I have been told that the following expression is intended to evaluate to 0, but that many implementations of Scheme evaluate it as 1:
(let ((cont #f))
(letrec ((x (call-with-current-continuation (lambda (c) (set! cont c) 0)))
(y (call-with-current-continuation (lambda (c) (set! cont c) 0))))
(if cont
(let ((c cont))
(set! cont #f)
(set! x 1)
(set! y 1)
(c 0))
(+ x y))))
I must admit that I cannot tell where even to start with this. I understand the basics of continuations and call/cc, but can I get a detailed explanation of this expression?
This is an interesting fragment. I came across this question because I was searching for discussions of the exact differences between letrec and letrec*, and how these varied between different versions of the Scheme reports, and different Scheme implementations. While experimenting with this fragment, I did some research and will report the results here.
If you mentally walk through the execution of this fragment, two questions should be salient to you:
Q1. In what order are the initialization clauses for x and y evaluated?
Q2. Are all the initialization clauses evaluated first, and their results cached, and then all the assignments to x and y performed afterwards? Or are some of the assignments made before some of the initialization clauses have been evaluated?
For letrec, the Scheme reports say that the answer to Q1 is "unspecified." Most implementations will in fact evaluate the clauses in left-to-right order; but you shouldn't rely on that behavior.
Scheme R6RS and R7RS introduce a new binding construction letrec* that does specify left-to-right evaluation order. It also differs in some other ways from letrec, as we'll see below.
Returning to letrec, the Scheme reports going back at least as far as R5RS seem to specify that the answer to Q2 is "evaluate all the initialization clauses before making any of the assignments." I say "seem to specify" because the language isn't as explicit about this being required as it might be. As a matter of fact, many Scheme implementations don't conform to this requirement. And this is what's responsible for the difference between the "intended" and "observed" behavior wrt your fragment.
Let's walk through your fragment, with Q2 in mind. First we set aside two "locations" (reference cells) for x and y to be bound to. Then we evaluate one of the initialization clauses. Let's say it's the clause for x, though as I said, with letrec it could be either one. We save the continuation of this evaluation into cont. The result of this evaluation is 0. Now, depending on the answer to Q2, we either assign that result immediately to x or we cache it to make the assignment later. Next we evaluate the other initialization clause. We save its continuation into cont, overwriting the previous one. The result of this evaluation is 0. Now all of the initialization clauses have been evaluated. Depending on the answer to Q2, we might at this point assign the cached result 0 to x; or the assignment to x may have already occurred. In either case, the assignment to y takes place now.
Then we begin evaluating the main body of the (letrec (...) ...) expression (for the first time). There is a continuation stored in cont, so we retrieve it into c, then clear cont and set! each of x and y to 1. Then we invoke the retrieved continuation with the value 0. This goes back to the last-evaluated initialization clause---which we've assumed to be y's. The argument we supply to the continuation is then used in place of the (call-with-current-continuation (lambda (c) (set! cont c) 0)), and will be assigned to y. Depending on the answer to Q2, the assignment of 0 to x may or may not take place (again) at this point.
Then we begin evaluating the main body of the (letrec (...) ...) expression (for the second time). Now cont is #f, so we get (+ x y). Which will be either (+ 1 0) or (+ 0 0), depending on whether 0 was re-assigned to x when we invoked the saved continuation.
You can trace this behavior by decorating your fragment with some display calls, for example like this:
(let ((cont #f))
(letrec ((x (begin (display (list 'xinit x y cont)) (call-with-current-continuation (lambda (c) (set! cont c) 0))))
(y (begin (display (list 'yinit x y cont)) (call-with-current-continuation (lambda (c) (set! cont c) 0)))))
(display (list 'body x y cont))
(if cont
(let ((c cont))
(set! cont #f)
(set! x 1)
(set! y 1)
(c 'new))
(cons x y))))
I also replaced (+ x y) with (cons x y), and invoked the continuation with the argument 'new instead of 0.
I ran that fragment in Racket 5.2 using a couple of different "language modes", and also in Chicken 4.7. Here are the results. Both implementations evaluated the x init clause first and the y clause second, though as I said this behavior is unspecified.
