Interview from : "Masterminds of Programming by Shane Warden, Federico Biancuzzi"

Chapter 4. Forth

Forth is a stack-based, concatenative language designed by Chuck Moore in the 1960s. Its main features are the use of a stack to hold data, and words that operate on the stack, popping arguments and pushing results. The language itself is small enough that it runs on anything from embedded machines to supercomputers, and expressive enough to build useful programs out of a few hundred words. Successors include Chuck Moore’s own colorForth, as well as the Factor programming language.

The Forth Language and Language Design

How do you define Forth?

Chuck Moore: Forth is a computer language with minimal syntax. It features an explicit parameter stack that permits efficient subroutine calls. This leads to postfix expressions (operators follow their arguments) and encourages a highly factored style of programming with many short routines sharing parameters on the stack.

I read that the name Forth stands for fourth-generation software. Would you like to tell us more about it?

Chuck: Forth is derived from “fourth,” which alludes to "fourth-generation computer language.” As I recall, I skipped a generation. FORTRAN/COBOL were first-generation languages; Algol/Lisp, second. These languages all emphasized syntax. The more elaborate the syntax, the more error checking is possible. Yet most errors occur in the syntax. I determined to minimize syntax in favor of semantics. And indeed, Forth words are loaded with meaning.

You consider Forth a language toolkit. I can understand that view, given its relatively simple syntax compared to other languages and the ability to build a vocabulary from smaller words. Am I missing anything else?

Chuck: No, it’s basically the fact that it’s extremely factored. A Forth program consists of lots of small words, whereas a C program consists of a smaller number of larger words.

By small word, I mean one with a definition typically one line long. The language can be built up by defining a new word in terms of previous words and you just build up that hierarchy until you have maybe a thousand words. The challenge there is 1) deciding which words are useful, and 2) remembering them all. The current application I’m working on has a thousand words in it. And I’ve got tools for searching for words, but you can only search for a word if you remember that it exists and pretty much how it’s spelled.

Now, this leads to a different style of programming, and it takes some time for a programmer to get used to doing it that way. I’ve seen a lot of Forth programs that look very much like C programs transliterated into Forth, and that isn’t the intent. The intent is to have a fresh start. The other interesting thing about this toolkit, words that you define this way are every bit as efficient or significant as words that are predefined in the kernel. There’s no penalty for doing this.

Does the externally visible structure consisting of many small words derive from Forth’s implementation?

Chuck: It’s a result of our very efficient subroutine call sequences. There’s no parameter passing because the language is stack-based. It’s merely a subroutine call and return. The stack is exposed. The machine language is compiled. A switch to and from a subroutine is literally one call instruction and one return instruction. Plus you can always reach down into the equivalent of an assembly language. You can define a word that will execute actual machine instructions instead of subroutine calls, so you can be as efficient as any other language, maybe more efficient than some.

You don’t have the C calling overhead.

Chuck: Right. This gives the programmer a huge amount of flexibility. If you come up with a clever factoring of a problem, you can not only do it efficiently, you can make it extraordinarily readable.

On the other hand, if you do it badly, you can end up with code that no one else can read—code your manager can’t understand, if managers can understand anything. And you can create a real mess. So it’s a two-edged sword. You can do very well; you can do very badly.

What would you say (or what code would you show) to a developer who uses another programming language to make him interested in Forth?

Chuck: It is very hard to interest an experienced programmer in Forth. That’s because he has invested in learning the tools for his language/operating system and has built a library appropriate for his applications. Telling him that Forth would be smaller, faster, and easier is not persuasive compared to having to recode everything. A novice programmer, or an engineer needing to write code, doesn’t face that obstacle and is much more receptive—as might be the experienced programmer starting a new project with new constraints, as would be the case with my multicore chips.

You mentioned that a lot of Forth programs you’ve seen look like C programs. How do you design a better Forth program?

Chuck: Bottom-up.

First, you presumably have some I/O signals that you have to generate, so you generate them. Then you write some code that controls the generation of those signals. Then you work your way up until finally you have the highest-level word, and you call it go and you type go and everything happens.

I have very little faith in systems analysts who work top-down. They decide what the problem is and then they factor it in such a way that it can be very difficult to implement.

Domain-driven design suggests describing business logic in terms of the customer’s vocabulary. Is there a connection between building up a vocabulary of words and using the terms of art from your problem domain?

Chuck: Hopefully the programmer knows the domain before he starts writing. I would talk to the customer. I would listen to the words he uses and I would try to use those words so that he can understand what the program’s doing. Forth lends itself to this kind of readability because it has postfix notation.

If I was doing a financial application, I’d probably have a word called “percent.” And you could say something like “2.03 percent”. And the argument’s percent is 2.03 and everything works and reads very naturally.

