┌───────────────────────┐ ▄▄▄▄▄ ▄▄▄▄▄ ▄▄▄▄▄ │ │ █ █ █ █ █ █ │ │ █ █ █ █ █▀▀▀▀ │ │ █ █ █ █ ▄ │ │ ▄▄▄▄▄ │ │ █ █ │ │ █ █ │ │ █▄▄▄█ │ │ ▄ ▄ │ │ █ █ │ │ █ █ │ │ █▄▄▄█ │ │ ▄▄▄▄▄ │ │ █ │ Interview: Doug McIlroy │ █ │ ~ tmp.0ut Staff └───────────────────█ ──┘ In October 2025, tmp.0ut Staff got the opportunity to talk with Doug McIlroy. Born in 1932, Doug is a computing pioneer who was on the ground floor for the creation of Unix and C, and contributed to many of the core utilities and languages that we use daily. His design approach and methodology is based a foundation of meticulous study of code and it's documentation. From the get go, Doug has been dedicated to understanding the machine on a fundamental level. Across multiple generations of computing, Doug has been thinking about what the next step is to elevate the domain as a whole, while doing more with less. Doug is attributed with the quote "The real hero of programming is the one who writes negative code". tmp.0ut staff and others were invited to Dartmouth College to do a week-long series of lectures called "Binary Fun Week", focusing on everything from file formats, malware, reverse engineering, exploit development, and Binary Golf. Our interview was conducted in the Dartmouth CS lab with several speakers present, as well as some students, who also asked questions of Doug and are included in this interview. ┌──────────────────────────────────────────────────────────────────────────────┘ ▄▄▄▄▄ █ [ So you wrote echo? ] █▀▀▀▀ █▄▄▄▄ ▄▄▄▄▄ That was the first C program I ever wrote. █ █ █▀▀▀█ █ █ [ o.0 wow ] ▄▄▄▄▄ █ █ █▀▀█▀ And little did I know that it was going to become a well-known program. █ █ ▄ █ [ echo was your program? ] █ █▄▄▄▄ ▄ ▄ First C program. When C was just invented. █ █ ▀▀▀▀█ ▄▄▄▄█ [ What year was this? ] <<<<< <<<<< <<<<< About '72. <<<<< ▄ ▄ █ █ [ There's a long list of programs you wrote, diff, sort. You wrote sort? ] █ █ █▄▄▄█ ▄▄ ▄ I didn't write the first one, but I put in most of the options for sort. █▀▄ █ Although I'm opposed to lots of options in programs like ls, which is █ ▀▄█ run through both upper and lower case flags. █ ▀█ ▄▄▄▄▄ █ sort was originally written by Ken Thompson. It had an amusing remark in █ the manual "wide options are available" but it didn't say what those ▄▄█▄▄ options were. ▄ ▄ █ █ ▄▀▀▀▄ I asked Ken, "well, what are they?" and he said, "well, all you have to █ █ do is go put them in the source." So, I did. >editor's note: $ ls -transubstantiationalist [ What was it like working with Ken Thompson and Dennis Ritchie? ] They were great. Ken is certainly one of the greatest programming geniuses I've known. He's just incredible. I mean, not only did he do Unix, he did the first chess machine to become master rated. His architecture for the chess machine was then imitated by Deep Blue and put in silicon. [ IBM. It was Gary Kasparov who got beat by Deep Blue, right? ] Yeah. At IBM. [ So Ken Thompson wrote one of the first chess programs ] Yep, and made a machine with which he had some grand adventures. He got invited to take the machine to Russia by Botvinnik, who was the world champion or a past world champion. [ Do you keep up with chess news? ] No. └──────────────────────────────────────────────────────────────────────────────┐ ▄▄▄▄▄ [ What was your education background before Bell Labs? ] █ █▀▀▀▀ █▄▄▄▄ I was an undergraduate in engineering physics at Cornell. This was ▄▄▄▄ back in the day, when Cornell thought it was going to set a new trend, █ █ where undergraduate engineers had a five-year curriculum. That lasted █ █ about a decade, I guess. Cornell realized that other colleges weren't █▄▄▄█ going along with it. If you could go to MIT for four years, you'd get ▄ ▄ a cheaper degree. █ █ █ █ █▄▄▄█ [ And you did go to MIT as well, didn't you? ] ▄▄▄▄▄ █ ▀ █ I did go on to MIT because they had the Whirlwind 2 computer, which █▄▄▄█ was a gigantic machine. ▄▄▄▄▄ █ █ █▀▀▀█ [ The Whirlwind 2? ] █ █ ▄▄▄▄▄ █ Whirlwind 2[1]. It was built in 1952. It was built there in 1954. The █ machine occupied three stories of a building. The basement was rotating █ equipment to convert AC to DC. The first floor was electronic power ▄▄▄▄▄ supplies. The second floor was the computer itself. You could walk █ inside the mainframe. There were 16 bits of the accumulator, each of █ which was an eight-foot-high relay rack. ▄▄█▄▄ ▄▄▄▄▄ █ █ [ Each bit was an eight-foot-high relay rack? ] █ █ █▄▄▄█ ▄▄ ▄ Yes. Then at the far end was one of its two great innovations, core █▀▄ █ memory. That was just one relay rack. The memory itself was about this █ ▀▄█ big. **holds hands up wide** █ ▀█ It was absolutely gigantic. The one at Cornell, the CQC, had 24 words of memory. It was not stored programming. The little ones had a thousand 16 bit words, 2K. The register was eight feet tall. Each one of those bits in the accumulator. In the CPU. [ So you said each register is eight feet tall? That's nuts man. ] Yeah. Jay Forrester, who was the project manager, told me once that he thought his greatest invention was marginal testing. [ Marginal testing? ] Do you know what that is? [ I don't. ] So every morning before work began, they would take the machine down for an hour with preventative maintenance. And during that hour, they would run the voltages up and down. The idea was that if you were going to break something, you broke it during testing, rather than during the day. [ Okay, so almost like stress testing. ] Yes. I may be getting my machines mixed up. I saw the Bell Labs relay computer but I never used it. [ When you were studying at MIT, you did a math degree? ] I did a math degree. [ It sounds like you did a pretty interesting dissertation, conical shells? ] Really boring. [ It sounds interesting. Conical shells? Like, nautilus shells? ] Yes. Just... You've got a sheet metal cone and you bend it, how does it bend? That's... Very classical... [ I'm sure you're oversimplifying it. ] So the notion was, it gets you a degree. Although, I was there because of the computer. Really early on, maybe 1948 or 49, von Neumann and Goldstein wrote this paper[2] about how many digits you had to carry to do matrix inversion. Their estimates were very pessimistic. To do a 10x10, you probably needed 40 decimal digits. Everybody at the time was inverting 10x10s with 8 digits. If von Neumann and Goldstein couldn't make a good estimate, what chance was there for me? So I didn't do it. It didn't work in numerical analysis. It was in 1979 that Wilkinson in England invented backwards error analysis. Finally, people began to be able to