Racket with #lang r5rs and #lang r6rs conforms to the spec for Q2, and so we get the "intended" result of re-assigning 0 to the other variable when the continuation is invoked. (When experimenting with r6rs, I needed to wrap the final result in a display to see what it would be.)
Here is the trace output:
(xinit #<undefined> #<undefined> #f)
(yinit #<undefined> #<undefined> #<continuation>)
(body 0 0 #<continuation>)
(body 0 new #f)
(0 . new)
Racket with #lang racket and Chicken don't conform to that. Instead, after each initialization clause is evaluated, it gets assigned to the corresponding variable. So when the continuation is invoked, it only ends up re-assigning a value to the final value.
Here is the trace output, with some added comments:
(xinit #<undefined> #<undefined> #f)
(yinit 0 #<undefined> #<continuation>) ; note that x has already been assigned
(body 0 0 #<continuation>)
(body 1 new #f) ; so now x is not re-assigned
(1 . new)
Now, as to what the Scheme reports really do require. Here is the relevant section from R5RS:
library syntax: (letrec <bindings> <body>)
Syntax: <Bindings> should have the form
((<variable1> <init1>) ...),
and <body> should be a sequence of one or more expressions. It is an error
for a <variable> to appear more than once in the list of variables being bound.
Semantics: The <variable>s are bound to fresh locations holding undefined
values, the <init>s are evaluated in the resulting environment (in some
unspecified order), each <variable> is assigned to the result of the
corresponding <init>, the <body> is evaluated in the resulting environment, and
the value(s) of the last expression in <body> is(are) returned. Each binding of
a <variable> has the entire letrec expression as its region, making it possible
to define mutually recursive procedures.
(letrec ((even?
(lambda (n)
(if (zero? n)
#t
(odd? (- n 1)))))
(odd?
(lambda (n)
(if (zero? n)
#f
(even? (- n 1))))))
(even? 88))
===> #t
One restriction on letrec is very important: it must be possible to evaluate
each <init> without assigning or referring to the value of any <variable>. If
this restriction is violated, then it is an error. The restriction is necessary
because Scheme passes arguments by value rather than by name. In the most
common uses of letrec, all the <init>s are lambda expressions and the
restriction is satisfied automatically.
The first sentence of the "Semantics" sections sounds like it requires all the assignments to happen after all the initialization clauses have been evaluated; though, as I said earlier, this isn't as explicit as it might be.
In R6RS and R7RS, the only substantial changes to this part of the specification is the addition of a requirement that:
the continuation of each <init> should not be invoked more than once.
R6RS and R7RS also add another binding construction, though: letrec*. This differs from letrec in two ways. First, it does specify a left-to-right evaluation order. Correlatively, the "restriction" noted above can be relaxed somewhat. It's now okay to reference the values of variables that have already been assigned their initial values:
It must be possible to evaluate each <init> without assigning or
referring to the value of the corresponding <variable> or the
<variable> of any of the bindings that follow it in <bindings>.
The second difference is with respect to our Q2. With letrec*, the specification now requires that the assignments take place after each initialization clause is evaluated. Here is the first paragraph of the "Semantics" from R7RS (draft 6):
Semantics: The <variable>s are bound to fresh locations, each
<variable> is assigned in left-to-right order to the result of evaluating
the corresponding <init>, the <body> is evaluated in the resulting
environment, and the values of the last expression in <body> are
returned. Despite the left-to-right evaluation and assignment order, each
binding of a <variable> has the entire letrec* expression as its region,
making it possible to define mutually recursive procedures.
So Chicken, and Racket using #lang racket---and many other implementations---seem in fact to implement their letrecs as letrec*s.
The reason for this being evaluated to 1 is because of (set! x 1). If instead of 1 you set x to 0 then it will result in zero. This is because the continuation variable cont which is storing the continuation is actually storing the continuation for y and not for x as it is being set to y's continuation after x's.
Related
Footnote #28 of SICP says the following:
Embedded definitions must come first in a procedure body. The management is not responsible for the consequences of running programs that intertwine definition and use.
What exactly does this mean? I understand:
"definitions" to be referring to both procedure creations and value assignments.
