My Lisp Experiences and the Development of GNU Emacs

Transcript of Richard Stallman's speech at the International Lisp Conference, 28 Oct 2002.


Since none of my usual speeches have anything to do with Lisp, none of them were appropriate for today. So I'm going to have to wing it. Since I've done enough things in my career connected with Lisp I should be able to say something interesting.

My first experience with Lisp was when I read the Lisp 1.5 manual in high school. That's when I had my mind blown by the idea that there could be a computer language like that. The first time I had a chance to do anything with Lisp was when I was a freshman at Harvard and I wrote a Lisp interpreter for the PDP-11. It was a very small machine—it had something like 8k of memory—and I managed to write the interpreter in a thousand instructions. This gave me some room for a little bit of data. That was before I got to see what real software was like, that did real system jobs.

I began doing work on a real Lisp implementation with JonL White once I started working at MIT. I got hired at the Artificial Intelligence Lab not by JonL, but by Russ Noftsker, which was most ironic considering what was to come—he must have really regretted that day.

During the 1970s, before my life became politicized by horrible events, I was just going along making one extension after another for various programs, and most of them did not have anything to do with Lisp. But, along the way, I wrote a text editor, Emacs. The interesting idea about Emacs was that it had a programming language, and the user's editing commands would be written in that interpreted programming language, so that you could load new commands into your editor while you were editing. You could edit the programs you were using and then go on editing with them. So, we had a system that was useful for things other than programming, and yet you could program it while you were using it. I don't know if it was the first one of those, but it certainly was the first editor like that.

This spirit of building up gigantic, complicated programs to use in your own editing, and then exchanging them with other people, fueled the spirit of free-wheeling cooperation that we had at the AI Lab then. The idea was that you could give a copy of any program you had to someone who wanted a copy of it. We shared programs to whomever wanted to use them, they were human knowledge. So even though there was no organized political thought relating the way we shared software to the design of Emacs, I'm convinced that there was a connection between them, an unconscious connection perhaps. I think that it's the nature of the way we lived at the AI Lab that led to Emacs and made it what it was.

The original Emacs did not have Lisp in it. The lower level language, the non-interpreted language—was PDP-10 Assembler. The interpreter we wrote in that actually wasn't written for Emacs, it was written for TECO. It was our text editor, and was an extremely ugly programming language, as ugly as could possibly be. The reason was that it wasn't designed to be a programming language, it was designed to be an editor and command language. There were commands like 5l, meaning move five lines, or i and then a string and then an ESC to insert that string. You would type a string that was a series of commands, which was called a command string. You would end it with ESC ESC, and it would get executed.

Well, people wanted to extend this language with programming facilities, so they added some. For instance, one of the first was a looping construct, which was < >. You would put those around things and it would loop. There were other cryptic commands that could be used to conditionally exit the loop. To make Emacs, we [1] added facilities to have subroutines with names. Before that, it was sort of like Basic, and the subroutines could only have single letters as their names. That was hard to program big programs with, so we added code so they could have longer names. Actually, there were some rather sophisticated facilities; I think that Lisp got its unwind-protect facility from TECO.

We started putting in rather sophisticated facilities, all with the ugliest syntax you could ever think of, and it worked—people were able to write large programs in it anyway. The obvious lesson was that a language like TECO, which wasn't designed to be a programming language, was the wrong way to go. The language that you build your extensions on shouldn't be thought of as a programming language in afterthought; it should be designed as a programming language. In fact, we discovered that the best programming language for that purpose was Lisp.

It was Bernie Greenberg, who discovered that it was [2]. He wrote a version of Emacs in Multics MacLisp, and he wrote his commands in MacLisp in a straightforward fashion. The editor itself was written entirely in Lisp. Multics Emacs proved to be a great success—programming new editing commands was so convenient that even the secretaries in his office started learning how to use it. They used a manual someone had written which showed how to extend Emacs, but didn't say it was a programming. So the secretaries, who believed they couldn't do programming, weren't scared off. They read the manual, discovered they could do useful things and they learned to program.

