This chapter describes several items that are included with current versions of Chez Scheme primarily for compatibility with older versions of the system.
The first two sections describe expansion-passing-style and extend-syntax macros. These features are supported directly by current versions of Chez Scheme, but support may be dropped in future versions. New programs should use the syntax-case facility described in The Scheme Programming Language, Second Edition [9] instead.
The third section describes a compatibility file distributed with Chez Scheme that contains definitions for forms and procedures no longer supported directly by Chez Scheme.
Earlier versions of Chez Scheme provided two complimentary macro systems: a low-level "expansion-passing-style" (EPS) macro system and a high-level extend-syntax system. The extend-syntax system was defined in terms of the lower-level system. extend-syntax is also supported, to a limited degree, by the syntax-case system. While it is not possible to intermix macros written using expansion-passing-style macros with syntax-case macros, it is possible to enable the EPS expander, allowing existing programs to be run without the need to rewrite macro definitions. The expander used by the compiler, interpreter, and direct calls to expand is determined by the value of the parameter current-expand. Its default value is sc-expand, the syntax-case expander. It must be set to eps-expand to enable support for expansion-passing-style macros.
An EPS macro is a procedure, or expander, associated with a keyword, that accepts two arguments: the expression to expand and an expander with which to perform further expansion. The expander associated with a keyword may choose to apply the expander passed to it to the entire expression it builds before returning, it may choose to apply the expander to selected pieces of the expression it builds, or it may choose not to use the expander. It may even use a different expander from the one given to it to expand the expression or pieces of the expression it builds. Expansion-passing style is described fully in the article "Expansion-passing style: a general macro mechanism" [12].
EPS macros are defined using the procedure install-expander or the syntactic form define-syntax-expander.
procedure: (install-expander keyword expander)
returns: unspecified
keyword must be a symbol, and expander must be a procedure. install-expander associates expander with the keyword keyword for purposes of macro expansion. When the expander (see eps-expand and eps-expand-once) finds a list-structured expression whose car is keyword, expander is passed the expression and an expander with which to perform further expansion.
New programs should employ syntax-case instead.
syntax: (define-syntax-expander keyword exp)
returns: unspecified
keyword must be a symbol, and exp must evaluate to a procedure.
define-syntax-expander associates the expansion procedure (expander) obtained by evaluating exp with the keyword keyword for purposes of macro expansion. When the expander (see eps-expand and eps-expand-once) finds a list-structured expression whose car is keyword, the expander is passed the expression and an expander with which to perform further expansion.
define-syntax-expander differs from install-expander in two ways: First, the keyword expression keyword is not evaluated. Second, a define-syntax-expander form is implicitly wrapped in an eval-when expression with all three options (compile, load, and eval) specified, so that syntax expander definitions appearing within a file are available to subsequent expressions in the file when the file is processed by compile-file.
extend-syntax is layered on top of define-syntax-expander, as can be seen by applying eps-expand-once to an extend-syntax form.
New programs should employ syntax-case instead.
(define-syntax-expander quote
(lambda (x e)
x))
(define-syntax-expander let*
; no syntax checking (should be written with extend-syntax!)
(lambda (x e)
(let ([bindings (cadr x)] [exps (cddr x)])
(if (null? bindings)
(e `(let () ,@exps) e)
(e `(let (,(car bindings))
(let* ,(cdr bindings) ,@exps)) e)))))
(eps-expand-once '(quote a)) 
(quote a)
(eps-expand-once
'(let* ([x 'a] [y x]) y)) (let ([x 'a]) (let* ([y x]) y))
This section describes extend-syntax, a powerful yet easy to use syntactic extension facility based on pattern matching [21]. Syntactic transformations written using extend-syntax are similar to those written using a define-syntax with syntax-case, except that the transformations produced by extend-syntax do not automatically respect lexical scoping.
extend-syntax is supported by the syntax-case expander as well as by the EPS expander. It is not always possible, however, to mix syntactic abstractions written using syntax-case with those written using extend-syntax; it is generally preferable to use one or the other wherever possible. Support for extend-syntax within the syntax-case expander is provided only as an aid to migrating to syntax-case.
syntax: (extend-syntax (name key ...) (pat fender template) ...)
returns: unspecified
The identifier name is the name, or syntax keyword, for the syntactic extension to be defined. When the system expander processes any list expression whose car is name, the syntactic transformation procedure generated by extend-syntax is invoked on this expression. The remaining identifiers key ... are additional keywords to be recognized within input expressions during expansion (such as else in cond or case).