"a procedure body" to be as defined in section 1.1.4. It's the code that comes after the list of parameters in a procedure's definition.
"Embedded definitions must come first in a procedure body" to mean 'any definitions that are made and used in a procedure's body must come before anything else'.
"intertwine definition and use" to mean 'calling a procedure or assigned value before it has been defined;'
However, this understanding seems to contradict the answers to this question, which has answers that I can summarise as 'the error that your quote is referring to is about using definitions at the start of a procedure's body that rely on definitions that are also at the start of that body'. This has me triply confused:
That interpretation clearly contradicts what I've said above, but seems to have strong evidence - compiler rules - behind it.
SICP seems happy to put definition in a body with other definitions that use them. Just look at the sqrt procedure just above the footnote!
At a glance, it looks to me that the linked question's author's real error was treating num-prod like a value in their definition of num rather than as a procedure. However, the author clearly got it working, so I'm probably wrong.
So what's really going on? Where is the misunderstanding?
In a given procedure's definition / code,
"internal definitions" are forms starting with define.
"a procedure body" is all the other forms after the forms which start with define.
"embedded definitions must come first in a procedure body" means all the internal define forms must go first, then all the other internal forms. Once a non-define internal form appears, there can not appear an internal define form after that.
"no intertwined use" means no use of any name before it's defined. Imagine all internal define forms are gathered into one equivalent letrec, and follow its rules.
Namely, we have,
(define (proc args ...)
;; internal, or "embedded", definitions
(define a1 ...init1...)
(define a2 ...init2...)
......
(define an ...initn...)
;; procedure body
exp1 exp2 .... )
Any ai can be used in any of the initj expressions, but only inside a lambda expression.(*) Otherwise it would refer to the value of ai while aj is being defined, and that is forbidden, because any ai names are considered not yet defined while any of the initj expressions are being evaluated.
(*) Remember that (define (foo x) ...x...) is the same as (define foo (lambda (x) ...x...)). That's why the definitions in that sqrt procedure in the book you mention are OK -- they are all lambda expressions, and any name's use inside a lambda expression will only actually refer to that name's value when the lambda expression's value -- a lambda function -- will be called, not when that lambda expression is being evaluated, producing that lambda function which is its value.
The book is a bit vague at first with the precise semantics of its language but in Scheme the above code is equivalent to
(define proc
(lambda (args ...)
;; internal, or "embedded", definitions
(letrec ( (a1 ...init1...)
(a2 ...init2...)
......
(an ...initn...) )
;; procedure body
exp1 exp2 ....
)))
as can be seen explained, for instance, here, here or here.
For example,
;; or, equivalently,
(define (my-proc x) (define my-proc
(lambda (x)
(define (foo) a) (letrec ( (foo (lambda () a))
(define a x) (a x) )
;; my-proc's body ;; letrec's body
(foo)) (foo))))
first evaluates the lambda expression, (lambda () a), and binds the name foo to the result, a function; that function will refer to the value of a when (foo) will be called, so it's OK that there's a reference to a inside that lambda expression -- because when that lambda expression is evaluated, no value of a is immediately needed, just the reference to its future value, under the name a, is present there; i.e. the value that a will have after all the names in that letrec are initialized, and the body of letrec is entered. Or in other words, when all the internal defines are completed and the body proper of the procedure my-proc is entered.
So we see that foo is defined, but is not used during the initializations; a is defined but is not used during the initializations; thus all is well. But if we had e.g.
(define (my-proc x)
(define (foo) x) ; `foo` is defined as a function
(define a (foo)) ; `foo` is used by being called as a function
a)
then here foo is both defined and used during the initializations of the internal, or "embedded", defines; this is forbidden in Scheme. This is what the book is warning about: the internal definitions are only allowed to define stuff, but its use should be delayed for later, when we're done with the internal defines and enter the full procedure's body.
This is subtle, and as the footnote and the question you reference imply, the subtlties can vary depending on a particular language's implementation.
These issues will be covered in far more detail later in the book (Chapters 3 and 4) and, generally, the text avoids using internal definitions so that these issues can be avoided until they are examined in detail.
