Lisp source code rewriting system

30-some years ago I did something similar, using macrolet.

(Actually, I used defmacro because we had only an early implementation of Common Lisp, which did not yet have macrolet. But macrolet is the right thing to use.)

I didn't translate macro-expanded code to what it was expanded from, but the idea is pretty much the same. You will come across some different difficulties, I expect, since your translation is even farther away from one-to-one.

I wrote a translator from (what was then) Franz Lisp to Common Lisp, to help with porting lots of existing code to a Lisp+Prolog-machine project. Franz Lisp back then was only dynamically scoped, while Common Lisp is (in general) lexically scoped.

And yes, obviously there is no general way to automatically translate Lisp code (in particular), especially considering that it can generate and then evaluate other code - but even ignoring that special case. Many functions are quite similar, but there is the lexical/dynamic difference, as well as significant differences in the semantics of some seemingly similar functions.

All of that has to be understood and taken for granted from the outset, by anyone wanting to make use of the results of translation.

Still, much that is useful can be done. And if the resulting code is self-documenting, telling you what it was derived from etc., then when in the resulting context you can decide just what to do with this or that bit that might be tricky (e.g., rewrite it manually, from scratch or just tweak it). In practice, lots of code was easily converted from Franz to Common - it saved much reprogramming effort.

The translator program was written in Common Lisp. It could be used interactively as well as in batch. When used interactively it provided, in effect, a Franz Lisp interpreter on top of Common Lisp.

The program used only macro-expansion (i.e., not expansion followed by Lisp evaluation). Macro-expansion is another way of saying reduction semantics, or rewriting.

Input Franz-Lisp code was macro-expanded via function-definition mapping macros to produce Common-Lisp code. Code that was problematic for translation was flagged (in code) with a description/analysis that described the situation.

The program was called FTOC. I think you can still find it, or at least references to it, by googling (ftoc lisp). (It was the first Lisp program I wrote, and I still have fond memories of the experience. It was a good way to learn both Lisp dialects and to learn Lisp in general.)

Have fun!

In general, I don't think you can do this. The expansion of an lisp macro is Turing complete, so you have to be able to predict the output of a program which could have arbitrary input.

There are some simple things that you could do. defmacros with backquoted forms in appear fairly similar in the output form and might be detected. This sort of heuristic would probably get you a long way.

What I don't understand is your use case. The macro-expanded version of a piece of code is usually only present in the compiled (or in emacs-lisp byte-compiled) form.

Ok so other people have pointed out the fact that this problem is impossible in general. There are two hard parts to this problem: one is that it could be a lot of work to find a preimage of some code fragment through a macro and it is also impossible to determine whether a macro was called or not—there are examples where one may write code which could have come from a macro without using that macro. Imagine for the sake of illustration an sha macro which expands to the SHA hash of the string literal passed to it. Then if you see some sha hash in your expanded code, it would obviously be silly to try to unexpand it. But it may be that the hash was put into the code as a literal, e.g. referencing a specific point in the history of a git repository so it would also be unhelpful to unexpand the macro.

Tractable subproblems

Let me preface this by saying that whilst these may be a little tractable, I still wouldn’t try to solve this problem.

Let’s ignore all the macros that do weird things (like the example above) and all the macros that are just as likely to not have been used in the original (e.g. cond vs if) and all the macros which generate complex code which seems like it would be difficult to unravel (e.g. loop, do, and backquote. Annoyingly these difficult cases are some of those which you would perhaps most want to unexpand). The type this leaves us with (that I’d like to focus on) are macros which basically just reduce boilerplate, e.g. save-excursion or with-XXXX. These are macros whose implementation consists of possibly making some fresh symbols (via gensym) and then having a big simple backquoted block of code. I still think it would be too hard to automatically go from defmacro to a function for unexpansion but I think you could attack some of these on a case-by-case basis. Do this by looking for the forms generated by the macro that delimit (I.e. begin/end) the expanded code. I can’t really offer much beyond that. This is still a hard problem and I don’t think any existing solutions (to other problems) will get you very far on your way.

A further complication I understand is that you do not start at the macroexpanded code but rather at the bytecode. Without knowing anything about the elisp compiler, I worry that more information would be lost in the compilation step and you would have to undo that as well, e.g. perhaps it is hard to determine which code goes inside a let or even when a let begins, or bytecode starts using goto type features even though elisp doesn’t have them.

