While I was always a sourced-base/personalized distribution personality type, this is also a big part of why I moved to Gentoo in early 2004 (for amd64, not Risc-V / other embedded per your example). While Pentium-IV's very deep pipelines and compiler flag sensitivities (and the name itself for the fastest Penguin) drove the for-speed perception of the compile-just-for-my-system style, it really plays well to all customization/configuation hacker mindsets.
That is a fantastic historical parallel. The early amd64 days were arguably Gentoo's killer app moment. While the binary distributions were wrestling with the logistical nightmare of splitting repositories and figuring out the /lib64 vs /lib standard, Gentoo users just changed their CHOST, bootstrapped and were running 64-bit native. You nailed the psychology of it, too. The speed marketing was always a bit of a red herring. The ability to say "I do not want LDAP support in my mail client" and have the package manager actually respect that is cool. It respects the user's intelligence rather than abstracting it away.
Since you've been on the ride since '04, I'm curious to hear your thoughts. How do you feel the maintenance burden compares today versus the GCC 3.x era? With the modern binhost fallback and the improvements in portage, I feel like we now spend less time fighting rebuild loops than back then? But I wonder if long time users feel the same.
> The ability to say "I do not want LDAP support in my mail client" and have the package manager actually respect that is cool.
I tried Gentoo around the time that OP started using it, and I also really liked that aspect of it. Most package managers really struggle with this, and when there is configuration, the default is usually "all features enabled". So, when you want to install, say, ffmpeg on Debian, it pulls in a tree of over 250 (!!) dependency packages. Even if you just wanted to use it once to convert a .mp4 container into .mkv.
I also liked the idea when I used Gentoo 15 years ago but you quickly realise it doesn't make much sense.
You are trading off having a system able to handle everything you will throw at it, and having the same binaries as everyone else for, well, basically nothing. You have a supposedly smaller exploitable surface but you have to trust that the Gentoo patches cutting these things out don't introduce new vulnerabilities and don't inadvertently shut off hardening features. You have slightly smaller packages but I'm hard pressed to think of a scenario where it would matter in 2026.
To me, the worst debuggability and the inability to properly communicate with the source project make it a bad idea. I find Arch's pledge to only ship strictly vanilla software much more sensible.
> Since you've been on the ride since '04, I'm curious to hear your thoughts. How do you feel the maintenance burden compares today versus the GCC 3.x era? With the modern binhost fallback and the improvements in portage, I feel like we now spend less time fighting rebuild loops than back then? But I wonder if long time users feel the same.
I'm another one on it since the same era :)
In general stable has become _really_ stable, and unstable is still mostly usable without major hiccups. My maintenance burden is limited nowadays compared to 10y ago - pretty much running `emerge -uDN @world --quiet --keep-going` and fixing issues if any, maybe once a month I get package failures but I run a llvm+libcxx system and also package tests, so likely I get more issues than the average user on GCC.
For me these days it's not about the speed anymore of course, but really the customization options and the ability to build pretty much anything I need locally. I also really like the fact that ebuilds are basically bash scripts, and if I need to further customize or reproduce something I can literally copy-paste commands from the package manager in my local folder.
The project has successfully implemented a lot of by-default optimizations and best practices, and in general I feel the codebases for system packages have matured to the point where it's odd to run in internal compiler errors, weird dependency issues, whole-world rebuilds etc. From my point of view it also helped a lot that many compilers begun enforcing more modern and stricter C/C++ standards over time, and at the same time we got Github, CI workflows, better testing tools etc.
I run `emerge -e1 @world` maybe once a year just to shake out stuff lurking in the shadows (like stuff compiled with clang 19 vs clang 21), but it's really normally not needed anymore. The configuration stays pretty much untouched unless I want to enable a new USE for a new package I'm installing.
I am replying here as a kind of "better place to attach".
Anyway, to answer grandparent, I basically never had rebuild loops in 19 years.. just emerge -uU world every day or sometimes every week. I have been running the same base system since..let's see:
I have never once had to rebuild the whole system from scratch in those 19 years. (I've just rsync'd the rootfs from machine to machine as I upgraded HW and gradually rebuilt because as many others here have said, for me it wasn't about "perf of everything" or some kind of reproducible system - "more customization + perf of some things".) The upgrade from monolithic X11 to split X11 was "fun", though. /s
I do engage in all sorts of package.mask/per-package use/many global use. I have my own portage/local overlay for things where I disagree with upstream. I even have an automated system to "patch" my disagreements in. E.g, I control how fast I upgrade my LLVM junk so I do it on my own timeline. Mostly I use gcc. I control that, too. Any really slow individual build, basically.
