Some time ago, the creator of Lua programming language, Roberto Ierusalimschy, visited our Moscow office. We asked him some questions that we prepared with the participation of Habr.com users also. And finally, we’d like to share full-text version of this interview.

— Let’s start with some philosophical matters. Imagine, if you recreated Lua from scratch, which three things would you change in Lua?

— Wow! That’s a difficult question. There’s so much history embedded in the creation and development of the language. It was not like a big decision at once. There are some regrets, several of which I had a chance to correct over the years. People complain about that all the time because of compatibility. We did it several times. I’m only thinking of small things.

— Global-by-default? Do you think this is the way?

— Maybe. But it’s very difficult for dynamic languages. Maybe the solution will be to have no defaults at all, but would be hard to use variables then.

For instance, you would have to somehow declare all the standard libraries. You want a one-liner, print(sin(x)), and then you’ll have to declare ‘print’ and also declare ‘sin’. So it’s kinda strange to have declarations for that kind of very short scripts.

Anything larger should have no defaults, I think. Local-by-default is not the solution, it does not exist. It’s only for assignments, not for usage. Something we assign, and then we use and then assign, and there’s some error — completely mystifying.

Maybe global-by-default is not perfect, but for sure local-by-default is not a solution. I think some kind of declaration, maybe optional declaration… We had this proposal a lot of times — some kind of global declaration. But in the end, I think the problem is that people start asking for more and more and we give up.

(sarcastically) Yes, we are going to put some global declaration — add that and that and that, put that out, and in the end we understand the final conclusion will not satisfy most people and we will not put all the options everybody wants, so we don’t put anything. In the end, strict mode is a reasonable compromise.

There is this problem: more often than not we’re using fields inside the modules for instance, then you have the same problems again. It’s just one very specific case of mistakes the general solution should probably include. So I think if you really want that, you should use a statically typed language.

— Global-by-default is also nice for small configuration files.

— Yes, exactly, for small scripts and so on.

— No tradeoffs here?

— No, there are always tradeoffs. There’s a tradeoff between small scripts and real programs or something like that.

— So, we’re getting back to the first big question: three things you would change if you had the chance. As I see it, you are quite happy with what we have now, is that right?

— Well, it’s not a big change, but still… Our bad debt that became a big change is nils in tables. It’s something I really regret. I did that kind of implementation, a kind of hack… Did you see what I did? I sent a version of Lua about six months or a year ago that had nils in tables.

— Nil values?

— Exactly. I think it was called nils in tables — what’s called null. We did some hack in the grammar to make it somewhat compatible.

— Why is it needed?

— I’m really convinced that this is a whole problem of holes… I think that most problems of nils in arrays would disappear, if we could have [nils in tables]… Because the exact problem is not nils in arrays. People say we can’t have nils in arrays, so we should have arrays separated from tables. But the real problem is that we can’t have nils in tables! So the problem is with the tables, not the way we represent arrays. If we could have nils in tables, then we would have nils in arrays without anything else. So this is something I really regret, and many people don’t understand how things would change if Lua allowed nils in tables.

— May I tell you a story about Tarantool? We actually have our own implementation of null, which is a CDATA to a null-pointer. We use it where gaps in memory are required. To fill positional arguments when we make remote calls and so on. But we usually suffer from it because CDATA is always converted to ‘true’. So nils in arrays would solve a lot of our problems.

— Yeah, I know. That’s exactly my point — this would solve a lot of problems for a lot of people, but there’s a big problem of compatibility. We don’t have the energy to release a version that is so incompatible and then break the community and have different documentation for Lua 5 and Lua 6 etc. But maybe one day we’ll release it. But it’s a really big change. I think it should have been like that since the beginning — if it was, it would be a trivial change in the language, except for compatibility. It breaks a lot of programs, in very subtle ways.

— What are the downsides except for compatibility?

