s7

s7 is a Scheme interpreter intended as an extension language for other applications. It exists as just two files, s7.c and s7.h, that want only to disappear into someone else's source tree. There are no libraries, no run-time init files, and no configuration scripts. It can be built as a stand-alone interpreter (see repl). s7test.scm is a regression test for s7. A tarball is available: s7 tarball. There is an svn repository at sourceforge (the Snd project): Snd, and a git repository (just s7): git@cm-gitlab.stanford.edu:bil/s7.git s7.git. Please ignore all other "s7" github sites. Christos Vagias created a web-assembly site with a repl: https://github.com/actonDev/s7-playground/.

s7 is an extension language of Snd and sndlib (snd), Rick Taube's Common Music (commonmusic at sourceforge), Kjetil Matheussen's Radium music editor, and Iain Duncan's Scheme for Max (or Pd). There are X, Motif, and openGL bindings in libxm in the Snd tarball, or at ftp://ccrma-ftp.stanford.edu/pub/Lisp/libxm.tar.gz.

Although it is a descendant of tinyScheme, s7 is closest as a Scheme dialect to Guile 1.8. I believe it is compatible with r5rs and r7rs: you can just ignore all the additions discussed in this file. It has continuations, ratios, complex numbers, macros, keywords, hash-tables, multiprecision arithmetic, generalized set!, unicode, and so on. It does not have syntax-rules or any of its friends, and it thinks there is no such thing as an inexact integer.

This file assumes you know about Scheme and all its problems, and want a quick tour of where s7 is different. (Well, it was quick once upon a time). The main difference: if it's in s7, it's a first-class citizen of s7, and that includes macros, environments, and syntactic values.


I originally used a small font for scholia, but now I have to squint to read that tiny text, so less-than-vital commentaries are shown in the normal font, but indented and on a sort of brownish background.


multiprecision arithmetic

All numeric types, integers, ratios, reals, and complex numbers are supported. The basic integer and real types are defined in s7.h, defaulting to int64_t and double. A ratio consists of two integers, a complex number consists of two reals. pi is predefined. s7 can be built with multiprecision support for all types, using the gmp, mpfr, and mpc libraries (set WITH_GMP to 1 in s7.c). If multiprecision arithmetic is enabled, the following functions are included: bignum, and bignum?, and the variable (*s7* 'bignum-precision). (*s7* 'bignum-precision) defaults to 128; it sets the number of bits each float takes. pi automatically reflects the current (*s7* 'bignum-precision):

> pi
3.141592653589793238462643383279502884195E0
> (*s7* 'bignum-precision)
128
> (set! (*s7* 'bignum-precision) 256)
256
> pi
3.141592653589793238462643383279502884197169399375105820974944592307816406286198E0

bignum? returns #t if its argument is a big number of some type; I use "bignum" for any big number, not just integers. To create a big number, either include enough digits to overflow the default types, or use the bignum function. Its argument is either a number which it casts to a bignum, or a string representing the desired number:

> (bignum "123456789123456789")
123456789123456789
> (bignum "1.123123123123123123123123123")
1.12312312312312312312312312300000000009E0

For read-time bignums:

(set! *#readers*
  (cons (cons #\B (lambda (str)
	            (bignum (string->number (substring str 1)))))
        *#readers*))

and now #B123 is the reader equivalent of (bignum 123).

In the non-gmp case, if s7 is built using doubles (s7_double in s7.h), the float "epsilon" is around (expt 2 -53), or about 1e-16. In the gmp case, it is around (expt 2 (- (*s7* 'bignum-precision))). So in the default case (precision = 128), using gmp:

> (= 1.0 (+ 1.0 (expt 2.0 -128)))
#t
> (= 1.0 (+ 1.0 (expt 2.0 -127)))
#f

and in the non-gmp case:

> (= 1.0 (+ 1.0 (expt 2 -53)))
#t
> (= 1.0 (+ 1.0 (expt 2 -52)))
#f

In the gmp case, integers and ratios are limited only by the size of memory, but reals are limited by (*s7* 'bignum-precision). This means, for example, that

> (floor 1e56) ; (*s7* 'bignum-precision) is 128
99999999999999999999999999999999999999927942405962072064
> (set! (*s7* 'bignum-precision) 256)
256
> (floor 1e56)
100000000000000000000000000000000000000000000000000000000

The non-gmp case is similar, but it's easy to find the edge cases:

> (floor (+ 0.9999999995 (expt 2.0 23)))
8388609

math functions

s7 includes:

The random function can take any numeric argument, including 0. Other math-related differences between s7 and r5rs:

> (exact? 1.0)
#f
> (rational? 1.5)
#f
> (floor 1.4)
1
> (remainder 2.4 1)
0.4
> (modulo 1.4 1.0)
0.4
> (lcm 3/4 1/6)
3/2
> (log 8 2)
3
> (number->string 0.5 2)
"0.1"
> (string->number "0.1" 2)
0.5
> (rationalize 1.5)
3/2
> (complex 1/2 0)
1/2
> (logbit? 6 1) ; argument order, (logbit? int index), follows gmp, not CL
#t

See cload and libgsl.scm for easy access to GSL, and similarly libm.scm for the C math library.

The exponent itself is always in base 10; this follows gmp usage. Scheme normally uses "@" for its useless polar notation, but that means (string->number "1e1" 16) is ambiguous — is the "e" a digit or an exponent marker? In s7, "@" is an exponent marker.

> (string->number "1e9" 2)  ; (expt 2 9)
512.0
> (string->number "1e1" 12) ; "e" is not a digit in base 12
#f
> (string->number "1e1" 16) ; (+ (* 1 16 16) (* 14 16) 1)
481
> (string->number "1.2e1" 3); (* 3 (+ 1 2/3))
5.0

The functions nan and nan-payload refer to the "payload" that can be associated with a NaN. s7's reader can read NaNs with these payloads: +nan.123 is a NaN with payload 123. s7 displays the NaN payload in the same way: +nan.123. A NaN without any payload (or payload 0) is written +nan.0.

> (nan 123)
+nan.123
> (nan-payload (nan 123))
123

The nan function (and s7's reader) always returns a positive NaN.

What is (/ 1.0 0.0)? s7 gives a "division by zero" error here, and also in (/ 1 0). Guile returns +inf.0 in the first case, which seems reasonable, but a "numerical overflow" error in the second. Slightly weirder is (expt 0.0 0+i). Currently s7 returns 0.0, Guile returns +nan.0+nan.0i, Clisp and sbcl throw an error. Everybody agrees that (expt 0 0) is 1, and Guile thinks that (expt 0.0 0.0) is 1.0. But (expt 0 0.0) and (expt 0.0 0) return different results in Guile (1 and 1.0), both are 0.0 in s7, the first is an error in Clisp, but the second returns 1, and so on — what a mess! This mess was made a lot worse than it needs to be when the IEEE decreed that 0.0 equals -0.0, so we can't tell them apart, but that they produce different results:

scheme@(guile-user)> (= -0.0 0.0)
#t
scheme@(guile-user)> (negative? -0.0)
#f
scheme@(guile-user)> (= (/ 1.0 0.0) (/ 1.0 -0.0))
#f
scheme@(guile-user)>  (< (/ 1.0 -0.0) -1e100 1e100 (/ 1.0 0.0))
#t

How can they be equal? In s7, the sign of -0.0 is ignored, and they really are equal. One other oddity: two floats can satisfy eq? and yet not be eqv?: (eq? +nan.0 +nan.0) might be #t (it is unspecified), but (eqv? +nan.0 +nan.0) is #f. The same problem afflicts memq and assq.

The random function takes a range and an optional state, and returns a number between zero and the range, of the same type as the range. It is perfectly reasonable to use a range of 0, in which case random returns 0. random-state creates a new random state from a seed. If no seed is passed, random-state returns the current state. If no state is passed to random, it uses some default state initialized from the current time. random-state? returns #t if passed a random state object.

> (random 0)
0
> (random 1.0)
0.86331198514245
> (random 3/4)
654/1129
> (random 1+i)
0.86300308872748+0.83601002730848i
> (random -1.0)
-0.037691127513267
> (define r0 (random-state 1234))
r0
> (random 100 r0)
94
> (random 100 r0)
19
> (define r1 (random-state 1234))
r1
> (random 100 r1)
94
> (random 100 r1)
19

copy the random-state to save a spot in a random number sequence, or save the random-state as a list via random-state->list, then to restart from that point, apply random-state to that list.

In the gmp s7, random calls gmp's random number generator. There are also many generators in GSL (see libgsl.scm). In the non-gmp s7, we use Marsaglia's MWC algorithm which I think is a good compromise between quality and speed. If you're calling it more than a billion times, it is probably a good idea to reseed it. The code I use for this in the auto-tester is:

(require libc.scm)
(define (reseed)
  (let ((seed (with-let *libc*
		(let ((res (clock_gettime CLOCK_MONOTONIC)))
		  (+ (* 1000000000 (cadr res)) (caddr res)))))
	(carry (#(1791398085 1929682203 1683268614 1965537969 1675393560 1967773755 1517746329
                  1447497129 1655692410 1606218150 2051013963 1075433238 1557985959 1781943330
                  1893513180 1631296680 2131995753 2083801278 1873196400 1554115554)
		(random 20))))
    (random-state seed carry)))

I can't find the right tone for this section; this is the 400-th revision; I wish I were a better writer!

In some Schemes, "rational" means "could possibly be expressed equally well as a ratio: floats are approximations". In s7 it's: "is actually expressed (at the scheme level) as a ratio (or an integer of course)"; otherwise "rational?" is the same as "real?":

(not-s7)> (rational? (sqrt 2))
#t

That 1.0 is represented at the IEEE-float level as a sort of ratio does not mean it has to be a scheme ratio; the two notions are independent.

But that confusion is trivial compared to the completely nutty "inexact integer". As I understand it, "inexact" originally meant "floating point", and "exact" meant integer or ratio of integers. But words have a life of their own. 0.0 somehow became an "inexact" integer (although it can be represented exactly in floating point). +inf.0 must be an integer — its fractional part is explicitly zero! But +nan.0... And then there's:

(not-s7)> (integer? 9007199254740993.1)
#t

When does this matter? I often need to index into a vector, but the index is a float (a "real" in Scheme-speak: its fractional part can be non-zero). In one Scheme:

(not-s7)> (vector-ref #(0) (floor 0.1))
ERROR: Wrong type (expecting exact integer): 0.0   ; [why?  "it's probably a programmer mistake"!]

Not to worry, I'll use inexact->exact:

(not-s7)> (inexact->exact 0.1)
3602879701896397/36028797018963968                  ; [why? "floats are ratios"!]

So I end up using the verbose (floor (inexact->exact ...)) everywhere, and even then I have no guarantee that I'll get a legal vector index. I have never seen any use made of the exact/inexact distinction — just wild flailing to try get around it. I think the whole idea is confused and useless, and leads to verbose and buggy code. If we discard it, we can maintain backwards compatibility via:

(define exact? rational?)
(define (inexact? x) (not (rational? x)))
(define inexact->exact rationalize) ; or floor
(define (exact->inexact x) (* x 1.0))

Standard Scheme's #i and #e are also useless because you can have any number after, for example, #b:

> #b1.1
1.5
> #b1e2
4.0
> #o17.5+i
15.625+1i

s7 uses #i for int-vector and does not implement #e. Speaking of #b and friends, what should (string->number "#xffff" 2) return?

define*, lambda*

define* and lambda* are extensions of define and lambda that make it easier to deal with optional, keyword, and rest arguments. The syntax is very simple: every parameter to define* has a default value and is automatically available as a keyword argument. The default value is either #f if unspecified, or given in a list whose first member is the parameter name. The last parameter can be preceded by :rest or a dot to indicate that all other trailing arguments should be packaged as a list under that parameter's name. A trailing or rest parameter's default value is () and can't be specified in the declaration. The rest parameter is not available as a keyword argument.

(define* (hi a (b 32) (c "hi")) (list a b c))

Here the parameter "a" defaults to #f, "b" to 32, etc. When the function is called, the parameter names are set from the values passed the function, then any unset parameters are bound to their default values, evaluated in left-to-right order. As the current argument list is scanned, any name that occurs as a keyword, :arg for example where the parameter name is arg, sets that parameter's new value. Otherwise, as values occur, they are plugged into the actual argument list based on their position, counting a keyword/value pair as one argument. This is called an optional-key list in CLM. So, taking the function above as an example:

> (hi 1)
(1 32 "hi")
> (hi :b 2 :a 3)
(3 2 "hi")
> (hi 3 2 1)
(3 2 1)

See s7test.scm for many examples. (s7's define* is very close to srfi-89's define*). To mark an argument as required, set its default value to a call on the error function:

> (define* (f a (b (error 'unset-arg "f's b parameter not set"))) (list a b))
f
> (f 1 2)
(1 2)
> (f 1)
error: f's b parameter not set

The combination of optional and keyword arguments is viewed with disfavor in the Lisp community, but the problem is in CL's implementation of the idea, not the idea itself. I've used the s7 style since around 1976, and have never found it confusing. The mistake in CL is to require the optional arguments if a keyword argument occurs, and to consider them as distinct from the keyword arguments. So everyone forgets and puts a keyword where CL expects a required-optional argument. CL then does something ridiculous, and the programmer stomps around shouting about keywords, but the fault lies with CL. If s7's way is considered too loose, one way to tighten it might be to insist that once a keyword is used, only keyword argument pairs can follow.

A natural companion of lambda* is named let*. In named let, the implicit function's arguments have initial values, but thereafter, each call requires the full set of arguments. Why not treat the initial values as default values?

> (let* func ((i 1) (j 2))
    (+ i j (if (> i 0) (func (- i 1)) 0)))
5
> (letrec ((func (lambda* ((i 1) (j 2))
                   (+ i j (if (> i 0) (func (- i 1)) 0)))))
    (func))
5

This is consistent with the lambda* arguments because their defaults are already set in left-to-right order, and as each parameter is set to its default value, the binding is added to the default value expression environment (just as if it were a let*). The let* name itself (the implicit function) is not defined until after the bindings have been evaluated (as in named let).

In CL, keyword default values are handled in the same way:

> (defun foo (&key (a 0) (b (+ a 4)) (c (+ a 7))) (list a b c))
FOO 
> (foo :b 2 :a 60)
(60 2 67) 

In s7, we'd use:

(define* (foo (a 0) (b (+ a 4)) (c (+ a 7))) (list a b c))

Also CL and s7 handle keywords as values in the same way:

> (defun foo (&key a) a)
FOO
> (defvar x :a)
X
> (foo x 1)
1
> (define* (foo a) a)
foo
> (define x :a)
:a
> (foo x 1)
1

Keywords (named arguments) also work in named let*:

> (let* loop ((i 0) (j 0))
      (if (> i 3)
          (+ i j)
          (loop :j 2 :i (+ i 1))))
6

To try to catch what I believe are usually mistakes, I added two error checks. One is triggered if you set the same parameter twice in the same call, and the other if an unknown keyword is encountered in the key position. To turn off these errors, add :allow-other-keys at the end of the parameter list. These problems arise in a case such as

(define* (f (a 1) (b 2)) (list a b))

You could do any of the following by accident:

(f 1 :a 2)  ; what is a?
(f :b 1 2)  ; what is b?
(f :c 3)    ; did you really want a to be :c and b to be 3?

In the last case, to pass a keyword deliberately, either include the argument keyword: (f :a :c), or make the default value a keyword: (define* (f (a :c) ...)), or set (*s7* 'accept-all-keyword-arguments) to some true value. See s7test.scm for many examples.

What if two functions share a keyword argument, and one wants to call the other, passing both arguments to the wrapper?

(define* (f1 a) a)                         ; the wrappee
(define* (f2 a :rest b :allow-other-keys)  ; the wrapper
  (+ a (apply f1 b)))
(f2 :a 3 :a 4)                             ; 7, b='(:a 4)
(let ((c :a))
  (f2 c 3 c 4))                            ; also 7

Since named let* is a form of lambda*, the prohibition of repeated variable names makes it different from let*: (let* ((a 1) (a 2)) a) is 2, but (let* loop ((a 1) (a 2)) a) is an error. If let* and named let* agreed in this, we'd have an inconsistency with lambda*. If all three allowed repeated variables, the decision as to which parameter is intended becomes messy: ((lambda* (a a) a) 2 :a 3), or (let* loop ((a 1) (a 2)) (loop 2 :a 3)). CL and standard scheme accept repeated variables in let*, so I think the current choice is the least surprising.

s7's lambda* arglist handling is not the same as CL's lambda-list. First, you can have more than one :rest parameter:

> ((lambda* (:rest a :rest b) (map + a b)) 1 2 3 4 5)
'(3 5 7 9)

and second, the rest parameter, if any, takes up an argument slot just like any other argument:

> ((lambda* ((b 3) :rest x (c 1)) (list b c x)) 32)
(32 1 ())
> ((lambda* ((b 3) :rest x (c 1)) (list b c x)) 1 2 3 4 5)
(1 3 (2 3 4 5))

CL would agree with the first case if we used &key for 'c', but would give an error in the second. Of course, the major difference is that s7 keyword arguments don't insist that the key be present. The :rest argument is needed in cases like these because we can't use an expression such as:

> ((lambda* ((a 3) . b c) (list a b c)) 1 2 3 4 5)
error: stray dot?
> ((lambda* (a . (b 1)) b) 1 2) ; the reader turns the arglist into (a b 1)
error: lambda* parameter '1 is a constant

Yet another nit: the :rest argument is not considered a keyword argument, so

> (define* (f :rest a) a)
f
> (f :a 1)
(:a 1)

macros

define-macro, define-macro*, define-bacro, define-bacro*, macroexpand, gensym, gensym?, and macro? implement the standard old-time macros. The anonymous versions (analogous to lambda and lambda*) are macro, macro*, bacro, and bacro*. See s7test.scm for many examples of macros including such perennial favorites as loop, dotimes, do*, enum, pushnew, and defstruct.

> (define-macro (and-let* vars . body)
    `(let ()
       (and ,@(map (lambda (v)
                     `(define ,@v))
                   vars)
            (begin ,@body))))

macroexpand can help debug a macro. I always forget that it wants an expression:

> (define-macro (add-1 arg) `(+ 1 ,arg))
add-1
> (macroexpand (add-1 32))
(+ 1 32)

gensym returns a symbol that is guaranteed to be unique. It takes an optional string argument giving the new symbol name's prefix. gensym? returns #t if its argument is a symbol created by gensym.

(define-macro (pop! sym)
  (let ((v (gensym)))
    `(let ((,v (car ,sym)))
       (set! ,sym (cdr ,sym))
       ,v)))

As in define*, the starred forms give optional and keyword arguments:

> (define-macro* (add-2 a (b 2)) `(+ ,a ,b))
add-2
> (add-2 1 3)
4
> (add-2 1)
3
> (add-2 :b 3 :a 1)
4

A macro is a first-class citizen of s7. You can pass it as a function argument, apply it to a list, return it from a function, call it recursively, and assign it to a variable. You can even set its setter!

> (define-macro (hi a) `(+ ,a 1))
hi
> (apply hi '(4))
5
> (define (fmac mac) (apply mac '(4)))
fmac
> (fmac hi)
5
> (define (fmac mac) (mac 4))
fmac
> (fmac hi)
5
> (define (make-mac)
    (define-macro (hi a) `(+ ,a 1)))
make-mac
> (let ((x (make-mac)))
    (x 2))
3
> (define-macro (ref v i) `(vector-ref ,v ,i))
ref
> (define-macro (set v i x) `(vector-set! ,v ,i ,x))
set
> (set! (setter ref) set)
set
> (let ((v (vector 1 2 3))) (set! (ref v 0) 32) v)
#(32 2 3)

To expand all the macros in a piece of code:

(define-macro (fully-macroexpand form)
  (list 'quote
    (let expand ((form form))
      (cond ((not (pair? form)) form)
            ((and (symbol? (car form))
                  (macro? (symbol->value (car form))))
             (expand (apply macroexpand (list form))))
            ((and (eq? (car form) 'set!)  ; look for (set! (mac ...) ...) and use mac's setter
                  (pair? (cdr form))
                  (pair? (cadr form))
                  (macro? (symbol->value (caadr form))))
	     (expand (apply macroexpand (list (cons (setter (symbol->value (caadr form)))
							 (append (cdadr form) (copy (cddr form))))))))
            (else (cons (expand (car form)) (expand (cdr form))))))))

This does not always handle bacros correctly because their expansion can depend on the run-time state.

A bacro is a macro that expands its body and evaluates the result in the calling environment.

(define setf
  (let ((args (gensym))
        (name (gensym)))
     (apply define-bacro `((,name . ,args)
			   (unless (null? ,args)
			     (apply set! (car ,args) (cadr ,args) ())
			     (apply setf (cddr ,args)))))))

The setf argument is a gensym (created when setf is defined) so that its name does not shadow any existing variable. Bacros expand in the calling environment, and a normal argument name might shadow something in that environment while the bacro is being expanded. Similarly, if you introduce bindings in the bacro expansion code, you need to keep track of which environment you want things to happen in. Use with-let and gensym liberally. stuff.scm has bacro-shaker which can find inadvertent name collisions, but it is flighty and easily confused. The calling environment itself is (outlet (curlet)) from within a bacro, so

(define-bacro (holler)
  `(format *stderr* "(~S~{ ~S ~S~^~})~%"
	   (let ((f (*function*)))
	     (if (pair? f) (car f) f))
	   (map (lambda (slot)
		  (values (symbol->keyword (car slot)) (cdr slot)))
		(map values ,(outlet (curlet))))))

(define (f1 a b)
  (holler)
  (+ a b))

(f1 2 3) ; prints out "(f1 :a 2 :b 3)" and returns 5

Since a bacro (normally) sheds its define-time environment:

(define call-bac
  (let ((x 2))
    (define-bacro (m a) `(+ ,a ,x))))

> (call-bac 1)
error: x: unbound variable

A macro here returns 3. The bacro can get its define-time environment (its closure) via funclet, so define-macro is a special case of define-bacro! We can define macros that work in all four ways: the expansion can happen in either the definition or calling environment, as can the evaluation of that expansion. In a bacro, both happen in the calling environment if we take no other action, and in a normal macro (define-macro), the expansion happens in the definition environment, and the evaluation in the calling environment. Here's a brief example of all four:

(let ((x 1) (y 2))
  (define-bacro (bac1 a)
     `(+ ,x y ,a))        ; expand and eval in calling env
  (let ((x 32) (y 64))
    (bac1 3)))            ; (with-let (inlet 'x 32 'y 64) (+ 32 y 3))
-> 99                     ;  with-let and inlet refer to environments

(let ((x 1) (y 2))        ; this is like define-macro
  (define-bacro (bac2 a)
    (with-let (sublet (funclet bac2) :a a)
      `(+ ,x y ,a)))      ; expand in definition env, eval in calling env
  (let ((x 32) (y 64))
    (bac2 3)))            ; (with-let (inlet 'x 32 'y 64) (+ 1 y 3))
-> 68

(let ((x 1) (y 2))
  (define-bacro (bac3 a)
    (let ((e (with-let (sublet (funclet bac3) :a a)
	       `(+ ,x y ,a))))
      `(with-let ,(sublet (funclet bac3) :a a)
	 ,e)))           ; expand and eval in definition env
  (let ((x 32) (y 64))
    (bac3 3)))           ; (with-let (inlet 'x 1 'y 2) (+ 1 y 3))
-> 6

(let ((x 1) (y 2))
  (define-bacro (bac4 a)
    (let ((e `(+ ,x y ,a)))
      `(with-let ,(sublet (funclet bac4) :a a)
	 ,e)))           ; expand in calling env, eval in definition env
  (let ((x 32) (y 64))
    (bac4 3)))           ; (with-let (inlet 'x 1 'y 2) (+ 32 y 3))
-> 37

Backquote (quasiquote) in s7 is almost trivial. Constants are unchanged, symbols are quoted, ",arg" becomes "arg", and ",@arg" becomes "(apply values arg)" — hooray for real multiple values! It's almost as easy to write the actual macro body as the backquoted version of it.