A key difference between between the code above the footnote:
(define (sqrt x)
(define (good-enough? guess)
(< (abs (- (square guess) x)) 0.001))
(define (improve guess)
(average guess (/ x guess)))
(define (sqrt-iter guess)
(if (good-enough? guess)
guess
(sqrt-iter (improve guess))))
(sqrt-iter 1.0))
and the code in the other question:
(define (pi-approx n)
(define (square x) (* x x))
(define (num-prod ind) (* (* 2 ind) (* 2 (+ ind 1)))) ; calculates the product in the numerator for a certain term
(define (denom-prod ind) (square (+ (* ind 2 ) 1))) ;Denominator product at index ind
(define num (product num-prod 1 inc n))
(define denom (product denom-prod 1 inc n))
is that all the defintions in former are procedure definitions, whereas num and denom are value defintions. The body of a procedure is not evaluated until that procedures is called. But a value definition is evaluated when the value is assigned.
With a value definiton:
(define sum (add 2 2))
(add 2 2) will evaluated when the definition is evaluated, if so add must already be defined. But with a procedure definition:
(define (sum n m) (add n m))
a procedure object will be assigned to sum but the procedure body is not yet evaluated so add does not need be defined when sum is defined, but must be by the time sum is called:
(sum 2 2)
As I say there is a lot sublty and a lot of variation so I'm not sure if the following is always true for every variation of scheme, but within 'SICP scheme' you can say..
Valid (order of evaluation of defines not significant):
;procedure body
(define (sum m n) (add m n))
(define (add m n) (+ m n))
(sum 2 2)
Also valid:
;procedure body
(define (sum) (add 2 2))
(define (add m n) (+ m n))
(sum)
Usually invalid (order of evaluation of defines is significant):
;procedure body
(define sum (add 2 2))
(define (add m n) (+ m n))
Whether the following is valid depends on the implementation:
;procedure body
(define (add m n) (+ m n))
(define sum (add 2 2))
And finally an example of intertwining definition and use, whether this works also depends on the implementation. IIRC, this would work with Scheme described in Chapter 4 of the book if the scanning out has been implemented.
;procedure body
(sum)
(define (sum) (add 2 2))
(define (add m n) (+ m n))
It is complicated and subtle, but the key points are:
value definitions are evaluated differently from procedure definitions,
behaviour inside blocks may be different from outside blocks,
there is variation between scheme implementations,
the book is designed so you don't have to worry much about this until Chapter 3,
the book will cover this in detail in Chapter 4.
You discovered one of the difficulties of scheme. And of lisp. Due to this kind of chicken and egg issue there is no single lisp, but lots of lisps appeared.
If the binding forms are not present in code in a letrec-logic in R5RS and letrec*-logic in R6RS, the semantics is undefined. Which means, everything depend on the will of the implementor of scheme.
See the paper Fixing Letrec: A Faithful Yet Efficient Implementation of Scheme’s Recursive Binding Construct.
Also, you can read the discussions from the mailing list from 1986, when there was no general consensus among different implementors of scheme.
Also, at MIT there were developed 2 versions of scheme -- the student version and the researchers' development version, and they behaved differently concerning the order of define forms.
In the book Structure and Interpretation of Computer Programs
by H. Abelson and G. J. Sussman with J. Sussman,
the accumulation or fold-right is introduced in Section 2.2.3 as follows:
(define (accumulate op initial sequence)
(if (null? sequence)
initial
(op (car sequence)
(accumulate op initial (cdr sequence)))))
I tried to use this to take the and of a list of Boolean variables, by writing:
(accumulate and
true
(list true true false))
However, this gave me the error and: bad syntax in DrRacket (with #lang sicp),
and I had to do this instead:
(accumulate (lambda (x y) (and x y))
true
(list true true false))
Why? I believe it has something to do with how and is a special form,
but I don't understand Scheme enough to say.
Perhaps I'm just missing some obvious mistake...
You answered your own question: and is a special form (not a normal procedure!) with special evaluation rules, and accumulate expects a normal procedure, so you need to wrap it inside a procedure.