— Besides compatibility, the downside is that we would need two new operations, two new functions. Like ‘delete key’, because assigning nil would not delete the key, so we would have a kind of primitive operation to delete the key and really remove it from the table. And ‘test’ to check where exactly to distinguish between nil and absent. So we need two primitive functions.

— Have you analyzed the impact of this on real implementations?

— Yes, we released a version of Lua with that. And as I said, it breaks code in many subtle ways. There are people who do table.insert(f(x)) — a call to a function. And it’s on purpose, it’s by design that when a function doesn’t want to insert anything, it returns nil. So instead of a separate check «do I want to insert?», then I call a table.insert, and knowing that if it’s nil, it won’t be inserted. As everything in every language, a bug becomes a feature, and people use the feature — but if you change it, you break the code.

— What about a new void type? Like nil, but void?

— Oh no, this is a nightmare. You just postpone the problem, if you put another, then you need another and another and another. That’s not the solution. The main problem — well, not main, but one of the problems — is that nil is already ingrained in a lot of places in the language. For instance, a very typical example. We say: you should avoid nils in arrays, holes. But then we have functions that return nil and something after nil, so we get an error code. So that construction itself assumes what nil represents… For instance, if I want to make a list of returns of that function, just to capture all of these returns.

— That’s why you have a hack for that. :)

— Exactly, but you don’t have to use hacks for so primitive and obvious [issue]. But the way the libraries are built… I once thought of that — maybe the libraries should return false instead of nil — but it’s a half-cooked solution, it solves only a small part of the problem. The real problem, as I said, is that we should have nils in tables. If not, maybe we should not use nils as frequently as we do now. It’s all kinda messy. So if you create a void, these functions would still return a nil, and we’d still have this problem unless we create a new type and the functions would return void instead of nil.

— Void could be used to explicitly tell that the key should be kept in a table — key with a void value. And nil can act as before.

— Yes, that’s what I mean. All the functions in the libraries should return void or nil.

— They can still return nil, why not?

— Because we’d still have the problem that you cannot capture some functions.

— But there won’t be a first key, only a second key.

— No, there won’t be a second key, because the counting will be wrong and you’ll have a hole in the array.

— Yes, so are you saying that you need a false metamethod?

— Yes. My dream is something like that:

{f(x)}

You should capture all returns of the function f(x). And then I can do %x or #x, and that will give me the number of returns of the function. That’s what a reasonable language should do. So creating a void will not solve that, unless we had a very strong rule that functions should never return nil, but then why do we need nil? Maybe we should avoid it.

— Roberto, will there be a much stronger static analysis support for Lua? Like «Lua Check on steroids». I know it won’t solve all the problems, of course. You’re saying this is a feature for 6.0, if ever, right? So if in a 5.x there will be a strong static analysis tool — if man-hours and man-years were invested — would it really help?

— No, I think a really strong static analysis tool is called… type system! If you want a really strong tool you should use a statically typed language, something like Haskell or even something with dependent types. Then you’ll have really strong analysis tools.

— But then you don’t have Lua.

— Exactly, Lua is for…

— Imprecise? I really enjoyed your giraffe picture on static and dynamic types.

— Yes, my last slide.

The final slide from Roberto Ierusalimschy's talk «Why (and why not) Lua?»
at Lua in Moscow 2019 conference

— For our next prepared question, let’s return to that picture. If I got it right, your position is that Lua is a small nice handy tool for solving not very large tasks.

— No, I think you can do some large tasks, but not with static analysis. I strongly believe in tests. By the way, I disagree with you on coverage, your opinion is we should not chase coverage… I mean, I fully agree that coverage does not imply full test, but non-coverage implies a zero percent test. So I gave a talk about a testing room — you were there in Stockholm. So I started my test with [a few] bugs — that’s the strangest thing — one of them was famous, the other was completely non-famous. It’s something completely broken in a header file from Microsoft, C and C++. So I search the web and nobody cares about it or even noticed it.