So Bernie saw that an application—a program that does something useful for you—which has Lisp inside it and which you could extend by rewriting the Lisp programs, is actually a very good way for people to learn programming. It gives them a chance to write small programs that are useful for them, which in most arenas you can't possibly do. They can get encouragement for their own practical use—at the stage where it's the hardest—where they don't believe they can program, until they get to the point where they are programmers.

At that point, people began to wonder how they could get something like this on a platform where they didn't have full service Lisp implementation. Multics MacLisp had a compiler as well as an interpreter—it was a full-fledged Lisp system—but people wanted to implement something like that on other systems where they had not already written a Lisp compiler. Well, if you didn't have the Lisp compiler you couldn't write the whole editor in Lisp—it would be too slow, especially redisplay, if it had to run interpreted Lisp. So we developed a hybrid technique. The idea was to write a Lisp interpreter and the lower level parts of the editor together, so that parts of the editor were built-in Lisp facilities. Those would be whatever parts we felt we had to optimize. This was a technique that we had already consciously practiced in the original Emacs, because there were certain fairly high level features which we re-implemented in machine language, making them into TECO primitives. For instance, there was a TECO primitive to fill a paragraph (actually, to do most of the work of filling a paragraph, because some of the less time-consuming parts of the job would be done at the higher level by a TECO program). You could do the whole job by writing a TECO program, but that was too slow, so we optimized it by putting part of it in machine language. We used the same idea here (in the hybrid technique), that most of the editor would be written in Lisp, but certain parts of it that had to run particularly fast would be written at a lower level.

Therefore, when I wrote my second implementation of Emacs, I followed the same kind of design. The low level language was not machine language anymore, it was C. C was a good, efficient language for portable programs to run in a Unix-like operating system. There was a Lisp interpreter, but I implemented facilities for special purpose editing jobs directly in C—manipulating editor buffers, inserting leading text, reading and writing files, redisplaying the buffer on the screen, managing editor windows.

Now, this was not the first Emacs that was written in C and ran on Unix. The first was written by James Gosling, and was referred to as GosMacs. A strange thing happened with him. In the beginning, he seemed to be influenced by the same spirit of sharing and cooperation of the original Emacs. I first released the original Emacs to people at MIT. Someone wanted to port it to run on Twenex—it originally only ran on the Incompatible Timesharing System we used at MIT. They ported it to Twenex, which meant that there were a few hundred installations around the world that could potentially use it. We started distributing it to them, with the rule that “you had to send back all of your improvements” so we could all benefit. No one ever tried to enforce that, but as far as I know people did cooperate.

Gosling did, at first, seem to participate in this spirit. He wrote in a manual that he called the program Emacs hoping that others in the community would improve it until it was worthy of that name. That's the right approach to take towards a community—to ask them to join in and make the program better. But after that he seemed to change the spirit, and sold it to a company.

At that time I was working on the GNU system (a free software Unix-like operating system that many people erroneously call “Linux”). There was no free software Emacs editor that ran on Unix. I did, however, have a friend who had participated in developing Gosling's Emacs. Gosling had given him, by email, permission to distribute his own version. He proposed to me that I use that version. Then I discovered that Gosling's Emacs did not have a real Lisp. It had a programming language that was known as “mocklisp,” which looks syntactically like Lisp, but didn't have the data structures of Lisp. So programs were not data, and vital elements of Lisp were missing. Its data structures were strings, numbers and a few other specialized things.

I concluded I couldn't use it and had to replace it all, the first step of which was to write an actual Lisp interpreter. I gradually adapted every part of the editor based on real Lisp data structures, rather than ad hoc data structures, making the data structures of the internals of the editor exposable and manipulable by the user's Lisp programs.

The one exception was redisplay. For a long time, redisplay was sort of an alternate world. The editor would enter the world of redisplay and things would go on with very special data structures that were not safe for garbage collection, not safe for interruption, and you couldn't run any Lisp programs during that. We've changed that since—it's now possible to run Lisp code during redisplay. It's quite a convenient thing.

This second Emacs program was “free software” in the modern sense of the term—it was part of an explicit political campaign to make software free. The essence of this campaign was that everybody should be free to do the things we did in the old days at MIT, working together on software and working with whomever wanted to work with us. That is the basis for the free software movement—the experience I had, the life that I've lived at the MIT AI lab—to be working on human knowledge, and not be standing in the way of anybody's further using and further disseminating human knowledge.

