Settings

Theme

Nimony: A New Compiler for Nim

nimony.org

177 points by Fervicus · 96 comments

Reader

23 threads
stakhanov

When I was doing the rounds roughly 10 years ago, trying to find a new programming language to learn, my money for "next big thing" landed on nim. I did some personal projects and absolutely loved it but, unfortunately, opportunities to work "on the clock" in nim never materialised and then rust started eating the world.

Thanks, Andreas Rumpf, for sticking with it for all these years: This seems like a great opportunity for me to get back into nim.

afavour

I’m curious if this closes the door on Nim’s capability to output C code which you then compile yourself. I first discovered Nim when I was looking for something that could work with Apple’s requirements to embed bitcode in iOS apps, it meant you had to use their specific version of LLVM and that ruled out a lot of languages. It always seemed like a really unique feature (though IIRC Haxe has something similar).

Apple has long since removed that requirement so it doesn’t really matter anyway and I wouldn’t blame the Nim folks if it’s gone, these seem like a great set of improvements.

  • 15155

    The real win with the compile-to-C strategy is embedded - a million embedded devices only publish vendor-provided (typically GCC) C toolchains.

    • elcritch

      This is a great part of Nim. I started using it with ESP32's for 5+ years or so now. Nimony is even more ideal for embedded.

      Great part is that Nim's stdlib networking was easy to port to esp-idf's lwIP based network stack. So you get access to production Nim libraries for lots of networking stuff, or async, etc.

  • txdv

    A lot of compilers nowadays compile to wasm, when you have wasm, you can compile to C.

    • 15155

      Wasm is great, but it's a terrible IR for C back-compilation if you have alternatives available IMO.

  • tokai

    Nimony still puts the C code in the .cache/ folder as nim does.

    • mathisfun123

      what does this imply? are you saying it does in fact still actually "transpile" to C?

      • tokai

        As I understand it both 'transpile' and does native code generation.

Symmetry

> Nimony is case sensitive like most other modern programming languages. The reason for this is implementation simplicity. This might also be changed in the future.

I've never been too annoyed by Nim's approach but I think this removes a complaint many people have.

  • aslilac

    it also removes one of nim’s most unique differentiators which seems like a loss to me

    • maplant

      If your unique differentiator for your programming language is that identifiers are case insensitive you need to find a new schtick

    • mathisfun123

      i've invented a new language which swaps the meaning of `*` and `÷` as a "unique differentiator". are you interested?

  • shevy-java

    I am not sure. I think Nim is a fine language, but it is a little bit cumbersome as well, due to the typing. People complain just about everything, so this is not a good metric to use as the only metric. Nim has a similar problem as crystal has: too few users. With the rise of AI, this problem has amplified. You can see it if you look at Ruby, how it declined in the last two years. I wonder what young people do these days. Has learning programming languages become worse here?

girvo

It’s super exciting. It’s basically adding all the features I wished Nim had, and I’ve been using Nim in anger since the v0.10 days!

polotics

This is great! Congrats to the Nim team for keeping the pedal to the metal!

Sum-types, borrow-checking... Nim 3 is beautiful.

  • ZoomZoomZoom

    BTW, Sum types (along with incremental compilation, SSO strings and other thing) have been recently added to Nim 2 with the 2.4.0 release: https://nim-lang.org/blog/2026/10/28/nim-240.html

    • torginus

      Is this a Python 2/3 situation?

      • christophilus

        Might be, but Nim is statically typed, and we have LLMs now, so migration should be pretty trivial, is my guess.

        • pjmlp

          Speaking from the stand of Java and .NET breaking changes, it is not, some folks will cling to the old world no matter what.

          Check how many deployments (products even) are still being done in Java 8, or .NET Framework.

          • simlevesque

            Nim 2 isn't as popular or ingrained as Java 8 or .NET Framework.

            • pjmlp

              True, that doesn't mean that folks will feel like migrating, unless they see a value in doing so.

        • torginus

          I'm not a Python guy, but my understanding was that the reason why Python 2 stuck around for so long was that some people preferred it over 3

          • stakhanov

            I think it's all about corpus weight.

            When Python did the migration, it had a good flow of new investments coming into it and delivered good reasons for why those should go to Python 3 rather than 2. At the same time, it also already had a base of existing investments in Python 2 delivering better ROI without the added cost of a migration, and the sum of that exceeded the critical mass it took to make it economical to keep both versions of the ecosystem around.

            With Perl you are seeing what it looks like when you have zero new investments coming in and no good reasons for why new investments should go to Raku. (Because: Just use Python). You do still have a non-negligible base of existing investments in Perl 5 and those deliver better ROI without the added cost of a migration. So my prediction is that Perl 5 will stick around, and Raku is dead on arrival.

