Forth: The programming language that writes itself

Circa 1980 BASIC was the dominant language for micros because you could fit BASIC in a machine with 4k of RAM. Although you got 64k to play with pretty quickly (1983 or so) it still was a pain in the ass to implement compilers on many chips, especially the 6502, which had so few registers and addressing modes that you're likely to use virtual machine techniques, like Wozniak's SWEET 16 or the atrocious p-code machine that turned a generation of programmers away from PASCAL.

FORTH was an alternative language for small systems. From the viewpoint of a BASIC programmer in 1981 the obvious difference between BASIC and all the other languages which that you could write your own functions to add "words" to the language. FORTH, like Lisp, lets you not only write functions but create new control structures based on "words" having both a compile-time and run-time meaning.

FORTH's answer to line numbers in BASIC was that it provided direct access to blocks (usually 1024 bytes) on the disk with a screen editor (just about a screenful on a 40x25) You could type your code into blocks and later load them into the interpreter. Circa 1986 I wrote a FORTH for the TRS-80 Color Computer running the OS-9 operating system and instead of using blocks it had POSIX-style I/O functions.

FORTH was faster than BASIC and better for systems work, but BASIC was dominant. Probably the best way to use FORTH was to take advantage of it's flexibility to create a DSL that you write your applications in.

I had that issue, and I think I still might have it in my closet. (Weren't those Robert Tinney covers amazing?)

I always wanted to try out Forth but had no real opportunity. Maybe I should now?

I implemented the Calculus of Constructions in untyped lambda calculus in order to shorten the 643 byte C program computing Loader's Number to 1850 bits (under 232 bytes) [1], as one of the milestones reached by my functional busy beaver function [2].

[1] https://codegolf.stackexchange.com/questions/176966/golf-a-n...

[2] https://oeis.org/A333479

Grounding programming languages in mathematics like this is essentially the goal of Strachey and Scott's denotational semantics, which has been very influential in programming language theory:

https://en.wikipedia.org/wiki/Denotational_semantics

A lot of languages(including forth) maps really poorly to LC. Read some of the Forth writing on portability at all costs:

http://www.ultratechnology.com/antiansi.htm

> And we should aim for that, the babel tower of languages doesn't really help anyone.

What exactly do you mean with this? That the amount of programming languages available isn't actually helpful, it's detrimental?

Why specifically untyped?

We just refer to them as “slot machines” in the US

I had to have a peek if it's all just web. Apparently, no.

https://concatenative.org/wiki/view/Factor/UI

> The Factor UI is a GUI toolkit together with a set of developer tools, written entirely in Factor, implemented on top of a combination of OpenGL and native platform APIs: X11, Win32 and Cocoa.

> UI gadgets are rendered using the cross-platform OpenGL API, while native platform APIs are used to create windows and receive events. The platform bindings can be also used independently; X11 binding has also been used in a Factor window manager, Factory, which is no longer maintained. The Cocoa binding is used directly by the webkit-demo vocabulary in Factor.

Fascinating. Probably dead and no mention of Wayland, but fascinating.

I have very fond memories of programming in PostScript within NeWS/HyperNeWS - it did quite a few things that I've never seen in any other environment.

Edit: To be fair relying on PostScript probably did limit the appeal, but I actually really liked it.

note to those interested: no apple silicon support.

I had a copy of that as well - I forget whether it was a Christmas gift or if I bought it. The demos were neat, but I was lacking in ideas when I had time to play with it, and the Apple didn't go to college with me.

But if I were going to do some "from the ground up, using first principles, with nobody else's libraries" embedded work, Forth would certainly be something I'd consider.

> As a child, I really was amused by the demo of GraFORTH on Apple ][, which included 3D wireframe animations, which at the time were magical.

I originally wrote GraFORTH (https://archive.org/details/a2_GraFORTH_1981_Lutus_Paul) to escape the slow world of integer BASIC on my first computer (an Apple II). Because it relied on large blocks of assembly code to produce nice graphics, it perhaps misled people about what Forth could do on its own.

Later I wrote a variation I called TransFORTH (https://mirrors.apple2.org.za/ftp.apple.asimov.net/documenta...) that supported floating-point. I intended to combine GraFORTH and TransFORTH, but my computer didn't have enough RAM.

Innocent times, different world, before the personal computing tail began wagging the dog.

>Yeah, PostScript is a dialect of Forth.

My understanding is they were developed independently.

