The APL Orchard

Kamila Szewczyk

8:01 PM

i also considered compiling APL to C++ or Rust

user

@KamilaSzewczyk Aren't some implementations of it somewhat faster than those two when it comes to array-related stuff?

Kamila Szewczyk

the first one was mildly disappointing - i tried to make an APL-style sum+iota function, and std::iota/std::reduce was horrible, the optimized formula method was the best (x(x+1)/2), and the streams-v3 and boost hana method was decent-ish

meanwhile when i threw it into the rust compiler, i got really high quality output

@user i was thinking of making something which isn't as flexible as APL, but thanks to that can be optimized better

like, supporting, say, rank 5 arrays at most

user

Oh, I see.

Kamila Szewczyk

or ditching eval

ngn

or making the language statically parsable

Kamila Szewczyk

8:04 PM

yes, that too

i wanted to somehow provide type definitions and compiler hints for functions

but i didn't come up with a good syntax yet which would fit APL's terse syntax

ngn

@KamilaSzewczyk 4 more ranks till k :)

user

I wonder what a statically typed version of APL would look like.

Kamila Szewczyk

at first i thought about a somewhat bad idea of utilizing the fact that assignment only produces constants, to modify it to convey type information if it can't be trivially deduced

we already know what is the signature of f, so x←f 2 doesn't need a type signature

ngn

type inferrence? there have been attempts..

Kamila Szewczyk

but if i did something like g←{...}, then i'd maybe require providing a prototype

rak1507

8:06 PM

@KamilaSzewczyk 'assignment only produces constants' what?!

Kamila Szewczyk

@rak1507 that's right, no mutable state, we're making APL purely functional.

rak1507

ok, cool, but how do you name a function

Kamila Szewczyk

like I just showed

you can't reassign functions

ngn

@KamilaSzewczyk oh, so proper dicts too

Kamila Szewczyk

i plan on bringing lisp-isms in, data structures, probably some more stuff

rak1507

8:08 PM

oh, I read constants as in 'things that aren't functions'

Kamila Szewczyk

@rak1507 think of it as of a function pointer

i'm afraid i'll end up making haskell but somewhat worse in the end

rak1507

I just got confused and associated 'constants' with like, mathematical constants or something

user

So how would you give it type hints?

rak1507

@KamilaSzewczyk monads in APL when

Kamila Szewczyk

@user and this i haven't worked out yet

user

8:09 PM

And generics would also be a bit troublesome.

Kamila Szewczyk

my initial horrible idea was making generics/guards just like they are in C++, but this is a fuzzy and potentially bad idea

Wezl

@KamilaSzewczyk but with more symbols!

Kamila Szewczyk

as for the type hints, I was thinking of something like g::u->u g←{...}

user

Monomorphization? That wouldn't be too bad.

Kamila Szewczyk

assuming u is some integer type, just like they are in K

or you could have type definitions near function bodies but this seems a bit readablen't to me

Wezl

8:11 PM

@KamilaSzewczyk but using actual arrows because why not

Kamila Szewczyk

yes :p

user

@KamilaSzewczyk "readablen't"?

Kamila Szewczyk

@user i like playing around with english vocabulary and grammar

user

@KamilaSzewczyk Er, ok.

Kamila Szewczyk

sorry if that's annoying, i found it a bit silly and creative, maybe a bit fun

Wezl

8:12 PM

me too

user

It'd be fun to see implicit parameters in APL too, or at least typeclasses

@KamilaSzewczyk Not annoying, just confusing

@user Though that would probably ruin the syntax.

ngn

new implementers always dream big (static typing and stuff), but it's hard enough to make even a boring interpreter

Kamila Szewczyk

@ngn i'm actually simplifying APL by quite a lot here

user

@ngn Shoot for the stars, and you may reach the moon.