For instance, there’s a mathematical function, modf(), where you have to pass a pointer to a double because it returns two doubles. We translate the integer part of the number or the fractional part. So this is a part of a standard library for a long time now. Then came C 99, and you need this function for floats. And the header file from Microsoft simply kept this function and declared another one as a macro. So it got this one into type casts. So it cast the double to float, ok, and then it cast the pointer to double for pointer to float!

— Something is wrong in this picture.

— This is a header file from Visual C++ and Visual C 2007. I mean, if you called this function once, with any parameters, and checked the results — it would be wrong unless it’s zero. Otherwise, any other value will be wrong. You would never ever use this function. Zero coverage. And then there’s a lot of discussions about testing… I mean, just call a function once, check the results! So it’s there, it’s been there for a long time, for many years nobody cared. One very famous was in Apple. Something like "if… what… goto… ok", it was something like that. Someone put another statement here. And then everything was going to ok. And there was a lot of discussions that you should have the rules, the brackets should be mandatory in your style, etc., etc. Nobody mentioned that there are a lot of other ifs here. That has never been executed…

— There’s also a security problem as far as I remember.

— Yes, exactly. Because they were only testing approved cases. They were not testing anything, because everything would be approved. It means there is not a single test case in the security application that checks whether it refuses some connection or whatever it is that it should refuse. So everyone discuss and say they should have brackets… They should have tests, minimum tests! Because nobody has ever tested that, that’s what I mean by coverage. It’s unbelievable how people don’t do basic tests. Because if they were doing all basic tests, then of course, it’s a nightmare to do all the coverage and execute all the lines, etc. People neglect even basic tests, so coverage is at least about the minimum. It is a way to call the attention to some parts of the program that you forgot about. It is a kind of guide on how to improve your tests a little.

— What’s test coverage in Tarantool? 83%! Roberto, what’s Lua test coverage?

— About 99.6. How many lines of code do you have? A million, hundreds of thousands? These are huge numbers. One percent of hundred thousand is a thousand lines of code that were never tested. You did not execute it at all. Your users don’t test anything.

— So there are like 17 percent of Tarantool features that are not currently used?

— I’m not sure if you want to unstack everything back to where we were… I think one of the problems with dynamic languages (and static languages for that matter) is that people don’t test stuff. Even if you have a static language, unless you have something — not even like Haskell, but Coq, — some proof system, you change that for that or that. No static analysis tool can catch these errors, so you do need tests. And if you have the tests, you detect global problems, rename misspellings, etc. All these kinds of errors. You should have these tests anyway, maybe sometimes it’s a little bit more difficult to debug, sometimes it’s not — depends on the language and the kind of bug. But the problem is that no static analysis tool can allow you to avoid tests. The tests, on the other hand… well, they never prove the absence of error, but I feel much more secure after all the tests.

— We have a question about testing Lua modules. As a developer, I want to test some local functions which may be used later. The question is: we want to have a coverage of about 99 percent, but for the API this module produces, the number of functional cases it should produce is much lower than the functionality it supports internally.

— Why is that, sorry?

— There is some functionality which is not reachable by the public interface.

— If there is functionality that is not reachable by the public interface, it shouldn’t be there, just erase it. Erase that code.

— Just kill it?

— Yes, sometimes I do that in Lua. There was some code coverage, I couldn’t get there or there or there, so I thought it was impossible and just removed the code. It’s not that common, but happened more than once. Those cases were impossible to happen, you just put an assertion to comment on why it cannot happen. If you cannot get inside your functions from the public API, it shouldn’t be there. We should code the public API with incorrect input, that’s essential for the tests.

— Remove code, removal is good, it reduces complexity. Reduced complexity increases maintainability and stability. Keep it simple.

— Yes, extreme programming had this rule. If it’s not in a test, then it doesn’t exist.

— What languages inspired you when you created Lua? Which paradigms or functional specialties or parts of these languages did you like?