The difference between Forth and Lisp could not be more pronounced. Forth source code has entirely implicit structure, you can't even tell which function is called on which arguments. Lisp has entirely explicit structure which makes it much easier to read and edit. Lisp needs only a single primitive (lambda) to create the entire programming language, whereas Forth needs many primitives which break the core idea of the language in order to be usable. All of what is elegant about Lisp is ultimately lacking in Forth.

It's a different solution for a different time.

Forth was an excellent way to write a powerful and expressive programming language that could self-host with a bare minimum of assembly language "bare metal" programming.

The fridge-sized computer that Forth was originally developed on had double-digit kilobytes of memory (maybe 8192kwords, with 16-bit words) and clocked instructions through at a whopping 300kHz or so. The microcontroller that drives the Caps Lock LED on your keyboard is a hundred times faster with a hundred times the memory.

These days we do not need to squeeze editor, compiler, and target binary into such a tiny machine. If you're developing for a microcontroller you just use C on your "big" computer, which is unimaginably more powerful.

In the olden days of the 1990s I used a development system for embedded stuff that was written in and targetted Forth on a Z80 with a whopping 64kB of RAM and 5.25" floppies, but that was at least ten years old and five years out of date at the time.

You're probably reading my words on a slice of glass the size of half a sandwich that contains more computing power than existed in the whole world when Forth was first written.

It's a shame because writing something like Forth from the ground up (and I mean, assembly code to load the registers to start the ACIA to begin transmitting text to the terminal) perhaps in an emulated early 80s home computer is a great way to get a sense of what the chip behind it all is doing, and I feel that makes you a better programmer in "real" languages like Go or Python or C.

One simpler explanation: in forth you are forced to keep the stack, and modifications to the stack, in your short term memory, albeit only really three numbers in most cases. Whereas with C et al you simply look down the page at the variables, far less taxing on your short term memory.

Well-written and designed high-level forth words often transcend that and tend to be, quite literally, readable however, in a way that is incredibly rare to see in C et al. Of course the argument is that other programmers shouldn't be expected to see the problem in the way the original problem solver did.

I was lucky, early in my career, to work at a place which used a lot of Perl and to read Damian Conway’s book, Object Oriented Perl. It was an amazing, mind-expanding book for me. It was filled with examples of different approaches to object-oriented programming, more than I ever dreamt existed, and it showed how to implement them all in Perl.

So much power! And right in line with Perl’s mantra, “there’s more than one way to do it.”

Unfortunately, our codebase contained more than one way of doing it. Different parts of the code used different, incompatible object systems. It was a lot of extra work to learn them all and make them work with each other.

It was a relief to later move to a language which only supported a single flavor of object-oriented programming.

What I heard is with Forth, basically no 2 environments are alike, but highly customized, meaning every forth programmer creates his own language in the end for his custom needs.

So collaborating is a bit hard like this. The only serious forth programmer that I know, lives alone in the woods doing his things.

So from a aesthetic point of view, I really like the language, but for getting things done, especially in a collaborative way?

But who knows, maybe someone will write the right tools for that to change?

I don't think "power" is really that helpful a metric in determining how useful a programming language is. If you think of programming from the standpoint of trying to specify the program you want out of all of the possibly programs you could write, one of the most helpful things a programming language can do is eliminate programs that you don't want by making them impossible to write. From that standpoint, constraints are a feature, not a drawback.

I worked at a place that had a big Forth codebase that was doing something mission critical. It was really neat and cool once you finally got it, and probably hundreds or maybe thousands of people had touched it, worked on it and learned it, but the ramp was pretty brutal for your average developer and thus someone decided it would be better to build the same thing over with a shitty almost-C-but-not-quite interpreted language. It certainly made it easier for more people to understand and build, even if the solution was less elegant.

Powerful languages invites people to do needlessly complex things. Needlessly complex things are harder to understand. Harder to understand is worse.

Code that matters is usually read and extended many more times than it is written, over time by different people, so being straightforward beats most other things in practice

I think it's a simple abstraction situation and the move for programming environments that include everything.

Geordi Laforge doesn't code much on the Enterprise. He simply asks the computer to build him a model of the anomaly so he can test out ideas. In a way, modern languages like Python (even before LLMs) let you get a lot closer to that reality. Sure you had to know some language basics, but this was pretty minimal and you'd use those basic building blocks to glue together libraries to make an application. Python has a good library for practically anything I do and since this is standard, it's expected that a task doesn't take too long. I can't tell my boss I'll need 3 years to code my own solution that uses my own libraries for numpy and scipy. You're expected to glue libraries together. This is why MIT moved SICP from scheme to Python. It's a different world.

