Mini-tutorial on explicit (and implicit) renaming macros in Chicken
The following is an attempt to explain, how to write hygienic macros in Chicken with explicit renaming. It supposes, that the reader knows, how to write other low-level-macros like define-macro in Chicken-3 and other Schemes or defmacro in Lisp. As you will see, explicit renaming macros are even easier to write than those with define-macro. This is because there is a brute force method to avoid variable capture.
Let's consider a trivial macro a la Chicken-3 without any considerations on hygiene and variable capture:
(define-macro swap! ; wrong (lambda (x y) `(let ((tmp ,x)) (set! ,x ,y) (set! ,y tmp))))
or, in short form
(define-macro (swap! x y) ; wrong `(let ((tmp ,x)) (set! ,x ,y) (set! ,y tmp)))
You might be surprised, but the replacement text, that within the backquotes, need not be changed in an explicit renaming macro, at least in a first step. What has to be changed, is the signature: explicit renaming transformers are always procedures of three variables, usually called form, rename, compare?, which has to be wrapped into an er-macro-transformer call:
(define-syntax swap! (er-macro-transformer (lambda (form rename compare?) ... `(let ((tmp ,x)) (set! ,x ,y) (set! ,y tmp)))))
and you have to destructure the form argument, which is the whole macro-from, (swap! x y) in the example, to fill in the ellipsis above. That is
(define-syntax swap! ; wrong (er-macro-transformer (lambda (form rename compare?) (let ((x (cadr form)) (y (caddr form))) `(let ((tmp ,x)) (set! ,x ,y) (set! ,y tmp))))))
In this form, the new macro is the same as our first attempt with define-macro. We haven't bothered about variable capture or hygiene at all. But this macro already works. You can test it, say with
(let ((x 'x) (y 'y)) (swap! x y) (list x y))
and see if it does what it is supposed to do. Of course, it is not correct, since if a client uses tmp as one of his or her arguments, the macro will crash.
In the classical macro systems, defmacro in Common Lisp or define-macro in some Schemes, you now have to think carefully, which macro variables are in danger of variable capture and use an uninterned symbol for them, tmp in the current example. Worse than that, contrary to Common Lisp, Scheme allows to use any name for variables, even keyword names as let, set!, ... so that define-macro can never create a hygienic swap! macro. Here is, where renaming comes into the play. You don't need gensym, use rename instead. And you needn't bother, which symbol to rename, rename all.
What does that mean, "all"? It's easy in the present example, everything within the backquote, except those symbols, which are already unquoted. But in more complicated examples, you should use expand, to see the replacement text of your macro calls.
(pp (expand '(swap! x y)))
will result in the replacement text
(let ((tmp x)) (set! x y) (set! y tmp))
Only x and y are arguments of your swap! call, hence everything else in the replacement text should be renamed: namely let, tmp and set!. So a hygienic version of swap! would be
(define-syntax swap! (er-macro-transformer (lambda (form rename compare?) (let ((x (cadr form)) (y (caddr form))) `(,(rename 'let) ((,(rename 'tmp) ,x)) (,(rename 'set!) ,x ,y) (,(rename 'set!) ,y ,(rename 'tmp)))))))
If you repeat the expand call above, you will get the same replacement text, but with numbers suffixed to the renamed symbols, something like
(let11 ((tmp12 x)) (set!13 x y) (set!13 y tmp12))
These renamed symbols have the same meaning as the original names without suffixes, but serve as aliases which the client cannot use under any circumstances.
Note, that the two appearances of tmp and set! are renamed to the same alias, the rename operator is referentially transparent. Note also, that these renamed symbols are much easier to interpret than gensym'ed ones!
I personally do not like these rename calls within the backquoted expression, because you loose the visual similarity between the backquoted expression and the resulting replacement text. Therefore I prefer this version, which is equivalent
(define-syntax swap! (er-macro-transformer (lambda (form rename compare?) (let ( (x (cadr form)) (y (caddr form)) (%tmp (rename 'tmp)) (%let (rename 'let)) (%set! (rename 'set!)) ) `(,%let ((,%tmp ,x)) (,%set! ,x ,y) (,%set! ,y ,%tmp))))))
You can think of the two characters ,% in the template as an identity operator.