> (define-macro (hi a) `(+ 1 ,a))
hi
> (procedure-source hi)
(lambda (a) (list-values '+ 1 a))

> (define-macro (hi a) `(+ 1 ,@a))
hi
> (procedure-source hi)
(lambda (a) (list-values '+ 1 (apply-values a)))

list-values and apply-values are quasiquote helper functions described below. There is no unquote-splicing macro in s7; ",@(...)" becomes "(unquote (apply-values ...))" at read-time. There shouldn't be any unquote either. In Scheme the reader turns ,x into (unquote x), so:

> (let (,'a) unquote)
a
> (let (, (lambda (x) (+ x 1))) ,,,,'3)
7

comma becomes a sort of symbol macro! I think I'll remove unquote; ,x and ,@x will still work as expected, but there will not be any "unquote" or "unquote-splicing" in the resultant source code.

hygienic macros

tldr: To make a macro hygienic, in the expanded code replace the names of things with the things themselves. For built-in values, use "#_", e.g. "#_abs" rather than "abs". For macro-definition-time variables, use ",abs" rather than "abs". If possible, replace other names with gensyms. As a last resort, use the explicit environment functions like let-ref and with-let.

s7 macros are not hygienic. For example,

> (define-macro (mac b)
    `(let ((a 12))
       (+ a ,b)))
mac
> (let ((a 1) (+ *)) (mac a))
144

This returns 144 because '+' has turned into '*', and 'a' is the internal 'a', not the argument 'a'. We get (* 12 12) where we might have expected (+ 12 1). Starting with the '+' problem, as long as the redefinition of '+' is local (that is, it happens after the macro definition), we can unquote the +:

> (define-macro (mac b)
    `(let ((a 12))
       (,+ a ,b))) ; ,+ picks up the definition-time +
mac
> (let ((a 1) (+ *)) (mac a))
24                 ; (+ a a) where a is 12

But the unquote trick won't work if we have previously loaded some file that redefined '+' at the top-level (so at macro definition time, + is *, but we want the built-in +). Although this example is silly, the problem is real in Scheme because Scheme has no reserved words and only one name space.

> (define + *)
+
> (define (add a b) (+ a b))
add
> (add 2 3)
6
> (define (divide a b) (/ a b))
divide
> (divide 2 3)
2/3
> (set! / -) ; a bad idea — this turns off s7's optimizer
-
> (divide 2 3)
-1

Obviously macros are not the problem here. Since we might be loading code written by others, it's sometimes hard to tell what names that code depends on or redefines. We need a way to get the pristine (start-up, built-in) value of '+'. One long-winded way in s7 uses unlet:

> (define + *)
+
> (define (add a b) (with-let (unlet) (+ a b)))
add
> (add 2 3)
5

But this is hard to read, and we might want all three values of a symbol, the start-up value, the definition-time value, and the current value. The latter can be accessed with the bare symbol, the definition-time value with unquote (','), and the start-up value with either unlet or #_<name>. That is, #_+ is a reader macro for (with-let (unlet) +).

> (define-macro (mac b)
    `(#_let ((a 12))
       (#_+ a ,b))) ; #_+ and #_let are start-up values
mac
> (let ((a 1) (+ *)) (mac a))
24                 ; (+ a a) where a is 12 and + is the start-up +

;;; make + generic (there's a similar C-based example below)
> (define (+ . args)
    (if (null? args) 0
        (apply (if (number? (car args)) #_+ #_string-append) args)))
+
> (+ 1 2)
3
> (+ "hi" "ho")
"hiho"

Conceptually, #_<name> could be implemented via *#readers*:

(set! *#readers*
  (cons (cons #\_ (lambda (str)
		    (with-let (unlet)
                      (string->symbol (substring str 1)))))
	*#readers*))

but s7 doesn't let you change the meaning of #\_; otherwise:

(set! *#readers* (list (cons #\_ (lambda (str) (string->symbol (substring str 1))))))

and now #_ provides no protection:

> (let ((+ -)) (#_+ 1 2))
-1

#t and #f (along with their stupid r7rs cousins #true and #false) are also not settable.

So, now we have only the variable capture problem ('a' has been captured in the preceding examples). This is the only thing that the gigantic "hygienic macro" systems actually deal with: a microscopic problem that you'd think, from the hype, was up there with malaria and the national debt. gensym is the standard approach:

> (define-macro (mac b)
    (let ((var (gensym)))
      `(#_let ((,var 12))
         (#_+ ,var ,b))))
mac
> (let ((a 1) (+ *)) (mac a))
13

;; or use lambda:
> (define-macro (mac b)
  `((lambda (b) (let ((a 12)) (#_+ a b))) ,b))
mac
> (let ((a 1) (+ *)) (mac a))
13

I think syntax-rules and its friends try to conjure up gensyms automatically, but the real problem is not name collisions, but unspecified environments. In s7 we have first-class environments, so you have complete control over the environment at any point:

(define-macro (mac b)
  `(with-let (inlet 'b ,b)
     (let ((a 12))
       (+ a b))))

> (let ((a 1) (+ *)) (mac a))
13

(define-macro (mac1 . b)         ; originally `(let ((a 12)) (+ a ,@b ,@b))
  `(with-let (inlet 'e (curlet)) ; this 'e will not collide with the calling env
     (let ((a 12))               ;   nor will 'a (so no gensyms are needed etc)
       (+ a (with-let e ,@b) (with-let e ,@b)))))

> (let ((a 1) (e 2)) (mac1 (display a) (+ a e)))
18  ; (and it displays "11")

(define-macro (mac2 x)           ; this will use mac2's definition environment for its body
  `(with-let (sublet (funclet mac2) :x ,x)
     (let ((a 12))
       (+ a b x))))              ; a is always 12, b is whatever b happens to be in mac2's env

> (define b 10)                  ; this is mac2's b
10
> (let ((+ *) (a 1) (b 15)) (mac2 (+ a b)))
37                               ; mac2 uses its own a (12), b (10), and + (+)
                                 ;   but (+ a b) is 15 because at that point + is *: (* 1 15)

Hygienic macros are trivial! Who needs syntax-rules? Here's an example of the while macro in stuff.scm before and after hygienification (using "exit" rather than "break"):

(define-macro (innocent-while test . body)
  `(call-with-exit
    (lambda (exit)
      (let loop ()
	(call-with-exit
	 (lambda (continue)
	   (do () ((not ,test) (exit))
	     ,@body)))
	(loop)))))

(define-macro (cynical-while test . body)
  (let ((loop (gensym)))
    `(#_call-with-exit
       (#_lambda (exit)
         (#_let ,loop ()
           (#_call-with-exit
	     (#_lambda (continue)
	       (#_do () ((#_not ,test) (exit))
	         ,@body)))
	   (,loop))))))

There are many more examples in s7test.scm (see especially the "or" macros toward the end of s7test.scm).

(define-macro (swap a b) ; assume a and b are symbols
  `(with-let (inlet 'e (curlet) 'tmp ,a)
     (set! (e ',a) (e ',b))
     (set! (e ',b) tmp)))

> (let ((b 1) (tmp 2)) (swap b tmp) (list b tmp))
(2 1)

(define-macro (swap a b) ; here a and b can be any settable expressions
  `(set! ,b (with-let (inlet 'e (curlet) 'tmp ,a)
	      (with-let e (set! ,a ,b))
	      tmp)))

> (let ((v (vector 1 2))) (swap (v 0) (v 1)) v)
#(2 1)
> (let ((tmp (cons 1 2))) (swap (car tmp) (cdr tmp)) tmp)
(2 . 1)

(set! (setter swap) (define-macro (set-swap a b c) `(set! ,b ,c)))

> (let ((a 1) (b 2) (c 3) (d 4)) (swap a (swap b (swap c d))) (list a b c d))
(2 3 4 1)

;;; but this is simpler:
(define-macro (rotate! . args)
  `(set! ,(args (- (length args) 1))
         (with-let (inlet 'e (curlet) 'tmp ,(car args))
	   (with-let e
	     ,@(map (lambda (a b) `(set! ,a ,b)) args (cdr args)))
	   tmp)))

> (let ((a 1) (b 2) (c 3)) (rotate! a b c) (list a b c))
(2 3 1)
(let ()
  (define (f23 y) (+ y 1))          ; the first f23
  (define-macro (m1 x) `(f23 ,x))   ; picks up f23 from whatever the local env is where m1 is expanded
  (m1 3) ; 4

  (let ((f23 (lambda (y) (+ y 2)))) ; the second f23
    (m1 3) ; 5
    (set! (symbol-initial-value :f23) f23))
    ;; this remembers the second f23 as #_:f23 or (symbol-initial-value :f23)

  (define-macro (m2 x) `(,f23 ,x)) ; ",f23" picks up f23 from m2's definition-time environment

  (define e1 #f)
  (let ((f23 (lambda (y) (+ y 3))))  ; the third f23
    (set! e1 (curlet))      ; save the environment holding the third f23
    (m2 3)) ; 4             ; y + 1 because at this point the definition env f23 is the first

  (set! (symbol-initial-value 'f23) f23) ; #_f23 now refers to the first f23
  (set! f23 (lambda (y) (+ y 4)))        ; the fourth f23

  (define-macro (m3 x) `((#_symbol-initial-value 'f23) ,x))
  ;; this picks up the first f23 by delaying the reference to it until evaluation time.
  ;; Using `(#_f23 ,x) here will fail unless f23's symbol-initial-value is set at the top-level
  ;; because the reader only knows about global values:

  (define-macro (m4 x) `(#_f23 ,x))
  ;; that is, when this definition is encountered, the reader sees the #_f23 without knowing anything about its
  ;; context (this is while reading the let form, before the symbol-initial-value is actually set).
  ;; When the let is later evaluated, the m4 code has already become `(#<undefined: f23> ,x)
  ;; so we get the error: "attempt to apply an undefined object #_f23 in (#_f23 3)?"

  (define-macro (m5 x) `((#_symbol-initial-value :f23) ,x))  ; this is a reference to the second f23
  ;; 'f23 and :f23 can have different initial-values

  (define-macro (m6 x) `((,e1 'f23) ,x)) ; use e1 to get the third f23
  ;; more hygienic: set e1's symbol-initial-value to itself above, then use (#_symbol-initial-value 'e1) here

  (let ((f23 (lambda (y) (+ y 5))))   ; the fifth f23
    (m1 3)   ; 8 = 3 + 5 from local (fifth) f23
    (m2 3)   ; 7 = 3 + 4 from definition environment f23 (the fourth)
    (m3 3)   ; 4 = 3 + 1 from the first f23 (before the (set! f23 ...))
    (catch #t (lambda () (m4 3)) (lambda (type info) (apply format #f info))))) ; error given above
    (m5 3)   ; 5 = 3 + 2 from second f23
    (m6 3))) ; 6 = 3 + 3 from third f23

On the subject of *#readers*, say we have:

(set! *#readers* (list (cons #\o (lambda (str) 42))  ; #o... -> 42
                       (cons #\x (lambda (str) 3)))) ; #x... -> 3

Now we load a file with:

(define (oct) #o123)

(let-temporarily ((*#readers* ()))
  (eval (with-input-from-string "(define (hex) #x123)" read)))

(define-constant old-readers *#readers*)
(set! *#readers* ())

(define (oct1) #o123)
(define (hex1) #x123)

(set! *#readers* old-readers)

(define (oct2) #o123)
(define (hex2) #x123)

Now we evaluate these functions, and get:

(oct): 42   ; oct is not read-time hygienic so #o123 -> 42
(oct1): 83  ; oct1 is protected by the top-level set, #o123 -> 83
(oct2): 42  ; same as oct
(hex): 291  ; hex is protected by let-temporarily + read
(hex1): 291 ; hex1 is like oct1
(hex2): 3   ; hex2 is like oct

Here is Peter Seibel's wonderful once-only macro:

(define-macro (once-only names . body)
  (let ((gensyms (map (lambda (n) (gensym)) names)))
    `(let (,@(map (lambda (g) (list g '(gensym))) gensyms))
       `(let (,,@(map (lambda (g n) (list list g n)) gensyms names))
          ,(let (,@(map list names gensyms))
             ,@body)))))

From the land of sparkling bacros:

(define once-only
  (let ((names (gensym))
	(body (gensym)))
    (apply define-bacro `((,(gensym) ,names . ,body)
      `(let (,@(map (lambda (name) `(,name ,(eval name))) ,names))
	 ,@,body)))))

Sadly, with-let is simpler.

setter

(setter proc)
(dilambda proc setter)

There are several kinds of setters, reflecting the many ways that set! can be called. First are the symbol setters:

> (let ((x 1))
    (set! (setter 'x) (lambda (name new-value) (* new-value 2)))
    (set! x 2)
    x)
4

Here the setter is a function that is called before the variable is set. It can take two or three arguments. In the two argument case shown above, the first is the variable name (a symbol), and the second is the new-value. The variable is set to the value returned by the setter function. When s7 sees (set! x 2) above, it calls the setter which returns 4. So x is set to 4.

In some cases you need the environment that the variable lives in (to get its current value for example), so you can include that in the setter function parameter list:

> (let ((x 1))
    (set! (setter 'x) (lambda (name new-value enviroment) (* new-value 2)))
    (set! x 2)
    x)
4

(define-macro (watch var) ; notification if 'var is set!
  `(set! (setter ',var)
      (lambda (s v e)
	(format *stderr* "~S set! to ~S~A~%" s v
                (let ((func (with-let e (*function*))))
                  (if (eq? func #<undefined>) "" (format #f ", ~S" func))))
	v)))

Since symbol setters are often implementing type restrictions, you can use the built-in type checking functions such as integer? as a short-hand for a setter that insists the new value be an integer:

> (let ((x 1))
    (set! (setter 'x) integer?)
    (set! x 3.14))
error: set! x: 3.14, is a real but should be an integer

;;; use typed-let from stuff.scm to do the same thing:
> (typed-let ((x 3 integer?))
    (set! x 3.14))
error: set! x: 3.14, is a real but should be an integer
;; see also typed-lambda in stuff.scm

C-side symbol setters go through s7_set_setter. There is an example below.

The second case is a function setter. Almost any function or macro can have an associated setter that is invoked when the function is the target of set!. In this case, the setter function does the set! itself (unlike a symbol setter):

> (setter cadr)
#f         ; by default cadr has no setter so (set! (cadr p) x) is an error
> (set! (setter cadr)  ; add a setter to cadr
        (lambda (lst val)
          (set! (car (cdr lst)) val)))
#<lambda (lst val)>
> (procedure-source (setter cadr))
(lambda (lst val) (set! (car (cdr lst)) val))
> (let ((lst (list 1 2 3)))
    (set! (cadr lst) 4)
    lst)
(1 4 3)

In some cases, the setter needs to be a macro:

> (set! (setter logbit?)
          (define-macro (m var index on) ; here we want to set "var", so we need a macro
	    `(if ,on
	         (set! ,var (logior ,var (ash 1 ,index)))
	         (set! ,var (logand ,var (lognot (ash 1 ,index)))))))
m
> (define (mingle a b)
    (let ((r 0))
      (do ((i 0 (+ i 1)))
          ((= i 31) r)
        (set! (logbit? r (* 2 i)) (logbit? a i))
        (set! (logbit? r (+ (* 2 i) 1)) (logbit? b i)))))
mingle
> (mingle 6 3) ; the INTERCAL mingle operator?
30

dilambda defines a function (or macro) and its setter without having to set! the setter by hand:

> (define f (let ((x 123))
                 (dilambda (lambda ()
                             x)
		           (lambda (new-value)
                             (set! x new-value)))))
f
> (f)
123 ; x = 123
> (set! (f) 32)
32  ; now x = 32
> (f)
32

Here is a pretty example of dilambda:

(define-macro (c?r path)
  ;; "path" is a list and "X" marks the spot in it that we are trying to access
  ;; (a (b ((c X)))) — anything after the X is ignored, other symbols are just placeholders
  ;; c?r returns a dilambda that gets/sets X

  (define (X-marks-the-spot accessor tree)
    (if (eq? tree 'X)
        accessor
        (and (pair? tree)
	     (or (X-marks-the-spot (cons 'car accessor) (car tree))
	         (X-marks-the-spot (cons 'cdr accessor) (cdr tree))))))

  (let ((body 'lst))
    (for-each
     (lambda (f)
       (set! body (list f body)))
     (reverse (X-marks-the-spot () path)))

    `(dilambda
      (lambda (lst)
	,body)
      (lambda (lst val)
	(set! ,body val)))))

> ((c?r (a b (X))) '(1 2 (3 4) 5))
3
> (let ((lst (list 1 2 (list 3 4) 5)))
   (set! ((c?r (a b (X))) lst) 32)
   lst)
(1 2 (32 4) 5)
> (procedure-source (c?r (a b (X))))
(lambda (lst) (car (car (cdr (cdr lst)))))
> ((c?r (a b . X)) '(1 2 (3 4) 5))
((3 4) 5)
> (let ((lst (list 1 2 (list 3 4) 5)))
   (set! ((c?r (a b . X)) lst) '(32))
   lst)
(1 2 32)
> (procedure-source (c?r (a b . X)))
(lambda (lst) (cdr (cdr lst)))
> ((c?r (((((a (b (c (d (e X)))))))))) '(((((1 (2 (3 (4 (5 6))))))))))
6
> (let ((lst '(((((1 (2 (3 (4 (5 6)))))))))))
    (set! ((c?r (((((a (b (c (d (e X)))))))))) lst) 32)
    lst)
(((((1 (2 (3 (4 (5 32)))))))))
> (procedure-source (c?r (((((a (b (c (d (e X)))))))))))
(lambda (lst) (car (cdr (car (cdr (car (cdr (car (cdr (car (cdr (car (car (car (car lst)))))))))))))))

I may remove dilambda and dilambda? someday; they are trivial:

(define (dilambda get set) (set! (setter get) set) get)
(define dilambda? setter)

When a function setter is called, (set! (func ...) val) is evaluated by s7 as ((setter func) ... val), so the setter function needs to handle both the inner arguments to the function and the new value.

(let ((x 123))
  (define (f a b) (+ x a b))
  (set! (setter f) (lambda (a b val) (set! x val)))
  (display (f 1 2)) (newline) ; "126"
  (set! (f 1 2) 32)
  (display (f 1 2)) (newline)) ; "35"

A third type of setter handles vector element type and hash-table key and value types. These are described under typed vectors and typed hash-tables.


Speaking of INTERCAL, COME-FROM:

(define-macro (define-with-goto-and-come-from name-and-args . body)
  (let ((labels ())
	(gotos ())
	(come-froms ()))

    (let collect-jumps ((tree body))
      (when (pair? tree)
	(when (pair? (car tree))
	  (case (caar tree)
	    ((label)     (set! labels (cons tree labels)))
	    ((goto)      (set! gotos (cons tree gotos)))
	    ((come-from) (set! come-froms (cons tree come-froms)))
	    (else (collect-jumps (car tree)))))
	(collect-jumps (cdr tree))))

    (for-each
     (lambda (goto)
       (let* ((name (cadr (cadar goto)))
	      (label (member name labels (lambda (a b) (eq? a (cadr (cadar b)))))))
	 (if label
	     (set-cdr! goto (car label))
	     (error 'bad-goto "can't find label: ~S" name))))
     gotos)

    (for-each
     (lambda (from)
       (let* ((name (cadr (cadar from)))
	      (label (member name labels (lambda (a b) (eq? a (cadr (cadar b)))))))
	 (if label
	     (set-cdr! (car label) from)
	     (error 'bad-come-from "can't find label: ~S" name))))
     come-froms)

    `(define ,name-and-args
       (let ((label (lambda (name) #f))
	     (goto (lambda (name) #f))
	     (come-from (lambda (name) #f)))
	 ,@body))))

applicable objects, generalized set!, generic functions

A procedure with a setter can be viewed as one generalization of set!. Another treats objects as having predefined get and set functions. In s7 lists, strings, vectors, hash-tables, environments, and any cooperating C or Scheme-defined objects are both applicable and settable. newLisp calls this implicit indexing, Kawa has it, Gauche implements it via object-apply, Guile via procedure-with-setter; CL's funcallable instance might be the same idea.

In (vector-ref #(1 2) 0), for example, vector-ref is just a type declaration. But in Scheme, type declarations are unnecessary, so we get exactly the same result from (#(1 2) 0). Similarly, (lst 1) is the same as (list-ref lst 1), and (set! (lst 1) 2) is the same as (list-set! lst 1 2). I like this syntax: the less noise, the better!

Well, maybe applicable strings look weird: ("hi" 1) is #\i, but worse, so is (cond (1 => "hi"))! Even though a string, list, or vector is "applicable", it is not currently considered to be a procedure, so (procedure? "hi") is #f. map and for-each, however, accept anything that apply can handle, so (map #(0 1) '(1 0)) is '(1 0). (On the first call to map in this case, you get the result of (#(0 1) 1) and so on). string->list, vector->list, and let->list are (map values object). Their inverses are (and always have been) equally trivial.

The applicable object syntax makes it easy to write generic functions. For example, s7test.scm has implementations of Common Lisp's sequence functions. length, copy, reverse, fill!, iterate, map and for-each are generic in this sense (map always returns a list).

> (map (lambda (a b) (- a b)) (list 1 2) (vector 3 4))
(5 -3 9)
> (length "hi")
2

Here's a generic FFT:

(define* (cfft data n (dir 1)) ; complex data
  (unless n (set! n (length data)))
  (do ((i 0 (+ i 1))
       (j 0))
      ((= i n))
    (if (> j i)
	(let ((temp (data j)))
	  (set! (data j) (data i))
	  (set! (data i) temp)))
    (do ((m (/ n 2) (/ m 2)))
        ((not (<= 2 m j))
         (set! j (+ j m)))
     (set! j (- j m))))
  (do ((ipow (floor (log n 2)))
       (prev 1)
       (lg 0 (+ lg 1))
       (mmax 2 (* mmax 2))
       (pow (/ n 2) (/ pow 2))
       (theta (complex 0.0 (* pi dir)) (* theta 0.5)))
      ((= lg ipow))
    (do ((wpc (exp theta))
         (wc 1.0)
         (ii 0 (+ ii 1)))
	((= ii prev)
	 (set! prev mmax))
      (do ((jj 0 (+ jj 1))
           (i ii (+ i mmax))
           (j (+ ii prev) (+ j mmax)))
          ((>= jj pow)
	   (set! wc (* wc wpc)))
        (let ((tc (* wc (data j))))
          (set! (data j) (- (data i) tc))
          (set! (data i) (+ (data i) tc))))))
  data)

> (cfft (list 0.0 1+i 0.0 0.0))
(1+1i -1+1i -1-1i 1-1i)
> (cfft (vector 0.0 1+i 0.0 0.0))
#(1+1i -1+1i -1-1i 1-1i)

And a generic function that copies one sequence's elements into another sequence:

(define (copy-into source dest) ; this is equivalent to (copy source dest)
  (do ((i 0 (+ i 1)))
      ((= i (min (length source) (length dest)))
       dest)
    (set! (dest i) (source i))))

but that is already built-in as the two-argument version of the copy function.

There is one place where list-set! and friends are not the same as set!: the former evaluate their first argument, but set! does not (with a quibble; see below):

> (let ((str "hi")) (string-set! (let () str) 1 #\a) str)
"ha"
> (let ((str "hi")) (set! (let () str) 1 #\a) str)
;((let () str) 1 #\a): too many arguments to set!
> (let ((str "hi")) (set! ((let () str) 1) #\a) str)
"ha"
> (let ((str "hi")) (set! (str 1) #\a) str)
"ha"

set! looks at its first argument to decide what to set. If it's a symbol, no problem. If it's a pair, set! looks at its car to see if it is some object that has a setter. If the car is itself a list, set! evaluates the internal expression, and tries again. So the second case above is the only one that won't work. And of course:

> (let ((x (list 1 2)))
    (set! ((((lambda () (list x))) 0) 0) 3)
    x)
(3 2)

By my count, around 20 of the Scheme built-in functions are already generic in the sense that they accept arguments of many types (leaving aside the numeric and type checking functions, take for example equal?, display, member, assoc, apply, eval, quasiquote, and values). s7 extends that list with map, for-each, reverse, and length, and adds a few others such as copy, fill!, sort!, object->string, object->let, vector-length and friends, and append. newLisp takes a more radical approach than s7: it extends operators such as '>' to compare strings and lists, as well as numbers. In map and for-each, however, you can mix the argument types, so I'm not as attracted to making '>' generic; you can't, for example, (> "hi" 32.1), or even (> 1 0+i).