— I designed Lua for a very specific purpose, it was not an academic project. That’s why when you ask me if I’d create it again, I say there’s lots of historical stuff on the language. I did not start with ‘Let me create the language I want or want to use or everybody needs etc. My problem was ‘This program here needs a configuration language for geologists and engineers, and I need to create some small language they could use with an easy interface. That’s why the API was always an integral part of the language, because it’s easier to be integrated. That was the goal. What I had in my background, it’s a lot of different languages at that time… about ten. If you want all of the background…

— I was interested in languages that you wanted to include in Lua.

— I was getting things from many different languages, whatever fitted the problem I had. The single biggest inspiration was the Modula language for syntax, but otherwise, it’s difficult to say because there are so many languages. Some stuff came from AWK, it was another small inspiration. Of course, Scheme and Lisp… I was always fascinated with Lisp since I started programming.

— And still no macros in Lua!

— Yes, there is much difference in syntax. Fortran, I think, was the first language… no, the first language I learned was Assembly, then came Fortran. I studied, but never used CLU. I did a lot of programming with Smalltalk, SNOBOL. I also studied, but never used Icon, it’s also very interesting. A lot came from Pascal and C. At the time I created Lua, C++ was already too complex for me — and that was before the templates, etc. It was 1991, and in 1993 Lua was started.

— The Soviet Union fell and you started creating Lua. :) Were you bored with semicolons and objects when you started working on Lua? I would expect that Lua would have a similar syntax to C, because it is integrated to C. But…

— Yes, I think it’s a good reason not to have similar syntax — so you don’t mix them, these are two different languages.

It’s something really funny and it’s connected to the answer you didn’t allow me [at the conference] to give on arrays starting at 1. My answer was too long.

When we started Lua, the world was different, not everything was C-like. Java and JavaScript did not exist, Python was in an infancy and had a lower than 1.0 version. So there was not this thing when all the languages are supposed to be C-like. C was just one of many syntaxes around.

And the arrays were exactly the same. It’s very funny that most people don’t realize that. There are good things about zero-based arrays as well as one-based arrays.

The fact is that most popular languages today are zero-based because of C. They were kind of inspired by C. And the funny thing is that C doesn’t have indexing. So you can’t say that C indexes arrays from zero, because there is no indexing operation. C has pointer arithmetic, so zero in C is not an index, it’s an offset. And as an offset, it must be a zero — not because it has better mathematical properties or because it’s more natural, whatever.

And all those languages that copied C, they do have indexes and don’t have pointer arithmetic. Java, JavaScript, etc., etc. — none of them have pointer arithmetic. So they just copied the zero, but it’s a completely different operation. They put zero for no reason at all — it’s like a cargo cult.

— You’re saying it’s logical if you have a language embedded in C to make it with C-like syntax. But if you have a C-embedded language, I assume you have C programmers who want the code to be in C and not some other language, which looks like C, but isn’t C. So Lua users were never supposed to use C daily? Why?

— Who uses C every day?

— System programmers.

— Exactly. That’s the problem, too many people use C, but should not be allowed to use it. Programmers ought to be certified to use C. Why is software so broken? All those hacks invading the world, all those security problems. At least half of them is because of C. It is really hard to program in C.

— But Lua is in C.

— Yes, and that’s how we learned how hard it is to program in C. You have buffer overflows, you have integer overflows that cause buffer overflows… Just get a single C program that you can be sure that no arithmetic goes wrong if people put any number anywhere and everything is checked. Then again, real portability issues — maybe sometimes in one CPU it works, but then it gets to the other CPU… It’s crazy.

For instance, very recently we had a problem. How do you know your C program does not do stack overflow? I mean stack depth, not stack overflow because you invaded… How many calls you have a right to do in a C program?

— Depends on a stack size.

— Exactly. What the standard says about that? If you code in C and then you do this function that calls this function that calls this function… how many calls can you do?

— 16 thousand?

