Yes, code is data, but that's not what makes Lisp cool
blog.rongarret.info
blog.rongarret.info
If you don't like how some Lisp application works, you can frequently just replace the contentious parts as the program is running. If something does not seem right, it is trivial to inspect all parts of the internal state. If you don't understand a variable, you can look up its documentation inside the running program.
It's a fantastic experience.
So, I didn't get a video made. My humble apologies.
Rather than just leave this hanging with no response, I'll still try and get a video while at the office this week. In the meantime, https://www.youtube.com/watch?v=VEcobuYr5wg shows a much more impressive example than what I was planning. This one actually shows "live coding" with a webpage. I was just going to show how the help system and some basic debugging in emacs works.
Try `"demos" run` then click on some of them, `\ each help`, `\ cpu edit` to run some demos, view the help, and edit a word. To reload the code while Factor is running, press f2. (no backticks)
I have done all of this with Java and Intellij
I thought it was very effective.
Most programs I see are text files. If I edit them in an editor, I don't know what data structure represents them, but it might be a doubly linked list of strings, or some other data structure.
If a compiler reads a program, it uses typically a stream of characters.
A string would be something like this: a delimiter, some characters, the delimiter. Example: "a + b". It could also be a string data object in some program.
But programs are not always represented as strings and many are not. Sometimes they are. For example an EVAL procedure may take a string as an argument - then code is represented as a string or passed at runtime as a string object.
When we say in Lisp code is data, we actually mean code is data, but not in the trivial form of text representation formats (text files, strings, ...).
It's an abbreviation. We expect usually not to mention, that it is not of the trivial version. Like when we speak of a picture with lots of color, we don't mention that this excludes pictures with only black, white and shades of grey.
Generally, if I see
a + b
If I see Lisp code in the form of
(+ a b)
I then would require that this is syntactical legal Lisp - THIS requires that in an s-expression, the + operator comes first - that it uses prefix syntax.
Wien we mean code as data, we mean that Lisp uaes an actual and explicit data format to encode its programs: s-expressions. The data format is similar to JSON or XML, in that it has an explicit external representation. It's also similar to JSON, because both have an internal representation in a programming language.
If we would use JSON as a format to encode programs, we would also be able to say, code is data.
S-expressions are slightly more effective as data encoding for programs than JSON, because it uses less delimiters and encodes a rich set of data structures.
So, yes. Code with a non-trivial and simple data-representation makes Lisp cool.
Every programming language uses an "actual" and explicit data format to encode its programs, the valid programs all parse to the corresponding AST.
The only thing unique about Lisp here is a construct to parse the serialization format(s-expressions) and perform actions on it i.e. macros. The fact that parsing s-expressions is so trivial makes it seem like manipulating code as data is somehow unique to Lisp.
where? Show me the data format which is used by C programmers to encode programs.
A data format with an external and internal representation, and both used by the programmer.
An AST of a C compiler isn't it, since the programmer isn't using it. Neither externally as text, nor internally as data. It's only used internally by a compiler - if it does use one. The C programmer does not write code with ASTs, nor does he even manipulate code via ASTs operations.
It's not that some languages have ASTs -> they are simply no neutral data format and the programmer isn't using this to write/read programs. In some languages or implementations ASTs are available to the programmer, but this is after a tool (the compiler) has parsed the program according to the language syntax rules. But that is not what Lisp does, s-expressions are more like a general data format which by the Lisp reader (a tool) is used to translate into a token tree. But the Lisp reader is not a compiler internal stage. It's a general tool, which turns externalized data into internalized data.
ASTs are also a different thing from s-expressions, since s-expressions don't represent abstract syntax of Lisp programs. S-expressions are language agnostic: it does know nothing about the syntax of Lisp constructs and doesn't represent no syntactic information. On the s-expression level (a b c) is no different from (+ 1 2) or (defun foo ()).
The C syntax isn't layered on top of a neutral data format, like Lisp is.
So, what additional power Lisp has as "code as data" vs the language that gives power to manipulate ASTs easily(with quasi-quotation)?