The somewhat non-standard generic sequence functions in s7 are:

(sort! sequence less?)
(reverse! sequence) and (reverse sequence)
(fill! sequence value (start 0) end)
(copy obj) and (copy source destination (start 0) end)
(object->string obj)
(object->let obj)
(length obj)
(append . sequences)
(map func . sequences) and (for-each func . sequences)
(equivalent? obj1 obj2)

copy returns a (shallow) copy of its argument. If a destination is provided, it need not match the source in size or type. The start and end indices refer to the source.

> (copy '(1 2 3 4) (make-list 2))
(1 2)
> (copy #(1 2 3 4) (make-list 5) 1) ; start at 1 in the source
(2 3 4 #f #f)
> (copy "1234" (make-vector 2))
#(#\1 #\2)
> (define lst (list 1 2 3 4 5))
(1 2 3 4 5)
> (copy #(8 9) (cddr lst))
(8 9 5)
> lst
(1 2 8 9 5)

reverse! is an in-place version of reverse. That is, it modifies the sequence passed to it in the process of reversing its contents. If the sequence is a list, remember to use set!: (set! p (reverse! p)). This is somewhat inconsistent with other cases, but historically, lisp programmers have treated the in-place reverse as the fast version, so s7 follows suit.

> (define lst (list 1 2 3))
(1 2 3)
> (reverse! lst)
(3 2 1)
> lst
(1)

Leaving aside the weird list case, append returns a sequence of the same type as its first argument.

> (append #(1 2) '(3 4))
#(1 2 3 4)
> (append (float-vector) '(1 2) (byte-vector 3 4))
(float-vector 1.0 2.0 3.0 4.0)

sort! sorts a sequence using the function passed as its second argument:

> (sort! (list 3 4 8 2 0 1 5 9 7 6) <)
(0 1 2 3 4 5 6 7 8 9)

sort! calls qsort or qsort_r; if the sequence is large (more than 1024 elements?), qsort_r may allocate an internal array; if the comparison function raises an error, this internal array will probably not be freed.

Underlying some of these functions are generic iterators, also built-into s7:

(make-iterator sequence carrier)
(iterator? obj)
(iterate iterator)
(iterator-sequence iterator)
(iterator-at-end? iterator)

make-iterator takes a sequence argument and returns an iterator object that traverses that sequence as it is called. The iterator itself can be treated as a function of no arguments, or (for code clarity) it can be the argument to iterate, which does the same thing. That is (iter) is the same as (iterate iter). The sequence that an iterator is traversing is iterator-sequence.

If the sequence is a hash-table or let, the iterator normally returns a cons of the key and value. There are many cases where this overhead is objectionable, so make-iterator takes a third optional argument, carrier, which should be the cons to use. Similarly, for int, float, and complex vectors, the carrier argument can be #t which tells s7 to use a mutable number of the correct type. In both cases, the returned carrier is the same across all iterator calls, so copy the carrier value if you need to save it.

When an iterator reaches the end of its sequence, it returns #<eof>. It used to return nil; I can't decide whether this change is an improvement. If an iterator over a list notices that its list is circular, it returns #<eof>. map and for-each use iterators, so if you pass a circular list to either, it will stop eventually. (An arcane consequence for method writers: specialize make-iterator, not map or for-each).

(define (find-if f sequence)
  (let ((iter (make-iterator sequence)))
    (do ((x (iter) (iter)))
	((or (eof-object? x) (f x))
	 (and (not (eof-object? x)) x)))))

But of course a sequence might contain #<eof>! So to be really safe, use iterator-at-end? instead of eof-object?.

The argument to make-iterator can also be a function or macro. In this case, to be acceptable to iterate, the closure's environment must have a variable named '+iterator+ with a non-#f value:

(define (make-circular-iterator obj)
  (let ((iter (make-iterator obj)))
    (make-iterator
     (let ((+iterator+ #t))
       (lambda ()
         (case (iter)
           ((#<eof>) ((set! iter (make-iterator obj))))
           (else)))))))

The +iterator+ variable is similar to the '+documentation+ variable used by documentation. It gives make-iterator some hope of catching inadvertent bogus function arguments that would otherwise cause an infinite loop. But unfortunately it can escape and infect other functions:

(with-let (let ((+iterator+ #t))
            (lambda () #<eof>))             ; we intended this to be our iterator
  (concatenate vector (lambda a (copy a)))) ; from stuff.scm
  ;; (lambda a (copy a)) is also considered an iterator by map (in sequences->list) because
  ;; the local +iterator+ is #t. "a" is () because there are no further arguments to
  ;; concatenate, so (lambda a (copy a)) is generating infinitely many ()'s and this
  ;; code eventually dies with a heap overflow!

multidimensional vectors

s7 supports vectors with any number of dimensions. It is here, in particular, that generalized set! shines. make-vector's first argument can be a list of dimensions, rather than an integer as in the one dimensional case:

(make-vector (list 2 3 4))
(make-vector '(2 3) 1.0)
(vector-dimensions (make-vector '(2 3 4))) -> (2 3 4)

The second example includes the optional initial element. (vect i ...) or (vector-ref vect i ...) return the given element, and (set! (vect i ...) value) and (vector-set! vect i ... value) set it. vector-length (or just length) returns the total number of elements. vector-dimensions returns a list of the dimensions; vector-rank returns the length of this list, and vector-dimension returns the nth member of the list (the size of the nth dimension).

> (define v (make-vector '(2 3) 1.0))
#2d((1.0 1.0 1.0) (1.0 1.0 1.0))
> (set! (v 0 1) 2.0)
#2d((1.0 2.0 1.0) (1.0 1.0 1.0))
> (v 0 1)
2.0
> (vector-length v)
6

This function initializes each element of a multidimensional vector:

(define (make-array dims . inits)
  (subvector (apply vector (flatten inits)) 0 (apply * dims) dims))

> (make-array '(3 3) '(1 1 1) '(2 2 2) '(3 3 3))
#2d((1 1 1) (2 2 2) (3 3 3))

make-int-vector, make-float-vector, make-complex-vector and make-byte-vector produce homogeneous vectors holding s7_ints, s7_doubles, s7_complexs or unsigned bytes.

(make-vector length-or-list-of-dimensions initial-value element-type-function)
(vector-dimensions vect)
(vector-dimension vect n)
(vector-rank obj)
(vector-typer obj)

(float-vector? obj)
(float-vector . args)
(make-float-vector len (init 0.0))
(float-vector-ref obj . indices)
(float-vector-set! obj indices[...] value)

(complex-vector? obj)
(complex-vector . args)
(make-complex-vector len (init 0.0))
(complex-vector-ref obj . indices)
(complex-vector-set! obj indices[...] value)

(int-vector? obj)
(int-vector . args)
(make-int-vector len (init 0))
(int-vector-ref obj . indices)
(int-vector-set! obj indices[...] value)

(byte-vector? obj)
(byte-vector . args)
(make-byte-vector len (init 0))
(byte-vector-ref obj . indices)
(byte-vector-set! obj indices[...] byte)
(byte? obj)

(string->byte-vector str)
(byte-vector->string str)

(subvector vector start end dimensions)
(subvector? obj)
(subvector-vector obj)
(subvector-position obj)

In addition to the dimension list mentioned above, make-vector accepts optional arguments giving the initial element and the element type. If the type is given, every attempt to set an element of the vector first calls the type function on the new value. If the type function is omitted (or set to #t), no type checking is performed. If the type function is a closure (rather than a C-defined or built-in function), its name must be accessible; it can't be an anonymous lambda (the signature and error handlers need this name). vector-typer returns or sets this type function; when set via vector-typer, there is no automatic check that the vector's current contents match that type function.

> (define v (make-vector 3 'x symbol?)) ; initial element: 'x, elements must be symbols
#(x x x)
> (vector-set! v 0 123)
error: vector-set! argument 3, 123, is an integer but should be a symbol?
> (define (10|12? val) (memv val '(10 12)))
10|12?
> (define v1 (make-vector 3 10 10|12?)) ; only allow values 10 or 12 (initially 10)
#(10 10 10)
> (set! (v1 0) 12)
12
> v1
#(12 10 10)
> (set! (v1 1) 32)
error: vector-set! argument 3, 32, is an integer but should be a 10|12?

To access a vector's elements with different dimensions than the original had, use (subvector original-vector 0 (length original-vector) new-dimensions):

> (let ((v1 #2d((1 2 3) (4 5 6))))
    (let ((v2 (subvector v1))) ; flatten the original (1D is the default)
      v2))
#(1 2 3 4 5 6)
> (let ((v1 #(1 2 3 4 5 6)))
    (let ((v2 (subvector v1 0 6 '(3 2))))
      v2))
#2d((1 2) (3 4) (5 6))

A subvector is a window onto some other vector's data. The data is not copied, just accessed differently. The new-dimensions parameter is a list giving the lengths of the dimensions. The start and end parameters refer to positions in the original vector. subvector-vector returns the underlying vector, and subvector-position returns the starting point of the subvector in the underlying data. subvector makes it easy to access rows or columns of a vector viewed as a matrix:

> (define V (vector 0 1 2 3 4 5 6 7 8 9 10 11))
#(0 1 2 3 4 5 6 7 8 9 10 11)
> (do ((i 0 (+ i 4))) ((= i 12)) (display (subvector V i (+ i 4))) (newline))
#(0 1 2 3)
#(4 5 6 7)
#(8 9 10 11)
> (do ((sV (subvector V 0 12 '(3 4))) (i 0 (+ i 1))) ((= i 4))
    (display (vector (sV 0 i) (sV 1 i) (sV 2 i))) (newline))
#(0 4 8)
#(1 5 9)
#(2 6 10)
#(3 7 11)

matrix multiplication:

(define (matrix-multiply A B)
  ;; assume square matrices and so on for simplicity
  (let ((size (car (vector-dimensions A))))
    (do ((C (make-vector (list size size) 0))
         (i 0 (+ i 1)))
	((= i size) C)
      (do ((j 0 (+ j 1)))
	  ((= j size))
	(do ((sum 0)
             (k 0 (+ k 1)))
	    ((= k size)
             (set! (C i j) sum))
	  (set! sum (+ sum (* (A i k) (B k j)))))))))

Conway's game of Life:

(define* (life (width 40) (height 40))
  (let ((state0 (make-vector (list width height) 0))
	(state1 (make-vector (list width height) 0)))

    ;; initialize with some random pattern
    (do ((x 0 (+ x 1)))
	((= x width))
      (do ((y 0 (+ y 1)))
	  ((= y height))
	(set! (state0 x y) (if (< (random 100) 15) 1 0))))

    (do () ()
      ;; show current state (using terminal escape sequences, borrowed from the Rosetta C code)
      (format *stderr* "~C[H" #\escape)           ; ESC H = tab set
      (do ((y 0 (+ y 1)))
	  ((= y height))
	(do ((x 0 (+ x 1)))
	    ((= x width))
	  (format *stderr*
                  (if (zero? (state0 x y))
	              "  "                        ; ESC 07m below = inverse
	              (values "~C[07m  ~C[m" #\escape #\escape))))
	(format *stderr* "~C[E" #\escape))        ; ESC E = next line

      ;; get the next state
      (do ((x 1 (+ x 1)))
	  ((= x (- width 1)))
	(do ((y 1 (+ y 1)))
	    ((= y (- height 1)))
	  (let ((n (+ (state0 (- x 1) (- y 1))
		      (state0    x    (- y 1))
		      (state0 (+ x 1) (- y 1))
		      (state0 (- x 1)    y)
		      (state0 (+ x 1)    y)
		      (state0 (- x 1) (+ y 1))
		      (state0    x    (+ y 1))
		      (state0 (+ x 1) (+ y 1)))))
	    (set! (state1 x y)
		  (if (or (= n 3)
			  (and (= n 2)
			       (not (zero? (state0 x y)))))
		      1 0)))))
      (copy state1 state0))))

Multidimensional vector constant syntax is modelled after CL: #nd(...) signals that the lists specify the elements of an 'n' dimensional vector: #2d((1 2 3) (4 5 6)) int-vector constants use #i, float-vectors use #r, complex-vectors use #c. Append the "nd" business after the type indication: #i2d((1 2) (3 4)). This syntax collides with the r7rs byte-vector notation "#u8"; s7 uses "#u" for byte-vectors. "#u2d(...)" is a two-dimensional byte-vector. For backwards compatibility, you can use "#u8" for one-dimensional byte-vectors.

> (vector-ref #2d((1 2 3) (4 5 6)) 1 2)
6
> (matrix-multiply #2d((-1 0) (0 -1)) #2d((2 0) (-2 2)))
#2d((-2 0) (2 -2))
> (int-vector 1 2 3)
#i(1 2 3)
> (make-float-vector '(2 3) 1.0)
#r2d((1.0 1.0 1.0) (1.0 1.0 1.0))
> (vector (vector 1 2) (int-vector 1 2) (float-vector 1 2))
#(#(1 2) #i(1 2) #r(1.0 2.0))

If any dimension has 0 length, you get an n-dimensional empty vector. It is not equal to a 1-dimensional empty vector.

> (make-vector '(10 0 3))
#3d()
> (equal? #() #3d())
#f

To save on costly parentheses, and make it easier to write generic multidimensional sequence functions, you can use this same syntax with lists.

> (let ((L '((1 2 3) (4 5 6))))
    (L 1 0))              ; same as (list-ref (list-ref L 1) 0) or ((L 1) 0)
4
> (let ((L '(((1 2 3) (4 5 6)) ((7 8 9) (10 11 12)))))
    (set! (L 1 0 2) 32)   ; same as (list-set! (list-ref (list-ref L 1) 0) 2 32) which is unreadable!
    L)
(((1 2 3) (4 5 6)) ((7 8 32) (10 11 12)))

Or with vectors of vectors, of course:

> (let ((V #(#(1 2 3) #(4 5 6))))
    (V 1 2))              ; same as (vector-ref (vector-ref V 1) 2) or ((V 1) 2)
6
> (let ((V #2d((1 2 3) (4 5 6))))
    (V 0))
#(1 2 3)

There's one difference between a vector-of-vectors and a multidimensional vector: in the latter case, you can't clobber one of the inner vectors.

> (let ((V #(#(1 2 3) #(4 5 6)))) (set! (V 1) 32) V)
#(#(1 2 3) 32)
> (let ((V #2d((1 2 3) (4 5 6)))) (set! (V 1) 32) V)
;not enough arguments for vector-set!: (#2d((1 2 3) (4 5 6)) 1 32)

Using lists to display the inner vectors may not be optimal, especially when the elements are also lists:

#2d(((0) (0) ((0))) ((0) 0 ((0))))

The "#()" notation is no better (the elements can be vectors), and I'm not a fan of "[]" parentheses. Perhaps we could use different colors? Or different size parentheses?

#2D(((0) (0) ((0))) ((0) 0 ((0))))
#2D(((0) (0) ((0))) ((0) 0 ((0))))

I'm not sure how to handle vector->list and list->vector in the multidimensional case. Currently, vector->list flattens the vector, and list->vector always returns a one dimensional vector, so the two are not inverses.

> (vector->list #2d((1 2) (3 4)))
(1 2 3 4)             ; should this be '((1 2) (3 4)) or '(#(1 2) #(3 4))?
> (list->vector '(#(1 2) #(3 4))) ; what about '((1 2) (3 4))?
#(#(1 2) #(3 4))      

This also affects format and sort!:

> (format #f "~{~A~^ ~}" #2d((1 2) (3 4)))
"1 2 3 4"
> (sort! #2d((1 4) (3 2)) >)
#2d((4 3) (2 1))

Perhaps subvector can help:

>(subvector (list->vector '(1 2 3 4)) 0 4 '(2 2))
#2d((1 2) (3 4))
> (let ((a #2d((1 2) (3 4)))
        (b #2d((5 6) (7 8))))
  (list (subvector (append a b) 0 8 '(2 4))
	(subvector (append a b) 0 8 '(4 2))
	(subvector (append (a 0) (b 0) (a 1) (b 1)) 0 8 '(2 4))
	(subvector (append (a 0) (b 0) (a 1) (b 1)) 0 8 '(4 2))))
(#2d((1 2 3 4) (5 6 7 8))
 #2d((1 2) (3 4) (5 6) (7 8))
 #2d((1 2 5 6) (3 4 7 8))
 #2d((1 2) (5 6) (3 4) (7 8)))

Another question: should we accept the multi-index syntax in a case such as (#("abc" "def") 0 2)? My first thought was that the indices should all refer to the same type of object, so s7 would complain in a mixed case like that. If we can nest any applicable objects and apply the whole thing to an arbitrary list of indices, ambiguities arise:

((lambda (x) x) "hi" 0)
((lambda (x) (lambda (y) (+ x y))) 1 2)

I think these should complain that the function got too many arguments, but from the implicit indexing point of view, they could be interpreted as:

(string-ref ((lambda (x) x) "hi") 0) ; i.e. (((lambda (x) x) "hi") 0)
(((lambda (x) (lambda (y) (+ x y))) 1) 2)

Add optional and rest arguments, and you can't tell who is supposed to take which arguments. Currently, you can mix types with implicit indices, but if you implicitly call an element of a sequence that is a function that is not known to be "safe" (unproblematic) you'll get an error. To insist that all objects are of the same type, use an explicit getter:

> (list-ref (list 1 (list 2 3)) 1 0) ; same as ((list 1 (list 2 3)) 1 0)
2
> ((list 1 (vector 2 3)) 1 0)
2
> (list-ref (list 1 (vector 2 3)) 1 0)
error: list-ref argument 1, #(2 3), is a vector but should be a proper list

hash-tables

Each hash-table keeps track of the keys it contains, optimizing the search wherever possible. Any s7 object can be the key or the key's value. If you pass a table size that is not a power of 2, make-hash-table rounds it up to the next power of 2. The table grows as needed. length returns the current size. If a key is not in the table, hash-table-ref returns #f. To remove a key, set its value to #f; to remove all keys, (fill! table #f).

> (let ((ht (make-hash-table)))
    (set! (ht "hi") 123)
    (ht "hi"))
123

hash-table (the function) parallels the functions vector, list, and string. Its arguments are the keys and values: (hash-table 'a 1 'b 2). Implicit indexing gives multilevel hashes:

> (let ((h (hash-table 'a (hash-table 'b 2 'c 3)))) (h 'a 'b))
2
> (let ((h (hash-table 'a (hash-table 'b 2 'c 3)))) (set! (h 'a 'b) 4) (h 'a 'b))
4

hash-code is like Common Lisp's sxhash. It returns an integer that can be associated with an s7 object when implementing your own hash-tables. s7test.scm has an example using vectors. In this case the eqfunc argument is ignored (hash-code assumes equal? is in use).

Since hash-tables accept the same applicable-object syntax that vectors use, we can treat a hash-table as, for example, a sparse array:

> (define make-sparse-array make-hash-table)
make-sparse-array
> (let ((arr (make-sparse-array)))
   (set! (arr 1032) "1032")
   (set! (arr -23) "-23")
   (list (arr 1032) (arr -23)))
("1032" "-23")

map and for-each accept hash-table arguments. On each iteration, the map or for-each function is passed an entry, '(key . value), in whatever order the entries are encountered in the table.

(define (hash-table->alist table)
  (map values table))

reverse of a hash-table returns a new table with the keys and values reversed. fill! sets all the values. Two hash-tables are equal if they have the same keys with the same values. This is independent of the table sizes, or the order in which the key/value pairs were added.

The second argument to make-hash-table (eq-func) is slightly complicated. If it is omitted (or #f), s7 chooses the hashing equality and mapping functions based on the keys in the hash-table. There are times when you know in advance what equality function you want. If it's one of the built-in s7 equality functions, eq?, eqv?, equal?, equivalent?, =, string=?, string-ci=?, char=?, or char-ci=?, you can pass that function as the second argument. In any other case, you need to give s7 both the equality function and the mapping function. The latter takes any object and returns the hash-table location for it (an integer). The problem here is that for the arbitrary equality function to work, objects that are equal according to that function have to be mapped to the same hash-table location. There's no way for s7 to intuit what this mapping should be except in the built-in cases. So to specify some arbitrary function, the second argument is a cons: '(equality-checker mapper).

Here's a brief example. In CLM, we have c-objects of type mus-generator (from s7's point of view), and we want to hash them using equal? (which will call the generator-specific equality function). But s7 doesn't realize that the mus-generator type covers 40 or 50 internal types, so as the mapper we pass mus-type: (make-hash-table 64 (cons equal? mus-type)).

If the hash key is a float (a non-rational number), hash-table-ref uses equivalent?. Otherwise, for example, you could use NaN as a key, but then never be able to access it!

To implement read-time hash-tables using #h(...):

(set! *#readers*
      (cons (cons #\h (lambda (str)
			(and (string=? str "h") ; #h(...)
			     (apply hash-table (read)))))
	    *#readers*))

(display #h(:a 1)) (newline)
(display #h(:a 1 :b "str")) (newline)

These can be made immutable by (immutable! (apply...)), or even better,

(let ((h (apply hash-table (read))))
  (if (> (*s7* 'safety) 1) (immutable! h) h))
(define-macro (define-memoized name&arg . body)
  (let ((arg (cadr name&arg))
	(memo (gensym "memo")))
    `(define ,(car name&arg)
       (let ((,memo (make-hash-table)))
	 (lambda (,arg)
	   (or (,memo ,arg)                             ; check for saved value
	       (set! (,memo ,arg) (begin ,@body)))))))) ; set! returns the new value

> (define (fib n)
  (if (< n 2) n (+ (fib (- n 1)) (fib (- n 2)))))
fib
> (define-memoized
   (memo-fib n)
     (if (< n 2) n (+ (memo-fib (- n 1)) (memo-fib (- n 2)))))
memo-fib
> (time (fib 34))         ; un-memoized time
1.168                        ;   0.70 on ccrma's i7-3930 machines
> (time (memo-fib 34))    ; memoized time
3.200e-05
> (outlet (funclet memo-fib))
(inlet '{memo}-18 (hash-table
    '(0 . 0) '(1 . 1) '(2 . 1) '(3 . 2) '(4 . 3) '(5 . 5)
    '(6 . 8) '(7 . 13) '(8 . 21) '(9 . 34) '(10 . 55) '(11 . 89)
    '(12 . 144) '(13 . 233) '(14 . 377) '(15 . 610) '(16 . 987)
    '(17 . 1597) '(18 . 2584) '(19 . 4181) '(20 . 6765) '(21 . 10946)
    '(22 . 17711) '(23 . 28657) '(24 . 46368) '(25 . 75025) '(26 . 121393)
    '(27 . 196418) '(28 . 317811) '(29 . 514229) '(30 . 832040) '(31 . 1346269)
    '(32 . 2178309) '(33 . 3524578) '(34 . 5702887)))

but the tail recursive version of fib is simpler and almost as fast as the memoized version, and the iterative version beats both.

The third argument, typers, sets type checkers for the keys and values in the table, much like the third argument to make-vector. It is a cons of the type functions, (cons symbol? integer?) for example. This says that all the keys must be symbols and all the values integers. hash-table-key-typer and hash-table-value-typer return or set these functions.