            With Nim it might be the other way around: Good reasons for why new investments should go to Nim 3 rather than Nim 2, and not all that much existing investment in Nim 2. If Nim 3 sees healthy adoption, it might well be more economical to re-deploy ecosystem maintenance to Nim 3 and migrate most of the existing code, rather than keep two versions of the ecosystem going.

            • ghshephard

              I know you realize this, but it is worth emphasizing that Perl 5 vs Perl 6 is a not comparable to Python 2 to Python 3. For 98%+ of code, Python 2 and Python 3 are mostly identical, and the conversion, even in massive code-bases, is completely mechanical for 99.5% of the code.

              Perl 5 and Perl 6 (Raku) are different languages.

              • nagaiaida

                it's also worth emphasizing basically all the changes raku made filed off an extremely sharp edge of perl

          • Beretta_Vexee

            It was largely a question of the existing codebase. It took years to convert all the obscure scientific libraries from Python 2 to Python 3.

            The migration was no easy task. So all the projects already written in Python 2 remained on version 2. Until the day they reached a dead end and were forced to migrate.

          • ghshephard

            It was inertia. I was working at a company in 2017 in which 100% of the very extensive code base was written in Python2. The Python3 migration was done in a weekend and was mostly a nothing burger, as, with few exceptions, the conversion was 100% mechanical and just worked. The print vs print() was the most visible change, and the str/bytes change was a genuine improvement.

            Except for no longer being confused about whether I was reading bytes or "strings" (Python3 eliminated that confusion), and having to remember for roughly a year to use "()" when using print, I don't believe I even really noticed that big a difference between the two.

            Honestly - I'm surprised that python managed to hold on long enough for "uv" to be released and solve all the packaging/venv pain - that was a real problem with python for 15+ years and wasn't solved until 2024.

    • brabel

      How does this differ from this new compiler?

      • michaelsbradley

        The Nimony compiler is a completely new codebase. Some language and compiler features were ported to the old codebase of the Nim 2 compiler.

eterps

Nice, it looks like it supports option and result types as well:

- https://nimony.org/stdlib/std/result.html

- https://nimony.org/stdlib/std/options.html

deepsquirrelnet

I did my first "real" nim project this year. I'd been tinkering with it here and there, but found a chance to use it. I'm coming from mostly python experience and found it to be a really nice experience.

pietroppeter

I think Nim is a great fit for the AI era. let me explain

Always appreciated how Nim gives you a low cognitive load: it is easy to reason about, does not get in the way of doing stuff, you focus on what you need to do it and rarely on how you do it (specifics of the language).

I think this is a great advantage in this era of automatic/AI assisted programming. The fact that the language is very readable also helps for the moments when you need/want to read the code.

Nimony btw is a very cool way to evolve the mainline compiler and looking forward to use this in main Nim (but it is also something nice to experiment with)

  • raffraffraff

    I thought it would be the opposite, because you can't say: "The AI was trained on a bazillion lines of idiomatic Nim 3 solving a wide range of common problems"

    • pietroppeter

      Yep that is the expectation one could have but it is not the experience I (and many others) had.

      The explanation I have is that it is so close to pseudocode that it only needs to know just a tiny bit about Nim peculiarities (not too many) to do a good job (plus static types and compiler help a lot).

    • tokai

      Ai writes nim pretty well. Its close enough to other languages that llms can manage to write performant code with few issues. There are some gotchas, but the fast compile times and good error messages makes it easy to iterate fast for the bots.

brainwipe

Hopeful for the new compiler but wished they'd called it Nimoy!

jphoward

I love Nim but my only problem with it previously was support for dictionaries/objects felt really weird? There was some sort of JSON structure as I recall, but it didn't seem anywhere near as easy to use as the analogous structures in Python/JS. Am I misremembering? Has this been improved?

  • PMunch

    Nim is very strict about types, which is probably what you remember. Things that are typically expressed as a dictionary/object in JavaScript and Python are better represented as a fully typed tree of objects in Nim. The language of course supports the data-structure dictionary/hashmap/table (called Table in Nim) which is "anything hashable" -> "any type" lookup. This is what's used for things like `myTable["hello"] = 5` for storing the number 5 for the key "hello". However because of the strict types you can't mix which types they map to without creating some sort of wrapper type that gives you the alternatives you're allowed to use. Many people reach for Nims JsonNode type to fill this gap for small collections of string/number/array/dictionary since it already is such a wrapper type.

    • jeremyjh

      I wonder if this is more palatable with first-class sumtypes/ADTs in Nim 3? Or is that already what JsonNode essentially was?

