Oxidizing OCaml: Locality
blog.janestreet.com
blog.janestreet.com
(I don't know neither Ocaml, nor Haskell, or any ML language), but the podcast is always fun to listen to (not only programming also)
Now would relisten this, and may actually understand it!
Podcasts (not only this one) have evolved a lot! (another fav of mine is "Go Time" and many others).
* Is there a process for upstreaming this into the mainline language, or is this essentially JaneStML now?
* Why choose such an obscure word as `exclave` to indicate return-value optimization? How about `return local` or something similarly approachable?
Many of their changes have been ported to the 'official' OCaml compiler.
;)
"Rinse and repeat That’s my std::launder joke "
best answer I've got so far :)
I'm not a C/C++ developer and I have to admit: For me reading anything about those languages is every time just mind-bending! (And I have a hard time not to use any curse words right now in addition).
It took me over 20 minutes to understand (more or less) what this `std::launder()` does. The spec is just gibberish! So I've tried a BS generator a.k.a "AI" than[1]. Together with a Stackoverflow entry[2] it finally made click.
This function "launders" pointers from compiler assumptions. It will give you a "clean", pristine pointer back, without for example the information about the concrete object attached to it previously (because for example casting an object to its base class won't give you a proper pointer to the base class; at least that's what as SO comments says… Sounds quite ill, but so is C++). It will also for example "launder" away compiler assumptions about the `const` property of an object pointer. It's needed for example when you do "in-place replacement of an object" (replacing directly the bytes in some memory area of another object). Just casting won't make the compiler "forget" where this memory came from / to which object it was once attached. So you need to launder the pointer to the new object so the compiler forgets what it assumed previously about that memory area.
But let's not talk about what it actually means when a language needs some mechanism like that… My "WTF per Minute" output just reached a new all time high, I guess.
[1] https://www.phind.com/search?cache=91e3f5ce-dde2-422e-a3eb-6... [2] https://stackoverflow.com/questions/66176720/why-introduce-s...
And I should use now `phind.com` - awesome!
1. Language design is hard. At Jane Street we have a great opportunity to design new features, test them extensively in a realistic environment, and then change them. Because we have access to the entire code base where the features are deployed, we can even change concrete syntax relatively easily. So by developing internally, releasing internally, and then upstreaming with experience, we can be more confident that the feature design is correct.
2. We get a faster turnaround between idea conception and internal deployment. Working solely with upstream, we would develop an idea, go through a long design discussion with upstream, implement, merge, wait for release, wait for the rest of Jane Street to be ready to upgrade, and upgrade. Now, we can implement an idea in parallel with its design, rolling it out internally in stages (as appropriate), and then upstream later. This is a big win for us, and well worth the extra time spent moving changes back and forth.
> Why choose such an obscure word as `exclave`
We discussed a lot of possible choices and eventually decided this was the best one. I personally think names should either be self-explanatory or memorable -- so that once they have been explained they aren't forgotten -- here we went with memorable. As a word, exclave also captures what is going on in terms of part of the parent region being contained within the child region. `return local` was a strong contender, but it implies that `exclave` is always about functions -- whereas the actual feature is a bit more general than that. It is also a bit easier, when adding new keywords, if you pick a word that isn't used much, and `return` is used a lot.
I suppose it could also be useful to allocate int64, floats, and maybe even bigints on the data stack? I suppose it's more difficult to put C blocks like bigints or bigarrays in a region.
By parent’s description, those codebase adjustments are done during the iterative process of feature development too. Upstreaming would just be the last iteration.
For years I've had a pet, personal theory on the hierarchy of naming; and these were my first 2 "levels". I never heard anyone else ever blurt it out in these terms. (My thought was that I was weird, or it was just so obvious that no one ever needed to.)
https://github.com/ocaml-flambda/ocaml-jst/blob/main/jane/do...
I'm curious whether the authors here considered this. I didn't see any discussion of related work in the post.
The annotations make enforcement. It's easy to make the compiler put something on the heap by accident.
> Even without explicit mode annotations, the compiler can statically determine which variables may escape their enclosing region. Such variables are assigned the global mode; all others are automatically inferred to be local.
There's a little more info about mode inference in the proposal:
https://github.com/ocaml-flambda/ocaml-jst/blob/main/jane/do...
So arguments to "public functions" do require explicit annotations in order to be local, but otherwise the compiler is able to infer locality and this acts as a transparent optimization.
There's lots of other choices that would be easier to recruit for, with wider community and industry support.
Where does the Jane Street commitment to Ocaml come from?
But I don't know what you mean by imperative - for loops and stuff ? Most languages these days support functional programming concepts. Unless you're coming from C it shouldn't be that groundbreaking.