> (define (10|12? val) (memv val '(10 12)))
10|12?
> (define hash (make-hash-table 8 #f (cons #t 10|12?))) ; any key is ok, but all values must be 10 or 12
(hash-table)
> (set! (hash 'a) 10)
10
> hash
(hash-table 'a 10)
> (set! (hash 'b) 32)
error: hash-table-set! value argument 3, 32, is an integer but should be a 10|12?
(define H (hash-table 'v1 1 'v2 2 'v3 3))
(let ((last-key #f))
  (define (valtyp val)
    (or (not last-key)
        (eq? last-key 'v1)
	(and (eq? last-key 'v2)
             (integer? val)
	     (<= 0 val 32))))
  (define (keytyp key)
    (set! last-key key)
    #t)
  (set! (hash-table-key-typer H) keytyp)
  (set! (hash-table-value-typer H) valtyp))

;; now (H 'v1) can be set to anything
;;     (H 'v2) must be an integer between 0 and 32
;;     (H 'v3) is immutable (but setting it to #f will remove it from H)

> (hash-table-set! H 'v1 11)
11
>(hash-table-set! H 'v2 12)
12
> (hash-table-set! H 'v3 13)
error: hash-table-set! third argument 13, is an integer, but the hash-table's value type checker, valtyp, rejects it
> (hash-table-set! H 'v2 112)
error: hash-table-set! third argument 112, is an integer, but the hash-table's value type checker, valtyp, rejects it

environments

An environment holds symbols and their values. The global environment, for example, holds all the variables that are defined at the top level. Environments are first class (and applicable) objects in s7.

(rootlet)               the top-level (global) environment
(curlet)                the current (innermost) environment
(funclet proc)          the environment at the time when proc was defined
(funclet? env)          #t if env is a funclet
(owlet)                 the environment at the point of the last error
(unlet)                 a let with built-in functions with their original value

(let-ref env sym)       get value of sym in env, same as (env sym)
(let-set! env sym val)  set value of sym in env to val, same as (set! (env sym) val)

(inlet . bindings)       make a new environment with the given bindings
(sublet env . bindings)  same as inlet, but the new environment is local to env
(varlet env . bindings)  add new bindings directly to env
(cutlet env . fields)    remove bindings from env

(let? obj)               #t if obj is an environment
(with-let env . body)    evaluate body in the environment env
(outlet env)             the environment that encloses the environment env (settable)
(let->list env)          return the environment bindings as a list of (symbol . value) cons's

(openlet env)            mark env as open (see below)
(openlet? env)           #t is env is open
(coverlet env)           mark env as closed (undo an earlier openlet)

(object->let obj)        return an environment containing information about obj
(let-temporarily vars . body)
> (inlet 'a 1 'b 2)
(inlet 'a 1 'b 2)
> (let ((a 1) (b 2)) (curlet))
(inlet 'a 1 'b 2)
> (let ((x (inlet :a 1 :b 2))) (x 'a))
1
> (with-let (inlet 'a 1 'b 2) (+ a b))
3
> (let ((x (inlet :a 1 :b 2))) (set! (x 'a) 4) x)
(inlet 'a 4 'b 2)
> (let ((x (inlet))) (varlet x 'a 1) x)
(inlet 'a 1)
> (let ((a 1)) (let ((b 2)) (outlet (curlet))))
(inlet 'a 1)
> (let ((e (inlet 'a (inlet 'b 1 'c 2)))) (e 'a 'b)) ; in C terms, e->a->b
1
> (let ((e (inlet 'a (inlet 'b 1 'c 2)))) (set! (e 'a 'b) 3) (e 'a 'b))
3
> (define* (make-let (a 1) (b 2)) (sublet (rootlet) (curlet)))
make-let
> (make-let :b 32)
(inlet 'a 1 'b 32)

As the names suggest, in s7 an environment is viewed as a disembodied let. Environments are equal if they contain the same symbols with the same values leaving aside shadowing, and taking into account the environment chain up to the rootlet. That is, two environments are equal if any local variable of either has the same value in both.

let-ref and let-set! return #<undefined> if the first argument is not defined in the environment or its parents. To search just the given environment (ignoring its outlet chain), use defined? with the third argument #t before calling let-ref or let-set!:

> (defined? 'car (inlet 'a 1) #t)
#f
> (defined? 'car (inlet 'a 1))
#t

This matters in let-set!: (let-set! (inlet 'a 1) 'car #f) is the same as (set! car #f)!

with-let evaluates its body in the given environment, so (with-let e . body) is equivalent to (eval `(begin ,@body) e), but probably faster. Similarly, (let bindings . body) is equivalent to (eval `(begin ,@body) (apply inlet (flatten bindings))), ignoring the outer (enclosing) environment (the default outer environment of inlet is rootlet). Or better,

(define-macro (with-environs e . body)
  `(apply let (map (lambda (a) (list (car a) (cdr a))) ,e) '(,@body)))

Or turning it around,

(define-macro (Let vars . body)
  `(with-let (sublet (curlet)
	       ,@(map (lambda (var)
			(values (symbol->keyword (car var)) (cadr var)))
		      vars))
     ,@body))

(Let ((c 4))
  (Let ((a 2)
        (b (+ c 2)))
  (+ a b c)))

It is faster to use (biglet 'a-function) than (with-let biglet a-function).

let-temporarily is somewhat similar to fluid-let in other Schemes. Its syntax looks like let, but it first saves the current value, then sets the variable to the new value (via set!), calls the body, and finally restores the original value. It can handle anything settable:

(let-temporarily (((*s7* 'print-length) 8)) (display x))

This sets s7's print-length variable to 8 while displaying x, then puts it back to its original value.

> (define ourlet
    (let ((x 1))
      (define (a-func) x)
      (define b-func (let ((y 1))
		       (lambda ()
		         (+ x y))))
    (curlet)))
(inlet 'x 1 'a-func a-func 'b-func b-func)
> (ourlet 'x)
1
> (let-temporarily (((ourlet 'x) 2))
    ((ourlet 'a-func)))
2
> ((funclet (ourlet 'b-func)) 'y)
1
> (let-temporarily ((((funclet (ourlet 'b-func)) 'y) 3))
    ((ourlet 'b-func)))
4

Despite the name, no new environment is created by let-temporarily: (let () (let-temporarily () (define x 2)) (+ x 1)) is 3.

sublet makes a new let with the bindings passed to it, then sets the new environment's outlet to the 'env' argument. This is similar to a let form inside another let.

> (sublet (curlet) 'b 2)
(inlet 'b 2)
> (let ((a 1)) (with-let (sublet (curlet) 'b 2) (+ a b)))
3

To add the bindings directly to the environment, use varlet. Both accept environment other than the first as well as individual bindings, adding all the argument's bindings to the new environment. inlet is very similar, but normally omits the environment argument. The arguments to sublet and inlet can be passed as symbol/value pairs, as a cons, or using keywords as if in define*. inlet can also be used to copy an environment without accidentally invoking that environment's copy method.

To implement read-time lets using #let(...):

(set! *#readers*
      (cons (cons #\l (lambda (str)
			(and (string=? str "let") ; #let(...)
			     (apply inlet (read)))))
	    *#readers*))

(display #let(:a 1)) (newline)
(display #let(:a 1 :b "str")) (newline)

Here's an example of varlet: we want to define two functions that share a local variable:

(varlet (curlet)            ; import f1 and f2 into the current environment
  (let ((x 1))              ; x is our local variable
    (define (f1 a) (+ a x))
    (define (f2 b) (* b x))
    (inlet 'f1 f1 'f2 f2))) ; export f1 and f2

One way to add reader and writer functions to an individual environment slot is:

(define e (inlet
            'x (let ((local-x 3)) ; x's initial value
		 (dilambda
		   (lambda () local-x)
		   (lambda (val) (set! local-x (max 0 (min val 100))))))))
> ((e 'x))
3
> (set! ((e 'x)) 123)
100

funclet returns a function's local environment. Here's an example that keeps a circular buffer of the calls to that function:

(define func (let ((history (let ((lst (make-list 8 #f)))
			      (set-cdr! (list-tail lst 7) lst))))
	       (lambda (x y)
		 (let ((result (+ x y)))
		   (set-car! history (list result x y))
		   (set! history (cdr history))
		   result))))

> (func 1 2)
3
> (func 3 4)
7
> ((funclet func) 'history)
#1=(#f #f #f #f #f #f (3 1 2) (7 3 4) . #1#)

It is possible in Scheme to redefine built-in functions such as car. To ensure that some code sees the original built-in function definitions, wrap it in (with-let (unlet) ...):

> (let ((caar 123))
    (+ caar (with-let (unlet)
              (caar '((2) 3)))))
125

Or perhaps better, to keep the current environment intact except for the changed built-ins:

> (let ((x 1)
        (display 3))
    (with-let (sublet (curlet) (unlet)) ; (curlet) picks up 'x, (unlet) the original 'display
      (display x)))
1

with-let and unlet are constants, so you can use them in any context without worrying about whether they've been redefined. As mentioned in the macro section, #_<name> is a built-in reader macro for (with-let (unlet) <name>), so for example, #_+ is the built-in + function, no matter what. (The environment of built-in functions that unlet accesses is not accessible from scheme code, so there's no way that those values can be clobbered).

cutlet removes bindings from an environment. If the environment is part of the outlet chain of a function, you'll probably get a segfault. Don't, for example, (cutlet (outlet (funclet func) 'x)) where func refers to x in its body. Similarly, don't mess up the outlet chain of a function (via (set! (outlet...))), and still expect that function to do something reasonable. (I may remove cutlet someday).


I think these functions can implement the notions of libraries, separate namespaces, or modules. Here's one way: first the library writer just writes his library. The normal user simply loads it. The abnormal user worries about everything, so first he loads the library in a local let to make sure no bindings escape to pollute his code, and then he uses unlet to make sure that none of his bindings pollute the library code:

(let ()
  (with-let (unlet)
    (load "any-library.scm" (curlet))
    ;; by default load puts stuff in the global environment
    ...))

Now Abnormal User can do what he wants with the library entities. Say he wants to use "lognor" under the name "bitwise-not-or", and all the other functions are of no interest:

(varlet (curlet)
  'bitwise-not-or (with-let (unlet)
                    (load "any-library.scm" (curlet))
                    lognor)) ; lognor is presumably defined in "any-library.scm"

Say he wants to make sure the library is cleanly loaded, but all its top-level bindings are imported into the current environment:

(varlet (curlet)
  (with-let (unlet)
    (let ()
      (load "any-library.scm" (curlet))
      (curlet)))) ; these are the bindings introduced by loading the library

To do the same thing, but prepend "library:" to each name:

(apply varlet (curlet)
  (with-let (unlet)
    (let ()
      (load "any-library.scm" (curlet))
      (map (lambda (binding)
	     (cons (symbol "library:" (symbol->string (car binding)))
		   (cdr binding)))
	    (curlet)))))

That's all there is to it! Here is the same idea as a macro:

(define-macro (let! init end . body)
  ;; syntax mimics 'do: (let! (vars&values) ((exported-names) result) body)
  ;;   (let! ((a 1)) ((hiho)) (define (hiho x) (+ a x)))
  `(let ,init
     ,@body
     (varlet (outlet (curlet))
       ,@(map (lambda (export)
		`(cons ',export ,export))
	      (car end)))
     ,@(cdr end)))

Well, almost, darn it. If the loaded library file sets (via set!) a global value such as abs, we need to put it back to its original form:

(define (safe-load file)
  (let ((e (with-let (unlet)         ; save the environment before loading
	     (let->list (curlet)))))
    (load file (curlet))
    (let ((new-e (with-let (unlet)   ; get the environment after loading
		   (let->list (curlet)))))
      (for-each                       ; see if any built-in functions were stepped on
       (lambda (sym)
	 (unless (assoc (car sym) e)
	   (format () "~S clobbered ~A~%" file (car sym))
	   (apply set! (car sym) (list (cdr sym)))))
       new-e))))

;; say libtest.scm has the line (set! abs odd?)

> (safe-load "libtest.scm")
"libtest.scm" clobbered abs
> (abs -2)
2

openlet marks its argument, either an environment, a closure, a c-object, or a c-pointer as open; coverlet as closed. I need better terminology here! An open object is one that the built-in s7 functions handle specially. If they encounter one in their argument list, they look in the object for their own name, and call that function if it exists. A bare-bones example:

> (abs (openlet (inlet 'abs (lambda (x) 47))))
47
> (define* (f1 (a 1)) (if (real? a) (abs a) ((a 'f1) a)))
f1
> (f1 :a (openlet (inlet 'f1 (lambda (e) 47))))
47

In CLOS, we'd declare a class and a method, and call make-instance, and then discover that it wouldn't work anyway. Here we have, in effect, an anonymous instance of an anonymous class. I think this is called a "prototype system"; javascript is apparently similar. A slightly more complex example:

(let* ((e1 (openlet
	   (inlet
	    'x 3
	    '* (lambda args
		 (apply * (if (number? (car args))
		     	      (values (car args) ((cadr args) 'x) (cddr args))
		              (values ((car args) 'x) (cdr args))))))))
       (e2 (copy e1)))
  (set! (e2 'x) 4)
  (* 2 e1 e2)) ; (* 2 3 4) => 24

Perhaps these names would be better: openlet -> with-methods, coverlet -> without-methods, and openlet? -> methods?.

let-ref and let-set! are problematic as methods. It is very easy to get into an infinite loop, especially with let-ref, since any reference to the let within the method body probably calls let-ref, which calls the let-ref method. We used to recommend coverlet here, but even that is not enough, so now let-ref and let-set! are immutable; they can't be used as methods. Use let-ref-fallback and let-set-fallback instead, if possible.

object->let returns an environment (more of a dictionary really) that contains details about its argument. It is intended as a debugging aid, underlying a debugger's "inspect" for example.

> (let ((iter (make-iterator "1234")))
    (iter)
    (iter)
    (object->let iter))
(inlet 'value #<iterator: string> 'type iterator? 'at-end #f 'sequence "1234" 'length 4 'position 2)

A c-object (in the sense of s7_make_c_type), can add its own info to this namespace via an object->let method in its local environment. snd-marks.c has a simple example using a class-wide environment (g_mark_methods), holding as the value of its 'object->let field the function s7_mark_to_let. The latter uses s7_varlet to add information to the namespace created by (object->let mark).

(define-macro (value->symbol expr)
  `(let ((val ,expr)
	 (e1 (curlet)))
     (call-with-exit
      (lambda (return)
	(do ((e e1 (outlet e))) ()
	  (for-each
	   (lambda (slot)
	     (if (equal? val (cdr slot))
		 (return (car slot))))
	   e)
	  (if (eq? e (rootlet))
	      (return #f)))))))

> (let ((a 1) (b "hi"))
    (value->symbol "hi"))
b

openlet alerts the rest of s7 that the environment has methods.

(begin
  (define fvector? #f)
  (define make-fvector #f)
  (let ((type (gensym))
	(->float (lambda (x)
		   (if (real? x)
		       (* x 1.0)
		       (error 'wrong-type-arg "fvector new value is not a real: ~A" x)))))
    (set! make-fvector
	  (lambda* (len (init 0.0))
	    (openlet
	     (inlet :v (make-vector len (->float init))
	            :type type
	   	    :length (lambda (f) len)
		    :object->string (lambda (f . args) "#<fvector>")
		    :let-set! (lambda (fv i val) (#_vector-set! (fv 'v) i (->float val)))
		    :let-ref-fallback (lambda (fv i) (#_vector-ref (fv 'v) i))))))
    (set! fvector? (lambda (p)
		     (and (let? p)
			  (eq? (p 'type) type))))))

> (define fv (make-fvector 32))
fv
> fv
#<fvector>
> (length fv)
32
> (set! (fv 0) 123)
123.0
> (fv 0)
123.0

Normally, every let's outlet chain goes back to the rootlet. If we want to create a let that breaks that chain, we can use let-ref-fallback:

(define lt (openlet (inlet 'a 1 'let-ref-fallback #<undefined>)))
> (lt 'abs)
#<undefined>

let-ref-fallback can be either a constant (most usefully #<undefined>) or a function of two arguments, the let and the symbol.

If an s7 function ignores the type of an argument, as in cons or vector for example, then that argument won't be treated as having any methods.

Since outlet is settable, there are two ways an environment can become circular. One is to include the current environment as the value of one of its variables. The other is: (let () (set! (outlet (curlet)) (curlet))).

If you want to hide an environment's fields from any part of s7 that does not know the field names in advance,

(openlet  ; make it appear to be empty to the rest of s7
  (inlet 'object->string  (lambda args "#<let>")
         'map             (lambda args ())
         'for-each        (lambda args #<unspecified>)
  	 'let->list       (lambda args ())
         'length          (lambda args 0)
	 'copy            (lambda args (inlet))
 	 'open #t
	 'coverlet        (lambda (e) (set! (e 'open) #f) e)
	 'openlet         (lambda (e) (set! (e 'open) #t) e)
	 'openlet?        (lambda (e) (e 'open))
         ;; your secret data here
         ))

(There are still at least two ways to tell that something is fishy).

Here's one way to add a method to a closure:

(define sf (openlet (let ((object->string (lambda (obj . arg)
				            "#<secret function!>")))
	              (lambda (x)
			(+ x 1)))))
> sf
#<secret function!>

multiple-values

In s7, multiple values are spliced directly into the caller's argument list.

> (+ (values 1 2 3) 4)
10
> (string-ref ((lambda () (values "abcd" 2))))
#\c
> ((lambda (a b) (+ a b)) ((lambda () (values 1 2))))
3
> (+ (call/cc (lambda (ret) (ret 1 2 3))) 4) ; call/cc has an implicit "values"
10
> ((lambda* ((a 1) (b 2)) (list a b)) (values :a 3))
(3 2)

(define-macro (call-with-values producer consumer)
  `(,consumer (,producer)))

(define-macro (multiple-value-bind vars expr . body)
  `((lambda ,vars ,@body) ,expr))

(define-macro (define-values vars expression)
  `(if (not (null? ',vars))
       (varlet (curlet) ((lambda ,vars (curlet)) ,expression))))

(define (curry function . args)
  (if (null? args)
      function
      (lambda more-args
        (if (null? more-args)
            (apply function args)
            (function (apply values args) (apply values more-args))))))

multiple-values are useful in several situations. For example, (if test (+ a b c) (+ a b d e)) can be written (+ a b (if test c (values d e))). There are a few special uses of multiple-values. First, you can use the values function to return any number of values, including 0, from map's function application:

> (map (lambda (x) (if (odd? x) (values x (* x 20)) (values))) (list 1 2 3))
(1 20 3 60)
> (map values (list 1 2 3) (list 4 5 6))
(1 4 2 5 3 6)

(define (remove-if func lst)
  (map (lambda (x) (if (func x) (values) x)) lst))

(define (pick-mappings func lst)
  (map (lambda (x) (or (func x) (values))) lst))

(define (shuffle . args)
  (apply map values args))

> (shuffle '(1 2 3) #(4 5 6) '(7 8 9))
(1 4 7 2 5 8 3 6 9)

(define (concatenate . args)
  (apply append (map (lambda (arg) (map values arg)) args)))

Second, a macro can return multiple values; these are evaluated and spliced, exactly like a normal macro, so you can use (values '(define a 1) '(define b 2)) to splice multiple definitions at the macro invocation point. If an expansion returns (values), nothing is spliced in. This is mostly useful in reader-cond and the #; reader, but unfortunately, it is tricky to use. The reader only knows about things globally defined when it encounters them, and a locally defined expansion is handled as a normal macro, so:

> (define-expansion (comment str) (values)) ; this must be at the top-level
comment
> (+ 1 (comment "one") 2 (comment "two"))
3

At the top-level (in the REPL), since there's nothing to splice into, you simply get your values back:

> (values 1 (list 1 2) (+ 3 4 5))
(values 1 (1 2) 12)

But this printout is just trying to be informative. There is no multiple-values object in s7. You can't (set! x (values 1 2)), for example. The values function tells s7 that its arguments should be handled in a special way, and the multiple-value indication goes away as soon as the arguments are spliced into some caller's arguments.

There are two helper functions for multiple values, apply-values and list-values, both intended primarily for quasiquote where (apply-values ...) implements what other schemes call unquote-splicing (",@..."). (apply-values lst) is like (apply values lst), and (list-values ...) is like (list ...) except in one special case. It is common in writing macros to create some piece of code to be spliced into the output, but if that code is nil, the resulting macro code should contain nothing (not nil). apply-values and list-values cooperate with quasiquote to implement this. As an example:

> (list-values 1 2 (apply-values) 3)
(1 2 3)
> (define (supply . args) (apply-values args))
supply
> (define (consume f . args) (apply f (apply list-values args)))
consume
> (consume + (supply 1 2) (supply 3 4 5) (supply))
15
> (consume + (supply))
0

It might seem simpler to return "nothing" from (values), rather than #<unspecified>, but that has drawbacks. First, (abs -1 (values)), or worse (abs (f x) (f y)) is no longer an error at the level of the program text; you lose the ability to see at a glance that a normal function has the right number of arguments. Second, a lot of code currently assumes that (values) returns #<unspecified>, and that implies that (apply values ()) does as well. But it would be nice if ((lambda* ((x 1)) x) (values)) returned 1!

Since set! does not evaluate its first argument, and there is no setter for "values", (set! (values x) ...) is not the same as (set! x ...). (string-set! (values string) ...) works because string-set! does evaluate its first argument. ((values + 1 2) (values 3 4) 5) is 15, as anyone would expect.

One problem with this way of handling multiple values involves cases where you can't tell whether an expression will return multiple values. Then you have, for example, (let ((val (expr)))...) and need to accept either a normal single value from expr, or one member of the possible set of multiple values. In lint.scm, I'm currently handling this with lambda: (let ((val ((lambda args (car args)) (expr))))...), but this feels kludgey. CL has nth-value which appears to do "the right thing" in this context; perhaps s7 needs it too.

A similar difficulty arises in (if (expr) ...) where (expr) might return multiple values. CL (or sbcl anyway) treats this as if it were wrapped in (nth-value 0 (expr)). Splicing the values in, on the other hand, could lead to disaster: there would be no way to tell from the code that the if statement was valid, or which branch would be taken! So, in those cases where a syntactic form evaluates an argument, s7 follows CL, and uses only the first of the values (this affects if, when, unless, cond, and case).

s7's signatures can indicate that a function returns multiple values: call-with-exit's signature is '(values procedure?), but call/cc's is '(#t procedure?) which seems inconsistent. Perhaps we could indicate the number and the expected types of those values via '((values integer? integer?)...); is this the function's "rarity"?

call-with-exit, with-baffle and continuation?

call-with-exit is call/cc without the ability to jump back into the original context, similar to "return" in C. This is cleaner than call/cc, and much faster.

(define-macro (block . body)
  ;; borrowed loosely from CL — predefine "return" as an escape
  `(call-with-exit (lambda (return) ,@body)))

(define-macro (while test . body)      ; while loop with predefined break and continue
  `(call-with-exit
    (lambda (break)
      (let continue ()
	(if (let () ,test)
	    (begin
	      (let () ,@body)
	      (continue))
	    (break))))))

(define-macro (switch selector . clauses) ; C-style case (branches fall through unless break called)
  `(call-with-exit
    (lambda (break)
      (case ,selector
	,@(do ((clause clauses (cdr clause))
	       (new-clauses ()))
	      ((null? clause) (reverse new-clauses))
	    (set! new-clauses (cons `(,(caar clause)
				      ,@(cdar clause)
				      ,@(map (lambda (nc)
					       (apply values (cdr nc))) ; doubly spliced!
					     (if (pair? clause) (cdr clause) ())))
				    new-clauses)))))))

(define (and-for-each func . args)
  ;; apply func to the first member of each arg, stopping if it returns #f
  (call-with-exit
   (lambda (quit)
     (apply for-each (lambda arglist
		       (if (not (apply func arglist))
			   (quit #<unspecified>)))
	    args))))

(define (find-if f . args)  ; generic position-if is very similar
  (call-with-exit
   (lambda (return)
     (apply for-each (lambda main-args
		       (if (apply f main-args)
			   (apply return main-args)))
	    args))))

> (find-if even? #(1 3 5 2))
2
> (* (find-if > #(1 3 5 2) '(2 2 2 3)))
6

The call-with-exit function's argument (the "continuation") is only valid within the call-with-exit function. In call/cc, you can save it, then call it later to jump back, but if you try that with call-with-exit (from outside the call-with-exit function's body), you'll get an error. This is similar to trying to read from a closed input port.

