tlack/xxl: a minimal vector programming language

XXL logo

XXL is a new programming language whose goals are simplicity, power, and speed.
We throw out most of conventional programming language thinking in order to
attain these goals.


Pretty buggy but passes tests. Useful for writing small utilities, for me. Not
yet suitable for real work.

Tested so far on Linux (64bit x86 & GCC or Clang), OS X (Clang), Windows
(Cygwin64), Android under termux, Intel Edison and Raspberry Pi. Should work on iOS as well
but untested.


JSON encoder

About 6 lines of code, without the tests:

json encoder screenshot with syntax highlighting

// enclose (c)urly(b)races, (s)quare(b)brackets, (q)uotes:
'ecb is {"{",x,"}"}; 'esb is {"[",x,"]"}; 'eq is {""",x,"""}; 
'jc is {join ","}; 'jac is {each y jc};  // join x with commas; apply y to each of x then join with commas
'pair is {encode,":",(y encode)};        // key:val pair for dict
'dict is {key as 'k; x val as 'v; [k],v >: pair jc ecb}; // get keys/vals, pair merge, commas, braces

// wrap non-scalar values in appropriate way:
'many is {as 'el type case ('char, {el str eq}, 'dict, {el dict}, {el jac encode esb})};
'encode is {ravel[many,str]};            // ravel calls x y[0] for arrays (len > 1), x y[1] for scalars

Here’s what’s going on in this monster. I’ll explain the parts, and then the sequence of how it comes together. As per the
fashion that we like, the code starts from the simple, base functions, and progresses to the more complex functions that
use them.

ecb/esb/eq enclose their argument in curly braces, square braces, or quotes, respectively. They use the name x
to refer to their argument.

jc (join comma) joins the contents of its argument together with a comma in between each item. jac does the same
thing, but after first calling its right argument (y) on each item – jac here is a mnemonic for “join and call”.

The astute might notice that jc and jac don’t refer tox, because the first verb inside a function will be invoked
automatically with x as the left argument. In small functions, x is almost always the first term in the function’s
code, so being able to omit it results in some expressivity. More on this later.

pair calls another function, encode (which we define later), without referring to x. It then appends a : to the
string that is returned from encode, and then appends that to the result of calling encode with the y argument.
This is a form of recursion.

dict takes a dictionary as an argument, gets its keys as a list using the built-in key verb, and saves them in a
new variable called k. The values of the dictionary (also a list) are extracted using the built-in val verb and
become v.

These two lists then become the left and right arguments for successive calls to jc using the “each pair” or >: verb.
Each of the key/value pairs will get commas between them. Then, we enclose the whole of the dictionary in curly braces,
like {"name":"Tyler","age":"2"}.

All the looping verbs have short names that end in :.

many looks complex, but isn’t. Its purpose is to handle multiple-item collections (called vectors in XXL). In our case, we have to worry about two main ones: character vectors (strings) and dictionaries.

First, we store the value of x as el (element). Then we extract the value’s type
using the type verb, and then use the case verb to decide how to treat all the different types.

If it’s a character vector (a string), we pass it through the str verb to remove its tag (a feature we’ll explain later).

If it’s a dictionary, we call our dict function. If it’s anything else, we recurse by calling encode again,
and then enclose the result in square braces (think arrays of numbers).

And finally the star of the show, encode. ravel is a verb that allows you to take one branch of logic for single-item
values (like the number 3), or a different branch for values that have many items, like an array or string (vectors).
Depending on the type of x, encode will dispatch either many (for multiple values) or str (for simple, single values).

MySQL Slow Query Watcher

Here we use command-line MySQL to get process list, turn into table, find interesting slow queries.

Note that the 0., 1. etc is XXL’s prompt in interactive mode. The number you see next to the
prompt allows you to refer later to that line of input or the result XXL produced for it. Don’t type
this part in.

0. 'slowtime is 2;
// get output, split it by line, then by column, save
1. "mysql -e "show full processlist" --batch" Shell.get split "n" each {split "t"} as 'lines; 
// the first line is column headings, massage data for each row, create table with keys:data
2. lines first each {make '} : (lines behead curtail each {make "issstiss"}) as 'procs;
// Use  except  with an anonymous function to filter rows we don't care about
3. procs except {@'Time

Originally used as a one-liner to fix a performance issue on a site.