You suggest that the reason you would like to unexpand macros is so you can decompile bytecode which sometimes comes up in the Emacs debugger and that this would be useful as even though the source code is available in theory, it isn’t always at your fingertips. I put it to you that if you want to make your life debugging elisp easier it would be more worthwhile to figure out how to have the Emacs debugger always take you to the source code for internal functions. This might involve installing extra debugging related packages or downloading the Emacs source code and setting some variable so Emacs knows where to find it or compiling Emacs yourself from source. I don’t really know about that but I bet getting thrown into bytecode instead of source would have been enough of a problem for Emacs developers over the past thirty years that a solution to that problem does exist.

If however what you really want to do is to try to implement a decompiler for elisp then I suppose that’s what you should do. A final observation is that while Lisp provides facilities which make manipulating Lisp code easy, this doesn’t help much with decompiling as all these facilities can be used in compilation so there are infinitely more patterns one might want to detect than in e.g. a C decompiler. Perhaps scheme style macros would be easier to unexpand, although they would still be hard.

If you’re decompiling because you want to give a better idea of which exact subexpression rather than line is being evaluated (normally Lisp debuggers work on expressions not lines anyway) in the debugger then perhaps it would actually be useful to see the code at the expanded level rather than the unexpanded one. Or perhaps it would be best to see both and maybe in between as well. Keeping track of what’s what through forwards macroexpansion is already difficult and fiddly. Doing it in reverse certainly won’t be easier. Good luck!

Edit: seeing as your not currently using Lisp anyway, I wonder if you might have more success using something like prolog for your unexpanding. You’d still have to manually write rules but I think it would be a large amount of work to try to derive rules from macro definitions.

I would like to take Emacs Lisp code that has been macro expanded and unmacro expand it.

Macros generate arbitrary expressions, which may contain macros recursively. You have no general way to revert the transformations, because it's not pattern-based.

Even if macros were pattern-based, they could still be infinite.

Even if macros were not infinite, they can certainly contain bugs in expansions of patterns that never matched. Given arbitrary code to try to unwind, it could match an expansion that looks like the code and try to revert to its pattern. Without bugs, you could still abuse this.

Even if you could revert macro expansion, some macros expand to the same code. An approach could be signalling a warning with a restart when all reversions expand equally minus the operator, such that if the restart doesn't handle the signal, it would choose the first expansion; and otherwise signalling an error with a restart, such that if the restart doesn't handle the signal, it errors. Or you could configure it to choose certain macros under certain conditions, such as in which package the code was found.

In practice, there are very few cases where reverting an expansion makes any sense. It could be a useful development tool that suggests macros, but I wouldn't generally rely on it for whole source transformations.

One way you could achieve what you want is through a controlled pattern matching. You could initially create patterns manually, which would already handle cases you care about directly, such as the ones you mention:

• (if (not <cond>) <expr>) and (if (not <cond>) (progn <&expr>)) to (unless <cond> <&expr>) You'd have to decide whether null would be equivalent to not. I personally don't mix the boolean meaning of nil with that of empty list or something else, e.g. no result, nothing found, null object, a designator, etc. But perhaps Lisp code as old as that in Emacs just uses them interchangeably.

• (if <cond> <expr>) and (if <cond> (progn <&expr>)) to (when <cond> <&expr>)

• If you feel like improving code overall, include cond with a single condition. And be careful with cond clauses with only the condition.

You should have a few dozen more, to see how the pattern matching behaves with more patterns to match in terms of time (CPU) and space (memory).

From the description of fare-quasiquote, optima doesn't support backtracking, which you probably want.

But you can do backtracking with optima by yourself, using recursion on complex inner patterns, and if nothing matches, return a control value to keep searching for matching patterns from the outer input.

Another approach is to treat a pattern as a description of a state machine, and handle each new token to advance the current state machines until one of them reaches the end, discarding the state machines that couldn't advance. This approach may consume more memory, depending on the amount of patterns, the similarity between patterns (if many have the same starting token, many state machines will be generated on a matching token), the length of the patterns and, last but not least, the length of the input (s-expression).

