LLVM is a trap. You bootstrap extra fast, you get all sorts of optimization passes and platforms for free, but you lose out on the ability to tune the final optimization passes and performance of the linking stages.
I think we'll see cranelift take off in Rust quite soon, though I also think it wouldn't be the juggernaut of a language if they hadn't stuck with LLVM those early years.
Go seems to have made the choice long ago to eschew outsourcing of codegen and linking and done well for it.
Is it? I think Rust is a great showcase for why it isn't. Of course it depends somewhat on your compiler implementation approach, but actual codegen-to-LLVM tends to only be a tiny part of the compiler, and it is not particularly hard to replace it with codegen-to-something-else if you so desire. Which is why there is now codegen_cranelift, codegen_gcc, etc.
The main "vendor lock-in" LLVM has is if you are exposing the tens of thousands of vendor SIMD intrinsics, but I think that's inherent to the problem space.
Of course, whether you're going to find another codegen backend (or are willing to write one yourself) that provides similar capabilities to LLVM is another question...
> You bootstrap extra fast, you get all sorts of optimization passes and platforms for free, but you lose out on the ability to tune the final optimization passes and performance of the linking stages.
You can tune the pass pipeline when using LLVM. If your language is C/C++ "like", the default pipeline is good enough that many such users of LLVM don't bother, but languages that differ more substantially will usually use fully custom pipelines.
> I think we'll see cranelift take off in Rust quite soon, though I also think it wouldn't be the juggernaut of a language if they hadn't stuck with LLVM those early years.
I'd expect that most (compiled) languages do well to start with an LLVM backend. Having a tightly integrated custom backend can certainly be worthwhile (and Go is a great success story in that space), but it's usually not the defining the feature of the language, and there is a great opportunity cost to implementing one.
It is a “trap” in the sense that it limits the design space for your language by forcing some of the choices implicit in C++. Rust went with slow build times like C++ and so was not held back in this dimension by LLVM.
Nothing about LLVM is a trap for C++ as that is what it was designed for.
As I can't update my comment, here goes some info out of Turbo Pascal 5.5 marketing brochure,
> Fast! Compiles 34 000 lines of code per minute
This was measured on a IBM PS/2 Model 60.
So lets put this into perspective, Turbo Pascal 5.5 was released in 1989.
IBM PS/2 Model 60 is from 1987, with a 80286 running at 10 MHz, limited by 640 KB, with luck one would expad it up to 1 MB and use HMA, in what concerns using it with MS-DOS.
Now projecting this to 2025, there is no reason that compiled languages, when using a limited set of optimizations like TP 5.5 on their -O0, can't be flying in their compilation times, as seen in good examples like D and Delphi, to use two examples of expressive languages with rich type systems.
Old versions of Turbo Pascal running in FreeDOS on the bare metal of a 21st century PC is how fast and responsive I wish all software could be, but never is. Press a key and before you have time to release it the operation you started has already completed.
A problem is how people have started depending on the optimizations. "This tower of abstractions is fine, the optimizer will remove it all" Result is some modern idioms run slow as molasses without optimization, and you can't really use o0 at all.
Turbo Pascal was an outlier in 1989 though. The funny thing is that I remember Turbo C++ being an outlier in the opposite direction.
In my computer science class (which used Turbo C++), people would try to get there early in order to get one of the two 486 machines, as the compilation times were a huge headache (and this was without STL, which was new at the time).
I recently saw an article about someone improving the machine code generation time of an assembler, here; I idly noticed that the scale was the same number of instructions we had in the budget to compile whole lines of code (expressions & all) "back in the day". It was weird. Of course, we're fighting bandwidth laws, so if you looked at the wall clock time, the machine code generation time was very good in an absolute sense.
> be, before C and C++ took over the zeitgeist of compilation times.
I wouldn’t put them together. C compilation is not the fastest but fast enough to not be a big problem. C++ is a completely different story: not only it orders of magnitude slower (10x slower probably not the limit) on some codebases compiler needs a few Gb RAM (so you have to set -j below the number of CPU cores to avoid OOM).
Back in 1999 - 2003, when I was working on a product mixing Tcl and C, the builds took one hour per platform, across Aix, Solaris, HP-UX, Red-Hat Linux, Windows 2000/NT build servers.
C++ builds can be very slow versus plain old C, yes, assuming people do all mistakes there can be done.
Like overuse of templates, not using binary libraries across modules, not using binary caches for object files (ClearMake style already available back in 2000), not using incremental compilation and incremental linking.
To this day, my toy GTK+ Gtkmm application that I used for a The C/C++ Users Journal article, and have ported to Rust, compiles faster in C++ than Rust in a clean build, exactly because I don't need to start from the world genesis for all dependencies.
I talked a bit about this at the Rust All Hands back in May.
A lot of Rust packages that people ust are setup more like header-only libraries. We're starting to see more large libraries that better fit the model of binary libraries, like Bevy and Gitoxide. I'm laying down a vague direction for something more binary-library like (calling them opaque dependencies) as part of the `build-std` effort (allow custom builds of the standard library) as that is special cased as a binary library today.
You only build the world occasionally, like when cloning or staring a new project, or when upgrading your rustc version. It isn't your development loop.
I do think that dynamic libraries are needed for better plugin support, though.
Unless a shared dependency gets updated, RUSTFLAGS changes, a different feature gets activated in a shared dependency, etc.
If Cargo had something like binary packages, it means they would be opaque to the rest of your project, making them less sensitive to change.
Its also hard to share builds between projects because of the sensitivity to differences.
I have seen projects where ninja instead of make (both generated by cmake) is able to cleverly invoke the compiler such that CPU is saturated and ram isn't exhausted, and make couldn't (it reached oom).
I have first hand experience of painfully slow C# compile times. Sprinkle in a few extra slow things like EDMX generated files (not C# but part of the MS ecosystem) and it has no business being in a list of fast compiling languages.
As much as I love F#, I wouldn't add it to a list of fast compilers; although I really appreciate that it supports actual file-scoped optimizations and features for compile-time inlining.
I wouldn't even put C# in there - minute builds are like nothing unusual, and my only experience with Java is Android but that was really bad too.
They aren't C++ levels bad, but they are slow enough to be distracting/flow breaking. Something like dart/flutter or even TS and frontend with hot reload is much leaner. Comparing to fully dynamic languages is kind of unfair in that regard.
I did not try Go yet but from what I've read (and also seeing the language design philosophy) I suspect it's faster than C#/Java.
Right, the work I get paid to do is often C# and literally yesterday I was cursing a slow C# build. Why is it slow? Did one of the other engineers screw up something about the build, maybe use a feature that's known to take longer to compile, or whatever? I don't know or care.
This idea that it's all sunshine and lollipops for other languages is wrong.
> Both Go and Ocaml have really, really fast compilers. They did the works from the get go, and now wont have to pay the price.
People tend to forget that LLVM was pretty much that for the C/C++ world. Clang was worlds ahead of GCC when first released (both in speed and in quality of error messages), and Clang was explicitly built from the ground up to take advantage of LLVM.
A trap? New language projects would never be able to develop the architecture support and optimizer. There's nothing new in the linking stage you couldn't submit as a patch to LLVM upstream.
Additionally any language that wants to be a C++ replacement, keeps a dependency on C++ as long as LLVM is part of its tooling infrastructure.
For me beyond the initial adoption hump, programming languages should bootstrap themselves, if nothing else, reduces the urban myth that C or C++ have to always be part of the equation.
Not at all, in most cases it is a convenience, writing usable compilers takes time, and it is an easy way to get everything else going, especially when it comes to porting across multiple platforms.
However that doesn't make them magical tools, without which is impossible to write compilers.
That is one area where I am fully on board with Go team's decisions.
I hope cranelift continues to grow too. LLVM tragically has become fork hell. Every cpu vendor now seems to have a fork with enhancements for their cpu arch in some closed source package. It’s hellish as soon as you want to use another language frontend or find the inevitable compiler bug.
So naturally a way to quickly reach the stage where a compiler is avaible for a brand new language, without having to write all compiler stages by hand.
A kind of middleware for writing compilers, if you wish.
MLIR is part of LLVM tooling, is the evolution of LLVM IR.