The other side, so to speak, of call-with-exit, is with-baffle. Sometimes we need a normal call/cc, but want to make sure it is active only within a given block of code. Normally, if a continuation gets away, there's no telling when it might wreak havoc on us. with-baffle blocks that — no continuation can jump into its body:

(let ((what's-for-breakfast ())
      (bad-dog 'fido))        ; bad-dog wonders what's for breakfast?
  (with-baffle                ; the syntax is (with-baffle . body)
   (set! what's-for-breakfast
	 (call/cc
	  (lambda (biscuit?)
	    (set! bad-dog biscuit?) ; bad-dog smells a biscuit!
	    'biscuit!))))
  (if (eq? what's-for-breakfast 'biscuit!)
      (bad-dog 'biscuit!))     ; now, outside the baffled block, bad-dog wants that biscuit!
  what's-for-breakfast)        ;   but s7 says "No!": baffled! ("continuation can't jump into with-baffle")

continuation? returns #t if its argument is a continuation, as opposed to a normal procedure. I don't know why Scheme hasn't had this function from the very beginning, but it's needed if you want to write a continuable error handler. Here is a sketch of the situation:

(catch #t
       (lambda ()
         (let ((res (call/cc
                      (lambda (ok)
                        (error 'cerror "an error" ok)))))
           (display res) (newline)))
       (lambda args
         (when (and (eq? (car args) 'cerror)
                    (continuation? (cadadr args)))
           (display "continuing...")
           ((cadadr args) 2))
         (display "oops")))

In a more general case, the error handler is separate from the catch body, and needs a way to distinguish a real continuation from a simple procedure.

(define (continuable-error . args)
  (call/cc
   (lambda (continue)
     (apply error args))))

(define (continue-from-error)
  (if (continuation? ((owlet) 'continue)) ; might be #<undefined> or a function as in the while macro
      (((owlet) 'continue))
      'bummer))

format, object->string

(object->string obj (write #t) (max-len (*s7* 'most-positive-fixnum)))
(format output-choice control-string . arguments)

object->string returns the string representation of its first argument. Its optional second argument can be #f or :display (use display), #t or :write (the default, use write), or :readable. In the latter case, object->string tries to produce a string that can be evaluated via eval-string to return an object equal to the original. The optional third argument sets the maximum desired string length; if object->string notices it has exceeded this limit, it returns the partial string.

> (object->string "hiho")
"\"hiho\""
> (format #f "~S" "hiho")
"\"hiho\""

s7's format function is very close to CL's and that in srfi-48.

> (format #f "~A ~D ~F" 'hi 123 3.14)
"hi 123 3.140000"

The format directives (tilde chars) are:

~%        insert newline
~&        insert newline if preceding char was not newline
~~        insert tilde
~\n       (tilde followed by newline): trim white space
~{        begin iteration (take arguments from a list, string, vector, or any other applicable object)
~}        end iteration
~^ ~|     jump out of iteration
~*        ignore the current argument
~C        print character (numeric argument = how many times to print it)
~P        insert 's' if current argument is not 1 or 1.0 (use ~@P for "ies" or "y")
~A        object->string as in display
~S        object->string as in write
~B        number->string in base 2
~O        number->string in base 8
~D        number->string in base 10 (~:D for ordinal)
~X        number->string in base 16
~E        float to string, (format #f "~E" 100.1) -> "1.001000e+02", (%e in C)
~F        float to string, (format #f "~F" 100.1) -> "100.100000",   (%f in C)
~G        float to string, (format #f "~G" 100.1) -> "100.1",        (%g in C)
~T        insert spaces (padding)
~N        get numeric argument from argument list (similar to ~V in CL)
~W        object->string with :readable (write readably: "serialization"; s7 is the intended reader)

The eight directives before ~W take the usual numeric arguments to specify field width and precision. These can also be ~N or ~n in which case the numeric argument is read from the list of arguments:

(format #f "~ND" 20 1234) ; => (format "~20D" 1234)
"                1234"

(format #f ...) simply returns the formatted string; (format #t ...) also sends the string to the current-output-port. (format () ...) sends the output to the current-output-port without returning the string (this mimics the other IO routines such as display and newline). Other built-in port choices are *stdout* and *stderr*.

Floats can occur in any base, so:

> #xf.c
15.75

This also affects format. In most Schemes, (format #f "~X" 1.25) is an error. In CL, it is equivalent to using ~A which is perverse. But

> (number->string 1.25 16)
"1.4"

and there's no obvious way to get the same effect from format unless we accept floats in the "~X" case. So in s7,

> (format #f "~X" 21)
"15"
> (format #f "~X" 1.25)
"1.4"
> (format #f "~X" 1.25+i)
"1.4+1.0i"
> (format #f "~X" 21/4)
"15/4"

That is, the output choice matches the argument. A case that came up in the Guile mailing lists is: (format #f "~F" 1/3). s7 currently returns "1/3", but Clisp returns "0.33333334".

The curly bracket directive applies to anything you can map over, not just lists:

> (format #f "~{~C~^ ~}" "hiho")
"h i h o"
> (format #f "~{~{~C~^ ~}~^...~}" (list "hiho" "test"))
"h i h o...t e s t"
> (with-input-from-string (format #f "(~{~C~^ ~})" (format #f "~B" 1367)) read) ; integer->list
(1 0 1 0 1 0 1 0 1 1 1)

Since any sequence can be passed to ~{~}, we need a way to truncate output and represent the rest of the sequence with "...", but ~^ only stops at the end of the sequence. ~| is like ~^ but it also stops after it handles (*s7* 'print-length) elements and prints "...". So, (format #f "~{~A~| ~}" #(0 1 2 3 4)) returns "0 1 2 ..." if (*s7* 'print-length) is 3.

I added object->string to s7 before deciding to include format. format excites a vague disquiet — why do we need this ancient unlispy thing? We can almost replace it with:

(define (objects->string . objects)
  (apply string-append (map (lambda (obj) (object->string obj #f)) objects)))

But how to handle lists (~{...~} in format), or columnized output (~T)? I wonder whether formatted string output still matters outside a REPL. Even in that context, a modern GUI leaves formatting decisions to a text or table widget.

(define-macro (string->objects str . objs)
  `(with-input-from-string ,str
     (lambda ()
       ,@(map (lambda (obj)
		`(set! ,obj (eval (read))))
	      objs))))

format is a mess. It is trying to cram two different choices into its first ("port") argument. Perhaps it should be split into format->string and format->port. format->string has no port argument and returns a string. format->port writes to its port argument (which must be an output port, not a boolean), and returns an empty string. Then:

(format #f ...) -> (format->string ...)
(format () ...) -> (format->port (current-output-port) ...)
(format #t ...) -> (display (format->string ...))
(format port ...) -> (display (format->string ...) port)

and the currently unavailable choice, format to port without creating a string: (format->port port ...).

hooks

(make-hook . fields)           ; make a new hook
(hook-functions hook)          ; the hook's list of 'body' functions

A hook is a function created by make-hook, and called (normally from C) when something interesting happens. In GUI toolkits hooks are called callback-lists, in CL conditions, in other contexts watchpoints or signals. s7 itself has several hooks: *error-hook*, *read-error-hook*, *unbound-variable-hook*, *missing-close-paren-hook*, *rootlet-redefinition-hook*, *load-hook*, and *autoload-hook*. make-hook is:

(define (make-hook . args)
  (let ((body ()))
    (apply lambda* args
      '(let ((result #<unspecified>))
         (let ((e (curlet)))
	   (for-each (lambda (f) (f e)) body)
           result))
       ())))

So the result of calling make-hook is a function (the lambda* that is applied to args above) that contains a list of functions, 'body. Each function in that list takes one argument, the hook. Each time the hook itself is called, each of the body functions is called, and the value of 'result is returned. That variable, and each of the hook's arguments are accessible to the hook's internal functions by going through the environment passed to the internal functions. This is a bit circuitous; here's a sketch:

> (define h (make-hook '(a 32) 'b))     ; h is a function: (lambda* ((a 32) b) ...)
h
> (set! (hook-functions h)              ; this sets ((funclet h) 'body)
           (list (lambda (hook) 	; each hook internal function takes one argument, the environment
                   (set! (hook 'result) ; this is the "result" variable above
                         (format #f "a: ~S, b: ~S" (hook 'a) (hook 'b))))))
(#<lambda (hook)>)
> (h 1 2)                               ; this calls the hook's internal functions (just one in this case)
"a: 1, b: 2"                            ; we set "result" to this string, so it is returned as the hook application result
> (h)
"a: 32, b: #f"

In C, to make a hook:

hook = s7_eval_c_string("(make-hook '(a 32) 'b)");
s7_gc_protect(s7, hook);

And call it:

result = s7_call(s7, hook, s7_list(s7, 2, s7_make_integer(s7, 1), s7_make_integer(s7, 2)));
(define-macro (hook . body)  ; return a new hook with "body" as its body, setting "result"
  `(let ((h (make-hook)))
     (set! (hook-functions h) (list (lambda (h) (set! (h 'result) (begin ,@body)))))
     h))

variable info

(documentation obj)
(signature obj)
(setter obj)
(arity obj)
(aritable? obj num-args)
(funclet proc)
(procedure-source proc)
(procedure-arglist proc)

funclet returns a procedure's environment.

> (funclet (let ((b 32)) (lambda (a) (+ a b))))
(inlet 'b 32)
> (funclet abs)
(rootlet)

setter returns or sets the setter function associated with a procedure (set-car! with car, for example).

procedure-source returns the procedure source (a list):

(define (procedure-arglist f) (cadr (procedure-source f)))

procedure-arglist returns the procedure's argument list. "procedure" here refers to functions and macros defined in s7, not built-in procedures.

documentation returns the documentation string associated with a procedure. This is normally provided via the '+documentation+ variable in the function's environment. If you'd rather, you can treat the initial string in the function's body as documentation.

(define func
  (let ((+documentation+ "helpful info"))
     (lambda (a) a)))

> (documentation func)
"helpful info"

(define (cl-func)
  "this is documentation"
  123)

> (documentation cl-func)
"this is documentation"

Since documentation is a method, a function's documentation can be computed at run-time:

(define func
  (let ((documentation (lambda (f) (format #f "this is func's funclet: ~S" (funclet f)))))
    (lambda (x)
      (+ x 1))))

> (documentation func)
"this is func's funclet: (inlet 'x ())"

arity takes any object and returns either #f if that object is not applicable, or a cons containing the minimum and maximum number of arguments acceptable. If the maximum reported is a really big number, that means any number of arguments is ok. aritable? takes two arguments, an object and an integer, and returns #t if the object can be applied to that many arguments. (For define* and friends, a key+value pair is considered to be one argument).

> (define* (add-2 a (b 32)) (+ a b))
add-2
> (procedure-source add-2)
(lambda* (a (b 32)) (+ a b))
> (arity add-2)
(0 . 2)
> (aritable? add-2 1)
#t
> (aritable? add-2 2)
#t
> (aritable? add-2 3) ; we can call (add-2 1 :b 2), but
#f ;   as mentioned above, the key+value pair is one argument

signature is a list describing the argument types and returned value type of the function. The first entry in the list is the return type, and the rest are argument types. #t means any type is possible, and 'values means the function returns multiple values.

> (signature round)
(integer? real?)                   ; round takes a real argument, returns an integer
> (signature vector-ref)
(#t vector? . #1=(integer? . #1#)) ; trailing args are all integers (indices)

If an entry is a list, any of the listed types can occur:

> (signature char-position)
((boolean? integer?) (char? string?) string? integer?)

which says that the first argument to char-position can be a string or a character, and the return type can be either boolean or an integer. To specify the types returned if multiple values are returned, use (values type1 ..). So the function:

(define (f int) (case ((0) (values 0 1)) ((1) ((values 'a 1)) (else 0))))

could declare its signature to be

(((values integer? integer?) (values symbol? integer?) integer?) integer?) ;; or would it be better to omit the 'values and just have a list of types?

If the function is defined in scheme, its signature is the value of the '+signature+ variable in its closure:

> (define f1 (let ((+documentation+ "helpful info")
                   (+signature+ '(boolean? real?)))
                (lambda (x)
                  (positive? x))))
f1
> (documentation f1)
"helpful info"
> (signature f1)
(boolean? real?)

We can do the same thing using methods:

> (define f1 (let ((documentation (lambda (f) "helpful info"))
                   (signature (lambda (f) '(boolean? real?))))
                (openlet  ; openlet alerts s7 that f1 has methods
                  (lambda (x)
                    (positive? x)))))
> (documentation f1)
"helpful info"
> (signature f1)
(boolean? real?)

signature can also be used to implement CL's 'the:

(define-macro (the value-type form)
  `((let ((+signature+ (list ,value-type)))
      (lambda ()
	,form))))

(display (+ 1 (the integer? (+ 2 3))))

but the optimizer currently doesn't know how to take advantage of this pattern.

You can obviously add your own methods:

(define my-add
  (let ((tester (lambda ()
		  (if (not (= (my-add 2 3) 5))
		      (format #t "oops: (myadd 2 3) -> ~A~%"
			      (my-add 2 3))))))
    (lambda (x y)
      (- x y))))

(define (auto-test) ; scan the symbol table for procedures with testers
  (let ((st (symbol-table)))
    (for-each (lambda (f)
		(let* ((fv (and (defined? f)
			       (symbol->value f)))
		       (testf (and (procedure? fv)
				   ((funclet fv) 'tester))))
		  (when (procedure? testf)  ; found one!
		    (testf))))
	      st)))

> (auto-test)
oops: (myadd 2 3) -> -1

Even the setter can be set this way:

(define flocals
  (let ((x 1))
    (let ((+setter+ (lambda (val) (set! x val))))
      (lambda ()
	x))))

> (flocals)
1
> (setter flocals)
#<lambda (val)>
> (set! (flocals) 32)
32
> (flocals)
32

(define (for-each-subset func args)
  ;; form each subset of args, apply func to the subsets that fit its arity
  (let subset ((source args)
               (dest ())
               (len 0))
    (if (null? source)
        (if (aritable? func len)   ; does this subset fit?
            (apply func dest))
        (begin
          (subset (cdr source) (cons (car source) dest) (+ len 1))
          (subset (cdr source) dest len)))))

eval

eval evaluates its argument, a list representing a piece of code. It takes an optional second argument, the environment in which the evaluation should take place. eval-string is similar, but its argument is a string.

> (eval '(+ 1 2))
3
> (eval-string "(+ 1 2)")
3

Leaving aside a few special cases, eval-string could be defined:

(define-macro* (eval-string x e)
  `(eval (with-input-from-string ,x read) (or ,e (curlet))))

IO and other OS functions

Besides files, ports can also represent strings and functions. The string port functions are:

(with-output-to-string thunk)         ; open a string port as current-output-port, call thunk, return string
(with-input-from-string string thunk) ; open string as current-input-port, call thunk
(call-with-output-string proc)        ; open a string port, apply proc to it, return string
(call-with-input-string string proc)  ; open string as current-input-port, apply proc to it
(open-output-string)                  ; open a string output port
(get-output-string port clear)        ; return output accumulated in the string output port
(open-input-string string)            ; open a string input port reading string
(open-input-function function)        ; open a function input port
(open-output-function function)       ; open a function output port
> (let ((result #f)
        (p (open-output-string)))
    (format p "this ~A ~C test ~D" "is" #\a 3)
    (set! result (get-output-string p))
    (close-output-port p)
    result)
"this is a test 3"

In get-output-string, if the optional 'clear' argument is #t, the port is cleared (the default in r7rs I think). Other functions:

Use length to get the length in bytes of an input port's file or string. port-line-number is settable (for fancy *#readers*). port-position is the position in bytes of the reader in the port. It is settable. port-string is the string contents of a string port. It is also settable. port-file is intended for use with the *libc* library. It returns a c-pointer containing the FILE* pointer associated with the file port (except in Windows):

(call-with-input-file "s7test.scm"
  (lambda (p)
    (with-let (sublet *libc* :file (port-file p))
      (fseek file 1000 SEEK_SET))))

The variable (*s7* 'print-length) sets the upper limit on how many elements of a sequence are printed by object->string and format.

When running s7 behind a GUI, you often want input to come from and output to go to arbitrary widgets. The function ports provide a way to redirect IO in C. See redirect display for an example. The function ports call a function rather than reading or writing the data to a string or file. See nrepl.scm and s7test.scm for examples. The function-port function is accessible as ((object->let function-port) 'function). These ports are even more esoteric than their C-side cousins. An example that traps current-ouput-port output:

(let* ((str ())
       (stdout-wrapper (open-output-function
			 (lambda (c)
			   (set! str (cons c str))))))
  (let-temporarily (((current-output-port) stdout-wrapper))
    (write-char #\a)
    ...))

The end-of-file object is #<eof>. When the read function encounters the constant #<eof> it returns #<eof>. This is neither inconsistent nor unusual: read returns either a form or #<eof>. If read encounters a form that contains #<eof>, it returns a form containing #<eof>, just as with any other constant.

> (with-input-from-string "(or x #<eof>)" read)
(or x #<eof>)
> (eof-object? (with-input-from-string "'#<eof>" read))
#f

If read hits the end of the input while reading a form, it raises an error (e.g. "missing close paren"). If it encounters #<eof> all by itself at the top level (this never happens), it returns that #<eof>. But this is specific to read, not (for example) load:

;; say we have "t234.scm" with:
(display "line 1") (newline)
#<eof>
(display "line 2") (newline)
;; end of t234.scm

> (load "t234.scm")
line 1
line 2

(with-input-from-file "t234.scm"
  (lambda ()
    (do ((c (read) (read)))
	((eof-object? c))
      (eval c))))
line 1

Built-in #<eof> has lots of uses, and as far as I can see, no drawbacks. For example, it is common to call read (or one of its friends) in a loop which first checks for #<eof>, then falls into a case statement. In s7, we can dispense with the extra if (and let), and include the #<eof> in the case statement: (case (read-char) ((#<eof>) (quit-reading)) ((#\a)...)). Another example: (or (eof-object? x) (eqv x 24)...) can be instead: (memv x '(#<eof> 24 ...). (All the discussions online that I've seen of #<eof> or its equivalent confuse the thing (end of file indication) with its name (#<eof>).

The default IO ports are *stdin*, *stdout*, and *stderr*. *stderr* is useful if you want to make sure output is flushed out immediately. The default output port is *stdout* which buffers output until a newline is seen.

An environment can be treated as an IO port, providing what Guile calls a "soft port":

(define (call-with-input-vector v proc)
  (let ((i -1))
    (proc (openlet (inlet 'read (lambda (p) (v (set! i (+ i 1)))))))))

Here the IO port is an open environment that redefines the "read" function so that it returns the next element of a vector. See stuff.scm for call-with-output-vector. The "proc" argument above can also be a macro, giving you a kludgey way to get around the dumb "lambda". Here are more useful examples:

(openlet          ; a soft port for format that sends its output to *stderr* and returns the string
  (inlet 'format (lambda (port str . args)
 	           (apply format *stderr* str args))))

(define (open-output-log name)
  ;; return a soft output port that does not hold its output file open
  (define (logit name str)
    (let ((p (open-output-file name "a")))
      (display str p)
      (close-output-port p)))
  (openlet
   (inlet :name name
	  :format (lambda (p str . args) (logit (p 'name) (apply format #f str args)))
	  :write (lambda (obj p)         (logit (p 'name) (object->string obj #t)))
	  :display (lambda (obj p)       (logit (p 'name) (object->string obj #f)))
	  :write-string (lambda (str p)  (logit (p 'name) str))
	  :write-char (lambda (ch p)     (logit (p 'name) (string ch)))
	  :newline (lambda (p)           (logit (p 'name) (string #\newline)))
	  :output-port? (lambda (p) #t)
	  :close-output-port (lambda (p) #f)
	  :flush-output-port (lambda (p) #f))))

(let ((p (open-output-log "logit.data")))
  (format p "this is a test~%")
  (format p "line: ~A~%" 2))

binary-io.scm in the Snd package has functions that read and write integers and floats in both endian choices in a variety of sizes.

If the compile time switch WITH_SYSTEM_EXTRAS is 1, several additional OS-related and file-related functions are built-in. This is work in progress; currently this switch adds:

(directory? str)         ; return #t if str is the name of a directory
(file-exists? str)       ; return #t if str names an existing file
(delete-file str)        ; try to delete the file, return 0 is successful, else -1
(getenv var)             ; return the value of an environment variable: (getenv "HOME")
(directory->list dir)    ; return contents of directory as a list of strings (if HAVE_DIRENT_H)
(system command)         ; execute command, if optional second arg is #t output is returned as a string

But maybe this is not needed; see cload.scm below for a more direct approach.

error handling

(error tag . info)           ; signal an error of type tag with addition information
(catch tag body err)         ; if error of type tag signalled in body (a thunk), call err with tag and info
(throw tag . info)           ; jump to corresponding catch

s7's error handling mimics that of Guile. An error is signalled via the error function, and can be trapped and dealt with via catch.