>ASTs are also a different thing from s-expressions, since s-expressions don't represent abstract syntax of Lisp programs. S-expressions are language agnostic: it does know nothing about the syntax of Lisp constructs and doesn't represent no syntactic information.
Irrelevant to the discussion of "code as data". Every language can read s-expr in language agnostic way.
> The C syntax isn't layered on top of a neutral data format, like Lisp is.
I never said that the languages are built on top of a neutral data format, but asking the benefits of being so. Unless i am mistaken, the only benefit is being easy to parse by tools(editors, compilers etc) and with quasi-quotation the power of macros.
From a point of Lisp code manipulation, quasi-quotation does not reach far and thus the convenience it provides is only shallow. It's useful to have, but its main purpose is to build up code based on some template mechanism. There are other things which are also important, like destructuring of code, which a typical macro mechanism will support in some way.
Another point: to have an AST you need to have the code parsed already. The AST is an output of a parser stage. That means you are bound to the syntax this parser implements and this AST can represent - or you need to be able to feed this parser with different syntax definitions and have it runtime extensible. Since Lisp macros see only code as internal data, the input format is much less limited. It can also transform arbitrary data into code. Or you need to be able to deal in some other with code which does not conform to what the parser expects.
> Irrelevant to the discussion of "code as data". Every language can read s-expr in language agnostic way.
'can' vs. usually done is a huge practical difference. s-expressions are not a standard input format for C compiler. I have not seen s-expressions or a similar data format in the C language standard used to represent code. I'd would be interested to see a reference to it, if you have one.
> I never said that the languages are built on top of a neutral data format,
But Lisp is.
> Unless i am mistaken, the only benefit is being easy to parse by tools(editors, compilers etc) and with quasi-quotation the power of macros.
Quasi-quotation is only a convenience feature - which is great, but they don't add any power to the language. Lisp can create the same effects without quasi-quotation.
The benefits are:
* operations on the data format can be supported in an text editor and then can be applied to both data and code
* the code structure is not limited to what syntax a programming language parser implements or what an AST can represent
* the manipulation of internal code as data is convenient and simple for many cases, since the typical operations can be applied
* the internal format used by Lisp has a default external format, but can also be used with different external formats
* the programming language implementation exposes this internal data in various stages to the user, which it then can conveniently use it. For example macros see this data and can transform it with the usual tools. Interpreters use it and can expose it to the developer. The compiler can use it as input. The pretty printer takes it as input to create textual representations.
The main point is that it is convenient to use. Thus code transformations and code manipulations of s-expressions are both common and simple in Lisp.
The effect is that user-written code manipulation is widely used in Lisp - even at runtime, where for example in C it is typically done in external tools in preprocessor stages or hidden from the user at compile time.
Destructuring doesn't require it to be a list/s-expr. You can destruct ASTs in any language that supports pattern-matching.
> or you need to be able to feed this parser with different syntax definitions and have it runtime extensible.
That's what languages should aim to do. For example rust's procedural macros can transform a TokenStream to another TokenStream and doesn't have to satisfy Rust language parser. Although that doesn't extend to lexing.
> 'can' vs. usually done is a huge practical difference. s-expressions are not a standard input format for C compiler
I am not talking about a specific language compiler, but modern languages are heading in the right direction without using s-expr.
> Quasi-quotation is only a convenience feature - which is great, but they don't add any power to the language. Lisp can create the same effects without quasi-quotation.
Yes, but then you write your own macro-expander.
> operations on the data format can be supported in an text editor and then can be applied to both data and code
Yes, but they will not be very helpful without additional tooling.
> the code structure is not limited to what syntax a programming language parser implements or what an AST can represent
And you have lost the previous point, now your "code as data" is not so well-supported.
> the programming language implementation exposes this internal data in various stages to the user, which it then can conveniently use it. For example macros see this data and can transform it with the usual tools. Interpreters use it and can expose it to the developer. The compiler can use it as input. The pretty printer takes it as input to create textual representations.
True, but these all can be provided without s-expressions, the compiler just has to be designed properly into reusable modules.