      • xigoi

        Sum types already existed in Nim 2, but Nimony makes them more convenient to use.

  • ZoomZoomZoom

    I suspect you're talking about heterogenous containers which are antithetical to a strongly typed language but nevertheless are supported via std/json's JsonNode dynamic data.

    In Nim regular hash maps (called tables here) have pretty regular accessors via `[]` `[]=` overloads (`foo["frob"]=42`), plus additional routines such as `.get(otherwise)`.

    https://nim-lang.github.io/Nim/tables.html

    Don't know why are you conflating objects in your question.

    • jphoward

      I'm 'conflating' them because they're called objects in javascript, so people unfamiliar with Python know what I mean. Ok thanks.

      • setr

        The difference isn’t between JS/Python; objects exist in many languages with various representations, JS happens to model objects using dictionaries.

        So it’s really the case that JS conflates objects/dictionaries, so only people familiar with only JS would know what you mean

        Edit: actually thinking about it, I’m fairly positive python also models objects as dictionaries; it just puts more effort into hiding it

netbioserror

Using Nim in production, AMA.

  • txdv

    Load, scale?

    Biggest pain point? Biggest win?

    • netbioserror

      It has a scoped responsibility, does mostly compute and data processing. Invoked by a PHP backend, single-threaded, we get parallelism at invocation rather than internally. So we can saturate server threads optionally if we need to.

      Biggest win: Single-owner scoped lifetimes by default with no annotations plus expression-oriented value semantics makes it eminently readable. Basically no memory semantics, all business logic. (It does processing of proprietary binary data including a lot of math.) Extraordinarily easy to maintain, refactor, add features. Valgrind confirms no orphaned blocks in some of our biggest test scenarios, it's impressive. Full compile-time evaluation is nice too, easy to slurp a whole file at compile-time and have it accessible in a normal binding, and writing generic procedures is obvious and simple.

      Biggest pain point: Hard to say, the only friction I've really ever run into is its weak tooling. I just use Make for build scripting (invoking the compiler directly instead of using Nimble) and Sublime's built-in features for refactoring. It's small enough that print-debugging works, but I occasionally dip into FlameGraph. Nimsuggest has some...dealbreaking quirks, but successors are in the works.

      • sirtaj

        What are the biggest things you miss from the standard and third party libraries?

        • netbioserror

          1) From std, a fully Nim-native tiny regex engine (the tinyre package fills this need, and I think it should be the default; the default regex engine brings in PCRE as a dependency).

          2) From std, compression (I pull the zippy package as a dependency, and have built against Brotli and other compression libaries; whether compression should be in std is I think something Araq may have commented on already).

          Aside from that, nothing. Nim's compiler has an extraordinarily powerful templating engine built-in called Source Code Filters: These build text replacement procedures at compile-time, and can handle most arbitrary Nim code, including generics and the compile-time `when` conditional, although certain sugary arrangements by the default parser can't work (Nim code in SCF must be prepended with a # to mark it, and control flow blocks must have end markers). I use SCF primarily for HTML reports, which is one of my primary outputs.

  • stakhanov

    Are you hiring?

pjmlp

Looking interesting, I like the improvements over what Nim used to be.

The big question is naturally the adoption curve, given the breaking changes.

  • DeathArrow

    It's not a big issue since Nim didn't have a large following. If anything, it might help Nim to take off if it's good enough and user friendly enough.

aktenlage

Does anyone already have experience, how the developer experience is? Error messages, debugging and so on?

kitd

Any recommendations on the best IDE for Nim 3?

flykespice

I have been eyeing nim for a while though I'm more used to curly-braced language families.

The first thing I evaluate in a language is how it looks like in production code, and the samples in github strikes me odd -- like most of the code identation is 2-space and no line break between variable declaration and other groups of statement (everything is lumped together).

Just check https://github.com/WyattBlue/auto-editor/blob/master/src/wav... to have an idea of what I'm talking about. Is that how normal nim code supposed to look like? (following its best practices)

sergiotapia

I've always had a soft spot for Nim, never had the chance to use it professionally. I personally can't wait to let it rip with AI. I think this is going to be the perfect language to use with AI. Fast as all holy hell, ergonomic, a language lead who cares deeply and does not quit. Major props to Araq for sticking it through all these years, I hope it's Nim's time to shine finally.

  • kombine

    It also seems like he embraced AI, which I find healthy given he wants to compete with established languages.

tokai

I have a hobby nim project at about 22K LOC, stdlib only. A quick and naive llm rewrite to nimony gave a ×1.3 runtime speed up. Pretty cool, will definitely keep up on nimony's development.

esafak

When is 1.0 slated to be released?

anthk

No i386? We do exist.