Simple Web Server (Counter)

Here's an example web server application that acts as a counter. You can run
this as./c examples/web-ctr.xxl. Source code in full:

0 as 'ctr;
(8080,"localhost") Net.bind {
	ctr+1 as '.ctr;  // dot in front of name means "parent"
	"HTTP/1.0 200 OKrnContent-Type: text/plainrn",
	"Connection: closernServer: xxl v0.0rnrnHello ",
	(.ctr repr),"rn" flat}

(We didn't show the XXL prompt here, just the code itself.)

The first line declares a variable or noun named ctr. Variables are defined
using a tag. After they are defined, you can refer to them without the tag.
Tags start with apostrophe ' and can contain anything in the range [A-Za-z_?].
Names may not contain numbers.

Names starting with . are referenced from the root of the XXL context tree,
rather than being resolved in the locals context to start, which makes them
behave something like globals in other languages.

The second line invokes the built in Net.bind verb, which creates a network socket
and calls the code you supply to handle the request after it comes in.

On the left of the verb come the listening socket configuration options. In this case,
port 8080 on localhost.

On the right side of the Net.bind verb we supply an anonymous function body
to act as the callback to be invoked when a connection is made and a request is
received. This callback is invoked with one argument, which is a list
containing the request body in the first member and the remote IP address in
the second.

At the moment, Net.bind only supports client-speaks-first-style network
protocols, such as HTTP.

Inside the callback, we display the request on the server's screen (x show), increase
and store the counter value, and then generate an HTTP response with a couple of headers.

Since we want to include the counter's current value, we append it to the string after
calling repr on the counter value only (since it's in parenthesis). The resulting list
needs to be flattened so we can spit it out as one long string to the client's socket, so
we call XXL's flat verb, which removes a level of nesting on a general list (akin to
raze in Q).



Here's some useful info, but no real docs yet.

See the various tests for more examples. In particular, you may appreciate
test-logic.h, and test-semantics.h.


  • Terse. It is much better to express yourself clearly in a few words rather
    than many. We provide powerful operators that allow you to do more while
    describing less.

  • Accessible. XXL avoids niche-y programmer terms like "map" and "reduce" in
    favor of common English

  • Minimalistic, but not spare. Provide tools people commonly need, even if
    they're a bit duplicative of other, more primitive approaches.

  • Fast. XXL cares deeply about performance, for the same reason a person should care
    about the quality of any tool they use for professional work.

  • Tangible. Abstractions provided by many systems and platforms make it much harder
    to reason about what your program is doing. XXL just helps you program the computer
    in the way the computer wants to be programmed. Those with a background in assembly
    might see some familiar tones in XXL.

These goals have not yet been attained.


I need a swiss army knife for data manipulation with predictable performance and concise
syntax. I want to use some of the more exotic features I've tasted in other languages,
but not all of them. In particular, I feel like programming has gotten a little too stack-


Minimal, easy to grasp syntax

XXL may seem odd at first glance but it's much simpler than other languages
which suffer from complex grammars, rules, and special cases. You don't have
to understand APL, K/Q or even functional programming to use XXL.

Clean, very easy to understand and parse left-to-right syntax with only three
special forms: comments, strings, and grouping (i.e., ( ), [ ] and { }).

Everything else is either a noun or a verb. Learn the verbs and you know the

Values and variables are called nouns. The first time you create a noun,
you usually refer to it using a tag: 'z is 1. You can use as if you
want to build up an expression and save it: 1,2,3+1 as 'numbers.

If you didn't use a tag name (like 'numbers) and instead just wrote numbers,
XXL would try to find an existing noun named numbers, and insert the contents
of that instead, which probably isn't what you want.

Verbs are either postfix (also called unary: having one variable, called x):
1,2,3 len
or infix (binary: two variables, called x and y)
1,2,3 + (4,5,6).

Verbs work with the things immediately to the left or left and right. There is
no grouping or precedence. This is much easier to keep track of mentally
and allows you to much easily scan code. Use (), [], or {} to group values
if you must, but many times you can write long expressions with few groupings.