If over the decades, they ever did anything that made it look like crazy amounts of rebuilds would happen, I'd tend to wait a few days/week or so and then figure something out. If some new dependency brings in a mountain of crap, I usually figure out how to block that.
gcc 3.3 to 3.4 was a big thing, and could cause some issues if people didnt follow the upgrade procedures, and also many c++ codebases would need minor adjustments.. this has been much much less of a problem since.
Additionally gentoo has become way more strict with use flag dependencies, and it also checks if binaries are depending on old libs, and doesnt remove them when updating a package, such that the "app depends on old libstdc++" doesnt happen anymore. It then automatically removes the old when nothing needs it anymore
I have been running gentoo since before 04, continously, and things pretty much just work. I would be willing to put money that I spend less time "managing my OS" than most who run other systems such as osx, windows, debian etc. Sure, my cpu gets to compile a lot, but thats about it.
And yes, the "--omg-optimize" was never really the selling point, but rather the useflags, where theres complete control. Pretty much nothing else comes close, and it is why gentoo is awesome
I'd say "the fastest" is a side effect of "allowing one to tune their systems to their utmost liking." -march=native, throw away unused bits and pieces, integrate modules into the kernel, replace bits and pieces with faster -- if more limited -- bits and pieces. And so on.
To be fair it was not that difficult to set create a pure 64 bit binary distro and there were a few of them. The real issue was to figure out how to do mixed 32/64 bit and this is where the fight about /lib directories originated. In a pure 64 bit distro the only way to run 32 bit binaries was to create a chroot with a full 32 bit installation. It took a while before better solutions were agreed to. This was an era of Flash and Acrobat Reader - all proprietary and all 32 bit only so people really cared about 32 bit.
Popularity is one thing { and probably more people use it than a non-existing new PLang you are only at the design phase of :-) }, but I think you misunderstood the backend idea. Nim has a javascript backend via `nim js` and on that backend you can use `emit` to do inline javascript, just as on the C/C++ backends you can use emit to put out C/C++. So, if you did do a zig backend, being able to emit inline zig would be part of that. It may still not be what you want, of course.
I see from `test/test_suite/compile_time_introspection/paramsof.c3t` that there is a way to get names & types of function parameters [1]. The language also seems to support default values { e.g. `int foo(int a, int b = 2) {...}` } and even call with keyword arguments/named parameters [2], but I couldn't find any `defaultof` or similar thing in the code. Does anyone know if this is just an oversight / temporary omission?
I don't think it is available no, and it's the first time I heard about such an idea. Thinking on it, this would allow such cursed code (love that :D). I'll put it up for discussion in the Discord as I'm interested in hearing whether `.defaultof` is a good idea or not.
One application of such a feature would be something like a "cligen.c3" (like the Nim https://github.com/c-blake/cligen or its /python/cg.py port or etc.). Mostly it just seems a more complete signature extraction, though. Any other kind of documentation system might benefit.
graalvm is literally 500x more overhead than a statically linked dash script.
Maybe not an issue for terminal UIs, but the article mentions both TUIs and CLI tools. A lot of people use CLI tools with a shell. As soon as you do `for file in *.c; do tool "$file"; done` (as a simple example), pure overhead on the order of even 10s of ms becomes noticeable. This is not theoretical. I recently had this trouble with python3, but I didn't want to rewrite all my f-strings into python2. So, it does arise in practice. (At least in the practice of some.)
Haven't checked graalvm in a long time. So, I got graalvm-jdk-25.0.1+8.1 for x86_64. It's a lot faster than Julia, and maybe 43ms is not slow in "human terms", but it's still pretty slow compared to some other competition. This was for a helloworld.jar [1]. On my laptop (i7-1370P, p-cores) using tim[2]:
That is just a normal JVM with optional Graal components if enabled, but not being used. The default memory allocation is based on a percentage of available memory and uncommitted (meaning its available for other programs). When people mention Graal they mean an AOT compiled executable that can be run without a JVM installed. Sometimes they may refer to Graal JIT as a replacement for C1/C2 available also in VM mode. You are using a plain HotSpot VM in server mode, as the optimized client mode was removed when desktop use-cases were deprioritized (e.g. JWS discontinued).