Note that Ocaml isn't lazy by default like Haskell - now that makes things... interesting I guess.
OCaml (which at FB was used to write the Flow typechecker and Hack compiler afair) favored higher order functions. I enjoyed playing with these in my Haskel-based introduction course at university, and I even wrote a similar language that favored them.
But I now heavily dislike if `a b c` means "call to a with functions b and c with unknown arguments". I need to read all three (a, b, c) to really get what's going on.
This was my biggest gripe with OCaml as it was used at FB.
(purely syntactically compare this to `a(parent => b(parent), child => c(child))`, for some verbosity I get a clearer syntax - a is the callee, and I got some information about what b and c possible do - this is a contrived example of course).
I think the strength of ML and Haskell got eroded as more "mainstream"y languages got decent type systems (TS/Flow, Hack, Kotlin, even Rust, Python is getting there).
The source code of the Viaweb editor was probably about 20-25% macros. Macros are harder to write than ordinary Lisp functions, and it's considered to be bad style to use them when they're not necessary. So every macro in that code is there because it has to be. What that means is that at least 20-25% of the code in this program is doing things that you can't easily do in any other language.
PG has created two other languages (ARC and Bel) which shows a high level of skill, and perhaps also shows that LISP is not his ideal language? He wrote this about designing a programming language: http://www.paulgraham.com/popular.html and he writes about the history of LISP here: http://www.paulgraham.com/icad.html which also talks about productivity:
if you were to compete with ITA and chose to write your software in C, they would be able to develop software twenty times faster than you [by using LISP].
ITA's hackers seem to be unusually smart`a b c` means `a(b, c)`. MLs just leaves out the parens.
You can't know just form looking at `a b c` what `b` and `c` are. They may be functions of course. But also any other value.
This would be exactly the same in for example JS. Nobody in JS ever complained that you can pass functions as arguments to functions and "don't see the arguments". (If you would write arguments you would actually call the function, and not pass it as a value.)
But in OCaml you have a static type-system that can tell you at any point what `a`, `b`, or `c` are! Just hover the symbol.
> I think the strength of ML and Haskell got eroded as more "mainstream"y languages got decent type systems (TS/Flow, Hack, Kotlin, even Rust, Python is getting there).
I don't think so. Especially given those examples.
TS/Flow doesn't have a sound type system. It may "type-check" fine and than explode at runtime. That's imho worse than dynamic typing as with dynamic typing you know at least that such fuck-up will happen. But TS delivers a false feeling of security about type safety where there is none.
Hack is "gradually typed" which is just another word for dynamically typed with some static checks. So not even close to MLs.
Python is also still dynamically typed, and that won't change. The bolted on type-checkers are unsound. So same situation as with TS.
Nothing can be said about Kotlin afaik, as there was no formal work on its type-system.
And Rust? Rust is a ML (with ugly C syntax)! So Rust is a prime example of how MLs are still one of the most modern and influential language designs to this day on.
With Rust and Scala you have even two MLs in the Top20 languages. If something than ML is slowly but finally approaching mainstream. Maybe not in its original form but still the tradition lives on.
Rust is more ML than Flow or Hack because of traits (type classes? Not sure what ML calls them or if it even has them, Haskell does). But it’s a vastly different language otherwise, so I think the “other languages have learned from ML” is more accurate description of what’s happened.
Some people write JS/TS in the ML style (point-free), but it’s not by far the most popular way. TC-39 came around to Hack-style pipelines for the same reason. For me the stylistic difference is a crucial indicator of preferences and priorities. Haskell’s operator “overuse” lives in a similar world.
I have no idea, maybe your original parent will elaborate :)
> Parameters with a recursive lifetime can only be used in a tail-call position within a function. If a function with only recursive params compiles successfully, it is guaranteed not to grow the stack.
For the self-recursive case, people are likely better off just writing the imperative version. But using these with mutual recursive calls lets you build safely build state machines that compile down to the same performant code a classic goto-based implementation would in C.
https://github.com/rust-lang/rust/tree/ef75860a0a72f79f97216...
Here's a pretty good article about the transition: https://roscidus.com/blog/blog/2014/06/06/python-to-ocaml-re...
What I would like to know is why they didn't take more direct inspiration form the "ML with Python syntax", namely Scala 3, and its "capture checking"[1].
Those OCaml "modi" look just like special cases of capture checking to me. Was that more general approach considered? And if, why was it dismissed?
[1] https://docs.scala-lang.org/scala3/reference/experimental/cc...
> Ideally, a language could provide a spectrum of allocation strategies freely interoperable within a single application. With modes, users can write OCaml with all the usual GC guarantees—but when performance is paramount, opt into the consideration of lifetimes, ownership, and concurrency.
where nim fits? how it can compete with other new fits and other old?