"While it does not have the aggressive optimization stance (or complexity) of larger compilers such as LLVM or gcc, its output is often not too far off: the same paper showed Cranelift generating code that ran ~2% slower than V8 (TurboFan) and ~14% slower than an LLVM-based system."
The artifact's executions speed doesn't seem to be Cranelift's priority. They're instead focusing on compilation speed and security. Those are still useful in Rust for debug builds at least. That's when we need a quick turnaround time and as much verification as possible.
Besides the potential typo noted by another comment, I'd also note that the benchmarks are on a WebAssembly compiler, which has already gone through the LLVM wringer and thus undergone many high-level optimizations, including those reliant on alias analysis (which Cranelift is quite conservative on, iirc). I'd also imagine that the translation from LLVM IR to WASM discards some useful information that the LLVM backend could otherwise have used to generate better assembly.
That's not to say Cranelift isn't a fantastic piece of tech, but I wouldn't take the "14%" or "24%" number at face value.
If you care about compilation speed b/c it's slowing down development - wouldn't it make sense to work on an interpreter?
Maybe I'm naiive, but it seems the simpler option
Compiling for executable-speed seems inherently orthogonal to compilation time
With an interpreter you have the potential of lots of extra development tools as you can instrument the code easily and you control the runtime.
Sure, in some corner cases people need to only be debugging their full-optimization-RELEASE binary and for them working on a interpreter, or even a DEBUG build just doesn't makes sense. But that's a tiny minority. Even there, you're usually optimizing a hot loop that's going to be compiling instantly anyway
"Trying to understand the Zig goals a bit better..
If you care about compilation speed b/c it's slowing down development - wouldn't it make sense to work on an interpreter?"
I've heard Rust classified as Safety, Performance, Usability in that order and Zig as Performance, Usability, Safety. From that perspective the build speed-ups make sense while an interpreter would only fit a triple where usability comes first.
I don't think it's right to say Zig is more performance-focused than Rust. Based on their designs, they should both be able to generate similar machine code. Indeed, benchmarks show that Rust and Zig perform very similarly: https://programming-language-benchmarks.vercel.app/rust-vs-z...
This is why the value proposition of Zig is a question for people, because if you got measurably better performance out of Zig in exchange for worse safety that would be one thing. But Rust is just as performant, so with Zig you're exchanging safety for expressivity, which is a much less appealing tradeoff; the learning from C is that expressivity without safety encourages programmers to create a minefield of exploitable bugs.
I don't think Zig is about expressivity over safety. Many people liken Zig to C, like with Rust to C++.
Zig is a lot safer than C by nature. It does not aim to be as safe as Rust by default. But the language is a lot simpler. It also has excellent cross-compiling, much faster compilation (which will improve drastically as incremental compilation is finished), and a better experience interfacing with/using C code.
So there are situations where one may prefer Zig, but if memory-safe performance is the most important thing, Rust is the better tool since that is what it was designed for.
The value proposition in Zig is that it has far and away the best C interoperability I've seen out of a language that wasn't also a near-superset of it. This makes migrations away from C towards Zig a lot more palatable than towards Rust unless security is your primary concern.
On the contrary I would not call Zig very C-like at all, and that's what makes the interoperability so impressive to me.
Zig feels like a small yet modern language that lacks many of C's footguns. It has far fewer implicit type conversions. There is no preprocessor. Comptime allows for type and reflection chicanery that even C++ would be jealous of.
Yet I can just include a C header file and 24 times out of 25, it just works. Same thing if you run translate-c on a c file, and as a bonus, it reveals the hoops that the tooling has to go through to preserve the fiddly edge cases while still remaining somewhat readable.
It's much easier to write ultra low latency code in Zig because doing so requires using local bump-a-pointer allocators, and the whole Zig standard library is built around that (everything that allocates must take the allocator as a param). Rust on the other hand doesn't make custom allocators easy to use, partially because proving they're safe to the borrow checker is tricky.
> expressivity without safety encourages programmers to create a minefield of exploitable bugs.
Also, psychologically, the combination creates a powerful and dangerously false sensation of mastery in practitioners, especially less experienced ones. Part of Zig's allure, like C's bizarre continued relevance, is the extreme-sport nature of the user experience. It's exhilarating to do something dangerous and get away with it.
Please — rust advocates need to realize that you're not winning anyone to rust with this kind of outlandish comments.
Zig is not an extreme sports experience. It has much better memory guarantees than C. It's clearly not as good as rust if that's your main concern but rust people need to also think long and hard about why rust has trouble competing with languages like Go, Zig and C++ nowadays.
I'm not a Rust advocate. I'm a "the whole debate is silly: 99% of the time you want a high level language with a garbage collector" advocate.
What's the use case for Zig? You're in that 1% of projects in which you need something beyond what a garbage collector can deliver and, what, you're in the 1% of the 1% in which Rust's language design hurts the vibe or something?
You can also get that "safer, but not Safe" feeling from modern C++. So what? It doesn't really matter whether a language is "safer". Either a language lets you prove the absence of certain classes of bug or it doesn't. It's not like C, Zig, and Rust form a continuum of safety. Two of these are different from the third in a qualitative sense.
> What's the use case for Zig? You're in that 1% of projects in which you need something beyond what a garbage collector can deliver and, what, you're in the 1% of the 1% in which Rust's language design hurts the vibe or something?
You want a modern language that has package management, bounds checks and pointer checks out of the box. Zig is a language you can pick up quickly whereas rust takes years to master. It's a good replacement to C++ if you're building a game engine for example.
> Either a language lets you prove the absence of certain classes of bug or it doesn't. It's not like C, Zig, and Rust form a continuum of safety. Two of these are different from the third in a qualitative sense.
Again repeating my critique from the previous comment – yes Zig brings in additional safety compared to C. Dismissing all of that out of hand does not convince anyone to use rust.
> Compiling for executable-speed seems inherently orthogonal to compilation time
That true at the limit. As is often the case there's a vast space in the middle, between the extremes of ultimate executable speed and ultimate compiler speed, where you can make optimizations that don't trade off against each other. Where you can make the compiler faster while producing the exact same bytecode.
This is why games ship interpreters and include a lot of code scripting the gameplay (e.g in Lua) rather than having it all written in C++. You get fast release builds with a decent framerate and the ability to iterate on a lot of the “business logic” from within the game itself.
That's only true for badly written game code though ;)
Even with C++ and heavy stdlib usage it's possible to have debug builds that are only around 3..5 times slower than release builds in C++. And you need that wiggle room anyway to account for lower end CPUs which your game should better be able to run on.
I've never done it, but I just find it hard to believe the slow down would be that large. Most of the computation is on GPU, and you can set your build up such that you link to libraries built at different compilation optimizations.. and they're likely the ones doing most of the heavy lifting. You're not rebuilding all of the underlying libs b/c you're not typically debugging them.
EDIT:
if you're targeting a console.. why would you not debug using higher end hardware? If anything it's an argument in favor of running on an interpreter with a very high end computer for the majority of development..
Yeah, around 3x slower is what I've seen when I was still working on reasonably big (around a million lines of 'orthodox C++' code) PC games until around 2020 with MSVC and without stdlib usage.
This was building the entire game code with the same build options though, we didn't bother to build parts of the code in release mode and parts in debug mode, since debug performance was fast enough for the game to still run in realtime. We also didn't use a big integrated engine like UE, only some specialized middleware libraries that were integrated into the build as source code.
We did spend quite some effort to keep both build time and debug performance under control. The few cases were debug performance became unacceptable were usually caused by 'accidentially exponential' problems.
> Most of the computation is on GPU
Not in our case, the GPU was only used for rendering. All game logic was running on the CPU. IIRC the biggest chunk was pathfinding and visibility/collision/hit checks (e.g. all sorts of NxM situations in the lower levels of the mob AI).
Ah okay, then that makes sense. It really depends on your build system. My natural inclination would be to not have a pathfinding system in DEBUG unless I was actively working on it. But it's not always very easy to set things up to be that way
The slowdown can be enormous if you use SIMD, I believe MSVC emits a write to memory after every single SIMD op in debug, thus the code is incredibly slow(more than 10x).
If you have any SIMD kernels you will suffer, and likely switch to release, with manual markings used to disable optimizations per function/file.
