Introduction to J Programming Language (2004) (opens in new tab)

I went to Trinity and Howland attempted to teach us J there. The entire thing went terribly. I agree with Iverson's assessment: learning J is a lot like learning math in that mathematicians often use weird and cryptic syntactic forms that are only tangentially related to other, similar forms. Unlike APL, this problem occurs in absurd detail in J. APL, at least, has strictly different symbols. J, on the other hand, uses strangely and opaquely-overloaded syntax. By way of example, consider the following:

A programmer may expect < to perform less-than. In J it does, when used as a dyad. Used monadically, it acts as a boxing operator (similar to Scheme's box). That's strange, but livable. The problem, though, is in J, <. means 'lesser of' (min), and <: means 'less than or equal'. Again, this is a potentially livable arrangement. Except that . means determinant (or dot product, depending on monadic or dyadic appearance) and : means either 'explicit' or 'monadic/dyadic', a sort of combinator that accepts a monadic and dyadic operator and yields a new one. That is, in Scheme, dyadic usage of : is akin to a case-lambda that checks if it uses one or two arguments.

If I present to you the program <., what does it do? Well, it certainly does something different from < ., which does something different from < :, which does something different from <:, which does something different from < . :, which again differs from <. :, and so on. This is exacerbated by operations such as 'table', written /, which bleed into other places. For example, </ inserts box, while <:/ inserts a decrement! And < :/ boxes up the result of an 'explicit' being 'inserted', whatever that means. And this syntactic problem plagues the entire vocabulary of the language:

http://www.jsoftware.com/help/dictionary/vocabul.htm

A clear optimization would be to just give these things their real names, use text, and write S-expressions. At least then the programs would be readable!

The largest problem, however, is that for being a functional language, J does little if anything to convey the important idea that have been developed in functional programming: recursion. It simply isn't how things 'are done' in J. So we have a language with bad syntax, good semantics, and a horrible problem from an educational perspective. It has some nice additions above and beyond other APLs, but after that it falls flat. I'd urge anyone who has read this far to learn APL instead of J; it will at least be readable by other APL programmers.

(I didn't mean to go on a rant, but I suffered through Howland's class on J at Trinity, and it wasn't until I had a professor show me Scheme in ~3 hours that I saw why functional programming was truly worthwhile.)

I have been teaching myself K, and one useful resource for me has been http://codegolf.stackexchange.com. There are a handful of active users who write solutions in J, so it might be a good way to compare your solutions with those of another J programmer.

The code and the interpreter together are both so terse that they fit in the cache. That buys you a lot of speed.

IIRC the interpreter is thin as hell. Almost no syntax or interpretation overhead. Also the semantics of J makes it very mechanical sympathetic. Some kind of macros based wrappers to express loops.

ps: I'm not sure it was J .. maybe K. And I can't find the C source anymore. Haa...

pps: check this http://lambda-the-ultimate.org/node/5075 they discuss performance.

> I'm an assembler hog and any language that doesn't compile to native code is just a toy to me.

First, what exactly is native code? Is it the PythonVM/JVM byte code? The x86 code it is JITted into? The microcode that interprets the "native" x86 code? You probably mean the x86/ARM/whatever bytecode, but that's not a trivial definition.

Practically speaking, compiling to "native code" is mostly useful given a known architecture. "rep movsb" used to be the fastest way to copy memory around; and then it was slow as mollasses; and then it was fast again. Table lookups were the fastest way to do 6-bit by 8-bit multiplications up until the 386 or 486 - and then the multiplier became faster than memory.

Somewhat surprisingly, J (and APL and K) tend to focus on promoting those things that have always been true - e.g. sequential access is much faster than non sequential; branch free code is better than branching code; small code/data that fit in cache/main memory is better than larger code/data that doesn't. The computation model fits modern CPUs and GPUs very well, despite being approximately the same one since the 1950s.

It's a different take on the "better algorithms beat faster implementations" - a good language that makes a good fast solution intuitive is often, in practice, better than one that can be faster with a lot of intricate work.

Let's say I create a new, interpreted language. It's called fgrep and it only has a couple instructions for searching text. But I spend 5 years optimizing those functions.

Now, if you need recursively search through a directory full of text files, would you use fgrep or would you write your own "high performance" search utility with a hard-coded (none of that slow run-time interpreting!) list of files and needle?

LtU discusses exactly the issue of how 'interpreted' can be faster: http://lambda-the-ultimate.org/node/5075 .

if it makes you feel any better in Dyalog APL[1] I'd write that as:

  s←{
  	sp←⍵[⍋ ⍵]
  	h t←(1↑sp) (1↓sp)
  	h , t[⍋ 12○(t-h)]
  }
  or I can make it ugly as hell:
  s←{h,t[⍋ 12○(t←1↓p)-h←1↑p←⍵[⍋ ⍵]]}

In Dyalog (and most APLs now) ←{...} creates function that automatically has ⍵ as it's left argument (so invoking "s points" ⍵ is points )

⍋ give the indexes in sorted order and [ ] is how you access elements so ⍵[⍋ ⍵] is how you sort an array

↑ and ↓ are take and drop so (1↑sp) takes the head and (1↓sp) drops the head leaving the tail "h t←(1↑sp) (1↓sp)" is just multiple assignments

J's verb trains are cute but kind of awful

"h , t[⍋ 12○(t-h)]" sorts the tail according to the value of each elements phase (I agree the circle functions are particularly bad) and cats its back on to the head (also strictly speaking i don't need the parens around "t-h" but I like them).

[1] well no actually I wouldn't because Dyalog's grade up (⍋) function doesn't work on complex numbers (though how to write a function ( gu ) that DOES is covered here http://dfns.dyalog.com/c_le.htm )

It would have only one page that said: "This is not for you".

According to http://en.wikipedia.org/wiki/List_of_programming_languages we're missing A, H, I, M, N, O, P, and U-Z, so half the alphabet.

However, this list is incomplete.

H - http://dcoj.wmh3.com/cscos/h/

M is sometimes used to refer to MUMPS (see http://en.wikipedia.org/wiki/MUMPS and http://www.mumps.org/ ). It was also a language at Microsoft, http://www.theregister.co.uk/2008/10/10/dial_m_for_microsoft .

N - http://link.springer.com/chapter/10.1007%2F978-3-642-76153-9...

Z - https://github.com/chrisdone/z

In other words, the answer to your question is either "yes" or "that's already happened."