Each clause after the list of keys consists of a pattern pat, an optional fender, and a template. The optional fender is omitted more often than not. The pat specifies the syntax the input expression must have for the clause to be chosen. Identifiers within the pattern that are not keywords (pattern variables) are bound to corresponding pieces of the input expression. If present, the fender is a Scheme expression that specifies additional constraints on the input expression (accessed through the pattern variables) that must be satisfied in order for the clause to be chosen. The template specifies what form the output takes, usually in terms of the pattern variables.
During expansion, the transformation procedure extend-syntax generates attempts to match the input expression against each pattern in the order the clauses are given. If the input expression matches the pattern, the pattern variables are bound to the corresponding pieces of the input expression and the fender for the clause, if any, is evaluated. If the fender returns a true value, the given expansion is performed. If input does not match the pattern or if the fender returns a false value, the transformation procedure tries the next clause. An error is signaled if no clause can be chosen.
Within the pattern, ellipsis (...) may be used to specify zero or more occurrences of the preceding pattern fragment, or prototype. Similarly, ellipses may be used in the output to specify the construction of zero or more expansion prototypes. In this case, the expansion prototype must contain part of an input pattern prototype. The use of patterns, templates, ellipses within patterns and templates, and fenders is illustrated in the following sequence of examples.
The first example, defining rec, uses a single keyword, a single clause with no fender, and no ellipses. The call to eps-expand-once in this example and the ones that follow assumes that the EPS expander has been enabled prior to the use of extend-syntax using (current-expand eps-expand).
(extend-syntax (rec)
[(rec id val)
(let ([id #f])
(set! id val)
id)])
(eps-expand-once 
(let ([eternal #f])
'(rec eternal (set! eternal
(lambda () (lambda ()
(eternal)))) (eternal)))
eternal)
The second example, defining when, shows the use of ellipses.
(extend-syntax (when)
[(when test exp1 exp2 ...)
(if test (begin exp1 exp2 ...) #f)])
(eps-expand-once (if (eq? msg 'print)
'(when (eq? msg 'print) (begin (display "hi")
(display "hi") (newline))
(newline))) #f)
The next example shows the definition of let. The definition of let shows the use of multiple ellipses, employing one for the identifier/value pairs and one for the expressions in the body. It also shows that the prototype need not be a single identifier, and that pieces of the prototype may be separated from one another in the template.
(extend-syntax (let)
[(let ([x v] ...) e1 e2 ...)
((lambda (x ...) e1 e2 ...) v ...)])
(eps-expand-once ((lambda (x y) (+ x y))
'(let ([x 3] [y 4]) 3
(+ x y))) 4)
The next example shows let*, whose syntax is the same as for let, but which is defined recursively in terms of let with two clauses (one for the base case, one for the recursion step) since it must produce a nested structure.
(extend-syntax (let*)
[(let* () e1 e2 ...)
(let () e1 e2 ...)]
[(let* ([x v] more ...) e1 e2 ...)
(let ([x v]) (let* (more ...) e1 e2 ...))])
The first pattern/template pair matches any let* expression with no identifier/value pairs and maps it into the equivalent begin expression. This is the base case. The second pattern/template pair matches any let* expression with one or more identifier/value pairs and transforms it into a let expression binding the first pair whose body is a let* expression binding the remaining pairs. This is the recursion step, which will eventually lead us to the base case because we remove one identifier/value pair at each step. Notice that the second pattern uses the pattern variable more for the second and later identifier/value pairs; this makes the pattern and template less cluttered and makes it clear that only the first identifier/value pair is dealt with explicitly.
It is interesting to try both eps-expand-once and expand on a let* expression. eps-expand-once shows the intermediate form in terms of let and a simpler let* expression; expand shows the expression completely expanded into applications of lambda expressions.
(eps-expand-once (let ([x 3])
'(let* ([x 3] [y x]) (let* ([y x])
(+ x y))) (+ x y)))
(expand ((lambda (x)
'(let* ([x 3] [y x]) ((lambda (y) (+ x y)) x))
(+ x y))) 3)
The definition for and requires three clauses. The first clause is necessary to recognize (and), and the second two define all other and forms recursively.
