Well, here's an example where not having first-class macros is confusing:

in takes any number of arguments, and returns whether the first argument is repeated in the rest of them:

  arc> (in 2 1 2 3)
  t
  arc> (in 0 1 2 3)
  nil

Ok, let's say I have a list, and I want to find out whether something is in it:

  arc> (apply in 2 (list 1 2 3))
  Error: "Function call on inappropriate object #(tagged mac #<procedure: in>) (0 1 2 3)"

Oh dear. That's not right.

The solution, at least in Arc, is to use mem:

  arc> (mem 2 (list 1 2 3))
  (2 3)

It returns the first sub-list of its second argument such that the car of the returned list is the element searched for -- or 'nil if the element isn't in the list. So yes, you can make it work, in this case. But it's surely more confusing than having (apply in 2 (list 1 2 3)) work.

Oh you picked a doozie. Apply won't work on macros even if you have first-class macros. Even kernel says, "apply expects an applicative." I spent weeks understanding why. The reason has to do with apply's special status as a dual of eval, and the fact that macros implicitly eval their result in the caller's lexical scope. Here's an example to think about:

  (mac foo(a b) `(cons ,a ,b))
  (apply foo '(1 (2 3)))
  => (cons 1 (2 3))
  => error: 2 is not a function

Further discussion (more than you could ever want): http://arclanguage.org/item?id=15659; http://arclanguage.org/item?id=15907

I guess the alternative would be to define "apply" as a macro which evals the cons of its first argument and the other list of arguments...

I addressed this approach at the top of the second link: http://arclanguage.org/item?id=15907

Basically you'd have to:

1. eval the list

2. cons the macro to the list, eval to yield the macroexpansion

3. implicitly eval the macroexpansion -- but not eval the elements of the spliced in list because they've already been eval'd once.

Step 3 throws a spanner in the works.

So you might say, fine, let's quote all the elements of the spliced in list. And that'll seem to work for the previous example, but you'll eventually run into an example like this:

  (apply = '(x 3))
  => (= 'x '3)
  => error: can't assign to (quote x)

So you have to perform step 3 without actually inserting a literal quote. I gave up at this point.

Ah I see.

I've started implementing a Lisp interpreter in C# and I also was going along the path of making everything "first class" by using f-expressions where appropriate. So far my only actual f-expressions are "lambda" and "if" - everything else is bootstrapped as macros or functions.

My "apply" currently expects a function but I was toying with making it "eval" the cons macro... but then "apply" has to become an f-expression instead of just a plain normal function which seemed inelegant so I left it as a function that requires a function.

Reading your story... I think I'll leave it permanently side-stepped!! :-)

Nice to meet ya! I've been working on a lisp as well: http://github.com/akkartik/wart#readme

BTW, in case you see this: I solved the problem for a large subset of macros. Details: http://arclanguage.org/item?id=16378

> You can't return one from a function, for instance.

Actually you can, for example:

    (assign macro-maker 
      (fn (x)
        (annotate 'mac (fn (y) `(cons ,x ',y)))))

    (assign m1 (macro-maker 1))

    (disp (m1 a)) ; print (1 . a)

    (macex '(m1 3)) ; print (cons 1 (quote a))

Above code are tested in tryarc.org.

> Non-first-class macros are the way to go for efficiency, at least, and first-class macros themselves are fairly useless.

It's true that non-first-class macros can be more efficiency. But first class macros can benefit from it's nature - a lexical function with environment. Construct such macros can be fancy.