We havn't used the compare? argument in the transformer in this example. What's its role?
Well, some macros use special symbols verbatim, else and => in cond for example. With compare? it's possible to check, if these symbols, which might appear verbatim in a list, have the same meaning as their renamed variants (rename 'else), (rename '=>). For example, if else appears as the car of a list, lst say, you can code something like
(if (compare? (car lst) (rename 'else)) ...)
You see, explicit renaming macros are a bit more verbose than classical ones with define-macro. But everybody who is able to write the latter, is able to write the former as well. It's even easier!
Implicit renaming macros
But it can still be improved. Above, I said "You need'nt bother which symbol to rename, rename all". Well, "all" is something which can perfectly be done by a machine. Consequently, since the advent of Chicken version 4.7, there is another low level macro system, based on ir-macro-transformer, which does this "rename all" in the background.
In explicit renaming macros, I had to tell the compiler explicitly, what to rename. Every symbol not renamed would break hygiene. In implicit renaming macros, the roles are reversed: I have to tell the compiler explicitly, which symbol should break hygiene, everything else is renamed, whence hygienic. A hygienic swap! macro now looks like this
(define-syntax swap! (ir-macro-transformer (lambda (form inject compare?) (let ((x (cadr form)) (y (caddr form))) `(let ((tmp ,x)) (set! ,x ,y) (set! ,y tmp))))))
Note, that the second argument to the transformer is now called inject. It's used to "inject" a symbol unrenamed into the replacement text. Note further, that there is no need to gensym tmp as in define-macro. And last but not least, it's obligatory to wrap the transformer into an ir-macro-transformer call: the compiler must know, which low-level system to compile, after all.
For implicit renaming macros, the third argument to the transformer is now used a bit differently, for example
(if (compare? 'else (car lst)) ...)
Destructuring the macro-code
Up to now, we have always destructured the macro-code, (swap! x y), by hand. define-macro did it automatically. Is there a way to do it automatically with explicit- and implicit-renaming-macros as well?
Off course, you can use match from the matchable package or an appropriate binding operater from the bindings package (bind or bind-case for example) to do the destructuring. Then instead of
(let ((x (cadr form)) (y (caddr form))) ...)
you could write for example
(match form ((_ x y) ...))
provided you had wrapped the code into something like
(import-for-syntax (only matchable match))
In simple macros like swap!, the use of match or bind-case is not that much help. But a lot of macros have complicated nested list structures as macro-code. Then automatic destructuring is almost a necessity. Moreover, it does error-checking for free ...
Hiding the matching process
To facilitate the use of low-level-macros even more, you can abstract away not only the renaming but also the matching process, by packaging it -- you guess it -- in a macro, conveniently one which looks much like syntax-rules and should hence be called ir-macro-rules. An explicit renaming version called er-macro-rules exists as well. To use them, they must be imported for syntax
(import-for-syntax (only low-level-macros ir-macro-rules er-macro-rules))
ir-macro-rules pairs possible macro-codes (where the macro name is conveniently replaced by an underscore) with expressions, which evaluate to the macro-expansion.
With it, our old friend swap! would look like this
(define-syntax swap! (ir-macro-rules () ((_ x y) `(let ((tmp ,x)) (set! ,x ,y) (set! ,y tmp)))))
The empty list in this macro shows, that the generated macro is hygienic. To implement unhygienic macros, you have to list those names, which should pollute the local namespace.
If you like it a bit more low-level and want to do the renaming yourself, you can code
(define-syntax swap! (er-macro-rules (%tmp %let %set!) ((_ x y) `(,%let ((,%tmp ,x)) (,%set! ,x ,y) (,%set! ,y ,%tmp)))))
where all renamed symbols are listet in er-macro-rules' first argument.
For details see low-level-macros.
You'll find there define-macro as well, and believe it or not
(define-macro (swap! x y) `(let ((tmp ,x)) (set! ,x ,y) (set! ,y tmp)))
is not only correct, but hygienic as well!
Low level macros are really great ...
Jun 09, 2009
Sep 13, 2011