really get good estimates on the accuracy of numerical computations. I was too early for that. ┌──────────────────────────────────────────────────────────────────────────────┘ ▄▄▄▄ █ █ [ How did you get to Bell Labs? ] █▀▀▀▄ █▄▄▄▀ ▄▄▄▄▄ I had two summer jobs before 1953 and 1957, I guess, before joining the █ labs in 58. When the recruiters showed up at MIT, I was well primed and █▀▀▀▀ eager to join. █▄▄▄▄ ▄ █ [ What was the path that led you to working on Multics and Unix? ] █ █▄▄▄▄ ▄ Oh, the usual thing is if you arrive at Bell Labs for a PhD, you're left █ to find out what you're going to do. At first, I played with integer █ linear programming and proved an incorrect theorem. Gomory at IBM proved █▄▄▄▄ the correct one, and called me one day to tell me about it. <<<<< <<<<< <<<<< Then I worked on macros. Macro-processing became very popular. In fact, <<<<< macros became such a thing in Bell Labs that the folks making switching ▄ systems invented a programming language, which was just a set of macros. █ They used this very heavily. █ █▄▄▄▄ ▄▄▄▄▄ I remember the director of that division once telling me, "we are the █ █ greatest users of computing at Bell Labs. We use more time than anybody █▀▀▀█ else. We depend on it much more." And that was because they're using a █ █ macro-processor, which runs slow by about a factor of a hundred, rather ▄▄▄▄ than a regular compiler. He was just blissfully unaware that macros were █ █ great for defining a language, but once they had it, they should have █▀▀▀▄ built a compiler. █▄▄▄▀ ▄▄▄▄▄ █ [ So between when you started in 1958 and when they created Unix... ] ▀▀▀▀█ ▄▄▄▄█ Oh yeah, that was eleven years. [ What else were you working on? ] Um...besides the macros, what else did I do? Lots of consulting. I worked with Dick Hamming, of the Hamming crew. [ Hamming? Did you keep your distance? ] Yes. They wrote a couple of papers about macros in there, but what else, what else did I do? Oh, worked out. Multics, that's it. Yeah. [ Multics, okay. ] Multics started in 1964, beginning with picking the hardware. At the same time, I had gotten myself onto the PL1 design committee. When Multics was choosing what language they were going to write in, there were several candidates around, and one... all of them, almost vaporware. PL1 was just getting its first compilers written. I was responsible for making sure that PL1 had enough facilities that it looked like a decent language to write in. I was responsible for bringing that in. Then I contracted with an outfit in California called Digitech, which had a terrific business of making Fortran compilers. They were really good. These Fortran compilers could even run on very small machines. Dynamically allocated internal tables. They could grow and move, so they'd get really the most out of every machine. They went for this contract, thinking they'd get in on the ground floor of a new language. But they put a new employee on, didn't supervise him, and the people who knew how to use their compilers didn't pay much attention to it, and he screwed up terribly. So there we were a year later with no compilers and Multics waiting for one. Bob Morris and I, when Bob Morris started it, decided we'd write a compiler in TMG. Do you know what TMG is? [ No. What is TMG? ] It was a compiler writing language from Bob McClure at Texas Instruments. He had sent it to me. He wrote it for the CDC 1620, which was a 36-bit machine, and he hand transliterated it on the green coding sheets into machine language for the IBM 704, 7090 by then. It was an interesting debugging thing. I knew the program logic was right, but there were opcodes like CAL and CLA (clear and add and clear and add logical), and you could get the wrong one if you were hand translating. So debugging was all finding clerical errors. And I had brought that up. We wrote the first Multics compiler in a couple of months in TMG. That lasted Multics until about 1970, when finally GE made a really good compiler for us. So that's what I was doing in that interval. TMG, Multics, and we had this big GE 645 in our attic. Three of us would be beavering the way out. MIT was the prime instigator of Multics. They had one in Cambridge, and they let us use their Cambridge time sharing system. We really, really stressed that time sharing system by writing the, by running a TMG written compiler, which had long compile times. They were optimized for short times. Their whole scheduling said, always prefer the short program. A really long program actually ran in the meantime between failures, and that meant you ran the long program twice. We were really screwing up MIT's computation center. The Multics project crept on and on, never got anywhere. Finally, higher management at Bell Labs said, "let's cut our losses". We were paying a million dollars a year rental or something like that on this machine. And it was supporting three people. Oh, that's enough about Multics, I guess. [ Then Unix folllowed? ] Ken Thompson in the background had been playing with these ideas for Little Machine. Thompson and Richie and Joe Ossanna lobbied hard to continue the work on operating systems. Our management, once burned, didn't like the idea. They thought it would be nice to have a PDP-10. Eventually Thompson finds this cast off PDP-7 and in three weeks builds the whole operating system. [ That's wild. Three weeks?? ] The assembler and the operating system. His wife was away for vacation. He allocated himself one week to build an assembler, one week to build an operating system, and one week to build the file system. [ That's nuts. ] They had been thinking about the file system. The design was ready to go. How he came up with the API for the operating system, I actually don't know. So clean. He threw out a lot of things. He had good ideas from Multics, like the hierarchical file system. Although he implemented it in an entirely different way. The shell, the separate shell that came from Multics, he had it up and running. [ What do you mean by "separate shell" that came from Multics? ] The command language for operating systems up until that time had been built into the operating system. The user space shell was a new idea of Multics. Not quite a totally new idea. The Dartmouth time sharing system had something, their idea was to just put a shell in user space, but they didn't think of it as a substitutable program. Everybody used the same shell, but it was in user space. Almost like the dynamic linker today. In Unix, several shells actually did get written. Although nowadays everybody uses