To see why and is a special form, consider these examples that demonstrate that and requires special evaluation rules (unlike procedures), because it short-circuits whenever it finds a false value:
; division by zero never gets executed
(and #f (/ 1 0))
=> #f
; division by zero gets executed during procedure invocation
((lambda (x y) (and x y)) #f (/ 1 0))
=> /: division by zero
From the R5RS standard:
Values might be defined as follows:
(define (values . things)
(call-with-current-continuation
(lambda (cont) (apply cont things))))
My first interpretation of this was that an expression like (+ (values 1 2)) is equivalent to (apply + '(1 2)) and would yield the result 3. But, according to my tests, this interpretation is not correct. Here is my interpretation of the code above: values is a function taking any number of arguments, bundled into a list called things. Then, the current continuation (the place where values is used) is called with the list things "unbundled".
What am I missing? The example above (+ (values 1 2)) gives an error or 1 depending on the interpreter I used.
See, when you type
(+ (values 1 2))
the continuation of the call to values is actually a single argument to +. So, it is either treated as 1 (the first element to the list, the first value produced by the procedure), or an error. R5RS says in this regard:
Except for continuations created by the call-with-values procedure, all continuations take exactly one value. The effect of passing no value or more than one value to continuations that were not created by call-with-values is unspecified.
On the other hand, call-with-values would correctly bind your list's elements to its consumer argument's formal arguments:
Calls its producer argument with no values and a continuation that, when passed some values, calls the consumer procedure with those values as arguments.
In order to understand what this definition of values means, you need to also understand the definition of call-with-current-continuation, which it is defined in terms of. And helpfully, the documentation for values mentions call-with-values, as an example of how to use the result of values.
So, you could use (values 1 2) in a context like:
(call-with-values (lambda () (values 1 2))
(lambda (x y) (+ x y)))
In the paper Fixing Letrec: A Faithful Yet Efficient Implementation
of Scheme’s Recursive Binding Construct by Dybvig et al. it is said that (emphasis mine):
A theoretical solution to these problems is to restrict letrec so
that its left-hand sides are unassigned and right-hand sides are lambda
expressions. We refer to this form of letrec as fix, since it amounts to
a generalized form of fixpoint operator. The compiler can handle fix
expressions efficiently, and there can be no violations of the letrec
restriction with fix. Unfortunately, restricting letrec in this manner is not an option for the implementor and would in any case reduce the generality and convenience of the construct.
I have not scrutinized the R5RS report, but I have used letrec and the equivalent "named let" in Scheme programs and the unfortunate consequences mentioned in the paper are not clear to me, can someone enlighten me ?
The R5RS letrec restriction says something like these are in violation:
(let ((x 10))
(letrec ((x x))
x))
(letrec ((y (+ x 5))
(x 5))
(list x y))
Thus, it's not specified what would happen and it would certainly not be portable Scheme. It may evaluate to 10 and (5 10), the implementation might signal an error or you get an undefined value that might result in an error getting signaled. I have tested Racket, Gambit, Chicken and Ikarus and not one of them signal anything in the first case and they all evaluate to an unspecified value. Ikarus is the only one that returned (5 10) in the latter while the others all got contract errors since an unspecified value as argument violates +'s contract. (Ikarus always evaluates operands right to left)
The R[567]RS reports all state that if all expressions are lambda expressions you have nothing to worry about and I think that is the clue. Another is that you should not try to shadow like you would do with (named) let.
There is a follow up on the original paper that is entitled Fixing letrec (reloaded) that has macros that implements the "fix".
With equational syntax,
letrec x = init-x
y = init-y
body....
the restriction is that no RHS init... expression can cause evaluation of (or assignment to) any of the LHS variables, because all init...s are evaluated while all the variables are still unassigned. IOW no init... should reference any of the variables directly and immediately. It is OK of course for any of the init...s to contain lambda-expressions which can indeed reference any of the variables (that's the purpose of letrec after all). When these lambda-expressions will be evaluated, the variables will be already assigned the values of the evaluated init... expressions.
