I'm learning common lisp. I have written a version of the once-only macro, which suffers from an unusual variable capture problem.
My macro is this:
(defmacro my-once-only (names &body body)
(let ((syms (mapcar #'(lambda (x) (gensym))
names)))
``(let (,,#(mapcar #'(lambda (sym name) ``(,',sym ,,name))
syms names))
,(let (,#(mapcar #'(lambda (name sym) `(,name ',sym))
names syms))
,#body))))
The canonical version of only-once is this:
(defmacro once-only ((&rest names) &body body)
(let ((gensyms (loop for n in names collect (gensym))))
`(let (,#(loop for g in gensyms collect `(,g (gensym))))
`(let (,,#(loop for g in gensyms for n in names collect ``(,,g ,,n)))
,(let (,#(loop for n in names for g in gensyms collect `(,n ,g)))
,#body)))))
The difference, as far as I can tell, is that the canonical version generates new symbols for every expansion of the macro using only-once. For example:
CL-USER> (macroexpand-1 '(once-only (foo) foo))
(LET ((#:G824 (GENSYM)))
`(LET (,`(,#:G824 ,FOO))
,(LET ((FOO #:G824))
FOO)))
T
CL-USER> (macroexpand-1 '(my-once-only (foo) foo))
`(LET (,`(,'#:G825 ,FOO))
,(LET ((FOO '#:G825))
FOO))
T
The variable my macro uses to store the value of foo is the same for every expansion of this form, in this case it would be #:G825. This is akin to defining a macro like the following:
(defmacro identity-except-for-bar (foo)
`(let ((bar 2))
,foo))
This macro captures bar, and this capture manifests when bar is passed to it, like so:
CL-USER> (let ((bar 1))
(identity-except-for-bar bar))
2
However, I cannot think of any way to pass #:G825 to a macro that uses my-only-once so that it breaks like this, because the symbols gensym returns are unique, and I cannot create a second copy of it outside of the macro. I assume that capturing it is unwanted, otherwise the canonical version wouldn't bother adding the additional layer of gensym. How could capturing a symbol like #:G826 be a problem? Please provide an example where this capture manifests.
We can demonstrate a behavioral difference between my-once-only and once-only:
Let's store our test form in a variable.
(defvar *form* '(lexalias a 0 (lexalias b (1+ a) (list a b))))
This test form exercises a macro called lexalias, which we will define in two ways. First with once-only:
(defmacro lexalias (var value &body body)
(once-only (value)
`(symbol-macrolet ((,var ,value))
,#body)))
(eval *form*) -> (0 1)
Then with my-once-only:
(defmacro lexalias (var value &body body)
(my-once-only (value)
`(symbol-macrolet ((,var ,value))
,#body)))
(eval *form*) -> (1 1)
Oops! The problem is that under my-once-only, both a and b end up being symbol-macrolet aliases for exactly the same gensym; the returned expression (list a b) ends up being something like (list #:g0025 #:g0025).
If you're writing a macro-writing helper that implements once-only evaluation, you have no idea how the symbol is going to be used by the code which calls the macro, whose author uses your once-only tool. There are two big unknowns: the nature of the macro and of its use.
As you can see, if you don't make fresh gensyms, it will not work correctly in all conceivable scenarios.
Often when I try to write a macro, I run up against the following difficulty: I need one form that is passed to the macro to be evaluated before being processed by a helper function that is invoked while generating the macro's expansion. In the following example, we are only interested in how we could write a macro to emit the code we want, and not in the uselessness of the macro itself:
Imagine (bear with me) a version of Common Lisp's lambda macro, where only the number of arguments is important, and the names and order of the arguments are not. Let's call it jlambda. It would be used like so:
(jlambda 2
...body)
where 2 is the arity of the function returned. In other words, this produces a binary operator.