> (catch 'wrong-number-of-args
    (lambda ()     ; code protected by the catch
      (abs 1 2))
    (lambda args   ; the error handler
      (apply format #t (cadr args))))
"abs: too many arguments: (1 2)"
> (catch 'division-by-zero
    (lambda () (/ 1.0 0.0))
    (lambda args (string->number "+inf.0")))
+inf.0

(define-macro (catch-all . body)
  `(catch #t (lambda () ,@body) (lambda args args)))

catch has 3 arguments: a tag indicating what error to catch (#t = anything), the code, a thunk, that the catch is protecting, and the function to call if a matching error occurs during the evaluation of the thunk. The tag is matched to the error type with eq?. The error handler takes a rest argument which will hold whatever the error function chooses to pass it. The error function itself takes at least 2 arguments, the error type, a symbol, and the error message. There may also be other arguments describing the error. The default action, in the absence of any catch, is to treat the message as a format control string, apply format to it and the other arguments, and send that info to the current-error-port:

(catch #t
  (lambda ()
    (error 'oops))
  (lambda args
    (format (current-error-port) "~A: ~A~%~A[~A]:~%~A~%"
      (car args)                        ; the error type
      (apply format #f (cadr args))     ; the error info
      (port-filename) (port-line-number); error file location
      (stacktrace))))                   ; and a stacktrace

Normally when reading a file, we have to check for #<eof>, but we can let s7 do that:

(define (copy-file infile outfile)
  (call-with-input-file infile
    (lambda (in)
      (call-with-output-file outfile
	(lambda (out)
	  (catch 'wrong-type-arg   ; s7 raises this error if write-char gets #<eof>
	    (lambda ()
	      (do () ()            ; read/write until #<eof>
		(write-char (read-char in) out)))
	    (lambda err
	      outfile)))))))

catch is not limited to error handling:

(define (map-with-exit func . args)
  ;; map, but if early exit taken, return the accumulated partial result
  ;;   func takes escape thunk, then args
  (let* ((result ())
	 (escape-tag (gensym))
	 (escape (lambda () (throw escape-tag))))
    (catch escape-tag
      (lambda ()
	(let ((len (apply max (map length args))))
	  (do ((ctr 0 (+ ctr 1)))
	      ((= ctr len) (reverse result))      ; return the full result if no throw
	    (let ((val (apply func escape (map (lambda (x) (x ctr)) args))))
	      (set! result (cons val result))))))
      (lambda args
	(reverse result))))) ; if we catch escape-tag, return the partial result

(define (truncate-if func lst)
  (map-with-exit (lambda (escape x) (if (func x) (escape) x)) lst))

> (truncate-if even? #(1 3 5 -1 4 6 7 8))
(1 3 5 -1)

But this is less useful than map (it can't map over a hash-table for example), and is mostly reimplementing built-in code. Perhaps s7 should have an extension of map (and more usefully, for-each) that is patterned after dynamic-wind: (dynamic-for-each init-func main-func end-func . args) where init-func is called with one argument, the length of the shortest sequence argument (for-each and map know this in advance); main-func takes n arguments where n matches the number of sequences passed; and end-func is called even if a jump out of main-func occurs (like dynamic-wind in this regard). In the dynamic-map case, the end-func takes one argument, the current, possibly partial, result list. dynamic-for-each then could easily (but maybe not efficiently) implement generic functions such as ->list, ->vector, and ->string (converting any sequence into a sequence of some other type). map-with-exit would be

(define (map-with-exit func . args)
  (let ((result ()))
    (call-with-exit
      (lambda (quit)
        (apply dynamic-map #f ; no init-func in this case
               (lambda main-args
                 (apply func quit main-args))
               (lambda (res)
                 (set! result res))
               args)))
    result))

With all the lambda boilerplate, nested catches are hard to read:

(catch #t
  (lambda ()
    (catch 'division-by-zero
      (lambda ()
	(catch 'wrong-type-arg
	  (lambda ()
	    (abs -1))
	  (lambda args (format () "got a bad arg~%") -1)))
      (lambda args 0)))
  (lambda args 123))

Perhaps we need a macro:

(define-macro (catch-case clauses . body)
  (let ((base (cons 'lambda (cons () body))))
    (for-each (lambda (clause)
	        (let ((tag (car clause)))
	          (set! base `(lambda ()
			        (catch ',(or (eq? tag 'else) tag)
			          ,base
			          ,@(cdr clause))))))
	      clauses)
    (caddr base)))

;;; the code above becomes:
(catch-case ((wrong-type-arg   (lambda args (format () "got a bad arg~%") -1))
	     (division-by-zero (lambda args 0))
	     (else             (lambda args 123)))
  (abs -1))

This is similar to r7rs scheme's "guard", but I don't want a pointless thunk for the body of the catch. Along the same lines:

(define (catch-if test func err)
  (catch #t
    func
    (lambda args
      (apply (if (test (car args)) err throw) args)))) ; if not caught, re-raise the error via throw

(define (catch-member lst func err)
  (catch-if (lambda (tag) (member tag lst)) func err))

(define-macro (catch* clauses . error)
  ;; try each clause until one evaluates without error, else error:
  ;;    (macroexpand (catch* ((+ 1 2) (- 3 4)) 'error))
  ;;    (catch #t (lambda () (+ 1 2)) (lambda args (catch #t (lambda () (- 3 4)) (lambda args 'error))))
  (define (builder lst)
    (if (null? lst)
	(apply values error)
	`(catch #t (lambda () ,(car lst)) (lambda args ,(builder (cdr lst))))))
  (builder clauses))

When an error is encountered, and when s7 is interrupted via begin_hook, (owlet) returns an environment that contains additional info about that error:

The error-history field depends on the compiler flag WITH_HISTORY. See ow! in stuff.scm for one way to display this data. The *s7* field 'history-size sets the size of the buffer.

To find a variable's value at the point of the error: ((owlet) var). To list all the local bindings from the error outward:

(do ((e (outlet (owlet)) (outlet e)))
    ((eq? e (rootlet)))
  (format () "~{~A ~}~%" e))

To see the current s7 stack, (stacktrace). To evaluate the error handler in the environment of the error:

(let ((x 1))
  (catch #t
    (lambda ()
      (let ((y 2))
        (error 'oops)))
    (lambda args
      (with-let (sublet (owlet) :args args)    ; add the error handler args
        (list args x y)))))    ; we have access to 'y'

To limit the maximum size of the stack, set (*s7* 'max-stack-size).

The hook *error-hook* provides a way to specialize error reporting. Its arguments are named 'type and 'data. It is called if there are no catches.

(set! (hook-functions *error-hook*)
      (list (lambda (hook)
              (apply format *stderr* (hook 'data))
              (newline *stderr*))))

*read-error-hook* provides two hooks into the reader. A major problem when reading code written for other Schemes is that each Scheme provides a plethora of idiosyncratic #-names (even special character names), and \ escapes in string constants. *read-error-hook* provides a way to handle these weird cases. If a #-name is encountered that s7 can't deal with, *read-error-hook* is called with two arguments, #t and the string representing the constant. If you set (hook 'result), that result is returned to the reader. Otherwise a 'read-error is raised and you drop into the error handler. Similarly, if some bizarre \ use occurs, *read-error-hook* is called with two arguments, #f and the offending character. If you return a character, it is passed to the reader; otherwise you get an error. lint.scm has an example.

*rootlet-redefinition-hook* is called when a top-level variable is redefined (via define and friends, not set!).

(set! (hook-functions *rootlet-redefinition-hook*)
      (list (lambda (hook)
              (format *stderr* "~A ~A~%" (hook 'name) (hook 'value)))))

will print out the variable's name and the new value.

The s7-built-in catch tags are 'wrong-type-arg, 'syntax-error, 'read-error, 'unbound-variable, 'out-of-memory, 'wrong-number-of-args, 'format-error, 'out-of-range, 'division-by-zero, 'io-error, and 'bignum-error.

autoload

If s7 encounters an unbound variable, it first looks to see if it has any autoload information about it. This info can be declared via autoload, a function of two arguments, the symbol that triggers the autoload, and either a filename or a function. If a filename, s7 loads that file; if a function, it is called with one argument, the current (calling) environment.

(autoload 'channel-distance "dsp.scm")
;; now if we subsequently call channel-distance but forget to load "dsp.scm" first,
;;   s7 loads "dsp.scm" itself, and uses its definition of channel-distance.
;;   The C-side equivalent is s7_autoload.

;; here is the cload.scm case, loading j0 from the math library if it is called:
(autoload 'j0
	  (lambda (e)
	    (unless (provided? 'cload.scm)
	      (load "cload.scm"))
	    (c-define '(double j0 (double)) "" "math.h")
	    (varlet e 'j0 j0)))

The entity (hash-table or environment probably) that holds the autoload info is named *autoload*. If after checking autoload, the symbol is still unbound, s7 calls *unbound-variable-hook*. The offending symbol is named 'variable in the hook environment. If after running *unbound-variable-hook*, the symbol is still unbound, s7 calls the error handler.

The autoloader knows about s7 environments used as libraries, so, for example, you can (autoload 'j0 "libm.scm"), then use j0 in scheme code. The first time s7 encounters j0, j0 is undefined, so s7 loads libm.scm. That load returns the C math library as the environment *libm*. s7 then automatically looks for j0 in *libm*, and defines it for you. So the result is the same as if you had defined j0 yourself in C code. You can use the r7rs library mechanism here, or with-let, or whatever you want! (In Snd, libc, libm, libdl, and libgdbm are automatically tied into s7 via autoload, so if you call, for example, frexp, libm.scm is loaded, and frexp is exported from the *libm* environment, then the evaluator soldiers on, as if frexp had always been defined in s7). You can also import all of (say) gsl into the current environment via (varlet (curlet) *libgsl*).

define-constant

define-constant defines a symbol whose value is always the same (within the current lexical scope), constant? returns #t if its argument is a constant, immutable! declares a sequence to be immutable (its elements can't be changed), and immutable? returns #t if its argument is immutable.

> (define v (immutable! (vector 1 2 3)))
#(1 2 3)
> (vector-set! v 0 23)
error: can't vector-set! #(1 2 3) (it is immutable)
> (immutable? v)
#t

> (define-constant var 32)
var
> (set! var 1)
;set!: can't alter immutable object: var
> (let ((var 1)) var)
;can't bind or set an immutable object: var, line 1

There is one complication here. (immutable! let) closes the let in the sense that you can't add locals to or delete locals from the let. You can still set! the locals. To make the locals themselves immutable:

(define (vars-immutable! L)
  (with-let L
    (for-each (lambda (f)
                (immutable! (car f)))
              (curlet)))
  L)

Now (vars-immutable! let) makes it an error to set! any of the locals, but you can add locals to the let. You can speed up evaluation by doing this because it tells the optimizer that the current entries in the let will not change. To completely petrify the let, (immutable! (vars-immutable! let)). To make a function's documentation immutable: (with-let (funclet 'f2) (immutable! '+documentation+)), and similarly for other function closure entries.

define-constant blocks any attempt to set! or shadow the constant (lexically speaking of course), so local constants behave as you'd expect:

> (let () (define-constant x 3) (let ((x 32)) x))
error: can't bind an immutable object: ((x 32))
> (let ((x 3)) (set! x (let () (define-constant x 32) x))) ; outer x is not a constant
32

But watch out for deferred bindings:

> (define (func a) (let ((cvar (+ a 1))) cvar))
func
> (define-constant cvar 23) ; cvar is now globally constant so it can't be shadowed
23
> (func 1)                  ; here we're trying to shadow cvar
error: can't bind an immutable object: ((cvar (+ a 1)))
> (let ((x 1))
     (define z (let ()
                 (define-constant x 3)
                 (lambda (y)
                   (let ((x y))  ; this x is the inner constant x
                     x))))
     (z 1))  ; so this is an error even though the outer x is not a constant
error: can't bind an immutable object: ((x y))

A function can also be a constant. In some cases, the optimizer can take advantage of this information to speed up function calls.

Constants are very similar to things such as keywords (no set, always return itself as its value), variable trace (informative function upon set or keeping a history of past values), typed variables (restricting a variable's values or doing automatic conversions upon set), and notification upon set (either in Scheme or in C; I wanted this many years ago in Snd). The notification function is especially useful if you have a Scheme variable and want to reflect any change in its value immediately in C (see notification in C). In s7, setter sets this function.

Each environment is a set of symbols and their associated values. setter places a function (or macro) between a symbol and its value in a given environment. The setter function takes two arguments, the symbol and the new value, and returns the value that is actually set. If the setter function accepts a third argument, the current (symbol-relative) environment is also passed (the weird argument order is an historical artifact).

(define e      ; save environment for use below
  (let ((x 3)  ; will always be an integer
	(y 3)  ; will always keep its initial value
	(z 3)) ; will report set!

    (set! (setter 'x) (lambda (s v) (if (integer? v) v x)))
    (set! (setter 'y) (lambda (s v) y))
    (set! (setter 'z) (lambda (s v) (format *stderr* "z ~A -> ~A~%" z v) v))

    (set! x 3.3) ; x does not change because 3.3 is not an integer
    (set! y 3.3) ; y does not change
    (set! z 3.3) ; prints "z 3 -> 3.3"
    (curlet)))

> e
(inlet 'x 3 'y 3 'z 3.3)
>(begin (set! (e 'x) 123) (set! (e 'y) #()) (set! (e 'z) #f))
;; prints "z 3.3 -> #f"
> e
(inlet 'x 123 'y 3 'z #f)
> (define-macro (reflective-let vars . body)
    `(let ,vars
       ,@(map (lambda (vr)
	        `(set! (setter ',(car vr))
		       (lambda (s v)
		         (format *stderr* "~S -> ~S~%" s v)
		         v)))
	      vars)
       ,@body))
reflective-let
> (reflective-let ((a 1)) (set! a 2))
2     ; prints "a -> 2"
>(let ((a 0))
     (set! (setter 'a)
      (let ((history (make-vector 3 0))
	    (position 0))
	(lambda (s v)
	  (set! (history position) v)
	  (set! position (+ position 1))
	  (if (= position 3) (set! position 0))
	  v)))
     (set! a 1)
     (set! a 2)
     ((funclet (setter 'a)) 'history))
#(1 2 0)

See also typed-let in stuff.scm. define-constant is more restrictive than a setter that raises an error: the latter does not block nested (possibly non-constant) bindings of the symbol. The setters are kind of ugly. Here's a macro that lets you put the let variable's setter after the initial value:

(define-macro (let/setter vars . body)
  ;; (let/setter ((name value [setter])...) ...)
  (let ((setters (map (lambda (binding)
			 (and (pair? (cddr binding))
			      (caddr binding)))
		       vars))
	(gsetters (gensym)))
    `(let ((,gsetters (list ,@setters))
	   ,@(map (lambda (binding)
		    (list (car binding) (cadr binding)))
		  vars))
       ,@(do ((s setters (cdr s))
	      (var vars (cdr var))
	      (i 0 (+ i 1))
	      (result ()))
	     ((null? s)
	      (reverse result))
	   (if (car s)
	       (set! result (cons `(set! (setter (quote ,(caar var))) (list-ref ,gsetters ,i)) result))))
       ,@body)))

(let ((a 3))
  (let/setter ((a 1)
	       (b 2 (lambda (s v)
		      (+ v a)))) ; this is the outer "a"
   (set! a (+ a 1))
   (set! b (+ a b))
   (display (list a b)) (newline)))

marvels and curiousities

*load-path* is a list of directories to search when loading a file. *load-hook* is a hook whose functions are called just before a file is loaded. The hook function argument, named 'name, is the filename. While loading, port-filename and port-line-number of the current-input-port can tell you where you are in the file. This data is also available after loading via pair-line-number and pair-filename.

(set! (hook-functions *load-hook*)
       (list (lambda (hook)
               (format () "loading ~S...~%" (hook 'name)))))

(set! (hook-functions *load-hook*)
      (cons (lambda (hook)
              (format *stderr* "~A~%"
                (system (string-append "./snd lint.scm -e '(begin (lint \"" (hook 'name) "\") (exit))'") #t)))
            (hook-functions *load-hook*)))

Here's a *load-hook* function that adds the loaded file's directory to the *load-path* variable so that subsequent loads don't need to specify the directory:

(set! (hook-functions *load-hook*)
  (list (lambda (hook)
          (let ((pos -1)
                (filename (hook 'name)))
            (do ((len (length filename))
                 (i 0 (+ i 1)))
	        ((= i len))
	      (if (char=? (filename i) #\/)
	          (set! pos i)))
            (if (positive? pos)
	        (let ((directory-name (substring filename 0 pos)))
	          (if (not (member directory-name *load-path*))
		      (set! *load-path* (cons directory-name *load-path*)))))))))

As in Common Lisp, *features* is a list describing what is currently loaded into s7. You can check it with the provided? function, or add something to it with provide. In my version of Snd, at startup *features* is:

> *features*
(snd-24.7 snd24 snd audio snd-s7 snd-motif gsl alsa xm clm6 clm
sndlib gcc linux autoload dlopen system-extras overflow-checks ieee-float
complex-numbers ratios s7-10.12 s7)
> (provided? 'gsl)
#t

The other side of provide is require. (require . things) finds each thing (via autoload), and if that thing has not already been loaded, loads the associated file. (require integrate-envelope) loads "env.scm", for example; in this case it is equivalent to simply using integrate-envelope, but if placed at the start of a file, it documents that you're using that function. In the more normal use, (require snd-ws.scm) looks for the file that has (provide 'snd-ws.scm) and if it hasn't already been loaded, loads it ("ws.scm" in this case). To add your own files to this mechanism, add the provided symbol via autoload. Since load can take an environment argument, *features* and its friends follow block structure. So, for example, (let () (require stuff.scm) ...) loads "stuff.scm" into the local environment, not globally.

*features* is an odd variable: it is spread out across the chain of environments, and can hold features in an intermediate environment that aren't in subsequent (nested) values. One simple way this can happen is to load a file in a let, but cause the load to happen at the top level. The provided entities get added to the top-level *features* value, not the current let's value, but they are actually accessible locally. So *features* is a merge of all its currently accessible values, vaguely like call-next-method in CLOS. We can mimic this behavior:

(let ((x '(a)))
  (let ((x '(b)))
    (define (transparent-memq sym var e)
      (let ((val (symbol->value var e)))
	(or (and (pair? val)
		 (memq sym val))
	    (and (not (eq? e (rootlet)))
		 (transparent-memq sym var (outlet e))))))
    (let ((ce (curlet)))
      (list (transparent-memq 'a 'x ce)
	    (transparent-memq 'b 'x ce)
	    (transparent-memq 'c 'x ce)))))

'((a) (b) #f)

Multi-line and in-line comments can be enclosed in #| and |#. (+ #| add |# 1 2).

Leaving aside this case and the booleans, #f and #t, you can specify your own handlers for tokens that start with "#". *#readers* is a list of pairs: (char . func). "char" refers to the first character after the sharp sign (#). "func" is a function of one argument, the string that follows the #-sign up to the next delimiter. "func" is called when #<char> is encountered. If it returns something other than #f, the #-expression is replaced with that value. Scheme has several predefined #-readers for cases such as #b1, #\a, and so on, but you can override these if you like. If the string passed in is not the complete #-expression, the function can use read-char or read to get the rest. Say we'd like #t<number> to interpret the number in base 12:

(set! *#readers* (cons (cons #\t (lambda (str) (string->number (substring str 1) 12))) *#readers*))

> #tb
11
> #t11.3
13.25

Or have #C(real imag) be read as a complex number:

(set! *#readers* (cons (cons #\C (lambda (str) (apply complex (read)))) *#readers*))

> #C(1 2)
1+2i

Here's a reader macro for read-time evaluation:

(set! *#readers*
  (cons (cons #\. (lambda (str)
		    (and (string=? str ".") (eval (read)))))
	*#readers*))

> '(1 2 #.(* 3 4) 5)
(1 2 12 5)

And a reader that implements #[...]# for literal hash-tables:

> (set! *#readers*
    (list (cons #\[ (lambda (str)
		      (let ((h (make-hash-table)))
		        (do ((c (read) (read)))
		            ((eq? c ']#) h) ; ]# is a symbol from the reader's point of view
		          (set! (h (car c)) (cdr c))))))))
((#\[ . #<lambda (str)>))
> #[(a . 1) (b . #[(c . 3)]#)]#
(hash-table '(b . (hash-table '(c . 3))) '(a . 1))

To return no value from a reader, use (values).

> (set! *#readers* (cons (cons #\; (lambda (str) (if (string=? str ";") (read)) (values))) *#readers*))
((#\; . #<lambda (str)>))
> (+ 1 #;(* 2 3) 4)
5

Here is CL's #+ reader:

(define (sharp-plus str)
  ;; str here is "+", we assume either a symbol or an expression involving symbols follows
  (let ((e (if (string=? str "+")
		(read)                                ; must be #+(...)
		(string->symbol (substring str 1))))  ; #+feature
	(expr (read)))  ; this is the expression following #+
    (if (symbol? e)
        (if (provided? e)
	    expr
	    (values))
	(if (not (pair? e))
	    (error 'wrong-type-arg "strange #+ chooser: ~S~%" e)
	    (begin      ; evaluate the #+(...) expression as in cond-expand
	      (define (traverse tree)
		(if (pair? tree)
		    (cons (traverse (car tree))
			  (case (cdr tree) ((())) (else => traverse)))
		    (if (memq tree '(and or not)) tree
			(and (symbol? tree) (provided? tree)))))
	      (if (eval (traverse e))
		  expr
		  (values)))))))

See also the #n= reader below.

(make-list length (initial-element #f)) returns a list of 'length' elements defaulting to 'initial-element'.

(char-position char-or-string searched-string (start 0))
(string-position substring searched-string (start 0))

char-position and string-position search a string for the occurrence of a character, any of a set of characters, or a string. They return either #f if none is found, or the position within the searched string of the first occurrence. The optional third argument sets where the search starts in the second argument.

If char-position's first argument is a string, it is treated as a set of characters, and char-position looks for the first occurrence of any member of that set. Currently, the strings involved are assumed to be C strings (don't expect embedded nulls to work right in this context).

(call-with-input-file "s7.c" ; report any lines with "static " but no following open paren
  (lambda (file)
    (let loop ((line (read-line file #t)))
      (or (eof-object? line)
	  (let ((pos (string-position "static " line)))
	    (if (and pos
		     (not (char-position #\( (substring line pos))))
 	        (if (> (length line) 80)
		    (begin (display (substring line 0 80)) (newline))
		    (display line))))
	    (loop (read-line file #t)))))))
(substring-uncopied str (start 0) end)

substring-uncopied exists because there are cases where you want substring, but don't need a copy made of the string. substring-uncopied does not GC protect the original string, but obviously depends on it; it is intended for very brief uses where there is no chance that the GC will be called. Normally the optimizer can find these cases, so for example, there's no need to use substring-uncopied in (string-length (substring str 1)).

Keywords exist mainly for define*'s benefit. The keyword functions are: keyword?, string->keyword, symbol->keyword, and keyword->symbol. A keyword is a symbol that starts or ends with a colon. The colon is considered to be a part of the symbol name. A keyword is a constant that evaluates to itself.

(symbol-table)
(symbol->value sym (env (curlet)))
(symbol->dynamic-value sym)
(symbol-initial-value sym) ; settable
(defined? sym (env (curlet)) ignore-rootlet)

defined? returns #t if the symbol is defined in the environment:

(define-macro (defvar name value)
  `(unless (defined? ',name)
     (define ,name ,value)))

If ignore-rootlet is #t, the search is confined to the given environment.

symbol->value returns the value (lexically) bound to the symbol, whereas symbol->dynamic-value returns the value dynamically bound to it.

symbol-initial-value is normally the built-in (start up) value of a function, accessed via #_abs for example. For other symbols this value can only be set once, and the value should be protected from the GC (s7 does not protect it).

symbol-table returns a vector containing the symbols currently in the symbol-table. Here we scan the symbol table looking for any function that doesn't have documentation:

(for-each
   (lambda (sym)
     (if (defined? sym)
         (let ((val (symbol->value sym)))
           (if (and (procedure? val)
                    (string=? "" (documentation val)))
               (format *stderr* "~S " sym)))))
  (symbol-table))

Or get a list of gensyms:

(map (lambda (sym) (if (gensym? sym) sym (values))) (symbol-table))

An automatic software tester (see also tauto.scm and auto-tester.scm in the tools directory):

(for-each
  (lambda (sym)
    (if (defined? sym)
	(let ((val (symbol->value sym)))
          (if (procedure? val)
	      (let ((max-args (cdr (arity val))))
	        (if (or (> max-args 4)
	   	        (memq sym '(exit abort)))
		    (format () ";skip ~S for now~%" sym)
		    (begin
		      (format () ";whack on ~S...~%" sym)
                      (let ((constants (list #f #t pi () 1 1.5 3/2 1.5+i)))
                        (let autotest ((args ()) (args-left max-args))
                          (catch #t (lambda () (apply func args)) (lambda any #f))
                          (if (> args-left 0)
 	                      (for-each
	                        (lambda (c)
	                          (autotest (cons c args) (- args-left 1)))
	                        constants)))))))))))
  (symbol-table))

help tries to find information about its argument.

> (help 'caadar)
"(caadar lst) returns (car (car (cdr (car lst)))): (caadar '((1 (2 3)))) -> 2"

gc calls the garbage collector. (gc #f) turns off the GC, and (gc #t) turns it on.

If you get an error complaining about a "free cell", this is usually a sign that the GC freed some object that it should have left alone. In straight scheme code, it's an s7 bug; please send me mail about it! In foreign code, it probably indicates that you need to protect some s7_pointer with s7_gc_protect.

(equivalent? x y)

Say we want to check that two different computations came to the same result, and that result might involve circular structures. Will equal? be our friend?

> (equal? 2 2.0)
#f
> (let ((x +nan.0)) (equal? x x))
#f
> (equal? .1 1/10)
#f    
> (= .1 1/10)
#f
> (= 0.0 0+1e-300i)
#f

No! We need an equality check that ignores epsilonic differences in real and complex numbers, and knows that NaNs are equal for all practical purposes. Leaving aside numbers, closed ports are not equal, yet nothing can be done with them. #() is not equal to #2d(). And two closures are never equal, even if their arguments, environments, and bodies are equal. Since there might be circles, it is not easy to write a replacement for equal? in Scheme. So, in s7, if one thing is basically the same as some other thing, they satisfy the function equivalent?.