the same one. [ What shell was it? Was it a named shell? ] It was a progentior of the Bourne shell. [ How long was it until they started networking Unix machines together? ] Well at the same time we had Sandy Fraser in the computer science department who was a real proselytizer and designer and builder of networks. Unfortunately, although he thought networking and digital data was the future, the big guns at AT&T said, "Digital data is 4% of our business, why should we worry about that?" Sandy didn't get near the support that he deserved. But anyway, he made these networks that we had used in the lab among our machines probably starting in 1972 or 73. We ran through about three generations of those. Eventually got onto CSNet, Computer Science Net, a version of ARPANET was not available to the industry. [ What was CSNet? ] CSNet was kind of an interim between ARPANET and the internet. [ Was it between colleges? ] Yes. CS departments. [ What was the communication medium? ] It was both... It had dial-up and... I even forgot the name of the protocol. But it was almost called dial-up. It had a little bit. And a bridge to the ARPANET. A bridge to ARPANET. [ What was the speed? ] Oh, I remember when we went to 1200 (baud). That was blindingly fast. [ What time period was that? Was that still early 70s? ] That would have been... I'm trying to think... When the cu utility was put into Unix. cu is call up another Unix system. [ I'm not familiar with that. I've never used that. ] Do you know my... So called Unix reader[3]? You'll find it on my website. Anyway, it has a very useful table of contents of all of research generations of Unix. Which I refer to when I need to know one of these things. I can tell you from that table of contents just when networking began. └──────────────────────────────────────────────────────────────────────────────┐ ▄▄▄▄▄ [ student: I'm still thinking about that whole Switching team at Bell █ █ Labs using macros as a programming language instead of just writing █▀▀▀▀ their own programming language and compiler. What was the reason for █ this? ] ▄▄▄▄▄ █ █ █▀▀█▀ I can't think of anything special. █ █ ▄▄▄▄▄ █ █ How can they write big systems that consist only of machine language? █ █ █▄▄▄█ ▄▄▄▄▄ How can people write big systems in general? █ █ ▀█ █▄▄▄█ We do, but as you say, they're often debugged into existence rather ▄▄▄▄▄ than cleanly designed into existence. █ █ █▀▀█▀ █ █ [ When you were working on macros, presumably, you were spending a lot ▄▄▄▄▄ of time debugging macros. Now, even today, C preprocessor macros can █ █ be super hard to debug. So, like, did you have tooling here that we █▀▀▀█ don't have anymore? ] █ █ ▄▄▄▄▄ █ █ █ I don't remember debugging difficulties. █ █ █ █ █ ▄▄▄▄▄ [ Ah! You don't remember, you must have been a good debugger. ] █ █ █ █ █ █ █ █ Well, one time I wrote a Lisp compiler, in macros. ▄▄▄▄▄ █ █ [ Oh, my God. ] ▄▄█▄▄ ▄▄ ▄ █▀▄ █ It was a pretty spectacular construction. I remember sitting in bed █ ▀▄█ with paper spread all around, all over the bed one night. I got the █ ▀█ thing was working within a couple more days. ▄▄▄▄▄ █ █ ▀█ It was about, as I recall, it was something that, typically, code █▄▄▄█ would issue at call depth, a macro call depth of 50 to 70. It was an optimizing compiler. That came with time. Bob Morris looked at some of my object code, and he made the astute comment, notice that transfer on non-zero sets the accumulator to zero. [ What do you mean? ] The TNZ instruction. It actually doesn't change the accumulator. But, if you flow through that instruction, you know that the accumulator is zero. Otherwise it wouldn't jump. It's conditioned. And, Lisp has lots of tests on zero. I could optimize the test for null away. That kind of little local optimizations. [ I was actually going to ask you about that. We are interested in weird file optimization and program optimization in general. You are talking like 2K of words, what was it like working within those early constraints? Especially for things like files transfer and storage. ] Well, you know, occasionally you'd look around for fields, fields buried in instructions that weren't used, you'd find a little extra memory that way. But, by and large, since every program is so constrained, that you wouldn't bother conceiving of something that needed a lot of space. The first IBM PL1 compilers had for the 360 came within several levels with completely different hardware implementations. They designed their compiler, so that would work on the smallest machine. This is a case where I think your question is really pertinent. They designed their compiler in such small pieces, that on the smallest machine, it could run in something like 60 overloads, one after another. On the bigger machines, everything would be in memory at once. A fascinating design, and very cool. [ Um, so, you were the inventor of the Unix pipes? ] Yes. [ That's pretty nuts, because we use pipes, like, on a daily basis. ] Everybody does. It's one of the most wonderful inventions in that regard. [ What was life like before pipes? ] Well, you'd write these programs, you would write a file, you'd write a shell script, one item would write in the file, and the next item would read from the file. The amusing thing about this is that I don't think either Ken or I were conscious of the fact that there's an authentic difference between writing a file and reading a file and piping. The difference is real time. [ Uh huh. ] If you want to have an interactive pair of programs, you can do it by piping them together, but you can't do it by writing a file and reading a file, because you have to write the whole file before you read it. And so you can't run it, you can't have interactive feedback from the output to the input. [ What was the design process for pipes? ] There's no design process. Well, while Ken and Dennis and Rudd Canaday on one side of the corridor were thinking about file systems, I happened to be on the other side of the corridor cooking up notations for composing programs. Due to lack of imagination, I ran into the difficulty of, if you say ls through wc, that's pretty cool, you can just write down ls | wc fine, no concatenation. But suppose ls has flag options, how do you distinguish options from concatenation of programs? For a long time, this baffled me. And then one day, I invented a notation, not the one that we're using now. [ Okay. ] You have multiple greater thans. [ Almost like a bit shift. ] Yeah. I had been lobbying for something to connect programs together for a long time. Already