How can a project started on punch cards still be useful on modern computers in the Internet era? Forth was designed on/for the IBM 1130 in 1968. That it is the language of choice for parallel processing in 2007 is surely amazing.

Chuck: It has evolved in the meantime. But Forth is the simplest possible computer language. It places no restrictions upon the programmer. He/she can define words that succinctly capture aspects of a problem in a lean, hierarchical manner.

Do you consider English readability as a goal when you design programs?

Chuck: At the very highest level, yes, but English is not a good language for description or functionality. It wasn’t designed for that, but English does have the same characteristic as Forth in the sense that you can define new words.

You define new words by explaining what they are in previously defined words mostly. In a natural language, this can be problematic. If you go to a dictionary and check that out, you find that often the definitions are circular and you don’t get any content.

Does the ability to focus on words instead of the braces and brackets syntax you might have in C make it easier to apply good taste to a Forth program?

Chuck: I would hope so. It takes a Forth programmer who cares about the appearance of things as opposed merely to the functionality. If you can achieve a sequence of words that flow together, it’s a good feeling. That’s really why I developed colorForth. I became annoyed at the syntax that was still present in Forth. For instance, you could limit a comment by having a left parenthesis and a right parenthesis.

I looked at all of those punctuation marks and said, “Hey, maybe there’s a better way.” The better way was fairly expensive in that every word in the source code had to have a tag attached to it, but once I swallowed that overhead, it became very pleasant that all of those funny little symbols went away and were replaced by the color of the word which was, to me, a much gentler way of indicating functionality.

I get interminable criticism from people who are color blind. They were really annoyed that I was trying to rule them out of being programmers, but somebody finally came up with a character set distinction instead of a color distinction, which is a pleasant way of doing it also.

The key is the four-bit tag in each word, which gives you 16 things that we’re to do, and the compiler can determine immediately what’s intended instead of having to infer it from context.

Second- and third-generation languages embraced minimalism, for example with meta-circular bootstrapping implementations. Forth is a great example of minimalism in terms of language concepts and the amount of hardware support required. Was this a feature of the times, or was it something you developed over time?

Chuck: No, that was a deliberate design goal to have as small a kernel as possible. Predefine as few words as necessary and then let the programmer add words as he sees fit.

The prime reason for that was portability. At the time, there were dozens of minicomputers and then there became dozens of microcomputers. And I personally had to put Forth on lots of them.

I wanted to make it as easy as possible. What happens really is there might be a kernel with 100 words or so that is just enough to generate a—I’ll call it an operating system, but it’s not quite—that has another couple hundred words. Then you’re ready to do an application.

I would provide the first two stages and then let the application programmers do the third, and I was usually the application programmer, too. I defined the words I knew were going to be necessary. The first hundred words would be in machine language probably or assembler or at least be dealing directly with the particular platform. The second two or three hundred words would be high-level words, to minimize machine dependence in the lower, previously defined level. Then the application would be almost completely machine independent, and it was easy to port things from one minicomputer to another.

Were you able to port things easily above that second stage?

Chuck: Absolutely. I would have a text editor, for instance, that I used to edit the source code. It would usually just transfer over without any changes.

Is this the source of the rumor that every time you ran across a new machine, you immediately started to port Forth to it?

Chuck: Yes. In fact, it was the easiest path to understanding how the machine worked, what its special features were based on how easy it was to implement the standard package of Forth words.

How did you invent indirect-threaded code?

Chuck: Indirect-threaded code is a somewhat subtle concept. Each Forth word has an entry in a dictionary. In direct-threaded code, each entry points to code to be executed when that word is encountered. Indirect-threaded code points to a location that contains the address of that code. This allows information besides the address to be accessed—for instance, the value of a variable.

This was perhaps the most compact representation of words. It has been shown to be equivalent to both direct-threaded and subroutine-threaded code. Of course these concepts and terminology were unknown in 1970. But it seemed to me the most natural way to implement a wide variety of kinds of words.

How will Forth influence future computer systems?

Chuck: That has already happened. I’ve been working on microprocessors optimized for Forth for 25 years, most recently a multicore chip whose cores are Forth computers.

What does Forth provide? As a simple language, it allows a simple computer: 256 words of local memory; 2 push-down stacks; 32 instructions; asynchronous operation; easy communication with neighbors. Small and low-power.

Forth encourages highly factored programs. Such are well-suited to parallel processing, as required by a multicore chip. Many simple programs encourage thoughtful design of each. And requiring perhaps only 1% the code that would otherwise be written.