There is nothing special about len, or +, or any of the built-in verbs.
They are simply verbs that have been created for you already. You can create
your own system-level verbs or replace ones that already exist.

You can create nouns in the middle of expressions and refer to them later, such
as this in this tedious example:
1,2,3+2 as 'num + num

Agnostic about whitespace. Don't let invisible special characters ruin your day.


We don't yet have a full tree data type yet, but operations on list-of-lists
are pretty diverse. After all, XXL is written in terms of its own verbs and values.

x nest [open,close] creates sublists in x between matching open
and close tags. Create parsers like an animal.

Note: In the examples below, 0., 1., etc are the XXL command line prompt.
Abridged output shown on the line below.

0. 1,4,4,0,5,5,6 nest [4,5]
[1i, [4i, (4,0,5i), 5i], 6i]

x ravel [funmany, funone] calls funone if x is scalar, funmany otherwise.
Pair with self to recurse. A napkin sketch of a markdown parser:

0. 'body is {behead curtail};
1. 'mdtxt is "My *Markdown* parser";
2. 'html is ['b:{"",x,""}]; // define some formatting templates
3. mdtxt nest ["*","*"] :: {ravel [{x body html.b},str]} flat
"My Markdown parser"

The contents of the html callbacks, the nest control parameters, are all just
data that you can build up or pull in from anywhere. Compare with traditional
control structures used to manipulate data which exist inside the source code
only and are not mutable at runtime.

Or use deep and wide for more explicit macro-control of

Erlang-inspired mailboxes

Thread safe for both readers and writers (at some cost to performance of
course). Fast enough to be usable for basic purposes (800 request/reply cycles
a second on a $5 Digital Ocean box).

0. [] as 'myservice;
1. myservice {x show}
2. myservice Mbox.send 55;

watch spawns a thread to act on mailbox messages. The 55 seen here is from
the show statement in the lambda. A y variable is available inside the
callback to maintain state; it starts as [] and it set to whatever your
function returns.

See doc/sect_mbox.xxl
for more.


Well, the beginnings of them anyway.

0. "name,age,jobnBob,30,ProgrammernJane,25,CFOnTyler,2,Dinosaur Hunter" as 'data;
1. data split "n" each {split ","} as 'fields;     // in a fantasy world where csv is never escaped
2. fields first:(fields behead) as 'emp             // : creates dicts or tables 
["name", "age", "job"]:[
["Bob", "30", "Programmer"],
["Jane", "25", "CFO"],
["Tyler", "2", "Dinosaur Hunter"]]

This is similar to the MySQL example above in that we are unpacking data and
making a table out of it.

Notice that the printed representation of the table is similar to what you'd
type in to regenerate it.

The in-memory table works very much like a regular value, and you can join it
with more dictionaries to append rows with , as you'd expect.

3. emp,["name":"Steve","age":"42","job":"Fact Checker"]
["name", "age", "job"]:[
["Bob", "30", "Programmer"],
["Jane", "25", "CFO"],
["Tyler", "2", "Dinosaur Hunter"],
["Steve", "42", "Fact Checker"]]

Or join it with a list that has a value for each column:

4. emp,["Arca","2","Tiny Dog"]
["name", "age", "job"]:[
["Bob", "30", "Programmer"],
["Jane", "25", "CFO"],
["Tyler", "2", "Dinosaur Hunter"],
["Arca", "2", "Tiny Dog"]]

Notice that we did not save the result of expression 3 (where we joined emp
with a dictionary containing Steve's record) so the change was not saved to
emp. Like most verbs in XXL, , does not modify the x argument, it creates
and returns a new value. Looks like Steve is going to have to check someone
else's facts from here on out.

Use amend (or equivalent the short operator !) to update emp in place.
emp![["age"],{::{base 10}}] would convert the ages to numbers so you can
ruthlessly compute them. In this example, "age" has to be put into a list here
because otherwise it would think you want to update three indices - "a", "g",
and "e".

Get individual rows with [email protected] or 0,2 from emp, or rows with "age" from emp or [email protected]"age" or, if the key "age" was a tag like 'age, you could use
simply emp.age.

Search for matching values using match (or ~) or filter rows with
{..} from emp or emp except {..}.

Still no joins, many rough edges, untested performance.