— I may be wrong, but I think the standard says nothing about that.

— I think there is nothing in the standard because it’s too dependent on the size.

— Of course, it depends on the size of each function. It may be huge, automatic arrays in the function frame… So the standard says nothing and there is no way to check whether a call will be valid. So you may have a single problem if you have three step calls, it can crash and still be a valid C program. Correct according to the standard — though it’s not correct because it crashes. So it’s very hard to program in C, because there are so many… Another good example: what is the result when you subtract two pointers? No one here works with C?

— No, so don’t grill them. But C++ supports different types.

— No, C++ has the same problem.

— What’s the type of declaration? ptrdiff_t?

— Exactly, ptrdiff_t is a signed type. So typically, if you have a standard memory the size of your word and you subtract two pointers in this space, you cannot represent all the sizes in the signed type. So, what does the standard say about that?

When you subtract two pointers, if the answer fits in a pointer diff, then that is the answer. Otherwise, you have undefined behavior. And how do you know if it fits? You don’t. So whenever you subtract two pointers, usually you know that’s out of standard, that if you’re pointing to anything larger than at least 2 bytes, then the larger size would be half the size of the memory, so everything is ok.

So you’re only having a problem if you’re pointing to bytes or characters. But when you do that, you have a real problem, you can’t do pointer arithmetic without worrying that you have a string larger than half of the memory. And then I can’t just compute the size and store in a pointer diff type because it’s wrong.

That’s what I mean about having a secure C or C++ program that’s really safe.

— Have you considered implementing Lua in a different language? Change it from C to something else?

— When we started, I considered C++, but as I said I gave up using it because of complexity — I cannot learn the whole language. It should be useful to have some stuff from C++ but… even today I don’t see any language that would do.

— Can you explain why?

— Because I have no alternatives. I can only explain why against other languages. I’m not saying C is perfect or even good, but it’s the best. To explain why, I need to compare it with other languages.

— What about JVM?

— Oh, JVM. Come on, it doesn’t fit in half the hardware… Portability is the main reason, but performance too. In JVM it’s a little better than .NET, but it’s not that different. A lot of things that Lua does we can’t do with JVM. You cannot control the garbage collector for instance. You have to use JVM garbage collector because you can’t have a different garbage collector implemented on top of JVM. JVM is also a huge consumer of memory. When any Java program starts to say hello, it’s like 10 MB or so. Portability is an issue not because it wasn’t ported, but because it cannot be ported.

— What about JVM modifications like Mobile JVM?

— That’s not JVM, that’s a joke. It’s like a micro edition of Java, not Java.

— How about other static languages like Go or Oberon? Could they be the basis for Lua if you created it today?

— Oberon… might be, it depends… Go, again, has a garbage collector and has a runtime too big for Lua. Oberon would be an option, but Oberon has some very strange things, like you almost don’t have constants, if I recall correctly. Yeah, I think they removed const from Pascal to Oberon. I had a book on Oberon and loved Oberon. Its system was unbelievable, it’s really something.

I remember that in 1994 I saw a demonstration of Oberon and Self. You know Self? It’s a very interesting dynamic language with jit-compilers etc… I saw these demos a week apart, and Self was very smart, they used some techniques from cartoons to disguise the slowness of the operations. Because when you opened something, it was like ‘woop!’ — first it reduces a little, then expands with some effects. It was implemented very well, these techniques they used to simulate movement…

Then a week later we saw a demo of Oberon, it was running on like 1/10^th of hardware for Self — there was this very old small machine. In Oberon you click and then just boom, everything works immediately, the whole system was so light.

But for me it’s too minimalistic, they removed constants and variant types.

— Haskell?

— I don’t know Haskell or how to implement Lua in Haskell.

— And what’s your attitude to languages like Python or R or Julia as a basis for future implementations of Lua?

— I think every one of these has its uses.

R seems to be good for statistics. It’s very domain specific, done by people in the area, so this is a strength.