Whenever I hear people boasting of millions of lines of code, I know they have greviously misunderstood their problem. There are no contemporary problems requiring millions of lines of code. Instead there are careless programmers, bad managers, or impossible requirements for compatibility.

Using Forth to program many small computers is an excellent strategy. Other languages just don’t have the modularity or flexibility. And as computers get smaller and networks of them are cooperating (smart dust?), this will be the environment of the future.

This sounds like one major idea of Unix: multiple programs, each doing just one thing, that interact. Is that still the best design today? Instead of multiple programs on one computer, might we have multiple programs across a network?

Chuck: The notion of multithreaded code, as implemented by Unix and other OSes, was a precursor to parallel processing. But there are important differences.

A large computer can afford the considerable overhead ordinarily required for multithreading. After all, a huge operating system already exists. But for parallel processing, almost always the more computers, the better.

With fixed resources, more computers mean smaller computers. And small computers cannot afford the overhead common to large ones.

Small computers will be networked, on chip, between chips and across RF links. A small computer has small memory. Nowhere is there room for an operating system. The computers must be autonomous, with a self-contained ability to communicate. So communication must be simple—no elaborate protocol. Software must be compact and efficient. An ideal application for Forth.

Those systems requiring millions of lines of code will become irrelevant. They are a consequence of large, central computers. Distributed computation needs a different approach.

A language designed to support bulky, syntactical code encourages programmers to write big programs. They tend to take satisfaction, and be rewarded, for such. There is no pressure to seek compactness.

Although the code generated by a syntactic language might be small, it usually isn’t. To implement the generalities implied by the syntax leads to awkward, inefficient object code. This is unsuitable for a small computer. A well-designed language has a one-one correlation between source code and object code. It’s obvious to the programmer what code will be generated from his source. This provides its own satisfaction, is efficient, and reduces the need for documentation.

Forth was designed partly to be compact in both source and binary output, and is popular among embedded developers for that reason, but programmers in many other domains have reasons to choose other languages. Are there aspects of the language design that add only overhead to the source or the output?

Chuck: Forth is indeed compact. One reason is that it has little syntax.

Other languages seem to have deliberately added syntax, which provides redundancy and offers opportunity for syntax checking and thus error detection.

Forth provides little opportunity for error detection due to its lack of redundancy. This contributes to more compact source code.

My experience with other languages has been that most errors are in the syntax. Designers seem to create opportunity for programmer error that can be detected by the compiler. This does not seem productive. It just adds to the hassle of writing correct code.

An example of this is type checking. Assigning types to various numbers allows errors to be detected. An unintended consequence is that programmers must work to convert types, and sometimes work to evade type checking in order to do what they want.

Another consequence of syntax is that it must accommodate all intended applications. This makes it more elaborate. Forth is an extensible language. The programmer can create structures that are just as efficient as those provided by the compiler. So all capabilities do not have to be anticipated and provided for.

A characteristic of Forth is its use of postfix operators. This simplifies the compiler and offers a one-one translation of source code to object code. The programmer’s understanding of his code is enhanced and the resulting compiled code is more compact.

Proponents of many recent programming languages (notably Python and Ruby) cite readability as a key benefit. Is Forth easy to study and maintain in relation to those? What can Forth teach other programming languages in terms of readability?

Chuck: Computer languages all claim to be readable. They aren’t. Perhaps it seems so to one who knows the language, but a novice is always bewildered.

The problem is the arcane, arbitrary, and cryptic syntax. All the parentheses, ampersands, etc. You try to learn why it’s there and eventually conclude there’s no good reason. But you still have to follow the rules.

And you can’t speak the language. You’d have to pronounce the punctuation like Victor Borge.

Forth alleviates this problem by minimizing the syntax. Its cryptic symbols @ and ! are pronounced “fetch” and “store.” They are symbols because they occur so frequently.

The programmer is encouraged to use natural-language words. These are strung together without punctuation. With good choice of words, you can construct reasonable sentences. In fact, poems have been written in Forth.

Another advantage is postfix notation. A phrase like “6 inches” can apply the operator “inches” to the parameter 6, in a very natural manner. Quite readable.

On the other hand, the programmer’s job is to develop a vocabulary that describes the problem. This vocabulary can get to be quite large. A reader has to know it to find the program readable. And the programmer must work to define helpful words.

All in all, it takes effort to read a program. In any language.

How do you define success in terms of your work?

Chuck: An elegant solution.

One doesn’t write programs in Forth. Forth is the program. One adds words to construct a vocabulary that addresses the problem. It is obvious when the right words have been defined, for then you can interactively solve whatever aspect of the problem is relevant.

For example, I might define words that describe a circuit. I’ll want to add that circuit to a chip, display the layout, verify the design rules, run a simulation. The words that do these things form the application. If they are well chosen and provide a compact, efficient toolset, then I’ve been successful.