With Forth, every program is a work of art that encapsulates an entire solution to a problem from scratch. It's creator chuck moore takes this to such a level that he also fabs his own chips to work with his forth software optimally. These languages had libraries, but they weren't easy to share and didn't have any kind of repository before Perl's CPAN. Perl really took off for awhile, but Python won out by having a simpler language with builtin OO (Perl's approach was a really hacky builtin OO or you download a library...).

To be honest though, I spent a decade trying many languages (dozens including common lisp, Prolog, APL, C, Ada, Smalltalk, Perl, C#, C++, Tcl, Lua, Rust...etc) looking for the best and although I never became experts in those languages, I kept coming to the conclusion that for my particular set of needs, Python was the best I could find. I wasted a lot of time reading common lisp books and just found it much easier to get the same thing done in Python. Your mileage will vary if you're doing something like building a game engine. A lot of people are just doing process automation and stuff like that and languages like Python are just better than common lisp due to the environment and tooling benefits. Also, although Python isn't as conceptually beautiful as lisp, I found it much easier to learn. The syntax just really clicked for me and some people do prefer it.

> Why is it that languages like this don't scale?

Stanislav Datskovskiy addressed this rather well:

https://www.loper-os.org/?p=69

It kinda happened with markup languages. HTML, SVG, and some other domain specific markup languages are all XML, which is a subset of SGML.

The thing there is those DSLs have their own specs.

Coding is a social activity. Reading code is hard. When there are multiple ways of doing things, it's extra hard. People want to have relatively standardized ways of doing things so they can share code and reason about it easier.

If there's a lisp or racket or a forth that's defined as a DSL, it might take off if it's standardized and it's the best solution for the domain.

Sadly our industry carries mostly about brick layers and usually tries to go into technologies that make it easier to deal with employees like replaceable servants at low wage prices.

The large scale salaries SV style isn't something that you will find all over the globe, in many countries the pay is similar across all office workers, regardless if they are working with Git, or Office.

I'm not entirely sure this is different from other languages but I believe a common complaint about lisp is every solution ends up writing a DSL for that solution, making it hard to understand for anyone else. So it's a super power if you're a small team and especially if you're a team of 1. But if you're a large team it doesn't scale.

I think those other languages have real advantages you aren't seeing.

—·—

The other day akkartik wrote an implementation of the program Knuth used to introduce literate programming to the CACM readers: https://basiclang.solarpunk.au/d/7-don-knuths-original-liter...

It just tells you the top N words by frequency in its input (default N=100) with words of the same frequency ordered alphabetically and all words converted to lowercase. Knuth's version was about 7 pages of Pascal, maybe 3 pages without comments. It took akkartik 50 lines of idiomatic, simple Lua. I tried doing it in Perl; it was 6 lines, or 13 without relying on any of the questionable Perl shorthands. Idiomatic and readable Perl would be somewhere in between.

    #!/usr/bin/perl -w
    use strict;

    my $n = @ARGV > 1 ? pop @ARGV : 100;
    my %freq;

    while (my $line = <>) {
      for my $w ($line =~ /(\w+)/g) {
        $freq{(lc $w)}++;
      }
    }

    for my $w (sort { $freq{$b} <=> $freq{$a} || $a cmp $b } keys %freq) {
      print "$w\t$freq{$w}\n";
      last unless --$n;
    }

Then I tried writing a Common Lisp version. Opening a file, iterating over lines, hashing words and getting 0 as default, and sorting are all reasonably easy in CL, but splitting a line into words is a whole project on its own. And getting a command-line argument requires implementation-specific facilities that aren't standardized by CL! At least string-downcase exists. It was a lark, so I didn't finish.

(In Forth you'd almost have to write something equivalent to Knuth's Pascal, because it doesn't come with even hash tables and case conversion.)

My experience with Smalltalk is more limited but similar. You can do anything you want in it, it's super flexible, the tooling is great, but almost everything requires you to just write quite a bit more code than you would in Perl, Python, Ruby, JS, etc. And that means you have more bugs, so it takes you longer. And it doesn't really want to talk to the rest of the world—you can forget about calling a Squeak method from the Unix command line.

Smalltalk and CL have native code compilers available, which ought to be a performance advantage over things like Perl. Often enough, though, it's not. Part of the problem is that their compilers don't produce highly performant code, but they certainly ought to beat a dumb bytecode interpreter, right? Well, maybe not if the program's hot loop is inside a regular expression match or Numpy array operation.