Kamila Szewczyk

so much that i'm afraid that it'll lose most of it's terseness properties :p

user

8:17 PM

@KamilaSzewczyk I mean, if you're going to make it like Haskell, it'll still be pretty terse, just not APL terse.

dzaima

@ngn but people who have already written interpreters will probably just stick to theirs, so with that mindset you're never gonna advance

Kamila Szewczyk

terseness is cool, but usefulness is better :p

i mean, limited ranks, no eval, statically compiled, static type hints so the grammar isn't ambiguous and you can actually parse it, no tacit programming

it's pretty evident that i'm trying to make something simpler and less flexible here :p

user

Terseness is useful. The less space your code takes up, the more space there is to do calculations, so you get faster code :P

ngn

@KamilaSzewczyk multidimensional arrays?

Kamila Szewczyk

this logic makes sense mostly for CPUs, not in terms of more space to do calculations

dzaima

8:20 PM

@KamilaSzewczyk properly using/validating type hints is hard though. The type of even + is pretty complicated

user

@dzaima That could be handled by not allowing schizophrenia, like Jelly does.

Kamila Szewczyk

@dzaima type of + depends on it's operands

you already know types of the operands

so you know the result type /shrug

ngn

@KamilaSzewczyk what is type? e.g. is it "int" or int64,int32,etc?

user

Here's a question about a type system for the J language.

dzaima

@KamilaSzewczyk so if you have a generic function, you won't be able to type-check it without concrete arguments?

Kamila Szewczyk

8:23 PM

1) types are the same => result is obvious

2) types aren't the same => it picks the bigger type and the result is of this type. the smaller argument is being promoted

@dzaima the user provides the signature of the function, and i have signatures of all existing functions/entities, why can't i type check it?

ngn

because of overflows?

dzaima

@KamilaSzewczyk is g::(u[],u[])->u[] g←{⍺+⍵} valid? Or do i need to somehow specify that both ⍺ and ⍵ have the same length?

Kamila Szewczyk

i won't handle overflows, at least yet

@dzaima i think i may let it bail out in the runtime

and check for obvious cases when it's called with arrays of different length

user

Might be better to keep it mostly dynamic and do some dirty casts to make the C compiler accept it.

Kamila Szewczyk

i want to make parsing decidable

oh also

hmm

user

8:29 PM

Couldn't it still be decidable if you don't have mutable state and there's no eval?

Kamila Szewczyk

yes

dzaima

also types themselves are complicated. you want to have types for: 1) any length vector; 2) any rank array; 3) 2-item vector of a character vector and an integer vector; maybe also 4) matrix with 2 rows and any number of columns

ngn

@user not without changing apl's syntax significantly

Kamila Szewczyk

@ngn how so?

user

Actually, you could still keep mutable state, just not allow assignment of a monadic function to a variable that previously held a dyadic function and stuff like that.

Kamila Szewczyk

8:31 PM

i don't want mutable state because it will interfere with optimizations

the only thing i can think of is the dfn/dop type problem

user

@ngn Why would you have to change APL's syntax?

Kamila Szewczyk

but maybe i could have different constructs to mark a dfn and a dop?

user

@KamilaSzewczyk Don't allow schizophrenia.

Kamila Szewczyk

what in the world is schizophrenia

ngn

@KamilaSzewczyk the classic example is: a b c. how do you parse that? a(b c) or (a b)c

Kamila Szewczyk

8:32 PM

@ngn give types of a, b and c

dfn/dop/value are valid answers

ngn

oh, ok

Kamila Szewczyk

this was exactly the reason why i wanted type hints

user

@KamilaSzewczyk It's more fun to say than function-operator overloading. Apparently, back in the day, they thought MPD/DID and schizophrenia were related.

Kamila Szewczyk

to avoid this horrible thing

user

Couldn't you tell the types of a, b, and c by looking at whether or not they contain ⍺⍺ if they're in braces?

ngn

8:35 PM

@user a, b and c are just identifiers. you need to know what's behind them at runtime.

Kamila Szewczyk

but... you know what's behind them

you're compiling APL

/interpreting

ngn

but yeah, with a bit of static type inference you can make it work for most cases

Kamila Szewczyk

doesn't matter for this case

user

@ngn But if these variables were already defined before, you'll know their types already.

ngn

would ⍺⍺ and ⍵⍵ have to always be functions?

Kamila Szewczyk

8:37 PM

can they be something else?

user

Unless you typed a b c befor eyou typed a←..., etc.

ngn

@user right

user

Haskell handles this problem quite well.