(extend-syntax (and)
[(and) #t]
[(and x) x]
[(and x y ...) (if x (and y ...) #f)])
The definition for cond requires four clauses. As with let*, cond must be described recursively, partly because it produces nested if expressions, and partly because one ellipsis prototype would not be sufficient to describe all possible cond clauses. The definition of cond also requires that we specify else as a keyword, in addition to cond. Here is the definition:
(extend-syntax (cond else)
[(cond) #f]
[(cond (else e1 e2 ...))
(begin e1 e2 ...)]
[(cond (test) more ...)
(or test (cond more ...))]
[(cond (test e1 e2 ...) more ...)
(if test
(begin e1 e2 ...)
(cond more ...))])
Two of the clauses are base cases and two are recursion steps. The first base case is an empty cond. The value of cond in this case is unspecified, so the choice of #f is somewhat arbitrary. The second base case is a cond containing only an else clause; this is transformed to the equivalent begin expression. The two recursion steps differ in the number of expressions in the cond clause. The value of cond when the first true clause contains only the test expression is the value of the test. This is similar to what or does, so we expand the cond clause into an or expression. On the other hand, when there are expressions following the test expression, the value of the last expression is returned, so we use if and begin.
(eps-expand-once (if test
'(cond (begin exp1 exp2)
[test exp1 exp2])) (cond))
(eps-expand-once
'(cond [test])) (or test (cond))
(expand (if test1
'(cond exp1
[test1 exp1] (if test2
[test2 exp2] exp2
[else exp3])) exp3))
To be absolutely correct about the syntax of let, we actually must require that the bound identifiers in the input are symbols. If we typed something like (let ([3 x]) x) we would not get an error from let because it does not check to verify that the objects in the identifier positions are symbols. Instead, lambda may complain, or perhaps the system evaluator long after expansion is complete. This is where fenders are useful.
(extend-syntax (let)
[(let ([x v] ...) e1 e2 ...)
(andmap symbol? '(x ...))
((lambda (x ...) e1 e2 ...) v ...)])
The andmap of symbol? over '(x ...) assures that each bound identifier is a symbol. A fender is simply a Scheme expression. Within that expression, any quoted object is first expanded by the same rules as the template part of the clause. In this case, '(x ...) is expanded to the list of identifiers from the identifier/value pairs.
extend-syntax typically handles everything you need it for, but some syntactic extension definitions require the ability to include the result of evaluating an arbitrary Scheme expression. This ability is provided by with.
syntax: (with ((pat exp) ...) template)
returns: processed template
with is valid only within an template inside of extend-syntax. with patterns are the same as extend-syntax patterns, with expressions are the same as extend-syntax fenders, and with templates are the same as extend-syntax templates.
with can be used to introduce new pattern identifiers bound to expressions produced by arbitrary Scheme expressions within extend-syntax templates. That is, with allows an escape from the declarative style of extend-syntax into the procedural style of full Scheme.
One common use of with is the introduction of a temporary identifier or list of temporary identifiers into a template. with is also used to perform complex transformations that might be clumsy or inefficient if performed within the extend-syntax framework.
For example, or requires the use of a temporary identifier. We could define or as follows.
(extend-syntax (or)
[(or) #f]
[(or x) x]
[(or x y ...)
(let ([temp x])
(if temp temp (or y ...)))])
This would work until we placed an or expression within the scope of an occurrence of temp, in which case strange things could happen, since extend-syntax does not respect lexical scoping. (This is one of the reasons that define-syntax is preferable to extend-syntax.)
(let ([temp #t])
(or #f temp)) #f
One solution is to use gensym and with to create a temporary identifier, as follows.
(extend-syntax (or)
[(or) #f]
[(or x) x]
[(or x y ...)
(with ([temp (gensym)])
(let ([temp x])
(if temp temp (or y ...))))])
Also, with can be used to combine elements of the input pattern in ways not possible directly with extend-syntax, such as the following folding-plus example.