And a decent native code compiler (GCC, HotSpot, LuaJIT, the Golang compilers, even ocamlopt) will beat any CL or Smalltalk compiler I have tried by a large margin. This is a shame because a lot of the extra hassle in Smalltalk and CL seems to be aimed at efficiency.

(Scheme might actually deliver the hoped-for efficiency in the form of Chez, but not Chicken. But Chicken can build executables and easily call C. Still, you'd need more code to solve this problem in Scheme than in Lua, much less Ruby.)

—·—

One of the key design principles of the WWW was the "principle of least power", which says that you should do each job with the least expressive language that you can. So the URL is a very stupid language, just some literal character strings glued together with delimiters. HTML is slightly less stupid, but you still can't program in it; you can only mark up documents. HTTP messages are similarly unexpressive. As much as possible of the Web is built out of these very limited languages, with only small parts being written in programming languages, where these limited DSLs can't do the job.

Lisp, Smalltalk, and Forth people tend to think this is a bad thing, because it makes some things—important things—unnecessarily hard to write. Alan Kay has frequently deplored the WWW being built this way. He would have made it out of mobile code, not dead text files with markup.

But the limited expressivity of these formats makes them easier to read and to edit.

I have two speech synthesis programs, eSpeak and Festival. Festival is written in Scheme, a wonderful, liberating, highly expressive language. eSpeak is in C++, which is a terrible language, so as much as possible of its functionality is in dumb data files that list pronunciations for particular letter sequences or entire words and whatnot. Festival does all of this configuration in Scheme files, and consequently I have no idea where to start. Fixing problems in eSpeak is easy, as long as they aren't in the C++ core; fixing problems in Festival is, so far, beyond my abilities.

(I'm not an expert in Scheme, but I don't think that's the problem—I mean, my Scheme is good enough that I wrote a compiler in it that implements enough of Scheme to compile itself.)

—·—

SQL is, or until recently was, non-Turing-complete, but expressive enough that 6 lines of SQL can often replace a page or three of straightforward procedural code—much like Perl in the example above, but more readable rather than less.

Similarly, HTML (or JSX) is often many times smaller than the code to produce the same layout with, say, GTK. And when it goes wrong, you can inspect the CSS rules applying to your DOM elements in a way that relies on them being sort of dumb, passive data. It makes them much more tractable in practice than Turing-complete layout systems like LaTeX and Qt3.

—·—

Perl and Forth both have some readability problems, but I think their main difficulty is that they are too error-prone. Forth, aside from being as typeless as conventional assembly, is one of the few languages where you can accidentally pass a parameter to the wrong call.

This sort of rhymes with what I was saying in 02001 in https://paulgraham.com/redund.html, that often we intentionally include redundancy in our expressions of programs to make them less error-prone, or to make the errors easily detectable.

One thing I note is that all of the languages you name are very far from the machine. Also Forth is not close to the modern machine. Note that it only has two integer types and the larger one can be aligned either way you make sure it is not.

They scale extremely effectively to large problems solved by a team size of one, maybe two.

The story goes that changing the language to fit how you're thinking about the problem is obstructive the rest of the people thinking about the same problem.

I'm pretty sure this story is nonsense. Popular though.

I don't think there's a unifying reason why programming languages languish in obscurity; it's certainly not because they're "too powerful." What does "powerful" even mean? I used to care more about comparing programming languages, but I mostly don't these days. Actually used/useful languages mostly just got lucky: C was how you wrote code for Unix; Python was Perl but less funny-looking; Ruby was Rails; JavaScript is your only choice in a web browser; Lisp had its heyday in the age of symbolic AI.

Forth and (R4RS) Scheme are simple to implement, so they're fun toys. Some other languages like Haskell have interesting ideas but don't excel at solving any particular problems. Both toy and general-purpose programming languages are plentiful.

frankly it's a miracle any of them scaled at all, such popularity mostly comes down to an arbitrary choice made decades ago by a lucky vendor instead of some grand overarching design

> However, no language should permit defining the value of 4 by 12, as there is no situation in which this can bring more good than harm in the long term.

A Skil saw should not permit you sticking your fingers in the spinning blade, yet most people know that this is a stupid and dangerous thing to do.

The quote makes more sense IMO for array languages like J that support a tacit style. J's "trains" just make things flow without a lot of variables. Aaron Hsu's Co Dfns compiler (spoken about on here and YouTube) also uses this style with Dyalog APL.

Forth is concatenative, so you can build the words on top of each other without worrying about a ton of variables. So I think it's partially true for Forth.

Yeah, I wouldn't have phrased it like in the article either. What I'd say is that Forth is more about naming processes than variables.

I believe, that you that sumes as mean.