If you target consoles, you’re already on the lowest end hardware you’ll run on. It’s extremely rare to have much (if any) headroom for graphically competitive games. 20% is a stretch, 3x is unthinkable.
Then I would do most of the development work that doesn't directly touch console APIs on a PC version of the game and do development and debugging on a high-end PC. Some compilers also have debug options which apply optimizations while the code still remains debuggable (e.g. on MSVC: https://learn.microsoft.com/en-us/visualstudio/debugger/how-...) - don't know about Sony's toolchain though. Another option might be to only optimize the most performance sensitive parts while keeping 90% of the code in debug mode. In any case, I would try everything before having to give up debuggable builds.
It doesn’t matter. You will hit (for example) rendering bugs that only happen on a specific version of a console with sufficient frequency that you’ll rarely use debug builds.
Debug builds are used sometimes, just not most of the time.
The gold standard for compiler speed so far is TCC. (Fabrice Bellard)
It is not even super optimized (single thread, no fancy tricks) but it is so far unbeaten by a large margin. Of course I use Clang for releases, but the codegen of tcc is not even awful.
I'd call it not terribly slow. IIRC it didn't hit 100k lines on an M2, while I saw over two million on my own compiler (I used multiple instances of tcc since I didn't want to deal with implementing debug info on mac)
One the one hand the go type system is a joke compared to typescript so the typescript compiler has a much harder job of type checking. On the other hand once type checking is done typescript just needs to strip the types and it's done while go needs to optimize and generate assembly.
Yes. I once was thinking I should write a x86-64 codegen for my own compiler, but using tcc I was able to reach 3 million lines in a second (using multi-threading and multiple instances of tcc) so I called it a day. I never did check out if the linking was in parallel
Not yet AFAIK, but unlike the 'inbetween' versions since C99, C23 actually has a number of really useful features, so I expect that a couple of projects will start using C23 features. The main problem is as always MSVC, but then maybe its time to ditch MSVC support since Microsoft also seems to have abandondend it (not just the C compiler, but also the C++ compiler isn't seeing lots of updates since everybody working on Visual Studio seems to have been reassigned to work in the AI salt mines).
(not sure how much Apple/Google even cared about the C frontend before though, but at least keeping the C frontend and stdlib uptodate by far doesn't require as much effort as C++).
Apple used to care before Swift, because Objective-C unlike C++, is a full superset from C.
Most of the work going into LLVM ecosystem is directly into LLVM tooling itself, clang was started by Apple, and Google picked up on it.
Nowadays they aren't as interested, given Swift, C++ on Apple platforms is mostly for MSL (C++14 baseline) and driver frameworks (also a C++ subset), Google went their own way after the ABI drama, and they care about what fits into their C++ style guide.
I know Intel is one of the companies that picked up some of the work, yet other compiler vendors that replaced their proprietary forks with clang don't seem that eager to contribute upstream, other than LLVM backends for their platforms.
I use tcc with Nim all the time and it yields almost interpreted REPL-like code-iteration speed (like 100-500ms per compile-link - almost ENTER-see-results).
As a warning, you need to be sure --lineDir=off in your nim.cfg or config.nims (to side-step the infinite loop mentioned in https://github.com/nim-lang/Nim/pull/23488). You may also need to --tlsEmulation=on if you do threads and you may want to --passL=-lm.
tcc itself is quite fast. Faster for me than a Perl5 interpreter start-up for me (with both on trivial files). E.g., on an
i7-1370P:
I've never had an issue with ccache and I've been using it forever. Are you using it over a networked filesystem or something like that for it to become confused about what needs recompiling?
There they go, making the same stupid mistakes as most other ORMs out there; by 1:1-mapping structs to tables as well as inventing their own SQL-dialect.
Are their ghostty builds statically linked by default or does zig follow in the footsteps of the C/C++ world and dynamically link everything? If statically linking (like rust) then a huge part of the incremental build times for apps with external dependencies is the lack of incremental linking in any of the major linkers. Wild linker is working on becoming the answer to this: https://github.com/davidlattimore/wild
ghostty statically links all its zig dependencies. The only dynamic libraries I can see are the usual suspects (libc, X11, Wayland, OpenGL, GTK, glib, libadwaita, etc).
I've really appreciated the incremental builds with zig - I really like having a single static binary, and I have an application which uses a few different libraries like SQLite and luau and it compiles at nearly Go speeds.
It's finally as snappy as recompiling the "Borland Pascal version of Turbo Vision for DOS" was on an Intel 486 in 1995, when I graduated from high school...
They C version of Turbo Vision was 5-10x slower to compile at that time too.
Turbo Vision is a TUI windowing framework, which was used for developing the Borland Pascal and C++ IDEs. Kinda like a character mode JetBrains IDE in 10 MB instead of 1000 MB...
Is there a straightforward path to building Zig with polyglot build systems like Bazel and Buck2? I'm worried Zig's reliance on Turing complete build scripts will make building (and caching) such code difficult in those deterministic systems. In Rust, libraries that eschew build.rs are far preferable for this reason. Do Zig libraries typically have a lot of custom build setup?
FYI, build scripts are completely optional. Zig can build and run individual source code files regardless of build scripts (`build.zig`). You may need to decipher the build script to extract flags, but that's pretty much it. You can integrate Zig into any workflow that accepts GCC and Clang. (Note: `zig` is also a drop-in replacement C compiler[1])
Tangential: I've been writing a Java library that is pretty much a JNI shim to a C library. Looking at building a dynamic library for each of the platform / architecture combo is looking like a huge pain in the arse; I'm toying with the idea of using the zig compiler to do the compile / cross compiles and simply be done with it.
do yourself a favor and try it, Zig ships with headers and libc sources and its own LLVM distribution and makes cross compilation so easy it's just ridiculous, you literally don't have to do anything
It's a bold strategy - let's see if it works out for them.
But from what I remember, this still uses the LLVM backend, right? Sure, you can beat LLVM on compilation speed and number of platforms supported, but when it comes to emitting great assembly, it is almost unbeatable.
It only uses the LLVM backend for release builds. For debug builds Zig now defaults to self-hosted (on supported platforms, only x86_64 right now). Debug builds are directly in the hot path of testing your software (this includes building test binaries), and the number of iterations you do on a debug build are multiple orders of magnitude higher than release builds, so this is a good tradeoff.
I don't really get the obsession w/ compiler performance. At the end of the day you are trading performance optimizations (inlining, dead code elimination, partial evaluation, etc) for compile times. If you want fast compile times then use a compiler w/ no optimization passes, done. The compile times are now linear w/ respect to lines of code & there is provably no further improvement you can make on that b/c any further passes will add linear or superlinear amount of overhead depending on the complexity of the optimization.
The reason to care about compile time is because it affects your iteration speed. You can iterate much faster on a program that takes 1 second to compile vs 1 minute.
Time complexity may be O(lines), but a compiler can be faster or slower based on how long it takes. And for incremental updates, compilers can do significantly better than O(lines).
In debug mode, zig uses llvm with no optimization passes. On linux x86_64, it uses its own native backend. This backend can be significantly faster to compile (2x or more) than llvm.
Zig's own native backend is designed for incremental compilation. This means, after the initial build, there will be very little work that has to be done for the next emit. It needs to rebuild the affected function, potentially rebuild other functions which depend on it, and then directly update the one part of the output binary that changed. This will be significantly faster than O(n) for edits.
> The reason to care about compile time is because it affects your iteration speed. You can iterate much faster on a program that takes 1 second to compile vs 1 minute.
Color me skeptical. I've only got 30 years of development under the belt, but even a 1 minute compile time is dwarfed by the time it takes to write and reason about code, run tests, work with version control, etc.
Further, using Rust as an example, even a project which takes 5 minutes to build cold only takes a second or two on a hot build thanks to caching of already-built artifacts.
Which leaves any compile time improvements to the very first time the project is cloned and built.
Consequently, faster compile times would not alter my development practices, nor allow me to iterate any faster.
I've worked with Delphi where a recompile takes a few seconds, and I've worked with C++ where a similar recompile takes a long time, often 10 minutes or more.
I found I work very differently in the two cases. In Delphi I use the compiler as a spell checker. With the C++ code I spent much more time looking over the code before compiling.