Where did you learn to write compilers? Was this something everybody at the time had to do?

Chuck: Well, I went to Stanford around ’60, and there was a group of grad students writing an ALGOL compiler—a version for the Burroughs 5500. It was only three or four of them, I think, but I was impressed out of my mind that three or four guys could sit down and write a compiler.

I sort of said, “Well, if they can do it, I can do it,” and I just did. It isn’t that hard. There was a mystique about compilers at the time.

There still is.

Chuck: Yeah, but less so. You get these new languages that pop up from time to time, and I don’t know if they’re interpreted or compiled, but well, hacker-type people are willing to do it anyway.

The operating system is another concept that is curious. Operating systems are dauntingly complex and totally unnecessary. It’s a brilliant thing that Bill Gates has done in selling the world on the notion of operating systems. It’s probably the greatest con game the world has ever seen.

An operating system does absolutely nothing for you. As long as you had something—a subroutine called disk driver, a subroutine called some kind of communication support, in the modern world, it doesn’t do anything else. In fact, Windows spends a lot of time with overlays and disk management all stuff like that which are irrelevant. You’ve got gigabyte disks; you’ve got megabyte RAMs. The world has changed in a way that renders the operating system unnecessary.

What about device support?

Chuck: You have a subroutine for each device. That’s a library, not an operating system. Call the ones you need or load the ones you need.

How do you resume programming after a short hiatus?

Chuck: I don’t find a short coding hiatus at all troublesome. I’m intensely focused on the problem and dream about it all night. I think that’s a characteristic of Forth: full effort over a short period of time (days) to solve a problem. It helps that Forth applications are naturally factored into subprojects. Most Forth code is simple and easy to reread. When I do really tricky things, I comment them well. Good comments help re-enter a problem, but it’s always necessary to read and understand the code.

What’s the biggest mistake you’ve made with regard to design or programming? What did you learn from it?

Chuck: Some 20 years ago I wanted to develop a tool to design VLSI chips. I didn’t have a Forth for my new PC, so I thought I’d try a different approach: machine language. Not assembler language, but actually typing the hex instructions.

I built up the code as I would in Forth, with many simple words that interacted hierarchically. It worked. I used it for 10 years. But it was difficult to maintain and document. Eventually I recoded it in Forth and it became smaller and simpler.

My conclusion was that Forth is more efficient than machine language. Partly because of its interactivity and partly because of its syntax. One nice aspect of Forth code is that numbers can be documented by the expression used to calculate them.

Hardware

How should people see the hardware they develop on: as a resource or as a limit? If you think of hardware as a resource, you might want to optimize the code and exploit every hardware feature; if you see it as a limit, you are probably going to write code with the idea that your code will run better on a new and more powerful version of the hardware, and that’s not a problem because hardware evolves rapidly.

Chuck: A very perceptive observation that software necessarily targets its hardware. Software for the PC certainly anticipates faster computers and can afford to be sloppy.

But for embedded systems, the software expects the system to be stable for the life of the project. And not a lot of software is migrated from one project to another. So here the hardware is a constraint, though not a limit. Whereas, for PCs, hardware is resource that will grow.

The move to parallel processing promises to change this. Applications that cannot exploit multiple computers will become limited as single computers stop getting faster. Rewriting legacy software to optimize parallel processing is impractical. And hoping that smart compilers will save the day is just wishful thinking.

What is the root of the concurrency problem?

Chuck: The root of the concurrency problem is speed. A computer must do many things in an application. These can be done on a single processor with multitasking. Or they can be done simultaneously with multiple processors.

The latter is much faster and contemporary software needs that speed.

Is the solution in hardware, software, or some combination?

Chuck: It’s not hard to glue multiple processors together. So the hardware exists. If software is programmed to take advantage of this the problem is solved. However, if the software can be reprogrammed, it can be made so efficient that multiprocessors are not needed. The problem is to use multiprocessors without changing legacy software. This is the intelligent compiler approach that has never been achieved.

I’m amazed that software written in the 1970s hasn’t/can’t be rewritten. One reason might be that in those days software was exciting; things being done for the first time; programmers working 18-hour days for the joy of it. Now programming is a 9–5 job as part of a team working to a schedule; not much fun.

So they add another layer of software to avoid rewriting the old software. At least that’s more fun than recoding a stupid word processor.

We have access to a big computational power in common computers, but how much actual computing (that is, calculating) are these systems doing? And how much are they just moving and formatting data?

Chuck: You are right. Most computer activity is moving data, not calculating. Not just moving data, but compressing, encrypting, scrambling. At high data rates, this must be done with circuitry so one wonders why a computer is needed at all.