Now imagine that, given the arity, jlambda produces a dummy lambda-list which it passes to the actual lambda macro, something like this:
(defun build-lambda-list (arity)
(assert (alexandria:non-negative-integer-p arity))
(loop for x below arity collect (gensym)))
(build-lambda-list 2)
==> (#:G15 #:G16)
The expansion of the above call to jlambda will look like this:
(lambda (#:G15 #:16)
(declare (ignore #:G15 #:16))
…body))
Let's say we need the jlambda macro to be able to receive the arity value as a Lisp form that evaluates to a non-negative integer (as opposed to receiving a non-negative integer directly) eg:
(jlambda (+ 1 1)
...body)
The form (+ 1 1) needs to be evaluated, then the result needs to be passed to build-lambda-list and that needs to be evaluated, and the result of that is inserted into the macro expansion.
(+ 1 1)
=> 2
(build-lambda-list 2)
=> (#:G17 #:18)
(jlambda (+ 1 1) ...body)
=> (lambda (#:G19 #:20)
(declare (ignore #:G19 #:20))
…body))
So here's a version of jlambda that works when the arity is provided as a number directly, but not when it's passed as a form to be evaluated:
(defun jlambda-helper (arity)
(let ((dummy-args (build-lambda-list arity)))
`(lambda ,dummy-args
(declare (ignore ,#dummy-args))
body)))
(defmacro jlambda (arity &body body)
(subst (car body) 'body (jlambda-helper arity)))
(jlambda 2 (print “hello”)) ==> #<anonymous-function>
(funcall *
'ignored-but-required-argument-a
'ignored-but-required-argument-b)
==> “hello”
“hello”
(jlambda (+ 1 1) (print “hello”)) ==> failed assertion in build-lambda-list, since it receives (+ 1 1) not 2
I could evaluate the (+ 1 1) using the sharp-dot read macro, like so:
(jlambda #.(+ 1 1) (print “hello”)) ==> #<anonymous-function>
But then the form cannot contain references to lexical variables, since they are not available when evaluating at read-time:
(let ((x 1))
;; Do other stuff with x, then:
(jlambda #.(+ x 1) (print “hello”))) ==> failure – variable x not bound
I could quote all body code that I pass to jlambda, define it as a function instead, and then eval the code that it returns:
(defun jlambda (arity &rest body)
(let ((dummy-args (build-lambda-list arity)))
`(lambda ,dummy-args
(declare (ignore ,#dummy-args))
,#body)))
(eval (jlambda (+ 1 1) `(print “hello”))) ==> #<anonymous-function>
But I can't use eval because, like sharp-dot, it throws out the lexical environment, which is no good.
So jlambda must be a macro, because I don't want the function body code evaluated until the proper context for it has been established by jlambda's expansion; however it must also be a function, because I want the first form (in this example, the arity form) evaluated before passing it to helper functions that generate the macro expansion. How do I overcome this Catch-22 situation?
EDIT
In response to #Sylwester 's question, here's an explanation of the context:
I'm writing something akin to an “esoteric programming language”, implemented as a DSL in Common Lisp. The idea (admittedly silly but potentially fun) is to force the programmer, as far as possible (I'm not sure how far yet!), to write exclusively in point-free style. To do this, I will do several things:
Use curry-compose-reader-macros to provide most of the functionality required to write in point-free style in CL
Enforce functions' arity – i.e. override CL's default behaviour that allows functions to be variadic
Instead of using a type system to determine when a function has been “fully applied” (like in Haskell), just manually specify a function's arity when defining it.
So I'll need a custom version of lambda for defining a function in this silly language, and – if I can't figure that out - a custom version of funcall and/or apply for invoking those functions. Ideally they'll just be skins over the normal CL versions that change the functionality slightly.
A function in this language will somehow have to keep track of its arity. However, for simplicity, I would like the procedure itself to still be a funcallable CL object, but would really like to avoid using the MetaObject Protocol, since it's even more confusing to me than macros.
A potentially simple solution would be to use a closure. Every function could simply close over the binding of a variable that stores its arity. When invoked, the arity value would determine the exact nature of the function application (i.e. full or partial application). If necessary, the closure could be “pandoric” in order to provide external access to the arity value; that could be achieved using plambda and with-pandoric from Let Over Lambda.