At the time, you could make a computer that was about the same price range as other computers that weren't meant for Lisp, except that it would run Lisp much faster than they would, and with full type checking in every operation as well. Ordinary computers typically forced you to choose between execution speed and good typechecking. So yes, you could have a Lisp compiler and run your programs fast, but when they tried to take car of a number, it got nonsensical results and eventually crashed at some point.

The Lisp machine was able to execute instructions about as fast as those other machines, but each instruction—a car instruction would do data typechecking—so when you tried to get the car of a number in a compiled program, it would give you an immediate error. We built the machine and had a Lisp operating system for it. It was written almost entirely in Lisp, the only exceptions being parts written in the microcode. People became interested in manufacturing them, which meant they should start a company.

There were two different ideas about what this company should be like. Greenblatt wanted to start what he called a “hacker” company. This meant it would be a company run by hackers and would operate in a way conducive to hackers. Another goal was to maintain the AI Lab culture [3]. Unfortunately, Greenblatt didn't have any business experience, so other people in the Lisp machine group said they doubted whether he could succeed. They thought that his plan to avoid outside investment wouldn't work.

Why did he want to avoid outside investment? Because when a company has outside investors, they take control and they don't let you have any scruples. And eventually, if you have any scruples, they also replace you as the manager.

So Greenblatt had the idea that he would find a customer who would pay in advance to buy the parts. They would build machines and deliver them; with profits from those parts, they would then be able to buy parts for a few more machines, sell those and then buy parts for a larger number of machines, and so on. The other people in the group thought that this couldn't possibly work.

Greenblatt then recruited Russell Noftsker, the man who had hired me, who had subsequently left the AI Lab and created a successful company. Russell was believed to have an aptitude for business. He demonstrated this aptitude for business by saying to the other people in the group, “Let's ditch Greenblatt, forget his ideas, and we'll make another company.” Stabbing in the back, clearly a real businessman. Those people decided they would form a company called Symbolics. They would get outside investment, not have scruples, and do everything possible to win.

But Greenblatt didn't give up. He and the few people loyal to him decided to start Lisp Machines Inc. anyway and go ahead with their plans. And what do you know, they succeeded! They got the first customer and were paid in advance. They built machines and sold them, and built more machines and more machines. They actually succeeded even though they didn't have the help of most of the people in the group. Symbolics also got off to a successful start, so you had two competing Lisp machine companies. When Symbolics saw that LMI was not going to fall flat on its face, they started looking for ways to destroy it.

Thus, the abandonment of our lab was followed by “war” in our lab. The abandonment happened when Symbolics hired away all the hackers, except me and the few who worked at LMI part-time. Then they invoked a rule and eliminated people who worked part-time for MIT, so they had to leave entirely, which left only me. The AI lab was now helpless. And MIT had made a very foolish arrangement with these two companies. It was a three-way contract where both companies licensed the use of Lisp machine system sources. These companies were required to let MIT use their changes. But it didn't say in the contract that MIT was entitled to put them into the MIT Lisp machine systems that both companies had licensed. Nobody had envisioned that the AI lab's hacker group would be wiped out, but it was.

So Symbolics came up with a plan [4]. They said to the lab, “We will continue making our changes to the system available for you to use, but you can't put it into the MIT Lisp machine system. Instead, we'll give you access to Symbolics' Lisp machine system, and you can run it, but that's all you can do.”

This, in effect, meant that they demanded that we had to choose a side, and use either the MIT version of the system or the Symbolics version. Whichever choice we made determined which system our improvements went to. If we worked on and improved the Symbolics version, we would be supporting Symbolics alone. If we used and improved the MIT version of the system, we would be doing work available to both companies, but Symbolics saw that we would be supporting LMI because we would be helping them continue to exist. So we were not allowed to be neutral anymore.

Up until that point, I hadn't taken the side of either company, although it made me miserable to see what had happened to our community and the software. But now, Symbolics had forced the issue. So, in an effort to help keep Lisp Machines Inc. going [5]—I began duplicating all of the improvements Symbolics had made to the Lisp machine system. I wrote the equivalent improvements again myself (i.e., the code was my own).

