Hacker News Reader

Comment by akkartik | Hacker News Reader

I think one reason to add characters and streams is so you can implement read and print, the R and the P in a Lisp REPL. The original Lisp assumes s-expressions are given, but provides no mechanism for creating them out of characters read from keyboard or printed to screen. Bel is taking the same axiomatic approach of the original Lisp but to a more realistic system.

Similarly, ccc exists so onerr can exist, which exists so saferead can exist, which exists so the REPL can show errors on screen.

That hangs together for me. But I still don't understand the distinction you're making between representing as lists vs being lists.

View on HN

> so the REPL can show errors on screen

And how do you formalize the screen? (BTW, this is just speculation on your part. The word "screen" doesn't actually appear anywhere in the article.)

You can't have it both ways. Either you are going to formalize everything, or you're going to punt at some point and yield to practical considerations. And formalizing I/O is particularly hard.

> I still don't understand the distinction you're making between representing as lists vs being lists.

(lambda () t) is a list. In McCarthy's original lisp, as in Bel, this list IS a function. This would work:

    (setq l '(lambda () t))
    (l) --> T

But in modern Lisps you have to EVALUATE this list in order to get a function that you can call. In Scheme:

    ts> (define l '(lambda () #t))                  
    l
    ts> l
    (lambda () #t)
    ts> (l)
    Error: illegal function 
    ts> (define p (eval l))
    p
    ts> p
    #<CLOSURE>
    ts> (p)
    #t
    ts> (procedure? l)
    #f
    ts> (procedure? p)
    #t
    ts> (list? l)
    #t
    ts> (list? p)
    #f

> (lambda () t) is a list.

This is not right, and it's a very basic thing to get wrong. I don't think you understand what you're criticizing.

Even replacing the Bel name, this is not right:

> (fn () t) is a list

You _do_ have to eval it.

    bel> (eval '(fn () t))
    (lit clo nil () t)

I see the returned list as akin to a serialization format. It's an s-expression rather than some opaque representation that can't be printed. That seems nice, more debuggable. I don't see what makes it harder to compile than an opaque representation. You can compile it just fine, you're just carrying around the source code in say a debug segment of the image.

> (BTW, this is just speculation on your part. The word "screen" doesn't actually appear anywhere in the article.)

Look, you asked what the point of streams was. I pointed out what the metacircular source code does with streams.

If someone asks what the purpose of stdout is in Unix, it seems reasonable (and a hoary tradition) to justify it in terms of printing to screen. stdout is about much more than a screen, but "screen" is an easy answer to "why does my program need stdout?"

> I don't think you understand what you're criticizing.

Oh, I'm pretty sure I do.

    Functions are represented using lists. For example, here is a function that takes one argument, and returns that plus 1.
    
    (lit clo nil (x) (+ x 1))
    
    The first element, lit, says that this is a literal object, not to be  evaluated.

I don't know where the disconnect is, but I'll stop here. The statement you just quoted disagrees with what you said on two occasions before:

> It's fine to represent functions as lists, but having functions actually be lists rather than a first-class primitive is a huge mistake.

> (lambda () t) is a list

You said this would work in Bel:

    (setq l '(lambda () t))
    (l) --> T

This is false even after fixing the Bel names for things.

The disconnect is that PG misuses the word "represented". Functions in Bel are not just represented as lists, they are lists. Specifically, they are lists whose first element is the symbol LIT and whose second element is the symbol CLO. Functions can also be represented as lists that start with the symbol FN.

    You would not ordinarily express a function using its literal representation. Usually you'd say
    
    (fn (x) (+ x 1))
    
    which yields the function above.  [i.e. (lit clo nil (x) (+ x 1))]

So a list that starts with FN represents a function, but it is not a function. It has to be evaluated to produce a function, but the function that it produces is just another list, one which starts with LIT rather than FN.