Can we learn something from this? Should we build hardware in a different way?

Don Knuth launched a challenge: check what happens inside a computer during one second of time. He said that what we would discover could change a lot of things.

Chuck: My computer chips recognize this by having a simple, slow multiply. It isn’t used very often. Passing data between cores and accessing memory are the important features.

On one hand you have a language that really enables people to develop their own vocabularies and not necessarily think about the hardware presentation. On the other hand, you have a very small kernel that’s very much tied to that hardware. It’s interesting how Forth can bridge the gap between the two. On some of these machines, is it true that you have no operating system besides your Forth kernel?

Chuck: No, Forth is really standalone. Everything that needs to exist is in the kernel.

But it abstracts away that hardware for people who write programs in Forth.

Chuck: Right.

The Lisp Machine did something similar, but never really was popular. Forth quietly has done that job.

Chuck: Well, Lisp did not address I/O. In fact, C did not address I/O and because it didn’t, it needed an operating system. Forth addressed I/O from the very beginning. I don’t believe in the most common denominator. I think that if you go to a new machine, the only reason it’s a new machine is because it’s different in some way and you want to take advantage of those differences. So, you want to be there at the input-output level so you can do that.

Kernighan and Ritchie might argue for C that they wanted a least common factor to make porting easier. Yet you found it easier to port if you didn’t take that approach.

Chuck: I would have standard ways of doing that. I would have a word—I think it was fetchp maybe—that would fetch 8 bits from a port. That would be defined differently on different computers, but it would be the same function at the stack.

In one sense then, Forth is equivalent to C plus the standard I/O library.

Chuck: Yeah, but I worked with the Standard FORTRAN Library in the early days, and it was awful. It just had the wrong words. It was extremely expensive and bulky. It was so easy to define half a dozen instructions to perform in I/O operation that you didn’t need the overhead of a predefined protocol.

Did you find yourself working around that a lot?

Chuck: In FORTRAN, yeah. When you’re dealing with, say, Windows, there’s nothing you can do. They won’t let you have access to the I/O. I have stayed away from Windows most deliberately, but even without Windows, the Pentium was the most difficult machine to put Forth on.

It had too many instructions. And it had too many hardware features like the lookaside buffers and the different kinds of caching you really couldn’t ignore. You had to wade your way through, and the initialization code necessary to get Forth running was the most difficult and the most bulky.

Even if it only had to be executed once, I spent most of my time trying to figure out how to do it correctly. We had Forth running standalone on a Pentium, so it was worth the trouble.

The process extended over 10 years probably, partly chasing the changes in the hardware Intel was making.

You mentioned that Forth really supports asynchronous operation. In what sense do you mean asynchronous operation?

Chuck: Well, there’s several senses. Forth has always had a multiprogramming ability, a multithreading ability called Cooperative.

We had a word called pause. If you had a task and it came to a place where it didn’t have anything to do immediately, it would say pause. A round-robin scheduler would assign the computer to the next task in the loop.

If you didn’t say pause, you could monopolize the computer completely, but that would never be the case, because this was a dedicated computer. It was running a single application and all the tasks were friendly.

I guess that was in the old days when all of the tasks were friendly. That’s one kind of asynchronism that these tasks could run, do their own thing without ever having to synchronize. One of the features, again, of Forth is that that word pause could be buried in lower-level words. Every time you tried to read or write disk, the word pause would be executed for you, because the disk team knew that it was going to have to wait for the operation to complete.

In the new chips, the new multicore chips that I’m developing, we’re taking that same philosophy. Each computer is running independently and if you have a task on your computer, and another task on the neighbor, they’re both running simultaneously but they’re communicating with each other. That’s the equivalent of what the tasks would’ve been doing in a threaded computer.

Forth just factors very nicely into those independent tasks. In fact, in the case of the multicore computer, I can use not exactly the same programs, but I can factor the programs in the same way to make them run in parallel.

When you had the cooperative multithreading, did each thread of execution have its own stack, and you switched between them?

Chuck: When you did a task switch, sometimes all you needed to do, depending on the computer, was save the word on top of the stack and then switch the stack pointer. Sometimes you actually had to copy out the stack and load the new one, but in that case, I would make it a point to have a very shallow stack.

Did you deliberately limit the stack depth?

Chuck: Yes. Initially, the stacks were arbitrarily long. The first chip I designed had a stack that was 256 deep because I thought that was small. One of the chips I designed had a stack 4 deep. I’ve settled now on about 8 or 10 as a good stack depth, so my minimalism has gotten stricter over time.

I would’ve expected it to go the other way.