In general, functions in my language will behave like so (potentially buggy pseudocode, purely illustrative):
Let n be the number of arguments provided upon invocation of the function f of arity a.
If a = 0 and n != a, throw a “too many arguments” error;
Else if a != 0 and 0 < n < a, partially apply f to create a function g, whose arity is equal to a – n;
Else if n > a, throw a “too many arguments” error;
Else if n = a, fully apply the function to the arguments (or lack thereof).
The fact that the arity of g is equal to a – n is where the problem with jlambda would arise: g would need to be created like so:
(jlambda (- a n)
...body)
Which means that access to the lexical environment is a necessity.
This is a particularly tricky situation because there's no obvious way to create a function of a particular number of arguments at runtime. If there's no way to do that, then it's probably easiest to write a a function that takes an arity and another function, and wraps the function in a new function that requires that is provided the particular number of arguments:
(defun %jlambda (n function)
"Returns a function that accepts only N argument that calls the
provided FUNCTION with 0 arguments."
(lambda (&rest args)
(unless (eql n (length args))
(error "Wrong number of arguments."))
(funcall function)))
Once you have that, it's easy to write the macro around it that you'd like to be able to:
(defmacro jlambda (n &body body)
"Produces a function that takes exactly N arguments and and evalutes
the BODY."
`(%jlambda ,n (lambda () ,#body)))
And it behaves roughly the way you'd want it to, including letting the arity be something that isn't known at compile time.
CL-USER> (let ((a 10) (n 7))
(funcall (jlambda (- a n)
(print 'hello))
1 2 3))
HELLO
HELLO
CL-USER> (let ((a 10) (n 7))
(funcall (jlambda (- a n)
(print 'hello))
1 2))
; Evaluation aborted on #<SIMPLE-ERROR "Wrong number of arguments." {1004B95E63}>.
Now, you might be able to do something that invokes the compiler at runtime, possibly indirectly, using coerce, but that won't let the body of the function be able to refer to variables in the original lexical scope, though you would get the implementation's wrong number of arguments exception:
(defun %jlambda (n function)
(let ((arglist (loop for i below n collect (make-symbol (format nil "$~a" i)))))
(coerce `(lambda ,arglist
(declare (ignore ,#arglist))
(funcall ,function))
'function)))
(defmacro jlambda (n &body body)
`(%jlambda ,n (lambda () ,#body)))
This works in SBCL:
CL-USER> (let ((a 10) (n 7))
(funcall (jlambda (- a n)
(print 'hello))
1 2 3))
HELLO
CL-USER> (let ((a 10) (n 7))
(funcall (jlambda (- a n)
(print 'hello))
1 2))
; Evaluation aborted on #<SB-INT:SIMPLE-PROGRAM-ERROR "invalid number of arguments: ~S" {1005259923}>.
While this works in SBCL, it's not clear to me whether it's actually guaranteed to work. We're using coerce to compile a function that has a literal function object in it. I'm not sure whether that's portable or not.
NB: In your code you use strange quotes so that (print “hello”) doesn't actually print hello but the whatever the variable “hello” evaluates to, while (print "hello") does what one would expect.
My first question is why? Usually you know how many arguments you are taking compile time or at least you just make it multiple arity. Making an n arity function only gives you errors when passwd with wrong number of arguments as added feature with the drawback of using eval and friends.
It cannot be solved as a macro since you are mixing runtime with macro expansion time. Imagine this use:
(defun test (last-index)
(let ((x (1+ last-index)))
(jlambda x (print "hello"))))
The macro is expanded when this form is evaluated and the content replaced before the function is assigned to test. At this time x doesn't have any value whatsoever and sure enough the macro function only gets the symbols so that the result need to use this value. lambda is a special form so it again gets expanded right after the expansion of jlambda, also before any usage of the function.
There is nothing lexical happening since this happens before the program is running. It could happen before loading the file with compile-file and then if you load it will load all forms with the macros already expanded beforehand.