Sometimes though you're forced to iterate over small changes. Might be some bug hunting where you add some debug code that allows you to narrow things a bit more, add some more code and so on. Or it might be some UI thing where you need to check to see how it looks in practice. In those cases the fast iteration really helps. I found those cases painful in C++.
For important code, where the details matter, then yeah, you're not going to iterate as fast. And sometimes forcing a slower pace might be beneficial, I found.
> Consequently, faster compile times would not alter my development practices, nor allow me to iterate any faster.
I think the web frontend space is a really good case for fast compile times. It's gotten to the point that you can make a change, save a file, the code recompiles and is sent to the browser and hot-reloaded (no page refresh) and your changes just show up.
The difference between this experience and my last time working with Ember, where we had long compile times and full page reloads, was incredibly stark.
As you mentioned, the hot build with caching definitely does a lot of heavy lifting here, but in some environments, such as a CI server, having minutes long builds can get annoying as well.
> Consequently, faster compile times would not alter my development practices, nor allow me to iterate any faster.
Maybe, maybe not, but there's no denying that faster feels nicer.
> Maybe, maybe not, but there's no denying that faster feels nicer.
Given finite developer time, spending it on improved optimization and code generation would have a much larger effect on my development. Even if builds took twice as long.
> I'm much more productive when I can see the results within 1 or 2 seconds.
That's my experience today with all my Rust projects. Even though people decry the language for long compile times. As I said, hot builds, which is every build while I'm hacking, are exactly that fast already. Even on the large projects. Even on my 4 year old laptop.
On a hot build, build time is dominated by linking, not compilation. And even halving a 1s hot build will not result in any noticeable change for me.
> Further, using Rust as an example, even a project which takes 5 minutes to build cold only takes a second or two on a hot build thanks to caching of already-built artifacts.
So optimizing compile times isn’t worthwhile because we already do things to optimize compile times? Interesting take.
What about projects for which hot builds take significantly longer than a few seconds? That’s what I assumed everyone was already talking about. It’s certainly the kind of case that I most care about when it comes to iteration speed.
> So optimizing compile times isn’t worthwhile because we already do things to optimize compile times?
That seems strange to you? If build times constituted a significant portion of my development time I might think differently. They don't. Seems the compiler developers have done an excellent job. No complaints. The pareto principle and law of diminishing returns apply.
> What about projects for which hot builds take significantly longer than a few seconds?
A hot build of Servo, one of the larger Rust projects I can think of off the top of my head, takes just a couple seconds, mostly linking. You're thinking of something larger? Which can't be broken up into smaller compilation units? That'd be an unusual project. I can think of lots of things which are probably more important than optimizing for rare projects. Can't you?
The part that seems strange to me is your evidence that multi-minute compile times are acceptable being couple-second compile times. It seems like everyone actually agrees that couple-second iteration is important.
It seems like you might have missed a few words in this comment, I'm honestly having trouble parsing it to figure out what you're trying to say.
Just for fun, I kicked off a cold build of Bevy, the largest Rust project in my working folder at the moment, which has 830 dependencies, and that took 1m 23s. A second hot build took 0.22s. Since I only have to do the cold build once, right after cloning the repository which takes just as long, that seems pretty great to me.
Are you telling me that you need faster build times than 0.22s on projects with more than 800 dependencies?
> > The reason to care about compile time is because it affects your iteration speed. You can iterate much faster on a program that takes 1 second to compile vs 1 minute.
> Color me skeptical. I've only got 30 years of development under the belt, but even a 1 minute compile time is dwarfed by the time it takes to write and reason about code, run tests, work with version control, etc.
If your counterexample to 1-minute builds being disruptive is a 1-second hot build, I think we’re just talking past each other. Iteration implies hot builds. A 1-minute hot build is disruptive. To answer your earlier question, I don’t experience those in my current Rust projects (where I’m usually iterating on `cargo check` anyway), but I did in C++ projects (even trivial ones that used certain pathological libraries) as well as some particularly badly-written Node ones, and build times are a serious consideration when I’m making tech decisions. (The original context seemed language-agnostic to me.)
I see. Perhaps I wasn't clear. I've never encountered a 1 minute hot build in Rust, and given my experience with large Rust codebases like Bevy I'm not even sure such a thing exists in a real Rust codebase. I was pointing out that no matter how slow a cold build is, hot builds are fast, and are what matters most for iteration. It seems we agree on that.
I too have encountered slow builds in C++. I can't think of a language with a worse tooling story. Certainly good C++ tooling exists, but is not the default, and the ecosystem suffers from decades of that situation. Thankfully modern langs do not.
Yeah, I agree. Much like how the time you spend thinking about the code massively outweighs the time you spend writing the code, the time you spend writing the code massively outweighs the time you spend compiling the code. I think the fascination with compiler performance is focusing on by far the most insignificant part of development.
> even a 1 minute compile time is dwarfed by the time it takes to write and reason about code, run tests, work with version control, etc.
You are far from the embedded world if you think 1 minute here or there is long. I have been involved with many projects that take hours to build, usually caused by hardware generation (fpga hdl builds) or poor cross compiling support (custom/complex toolchain requirements). These days I can keep most of the custom shenanigans in the 1hr ballpark by throwing more compute at a very heavy emulator (to fully emulate the architecture) but that's still pretty painful. One day I'll find a way to use the zig toolchain for cross compiles but it gets thrown off by some of the c macro or custom resource embedding nonsense.
Edit: missed some context on lazy first read so ignore the snark above.
> Edit: missed some context on lazy first read so ignore the snark above.
Yeah, 1 minute was the OP's number, not mine.
> fpga hdl builds
These are another thing entirely from software compilation. Placing and routing is a Hard Problem(TM) which evolutionary algorithms only find OK solutions for in reasonable time. Improvements to the algorithms for such carry broad benefits. Not just because they could be faster, but because being faster allows you to find better solutions.
When you are writing and thinking about code, you are making progress. You are building your mental model and increasing understanding. Compile time is often an excuse to check hacker news while you wait for the artifact to confirm your assumptions. Of course faster compile times are important, even if dwarfed relatively by writing the code. It’s not the same kind of time.
But if those passes (or indeed any other step of the process) are inefficiently implemented, they could be improved. The compiler output would be the same, but now it would be produced more quickly, and the iteration time would be shorter.
> If you want fast compile times then use a compiler w/ no optimization passes, done. The compile times are now linear w/ respect to lines of code & there is provably no further improvement you can make on that b/c any further passes will add linear or superlinear amount of overhead depending on the complexity of the optimization.
Umm this is completely wrong. Compiling involves a lot of stuff, and the language design, as well as compiler design can make or break them. Parsing is relatively easy to make fast and linear, but the other stuff (semantic analysis) is not. Hence why we have a huge range of compile times across programming languages that are (mostly) the same.
People are (reasonably) objecting to Clang's performance even in -O0. Still, it is a great platform for sophisticated optimizations and code analysis tools.
Honestly a bit surprised that he waits 16 seconds every minor change to Ghostty while developing. That sounds painful. Glad Zig build times are getting faster though, and I understand he is still on LLVM.
I work in a large C++ codebase where incremental builds can take minutes. It's flow-interrupting, but doesn't make it impossible to be productive. You just can't rely too heavily on tight printf debugging cycles or you waste a lot of time waiting.
I'm surprised that whatever IO primitive is used on Mac doesn't work on OpenBSD. Mac has select and some variant of kqueue, right? What is ghostty doing there that doesn't work on OpenBSD?
Might not be that different to add OpenBSD. Someone would begin here: https://github.com/mitchellh/libxev/blob/main/src/backend/kq...
It's about 1/3 tests and 2/3 mostly-designed-to-be-portable code. Some existing gaps for FreeBSD, but fixing those (and adding OpenBSD to some switch/enums) should get you most of the way there.
For me, Zig the language is "meh", but Zig the toolchain is very interesting. It's really ambitious, to the point I think it can't work in reality in the long term, but I want it to succeed.
LLVM is a trap. You bootstrap extra fast, you get all sorts of optimization passes and platforms for free, but you lose out on the ability to tune the final optimization passes and performance of the linking stages.
I think we'll see cranelift take off in Rust quite soon, though I also think it wouldn't be the juggernaut of a language if they hadn't stuck with LLVM those early years.
Go seems to have made the choice long ago to eschew outsourcing of codegen and linking and done well for it.