After a while [6], I came to the conclusion that it would be best if I didn't even look at their code. When they made a beta announcement that gave the release notes, I would see what the features were and then implement them. By the time they had a real release, I did too.

In this way, for two years, I prevented them from wiping out Lisp Machines Incorporated, and the two companies went on. But, I didn't want to spend years and years punishing someone, just thwarting an evil deed. I figured they had been punished pretty thoroughly because they were stuck with competition that was not leaving or going to disappear [7]. Meanwhile, it was time to start building a new community to replace the one that their actions and others had wiped out.

The Lisp community in the 70s was not limited to the MIT AI Lab, and the hackers were not all at MIT. The war that Symbolics started was what wiped out MIT, but there were other events going on then. There were people giving up on cooperation, and together this wiped out the community and there wasn't much left.

Once I stopped punishing Symbolics, I had to figure out what to do next. I had to make a free operating system, that was clear—the only way that people could work together and share was with a free operating system.

At first, I thought of making a Lisp-based system, but I realized that wouldn't be a good idea technically. To have something like the Lisp machine system, you needed special purpose microcode. That's what made it possible to run programs as fast as other computers would run their programs and still get the benefit of typechecking. Without that, you would be reduced to something like the Lisp compilers for other machines. The programs would be faster, but unstable. Now that's okay if you're running one program on a timesharing system—if one program crashes, that's not a disaster, that's something your program occasionally does. But that didn't make it good for writing the operating system in, so I rejected the idea of making a system like the Lisp machine.

I decided instead to make a Unix-like operating system that would have Lisp implementations to run as user programs. The kernel wouldn't be written in Lisp, but we'd have Lisp. So the development of that operating system, the GNU operating system, is what led me to write the GNU Emacs. In doing this, I aimed to make the absolute minimal possible Lisp implementation. The size of the programs was a tremendous concern.

There were people in those days, in 1985, who had one-megabyte machines without virtual memory. They wanted to be able to use GNU Emacs. This meant I had to keep the program as small as possible.

For instance, at the time the only looping construct was while, which was extremely simple. There was no way to break out of the while statement, you just had to do a catch and a throw, or test a variable that ran the loop. That shows how far I was pushing to keep things small. We didn't have caar and cadr and so on; “squeeze out everything possible” was the spirit of GNU Emacs, the spirit of Emacs Lisp, from the beginning.

Obviously, machines are bigger now, and we don't do it that way any more. We put in caar and cadr and so on, and we might put in another looping construct one of these days. We're willing to extend it some now, but we don't want to extend it to the level of common Lisp. I implemented Common Lisp once on the Lisp machine, and I'm not all that happy with it. One thing I don't like terribly much is keyword arguments [8]. They don't seem quite Lispy to me; I'll do it sometimes but I minimize the times when I do that.

That was not the end of the GNU projects involved with Lisp. Later on around 1995, we were looking into starting a graphical desktop project. It was clear that for the programs on the desktop, we wanted a programming language to write a lot of it in to make it easily extensible, like the editor. The question was what it should be.

At the time, TCL was being pushed heavily for this purpose. I had a very low opinion of TCL, basically because it wasn't Lisp. It looks a tiny bit like Lisp, but semantically it isn't, and it's not as clean. Then someone showed me an ad where Sun was trying to hire somebody to work on TCL to make it the “de-facto standard extension language” of the world. And I thought, “We've got to stop that from happening.” So we started to make Scheme the standard extensibility language for GNU. Not Common Lisp, because it was too large. The idea was that we would have a Scheme interpreter designed to be linked into applications in the same way TCL was linked into applications. We would then recommend that as the preferred extensibility package for all GNU programs.

There's an interesting benefit you can get from using such a powerful language as a version of Lisp as your primary extensibility language. You can implement other languages by translating them into your primary language. If your primary language is TCL, you can't very easily implement Lisp by translating it into TCL. But if your primary language is Lisp, it's not that hard to implement other things by translating them. Our idea was that if each extensible application supported Scheme, you could write an implementation of TCL or Python or Perl in Scheme that translates that program into Scheme. Then you could load that into any application and customize it in your favorite language and it would work with other customizations as well.