With compile you can make a function from data. It is probably as evil as eval is so you shouldn't be using it for common tasks, but they exist for a reason:
;; Macro just to prevent evaluation of the body
(defmacro jlambda (nexpr &rest body)
`(let ((dummy-args (build-lambda-list ,nexpr)))
(compile nil (list* 'lambda dummy-args ',body))))
So the expansion of the first example turns into this:
(defun test (last-index)
(let ((x (1+ last-index)))
(let ((dummy-args (build-lambda-list x)))
(compile nil (list* 'lambda dummy-args '((print "hello")))))))
This looks like it could work. Lets test it:
(defparameter *test* (test 10))
(disassemble *test*)
;Disassembly of function nil
;(CONST 0) = "hello"
;11 required arguments <!-- this looks right
;0 optional arguments
;No rest parameter
;No keyword parameters
;4 byte-code instructions:
;0 (const&push 0) ; "hello"
;1 (push-unbound 1)
;3 (calls1 142) ; print
;5 (skip&ret 12)
;nil
Possible variations
I've made a macro that takes a literal number and makes bound variables from a ... that can be used in the function.
If you are not using the arguments why not make a macro that does this:
(defmacro jlambda2 (&rest body)
`(lambda (&rest #:rest) ,#body))
The result takes any number of arguments and just ignores it:
(defparameter *test* (jlambda2 (print "hello")))
(disassemble *test*)
;Disassembly of function :lambda
;(CONST 0) = "hello"
;0 required arguments
;0 optional arguments
;Rest parameter <!-- takes any numer of arguments
;No keyword parameters
;4 byte-code instructions:
;0 (const&push 0) ; "hello"
;1 (push-unbound 1)
;3 (calls1 142) ; print
;5 (skip&ret 2)
;nil
(funcall *test* 1 2 3 4 5 6 7)
; ==> "hello" (prints "hello" as side effect)
EDIT
Now that I know what you are up to I have an answer for you. Your initial function does not need to be runtime dependent so all functions indeed have a fixed arity, so what we need to make is currying or partial application.
;; currying
(defmacro fixlam ((&rest args) &body body)
(let ((args (reverse args)))
(loop :for arg :in args
:for r := `(lambda (,arg) ,#body)
:then `(lambda (,arg) ,r)
:finally (return r))))
(fixlam (a b c) (+ a b c))
; ==> #<function :lambda (a) (lambda (b) (lambda (c) (+ a b c)))>
;; can apply multiple and returns partially applied when not enough
(defmacro fixlam ((&rest args) &body body)
`(let ((lam (lambda ,args ,#body)))
(labels ((chk (args)
(cond ((> (length args) ,(length args)) (error "too many args"))
((= (length args) ,(length args)) (apply lam args))
(t (lambda (&rest extra-args)
(chk (append args extra-args)))))))
(lambda (&rest args)
(chk args)))))
(fixlam () "hello") ; ==> #<function :lambda (&rest args) (chk args)>
;;Same but the zero argument functions are applied right away:
(defmacro fixlam ((&rest args) &body body)
`(let ((lam (lambda ,args ,#body)))
(labels ((chk (args)
(cond ((> (length args) ,(length args)) (error "too many args"))
((= (length args) ,(length args)) (apply lam args))
(t (lambda (&rest extra-args)
(chk (append args extra-args)))))))
(chk '()))))
(fixlam () "hello") ; ==> "hello"
If all you want is lambda functions that can be applied either partially or fully, I don't think you need to pass the amount of parameters explicitly. You could just do something like this (uses Alexandria):
(defmacro jlambda (arglist &body body)
(with-gensyms (rest %jlambda)
`(named-lambda ,%jlambda (&rest ,rest)
(cond ((= (length ,rest) ,(length arglist))
(apply (lambda ,arglist ,#body) ,rest))
((> (length ,rest) ,(length arglist))
(error "Too many arguments"))
(t (apply #'curry #',%jlambda ,rest))))))
CL-USER> (jlambda (x y) (format t "X: ~s, Y: ~s~%" x y))
#<FUNCTION (LABELS #:%JLAMBDA1046) {1003839D6B}>
CL-USER> (funcall * 10) ; Apply partially
#<CLOSURE (LAMBDA (&REST ALEXANDRIA.0.DEV::MORE) :IN CURRY) {10038732DB}>
CL-USER> (funcall * 20) ; Apply fully
X: 10, Y: 20
NIL
CL-USER> (funcall ** 100) ; Apply fully again
X: 10, Y: 100
NIL
CL-USER> (funcall *** 100 200) ; Try giving a total of 3 args
; Debugger entered on #<SIMPLE-ERROR "Too many arguments" {100392D7E3}>
Edit: Here's also a version that lets you specify the arity. Frankly, I don't see how this could possibly be useful though. If the user cannot refer to the arguments, and nothing is done with them automatically, then, well, nothing is done with them. They might as well not exist.