> LLVM is a trap.
Is it? I think Rust is a great showcase for why it isn't. Of course it depends somewhat on your compiler implementation approach, but actual codegen-to-LLVM tends to only be a tiny part of the compiler, and it is not particularly hard to replace it with codegen-to-something-else if you so desire. Which is why there is now codegen_cranelift, codegen_gcc, etc.
The main "vendor lock-in" LLVM has is if you are exposing the tens of thousands of vendor SIMD intrinsics, but I think that's inherent to the problem space.
Of course, whether you're going to find another codegen backend (or are willing to write one yourself) that provides similar capabilities to LLVM is another question...
> You bootstrap extra fast, you get all sorts of optimization passes and platforms for free, but you lose out on the ability to tune the final optimization passes and performance of the linking stages.
You can tune the pass pipeline when using LLVM. If your language is C/C++ "like", the default pipeline is good enough that many such users of LLVM don't bother, but languages that differ more substantially will usually use fully custom pipelines.
> I think we'll see cranelift take off in Rust quite soon, though I also think it wouldn't be the juggernaut of a language if they hadn't stuck with LLVM those early years.
I'd expect that most (compiled) languages do well to start with an LLVM backend. Having a tightly integrated custom backend can certainly be worthwhile (and Go is a great success story in that space), but it's usually not the defining the feature of the language, and there is a great opportunity cost to implementing one.
It is a “trap” in the sense that it limits the design space for your language by forcing some of the choices implicit in C++. Rust went with slow build times like C++ and so was not held back in this dimension by LLVM.
Nothing about LLVM is a trap for C++ as that is what it was designed for.
Both Go and Ocaml have really, really fast compilers. They did the works from the get go, and now wont have to pay the price.
I cant see myself ever again working on a system with compile times that take over a minute or so (prod build not counting).
I wish more projects would have they own "dev" compiler that would not do all the shit llvm does, and only use llvm for the final prod build.
They show a return to how fast compilers used to be, before C and C++ took over the zeitgeist of compilation times.
Eiffel, Common Lisp, Java and .NET (C#, F#, VB) are other examples, where we can enjoy fast development loops.
By combining JIT and AOT workflows, you can get best of both worlds.
I think the main reason this isn't as common is the effort that it takes to keep everything going.
I am on the same camp regarding build times, as I keep looking for my Delphi experience when not using it.
As I can't update my comment, here goes some info out of Turbo Pascal 5.5 marketing brochure,
> Fast! Compiles 34 000 lines of code per minute
This was measured on a IBM PS/2 Model 60.
So lets put this into perspective, Turbo Pascal 5.5 was released in 1989.
IBM PS/2 Model 60 is from 1987, with a 80286 running at 10 MHz, limited by 640 KB, with luck one would expad it up to 1 MB and use HMA, in what concerns using it with MS-DOS.
Now projecting this to 2025, there is no reason that compiled languages, when using a limited set of optimizations like TP 5.5 on their -O0, can't be flying in their compilation times, as seen in good examples like D and Delphi, to use two examples of expressive languages with rich type systems.
Reminds me of this old series of posts on the Turbo Pascal compiler (have been shared a few times on HN in the past):
A Personal History of Compilation Speed (2 parts): https://prog21.dadgum.com/45.html
"Full rebuilds were about as fast as saying the name of each file in the project aloud. And zero link time. Again, this was on an 8MHz 8088."
Things That Turbo Pascal is Smaller Than: https://prog21.dadgum.com/116.html
Old versions of Turbo Pascal running in FreeDOS on the bare metal of a 21st century PC is how fast and responsive I wish all software could be, but never is. Press a key and before you have time to release it the operation you started has already completed.
A problem is how people have started depending on the optimizations. "This tower of abstractions is fine, the optimizer will remove it all" Result is some modern idioms run slow as molasses without optimization, and you can't really use o0 at all.
Indeed. Also,
- Turbo Pascal was compiling at o-1, at best. For example, did it ever in-line function calls?
- its harder to generate halfway decent code for modern CPUs with deep pipelines, caches, and branch predictors, than it was for the CPUs of the time.
Can you give an example? AFAICT monomorphization takes the major portion of time, and it's not even a result of some complicated abstraction.
Our Delphi codebase at work is 1.7 million lines of code, takes about 40 seconds on my not very spicy laptop to do a full release build.
That's with optimizations turned on, including automatic inlining, as well as a lot of generics and such jazz.
Turbo Pascal was an outlier in 1989 though. The funny thing is that I remember Turbo C++ being an outlier in the opposite direction.
In my computer science class (which used Turbo C++), people would try to get there early in order to get one of the two 486 machines, as the compilation times were a huge headache (and this was without STL, which was new at the time).
I recently saw an article about someone improving the machine code generation time of an assembler, here; I idly noticed that the scale was the same number of instructions we had in the budget to compile whole lines of code (expressions & all) "back in the day". It was weird. Of course, we're fighting bandwidth laws, so if you looked at the wall clock time, the machine code generation time was very good in an absolute sense.
> be, before C and C++ took over the zeitgeist of compilation times.
I wouldn’t put them together. C compilation is not the fastest but fast enough to not be a big problem. C++ is a completely different story: not only it orders of magnitude slower (10x slower probably not the limit) on some codebases compiler needs a few Gb RAM (so you have to set -j below the number of CPU cores to avoid OOM).
Back in 1999 - 2003, when I was working on a product mixing Tcl and C, the builds took one hour per platform, across Aix, Solaris, HP-UX, Red-Hat Linux, Windows 2000/NT build servers.
C++ builds can be very slow versus plain old C, yes, assuming people do all mistakes there can be done.
Like overuse of templates, not using binary libraries across modules, not using binary caches for object files (ClearMake style already available back in 2000), not using incremental compilation and incremental linking.
To this day, my toy GTK+ Gtkmm application that I used for a The C/C++ Users Journal article, and have ported to Rust, compiles faster in C++ than Rust in a clean build, exactly because I don't need to start from the world genesis for all dependencies.
That’s not really an apples to apples comparison, is it?
I dunno why, the missing apple on Rust side is not embracing binary libraries like C and C++ do.
Granted there are ways around it for similar capabilities, however they aren't the default, and defaults matter.
I talked a bit about this at the Rust All Hands back in May.
A lot of Rust packages that people ust are setup more like header-only libraries. We're starting to see more large libraries that better fit the model of binary libraries, like Bevy and Gitoxide. I'm laying down a vague direction for something more binary-library like (calling them opaque dependencies) as part of the `build-std` effort (allow custom builds of the standard library) as that is special cased as a binary library today.
You only build the world occasionally, like when cloning or staring a new project, or when upgrading your rustc version. It isn't your development loop.
I do think that dynamic libraries are needed for better plugin support, though.
> You only build the world occasionally
Unless a shared dependency gets updated, RUSTFLAGS changes, a different feature gets activated in a shared dependency, etc.
If Cargo had something like binary packages, it means they would be opaque to the rest of your project, making them less sensitive to change. Its also hard to share builds between projects because of the sensitivity to differences.
One hour? How did you develop anything?
That’s probably why tcl is in there, you use uncompiled scripting to orchestrate the native code, which is the part that takes hours to compile.
I have seen projects where ninja instead of make (both generated by cmake) is able to cleverly invoke the compiler such that CPU is saturated and ram isn't exhausted, and make couldn't (it reached oom).
I have first hand experience of painfully slow C# compile times. Sprinkle in a few extra slow things like EDMX generated files (not C# but part of the MS ecosystem) and it has no business being in a list of fast compiling languages.
Painfully slow? On a modern machine reasonable sized solution will compile almost instantly.
I can do the same with some C code I worked on at enterprise scale.
Lets apply the same rules then.
This.
The C# compiler is brutally slow and the language idioms encourage enormous amounts of boilerplate garbage, which slows builds even further.
As much as I love F#, I wouldn't add it to a list of fast compilers; although I really appreciate that it supports actual file-scoped optimizations and features for compile-time inlining.
I wouldn't even put C# in there - minute builds are like nothing unusual, and my only experience with Java is Android but that was really bad too.
They aren't C++ levels bad, but they are slow enough to be distracting/flow breaking. Something like dart/flutter or even TS and frontend with hot reload is much leaner. Comparing to fully dynamic languages is kind of unfair in that regard.