An advantage of this approach is that you can use it interactively to see which patterns have matched the most tokens, and you can give weights to patterns instead of just taking the first that matches.

A disadvantage is that, most probably, you'll have to spend effort to develop it.

EDIT: I just lousily described a kind of trie or radix tree.

Once you got something working, maybe try to obtain patterns automatically. This is really hard, you must probably limit it to simple backquoting and accept the fact you can't generalize for anything that contains more complex code.

I believe the hardest will be code walking, which is hard enough with source code, but much more with macro-expanded code. Perhaps if you could expand the whole picture a bit further to understand the goal, maybe someone could suggest a better approach other than operating on macro-expanded code.

However one would think that this kind of thing, S-expression transformation, is right up Lisp's alley. And defmacro is I believe available in Lisp as it is in Emacs Lisp.

So surely there are program transformation systems, or term-rewriting systems that can be adapted here.

There's a huge step from expanding code with defmacro and all that generality. Most Lisp developers will know about hygienic macros, at least in terms of symbols as variables.

But there's still hygienic macros in terms of symbols as operators¹, code walking, interaction with a containing macro (usually using macrolet), etc. It's way too complex.

1.

Common Lisp evaluates the operator in a compound form in the lexical environment, and probably everyone makes macros that assume that the global macro or function definition of a symbol will be used.

But it might not be so:

(defmacro my-macro-1 ()
  `1)

(defmacro my-macro-2 ()
  `(my-function (my-macro-1)))

(defun my-function (n)
  (* n 100))

(macrolet ((my-macro-1 ()
             `2))
  (flet ((my-function (n)
           (* n 1000)))
    (my-macro-2)))

That last line will expand to (my-function (my-macro-2)), which will be recursively expanded to (my-function 2). When evaluated, it will yield 2000.

For proper operator hygiene, you'd have to do something like this:

(defmacro my-macro-2 ()
  ;; capture global bindings of my-macro-1 and my-function-1 by name
  (flet ((my-macro-1-global (form env)
           (funcall (macro-function 'my-macro-1) form env))
         (my-function-global (&rest args)
           ;; hope the compiler can optimize this
           (apply 'my-function args)))
    ;; store them globally in uninterned symbols
    ;; hopefully, no one will mess with them
    (let ((my-macro-1-symbol (gensym (symbol-name 'my-macro-1)))
          (my-function-symbol (gensym (symbol-name 'my-function))))
      (setf (macro-function my-macro-1-symbol) #'my-macro-1-global)
      (setf (symbol-function my-function-symbol) #'my-function-global)
      `(,my-function-symbol (,my-macro-1-symbol)))))

With this definition, the example will yield 100.

Common Lisp has some restrictions to avoid this, but it only states the consequences are undefined when (re)defining symbols in the common-lisp package, globally or locally. It doesn't require errors or warnings to be signaled.

I don't think it is possible to do this in general, but you can undo a pattern back into a macro use for every match if you supply code for each unmacroing. Code that mixed cond and if will end up being just if and your code would remove all if into cond making the reverse not the same as the starting point. The more macros you have and the more they expand into each other the more uncertain of the end result will be of the starting point.

You could have rules such that if is not translated into cond unless you used one of the features, like more than one predicate or implicit progn, but you have no idea if the coder actually did use cond everywhere because he liked in consistent regardless. Thus your unmacroing will acyually be more of a simplification.

I don't believe there's a general solution to that, and you certainly can't guarantee that the structure of the output would match that of the original code, and I'm not going near the idea of auto-generating patterns and desired transformations from macro definitions; but you might achieve a simple version of this with Emacs' own pcase pattern matching facility.

Here's the simplest example I could think of:

With reference to the definition of when:

(defmacro when (cond &rest body)
  (list 'if cond (cons 'progn body)))

We can transform code using a pcase pattern like so:

(let ((form '(if (and foo bar baz) (progn do (all the) things))))
  (pcase form
    (`(if ,cond (progn . ,body))
     `(when ,cond ,@body))
    (_ form)))

=> (when (and foo bar baz) do (all the) things)

Obviously if the macro definitions change, then your patterns will cease to work (but that's a pretty safe kind of failure).

Caveat: This is the first time I've written a pcase form, and I don't know what I don't know. It seems to work as intended, though.