Rest of your points just suggest that the tooling part is easier and simpler to do and understand, which ofcourse i agree with. But not all programmers prefer Lisp like syntax and most languages haven't really cared about it or see as undesired(macros bring some undesirable properties as languages like loop pop-up) flexibility.
It's not about requiring. It's provided in Lisp. Macros already provide destructuring in their arglists.
> That's what languages should aim to do. For example rust's procedural macros can transform a TokenStream to another TokenStream and doesn't have to satisfy Rust language parser. Although that doesn't extend to lexing.
A tokenizer is not a language parser.
> I am not talking about a specific language compiler, but modern languages are heading in the right direction without using s-expr.
That's nothing new or modern. Code transformations and code representation has been used for decades.
> Yes, but then you write your own macro-expander.
No, quasi-quotations have nothing to do with a macro-expander. It's a general feature in Lisp - independent of macros. One alternative is to procedurally build up code.
There is no functionality difference between
`(a ,a)
(list 'a a)
>> operations on the data format can be supported in an text editor and then can be applied to both data and code
> Yes, but they will not be very helpful without additional tooling.
There is nothing wrong with tool building.
>> the code structure is not limited to what syntax a programming language parser implements or what an AST can represent
> And you have lost the previous point, now your "code as data" is not so well-supported.
But I gain a lot of flexibility.
> True, but these all can be provided without s-expressions, the compiler just has to be designed properly into reusable modules.
That's what Lisp already provides.
> Rest of your points just suggest that the tooling part is easier and simpler to do and understand, which ofcourse i agree with. But not all programmers prefer Lisp like syntax and most languages haven't really cared about it or see as undesired(macros bring some undesirable properties as languages like loop pop-up) flexibility.
All true.
It's not a tokenizer it can transform a token stream to another token stream with a given function which can be some language parser/just rust code producing desired rust ast.
> No, quasi-quotations have nothing to do with a macro-expander.
My bad, i was confusing it with mixing evaluation.
Almost every year I come crawling back to emacs and I start putting the pieces back together (first, by hand, then Spacemacs a couple of times, then by hand again but in a different way). It has such an incredible allure and it's a joy to write. I don't know what else I would write in Lisp right now so I just abandon my config and rebuild it to get that feeling again.
Small critique on presentation. The title clearly indicates there are two parts to this, a brief exploration on "code is data", followed by a dip into what makes lisp cool. Having some visual indicator saying "we are now at the cool part" would be good.
Or, perhaps, you're saying that it's not just homoiconicity. It's that it's the right homoiconicity - lists. Lists are well-suited for being processed as data in Lisp (kind of the point of Lisp, really), and lists are great for representing code (abstract syntax trees), and lists are really simple to parse if they're represented cleanly.
That is exactly right.
I disagree with this statement. Support for your argument seems to stem from the other statement you then make: "What makes Lisp cool is not that programs are data (because all programs are data), but that they are a particular kind of data.".
Personally I think it's not a good argument, for the following reason: I would guess that 99% of the people who argue that "code is data" is what makes Lisp cool, mean exactly that -- that code is expressed as lists of lists (and they are thinking of all the nice implications of that).
So effectively, you are saying that the statement is invalid because it could mean something else, but almost all of the people who make that statement are making the exact same argument you are for why Lisp is cool.
I opened your article expecting to find an explanation of why some other feature or set of features of Lisp was way more important/useful than just code being written as lists of lists, only to find that you were arguing for the importance of that exact feature.
So while it is worth pointing out that having code as a list of lists has awesome implications, I guess I would have gone about discussing in a different way.
"code is data" has a specific meaning to Lisp programmers, and it is completely valid.
https://news.ycombinator.com/item?id=16388578
and the subsequent replies. I was surprised that this message resonated. I thought it was obvious, but it's obviously not.
> I would guess that 99% of the people who argue that "code is data" is what makes Lisp cool, mean exactly that -- that code is expressed as lists of lists
But "code is data" and "code is expressed as lists of lists" are not the same statement. And the difference matters.