I did not try Go yet but from what I've read (and also seeing the language design philosophy) I suspect it's faster than C#/Java.
Right, the work I get paid to do is often C# and literally yesterday I was cursing a slow C# build. Why is it slow? Did one of the other engineers screw up something about the build, maybe use a feature that's known to take longer to compile, or whatever? I don't know or care.
This idea that it's all sunshine and lollipops for other languages is wrong.
The fun of tracking down slow autoconf builds with spaghetti Makefiles in C, that take a full hour to do a clean build.
Lets not mix build tools with compilers.
Go is slower than Java (Java's (JIT) optimising compiler and GCs are much more advanced than Go's).
Android is an anti-pattern in Java world, and having a build system based on a relatively slow scripting language hardly helps.
Many things that Google uses to sell Java (the language) over Kotlin also steems from how bad they approach the whole infrastructure.
Try using Java on Eclipse with compilation on save.
It never took more than a few seconds to me.
> Both Go and Ocaml have really, really fast compilers. They did the works from the get go, and now wont have to pay the price.
People tend to forget that LLVM was pretty much that for the C/C++ world. Clang was worlds ahead of GCC when first released (both in speed and in quality of error messages), and Clang was explicitly built from the ground up to take advantage of LLVM.
Yet Ocaml is unusable. Focus on user experience before focusing on speed. Golang did both tbh.
A trap? New language projects would never be able to develop the architecture support and optimizer. There's nothing new in the linking stage you couldn't submit as a patch to LLVM upstream.
Additionally any language that wants to be a C++ replacement, keeps a dependency on C++ as long as LLVM is part of its tooling infrastructure.
For me beyond the initial adoption hump, programming languages should bootstrap themselves, if nothing else, reduces the urban myth that C or C++ have to always be part of the equation.
Not at all, in most cases it is a convenience, writing usable compilers takes time, and it is an easy way to get everything else going, especially when it comes to porting across multiple platforms.
However that doesn't make them magical tools, without which is impossible to write compilers.
That is one area where I am fully on board with Go team's decisions.
I hope cranelift continues to grow too. LLVM tragically has become fork hell. Every cpu vendor now seems to have a fork with enhancements for their cpu arch in some closed source package. It’s hellish as soon as you want to use another language frontend or find the inevitable compiler bug.
That is the magic of MIT like licenses.
How much do you think the BSDs get back from Apple and Sony?
How is it a trap if languages are able to move away?
Not all do, some stick with their decision to live with LLVM, e.g. Swift, Julia, Mojo,...
But isn't there a reason why so many start with LLVM ? It's highly optimised. I think mojo is not LLVM but mlir
It is the modern version of a compiler construction toolkit, which by the way isn't a new idea, even though so many think LLVM was the first one,
One example from 1980s, there are others to pick from as example,
https://en.wikipedia.org/wiki/Amsterdam_Compiler_Kit
So naturally a way to quickly reach the stage where a compiler is avaible for a brand new language, without having to write all compiler stages by hand.
A kind of middleware for writing compilers, if you wish.
MLIR is part of LLVM tooling, is the evolution of LLVM IR.
Mojo uses mlir for high-level ir and most optimizations but codegen is still done with llvm, albeit in unconventional manner.
Huh, are there plans on using cranelift for release builds?
No
If it's a small overhead I might
"While it does not have the aggressive optimization stance (or complexity) of larger compilers such as LLVM or gcc, its output is often not too far off: the same paper showed Cranelift generating code that ran ~2% slower than V8 (TurboFan) and ~14% slower than an LLVM-based system."
-- https://cranelift.dev/
The Cranelift website does have that quote, but the linked paper says
> The resulting code performs on average 14% better than LLVM -O0, 22% slower than LLVM -O1, 25% slower than LLVM -O2, and 24% slower than LLVM -O3
So it is more like 24% slower, not 14%. Perhaps a typo (24/14), or they got the direction mixed up (it is +14 vs -24), or I'm reading that wrong?
Regardless, those numbers are on a particular set of database benchmarks (TPC-H), and I wouldn't read too much into them.
Even 14% would be unacceptably slow for a system language.
I don’t think that means it’s not doable, though.
The artifact's executions speed doesn't seem to be Cranelift's priority. They're instead focusing on compilation speed and security. Those are still useful in Rust for debug builds at least. That's when we need a quick turnaround time and as much verification as possible.
Interesting. Thanks for those numbers. I'd be interested in trying some real-world applications myself.
Besides the potential typo noted by another comment, I'd also note that the benchmarks are on a WebAssembly compiler, which has already gone through the LLVM wringer and thus undergone many high-level optimizations, including those reliant on alias analysis (which Cranelift is quite conservative on, iirc). I'd also imagine that the translation from LLVM IR to WASM discards some useful information that the LLVM backend could otherwise have used to generate better assembly.
That's not to say Cranelift isn't a fantastic piece of tech, but I wouldn't take the "14%" or "24%" number at face value.
GCC is the way.
Trying to understand the Zig goals a bit better..
If you care about compilation speed b/c it's slowing down development - wouldn't it make sense to work on an interpreter? Maybe I'm naiive, but it seems the simpler option
Compiling for executable-speed seems inherently orthogonal to compilation time
With an interpreter you have the potential of lots of extra development tools as you can instrument the code easily and you control the runtime.
Sure, in some corner cases people need to only be debugging their full-optimization-RELEASE binary and for them working on a interpreter, or even a DEBUG build just doesn't makes sense. But that's a tiny minority. Even there, you're usually optimizing a hot loop that's going to be compiling instantly anyway
"Trying to understand the Zig goals a bit better.. If you care about compilation speed b/c it's slowing down development - wouldn't it make sense to work on an interpreter?"
I've heard Rust classified as Safety, Performance, Usability in that order and Zig as Performance, Usability, Safety. From that perspective the build speed-ups make sense while an interpreter would only fit a triple where usability comes first.
I don't think it's right to say Zig is more performance-focused than Rust. Based on their designs, they should both be able to generate similar machine code. Indeed, benchmarks show that Rust and Zig perform very similarly: https://programming-language-benchmarks.vercel.app/rust-vs-z...
This is why the value proposition of Zig is a question for people, because if you got measurably better performance out of Zig in exchange for worse safety that would be one thing. But Rust is just as performant, so with Zig you're exchanging safety for expressivity, which is a much less appealing tradeoff; the learning from C is that expressivity without safety encourages programmers to create a minefield of exploitable bugs.
I don't think Zig is about expressivity over safety. Many people liken Zig to C, like with Rust to C++.
Zig is a lot safer than C by nature. It does not aim to be as safe as Rust by default. But the language is a lot simpler. It also has excellent cross-compiling, much faster compilation (which will improve drastically as incremental compilation is finished), and a better experience interfacing with/using C code.
So there are situations where one may prefer Zig, but if memory-safe performance is the most important thing, Rust is the better tool since that is what it was designed for.
The value proposition in Zig is that it has far and away the best C interoperability I've seen out of a language that wasn't also a near-superset of it. This makes migrations away from C towards Zig a lot more palatable than towards Rust unless security is your primary concern.
With Zig, you're not so much migrating away from C as migrating to a better syntax for C.
On the contrary I would not call Zig very C-like at all, and that's what makes the interoperability so impressive to me.
Zig feels like a small yet modern language that lacks many of C's footguns. It has far fewer implicit type conversions. There is no preprocessor. Comptime allows for type and reflection chicanery that even C++ would be jealous of.
Yet I can just include a C header file and 24 times out of 25, it just works. Same thing if you run translate-c on a c file, and as a bonus, it reveals the hoops that the tooling has to go through to preserve the fiddly edge cases while still remaining somewhat readable.
It's much easier to write ultra low latency code in Zig because doing so requires using local bump-a-pointer allocators, and the whole Zig standard library is built around that (everything that allocates must take the allocator as a param). Rust on the other hand doesn't make custom allocators easy to use, partially because proving they're safe to the borrow checker is tricky.
Excellent way to put it.
> expressivity without safety encourages programmers to create a minefield of exploitable bugs.
Also, psychologically, the combination creates a powerful and dangerously false sensation of mastery in practitioners, especially less experienced ones. Part of Zig's allure, like C's bizarre continued relevance, is the extreme-sport nature of the user experience. It's exhilarating to do something dangerous and get away with it.