  • drob518

    386 only? You’re very rare indeed these days. Didn’t the Linux kernel also recently move past 386 only systems?

    • ptx

      There are still 32-bit applications running on 64-bit operating systems which you might want to extend.

    • anthk

      i686 then.

DeathArrow

Cool, now we can fight the borrow checker in Nim, too.

  • suncore

    If I understand correctly, it still does have reference counting so you don't have to fight the borrow checker if you don't want to.

jimbob45

Nim still has the GC, right? It always seemed like a slightly better Golang, much worse Dlang in that respect.

  • miguel_martin

    Nim's better than dlang in this respect. Default memory management is similar to C++ and Rust. See https://nim-lang.org/docs/destructors.html

  • jibalt

    Optionally but not by default ... things have changed a lot:

    https://nim-lang.org/docs/mm.html

    • vips7L

      Reference counting by default is GC.

      • miguel_martin

        ref counting is only performed if you use `ref` types, e.g.

          type 
            RefObj = ref object
              foo: string
              bar: int
        
            ValueObj = object
              foo: string
              bar: int
        
        
        `RefObj` is ref counted equivalent to a shared_ptr in C++ land (non-atomic counter unless using --mm:atomicArc or Nim3/Nimony has this as a default I believe). `ValueObj` is equivalent to struct in C, C++, etc.
      • flykespice

        IIRC, ref counting is at least more deterministic than traditional GC because it only runs when you destruct the object (it goes out scope, object gets reassigned whatnot).

        In Comparison, traditional GCs like JVM runs on a background thread, it's hard to determine when they run and generally blocks the whole program while collecting (stop-the-world).

        So at least they are a little better in that regard?

lobf

Nim is cool, but holy moly this web design is disgusting. Absolute one-shot prompting.

  • heavensteeth

       <button class="nav-btn" onclick="navigateToPage('index.html')">Home</button>
    
       ... 
    
       // Navigation functionality
       function navigateToPage(filename) {
           window.location.href = filename;
       }
    
    What shockingly poor code. Just disgraceful.

    This is bad for accessibility, bad for compatibility, and serves no purpose at all.

    • arcanemachiner

      Ha, that is bad. This is what happens when you use the cheap LLMs. (At least, I hope Opus 5.5 isn't pulling stunts that are this bad... It's not that bad when I look at my code.)

  • badsectoracula

    Everyone has a different taste.

    Personally i like the site's design, it looks like from late 2000s / early 2010s with its gradient rounded buttons, 1px offsets and chiseled text - i even clicked randomly on buttons because of how "buttony" they look :-P.

    IME this isn't what LLMs make by default, so if nothing else, i doubt it was a single prompt and most likely it was adjusted to fit the author's taste (that it sounds like doesn't align with yours).

  • pietroppeter

    I think there was very little effort put on this specific website. The effort has gone in the compiler itself (and the mainline compiler). Not that it would not benefit for a better website

  • konart

    idk, looks like a classic design from 2010. See QIP or AIMP Classic.

  • ilaksh

    It looks good on my phone and everything loads virtually instantly.

    It has a clean theme. The only knock I can see is that the buttons are a little retro. But I like them .

    Because of how fast every link loads and uncluttered the design is (not a single ad of any sort), I think it's better than 90% of websites.

    The fact that your fashion sense weighs other trend aesthetics so much over those things..

  • gsoff

    border-left: 3px solid color...;

    this gives the design a cheap ai-template vibe

  • asibahi

    Not everything you dont like is LLM-made.

lcnmrn

This feels archaic to me after seeing what is possible with Stricta. You get full Python syntax as low level as possible, safe, strict, no ARC/GC, memory access. https://github.com/lucianmarin/stricta

  • mirrir

    yes but that's your own project. its odd to not disclose that while calling other's work archaic.

  • platinumrad

    > Rust-Style Memory Safety (No GC / No ARC): Compile-time ownership and borrow checking with inferred BORROW / MUT_BORROW / MOVE parameter modes, non-lexical lifetimes (NLL), monomorphized deterministic @_drop_T / @_clone_T generation, compiler-lowered { i1, T } exception propagation, and runtime bounds traps—with zero garbage collection or reference counting overhead.

    There is either a catch here or you(r LLM) has solved multiple open research problems.

    • lcnmrn

      Get the code and ask your LLM if it did, otherwise provide an edge case to be fixed.

  • fishgoesblub

    > All Rights Reserved

    No thanks.

  • user8765439

    wow a 16k line single c file, revolutionary slop

Keyboard Shortcuts

j
Next item
k
Previous item
o / Enter
Open selected item
?
Show this help
Esc
Close modal / clear selection