Chuck: Well, in my VLSI design application, I do have a case where I’m recursively following traces across the chip, in which case, I have to set the stack depths to about 4,000. To do that might require a different kind of stack, a software-implemented stack. But, in fact, on the Pentium it can be a hardware stack.

Application Design

You brought up the idea that Forth is an ideal language for many small computers networked together—smart dust, for example. For which kinds of applications do you think these small computers are the most appropriate?

Chuck: Communication certainly, sensing certainly. But I’m just beginning to learn how independent computers can cooperate to achieve a greater task.

The multicore computers we have are brutally small. They have 64 words of memory. Well, to put it differently, they have 128 words of memory: 64 RAM, 64 ROM. Each word can hold up to four instructions. You might end up with 512 instructions in a given computer, period, so the task has to be rather simple. Now how do you take a task like the TCP/IP stack and factor it amongst several of these computers in such a way that you can perform the operation without any computer needing more than 512 instructions? That’s a beautiful design problem, and one that I’m just approaching now.

I think that’s true of almost all applications. It’s much easier to do an application if it’s broken up into independent pieces as it is trying to do it in serial on a single processor. I think that’s true of video generation. Certainly I think it’s true of compressing and uncompressing images. But I’m just learning how to do that. We’ve got other people here in the company that are also learning and having a good time at it.

Is there any field of endeavor where this is not appropriate?

Chuck: Legacy software, certainly. I’m really worried about legacy software, but as soon as you’re willing to rethink a problem, I think it is more natural to think of it this way. I think it corresponds more closely to the way we think the brain works with Minsky’s independent agents. An agent to me is a small core. It may be that consciousness arises in the communication between these, not in the operation of any one of them.

Legacy software is an unappreciated but serious problem. It will only get worse—not only in banking but in aerospace and other technical industries. The problem is the millions of lines of code. Those could be recoded, say in thousands of lines of Forth. There’s no point in machine translation, which would only make the code bigger. But there’s no way that code could be validated. The cost and risk would be horrendous. Legacy code may be the downfall of our civilization.

It sounds like you’re betting that in the next 10 to 20 years we’ll see more and more software arise from the loose joining of many small parts.

Chuck: Oh, yes. I’m certain that’s the case. RF communication is so nice. They talk about micro agents inside your body that are fixing things and sensing things, and these agents can only communicate via RF or maybe acoustic.

They can’t do much. They’re only a few molecules. So this has got to be how the world goes. It’s the way our human society is organized. We have six and half billion independent agents out there cooperating.

Choosing words poorly can lead to poorly designed, poorly maintainable applications. Does building a larger application out of dozens or hundreds of small words lead to jargon? How do you avoid that?

Chuck: Well, you really can’t. I find myself picking words badly. If you do that, you can confuse yourself. I know in one application, I had this word—I forget what it was now—but I had defined and then I had modified it, and it ended up meaning the opposite of what it said.

It was like you had a word called right that makes things go to the left. That was hideously confusing. I fought it for a while and finally renamed the word because it was just impossible to understand the program with that word throwing so much noise into your cognition. I like to use English words, not abbreviations. I like to spell them out. On the other hand, I like them to be short. You run out of short meaningful English words after a while and you’ve got to do something else. I hate prefixes—a crude way to try to create namespaces so you can use the same old words over and over. They just look to me like a cop out. It’s an easy way to distinguish words, but you should’ve been smarter.

Very often Forth applications will have distinct vocabularies where you can reuse words. In this context, the word does this; in that context, it does something else. In the case of my VLSI design, all of this idealism failed. I needed at least a thousand words, and they’re not English words; they’re signal names or something, and I quickly had to revert to definitions and weirdly spelled words and prefixes and all of that stuff. It isn’t all that readable. But on the other hand, it’s full of words like nand and nor and xor for the various gates that are involved. Where possible, I use the words.

Now, I see other people writing Forth; I don’t want to pretend to be the only Forth programmer. Some of them do a very good job of coming up with names for things; others do a very bad job. Some come up with a very readable syntax, and others don’t think that that’s important. Some come up with very short definitions of words, and some have words that are a page long. There are no rules; there’s only stylistic conventions.

Also, the key difference between Forth and C and Prolog and ALGOL and FORTRAN, the conventional languages tried to anticipate all possible structures and syntax and build it into the language in the first place. That has led to some very clumsy languages. I think C is a clumsy language with its brackets and braces and colons and semicolons and all of that. Forth eliminated all of that.

I didn’t have to solve the general problem. I just had to provide a tool that someone else could use to solve whatever problem they encountered. The ability to do anything and not the ability to do everything.

Should microprocessors include source code so that they can be fixed even decades later?