You are correct and I apologize for the misimpression.
`native-image -jar helloworld.jar helloworld` did take a whopping 17 seconds to compile, what might be the smallest possible project. That does makes me worry for iterations trying to get better perf in a context where startup overhead matters, BUT the executable it produced did run much faster - only about 1.8x slower than `tcc -run`:
Perl has 2 more shared libraries for ld.so to link, but is somehow faster. So, there may still be some room for improvement, but anyway, thank you for the correction.
(Also, I included 4 of the faster comparative programs to show additionally that the error bars are vaguely credible. In truth, on time shared OSes, the distributions have heavier tails than Gaussian and so a single +- is inadequate.)
--
EDIT: So, the ld.so/dynamic linking overhead was bothering me. I had to get a musl + zlib build environment going, but I did after a few minutes and then found this result with a fully statically linked binary executable:
398.2 +- 4.2 μs ./helloworld-sta>/n
(I should have noted earlier that /n -> /dev/null is just a convenience symlink I put on all my systems. Also, this is all on Linux 6.18.2 and the same CPU as before.)
Also, only around 4.2 MiB of RSS compared to ~1.8 MiB for dash & awk. So, 2.4x the space and only ~4X the time of static awk & dash. That might sound like criticism, but those are usually the efficiency champs. The binary size is kind of hefty (~2x like the RAM use):
$ size *sta
text data bss dec hex filename
2856644 3486552 3184 6346380 60d68c helloworld-sta
So, I guess, in 2026, Java start-up overhead is pretty acceptable. I hope that these "real numbers" can maybe add some precision to the discussion. Someone saying "mere milliseconds" just does not mean as much to me as 400 +- 4 microseconds, and perhaps there are others like me.
Thanks for the corrected evaluation. Just for your awareness, the HotSpot team is working on offering a spectrum of AOT/JIT options under the Leyden project [1]. Currently one has to choose between either a fully open world (JIT) or closed world (AOT) evaluation. The mixed world allows for a tunable knob, e.g. pre-warmup by AOT while retaining dynamic class loading and fast compile times for the application. This will soften the hard edges so developers can choose their constraints that best fit their application's deployment/usage model.
I'm not sure this opt-in/out "philosophical razor" is as sharp as one would like it to be. I think "optionality" alone oversimplifies and a person trying to adopt that rule for taxonomy would just have a really hard time and that might be telling us something.
For example, in Nim, at the compiler CLI tool level, there is opt-in/opt-out via the `--mm=whatever` flag, but, at the syntax level, Nim has both `ref T` and `ptr T` on equal syntactic footing . But then in the stdlib, `ref` types (really things derived from `seq[T]`) are used much more (since it's so convenient). Meanwhile, runtimes are often deployment properties. If every Linux distro had their libc link against -lgc for Boehm, people might say "C is a GC'd language on Linux". Minimal CRTs vary across userspaces and OS kernel/userspace deployment.. "What you can rely on/assume", I suspect the thrust behind "optionality", just varies with context.
Similar binding vagueness between properties (good, bad, ugly) of a language's '"main" compiler' and a 'language itself' and 'its "std"lib' and "common" runtimes/usage happen all the time (e.g. "object-oriented", often diluted by the vagueness of "oriented"). That doesn't even bring in "use by common dependencies" which is an almost independent axis/dimension and starts to relate to coordination problems of "What should even be in a 'std'-lib or any lib, anyway?".
I suspect this rule is trying to make the adjective "GC'd" do more work in an absolute sense than it realistically can given the diversity of PLangs (sometimes not so visible considering only workaday corporate PLangs). It's not always easy to define things!
> I think "optionality" alone oversimplifies and a person trying to adopt that rule for taxonomy would just have a really hard time and that might be telling us something.
I think optionality is what gives that definition weight.
Think of it this way. You come to a project like a game engine, but you find it's written in some language and discover for your usage you need no/minimal GC. How hard is it to minimize or remove GC. Assume that changing build flags will also cause problems elsewhere due to behavior change.