Except that when a Lisp programmer says "code is data" that's exactly what they mean. Lisp programs are lists of Lisp lists (say that 5 times fast xD). This gives Lisp code generation powers that leave any string-based code manipulation in the dust (I know you already know this).
I'm not against making the clarification, I just disagree with the statement "even the most casual reflection on what this slogan actually means will reveal that it can't possibly be the right answer.".
I realize that some may not understand what is meant when a Lisp programmer says "code is data", and it is worth explaining what they mean when they say it. I just disagree that that makes the expression invalid for Lisp programmers to use in that manner.
That's actually not true either. Here is a Lisp program:
(defun true () t)
It's a list of four elements, only one of which is a list. (You actually could define a language where all programs were lists of lists, and you could even make it Turing-complete. But it would be completely unusable because it would consist of nothing but parentheses!)
Now, obviously I know you know this. And equally obviously I know that what you meant was different from what you actually said. But that's only because I already knew what you were trying to say. If you're trying to communicate with someone who doesn't already know what you're trying to say (and isn't that the whole point of communication?) then you have to actually say what you mean or you will fail.
No, it's !@#$%^& not.
This pisses me off every single time I see it.
Parsing s-expressions requires on-the-fly allocation of memory and something resembling recursion.
If you can't allocate memory--good luck.
If you don't have recursion in your language, you're going to have to build it yourself--good luck.
And god help you if you have dot notation--(a b) is a tree of (a . (b . null)) while (a . b) is an actual pair.
Clojure shooting the idea of building things out of cons pairs and using lists composed of arrays is one of its most important design choices.
Now, defun/define/etc will require more memory management than people would care to admit to. However, just parsing the s-expression is dead simple. Recursion will make it easier, but I'm actually not sure it is required here. For the most part, just keeping a manual stack of what you're dealing with would be easy enough and won't take much effort.
Again I'll stress, though; that anyone that thinks implementing a lisp is dead easy, is likely overstating how easy it can be done. Easier than you could think. Not guaranteed easier than you think.
Parsing any code requires on-the-fly allocation of memory because your function can contain 10 or 100 or 1000 lines of code and, when you are at the beginning, you don't know how long the whole function is.
Parsing any code requires something resembling recursion because your function can contain 3 or 10 or 20 nested scopes and, when you are at the beginning, you don't know how deep the whole function is.
Of course that you can say that you can pre-read the code and compute how much memory you need beforehand, but that applies to Lisp as much as it applies to any other language.
This is not necessarily true. Parsing a language can be bounded by the grammar. If you don't have indefinitely nested structures (like arithmetic open parentheses), your parsing is linear and you can bound the memory (ie. your parsing memory cannot be longer than your input string, for example).
I believe that Tcl and Microsoft Basic fell into this category. (I believe that Ousterhout used to actually generate code in the input buffer behind the current parsing point precisely because Tcl was designed to do this).
> Parsing any code requires something resembling recursion because your function can contain 3 or 10 or 20 nested scopes and, when you are at the beginning, you don't know how deep the whole function is.
Indefinitely nested lexical scoping is not a requirement for a language. Famously Javascript screws up all manner of programmers because they expect that to be true, but it is not. And many languages start out with a single global scope and grow lexical scoping only after they gain some complexity.
I understand lisp isn't trivial to implement if all you have is assembly with no OS or libraries, but beyond that it's pretty damn easy. Even in C it only takes a few dozen lines.
> Javascript
I'm not a JS programmer, but I think JS has had infinite lexical scope at function-level since the beginning. It does not have scope at block-level, which is the confusing part.
Very nice article.
Anyway, even saying "code is linked list" does not really cut it as you can represent 'ordinary' programming language program as a list of tokens. What makes LISP special is that code is a first class citizen and the mapping between code you edit and the programs run-time memory representation is trivial. Combined, you get this special feeling that "code is data".
When you get into the LISP interpreter itself, which is just a linked-list structure describing how to process linked-list structures (like itself) as code, well... Here's William Byrd on "The Most Beautiful Program Ever Written" (ye LISP interpreter) https://www.youtube.com/watch?v=OyfBQmvr2Hc