In Bel you have to eval (fn (x) (+ x 1)) before you can call it. Do you agree? Following your terms in https://news.ycombinator.com/item?id=40409467, in this respect Bel feels like a modern lisp. In other words, this quoted passage of yours is not correct:

    (lambda () t) is a list. In McCarthy's original lisp, as in Bel, this list IS a function. This would work:

    (setq l '(lambda () t))
    (l) --> T

This would not work in Bel.

Yes, that's right, but you have missed the point. In McCarthy's Lisp, functions were lists whose first element was the symbol LAMBDA. In Bel, functions are lists whose first element is LIT and whose second element is CLO. But the details of what is in those lists is irrelevant. The point is, in both langauges, functions are lists. You can call CAR on a function and get a result. In McCarthy's Lisp, that result will be LAMBDA and in Bel it will be LIT. If you call CAR on a function in a modern Lisp you will get an error.

    ? (car (lambda () t))
    > Error: The value #<Anonymous Function #x302000E7B61F> is not of the expected type LIST.

Contrast with:

    ? (car '(lambda () t))
    LAMBDA

In addition, there is a special form in BEL called FN which, when evaluated, produces a function, i.e. a list whose first element is LIT and whose second element is CLO. McCarthy's original lisp had no such function. The special handling of functions, i.e. lists starting with LAMBDA, was hard-coded into the definition of EVAL. But that is also an irrelevant detail.

Ok, I think we're past the misunderstanding now.

I don't understand why this makes it hard to compile Bel. Application code won't check the first two elements of the list to determine if a term is a function. Instead it will say ((isa 'function) ___). It seems fine that a program using the former approach will run slower.

> I don't understand why this makes it hard to compile Bel.

Because Bel is supposed to be a specification, not an implementation. And if you specify that functions are lists then you are no longer free to implement them in a way that does not preserve the illusion that they are lists. Worse, in Bel, lists are mutable. That makes reasoning about programs nearly impossible. Every time you call CONS (which in Bel is spelled JOIN for some reason) you may or may not be producing a function. Every time you mutate a pair you may or may not be creating a function, or destroying one, or changing a function's environment.

There is not and cannot be any barrier between the mutable and the immutable in Bel. It is not possible to add a data type that guarantees invariants. Mutable pairs are all you have to build with, and so whatever you build, not matter how carefully you design it, the rug can get pulled out from under you at any time by an errant call to XAR or XDR.

Generalizing from the representation of functions to the absence of invariants makes your point crystal clear, thanks.

What you want in this kind of situation is for those symbols to be internal objects that cannot be written. Or perhaps symbols in a private system package, e.g.:

  (sys:lit sys:clo nil (x) (+ x 1))

No program using symbols in the ordinary namespace will produce this object by accident. It has to use the sys: stuff and when it does that, the language specification can say that it's sticking the proverbial fork into the proverbial toaster.

That doesn't help at all. These symbols don't have to be produced by the reader. All you need to do to get access to these symbols is to call FN once. That gives you a list whose first element is LIT and second element is CLO. So all you need to "stick a fork" in the Bel toaster is to be able to call FN and XAR.

Yes. In a nutshell, the problem is that (consp x) being true does not rule out (functionp x), which it should. (Unless consp is hacked in an unspeakable way.)

In a real implementation, you'd want a new kind of cons cell with a separate type tag so that it fails consp. It's okay if car and cdr like it.

> (consp x) being true does not rule out (functionp x), which it should

Exactly. That's what "punning lists" means.

> It's okay if car and cdr like it.

That's arguable. But the real problem is if XAR and XDR accept it. Functions, however they are implemented, should not be mutable. That way lies madness.

Why would you need read and print in a language that isn't yet intended to program computers?

You don't need I/O in order to describe input and output in an abstract language. Literals, including quoted lists, are enough to encode input into the program as part of its body. The result of a function can be described using a notation like (+ 2 2) -> 4. It's not actually running so nothing needs to print that.

Reply on news.ycombinator.com