I had in mind not only pipelines, but cellular automata, where you connect them up to at least a player array. [ Wow. ] I'd been lobbying for putting in something, but Ken, Ken, just wouldn't do it because it made the system more complicated, perhaps. But when I came up with a notation for the shell, he said, to my surprise, "I'll do it." And he did it in one night. [ Of course he did. ] And not only did he do it, but it's no surprise that he could do it in one night because of the architecture of the system. It already had shared buffers between processes. All he had to do was to have, instead of just one read-write pointer in the buffer, he had to have two pointers, one for one process and one for the other. In the same night, he recruited Dennis. They realized that lots of programs which had named inputs and outputs would naturally be pipeable together. You had to arrange for them to have either a missing output file or a missing input file. They changed a whole lot of programs that night, too. And the next morning they told us, it's working. Then we had an orgy of pipelines that were demonstrated. Look what I can do, look what I can do. [ I bet. ] I think the most fun one was Joe Ossanna. He wrote nroff and troff[4]. I wrote roff. He had a two column gimmick. He had set alternate pages on the left half of the page, on the right half of the page. Then he had a final post processor called Overlay, which would OR together these pairs of pages. He said, oh, I can do that in a pipe now, and furthermore, I can pipe two of them in a row and get four column output. I thought that was the prettiest one. By the end of the week, our secretaries were piping stuff to the line printer. [ That's so sick. ] The original description in the manual took a page. Because of the double greater-thans, there were some really ugly, ugly things there in order to properly parse it. The manual now manages to describe pipes in the shell in about that much space. **holds hands up to show a small block** That's all there is to it. And that's what's wonderful about it. You don't have to learn anything. Finally the, the pipe symbol itself came up. Because Ken was scheduled to give a talk about Unix in London. He was embarrassed to talk about this ugly notation. So he cooked that one up, and it really stuck. Ken gets a lot of credit for the symbol. [ One corner case of pipelines is unintuitive is the exit status of a pipeline. You can have programs that exit non-zero piped into another program that exits zero && another program, and it works. What was the thinking there? ] I think whatever happened happened accidentally. [ Okay. You were just interested in functionality. Like get this pipe to the other, get this data over here to this program faster, the exit status doesn't matter? ] Yes. Have you ever heard of the programming language called Pogol? [ Pogol? No. Uh, no. What's that? ] It's not very well known. It came from NSA. [ Of course it did. ] It's a data processing language. It came from 1970 to 72 thereabouts. Tthe way you write programs in it is you write little bits of code that write unnamed files and read from them. Then you take all these little bits of programs, compile them together, and it optimizes the files out of existence. [5] [ I can't even find a Wikipedia page on it. Dude, the AI in Google says that it doesn't exist. Wait, I found a paper on it... ] [ 1973, Early development occurring in the Department of Defense. ] [ So this sounds a lot like pipes, yeah. ] They were ahead of pipes, but they didn't have the pipe notation. You still had to pretend there was a file. [ There's very little about Pogol out there. ] Very little, just the one paper by Gloria Lambert.[6] ┌──────────────────────────────────────────────────────────────────────────────┘ ▄▄▄▄▄ █ [ What do you know about executable formats for Unix? Like a.out, COFF, █▀▀▀▀ then eventually ELF. Can you tell us about the transition to using file █▄▄▄▄ formats like ELF? ] ▄ █ █ By the time ELF was embedded, it was going on in the Unix support group. █▄▄▄▄ I'm completely unaware of how it was designed. ▄▄▄▄▄ █ █▀▀▀ I have a question about ELF, which you people all know about. Um, I had █ one program that had a 2^15 * 2^15 byte of array in it. [ Uh huh. ] I initialized element 00 of that array, one very first element. It took forever to load this program. In C, even though only one element is initialized, C compiles it, initializes the whole thing and puts it into ELF. As far as I know, there is not a way to initialize a piece of an array and have the other uninitialized in the ELF file. [ I think it's like a more of a compiler thing than an ELF thing. ] It seems like. It's how the compiler uses ELF, but there didn't seem to be a way for the compiler to force ELF to do that. Because that single element has to be adjacent to all the other elements. I don't think you can tell ELF that these two data things have to be contiguous. [ Yeah you will usually have to do this manually or at runtime. ] This billion-byte binary, uh, is, uh, really quite a surprise. └──────────────────────────────────────────────────────────────────────────────┐ ▄▄▄▄ [ So, before there was Dwarf, what was the debugging procedure like? What █ █ was the debugging information? How did the debugger figure this out? ] █ █ █▄▄▄█ ▄▄▄▄▄ DWARF? I never used DWARF. █ █▀▀▀▀ █▄▄▄▄ [ When you wanted to debug something, how was your debugger figuring out ▄▄▄▄ the information about the debug target? Did debuggers map assembly to █ █ symbols? If so how? And if it crashed, would it map to the source? ] █▀▀▀▄ █▄▄▄▀ ▄ ▄ You would use whatever was in the symbol table of the C program. And █ █ that goes way back. █ █ █▄▄▄█ ▄▄▄▄▄ [ Did Ken write the original debugger for Unix? ] █ █ ▀█ █▄▄▄█ Um, I think the first debugger was probably written by Dennis. ▄▄▄▄▄ █ █ ▀█ [ What was it called? ] █▄▄▄█ ▄▄▄▄▄ █ db █ ▄▄█▄▄ ▄▄ ▄ [ Obviously, right. ] █▀▄ █ █ ▀▄█ █ ▀█ It became adb. ▄▄▄▄▄ █ █ ▀█ [ Ah there was another adb, before the Android debugger :P. Advanced █▄▄▄█ Debugger? Written by Stephen Bourne. ] Yes. It's wonderful to be in the prescence of people who know the insides of this stuff. [ Nerds? ] Yeah. I am really remote from it now. [ So what about backtraces? In those original debuggers, you want to see the call stack. So you've got to unwind the stack. What was the procedure for doing it? ] It's been so long since I used the debugger. I think you could ask for a whole stack trace. We could also ask for elements in the current stack frame and then move a stack frame. I