As long as the extensibility languages are weak, the users have to use only the language you provided them. Which means that people who love any given language have to compete for the choice of the developers of applications—saying “Please, application developer, put my language into your application, not his language.” Then the users get no choices at all—whichever application they're using comes with one language and they're stuck with [that language]. But when you have a powerful language that can implement others by translating into it, then you give the user a choice of language and we don't have to have a language war anymore. That's what we're hoping Guile, our scheme interpreter, will do. We had a person working last summer finishing up a translator from Python to Scheme. I don't know if it's entirely finished yet, but for anyone interested in this project, please get in touch. So that's the plan we have for the future.

I haven't been speaking about free software, but let me briefly tell you a little bit about what that means. Free software does not refer to price; it doesn't mean that you get it for free. (You may have paid for a copy, or gotten a copy gratis.) It means that you have freedom as a user. The crucial thing is that you are free to run the program, free to study what it does, free to change it to suit your needs, free to redistribute the copies of others and free to publish improved, extended versions. This is what free software means. If you are using a nonfree program, you have lost crucial freedom, so don't ever do that.

The purpose of the GNU project is to make it easier for people to reject freedom-trampling, user-dominating, nonfree software by providing free software to replace it. For those who don't have the moral courage to reject the nonfree software, when that means some practical inconvenience, what we try to do is give a free alternative so that you can move to freedom with less of a mess and less of a sacrifice in practical terms. The less sacrifice the better. We want to make it easier for you to live in freedom, to cooperate.

This is a matter of the freedom to cooperate. We're used to thinking of freedom and cooperation with society as if they are opposites. But here they're on the same side. With free software you are free to cooperate with other people as well as free to help yourself. With nonfree software, somebody is dominating you and keeping people divided. You're not allowed to share with them, you're not free to cooperate or help society, anymore than you're free to help yourself. Divided and helpless is the state of users using nonfree software.

We've produced a tremendous range of free software. We've done what people said we could never do; we have two operating systems of free software. We have many applications and we obviously have a lot farther to go. So we need your help. I would like to ask you to volunteer for the GNU project; help us develop free software for more jobs. Take a look at gnu.org/help to find suggestions for how to help. If you want to order things, there's a link to that from the home page. If you want to read about philosophical issues, look in /philosophy. If you're looking for free software to use, look in /directory, which lists about 1900 packages now (which is a fraction of all the free software out there). Please write more and contribute to us. My book of essays, “Free Software and Free Society,” is on sale and can be purchased at www.gnu.org [9]. Happy hacking!

Footnotes

  1. Guy Steele designed the original symmetrical Emacs command set; then he and I began implementing Emacs (on top of TECO), but after one long joint development session, Steele began drifting away, so I finished Emacs. Others particularly including Eugene C. Cicciarelli and Mike McMahon contributed substantially later on.
  2. Bernie Greenberg says that Dan Weinreb's implementation of Emacs for the Lisp Machine came before Greenberg's implementation for Multics. I apologize for the mistake.
  3. Greenblatt's plan, as I understood it, was to hire lab people part time, so that they could continue working at the AI Lab. Symbolics hired them full time instead, so they stopped working at MIT.
  4. The background of this plan, which I did not state explicitly in the talk, is that during an initial period the ex-AI-Lab hackers, whether at Symbolics or LMI, continued contributing their changes to the MIT Lisp Machine system—even though the contract did not require this. Symbolics' plan was to rupture this cooperation unilaterally.
  5. It was not that I cared particularly about the fate of LMI, but rather I did not want to let Symbolics gain through its aggression against the AI Lab.
  6. This statement has been misconstrued as saying that I never, ever looked at Symbolics' code. Actually it says I did look, at first. The Symbolics source code was available at MIT, where I was entitled to read it, and at first that's how I found out about their changes.

    But that meant I had to make a special effort to solve each problem differently, in order to avoid copying Symbolics code. After a while, I concluded it was better not to even look. That way I could write code in whatever way was best, without concern for what might be in Symbolics' code.

  7. Symbolics at one point protested to MIT that my work, by thwarting their plan, had cost Symbolics a million dollars.
  8. I don't mind if a very complex and heavyweight function takes keyword arguments. What bothers me is making simple basic functions such as “member” use them.
  9. In 2021, this book can be purchased from GNU Press.