The authors say, to require all the RHSes to be lambda-expressions would simplify the implementation, because there's no chance for misbehaving code causing premature evaluation of LHS variables inside some RHS. But unfortunately, this changes letrec's semantics and thus is not an option. It would also prohibit simple use of outer variables in RHSes and thus this new cut-down letrec would also be less general and less convenient.
You also mention named let but it is not equivalent to letrec: its variables are bound as-if by let, only the looping function itself is bound via letrec:
(let ((x 1)(y 2))
(let g ((x x) (y x))
(if (> x 0) (g (- x y) y) (display x)))
(display x))
01
;Unspecified return value
(let ((x 1)(y 2))
(letrec ((g (lambda (x y)
(if (> x 0) (g (- x y) y) (display x)))))
(g x x)) ; no 'x' in letrec's frame, so refer to outer frame
(display x))
01
;Unspecified return value
The following Scheme code
(let ((x 1))
(define (f y) (+ x y))
(set! x 2)
(f 3) )
which evaluates to 5 instead of 4. It is surprising considering Scheme promotes static scoping. Allowing subsequent mutation to affect bindings in the closed environment in a closure seems to revert to kinda dynamic scoping. Any specific reason that it is allowed?
EDIT:
I realized the code above is less obvious to reveal the problem I am concerned. I put another code fragment below:
(define x 1)
(define (f y) (+ x y))
(set! x 2)
(f 3) ; evaluates to 5 instead of 4
There are two ideas you are confusing here: scoping and indirection through memory. Lexical scope guarantees you that the reference to x always points to the binding of x in the let binding.
This is not violated in your example. Conceptually, the let binding is actually creating a new location in memory (containing 1) and that location is the value bound to x. When the location is dereferenced, the program looks up the current value at that memory location. When you use set!, it sets the value in memory. Only parties that have access to the location bound to x (via lexical scope) can access or mutate the contents in memory.
In contrast, dynamic scope allows any code to change the value you're referring to in f, regardless of whether you gave access to the location bound to x. For example,
(define f
(let ([x 1])
(define (f y) (+ x y))
(set! x 2)
f))
(let ([x 3]) (f 3))
would return 6 in an imaginary Scheme with dynamic scope.
Allowing such mutation is excellent. It allows you to define objects with internal state, accessible only through pre-arranged means:
(define (adder n)
(let ((x n))
(lambda (y)
(cond ((pair? y) (set! x (car y)))
(else (+ x y))))))
(define f (adder 1))
(f 5) ; 6
(f (list 10))
(f 5) ; 15
There is no way to change that x except through the f function and its established protocols - precisely because of lexical scoping in Scheme.
The x variable refers to a memory cell in the internal environment frame belonging to that let in which the internal lambda is defined - thus returning the combination of lambda and its defining environment, otherwise known as "closure".
And if you do not provide the protocols for mutating this internal variable, nothing can change it, as it is internal and we've long left the defining scope:
(set! x 5) ; WRONG: "x", what "x"? it's inaccessible!
EDIT: your new code, which changes the meaning of your question completely, there's no problem there as well. It is like we are still inside that defining environment, so naturally the internal variable is still accessible.
More problematic is the following
(define x 1)
(define (f y) (+ x y))
(define x 4)
(f 5) ;?? it's 9.
I would expect the second define to not interfere with the first, but R5RS says define is like set! in the top-level.
Closures package their defining environments with them. Top-level environment is always accessible.
The variable x that f refers to, lives in the top-level environment, and hence is accessible from any code in the same scope. That is to say, any code.
No, it is not dynamic scoping. Note that your define here is an internal definition, accessible only to the code inside the let. In specific, f is not defined at the module level. So nothing has leaked out.
Internal definitions are internally implemented as letrec (R5RS) or letrec* (R6RS). So, it's treated the same (if using R6RS semantics) as:
(let ((x 1))
(letrec* ((f (lambda (y) (+ x y))))
(set! x 2)
(f 3)))
My answer is obvious, but I don't think that anyone else has touched upon it, so let me say it: yes, it's scary. What you're really observing here is that mutation makes it very hard to reason about what your program is going to do. Purely functional code--code with no mutation--always produces the same result when called with the same inputs. Code with state and mutation does not have this property. It may be that calling a function twice with the same inputs will produce separate results.