eventually gave up on using...because I don't write really big things, by and large. Recompile it and throw in some printfs. [ That's the way. ] Well, typically what I'll do is throw in a lot of printfs that print a single character with no new line. And that gives you a trace. That's often enough to locate where the trouble is. [ So we keep asking, how did you debug this and that, and you keep saying "Oh I didn't need to, just printf()". It's starting to make sense how you would approach these types of problems with extremely limited resources. ] Well, you know, I didn't come up with that technique in the days of limited resources. That's something, in the days of limited resources, I actually used debuggers. Debugging may be better with a limiter, so there is not so much to look at. I mean, there's so much, so much now. [ What was the most interesting bug that you think you've seen overall? ] The most interesting bug... I think it's one that I found in the Fortran 2 compiler. It was interesting just because it was a programming technique that I never saw before or since. Do you know about the 7090 machine code? The 7090 had some instructions with four parts. Two addresses, or an address with an offset, an opcode, which was only three bits, and an index register, which was only three bits. As evolution of the family of computers, on the 704, you could store a whole machine word, or you could store one of those address fields. On the 709, you could store the whole machine word, or either one of those. And on the 7094, they even added, you can store into the opcode, or the index register tag. I never saw a store tag used in any program except one. Somebody brought me a piece of Fortran code with the computed go-to. You list a whole bunch of addresses, and you index into that list. This computer's go-to failed. It had 74 branches. If you took one of the branches out, it worked. If you added a branch, it worked. At that time, we had a copy of the Fortran compiler, this was before debuggers. It's the only time I have used a 7090 for my console. I stayed up all night with it, and finally located the bug. There was a table in which they kept the addresses for a computed go-to. They wanted the table to grow to arbitrary length, but they didn't want it to take arbitrary amount of memory, so they paged it in 75 word pages. I cannot remember the exact logic. But, instead of doing any rational indexing, they used the store tag to turn off indexing after 74 elements. So, in fact, they forgot to do that. They did it in one place, but they needed to do it in two places, and they didn't find the second place. Needless to say, they'd never tried an example of 74. So a computed go-to in Fortran 2 would fail if there were 74 mod 75 branches, and I was really proud of finding this bizarre trick of the store tag instruction. Yeah, I think that's my favorite. [ That sounds insane to figure out. ] Yeah. ┌──────────────────────────────────────────────────────────────────────────────┘ ▄ ▄ █ █ [ Earlier this morning we were discussing the trusting trust compiler ▀█▄█▀ backdoor ] █ ▄▄▄▄▄ █ A little louder. One of my hearing aids is just burned out. █ ▄▄█▄▄ ▄▄▄▄▄ [ The trusting trust compiler backdoor. Ken Thompson actually implemented █ █ it? ] █▀▀█▀ █ █ ▄ ▄ Yes, he did. [7][8] █ █ █ █ █▄▄▄█ [ And then it nearly escaped or something like this? ] ▄▄▄▄▄ █ ▀▀▀▀█ Yes. There was another group that maintained the so-called programmer's ▄▄▄▄█ workbench. They tried to keep abreast of how it worked, and apparently ▄▄▄▄▄ Ken had this new compiler in his home directory and they found it. They █ were like "Oh, we want to be up with the latest stuff". The interesting █▀▀▀▀ thing is, I don't know how they did it, but they discovered it to be █▄▄▄▄ bugged. ▄▄▄▄▄ █ ▀▀▀▀█ They actually tried it, and whether they discovered it by diffing it ▄▄▄▄█ with the old code to see what was doing, or by measuring the size of the code, I don't know. But they actually discovered it and called up Ken and he said, "get that thing out of there". So, it was lucky that it didn't really spread. I had a similar case when Jim Reeds and I were working on IX [9], a multi- level secure Unix system that handled Bell-LaPadula [10] classification schemes. Bell-LaPadula was a multi-level thing. We observed that if you turn it upside down, Bell-LaPadula makes it illegal to read higher level stuff into lower level memory. Instead, if you prevent writing higher level stuff into lower level memory, you can use it for integrity. So, we used the two together in the lattice and started in the middle. Going up, you had privacy control. Going down, you had integrity control. For changing system components, it's the integrity control you want. You don't want people to be able to reach. So, system pieces that everybody could use were below the floor where you normally started. Highly secret stuff that people weren't supposed to look up were up above the floor. Well, originally we put the floor at zero. The first time we turned on the floor up in the middle, it failed and said there's a security violation. It was probably a certain thing that we had made a mistake in the implementation somewhere, and Jim worked for a couple of days over it. Then I get this phone call from them and said, "We caught a virus!" Now, what was a virus doing on our machine? Well, this goes back to Tom Duff a couple of months before. He wrote about shell viruses. It was published in the Usenix journal. [ What year was that? ] That would have been 1987 or 88. There were two articles about shell viruses in that issue in the Usenix journal. Tom Duff wrote one [11] and I wrote another [12]. He had actually implemented the virus and to his dismay, it got loose on our network. So he implemented a worm that would get rid of it. However, Jim and I were working on this, had a disconnected system at the time that we were developing IX on. We happened to be connected when he sent the virus, when the virus got loose. Then, we were disconnected when the disinfectant was running around. It had been in our machine all along and we didn't know it. But the minute we turned on integrity control, we discovered, it revealed itself. So that was a wonderful triumph of our new system. [ So Tom Duff wrote the virus? ] Yeah. [ He worked there as well? ] Yes, he did. [ How did he accidentally? ] I think he was working there when he got his Emmy [13] for the, for the alpha channel [14]. But, he didn't invent the alpha channel. He was working with us. [ And you say he invented, he created a worm