Python is nice, but I had personal problems with it. I thought I mentioned it in my talk or the interview. That thing about not knowing the whole language or not using it, the subset fallacy.

We use Python in our courses, teaching basic programming — just a small part, loops and integers. Everybody was happy, and then they said it would be nice to have some graphical applications, so we needed some graphical library. And almost all graphical libraries, you get the API… But I don’t know Python enough, this is much-advanced stuff. It has the illusion it’s easy and I have all these libraries for everything, but it’s either easy or you have everything.

So when you start using the language, then you start: oh, I have to learn OOP, inheritance, whatever else. Every single library. It looks like authors take pride in using more advanced language features in their API to show I don’t know what. Function calls, standard types, etc. You have this object, and then if you want another thing then you have to create another object…

Even the pattern matching, you can do some simple stuff, but usually the standard pattern matching is not something you do. You do a matching, an object returns a result and then you call methods on that object result to get the real result of the match. Sometimes there is a simpler way to use but it’s not obvious, it’s not the way most people use.

Another example: I was teaching a course on pattern matching and wanted to use Perl-like syntax, and I couldn’t use Lua because of a completely different syntax. So I thought Python would be the perfect example. But in Python there are some direct functions for some basic stuff but for anything more complex you’d have to know objects and methods etc. I just wanted to do something and have the result.

— What did you end up using?

— I used Python and explained to them. But even Perl is much simpler, you do the match and the results are $1, $2, $3, it’s much easier, but I don’t have the courage to use Perl, so…

— I was using Python for two years before I noticed there were decorators. (question by Yaroslav Dynnikov from Tarantool team)

— Yes, and when you want to use a library, then you have to learn this stuff and you don’t understand API etc. Python gives an illusion that it’s easy but it’s quite complex.

...And Julia, I don’t know much about Julia, but it reminded me of LuaJIT in the sense that sometimes it looks like user’s pride. You can have very good results but you have to really understand what’s going on. It’s not like you write code and get good results. No, you write code and sometimes the results are good, sometimes they are horrible. And when the results are horrible, you have a lot of good tools that show you the intermediate language that was once generated, you check it and then you go through all this almost assembly code. Then you realize: oh, it’s not optimizing that because of that. That’s the problem of programmers, they like games and sometimes they like stuff because it’s difficult, not because it’s easy.

I don’t know much about Julia, but I once saw a talk about it. And the guy talking, he was the one to have this point of view: see how nice it is, we wrote this program and it’s perfect. I don’t remember much, something about matrix multiplication I guess. And then the floats are perfect, then the doubles are perfect, and then they put complex [numbers]… and it was a tragedy. Like a hundred times slower.

(sarcastically) ‘See how nice it is, we have this tool, we can see the whole assembly [listing], and then you go and change that and that and that. See how efficient this is’. Yes, I see, I can program in assembly directly.

But that was just one talk. I studied a little R and have some user experience with Python for small stuff.

— What do you think of Erlang?

— Erlang is a funny language. It has some really good uses, fault tolerance is really interesting. But they claim it’s a functional language and the whole idea of the functional language is that you don’t have a state.

And Erlang has a huge hidden state in the messages that are sent and not yet received. So each little process is completely functional but the program itself is completely non-functional.

It’s a mess of hidden data that is much worse than global variables because if it were global variables, you would print them. Messages that are the real state of your system. Every single moment, what’s the state of the system? There are all these messages sent here and there. It’s completely non-functional, at all.

— So Erlang lies about being functional and Python lies about being simple. What does Lua lie about?

— Lua lies a bit about being small. It’s still smaller than most other languages, but if you want a really small language then Lua is larger than you want it to be.

— What’s a small language then?

— Forth is, I love Forth.

— Is there a room for a smaller version of Lua?

— Maybe, but it’s difficult. I love tables but tables are not very small. If you want to represent small stuff, the whole idea behind tables will not suit you. It would be syntax of Lua, we’d call it Lua but it’s not Lua.