> Similar binding vagueness between properties (good, bad, ugly) of a language's '"main" compiler' and a 'language itself' and 'its "std"lib' and "common" runtimes/usage happen all the time (e.g. "object-oriented", often diluted by the vagueness of "oriented")
Vagueness is the intrinsic quality of human language. You can't escape it.
The logic is fuzzy, but going around saying stuff like "Rust is a GC language" because it has an optional, rarely used Arc/Rc, is just off the charts level of wrong.
PMunch and summarity both already said this, but because maybe code speaks louder than words (like pictures?)... This works:
from strutils as su import nil
echo su.split "hi there"
(You can put some parens () in there if you like, but that compiles.) So, you can do Python-style terse renames of imports with forced qualification. You just won't be able to say "hi there".(su.split) or .`su.split` or the like.
You can revive that, though, with a
template suSplit(x): untyped = su.split x
echo "hi there".suSplit`
That most Nim code you see will not do this is more a cultural/popularity thing that is kind of a copy-paste/survey of dev tastes thing. It's much like people using "np" as the ident in `import numpy as np`. I was doing this renaming import before it was even widely popular, but I used capital `N` for `numpy` and have had people freak out at me for such (and yet no one freaking out at Travis for not just calling it `np` in the first place).
So, it matters a little more in that this impacts how you design/demo library code/lib symbol sets and so on, but it is less of a big deal than people make it out to be. This itself is much like people pretending they are arguing about "fundamental language things", when a great deal of what they actually argue about are "common practices" or conventions. Programming language designers have precious little control over such practices.
The only way to really test out a programming language is by trying it out or reading how someone else approached a problem that you're interested in/know about.
There are over 2200 nimble packages now. Maybe not an eye-popping number, but there's still a good chance that somewhere in the json at https://github.com/nim-lang/packages you will find something interesting. There is also RosettaCode.org which has a lot of Nim example code.
This, of course, does not speak to the main point of this subthread about the founder but just to some "side ideas".
In this case, it compiles & runs fine with floats (if you just delete the type constraint "Fibable") because the string "1" can be implicitly converted into float(1) { or 1.0 or 1f64 or float64(1) or 1'f64 or ..? }. You can think of the "1" and "2" as having an implicit "T(1)", "T(2)" -- which would also resolve your "doesn't work for me" if you prefer the explicitness. You don't have to trust me, either. You can try it with `echo fib(7.0)`.
Nim is Choice in many dimensions that other PLang's are insistently monosyllabic/stylistic about - gc or not or what kind, many kinds of spelling, new operator vs. overloaded old one, etc., etc., etc. Some people actually dislike choice because it allows others to choose differently and the ensuing entropy creates cognitive dissonance. Code formatters are maybe a good example of this? They may not phrase opposition as being "against choice" as explicitly as I am framing it, but I think the "My choices only, please!" sentiment is in there if they are self-aware.
But given the definition of Fibable, it could be anything that supports + and - operators. That could be broader than numbers. You could define it for sets for example. How do you add the number 1 to the set of strings containing (“dog”, “cat”, and “bear”)? So I suppose I do have a complaint about Fibable, which is that it’s underconstrained.
Granted, I don’t know nim. Maybe you can’t define + and - operators for non numbers?
Araq was probably trying to keep `Fibable` short for the point he was trying to make. So, your qualm might more be with his example than anything else.
You could add a `SomeNumber` predicate to the `concept` to address that concern. `SomeNumber` is a built-in typeclass (well, in `system.nim` anyway, but there are ways to use the Nim compiler without that or do a `from system import nil` or etc.).
Unmentioned in the article is a very rare compiler/PLang superpower (available at least in Nim 1, Nim 2) - `compiles`. So, the below will print out two lines - "2\n1\n":
when compiles SomeNumber "hi": echo 1 else: echo 2
when compiles SomeNumber 1.0: echo 1 else: echo 2
Last I knew "concept refinement" for new-style concepts was still a work in progress. Anyway, I'm not sure what is the most elegant way to incorporate this extra constraint, but I think it's a mistake to think it is unincorporatable.
To address your question about '+', you can define it for non-SomeNumber, but you can also define many new operators like `.+.` or `>>>` or whatever. So, it's up to your choice/judgement if the situation calls for `+` vs something else.