to disinfect it? ] Yes. [ That's kind of funny. ] Well, it was probably not a worm. It was probably all reaching out from one computer to the others. [ What was the mechanism of action here? Was there a bug that the worm was exploiting to fix the other? ] No, I don't think that. When I said worm, I probably misspoke. He did something to reach out and touch every other machine and remove the virus. [ Was that the first virus that you ever saw on a computer that you used? ] No. Um. It's probably the first one I ever saw in real life. Yes. └──────────────────────────────────────────────────────────────────────────────┐ ▄ ▄ [ You were at Bell Labs when GNU first became a thing, right? ] █ █ █ █ █▄▄▄█ Yes. ▄▄ ▄ █▀▄ █ █ ▀▄█ [ What was the response internally? ] █ ▀█ ▄▄▄▄▄ █ Good question. Uh, by then we had so many imitators. The policy was █ that if all they were doing was re-implementing those things that had ▄▄█▄▄ already done before, it was fine. So, there was nothing very special ▄ ▄ about GNU except, except for all the, uh, overburden of policy on it. █ █ That was not of interest to the lawyers. ▄▀▀▀▄ █ █ >>>>> [ So you're saying as long as they were just re-implementing, what you'd >>>>> already done, it was fine. What would not have been fine? ] >>>>> >>>>> ▄▄▄▄▄ Well, one example was a guy in Boston, his name escapes me now, was █ ▀ selling a collection of Unix utility lookalikes. Well, that's what he █ called them. When we looked at those, I looked at them in fact. If you █▄▄▄█ looked at the binary, it was in one-to-one correspondence with the ▄ source. █ █ █▄▄▄▄ [ Wow. ] ▄▄▄▄▄ █ █ █ █ It happened at the time that one of our patent lawyers was negotiating █▄▄▄█ cross-licensing with DEC. DEC was giving him a hard time and said, ▄▄ ▄ "you don't have anything of interest to us." █▀▄ █ █ ▀▄█ █ ▀█ He went to a meeting armed with some material that he thought would ▄▄▄▄▄ prove that he had stuff of interest to them. But, unbeknownst to me, █ he went armed with another thing. █▀▀▀▀ █▄▄▄▄ ▄▄▄▄▄ He went armed with the fact that I had told him about this, um, robbery █ of our code. It turned out that DEC had brought along an expert witness ▀▀▀▀█ who was this guy. ▄▄▄▄█ [ Interesting. ] And our lawyer, at the end of the day when they broke up and had some coffee together or whatever, he somehow managed to bring the subject matter around to furloughing the code. And he said, I happen to have an example in my briefcase, would you like to see it? "Oh, yeah!" He brings it out and here's the source, here's the object. The guy said you know, that sure looks like it came from that source. Would it interest you if I told you that this is ours and that's yours? He didn't show up the next day. [ He had no idea that guy was going to be in that meeting? It was just a coincidence? ] I, he may, he probably did have no idea. [ That's incredible. ] He had no trouble after that in getting the guy to cease and desist. ┌──────────────────────────────────────────────────────────────────────────────┘ ▄▄▄▄▄ █ [ What was your favorite program that you wrote for early Unix? ] █▀▀▀ █ ▄▄▄▄▄ Well, I mentioned echo, but that's just because it's amusing. █ █ █▀▀▀█ █ █ Uh, I'll mention two of them. One was a plumbing program, which took a ▄ ▄ long list of plumbing connections to make among processes, of pipes, and █ █ would pick them all up. ▀█▄█▀ █ ▄▄▄▄▄ The interesting thing about this program is you can only have a limited █ █ number of pipes open every time, and we had to do something to get that. █ █ That was fun, but not useful. The one nice program is Speak [15]. █▄▄▄█ ▄▄▄▄▄ █ █ [ How did the design of that go? ] █▀▀█▀ █ █ ▄▄▄▄▄ Joe Ossanna got wind of this new device that you could buy from Federal █ Screw Works. Turns out a guy in the back room of Federal Screw Works had █ come up with a speech synthesizer. It has nothing to do with the rest of ▄▄█▄▄ our business. ▄▄▄▄▄ █ █ You fed it phonemes, and it would pronounce them and string them █ together. It's hard to do, we had a whole department working on speech ▄▄▄▄▄ synthesis. And they could do it, but it usually took something like a █ minute of computer time to produce a second of speech. If you had the █▀▀▀▀ Votrax [16], oh, you just fed it phonemes in real time. █▄▄▄▄ <<<<< <<<<< So Joe said, "we need one of those" and he put it on a machine. Lee <<<<< McMahon promptly became really quite expert fashion in phoneme streams. <<<<< But it's still like hard labor. You had to prepare it. ▄▄▄▄▄ █ █ █▀▀▀▀ I looked at it and I realized that an awful lot of English, especially █ all the big words, things that fill up the dictionary can be pronounced ▄▄▄▄▄ phonetically. So I put together in one night this program that had a █ █ list of phonetics, which were fragments of words, and it would take the █▀▀█▀ longest fragment, the longest matching fragment, and pronounce it. And █ █ maybe rewrite part of it and pronounce part of it. ▄▄▄▄▄ █ █ █ █ I put together a list of rules for it, and tried it the next day. Well, █▄▄▄█ Morris walks up to the machine and types in "Orlok", and it came out ▄▄▄▄▄ perfectly. He said, oh, "it's a success" and it really actually was a █ success. We put itpermanently online. Anybody could walk up to the █ ▀█ computer, type into the computer, and it'd talk back to them. █▄▄▄█ ▄▄▄▄▄ █ █ [ And you wrote that in one night? ] █▀▀█▀ █ █ ▄▄▄▄▄ I wrote it in one night, then, spent the next six months perfecting it. █ █ But, it went into business right away. █▀▀▀█ █ █ ▄▄▄▄▄ [ Who is this guy at Federal Screw Works? This guy sounds like a genius. ] █ █ █ █ █ █ █ █ Right. Yeah. It was really amazing. ▄▄▄▄▄ █ ▀▀▀▀█ [ What even is Federal Screw Works? ] ▄▄▄▄█ The department head for the speech department came in and he listened to it. And he said, this is a breakthrough. But of course, one of the things the speech department was trying to do was make it a natural speech. The Vortrax clearly wasn't natural, but it's predictable, and it turns out that for many tasks, that's all that counts. To foreign speakers of English, non-native speakers, it turned out the Votrax was actually, because I put a little gap between words, ten milliseconds or something like that between words. That segmented it, whereas natural speech is not segmented. That's one of the real problems of listening to a new or foreign language. So foreigners would listen to it and say, that's easier to understand than people. Although, I could not do syllable stress. I never got that. It was just, every syllable was at the same, same stress level. But it was fun. [ What