(defmacro jlambda (arity &body body)
(with-gensyms (rest %jlambda n)
`(let ((,n ,arity))
(named-lambda ,%jlambda (&rest ,rest)
(cond ((= (length ,rest) ,n)
,#body)
((> (length ,rest) ,n)
(error "Too many arguments"))
(t (apply #'curry #',%jlambda ,rest)))))))
CL-USER> (jlambda (+ 1 1) (print "hello"))
#<CLOSURE (LABELS #:%JLAMBDA1085) {1003B7913B}>
CL-USER> (funcall * 2)
#<CLOSURE (LAMBDA (&REST ALEXANDRIA.0.DEV::MORE) :IN CURRY) {1003B7F7FB}>
CL-USER> (funcall * 5)
"hello"
"hello"
Edit2: If I understood correctly, you might be looking for something like this (?):
(defvar *stack* (list))
(defun jlambda (arity function)
(lambda ()
(push (apply function (loop repeat arity collect (pop *stack*)))
*stack*)))
CL-USER> (push 1 *stack*)
(1)
CL-USER> (push 2 *stack*)
(2 1)
CL-USER> (push 3 *stack*)
(3 2 1)
CL-USER> (push 4 *stack*)
(4 3 2 1)
CL-USER> (funcall (jlambda 4 #'+)) ; take 4 arguments from the stack
(10) ; and apply #'+ to them
CL-USER> (push 10 *stack*)
(10 10)
CL-USER> (push 20 *stack*)
(20 10 10)
CL-USER> (push 30 *stack*)
(30 20 10 10)
CL-USER> (funcall (jlambda 3 [{reduce #'*} #'list])) ; pop 3 args from
(6000 10) ; stack, make a list
; of them and reduce
; it with #'*
Looking at Practical Common Lisp, we're looking at a simple automated unit test framework. We're trying to write a macro to be used as such:
(check (= (+ 1 2) 3) (= (- 1 4) 9))
This should expand to something using a previously defined function report-result. The suggested implementation is:
(defmacro check (&body forms)
`(progn
,#(loop for f in forms collect `(report-result ,f ',f))))
However, that expansion seems rather procedural to me. I wanted to replace the loop with a mapcar to expand to something like this:
(mapcar #'(lambda (form) (report-result form 'form)) (list form-1 ... form-n))
However, I'm clearly lacking the macro-writing skills to do so. Can someone come up with one such macro?
In case it's relevant, this is the definition of report-result:
(defun report-result (result form)
(format t "~:[FAIL~;pass~] ... ~a~%" result form))
It's indeed fairly simple: you just place the collect expression into the body of your mapcar:
(defmacro check (&body forms)
`(progn
,#(mapcar #'(lambda (form)
`(report-result ,form ',form))
forms)))
You don't really need to know anything about the "macro-y" stuff that's going on, in order to do the replacement you want, which is simply replacing a loop with some other equivalent expression: it will work just as well in a macro context as it would outside.