Chuck: You’re right, including the source with microcomputers will document them nicely. Forth is compact, which facilitates that. But the next step is to include the compiler and editor so that the microcomputer code can be examined and changed without involving another computer/operating system that may have been lost. colorForth is my attempt to do that. A few K of source and/or object code is all that’s required. That can easily be stored on flash memory and be usable in the far future.

What is the link between the design of a language and the design of a software written with that language?

Chuck: A language determines its use. This is true of human-human languages. Witness the difference between Romance (French, Italian), Western (English, German, Russian) and Eastern (Arabic, Chinese) languages. They affect their cultures and their worldview. They affect what is said and how it’s said. Of these, English is particularly terse and increasingly popular.

So too with human-computer languages. The first languages (COBOL, FORTRAN) were too verbose. Later languages (Algol, C) had excessive syntax. These languages necessarily led to large, clumsy descriptions of algorithms. They could express anything, but do it badly.

Forth addresses these issues. It is relatively syntax-free. It encourages compact, efficient descriptions. It minimizes the need for comments, which tend to be inaccurate and distract attention from the code itself.

Forth also has a simple, efficient subroutine call. In C, a subroutine call requires expensive setup and recovery. This discourages its use. And encourages elaborate parameter sets that amortize the cost of the call, but lead to large, complex subroutines.

Efficiency allows Forth applications to be very highly factored, into many, small subroutines. And they typically are. My personal style is one-line definitions—hundreds of small subroutines. In such a case, the names assigned this code become important, both as a mnemonic device and as a way to achieve readability. Readable code requires less documentation.

The lack of syntax allows Forth a corresponding lack of discipline. This, to me, allows individual creativity and some very pleasant code. Others view it as a disadvantage, fearing management loss of control and lack of standardization. I think that’s more of a management failure than the fault of the language.

You said “Most errors are in syntax.” How do you avoid the other types of errors in Forth programs, such as logic errors, maintainability errors, and bad style decisions?

Chuck: Well, the major error in Forth has to do with stack management. Typically, you leave something on the stack inadvertently and it’ll trip you up later. We have a stack comment associated with words, which is very important. It tells you what is on the stack upon entry and what is on the stack upon exit. But that’s only a comment. You can’t trust it.

Some people did actually execute those and use them to do verification and stack behavior.

Basically, the solution is in the factoring. If you have a word whose definition is one line long, you can read through it thinking how the stack acts and conclude at the end that it’s correct. You can test it and see if it works the way you thought it did, but even so, you’re going to get caught up in stack errors. The words dup and drop are ubiquitous and have to be used correctly. The ability to execute words out of context just by putting their input parameters and looking at their output parameters is hugely important. Again, when you’re working bottom-up, you know that all of the words you’ve already defined work correctly because you tested them.

Also, there are only a few conditionals in Forth. There’s an if-else-then construction, a begin-while construct. My philosophy, which I regularly try to teach, is that you minimize the number of conditionals in your program. Rather than having a word that tests something and either does this or that, you have two words: one that does this and one that does that, and you use the right one.

Now it doesn’t work in C because the calling sequences are so expensive that they tend to have parameters that let the same routine do different things based upon the way it’s called. That’s what leads to all of the bugs and complications in legacy software.

In trying to work around deficiencies of the implementation?

Chuck: Yeah. Loops are unavoidable. Loops can be very, very nice. But a Forth loop, at least a colorForth loop, is a very simple one with a single entry and a single exit.

What advice would you give a novice to make programming more pleasant and effective?

Chuck: Well, surely not to your surprise, I would say you should learn to write Forth code. Even if you aren’t going to be writing Forth code professionally, exposure to it will teach you some of these lessons and give you a better perspective on whatever language you use. If I were writing a C program, I have written almost none, but I would write it in the style of Forth with a lot of simple subroutines. Even if there were a cost involved there, I think it would be worth it in maintainability.

The other thing is keep it simple. The inevitable trend in designing an aircraft or in writing an application, even a word processor, is to add features and add features and add features until the cost becomes unsupportable. It would be better to have half a dozen word processors that would focus on different markets. Using Word to compose an email is silly; 99% of all of the facilities available are unnecessary. You ought to have an email editor. There used to be such, but the trend seems to be away from that. It’s not clear to me why.

Keep it simple. If you’re encountering an application, if you’re on part of a design team, try to persuade other people to keep it simple. Don’t anticipate. Don’t solve a problem that you think might occur in the future. Solve the problem you’ve got. Anticipating is very inefficient. You can anticipate 10 things happening, of which only one will, so you’ve wasted a lot of effort.

How do you recognize simplicity?