  • Variable names can't contain numbers, so you can build up some pretty clean and
    short expressions that mirror mathematics. 3u b5 means 3 u b 5, or
    b(u(3),5), assuming u is a single argument (unary) function and b is a two argument (binary) function.
    If you're recoiling in horror right now at losing your precious numerals in variable
    names, consider how often you're just poorly naming a temp variable.

  • Variable names can contain ?, so you can name your predicate functions in a
    pleasant manner.

  • Values can have "tags" associated with them, allowing you to create an
    OOP-like concept of structure within data, while all regular operations work
    seamlessly as if you were using the underlying data type. Consider the
    classic OOP example of a "point", which in XXL would just be a tagged int
    vector like 'point#(100,150).

  • Vector-oriented and convenient manipulation on primitive values. For instance,
    1,2,3 * (4,5,6) is perfectly legal and returns 4,10,18. This is a huge time saver
    and eliminates many loops. It's also pretty fast, which somewhat makes up for the
    dangerously slow interpreter.

  • Diverse integer type options, including 128bit octaword (1 make 'octa). At
    this time we don't display the int type when displaying the representation of
    the numeric value. Use the x type verb to discovere what size of int you
    are working with.

  • Fast-enough unboxed binary data files. 25 million int/s to disk per second on $5/mo
    Digital Ocean droplet via 1024*1024*50 count Xd.set "/tmp/50m.xd"

  • Threading support, kinda. XXL has the notion of threads available internally, but
    the only way they are usable from inside XXL code is via mb {..}.
    This is because modifications to contexts' data items are not locked or synchronized,
    So chaos will surely ensue if you modify global data from threads. Thus, the mailbox
    acts as both a safe communication mechanism for threads, and a way to discourage global
    state mutation. See doc/sect_mbox.xxl
    for more.

  • Built in simple networking. Client-speaks-first protocols are a snap to implement.
    World's easiest echo server: [8888,""]Net.bind{"Echo: ",x}. Net.bind creates one
    thread to service each port/service.

  • Vim syntax file in tools/syntax/vim-syntax.vim (which is generated by vim-build.xxl
    in that same directory).

  • No stinkin loops (and no linked lists, either)

  • Supports \ to exit the REPL, as god intended (quit and exit too)

  • BSD license

Not yet implemented

XXL is still very much a work in progress and is mostly broken. That said, here are the
major features I anticipate finishing soon-ish.

  • Currently has severe memory leaks and performance problems
  • Floats! We've got floats now - no comparison tolerance yet tho.
  • Dictionary literals (dictionaries do work and exist as a primitive type, just
    can't decide on a literal syntax for them)

    Settled on and implemented ['a:1,'b:2] by creating : as the make-dict operator
    and making , smart about dicts.
  • JSON
  • FancyRepl(tm)
  • Dates/times (need to figure out core representation and a way to express them as literals
    in code..)
  • In-memory tables Few operations yet, and bugs remain, but the notion of tables has
    slowly seeped in. See tests.
  • ~~Non-pointer tags implementation to avoid logic to enc/decode to binary for disk and IPC~
    Temporarily replaced tag representation with 128bit pointer string. Fast to compare,
    almost as fast as possible to create (memcpy), zero overhead on IPC ingest/excrete,
    not terribly large on today's memory sizes.
  • I/O (files, sockets, mmap)
  • Logged updates (I like Kdb's approach to this)
  • Mailboxes/processes (implemented as a writer-blocks general list) Available in "Mbox"
  • Streams, laziness (perhaps based on mailboxes? studying other systems now)
  • Tail call optimization in functions using the self keyword
    as per Kuc's approach
  • Sorted vectors
  • Grouped/index types
  • Well-supported Apter trees (see also APL's approach).

Well known bugs

Work in progress on these:

  • Interpreter speed and memory leaks. These are at times severe.

  • The interpreter still recurses too much in some scenarios, even though its
    main loop is self-managed on the heap. C call stack depth gets too deep. I
    have a simple plan to resolve this, but it requires a (much needed) rewrite
    to applyexpr(), which can be thought of as the interpreter loop.