I think the example was chosen only for familiarity and is otherwise not great. Though it was the familiarity itself that probably helped you to so easily criticize it. So, what do I know? :-)
FWIW, the "catenation operator" in the Nim stdlib is ampersand `&`, not `+` which actually makes it better than most PLangs at visually disambiguating things like string (or other dynamic array, `seq[T]` in Nim) concatenation from arithmetic. So, `a&b` means `b` concatenated onto the end of `a` while `a+b` is the more usual commutative operation (i.e. same as `b+a`). Commutativity is not enforced by the basic dispatch on `+`, though such might be add-able as a compiler plugin.
Mostly, it's just a very flexible compiler / system.. like a static Lisp with a standard surface syntax closer to Python with a lot of parentheses made optional (but I think much more flexible and fluid than Python). Nim is far from perfect, but it makes programming feel like so much less boilerplate ceremony than most alternatives and also responds very well to speed/memory optimization effort.
There is no direct argument/guidence that I saw for "when to use them", but masked arrays { https://numpy.org/doc/stable/reference/maskedarray.html } (an alternative to sentinels in array processing sub-languages) have been in NumPy (following its antecedents) from its start. I'm guessing you could do a code-search for its imports and find arguments pro & con in various places surrounding that.
From memory, I have heard "infecting all downstream" as both "a feature" and "a problem". Experience with numpy programs did lead to sentinels in the https://github.com/c-blake/nio Nim package, though.
Another way to try to investigate popularity here is to see how much code uses signaling NaN vs. quiet NaN and/or arguments pro/con those things / floating point exceptions in general.
I imagine all of it comes down to questions of how locally can/should code be forced to confront problems, much like arguments about try/except/catch kinds of exception handling systems vs. other alternatives. In the age of SIMD there can be performance angles to these questions and essentially "batching factors" for error handling that relate to all the other batching factors going on.
Today's version of this wiki page also includes a discussion of Integer Nan: https://en.wikipedia.org/wiki/NaN . It notes that the R language uses the minimal signed value (i.e. 0x80000000) of integers for NA.
To be clear, I am not taking some specific position, but I think all these topics inform answers to your question. I think it's something with trade-offs that people have a tendency to over-simplify based on a limited view.
>To be clear, I am not taking some specific position, but I think all these topics inform answers to your question. I think it's something with trade-offs that people have a tendency to over-simplify based on a limited view.
That's fair, I wasn't dimsissing the practice but rather just commenting that it's a shame the author didn't clarify their preference.
I don't think the popularity angle is a good proxy for usefulness/correction of the practice. Many factors can influence popularity.
Performance is a very fair point, I don't know enough to understand the details but I could see it being a strong argument. It is counter intuitive to move forward with calculations known to be useless, but maybe the cost of checking all calculations for validity is larger than the savings of skipping early the invalid ones.
There is a catch though. Numpy and R are very oriented to calculation pipelines, which is a very different usecase to general programming, where the side effects of undetected 'corrupt' values can be more serious.
The conversation around Nim for the past 20 years has been rather fragmented - IRC channels, Discord channels (dozens, I think), later the Forum, Github issue threads, pull request comment threads, RFCs, etc. Araq has a tendency to defend his ideas in one venue (sometimes quite cogently) and leave it to questioners to dig up where those trade-off conversations might be. I've disliked the fractured nature of the conversation for the 10 years I've known about it, but assigned it to a kind of "kids these days, whachagonnado" status. Many conversations (and life!) are just like that - you kind of have to "meet people where they are".
Anyway, this topic of "error handling scoping/locality" may be the single most cross-cutting topic across CPUs, PLangs, Databases, and operating systems (I would bin Numpy/R under Plangs+Databases as they are kind of "data languages"). Consequently, opinions can be very strong (often having this sense of "Everything hinges on this!") in all directions, but rarely take a "complete" view.
If you are interested in "fundamental, not just popularity" discussions, and it sounds like you are, I feel like the database community discussions are probably the most "refined/complete" in terms of trade-offs, but that could simply be my personal exposure, and DB people tend to ignore CPU SIMD because it's such a "recent" innovation (hahaha, Seymore Cray was doing it in the 1980s for the Cray-3 Vector SuperComputer). Anyway, just trying to help. That link to the DB Null page I gave is probably a good starting point.
reply