> (equivalent? 2 2.0)
#t
> (equivalent? 1/0 1/0)         ; NaN
#t
> (equivalent? .1 1/10)
#t                              ; floating-point epsilon here is 1.0e-15 or thereabouts
> (equivalent? 0.0 1e-300)
#t
> (equivalent? 0.0 1e-14)
#f                              ; its not always #t!
> (equivalent? (lambda () #f) (lambda () #f))
#t

The *s7* field equivalent-float-epsilon sets the floating-point fudge factor. I can't decide how bignums should interact with equivalent?. Currently, if a bignum is involved, either here or in a hash-table, s7 uses equal?. Finally, if either argument is an environment with an 'equivalent? method, that method is invoked.

define-expansion defines a macro that expands at read-time. It has the same syntax as define-macro, and (in normal use) the same result, but it is much faster because it expands only once. Similarly, define-expansion* defines a read-time macro*. (See also define-with-macros in s7test.scm for a way to expand macros in a function body at definition time). Since the reader knows almost nothing about the code it is reading, you need to make sure the expansion is defined at the top level and that its name is unique. The reader does know about global variables, so:

(define *debugging* #t)

(define-expansion (assert assertion)
  (if *debugging*          ; or maybe better, (eq? (symbol->value '*debugging*) #t)
      `(unless ,assertion
	 (format *stderr* "~A: ~A failed~%" (*function*) ',assertion))
      (values)))

Now the assertion code is only present in the function body (or wherever) if *debugging* is #t; otherwise assert expands into nothing. Another very handy use is to embed a source file line number into a message; see for example lint-format in lint.scm. Leaving aside read-time expansion and splicing, the real difference between define-macro and define-expansion is that the expansion's result is not evaluated. I'm no historian, but I believe that means that define-expansion creates a (gasp!) f*xpr. In fact:

(define-macro (define-f*xpr name-and-args . body)
  `(define ,(car name-and-args)
     (apply define-expansion
       (append (list (append (list (gensym)) ',(cdr name-and-args))) ',body))))

> (define-f*xpr (mac a) `(+ ,a 1))
mac
> (mac (* 2 3))
(+ (* 2 3) 1)

You can do something similar with a normal macro, or make the indirection explicit:

> (define-macro (fx x) `'(+ 1 ,x)) ; quote result to avoid evaluation
fx
> (let ((a 3)) (fx a))
(+ 1 a)
> (define-expansion (ex x) `(+ 1 ,x))
ex
> (let ((x ex) (a 3)) (x a))       ; avoid read-time splicing
(+ 1 a)
> (let ((a 3)) (ex a))             ; spliced in at read-time
4

As this example shows, the reader knows nothing about the program context, so if it does not see a list whose first element is a expansion name, it does not do anything special. In the (x a) case above, the expansion happens when the code is evaluated, and the expansion result is simply returned, unevaluated.

You can also use macroexpand to cancel the evaluation of a macro's expansion:

(define-macro (rmac . args)
  (if (null? args)
      ()
      (if (null? (cdr args))
	  `(display ',(car args))
	  (list 'begin
                `(display ',(car args))
		(apply macroexpand (list (cons 'rmac (cdr args))))))))

> (macroexpand (rmac a b c))
(begin (display 'a) (begin (display 'b) (display 'c)))
> (begin (rmac a b c d) (newline))
abcd

The main built-in expansion is reader-cond. The syntax is based on cond: the car of each clause is evaluated (in the read-time context), and if it is not false, the remainder of that clause is spliced into the code as if you had typed it from the start.

> '(1 2 (reader-cond ((> 1 0) 3) (else 4)) 5 6)
(1 2 3 5 6)
> ((reader-cond ((> 1 0) list 1 2) (else cons)) 5 6)
(1 2 5 6)

Here is reader-if:

(define-expansion (reader-if test true . false)
  (let ((test-val (eval test)))
    (if test-val
	true
	(and (pair? false)
	     (car false)))))

Whenever (*s7* 'profile) is positive, profiling is turned on. As the program runs, the profiler collects data about each function it can identify. At any time, you can call show-profile to see that data. The first timing is inclusive (it includes the time spent in any nested calls), the second is exclusive (it is the time spent just in the current function). In Linux and *BSD, we use clock_gettime() which is reasonably fast, but there is some profiler overhead. In other systems, we use clock() which is amazingly slow. The optimizer sometimes recasts tail recursion and similar cases as while loops, so the number of calls listed may be less than you'd expect, but the overall time should be correct. To clear out the current data, call clear-profile.

*s7* is a let that gives access to some of s7's internal state:

version                       a string describing the current s7: e.g. "s7 10.0, 13-Jan-2022"
major-version                 an integer (10 in the example above)
minor-version                 an integer (0 in the example above)

print-length                  number of elements to print of a non-string sequence
max-string-length             maximum size arg to make-string and read-string
max-format-length             maximum size arg to ~N or the width and precision fields for floats in format
max-list-length               maximum size arg to make-list
max-port-data-size            maximum size of a port data buffer
max-vector-length             maximum size arg to make-vector and make-hash-table
max-vector-dimensions         make-vector dimensions limit
default-hash-table-length     default size for make-hash-table (8, tables resize as needed)
initial-string-port-length    128, initial size of a input string port's buffer
output-port-data-size         2048, size of an output port's buffer

history			      a circular buffer of recent eval entries stored backwards (use set! to add an entry)
history-size                  eval history buffer size if s7 built WITH_HISTORY=1
history-enabled               is history buffer receiving additions (if WITH_HISTORY=1 as above)
debug                         determines debugging level (see debug.scm), default=0
profile                       profile switch (0=default, 1=gather profiling info)
profile-info                  the current profiling data; see profile.scm
profile-prefix                name (a symbol) used to identify the current environment in profile data

default-rationalize-error     1e-12
equivalent-float-epsilon      1e-15
hash-table-float-epsilon      1e-12 (currently limited to less than 1e-3).
bignum-precision              bits for bignum floats (128)
float-format-precision        digits to print for floats (16)
default-random-state          the default arg for random
most-positive-fixnum          if not using gmp, the most positive integer ("fixnum" comes from CL)
most-negative-fixnum          as above, but negative
number-separator              #\null
symbol-quote?                 #f, so in (quote x) "quote" is a #_quote (a c-function); set to #t to get quote as a symbol
symbol-printer                #f, a function to print symbols whose names contain unusual character
make-function                 #f, a function that is called when a new function is created

safety                        0 (see below)
undefined-identifier-warnings #f
undefined-constant-warnings   #f
accept-all-keyword-arguments  #f
autoloading?                  #t
openlets                      #t, whether any let can be open globally (this overrides all openlets)
expansions?                   #t, whether expansions are handled at read-time
muffle-warnings?              #f, if #t s7_warn does not output anything

cpu-time                      run time so far (proportional to cpu cycles consumed, not wall-clock seconds)
file-names or filenames       currently loaded files (a list)
catches                       a list of the currently active catch tags
c-types                       a list of c-object type names (from s7_make_c_type, etc)

stack			      the current stack entries
stack-top                     current stack location
stack-size                    current stack size
max-stack-size                maximum stack size
stacktrace-defaults           stacktrace formatting info for error handler

rootlet-size                  the number of globals
heap-size                     total cells currently available
max-heap-size                 maximum heap size
free-heap-size                the number of currently unused cells
gc-stats                      0 (or #f), 1: show GC activity, 2: heap, 4: stack, 8: protected_objects, #t = 1
gc-freed                      number of cells freed by the last GC pass
gc-total-freed                number of cells freed so far by the GC; the total allocated is probably close to
                                (with-let *s7* (+ (- heap-size free-heap-size) gc-total-freed))
gc-info                       a list: calls total-time ticks-per-second (see profile.scm)
gc-temps-size                 number of cells just allocated that are protected from the GC (256)
gc-resize-heap-fraction       when to resize the heap (0.8); these two are aimed at GC experiments
gc-resize-heap-by-4-fraction  when to get panicky about resizing the heap
gc-protected-objects          vector of the objects permanently protected from the GC
memory-usage                  a let (environment) describing current s7 memory allocations

Use the standard environment syntax to access these fields: (*s7* 'stack-top). stuff.scm has the function *s7*->list that returns most of these fields in a list.

The compile-time defaults for some of these fields can be set:

heap-size:               INITIAL_HEAP_SIZE        (64000)
stack-size:              INITIAL_STACK_SIZE       (4096)
gc-temps-size:           GC_TEMPS_SIZE            (256)
bignum-precision:        DEFAULT_BIGNUM_PRECISION (128)
history-size:            DEFAULT_HISTORY_SIZE     (8)
print-length:            DEFAULT_PRINT_LENGTH     (12)
gc-resize-heap-fraction: GC_RESIZE_HEAP_FRACTION  (0.8)
output-port-data-size:   OUTPUT_PORT_DATA_SIZE    (2048)

See also WITH_WARNINGS, S7_ALIGNED, and GC_TRIGGER_SIZE.

(set! (*s7* 'autoloading) #f) turns off the autoloader.

The 'safety variable is an integer. Currently:

0: default.
1: no remove_from_heap (a GC optimization)
   infinite loop check in eval, sort! and some iterators
   immutable object check in reverse!, sort!, and fill!
   more info in (*s7* 'history) for s7_apply_function, s7_call and s7_eval
   less aggressive optimization in with-let and lambda
   warnings about syntax redefinition
   incoming s7_pointer checks in some FFI functions
   bignum int to s7_int conversion checks
2: vector, string, and pair constants are immutable (but checks for this are currently sparse)

The debug variable controls where debug.scm is active. If it is (if debug > 0), it inserts trace calls in functions and so on. It uses dynamic-unwind to establish a catcher for the return value. (dynamic-unwind function arg) causes function to be called after the traced function has returned, passing it arg and the returned value.

(*s7* 'stacktrace-defaults) is a list of four integers and a boolean that tell the error handler how to format stacktrace information. The four integers are: how many frames to display, how many columns are devoted to code display, how many columns are available for a line of data, and where to place comments. The boolean sets whether the entire output should be displayed as a comment. The defaults are '(3 45 80 45 #t).

This will display s7 memory usage sort of like the top program:

(format *stderr* "~C[~D;~DH" #\escape 0 0)
(format *stderr* "~C[J" #\escape)
(display (with-output-to-string (lambda() (*s7* 'memory-usage))))

(Ideally we'd only redisplay the changed fields).

The standard time macro:

(define-macro (time expr)
  `(let ((start (*s7* 'cpu-time)))
     (let ((res (list ,expr))) ; expr might return multiple values
       (list (car res)
	     (- (*s7* 'cpu-time) start)))))

Add automatic log10 recalculation to (*s7* 'bignum-precision):

(define log10 (log (bignum 10)))
(define bignum-precision (dilambda (lambda ()
				     (*s7* 'bignum-precision))
				   (lambda (val)
				     (set! (*s7* 'bignum-precision) val)
				     (set! log10 (log (bignum 10)))
				     val)))

The stack, history and gc-protected-objects fields are intended for debugging. Don't keep these hanging around and expect good things to happen!

The *s7* field 'number-separator refers to what some languages call "numeric literal separator", a character that can appear in a number as a separator to make it more readable: "123,321" as opposed to "123321". If the compile time flag WITH_NUMBER_SEPARATOR is set, and (*s7* 'number-separator) is not #\null (the default), then that character can appear anywhere in a number (as long as it is between two digits), and the reader will ignore it. The *features* list will have the entry 'number-separator if s7 was compiled with WITH_NUMBER_SEPARATOR defined. (Number separators don't work with bignums).

(*s7* 'symbol-printer) is invoked by object->string when it has to print a symbol whose name can normally only be handled by the symbol function. (*s7* 'make-function) is a function called when a Scheme function is created. It is aimed at argument type checks and so on. See s7test.scm for examples.

(c-object? obj)
(c-object-type obj)

(c-pointer? obj)
(c-pointer int type info weak1 weak2)
(c-pointer-type obj)
(c-pointer-info obj)
(c-pointer-weak1 obj) ; also weak2
(c-pointer->list obj)

c-object? returns #t is its argument is a c-object. c-object-type returns the object's type tag (otherwise #f of course). This tag is also the position of the object's type in the (*s7* 'c-types) list. (*s7* 'c-types) returns a list of the types created by s7_make_c_type.

You can wrap up raw C pointers and pass them around in s7 code. The function c-pointer returns a wrapped pointer, and c-pointer? returns #t if passed one. (define NULL (c-pointer 0)). If the type field is a symbol, it is used to check types in s7_c_pointer with_type. If the 'info field of a c-pointer is a let, that pointer can participate in the generic functions mechanism, much like a c-object:

> (let ((ptr (c-pointer 1 'abc
                (inlet 'object->string
		  (lambda (obj . args)
		    (let ((lt (object->let obj)))
		      (format #f "I am pointer ~A of type '~A!"
			      (lt 'c-pointer)        ; we need c-pointer-type etc
			      (lt 'c-type))))))))
    (openlet ptr)
    (object->string ptr))
"I am pointer 1 of type 'abc!"

c-pointer->list returns (list pointer-as-int type info). The "weak1" and "weak2" fields are intended for custom "weak" references. The weak fields values are not marked during the GC sweep, much like a key in a weak-hash-table. If either value is GC'd, that field is set to #f by the GC. The weak fields are ignored by equal? and equivalent? when comparing c-pointers, and by object->string of a c-pointer even if :readable is specified.

There are several tree-oriented functions currently built into s7:

(tree-cyclic? tree) returns #t if tree contains a cycle.
(tree-leaves tree) returns the number of leaves in tree.
(tree-memq obj tree) returns #t if obj is in tree (using eq?).
(tree-set-memq set tree) returns #t if any member of the set (using eq?) is in tree.
(tree-count obj tree) returns how many times obj is in tree.

s7 originally had Scheme-level multithreading support, but I removed it in August, 2011. It turned out to be less useful than I hoped, mainly because s7 threads shared the heap and therefore had to coordinate all cell allocations. It was faster and simpler to use multiple processes each running a separate s7 interpreter, rather than one s7 running multiple s7 threads. In CLM, there was also contention for access to the output stream. In GUI-related situations, threads were not useful mainly because the GUI toolkits are not thread safe. Last but not least, the effort to make the non-threaded s7 faster messed up parts of the threaded version. Rather than waste a lot of time fixing this, I chose to flush multithreading. s7 is thread-safe:

#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
#include "s7.h"

#define NUM_THREADS 16
static pthread_t threads[NUM_THREADS];
static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;

static void *run_thread(void *obj)
{
  s7_scheme *sc = (s7_scheme *)obj;
  const char *str = s7_object_to_c_string(sc, s7_make_integer(sc, 123));
  s7_eval_c_string(sc, "(let () \
                          (define (f) \
                            (do ((i 0 (+ i 1))) ((= i 10)) \
                              (do ((k 0 (+ k 1))) ((= k 1000000))) \
                              (format *stderr* \"~D \" i))) \
                          (f))");
  pthread_mutex_lock(&lock);
  fprintf(stderr, "%s\n", str);
  pthread_mutex_unlock(&lock);
}

int main(int argc, char **argv)
{
  for (int32_t i = 0; i < NUM_THREADS; i++)
    pthread_create(&threads[i], NULL, run_thread, (void *)s7_init());
  for (int32_t i = 0; i < NUM_THREADS; i++)
    pthread_join(threads[i], NULL);
  exit(0);
}

/* linux: gcc -o threads threads.c s7.o -Wl,-export-dynamic -pthread -lm -I. -ldl
 * mac: clang -o threads threads.c s7.o -pthread -lm -I. -ldl
 */

Here's an example using gdbm to handle a variable global to the threads:

#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
#include <gdbm.h>
#include "s7.h"

#define GDBM_DB "test.gdbm"
GDBM_FILE gdb;

#define NUM_THREADS 1024
static pthread_t threads[NUM_THREADS];

static void *run_thread(void *obj)
{
  s7_scheme *sc = (s7_scheme *)obj;
  datum key, rtn;
  key.dptr = "global_int";
  key.dsize = 10;
  rtn = gdbm_fetch(gdb, key);
  if (rtn.dptr)
    {
      /* this makes a local copy of the global variable and displays its value */
      /* s7_define_variable(sc, "global-int", s7_make_integer(sc, strtol((const char *)rtn.dptr, NULL, 10))); */
      /* s7_display(sc, s7_name_to_value(sc, "global-int"), s7_current_error_port(sc)); */

      /* this increments the global variable and displays it */
      datum val;
      char buf[128];
      int bytes;
      long int ctr = strtol((const char *)rtn.dptr, NULL, 10);
      bytes = snprintf(buf, 128, "%ld", ++ctr);
      val.dptr = buf;
      val.dsize = bytes + 1;
      gdbm_store(gdb, key, val, GDBM_REPLACE);
      fprintf(stderr, "%s ", buf);

      free(rtn.dptr);
    }
  else fprintf(stderr, "oops ");
  s7_free(sc);
}

int main(int argc, char **argv)
{
  int32_t i, k, rtn, last_i = 0;
  datum key, val;
  key.dptr = "global_int";
  key.dsize = 10;
  val.dptr = "0";
  val.dsize = 2;
  gdb = gdbm_open(GDBM_DB, 1024, GDBM_NEWDB, 0664, NULL);
  gdbm_store(gdb, key, val, GDBM_REPLACE);
  for (i = 0; i < NUM_THREADS; i++)
    {
      rtn = pthread_create(&threads[i], NULL, run_thread, (void *)s7_init());
      if (rtn)
	{
	  fprintf(stderr, "failed to create thread %d\n", i);
	  exit(0);
	}
      if ((i - last_i) > 16)
	{
	  for (k = last_i; k < i; k++)
	    pthread_join(threads[k], NULL);
	  last_i = i;
	}}
  gdbm_close(gdb);
}

/* linux: gcc -o gthreads gthreads.c s7.o -O -g -Wl,-export-dynamic -pthread -lgdbm -lm -I. -ldl
 */

See libgdbm for a Scheme side example.

Some other differences from r5rs:

In s7 if a built-in function like gcd is referred to in a function body, the optimizer is free to replace it with #_function. That is, (gcd ...) can be changed to (#_gcd ...) at s7's whim, if gcd has its original value at the time the optimizer sees the expression using it. A subsequent (set! gcd +) does not affect this optimized call. I think I could wave my hands and mumble about "aggressive lexical scope" or something, but actually the choice here is that speed trumps that ol' hobgoblin consistency. If you want to change gcd to +, do it before loading code that calls gcd. I think most Schemes handle macros this way: the macro call is replaced by its expansion using its current definition, and a later redefinition does not affect earlier uses. Guile behaves like s7:

(define (add1 x) (+ x 1))
(set! + -)
(display (add1 3))) ; 4 in both s7 and Guile 3.0.4

But if a Scheme function is involved, things get messy:

(define (fib n) (if (< n 2) n (+ (fib (- n 1)) (fib (- n 2)))))
(define oldfib fib)
(set! fib 32)
(display (oldfib 10))) ; s7 says 55, Guile says "wrong type to apply: 32"

I can't decide which way is correct: s7 looks more consistent, but:

(define (fib n) 32)
(set! fib (lambda (n) (if (< n 2) n (+ (fib (- n 1)) (fib (- n 2))))))
(define oldfib fib)
(set! fib 32)
(display (oldfib 10)) ; "attempt to apply an integer 32 to..."

So s7 is inconsistent too! (Actually this was consistent until Jan 2021 when I suddenly thought it was a mistake and "fixed" it; now I'm having second thoughts).

Here are some changes I'd make to s7 if I didn't care about compatibility with other Schemes:

(most of these are removed if you set the compiler flag WITH_PURE_S7), and perhaps:

Currently WITH_PURE_S7:

With the move to s7_setter and s7_set_setter (setter in Scheme), dilambda and dilambda? have been reduced to trivial conveniences, so perhaps they can also be removed.

string-copy has 3 extra arguments to allow strings to be copied directly into other strings. In vectors, we can use subvector, but substring returns a new string (copying its argument) unless the optimizer notices that the copy is not needed. Copy almost works, but its start and end arguments refer to the source, not the destination. substring should be like subvector, but that is not backwards compatible.

There are several less-than-ideal names. get-output-string should be current-output-string. write-char behaves like display, not write. provided? should be feature? or *features* should be *provisions*. list-ref, list-set!, and list-tail actually only apply to pairs. let-temporarily should be templet, set-temporarily, or maybe just "with". define-expansion should be define-reader-macro, but that name collides with reader macros in Common Lisp. *cload-directory* should be *cload-path*. There should not be two names for the same thing: call/cc and call-with-current-continuation: flush the latter! The CL-inspired "log*" names such as logand look very old-fashioned. Standard scheme opts for the name "bitwise*"; why not "integerwise" or "bytevectorwise"? The "wise" business is just noise; are they thinking of The Hobbit? (define & logand) (define | logior) (define ~ lognot), but ^ for logxor (as in C) is not ideal; ^ should be expt. Finally, I think the notion of a current input or output port is a mistake: the IO functions should always get an explicit port.

cond-expand is dumb and its name is dumber. Take libgsl.scm; different versions of the GSL library have different functions. We need to know when we're building the FFI what GSL version we're dealing with. It would be nuts to start pushing and checking dozens of library version symbols when all we actually want is (> version 23.19). In place of cond-expand, s7 uses reader-cond, so the read-time decision involves normal Scheme evaluation.

In the section about cond-expand, the r7rs spec says "If none of the <feature requirement>s evaluate to #t, then if there is an else clause, its <expression>s are included. Otherwise, the cond-expand has no effect." I read that to mean that (begin 23 (cond-expand (surreals 1))) should evaluate to 23, and (abs -1 (cond-expand (surreals 1))) should be 1. Currently s7 returns #<unspecified> for the first, and an error, "abs: too many arguments: (abs -1 #<unspecified>)" for the second. reader-cond behaves in a way that fits the r7rs spec: (begin 23 (reader-cond ((provided? 'surreals) 1))) returns 23, and (abs -1 (reader-cond ((provided? 'surreals) 1))) returns 1, but these examples make me unhappy. Even worse: (define (f a (reader-cond ((provided? 'surreals) b))) a) which adds the argument "b" if surreals are provided. Maybe reader-cond (and expansions in general) should not magically evaporate in such cases. (I just noticed that the corrected r7rs.pdf says that the result in cond-expand is unspecified).

Then there's the case case: a case clause without a result appears to be an error in r7rs. But the notation used to indicate that is the same as that used for begin, so if we allow (begin), we should allow case clauses to have no explicit result. In cond, the "implicit progn" (in CL terminology) includes the test expression, so a clause without a result returns the test result (if true of course). In the case case, s7 returns the selector. (case x ((0 1))) is equivalent to (case x ((0 1) => values)), just as (cond (A)) is equivalent to (cond (A => values)). One application is method lookup: ((case (obj 'abs) ((#<undefined>) abs) (else)) ...); we would otherwise have to save the lookup result or do it twice. This choice has a ripple effect on do: if no result is specified for do, s7 returns the test result. It also affects hash-tables. Currently hash-table-ref returns #f if the key is not in the table, mimicking assoc and aimed at cond with =>, but if we also use case and #<undefined>, it seems more useful and maybe intuitive to mimic let-ref instead. But if hash-table-ref returns #<undefined>, it's harder to use hash-tables as sets. Hmm. In any case, the fall-through value of case should be (and is in s7) #<unspecified>: case is a form of if, so (if #f #f), (cond (#f #f)), and (case #t ((#f) #f)) should be equal.

Better ideas are always welcome!

Here are the built-in s7 variables:

And the built-in constants:

Is it odd that the "+" in +nan.0 can't be omitted, but as used in a complex number, someone drops a "+": 1+nan.0i?

