any code assuming that the size is a fixnum.

Macros

Many low-level details of the Emacs implementation are hidden in macros.  For example, Emacs stuffs some type information into the low-order bits of pointers.  It uses macros to add or remove this information.  For this build, I redefine these macros to do nothing.  This makes the GCC Gimple representation much closer to the abstract meaning of the program, and thus simpler to translate.

There are also some macros that are useful to redefine so that we can more easily hook into them from the translator.  For example, Emacs has a C macro INTEGERP that is used to check whether its argument is an integer.  Normally this macro uses bit twiddling to get its answer, but I redefine it like so:

#undef INTEGERP
extern Lisp_Object *INTEGERP (Lisp_Object)
    __attribute__((lisp_form("integerp")));

Example

The translator is not nearly complete, but it can already do a fair job at translating simple functions.  For example, here is “forward-point” from the Emacs C code:

DEFUN ("forward-point", Fforward_point, Sforward_point, 1, 1, 0,
       doc: /* Return buffer position N characters after (before if N negative) point.  */)
  (Lisp_Object n)
{
  CHECK_NUMBER (n);

  return make_number (PT + XINT (n));
}

Here is what the translator comes up with:

(defun Fforward_point (n)
  (let (
    temp-var-0
    Qintegerp.316
    temp-var-1
    current_buffer.317
    temp-var-2
    )
    (block nil (tagbody
      bb-0
        ; no gimple here
      bb-1
        ; no gimple here
      bb-2
        (setf temp-var-0 (integerp n))
        (if (== temp-var-0 nil)
          (go bb-3)
          (go bb-4))
      bb-3
        (setf Qintegerp.316 Qintegerp)
        (wrong_type_argument Qintegerp.316 n)
      bb-4
        (setf current_buffer.317 current_buffer)
        (setf temp-var-2 (buffer-pt current_buffer.317))
        (setf temp-var-1 (+ temp-var-2 n))
        (return temp-var-1)
  ))))

(defun elisp:forward-point (arg0)
  (Fforward_point arg0))

The output looks pretty weird, because the translator works after GCC’s CFG is built, and so the most straightforward translation is to use this mess with tagbody.  I doubt this matters much, but in any case the translator is readily hackable — it is still less than 400 lines of Python, including comments.

One thing to note is the translation of “PT“.  This is actually a macro that refers to the current buffer:

#define PT (current_buffer->pt + 0)

The translator properly turns this into a reference to “buffer-pt“.

Another detail is the handling of packages.  My plan is to put the Emacs implementation into one package, and then any elisp into a second package called “elisp“.  A DEFUN in the C code will actually generate two functions: the internal one, and the elisp-visible one; hence the “elisp:” in the translation.

Next Steps

There’s still a good amount of work to be done.  The converter punts on various constructs; type translation is implemented but not actually wired up to anything; the translator should emit definitions for alien functions; and plenty more.

Recently I’ve been thinking about how to rebase Emacs on Common Lisp.

First, why rebase?  Performance is the biggest reason.  Emacs Lisp is a very basic lisp implementation.  It has a primitive garbage collector and basic execution model, and due to how it is written, it is quite hard to improve this in place.

Seccond, why Common Lisp?  Two reasons: first, Emacs Lisp resembles Common Lisp in many ways; elisp is like CL’s baby brother.  Second, all of the hard problems in Lisp execution have already been solved excellently by existing, free-software CL implementations.  In particular, the good CL implementations have much better garbage collectors, native compilation, threads, and FFI; we could expose the latter two to elisp in a straightforward way.

By “rebase” I mean something quite ambitious — rewrite the C source of Emacs into Common Lisp.  I think this can largely be automated via a GCC plugin (e.g., written using David Malcolm’s Python plugin).  Full automation would let the CL port be just another way to build Emacs, following the upstream development directly until all the Emacs maintainers can be convinced to drop C entirely (cough, cough).

Part of the rewrite would be dropping code that can be shared with CL.  For example, we don’t need to translate the Emacs implementation of “cons“, we can just use the CL one.

Some CL glue would be needed to make this all work properly.  These days it can’t be quite as small as elisp.lisp, but it still would not need to be very big.  The trickiest problem is dealing with buffer-local variables; but I think that can be done by judicious use of define-symbol-macro in the elisp reader.

Emacs might be the only program in the world that would see a performance improvement from rewriting in CL :-).  The reason for this is simple: Emacs’ performance is largely related to how well it executes lisp code, and how well the GC works.