Kamila Szewczyk

@user my idea was to defer parsing {}'s and overall definitions before building the "symbol table"

dzaima

@KamilaSzewczyk 1(2{⍺ ⍺⍺ ⍵⍵ ⍵}3)4 is a 4-element vector

ngn

8:37 PM

in ngn/apl i did this inference scope by scope. i wouldn't recurse into an inner pair of { } before i've processed the outer one

Kamila Szewczyk

yeah, i had the same idea

Adám

@user Even for Dyalog's dfns it can mostly be done. See kk

ngn

this helps compile things like mutual recursion: {f←{g ...}⋄g←{f ..}}

Kamila Szewczyk

yep, that's what i thought of

so far even thinking of implementing an APL dialect is exhausting

Adám

Btw, I think the strict naming convention, as used in BQN, is the best way to make it possible to statically parse APL.

It helps the human reader too.

Kamila Szewczyk

8:40 PM

like?

goof

What's the strict naming convention?

Kamila Szewczyk

is it what i'm thinking of? (semantic names)

Adám

abrudz.github.io/style#nc

user

Lowercase for functions/uppercase for operators?

ngn

or underscores

Adám

8:41 PM

It doesn't really matter what the exact scheme is, as long as you can distinguish arrays, function, monadic operators, and dyadic operators.

user

Are you going to use dfn/dop syntax or tradfns (or both)?

Adám

I find that array, Function, _MonOp, _DyaOp_ works well.

ngn

@KamilaSzewczyk it will only take a few years :)

Kamila Szewczyk

depressing.

Adám

Though making arrays uppercase and all other lowercase would work well too.

Kamila Szewczyk

8:43 PM

a few days ago i hacked together a bit of code which supported strands, scalars, +-×÷ and monadic dfn's

and assignment, of course

ngn

@KamilaSzewczyk that took me about a week or two, and the whole interpreter ~3years, so you can probably do the whole interpreter in a year :)

Kamila Szewczyk

oh well, i didn't really care about the style and i wrote it in JS, and i used a PEG

i got crazy with the operators by giving them around 6 or 7 meanings because h.

Wezl

CMP: What are the minimum features for an APL dialect to be considered an "APL dialect"?

ngn

@Wezl depends on who you ask

Wezl

well I'm asking everyone

I need opinions

Kamila Szewczyk

8:46 PM

another chat mini something: how much of an overstretch would be calling my APL dialect "APL"

Adám

Well, every APL dialect I know of has +/ mean "sum"…

ngn

@KamilaSzewczyk 7.34

Wezl

there was a standard at one point

Kamila Szewczyk

thanks for a very precise answer

ngn

:D

dzaima

8:47 PM

@Wezl for me - 1) you can do scalar + array which vectorizes; 2) there's a reduce operator

Kamila Szewczyk

what is the scale?

Adám

@KamilaSzewczyk I think that'd only make sense if most basic APL\360 statements work with no modification.

Kamila Szewczyk

do we assume ⎕io←0 or ⎕io←1 or ⎕io←0.5?

goof

0 to many

ngn

@KamilaSzewczyk millistretches

Wezl

8:48 PM

@Adám but what's "basic"?

Kamila Szewczyk

damn, that's really low

Adám

@Wezl Yeah.

goof

I'm so disappointed that there's no xterm escape sequence to insert a new line of text without disturbing the line the cursor is on.. that sounds like such a basic scenario

(For any application that wants to output lines while keeping a simple input prompt at the bottom)

ngn

@goof rlwrap not good enough?

@KamilaSzewczyk one language used centistretches instead and it tore off apl, flying in the direction of lisp :)

goof

@ngn It's a cool hack / useful tool but I'm disappointed that there's no "three lines of C max" solution for this :P

ngn

8:58 PM

sorry, on behalf of c :)

goof

Like all of unix, you can easily get 95% of the functionality you need, and then you need 20000 lines of libraries and crap for the next 1% :P

Even if that 1% is just a small tweak to what exists there already

Does not compose well, does not extend well..

cannadayr

@Wezl infix notation maybe

Adám

9:12 PM

@cannadayr even for user functions?