kind of other things did you develop for Unix? We talked about pipes, and echo, you did tee and tr as well? I did tee and tr yeah. Those are just obvious things. You need one night and you put them together. Spell was a big one. [ What was that one? ] Oh, it's a spell checker. Ah, have you ever, have you ever seen the spell checking pipeline that I taunted Donald Knuth with? [17] [ I don't think so. ] There's actually a nice cartoon online about it. Anyway, Steve Johnson invented this pipeline for doing spell checking. It works wonderfully if the words are in the dictionary, but you have to fill your dictionary out with all the plurals and derived forms. That should be done automatically rather than, rather than filling up 200,000 word dictionary. So I did the automation part, all of the businesses, suffixing, prefixing, and turned this pipeline into a nice spell checker. [ Nice. So this is early 70s that you wrote that? ] Probably around 76. The pipeline is a wonder[18]. You will really want to see that. You should see how big the Donald Knuth program is in comparison. One of the things about the spell check program is, I ended up with a vocabulary of 30,000 words. But, you know, 30,000 words at an average word length of seven, something like that, is rather too big for a 64K memory. So there's some great data compression in there[19]. John Bentley wrote one of his, um, programming pearls columns about it[20]. [ Who is John Bentley? ] John Bentley was another Bell Labs person who came from Carnegie Mellon. After he got a teaching prize at Carnegie Mellon, the president presents it to him and says, "you know, every time you realize you're getting paid for research, not teaching, this prize makes up for that a bit."" John really didn't appreciate that comment, so he came to work for Bell Labs. Where he continued to teach a lot by writing these columns for the CACM. [ What was it like being around so many really smart people at Bell Labs? ] People being recruited also would say, I wasn't interested in coming here. Thompson said this. I wasn't interested in coming here, but then I walked down the corridor and saw the names on the walls, that changed my mind. It was a really nice collegial atmosphere. I discovered when I came to Dartmouth that colleges aren't as collegial. Especially the poor assistant professors struggling for tenure. They can't take a minute off to get diverted to something else. You could walk into anybody's office at Bell Labs and say, who's an expert? Then say, I need help here, and they may help. You weren't paid for how many papers you wrote. You were paid for how useful you were to the company. [ What year did you come to Dartmouth? ] Uh, 97. [ So then were you at Bell Labs for Plan 9? ] Yes. I didn't participate in the project, but that was while I was there. It's really too bad that it didn't catch on. Well, there are a thousand of them out there still, but. [ I would say like, I mean, 9P still exists. And like, ProcFS came from Plan. I mean, like the, the big ideas got imported. But yeah, it is sort of unfortunate, that like the grand re-architecting of things is not adopted. ] Some pieces do still exist, like the idea that multiple computers are not different from one computer. We still have sockets and files. And Plan 9 even managed to overcome big and little and the differences between machines. [ How? ] When you move data between them, they would automatically do the necessary swapping. [ For like executable code? ] If you were doing executable code you had to recompile for different machines. But they did that very nicely. No ifdefs, all done with, um, makefiles. [ Would it do like automatic detection of the data for the machine type? Well, how would you know if it's just data? ] It would. A machine of course knew what it was, so it would establish connections. If you talked to a different machine, right. But if you were just given some data. They both had to be running Plan 9. [ Gotcha. So what got you into programming Haskell? ] Well, the thing is, it's a beautiful imitation of mathematics. And, um, I happen to be an addict of stream processing. Haskell is the best stream processing language I know of. With lazy evaluation, a stream is exactly like a list. And, and so there's some lovely demos you can do. [ What do you use it for? ] I happen to be a fan of map projections. One class of map projections is the doubly periodic. It's doubly periodic. You have a world map and you can paste them together and tile, and tile the globe. It continues like the Earth goes without ends, repeated, copies of the same map in two dimensions. If you know advanced calculus, you can do this with conformal mapping, elliptic integrals. I have used Haskell to compute the coefficients of the power series expansions of the elliptic integrals. I also use Haskell to help make Christmas cards for myself. [ So it's easier for you to do this type of math in Haskell? ] Knuth tells you how to, in uh, volume one, I guess it is, how to do various power series calculations. And the most complicated one is called reversion, which is a function to find the power series that is the functional inverse of another power series. Knuth does reversion in a half page of pseudocode. I do it in one line in Haskell. Real code, working code. [ Yeah. Real code, you call it. This is out of The Art of Computer Programming? ] Yes. [ When we look at what languages have caught on, the family of things that spawned off from C remains super dominant. Way more so than the things that came from ML like Haskell. ] Yeah. [ And similarly, when we look at the operating systems in use today, and the family of things that spawned off from Unix, they integrate ideas from the future, but the older idea remains. What do you think accounts for the stickiness of C in Unix? ] Well, partly, the assignment statement really makes code go faster. No, ML does not have an assignment statement. It isn't done. [ It has "let" ] Yeah, but, but you, but you are never proud when you have to write one. [ Yes. ] Haskell doesn't have one at all. Uh, Lisp, Lisp was another one that just, uh, had a dirty assignment statement. [ Unpopular opinion, but I never really liked Lisp. ] Well, it's, too many right parentheses. That's the nice thing about pipes. They do away with the right parentheses. [ But they also do away with the tree. So because it's a line, you don't need the parentheses anymore, but often you want the tree... ] It's sad to say, we've, nobody yet has come up with a nice way of expressing. I don't know how