If you want to expand to a mapcar you can, but there's no real reason to do so, since the list's size is known at compile time. Here's what that would look like:
(defmacro check (&body forms)
`(let ((results (list ,#(mapcar #'(lambda (form)
`(list ,form ',form))
forms))))
(mapcar #'(lambda (result)
(report-result (car result) (cadr result)))
results)))
Which expands like so
> (macroexpand-1 '(check (+ 1 2) (* 2 3)))
(let ((results (list (list (+ 1 2) '(+ 1 2))
(list (* 2 3) '(* 2 3)))))
(mapcar #'(lambda (result) (report-result (car result) (cadr result)))
results))
Which as you can see is rather awkward: the macro already has the forms like (+ 1 2) available to it, but in order to preserve them to runtime for the mapcar lambda to see, you have to emit the input form twice. And you have to produce the whole list to map over, rather than just producing a list that's "finished" to begin with. Additionally, this produces a list as output, and requires having all the inputs and outputs in memory at once: the original macro with progn produced the inputs and outputs one at a time, and discarded them when finished.
I'm reading Paul Graham's ANSI Common Lisp. In the chapter about macros he shows the following example:
(defmacro in (obj &rest choices)
(let ((insym (gensym)))
`(let ((,insym ,obj))
(or ,#(mapcar #'(lambda (c) `(eql ,insym ,c))
choices)))))
(Returns true if the first argument is equal to any of the other arguments)
He holds that it can't be written as a function. Wouldn't this function have the same functionality?
(defun in (obj &rest choices)
(reduce (lambda (x y)
(or x (eql y obj)))
choices
:initial-value nil))
The difference I see is that the macro will only evaluate arguments till it finds an eql argument. Is that it?
The point is, that the macro version evaluates the arguments lazily (it expands into an OR) stopping if a match is found. This cannot be achieved with a function, since a funcall will always evaluate all arguments first.
> (macroexpand '(in 42
(long-computation-1)
(long-computation-2)
(long-computation-3)))
(LET ((#:G799 42))
(OR (EQL #:G799 (LONG-COMPUTATION-1))
(EQL #:G799 (LONG-COMPUTATION-2))
(EQL #:G799 (LONG-COMPUTATION-3))))
To get the same effect you would need to write:
(defun in (obj &rest choices)
(reduce (lambda (x y)
(or x (eql (funcall y) obj)))
choices
:initial-value nil))
and use it this way:
(in 42
(function long-computation-1)
(function long-computation-2)
(function long-computation-3))
In this page there is a comment after the post that gives a very short implementation of amb as a procedure:
(define (amb-backtrack)
(error "no solution found"))
(define (amb . args)
(call/cc (lambda (return)
(let ((backtrack amb-backtrack))
(map (lambda (x)
(call/cc (lambda (k)
(set! amb-backtrack k)
(return x))))
args)
(backtrack 'fail)))))
But I usually see amb implemented as a macro -- in the schemers.org FAQ, and also in Dorai Sitaram's book:
(define amb-fail '*)
(define initialize-amb-fail
(lambda ()
(set! amb-fail
(lambda ()
(error "amb tree exhausted")))))
(initialize-amb-fail)
(define-macro amb
(lambda alts...
`(let ((+prev-amb-fail amb-fail))
(call/cc
(lambda (+sk)
,#(map (lambda (alt)
`(call/cc
(lambda (+fk)
(set! amb-fail
(lambda ()
(set! amb-fail +prev-amb-fail)
(+fk 'fail)))
(+sk ,alt))))
alts...)
(+prev-amb-fail))))))
So -- the macro version is longer, and a little harder to understand. I could not see any advantages of it over the procedure version, and of course I would rather use a procedure than a macro. Is there anything I missed?
The difference is that a procedure call always evaluates all the arguments.
(amb 1 (very-expensive-computation))
will, with the procedure version of amb, perform the very-expensive-computation, then yield 1. If yielding 1 suffices for all further computation, then you've wasted a lot of time on a computation whose result value is never used. Even worse things happen, as #Eli Barzilay mentions in a comment, when amb is used to model an infinite non-determinism such as generating all natural numbers.
The macro version avoids this and its behavior is therefore closer to that of non-deterministic programming languages such as Prolog.