Marshall

@dzaima I think we ultimately want something higher-level to work with (hopefully most of the implementation would take place in the new language though). Here's another sketch:

dzaima.github.io/paste/…

I'd like it to be mostly APL- or K-like other than the types and control structures.

The {} parameters are always computed and expanded at compile time, and it should be possible to pass anything that's known at compile time in that way. So you can do all sorts of functional programming and type computations at compile time, but not at runtime.

ngn

@Marshall what should be the result from compiling that - c code? an .o file? a buffer filled with x86_64 code?

Marshall

@ngn I think we'd want multiple backends. Probably C with vector intrinsics is easiest to start with, since we don't have to do register allocation.

dzaima

@Marshall i also decided to switch to pointers & size for the basic functions. everyone could easily argue forever about what specific syntax should things have, so i'll decide to not

ngn

@Marshall that's a very ambitious goal

dzaima

9:23 PM

@ngn starting out, a single backend isn't that ambitions, and it definitely isn't unreachable

ngn

maybe LLVM IR or something for multiple targets?

dzaima

@ngn that'd be an "easy" next step after C generation

Marshall

@ngn We definitely want some sort of IR. I don't know how LLVM does vector instructions so I don't know if it's suitable.

ngn

idk either :(

dzaima

@Marshall my educated guess is that it does exactly what you see in clang - there are types for vector registers of certain sizes, and builtins that act on them

Marshall

9:26 PM

@dzaima That works as long as it supports all of x86.

coltim

@Marshall I think they have a bunch (or maybe you have to write your own wrappers). I think that's what ISPC does

dzaima

@Marshall i think it might be better to have a custom internal IR. could take ideas (heavily) from LLVM or whatever, but LLVM IR itself will probably be too cumbersome

Marshall

Although the thing about our version is that the source code should describe exactly the instructions in the object code, so we don't want LLVM to do any sort of optimization.

@dzaima If we can use a subset of LLVM then that would be nice.

All we really want is x86 but stack-based and with structured loops.

ngn

this is the IR my addII() function compiles to:

define internal i32 @addII(i32* noalias %0, i32* noalias %1, i32* noalias %2, i64 %3) #6 {
  %5 = ptrtoint i32* %0 to i64
  %6 = and i64 %5, 31
  %7 = icmp eq i64 %6, 0
  tail call void @llvm.assume(i1 %7)
  %8 = ptrtoint i32* %1 to i64
  %9 = and i64 %8, 31
  %10 = icmp eq i64 %9, 0
  tail call void @llvm.assume(i1 %10)
  %11 = ptrtoint i32* %2 to i64
  %12 = and i64 %11, 31
  %13 = icmp eq i64 %12, 0
  tail call void @llvm.assume(i1 %13)
  %14 = add i64 %3, 7
  %15 = and i64 %14, 4611686018427387896

dzaima

@Marshall i don't think LLVM is stack-based, nor does it have structured loops?

cannadayr

9:30 PM

@Adám do you have an example? (not sure what you mean)

@Marshall what are yall working on?

dzaima

21 hours ago, by Marshall

If anyone wants to get together on a b-ish (Whitney's b) vector language I'd definitely contribute.

cannadayr

ty.

Marshall

@cannadayr The idea is that this would just be for implementing SIMD kernels for array languages. C-like languages make this difficult, especially if you want to write code that handles multiple types.

cannadayr

@Marshall yea the closest i can think of is something like SAC

Marshall

@dzaima The SSA thing is probably easier than stack-based actually.

ngn

9:35 PM

does it deserve an own chat room? i guess most people here wouldn't be too interested in such low-level stuff

ktye

yes, stack based vector language! raw.githubusercontent.com/ktye/i/master/_/ps/k.ps

cannadayr

i hope this doesn't mean i have to move on from BQN after i finally might get the VM working ;)

dzaima

@cannadayr BQN will still exist. what'll be happening is there will be a better language in which to implement BQN (or its builtins) in

Marshall

@cannadayr If erlang has SIMD bindings then you could make an erlang backend for it and have an actually fast BQN.

cannadayr

@Marshall ill do some benchmarking, but my idea is that ebqn will be a control-plane language only -> small datasets, mostly to communicate state changes between nodes

Marshall

9:38 PM

This definitely doesn't compete with BQN. It's a DSL, really.

cannadayr

what i would want after that (who knows, im not made of time), is a data plane language, which is what this sounds like

Marshall

@ngn Yes, and we probably want some system where we can start making design docs instead of sticking everything in chat.

ngn

@Marshall a wiki?