many different shells allow trees, including the Bourne shell. Uh, the, I'm sorry, bash. None of them are very appealing syntaxes. It still eludes us. [ student: I am working on a domain specific language for on the fly testing of REDACTED. I want to send it a request to multiple REDACTED, then send back through other REDACTED, then bounce that to REDACTED, then have it all go backwards through a bunch of stuff. My problem is the syntax, just thinking about how to make this not annoying to write. I still have no idea what the solution is. ] Yeah. I sympathize with you. I've, I wrote my first proposal for such a thing that never, has never seen the light of day. Because it's so ugly. And I keep going back to it and nothing ever gets better. [ student: Well, maybe it's just inherently graphical. ] Well, there's one argument you can make. You can make it for structured programming. Structured programming only covers a very small fraction of all possible graphs. But it turns out, this is an extremely useful subset, and it's pretty clean by indentation. We haven't yet found that subset for, other than linear pipelines for connecting processes, and I don't know why. [ student: I mean, I guess the question is, does that, is that subset even significantly different from the space of all graphs? Yeah. [ student: When you're writing a compiler, you want things to be inter- operable, and you have to figure out how to do that by reading the ABIs and understanding the constraints and all that. If you're starting fresh, like at the beginning of Unix, when none of this exists, so you get to make it all up, which sounds very fun. But it also sort of implies that you need a long design phase and think about how you can make it so that people actually want to write machine code for this thing. It sounds like you had long design phases and then short bursts of writing code. ] Yes. [ student: This is very much not how, uh, a lot of people work today. Was there a specified ABI for Unix? Was it well documented or was it just like "eh, whatever works"? ] Yeah, that's a huge contrast between Multics and Unix. Multics had a five-foot design shelf, uh, and it still wasn't working. In Unix, the first manual occurred two years after the system. └──────────────────────────────────────────────────────────────────────────────┐ ▄▄▄▄▄ Our interview had to come to an end because it was lunch time, followed █ by afternoon talks. We invited Doug to come hang out, and he attended █▀▀▀▀ lectures by elfmaster, netspooky, and sblip. █▄▄▄▄ ▄▄ ▄ █▀▄ █ It was an honor to talk to such a legend. Thank you to Sergey Bratus for █ ▀▄█ inviting us all in to Dartmouth! █ ▀█ ▄▄▄▄ █ █ Gigantic shoutout goes to sblip for recording and leading this interview. █ █ █▄▄▄█ ┌──────────────────────────────────────────────────────────────────────────────┘ ▄▄▄▄▄ █ ▀ [1] Whirlwind 2 Info █ https://www.radomes.org/museum/parsehtml.php?html=fsq-7.html&type=equip_html █▄▄▄█ https://x.com/MITLL/status/1910440210791379145 ▄▄▄▄▄ █ [2] von Neumann, J. and Goldstine H.H. (1947) Numerical Inverting of Matrices of █ High Order. Bulletin of the American Mathematical Society, 53, 1021-1100. ▄▄█▄▄ https://doi.org/10.1090/S0002-9904-1947-08909-6 ▄▄▄▄▄ █ [3] A Research UNIX Reader: Annotated Excerpts from the Programmer’s Manual, █ 1971-1986 M. Douglas McIlroy █ https://www.cs.dartmouth.edu/~doug/reader.pdf ▄▄▄▄▄ full link: https://archive.org/details/a_research_unix_reader/mode/2up █ █ █▀▀▀█ [4] troff History █ █ https://www.troff.org/history.html ▄▄▄▄▄ Editor's Note: nroff is new roff, troff is typesetter roff █ █ [5] https://en.wikipedia.org/wiki/Dataflow_programming █ ▄▄▄▄▄ [6] Gloria Lambert (1973). "Large scale file processing: POGOL". _POPL '73: █ Proceedings of the 1st annual ACM SIGACT-SIGPLAN symposium on Principles █ of programming languages_. pp. 226–234. ▄▄█▄▄ https://dl.acm.org/doi/abs/10.1145/512927.512948 ▄▄▄▄▄ █ █ >"POGOL, an otherwise conventional data-processing language developed at █ █ >NSA, compiled large-scale applications composed of multiple file-to-file █▄▄▄█ >operations, e.g. merge, select, summarize, or transform, into efficient ▄▄ ▄ >code that eliminated the creation of or writing to intermediate files to █▀▄ █ >the greatest extent possible." █ ▀▄█ █ ▀█ [7] Reflections on Trusting Trust (1984) by Ken Thompson ▄▄▄▄▄ https://www.cs.cmu.edu/~rdriley/487/papers/Thompson_1984_ReflectionsonTrustingTrust.pdf █ ▀▀▀▀█ [8] Computerphile video on Trusting Trust ▄▄▄▄█ https://www.youtube.com/watch?v=SJ7lOus1FzQ [9] Multilevel Security in the UNIX Tradition - M. D. McIlroy, J. A. Reeds (1992) https://www.tuhs.org/Archive/Documentation/TechReports/Bell_Labs/CSTRs/163c.pdf [10] https://en.wikipedia.org/wiki/Bell%E2%80%93LaPadula_model [11] Tom Duff, Experience with Viruses on UNIX Systems (1989) https://www.usenix.org/legacy/publications/compsystems/1989/spr_duff.pdf [12] Doug McIlroy - Virology 101 (1989) https://www.usenix.org/legacy/publications/compsystems/1989/spr_mcilroy.pdf [13] 1996 Academy Award Winners https://www.imdb.com/event/ev0000003/1996/1/#scientific_and_engineering_award [14] Thomas Porter, Tom Duff. "Compositing Digital Images" Computer Graphics Volume 18, Number 3 (July 1984) https://keithp.com/~keithp/porterduff/p253-porter.pdf [15] https://en.wikipedia.org/wiki/Speak_(Unix) [16] https://en.wikipedia.org/wiki/Votrax [17] https://www.leancrew.com/all-this/2011/12/more-shell-less-egg/ [18] Editors note - this is the pipeline he describes:

   tr -cs A-Za-z '\n' |
   tr A-Z a-z |
   sort |
   uniq -c |
   sort -rn |
   sed ${1}q

His explanation: If you are not a UNIX adept, you may need a little explanation, but not much, to understand this pipeline of processes. The plan is easy: 1. Make one-word lines by transliterating the complement (-c) of the alphabet into newlines (note the quoted newline), and squeezing out (-s) multiple newlines. 2. Transliterate upper case to lower case. 3. Sort to bring identical words together. 4. Replace each run of duplicate words with a single representative and include a count (-c). 5. Sort in reverse (-r) numeric (-n) order. 6. Pass through a stream editor; quit (q) after printing the number of lines designated by the script’s first parameter (${1}). [19] "yes, it's me, the author of spell" https://x.com/abhi9u/status/1887010136155414602 https://news.ycombinator.com/item?id=42962394 [20] Jon Bentley, Don Knuth, and Doug McIlroy. 1986. Programming pearls: a literate program. Commun. ACM 29, 6 (June 1986), 471–483. https://doi.org/10.1145/5948.315654 --[ PREV | HOME | NEXT ]--