It would be just like Java micro edition. You call it Java but does it have multi-threading? No, it doesn’t. Does It have a reflection? No, it doesn’t. So why use it? It has a syntax of Java, the same type system but it’s not Java at all. It’s a different language that is easier to learn if you know Java but it’s not Java.

If you want to make a small language that looks like Lua but Lua without tables is not… Probably you should have to declare tables, something like FFI to be able to be small.

— Are there any smaller adaptations of Lua?

— Maybe, I don’t know.

— Is Lua ready for pure functional programming? Can you do it with Lua?

— Of course, you can. It’s not particularly efficient but only Haskell is really efficient for that. If you start using monads and stuff like that, create new functions, compose functions etc… You can do [that] with Lua, it runs quite reasonably, but you need implementation techniques different from normal imperative languages to do something really efficient.

— Actually, there is a library for functional programming in Lua.

— Yes, it’s reasonable and usable, if you do really need performance; you can do a lot of stuff with it. I love functional stuff and I do it all the time.

— My question is more about the garbage collector, because we only have only mutable objects and we have to use them efficiently. Will Lua be good for that?

— I think a new incarnation of garbage collector will help a lot, but again…

— Young die young? The one that seems to work with young objects?

— Exactly, yes. But as I said even with the standard garbage collector we don’t have optimal performance but it can be reasonable. More often you don’t even need that performance for most actions unless you are writing servers and having big operations.

— What functional programming tasks do you perform in Lua?

— A simple example. My book, I’m writing my own format and I have a formatter that transforms that in LaTex or DocBook. It’s completely functional, it has a big pattern matching… It’s slightly inspired by LaTex but much more uniformed. There’s @ symbol instead of backslash, a name of a macro and one single argument in curly brackets. So I have gsub that recognizes this kind of stuff and then it calls a function, the function does something and returns something. It’s all functional, just functions on top of functions on top of functions, and the final function gives a big result.

— Why don’t you program with LaTeX?

— Plain LaTeX? First, it’s too tricky for a lot of stuff and so difficult. I have several things that I don’t know how to do in LaTex. For example, I want to put a piece of inline code inside a text. Then there is a slash verb, standard stuff. But slash verb gives fixed space. And the space between stuff is never right. All real spaces are variable, it depends on how the line is adjusted, so it expands in some spaces and compacts in others depending on a lot of stuff. And those spaces are fixed, so sometimes they look too large, sometimes too small. It also depends on what you put in code.

— But you still render your own format to LaTeX?

— Yes, but with a lot of preprocessing. I write my own verb but then it changes and becomes not a verb but a lot of stuff. For example, when I write 3+1 I write a very small space here. In verb, if I don’t put any space here, it shrinks, and if I do, it’s too large. So I do the preprocessing, inserting a variable space. It’s very small but can be a little larger if it needs to adjust. But if I put ‘and’ after 1 then I put a larger space. This function here does all that. This is a small example but there are other things…

— Do you have a source?

— I do have the source, it is in the git. The program’s called 2html. The current version only generates HTML… Sorry, that’s a kind of a mess. I created it for a book but also another one for the manual. The one in the git is for the manual. But the other one is more complicated and not public, I can’t make it public. But the main problem is that TeX is not there. It’s almost impossible to process TeX files without TeX itself.

— Is it not machine-readable?

— Yes, it’s not machine-readable. I mean, it is readable because TeX reads it. It’s so hard to test, so many strange rules etc. So this is much more uniformed and as I said I generate DocBook format, sometimes I need it. That started when I had this contract for a book.

— So you use 2html to generate DocBook?

— Yes, it generates DocBook directly.

— Ok, thank you very much for the interview!

If you have any more questions, you can ask them in the Lua Mailing List. See you soon!

Lua in Moscow 2019: Interview with Roberto Ierusalimschy