Chuck: There’s I think a budding science of complexity, and one of their tenets is how to measure complexity. The description that I like, and I don’t know if there’s any other one, is that the shortest description or if you have two concepts, the one with the shorter description is the simpler. If you can come up with a shorter definition of something, you come up with a simpler definition.

But that fails in a subtle way that any kind of description depends on the context. If you can write a very short subroutine in C, you might say this is very simple, but you’re relying upon the existence of the C compiler and the operating system and the computer that’s going to execute it all. So really, you don’t have a simple thing; you have a pretty complex thing when you consider the wider context.

I think it’s like beauty. You can’t define it, but you can recognize it when you see it—simple is small.

How does teamwork affect programming?

Chuck: Teamwork—much overrated. The first job of a team is to partition the problem into relatively independent parts. Assign each part to an individual. The team leader is responsible for seeing that the parts come together.

Sometimes two people can work together. Talking about a problem can clarify it. But too much communication becomes an end in itself. Group thinking does not facilitate creativity. And when several people work together, inevitably one does the work.

Is this valid for every type of project? If you have to write something as feature-rich as OpenOffice.org…it sounds pretty complex, no?

Chuck: Something like OpenOffice.org would be factored into subprojects, each programmed by an individual with enough communication to assure compatibility.

How do you recognize a good programmer?

Chuck: A good programmer writes good code quickly. Good code is correct, compact, and readable. “Quickly” means hours to days.

A bad programmer will want to talk about the problem, will waste time planning instead of writing, and will make a career out of writing and debugging the code.

What is your opinion of compilers? Do you think they mask the real skills of programmers?

Chuck: Compilers are probably the worst code ever written. They are written by someone who has never written a compiler before and will never do so again.

The more elaborate the language, the more complex, bug-ridden, and unusable is the compiler. But a simple compiler for a simple language is an essential tool—if only for documentation.

More important than the compiler is the editor. The wide variety of editors allows each programmer to select his own, to the great detriment of collaborative efforts. This fosters the cottage industry of translating from one to another.

Another failing of compiler writers is the compulsion to use every special character on the keyboard. Thus keyboards can never become smaller and simpler. And source code becomes impenetrable.

But the skills of a programmer are independent of these tools. He can quickly master their foibles and produce good code.

How should software be documented?

Chuck: I value comments much less than others do. Several reasons:

• If comments are terse, they are often cryptic. Then you have to guess what they mean.
• If comments are verbose, they overwhelm the code they’re embedded in and trying to explain. It’s hard to find and relate code to comment.
• Comments are often badly written. Programmers aren’t known for their literary skills, especially if English is not their native language. Jargon and grammatical errors often make them unreadable.
• Most importantly, comments are often inaccurate. Code may change without comments being updated. Although code may be critically reviewed, comments rarely are. An inaccurate comment causes more trouble than no comment. The reader must judge whether the comment or the code is correct.

Comments are often misguided. They should explain the purpose of the code, not the code itself. To paraphrase the code is unhelpful. And if it is inaccurate, downright misleading. Comments should explain why the code is present, what it is intended to accomplish, and any tricks employed in accomplishing it.

colorForth factors comments into a shadow block. This removes them from the code itself, making that code more readable. Yet they are instantly available for reading or updating. It also limits the size of comments to the size of the code.

Comments do not substitute for proper documentation. A document must be written that explains in prose the code module of interest. It should expand greatly the comments and concentrate on literate and complete explanation.

Of course, this is rarely done, is often unaffordable, and is easily lost since it is separate from the code.

Quoting from http://www.colorforth.com/HOPL.html :

“The issue of patenting Forth was discussed at length. But since software patents were controversial and might involve the Supreme Court, NRAO declined to pursue the matter. Whereupon, rights reverted to me. I don’t think ideas should be patentable. Hindsight agrees that Forth’s only chance lay in the public domain. Where it has flourished.”

Software patents are still controversial today. Is your opinion about patents still the same?

Chuck: I’ve never been in favor of software patents. It’s too much like patenting an idea. And patenting a language/protocol is especially disturbing. A language will only be successful if it’s used. Anything that discourages use is foolish.

Do you think that patenting a technology prevents or limits its diffusion?

Chuck: It is difficult to market software, which is easy to copy. Companies go to great lengths to protect their product, sometimes making it unusable in the process. My answer to that problem is to sell hardware and give away the software. Hardware is difficult to copy and becomes more valuable as software is developed for it.

Patents are one way of addressing these issues. They have proven a wonderful boon to innovation. But there’s a delicate balance required to discourage frivolous patents and maintain consistency with prior art/patents. And there are huge costs associated with granting and enforcing them. Recent proposals to reform patent law threaten to freeze out the individual inventor in favor of large companies. Which would be tragic.