Marshall

@ngn Sure.

Adám

@cannadayr k does not allow things like f←{2×⍺+1+⍵} ⋄ 2 f 3

Marshall

9:43 PM

We could open a TopAnswers room as well if we want to keep talking.

cannadayr

@Adám fair enough. i guess we need 3 categories. APL languages. non-apl languages. and languages that are neither APL or not-apl

ngn

@Marshall as long as it's publicly viewable and backupable, i don't mind

cannadayr

i would be interested in following but not sure ill be able to contribute much

dzaima

@Marshall having *: just be implemented for you is cheating a bit. imo there should be primitive builtins for operations currently available in at least one targeted architecture, everything else needs to be manually implemented (as that's what you'd need to do anyways)

Marshall

@dzaima It would be implemented within the language. I think there's going to be one best way to multiply with overflow, so you shouldn't have to reimplement it for every program you write.

dzaima

9:46 PM

@Marshall hm, i guess it could be in the stdlib of sorts

oh, you'd want a stdlib anyways, for implementing things not available in only some architectures

Marshall

Here's our room: topanswers.xyz/apl?q=1623

2

pitr

10:06 PM

@Adám only got notif now. Any time should work for me, I don’t have many meetings. I’m in Berlin time zone

Adám

@pitr Let's do Wednesday at 16:00 UTC/17:00 CET.

@cannadayr Did you want to participate too?

cannadayr

@Adám maybe! ill mostly stay on sidelines to not try and derail the convo

Wezl

@Marshall I like etherpad (especially bc I can't use topanswers)

Marshall

10:21 PM

@Wezl Why not?

Wezl

@Marshall terms and conditions

Adám

Wezl

I never said I was going to break the CoC

Marshall

@Wezl I can't find any terms or conditions for using TopAnswers. What do you mean?

Wezl

last I checked you had to be 16+ to join

dzaima

10:36 PM

@Wezl can't find anything anywhere saying that. maybe you're thinking of Codidact?

pitr

@Adám sounds good. Meet here?

Marshall

@Wezl I don't see where you'd be presented with T&C either, as you just join by clicking a button and then optionally set profile information.

Wezl

...last I checked

Adám

Announcement: Live coding session, prototyping load-on-demand in APL; Wednesday 16:00 UTC at twitch.tv/adambrudz

@pitr @rak1507 @KamilaSzewczyk @RGS @cannadayr ^

dzaima

@Wezl something like that would be nice, but it doesn't seem very suited for code

Wezl

10:42 PM

It's almost ok if you set the font to monospace, (gear menu)

chrispsn

10:55 PM

@Marshall error: "ERROR: invalid question CONTEXT: PL/pgSQL function raise_error(integer,text) line 1 at RAISE SQL function "login_question" statement 2 SQL function "login_question" statement 1"

Marshall

@chrispsn Try again?

chrispsn

Now working - ty

Marshall

@Adám told me to flag the question to call a mod to change the question type, but it seems flagging also makes it disappear?

Adám

@Marshall Oh, maybe I remember wrong (I'm fairly new to TA too), try removing your post flag and flagging your comment instead.

rak1507

@Wezl ...or lie

Wezl

11:06 PM

@rak1507 well yes, but I don't think I can sign up because I have to access it through a proxy anyway

rak1507

oh do you have some sort of internet controls that block .xyz? it's blocked for me at school

(unlike the APL orchard, which is probably bad for my education but oh well ;) )

Wezl

I don't think it specifically blocks .xyz, but it block lots like the pictures I proxy through Google Docs or the ->irc->matrix->element bridges, and even the main etherpad website (I use an alternative instance) (and twitch.tv btw)

open source and self hosting have been good to me though

rak1507

Ouch, that's a lot

RGS

11:33 PM

@Adám Thanks for this. Is there an expected duration for the stream?

Transcript for