(*function*) returns the name (or name and location) of the function currently being called. (define (example) (*function*)) returns 'example. Here is an example using a bacro (to access the call-time environment) and an openlet to implement a probe; it reports any operation that the probe participates in, using *function* to get the calling function name:

(define (probe-eval val)
  (let ((all-let (inlet)))
    (for-each
     (lambda (sym)
       (unless (immutable? sym) ; apply-values etc
	 (let ((func (symbol->value sym (rootlet))))
	   (when (procedure? func)
	     (varlet all-let sym
		     (apply bacro 'args
		       `((let-temporarily (((*s7* 'openlets) #f))
			   (let ((clean-args (map (lambda (arg)
						    (if (eq? arg probe-eval)
							(probe-eval 'value)
							arg))
						  args)))
			   (format *stderr* "(~S ~{~S~^ ~}) ; ~S~%"
                                   ,sym clean-args
                                   (*function* (outlet (outlet (curlet)))))
			   (apply ,func clean-args))))))))))
     (symbol-table))
    (varlet all-let 'value val)
    (openlet all-let)))

(define (call-any x)
  (+ x 21))

(call-any (probe-eval 42)) ; prints "(+ 42 21) ; call-any", returns 63

The second argument to *function* is the let from which to start searching for a function. In the example above, we start the search from the let outside the bacro, since we hope to find the bacro's caller. As a convenience, *function* takes an optional third argument specifying what information you want about the current function. An example: (*function* (curlet) 'name). name returns the name (a symbol) of the current function. line returns the function's definition line number. file returns the function's definition file. Other possibilities are signature, documentation, arity, arglist, value, and source. funclet returns the current function's funclet.

Schemes differ in their treatment of (). s7 considers it a constant that evaluates to itself, so you don't need to quote it. (eq? () '()) is #t. This is consistent with, for example, (eq? #f '#f) which is also #t. The standard says "the empty list is a special object of its own type", so surely either choice is acceptable in that regard (but, sigh, the standard stupidly goes on to deny that () can evaluate to itself). (I'm told that "is an error" means "is not portable" in the standard's weasely abuse of English; if they mean "is not portable" why not say that?). Some of the confusion appears to be caused by the word "list". I would describe the evaluator: "if it gets a constant (and () is a constant) it returns that constant; if a symbol, it returns the value associated with that symbol; if a pair, it looks at the pair's car to decide what to do".

Similarly, in s7, vector constants do not have to be quoted. A list constant is quoted to keep it from being evaluated, but #(1 2 3) is as unproblematic as "123" or 123.

These examples bring up another odd corner of scheme: else. In (cond (else 1)) the 'else is evaluated (like any cond test), so its value might be #f; in (case 0 (else 1)) it is not evaluated (like any case key), so it's just a symbol. Since setters are local in s7, someone can (let ((else #f)) (cond (else 1))) even if we protect the rootlet 'else. Of course, in scheme this kind of trouble is pervasive, so rather than make 'else a constant I think the best path is to use unlet: (let ((else #f)) (cond (#_else 1))). This is 1 (not ()) because the initial value of 'else can't be changed.

s7 treats '<datum> as (#_quote <datum>) which is not the same as the r7rs standard (quote <datum>). The name 'quote' can be captured by the local context, whereas the function #_quote can't. Apostrophe should be something that does not require you to worry about name capture; the form 'a does not contain the name 'quote', so it is annoying that its result can be altered by redefining 'quote'. Here are a few examples of the difference (I'm using guile because it is available on this machine):

(let (' 1) quote)                  ; guile 1, s7 error (#_quote is not a symbol)
(let ((quote 32)) (length '(1 2))) ; guile error Wrong type to apply: 1, s7 2
(let ((quote cos)) '0)             ; guile 1, s7 0
(let ((quote -) (x 1)) 'x)         ; guile -1, s7 x
(let ('(lambda (x) (+ x 1))) '1)   ; guile 2, s7 error (#_quote is not a symbol)
((lambda 'x (+ 'x 1)) cos 0)       ; guile 2, s7 error (lambda parameter #_quote is a constant)
(define-macro (m x) `(length ',x)) (let ((quote 32)) (m (1 2 3))) ; guile "wrong type to apply: 1", s7 3

Perhaps these simple examples will clarify s7's way of handling the apostrophe.

(syntax? #_quote) -> #t
(syntax? 'quote) -> #f      ; the symbol quote
(equal? quote #_quote) -> #t
(equal? 'quote quote) -> #f ; quote is not self-evaluating
(equal? 'quote #_quote) -> #f
(equal? '#f (quote #f)) -> #t

If you want the standard scheme approach, (set! (*s7* 'symbol-quote?) #t).

s7 handles circular lists and vectors and dotted lists with its customary aplomb. You can pass them to memq, or print them, for example; you can even evaluate them. The print syntax is borrowed from CL:

> (let ((lst (list 1 2 3)))
    (set! (cdr (cdr (cdr lst))) lst)
    lst)
#1=(1 2 3 . #1#)
> (let* ((x (cons 1 2))
         (y (cons 3 x)))
    (list x y))
(#1=(1 . 2) (3 . #1#))

But should this syntax be readable as well? I'm inclined to say no because then it is part of the language, and it doesn't look like the rest of the language. (I think it's kind of ugly). Perhaps we could implement it via *#readers*:

(define circular-list-reader
  (let ((known-vals #f)
	(top-n -1))
    (lambda (str)

      (define (replace-syms lst)
	;; walk through the new list, replacing our special keywords
        ;;   with the associated locations

	(define (replace-sym tree getter)
	  (if (keyword? (getter tree))
	      (let ((n (string->number (symbol->string (keyword->symbol (getter tree))))))
		(if (integer? n)
		    (let ((lst (assoc n known-vals)))
		      (if lst
			  (set! (getter tree) (cdr lst))
			  (format *stderr* "#~D# is not defined~%" n)))))))

	(let walk-tree ((tree (cdr lst)))
	  (if (pair? tree)
	      (begin
		(if (pair? (car tree)) (walk-tree (car tree)) (replace-sym tree car))
		(if (pair? (cdr tree)) (walk-tree (cdr tree)) (replace-sym tree cdr))))
	  tree))

      ;; str is whatever followed the #, first char is a digit
      (let* ((len (length str))
	     (last-char (str (- len 1))))
	(and (memv last-char '(#\= #\#))             ; is it #n= or #n#?
	    (let ((n (string->number (substring str 0 (- len 1)))))
	      (and (integer? n)
		  (begin
		    (if (not known-vals)            ; save n so we know when we're done
			(begin
			  (set! known-vals ())
			  (set! top-n n)))

		    (if (char=? last-char #\=)      ; #n=
			(and (eqv? (peek-char) #\() ; eqv? since peek-char can return #<eof>
			    (let ((cur-val (assoc n known-vals)))
			      ;; associate the number and the list it points to
			      ;;    if cur-val, perhaps complain? (#n# redefined)
			      (let ((lst (catch #t
					   read
					   (lambda args             ; a read error
					     (set! known-vals #f)   ;   so clear our state
					     (apply throw args))))) ;   and pass the error on up
				(if cur-val
                                    (set! (cdr cur-val) lst)
				    (set! known-vals
					  (cons (set! cur-val (cons n lst)) known-vals))))

			      (if (= n top-n)            ; replace our special keywords
				  (let ((result (replace-syms cur-val)))
				    (set! known-vals #f) ; '#1=(#+gsl #1#) -> '(:1)!
				    result)
				  (cdr cur-val))))
			                         ; #n=<not a list>?
			;; else it's #n# — set a marker for now since we may not
			;;   have its associated value yet.  We use a symbol name that
                        ;;   string->number accepts.
			(symbol->keyword
                          (symbol (number->string n) (string #\null) " "))))))
		                                 ; #n<not an integer>?
	    )))))                                ; #n<something else>?

(do ((i 0 (+ i 1)))
    ((= i 10))
  ;; load up all the #n cases
  (set! *#readers*
    (cons (cons (integer->char (+ i (char->integer #\0))) circular-list-reader)
          *#readers*)))

> '#1=(1 2 . #1#)
#1=(1 2 . #1#)
> '#1=(1 #2=(2 . #2#) . #1#)
#2=(1 #1=(2 . #1#) . #2#)

And of course, we can treat these as labels:

(let ((ctr 0)) #1=(begin (format () "~D " ctr) (set! ctr (+ ctr 1)) (if (< ctr 4) #1# (newline))))

which prints "0 1 2 3" and a newline.


Length returns +inf.0 if passed a circular list, and returns a negative number if passed a dotted list. In the dotted case, the absolute value of the length is the list length not counting the final cdr. (define (circular? lst) (infinite? (length lst))).

cyclic-sequences returns a list of the cyclic sequences in its argument, or nil. (define (cyclic? obj) (pair? (cyclic-sequences obj))).

Here's an amusing use of circular lists:

(define (for-each-permutation func vals)
  ;; apply func to every permutation of vals:
  ;;   (for-each-permutation (lambda args (format () "~{~A~^ ~}~%" args)) '(1 2 3))
  (define (pinner cur nvals len)
    (if (= len 1)
        (apply func (car nvals) cur)
        (do ((i 0 (+ i 1)))                       ; I suppose a named let would be more Schemish
            ((= i len))
          (let ((start nvals))
            (set! nvals (cdr nvals))
            (let ((cur1 (cons (car nvals) cur)))  ; add (car nvals) to our arg list
              (set! (cdr start) (cdr nvals))      ; splice out that element and
              (pinner cur1 (cdr start) (- len 1)) ;   pass a smaller circle on down, "wheels within wheels"
              (set! (cdr start) nvals))))))       ; restore original circle
  (let ((len (length vals)))
    (set-cdr! (list-tail vals (- len 1)) vals)    ; make vals into a circle
    (pinner () vals len)
    (set-cdr! (list-tail vals (- len 1)) ())))    ; restore its original shape

s7 and Snd use "*" in a variable name, *features* for example, to indicate that the variable is predefined. It may occur unprotected in a macro, for example. The "*" doesn't mean that the variable is special in the CL sense of dynamic scope, but some clear marker is needed for a global variable so that the programmer doesn't accidentally step on it.

Although a variable name's first character is more restricted, currently only #\null, #\newline, #\tab, #\space, #\), #\(, #\", and #\; can't occur within the name. I did not originally include double-quote in this set, so wild stuff like (let ((nam""e 1)) nam""e) would work, but that means that '(1 ."hi") is parsed as a 1 and the symbol ."hi", and (string-set! x"hi") is an error. The first character should not be #\#, #\', #\`, #\,, #\:, or any of those mentioned above, and some characters can't occur by themselves. For example, "." is not a legal variable name, but ".." is. These weird symbols have to be printed sometimes:

> (list 1 (string->symbol (string #\; #\" #\\)) 2)
(1 ;"\ 2)            
> (list 1 (string->symbol (string #\.)) 2)
(1 . 2)

which is a mess. Guile prints the first as (1 #{\;\"\\}# 2). In CL and some Schemes:

[1]> (list 1 (intern (coerce (list #\; #\" #\\) 'string)) 2) ; thanks to Rob Warnock
(1 |;"\\| 2)        
[2]> (equalp 'A '|A|) ; in CL case matters here
T

This is clean, and has the weight of tradition behind it, but I think I'll use "symbol" instead:

> (list 1 (string->symbol (string #\; #\" #\\)) 2)
(1 (symbol ";\"\\") 2)       

This output is readable, and does not eat up perfectly good characters like vertical bar, but it means we can't easily use variable names like "| e t c |". We could allow a name to contain any characters if it starts and ends with "|", but then one vertical bar is trouble. We can define a reader that turns #symbol<...> into (symbol "..."), making it possible to use odd names more widely:

(set! *#readers*
      (list (cons #\s
		  (lambda (str)
		    (let ((len (length str)))
		      (and (string=? (substring str 0 7) "symbol<")
			   (if (char=? (str (- len 1)) #\>) ; pointless use of #symbol!
			       (symbol (substring str 7 (- len 1)))
			       (do ((sym (substring str 7))
				    (c (read-char) (read-char)))
				   ((memq c (list #\> #<eof>))
				    (string->symbol sym))
				 (set! sym (string-append sym (string c)))))))))))

> (let ((#symbol<a b c> 32)) (+ #symbol<a b c> 1))
33

But there's a problem: if we try to call object->string with :readable on these symbol tokens, it does not know that we want it to use our "#symbol<...>" reader-macro. We need to set the *s7* field 'symbol-printer:

> (define f (apply lambda (list () (list 'let (list (list (symbol "a b") 3)) (symbol "a b")))))
f
> (f)
3
> (object->string f :readable)
"(lambda () (let (((symbol \"a b\") 3)) (symbol \"a b\")))" ; not actually readable!
> (set! (*s7* 'symbol-printer) (lambda (obj) (string-append "#symbol<" (symbol->string obj) ">")))
#<lambda (obj)>
> (object->string f :readable)
"(lambda () (let ((#symbol<a b> 3)) #symbol<a b>))"

The symbol function accepts any number of string arguments which it concatenates to form the new symbol name.

These symbols are not just an optimization of string comparison:

> (define-macro (hi a)
  (let ((funny-name (string->symbol ";")))
    `(let ((,funny-name ,a)) (+ 1 ,funny-name))))
hi
> (hi 2)
3
> (macroexpand (hi 2))
(let ((; 2)) (+ 1 ;))    ; for a good time, try (string #\")

> (define-macro (hi a)
  (let ((funny-name (string->symbol "| e t c |")))
    `(let ((,funny-name ,a)) (+ 1 ,funny-name))))
hi
> (hi 2)
3
> (macroexpand (hi 2))
(let ((| e t c | 2)) (+ 1 | e t c |))
> (let ((funny-name (string->symbol "| e t c |"))) ; now use it as a keyword arg to a function
    (apply define* `((func (,funny-name 32)) (+ ,funny-name 1)))
    ;; (procedure-source func) is (lambda* ((| e t c | 32)) (+ | e t c | 1))
    (apply func (list (symbol->keyword funny-name) 2)))
3

I hope that makes you as happy as it makes me!

The built-in syntactic forms, such as "begin", are almost first-class citizens.

> (let ((progn begin))
    (progn
      (define x 1)
      (set! x 3)
      (+ x 4)))
7
> (let ((function lambda))
    ((function (a b) (list a b)) 3 4))
(3 4)
> (apply begin '((define x 3) (+ x 2)))
5
> ((lambda (n) (apply n '(((x 1)) (+ x 2)))) let)
3

(define-macro (symbol-set! var val) ; like CL's set
  `(apply set! ,var ',val ()))      ; trailing nil is just to make apply happy — apply*?

(define-macro (progv vars vals . body)
 `(apply (apply lambda ,vars ',body) ,vals))

> (let ((s '(one two)) (v '(1 2))) (progv s v (+ one two)))
3

We can snap together program fragments ("look Ma, no macros!"):

(let* ((x 3)
       (arg '(x))
       (body `((+ ,x x 1))))
  ((apply lambda arg body) 12)) ; "legolambda"?

(define (engulph form)
  (let ((body `(let ((L ()))
		 (do ((i 0 (+ i 1)))
		     ((= i 10) (reverse L))
		   (set! L (cons ,form L))))))
    (define function (apply lambda () (list (copy body))))
    (function)))

(let ()
  (define (hi a) (+ a x))
  ((apply let '((x 32)) (list (procedure-source hi))) 12)) ; one function, many closures?

(let ((ctr -1))  ; (enum zero one two) but without using a macro
  (apply begin
    (map (lambda (symbol)
           (set! ctr (+ ctr 1))
           (list 'define symbol ctr)) ; e.g. '(define zero 0)
         '(zero one two)))
  (+ zero one two))

But there's a prettier way to implement enum ("transparent-for-each"):

> (define-macro (enum . args)
    `(for-each define ',args (iota (length ',args))))
enum
> (enum a b c)
#<unspecified>
> b
1

Now we notice that (case 0.0 ((0.0) 1) (else 0)) is 1, but how to get pi into a key list?

> (apply case 'pi `(((,pi) 1) (else 0)))
1
> (let ((lst '(1 2))) (apply case 'lst `(((,lst) 1) (else 0))))
1         ; same trick puts a list in the keys
> (apply case '+nan.0 `(((,+nan.0) 1) (else 0)))
0         ; (eqv? +nan.0 +nan.0) is #f

(apply define ...) is similar to CL's set.

> ((apply define-macro '((m a) `(+ 1 ,a))) 3)
4
> ((apply define '((hi a) (+ a 1))) 3)
4

Apply let is very similar to eval:

> (apply let '((a 2) (b 3)) '((+ a b)))
5
> (eval '(+ a b) (inlet 'a 2 'b 3))
5
> ((apply lambda '(a b) '((+ a b))) 2 3)
5
> (apply let '((a 2) (b 3)) '((list + a b))) ; a -> 2, b -> 3
(+ 2 3)

The redundant-looking double lists are for apply's benefit. We could use a trailing null instead (mimicking apply* in some ancient lisps):

> (apply let '((a 2) (b 3)) '(list + a b) ())
(+ 2 3)

Scheme claims that it evaluates the car of an expression, then calls the result with the rest of the expression. So ((if x + -) y z) calls either (+ y z) or (- y z) depending on x. But only s7, as far as I know, handles ((if x or and) y z).

catch, dynamic-wind, and many of the other functions that take function arguments in standard Scheme, accept macros in s7, and dynamic-wind accepts #f as the initial and final entries.

Currently, you can't set! a built-in syntactic keyword to some new value: (set! if 3). let-temporarily uses set!, so (let-temporarily ((if 3))...) is also unlikely to work.

Speaking of speed... It is widely believed that a Scheme with first class everything can't hope to compete with any "real" Scheme. Humph I say. Take this little example (which is not so misleading that I feel guilty about it):

(define (do-loop n)
  (do ((i 0 (+ i 1)))
      ((= i n))
    (if (zero? (modulo i 1000))
	(display ".")))
  (newline))

(for-each do-loop (list 1000 1000000 10000000))

In s7, that takes 0.09 seconds on my home machine. In tinyScheme, from whence we sprang, it takes 85 seconds. In the chicken interpreter, 5.3 seconds, and after compilation (using -O2) of the chicken compiler output, 0.75 seconds. So, s7 is comparable to chicken in speed, even though chicken is compiling to C. I think Guile 2.0.9 takes about 1 second. The equivalent in CL: clisp interpreted 9.3 seconds, compiled 0.85 seconds; sbcl 0.21 seconds. Similarly, s7 computes (fib 40) in 0.8 seconds, approximately the same as sbcl. Guile 2.2.3 takes 7 seconds.

s7's timing tests are in its tools directory. The script valcall.scm runs them through callgrind. The results can be found at the end of s7.c. If you're interested in the standard Scheme benchmarks, it is possible to add s7 to that package. First, s7-prelude.scm and s7-postlude.scm need to be added to the benchmarks src directory. s7-postlude.scm can be empty. My version of s7-prelude.scm is:

(define (this-scheme-implementation-name) "s7")
(define exact-integer? integer?)
(define (exact-integer-sqrt i) (let ((sq (floor (sqrt i)))) (values sq (- i (* sq sq)))))
(define inexact exact->inexact)
(define exact inexact->exact)
(define (square x) (* x x))
(define (vector-map f v) (copy v)) ; for quicksort.scm
(define-macro (import . args) #f)
(define (jiffies-per-second) 1000)
(define (current-jiffy) (round (* (jiffies-per-second) (*s7* 'cpu-time))))
(define (current-second) (floor (*s7* 'cpu-time)))
(define make-bytevector make-byte-vector)
(define bytevector-u8-set! byte-vector-set!)
(set! (*s7* 'symbol-quote?) #t)

If you want to run gcbench, add the define-record-type macro from r7rs.scm. Here are the diffs for the bench script:

141a142
>     S7=${S7:-"/home/bil/motif-snd/repl"}
187a189
>   s7               for s7
406a409,421
> # Definitions specific to s7
>
> s7_comp ()
> {
>     :
> }
>
> s7_exec ()
> {
>     time ${S7} "$1" < "$2"
> }
>
> # -----------------------------------------------------------------------------
940a957,966
>
>         s7)    NAME='s7'
>                COMP=s7_comp
>                EXEC=s7_exec
>                COMPOPTS=""
>                EXTENSION="scm"
>                EXTENSIONCOMP="scm"
>                COMPCOMMANDS=""
>                EXECCOMMANDS=""
>                ;;

I call the standalone version of s7 "repl", so its path is /home/bil/motif-snd/repl. To build repl, get s7.tar.gz from https://ccrma.stanford.edu/software/s7/s7.tar.gz; if not using gcc or clang, add the empty file mus-config.h to the tarball's contents, then (in Linux):

gcc s7.c -o repl -DWITH_MAIN -I. -O2 -g -ldl -lm -Wl,-export-dynamic
;; tcc -o s7 s7.c -I. -lm -DWITH_MAIN -ldl -rdynamic -DWITH_C_LOADER

For timing tests, I add "-fomit-frame-pointer -funroll-loops -march=native". mus-config.h normally has

#define HAVE_COMPLEX_NUMBERS 1
#define HAVE_COMPLEX_TRIG 1

but s7.c has defaults, so mus-config.h can be empty, or absent. Finally, go back to the benchmarks directory and

bench s7 all

The benchmark compiler.scm assumes that small integers can be compared with eq? (via assq), which is incorrect. pi.scm and chudnovsky.scm need the gmp version of s7. As of 24-Oct-23, s7 treats '<> as (#_quote <>) so dynamic.scm doesn't run, and peval.scm gets the wrong result, but (set! (*7s* 'symbol-quote?) #t) fixes this problem. I ran the bench script on an AMD 3950X machine, and got these results (in seconds): ack: 6.6, array1: 6.4, browse: 11.2, bv2string: 4.1, cat: 0.4, compiler: 16.9, conform: 30.0, cpstak: 42.8, ctak: 16.6, deriv: 9.7, destruc: 8.6, diviter: 3.7, divrec: 4.6, dynamic: 12.6, earley: 25.5, equal: 0.3, fft: 12.5, fib: 6.1, fibc: 8.6, fibfp: 1.1, gcbench: 12.9, graphs: 72.5, lattice: 63.4, matrix: 21.0, maze: 11.4, mazefun: 9.8, mbrot: 12.6, mbrotZ: 8.0, mperm: 18.9, nboyer: 20.1, nqueens: 27.0, ntakl: 8.0, nucleic: 8.3, paraffins: 4.4, parsing: 20.7, peval: 15.2, pnpoly: 9.8, primes: 10.2, puzzle: 10.2, quicksort: 40.0, ray: 8.3, read1: 0.2, sboyer: 19.1, scheme: 29.5, simplex: 26.9, slatex: 4.2, string: 0.3, sum1: 0.2, sum: 4.1, sumfp: 2.2, tail: 0.1, tak: 7.1, takl: 8.1, triangl: 16.4, wc: 4.9. In the gmp case, chudnovsky: 0.017, pi: .01.

In s7, there is only one kind of begin statement, and it can contain both definitions and expressions. These are evaluated in the order in which they occur, and in the environment at the point of the evaluation. I think of it as being a little REPL. begin does not introduce a new frame in the current environment, so defines happen in the enclosing environment. Finally, begin, explicit or otherwise, does not pretend to emulate letrec*.

If we allow defines anywhere, the notion of "lexical scope" becomes problematic. Scheme is already a mess in that regard: take

(let ((x 1))
  (do ((y x x)
       (x 3))
      ((> y 1) y)))

In (y x x) the first x is the outer one, and the second is the following do variable, so this returns 3! But sticking to define, in

(let ((x 1))
  (define y x)
  (define x 2)
  y)

s7 returns 1 even though technically the second x is in y's environment. Since we treat this as a REPL, y gets its value from the only x defined at the point it is defined. However,

(let ((x 1))
  (define y (lambda () x))
  (define x 2)
  (y))

returns 2 in s7 because the x in y's function body is not evaluated until after the second x is defined. The define propagates backwards, but: (list x (define x 0)), or (list x (begin (define x 0) x)).

The r7rs compatibility code is in r7rs.scm. I used to include it here, but as r7rs grew, this section got too large. In general, all the conversion routines in r7rs are handled in s7 via generic functions, records are classes, and so on.

r7rs.scm has the command-line function that makes it easier (on Linux anyway) to run s7 as a scripting engine. Say we have a s7's repl.c compiled and loaded to the file "repl", and a file named "runit" with these contents:

#!repl
!#
(load "r7rs.scm")
(display (command-line))
(newline)
(exit)

Now make runit executable via chmod, and run it with some arguments:

runit 123 abc

and it prints: ("repl" "runit" "123" "abc")