  • The join verb (,) is still finnicky about joining like-types of data with
    general lists. In particular, I often find myself a bit puzzled by results
    like ['a,2],3 - should this be a two-item list with two items in the first
    sublist, or a three element list? Before you answer, consider ['a,2] as 'q;q,3. When in doubt, build parts separately and combine.

  • You can't name a variable you define x or y, since the interpreter treats
    those specially. There's no real reason for this other than the idea that
    it's faster to resolve this short literal symbol manually using what we know
    about the calling frame, rather than setting it as a regular variable in the
    dictionary that is used by the context to resolve symbols. The setting part
    works, it's just resolving x afterward doesn't consider the dictionary.

Open questions

  • Date and time stamps: Nanosecond precision from uint64 sufficient? What about literal
    representation? Can we use 1997. or something non-pollutey like that?
    Need to patch up apply() to allow you to index 'day and similar.

Maybe later

  • JS/Emscripten (both in terms of Node.js interop and use of XXL in the browser)
  • GUI

Probably not


Pretty rough right now. You'll need a C compiler and a minimally POSIX
environment but that's about it.

Customize your build

XXL is meant to be tailored to the device you're using it on. You can remove
You can customize the features that XXL includes out of the box.

We use a build script called ./c instead of a Makefile. It tries to make some
guesses about how your system is configured and what options are best to use.
We use a shell script to keep it simple (without having to resort to Autoconf).
When in doubt, you can make your own decisions by editing the c source. It's
very simple.

Around line 15 of this file you'll see a line like:


Each of those "-D" statements enables a feature. You can remove those features
if you don't want any of them.

In particular, on platforms that do not support glob.h you should remove -DSTDLIBGLOB.

If you're on a 32 bit system that doesn't provide support for 128bit integers, which we call
octawords in XXL, remove -DOCTA.

If you don't have support for sockets on your platform, remove -DSTDLIBNET.

No filesystem? Remove -DSTDLIBFILE.

Without threading? Remove -DSTDLIBMBOX and -DTHREAD in DEFS a few lines above.

Not Unix-like? Remove -DSTDLIBSHELL.

No shared libraries, or don't care for dlopen? Remove -DSTDLIBSHAREDLIB.

Regenerating auto-built .h files

Some of the .h files in the source are automatically generated by some
primitive .js scripts. If you don't have Node installed, or don't care to
rebuild those (it's not necessary unless you change the definitions of the
built in data types), just comment out "BUILDH" at the top of the ./c

Try something like:

git clone
cd xxl

The default build has a TON of debugging info turned on which it will violently
spit at you while you use it. Commented out the "DEBUG=" line in ./c to make the
system more silent and friendly - but perhaps less predictable when things go
wrong (which they will.. often).

Turning off debugging also allows XXL to run much more quickly.


XXL is about 3,000 lines of hand-written C, plus 2,000 lines of auto-generated
.h files. Stripped executable is about 400kb.

I dislike large systems and aim to keep XXL small, but I also want programmers
to be able to easily get at functionality they need, both through a decent set
of built-ins, and through something akin to npm (still pondering this).

I believe XXL could be reduced to 1,000 lines or less of JavaScript or another
language that offers a more flexible type system than C's. I also wrote different
versions of similar code a lot for speed; implementing everything in terms
of each() (and other forms of iteration) would probably require much less code.


K4/Q by Kx Systems™, the quirky
Klong by Nils Holm, Erlang
(process model, introspection, parse tree/transforms),
Kerf's approachablity,
Io (self-similarity of objects and asceticism)
and C (simplicity, performance, rectangularity of data).

Imminent Rename

(I was trying to make this an eminent domain joke.)

I'm considering renaming XXL to tx, so if you see some mixing of names here
and there, please be patient with me as I consider this change. Input

Contact, Community, Chat

Visit with the XXL crew and many other smart folks on #kq and #xxl on Freenode IRC.


Wetly birthed by @tlack [email protected] at in sunny Miami

Many thanks to co-conspirators:

Eblin Lopez @, Mike Martinez @, Khalife Nizar @ Ironhack for
keeping me grounded and much input. 🙂

@joebo for help with Android.


No warranty. Trademarks right of their respective owners.

This project is not related in any way to the interesting XL project by
Dinechin et al


BSD-L (3 clause)