Please — rust advocates need to realize that you're not winning anyone to rust with this kind of outlandish comments.
Zig is not an extreme sports experience. It has much better memory guarantees than C. It's clearly not as good as rust if that's your main concern but rust people need to also think long and hard about why rust has trouble competing with languages like Go, Zig and C++ nowadays.
I'm not a Rust advocate. I'm a "the whole debate is silly: 99% of the time you want a high level language with a garbage collector" advocate.
What's the use case for Zig? You're in that 1% of projects in which you need something beyond what a garbage collector can deliver and, what, you're in the 1% of the 1% in which Rust's language design hurts the vibe or something?
You can also get that "safer, but not Safe" feeling from modern C++. So what? It doesn't really matter whether a language is "safer". Either a language lets you prove the absence of certain classes of bug or it doesn't. It's not like C, Zig, and Rust form a continuum of safety. Two of these are different from the third in a qualitative sense.
> What's the use case for Zig? You're in that 1% of projects in which you need something beyond what a garbage collector can deliver and, what, you're in the 1% of the 1% in which Rust's language design hurts the vibe or something?
You want a modern language that has package management, bounds checks and pointer checks out of the box. Zig is a language you can pick up quickly whereas rust takes years to master. It's a good replacement to C++ if you're building a game engine for example.
> Either a language lets you prove the absence of certain classes of bug or it doesn't. It's not like C, Zig, and Rust form a continuum of safety. Two of these are different from the third in a qualitative sense.
Again repeating my critique from the previous comment – yes Zig brings in additional safety compared to C. Dismissing all of that out of hand does not convince anyone to use rust.
> wouldn't it make sense to work on an interpreter
I like the Julia approach. Full on interactivity, the “interpreter” just compiles and runs your code.
SBCL, my favorite Common Lisp, works the same way I think
> Compiling for executable-speed seems inherently orthogonal to compilation time
That true at the limit. As is often the case there's a vast space in the middle, between the extremes of ultimate executable speed and ultimate compiler speed, where you can make optimizations that don't trade off against each other. Where you can make the compiler faster while producing the exact same bytecode.
Limits are more important than implied. Language benchmarks, computer game performance.
I wouldn't say games are a corner case.
For games you have to compile release builds almost all the time anyway, otherwise performance wouldn’t be playable.
This is why games ship interpreters and include a lot of code scripting the gameplay (e.g in Lua) rather than having it all written in C++. You get fast release builds with a decent framerate and the ability to iterate on a lot of the “business logic” from within the game itself.
That's only true for badly written game code though ;)
Even with C++ and heavy stdlib usage it's possible to have debug builds that are only around 3..5 times slower than release builds in C++. And you need that wiggle room anyway to account for lower end CPUs which your game should better be able to run on.
Just curious - is this from personal experience?
I've never done it, but I just find it hard to believe the slow down would be that large. Most of the computation is on GPU, and you can set your build up such that you link to libraries built at different compilation optimizations.. and they're likely the ones doing most of the heavy lifting. You're not rebuilding all of the underlying libs b/c you're not typically debugging them.
EDIT:
if you're targeting a console.. why would you not debug using higher end hardware? If anything it's an argument in favor of running on an interpreter with a very high end computer for the majority of development..
Yeah, around 3x slower is what I've seen when I was still working on reasonably big (around a million lines of 'orthodox C++' code) PC games until around 2020 with MSVC and without stdlib usage.
This was building the entire game code with the same build options though, we didn't bother to build parts of the code in release mode and parts in debug mode, since debug performance was fast enough for the game to still run in realtime. We also didn't use a big integrated engine like UE, only some specialized middleware libraries that were integrated into the build as source code.
We did spend quite some effort to keep both build time and debug performance under control. The few cases were debug performance became unacceptable were usually caused by 'accidentially exponential' problems.
> Most of the computation is on GPU
Not in our case, the GPU was only used for rendering. All game logic was running on the CPU. IIRC the biggest chunk was pathfinding and visibility/collision/hit checks (e.g. all sorts of NxM situations in the lower levels of the mob AI).
Ah okay, then that makes sense. It really depends on your build system. My natural inclination would be to not have a pathfinding system in DEBUG unless I was actively working on it. But it's not always very easy to set things up to be that way
The slowdown can be enormous if you use SIMD, I believe MSVC emits a write to memory after every single SIMD op in debug, thus the code is incredibly slow(more than 10x). If you have any SIMD kernels you will suffer, and likely switch to release, with manual markings used to disable optimizations per function/file.
If you target consoles, you’re already on the lowest end hardware you’ll run on. It’s extremely rare to have much (if any) headroom for graphically competitive games. 20% is a stretch, 3x is unthinkable.
Then I would do most of the development work that doesn't directly touch console APIs on a PC version of the game and do development and debugging on a high-end PC. Some compilers also have debug options which apply optimizations while the code still remains debuggable (e.g. on MSVC: https://learn.microsoft.com/en-us/visualstudio/debugger/how-...) - don't know about Sony's toolchain though. Another option might be to only optimize the most performance sensitive parts while keeping 90% of the code in debug mode. In any case, I would try everything before having to give up debuggable builds.
It doesn’t matter. You will hit (for example) rendering bugs that only happen on a specific version of a console with sufficient frequency that you’ll rarely use debug builds.
Debug builds are used sometimes, just not most of the time.
That only kind of works when not using console specific hardware.
Naturally you can abstract them as middleware does it, however it is where tracking performance bugs usually lies.
And if using Visual Studio, it is even possible to debug release builds, with dynamic debugging feature.
The gold standard for compiler speed so far is TCC. (Fabrice Bellard)
It is not even super optimized (single thread, no fancy tricks) but it is so far unbeaten by a large margin. Of course I use Clang for releases, but the codegen of tcc is not even awful.
I would rather vouch for Delphi, as a much more expressive, and safer language, while having blazing fast compile times.
True. Incremental compilation in Delphi is quite something else.
The Go compiler is also pretty fast.
I'd call it not terribly slow. IIRC it didn't hit 100k lines on an M2, while I saw over two million on my own compiler (I used multiple instances of tcc since I didn't want to deal with implementing debug info on mac)
Maybe typescript will catch up
It's being rewritten in Go so yes it will
The compiler might be rewritten in Go, but the Typescript and Go languages themselves are very different.
Go’s main objectives were fast builds and a simple language.
Typescript is tacked onto another language that doesn’t really care about TS with three decades of warts, cruft and hacks thrown together.
It's hard to say.
One the one hand the go type system is a joke compared to typescript so the typescript compiler has a much harder job of type checking. On the other hand once type checking is done typescript just needs to strip the types and it's done while go needs to optimize and generate assembly.
Yes. I once was thinking I should write a x86-64 codegen for my own compiler, but using tcc I was able to reach 3 million lines in a second (using multi-threading and multiple instances of tcc) so I called it a day. I never did check out if the linking was in parallel
I thought DMD which compiles both C and D was the gold standard!?
> The gold standard for compiler speed so far is TCC.
I only wish it supported C23...
Any big projects that doesn't compile without C23 support?
Not yet AFAIK, but unlike the 'inbetween' versions since C99, C23 actually has a number of really useful features, so I expect that a couple of projects will start using C23 features. The main problem is as always MSVC, but then maybe its time to ditch MSVC support since Microsoft also seems to have abandondend it (not just the C compiler, but also the C++ compiler isn't seeing lots of updates since everybody working on Visual Studio seems to have been reassigned to work in the AI salt mines).
And Rust, as the official system programming language for Azure.
However there are several C++23 goodies on latest VC++, finally.
Also lets not forget Apple and Google no longer are that invested into clang, rather LLVM.
It is up for others to bring clang up to date regarding ISO.
Clang's support for C23 looks actually pretty good:
https://en.cppreference.com/w/c/compiler_support/23.html
(not sure how much Apple/Google even cared about the C frontend before though, but at least keeping the C frontend and stdlib uptodate by far doesn't require as much effort as C++).
Apple used to care before Swift, because Objective-C unlike C++, is a full superset from C.
Most of the work going into LLVM ecosystem is directly into LLVM tooling itself, clang was started by Apple, and Google picked up on it.