(extend-syntax (folding-plus)
[(folding-plus x y)
(and (number? 'x) (number? 'y))
(with ([val (+ 'x 'y)])
val)]
[(folding-plus x y) (+ x y)])
folding-plus collapses into the value of (+ x y) if both x and y are numeric constants. Otherwise, folding-plus is transformed into (+ x y) for later evaluation. The fender checks that the operands are numbers at expansion time, and the with performs the evaluation. As with fenders, expansion is performed only within a quoted expressions, since quote sets the data apart from the remainder of the Scheme expression.
(eps-expand-once '(folding-plus 3 4.3)) 7.3
(eps-expand-once '(folding-plus 3 x)) (+ 3 x)
The example below binds a list of pattern variables to a list of temporary symbols, taking advantage of the fact that with allows us to bind patterns to expressions. This list of temporaries helps us to implement the sigma syntactic extension. sigma is similar to lambda, except it assigns the identifiers in the identifier list instead of creating new bindings. It may be used to perform a series of assignments in parallel.
(extend-syntax (sigma)
[(sigma (x ...) e1 e2 ...)
(with ([(t ...) (map (lambda (x) (gensym)) '(x ...))])
(lambda (t ...)
(set! x t) ...
e1 e2 ...))])
(eps-expand-once (lambda (G0 G1)
'(sigma (x y) (set! x G0)
(list x y))) (set! y G1)
(list x y))
(let ([x 'a] [y 'b])
((sigma (x y) (list x y)) y x)) (b a)
The final example below uses extend-syntax to implement define-structure, following a similar example using syntax-case in The Scheme Programming Language, Second Edition.
The definition of define-structure makes use of two pattern/template clauses. Two clauses are needed to handle the optionality of the second subexpression. The first clause matches the form without the second subexpression and merely converts it into the equivalent form with the second subexpression present, but empty.
The definition also makes heavy use of with to evaluate Scheme expressions at expansion time. The first four with clauses are used to manufacture the identifiers that name the automatically defined procedures. (The procedure format is particularly useful here, but it could be replaced with string-append!, using symbol->string as needed.) The first two clauses yield single identifiers (for the constructor and predicate), while the next two yield lists of identifiers (for the field access and assignment procedures). The fifth with clause (the final clause in the outer with) is used to count the total length vector needed for each instance of the structure, which must include room for the name and all of the fields. The final with clause (the only clause in the inner with) is used to create a list of vector indexes, one for each field (starting at 1, since the structure name occupies position 0).
(extend-syntax (define-structure)
[(define-structure (name id1 ...))
(define-structure (name id1 ...) ())]
[(define-structure (name id1 ...) ([id2 val] ...))
(with ([constructor
(string->symbol (format "make-~a" 'name))]
[predicate
(string->symbol (format "~a?" 'name))]
[(access ...)
(map (lambda (x)
(string->symbol
(format "~a-~a" 'name x)))
'(id1 ... id2 ...))]
[(assign ...)
(map (lambda (x)
(string->symbol
(format "set-~a-~a!" 'name x)))
'(id1 ... id2 ...))]
[count (length '(name id1 ... id2 ...))])
(with ([(index ...)
(let f ([i 1])
(if (= i 'count)
'()
(cons i (f (+ i 1)))))])
(begin
(define constructor
(lambda (id1 ...)
(let* ([id2 val] ...)
(vector 'name id1 ... id2 ...))))
(define predicate
(lambda (obj)
(and (vector? obj)
(= (vector-length obj) count)
(eq? (vector-ref obj 0) 'name))))
(define access
(lambda (obj)
(vector-ref obj index)))
...
(define assign
(lambda (obj newval)
(vector-set! obj index newval)))
...)))])
Current versions of Chez Scheme are distributed with a compatibility file containing definitions of various syntactic forms and procedures supported by earlier versions of Chez Scheme for which support has since been dropped. This file, examples/compat.ss, is typically installed in /usr/local/lib/scheme on Unix systems.
In some cases, the forms and procedures found in compat.ss have been dropped because they were infrequently used and easily written directly in Scheme. In other cases, the forms and procedures have been rendered obsolete by improvements in the system. In such cases, new code should be written to use the newer features, and older code should be rewritten if possible to use the newer features as well.
 |
Chez Scheme User's Guide © 1998 R. Kent Dybvig Cadence Research Systems http://www.scheme.com Illustrations © 1998 Jean-Pierre Hébert about this book |