Nowadays they aren't as interested, given Swift, C++ on Apple platforms is mostly for MSL (C++14 baseline) and driver frameworks (also a C++ subset), Google went their own way after the ABI drama, and they care about what fits into their C++ style guide.
I know Intel is one of the companies that picked up some of the work, yet other compiler vendors that replaced their proprietary forks with clang don't seem that eager to contribute upstream, other than LLVM backends for their platforms.
Such is the wonders of Apache 2.0 license.
Ladybird.
Isn't that C++?
There is no singular "gold standard".
vlang is really fast, recompiling itself entirely within a couple of seconds.
And Go's compiler is pretty fast for what it does too.
No one platform has a unique monopoly on build efficiency.
And also there are build caching tools and techniques that obviate the need for recompliation altogether.
> vlang is really fast, recompiling itself entirely within a couple of seconds.
Does V still just output C and use TCC under the hood?
Yes. [0]
https://github.com/vlang/v/blob/master/Makefile#L18
So in theory, I can make Nim do the same?
I use tcc with Nim all the time and it yields almost interpreted REPL-like code-iteration speed (like 100-500ms per compile-link - almost ENTER-see-results).
As a warning, you need to be sure --lineDir=off in your nim.cfg or config.nims (to side-step the infinite loop mentioned in https://github.com/nim-lang/Nim/pull/23488). You may also need to --tlsEmulation=on if you do threads and you may want to --passL=-lm.
tcc itself is quite fast. Faster for me than a Perl5 interpreter start-up for me (with both on trivial files). E.g., on an i7-1370P:
{ EDIT: /n -> /dev/null and true.c is just "int main(int ac,char*av){return 0;}". }Ah nice, I’ve been meaning to try TCC with Nim.
We are talking about compilation speed, not the harness that you could use around it.
I never tried vlang but I would say this is pretty niche, while C is a standard.
AFAIK tcc is unbeaten, and if we want to be serious about compilation speed comparison, I’d say tcc is a good baseline.
Vlang uses tcc though...
Build caching is nice until you have to disable it because it broke something somewhere. At which point it becomes very hard to turn it back on.
In other words, one needs to have absolute trust in such tools to be able to rely on them.
I've never had an issue with ccache and I've been using it forever. Are you using it over a networked filesystem or something like that for it to become confused about what needs recompiling?
V Lang is under appreciated. If you want to really see its simplicity and expressiveness, one look at its ORM will be enough. It is a thing of beauty:
https://modules.vlang.io/orm.html
There they go, making the same stupid mistakes as most other ORMs out there; by 1:1-mapping structs to tables as well as inventing their own SQL-dialect.
Are their ghostty builds statically linked by default or does zig follow in the footsteps of the C/C++ world and dynamically link everything? If statically linking (like rust) then a huge part of the incremental build times for apps with external dependencies is the lack of incremental linking in any of the major linkers. Wild linker is working on becoming the answer to this: https://github.com/davidlattimore/wild
ghostty statically links all its zig dependencies. The only dynamic libraries I can see are the usual suspects (libc, X11, Wayland, OpenGL, GTK, glib, libadwaita, etc).
I've really appreciated the incremental builds with zig - I really like having a single static binary, and I have an application which uses a few different libraries like SQLite and luau and it compiles at nearly Go speeds.
That said, there are definitely still bugs to their self hosted compiler. For example, for SQLite I have to use llvm - https://github.com/vrischmann/zig-sqlite/issues/195 - which kinda sucks.
Great.
It's finally as snappy as recompiling the "Borland Pascal version of Turbo Vision for DOS" was on an Intel 486 in 1995, when I graduated from high school...
They C version of Turbo Vision was 5-10x slower to compile at that time too.
Turbo Vision is a TUI windowing framework, which was used for developing the Borland Pascal and C++ IDEs. Kinda like a character mode JetBrains IDE in 10 MB instead of 1000 MB...
https://en.m.wikipedia.org/wiki/Turbo_Vision
Is there a straightforward path to building Zig with polyglot build systems like Bazel and Buck2? I'm worried Zig's reliance on Turing complete build scripts will make building (and caching) such code difficult in those deterministic systems. In Rust, libraries that eschew build.rs are far preferable for this reason. Do Zig libraries typically have a lot of custom build setup?
For bazel:
https://github.com/aherrmann/rules_zig
Real world projects like ZML uses it:
https://github.com/zml/zml
FYI, build scripts are completely optional. Zig can build and run individual source code files regardless of build scripts (`build.zig`). You may need to decipher the build script to extract flags, but that's pretty much it. You can integrate Zig into any workflow that accepts GCC and Clang. (Note: `zig` is also a drop-in replacement C compiler[1])
[1]: https://andrewkelley.me/post/zig-cc-powerful-drop-in-replace...
Tangential: I've been writing a Java library that is pretty much a JNI shim to a C library. Looking at building a dynamic library for each of the platform / architecture combo is looking like a huge pain in the arse; I'm toying with the idea of using the zig compiler to do the compile / cross compiles and simply be done with it.
do yourself a favor and try it, Zig ships with headers and libc sources and its own LLVM distribution and makes cross compilation so easy it's just ridiculous, you literally don't have to do anything
If it is a plain shim, and you can use modern Java, I would rather make use of Panama and jextract.
It's a bold strategy - let's see if it works out for them.
But from what I remember, this still uses the LLVM backend, right? Sure, you can beat LLVM on compilation speed and number of platforms supported, but when it comes to emitting great assembly, it is almost unbeatable.
It only uses the LLVM backend for release builds. For debug builds Zig now defaults to self-hosted (on supported platforms, only x86_64 right now). Debug builds are directly in the hot path of testing your software (this includes building test binaries), and the number of iterations you do on a debug build are multiple orders of magnitude higher than release builds, so this is a good tradeoff.
The post says the llvm is used in all builds they measured since the self hosted backend still can’t compile ghostty.
Incorrect, the last build is substantially faster and is using the "native" debug backend.
I don't really get the obsession w/ compiler performance. At the end of the day you are trading performance optimizations (inlining, dead code elimination, partial evaluation, etc) for compile times. If you want fast compile times then use a compiler w/ no optimization passes, done. The compile times are now linear w/ respect to lines of code & there is provably no further improvement you can make on that b/c any further passes will add linear or superlinear amount of overhead depending on the complexity of the optimization.
The reason to care about compile time is because it affects your iteration speed. You can iterate much faster on a program that takes 1 second to compile vs 1 minute.
Time complexity may be O(lines), but a compiler can be faster or slower based on how long it takes. And for incremental updates, compilers can do significantly better than O(lines).
In debug mode, zig uses llvm with no optimization passes. On linux x86_64, it uses its own native backend. This backend can be significantly faster to compile (2x or more) than llvm.
Zig's own native backend is designed for incremental compilation. This means, after the initial build, there will be very little work that has to be done for the next emit. It needs to rebuild the affected function, potentially rebuild other functions which depend on it, and then directly update the one part of the output binary that changed. This will be significantly faster than O(n) for edits.
> The reason to care about compile time is because it affects your iteration speed. You can iterate much faster on a program that takes 1 second to compile vs 1 minute.
Color me skeptical. I've only got 30 years of development under the belt, but even a 1 minute compile time is dwarfed by the time it takes to write and reason about code, run tests, work with version control, etc.
Further, using Rust as an example, even a project which takes 5 minutes to build cold only takes a second or two on a hot build thanks to caching of already-built artifacts.
Which leaves any compile time improvements to the very first time the project is cloned and built.
Consequently, faster compile times would not alter my development practices, nor allow me to iterate any faster.
I've worked with Delphi where a recompile takes a few seconds, and I've worked with C++ where a similar recompile takes a long time, often 10 minutes or more.
I found I work very differently in the two cases. In Delphi I use the compiler as a spell checker. With the C++ code I spent much more time looking over the code before compiling.
Sometimes though you're forced to iterate over small changes. Might be some bug hunting where you add some debug code that allows you to narrow things a bit more, add some more code and so on. Or it might be some UI thing where you need to check to see how it looks in practice. In those cases the fast iteration really helps. I found those cases painful in C++.
For important code, where the details matter, then yeah, you're not going to iterate as fast. And sometimes forcing a slower pace might be beneficial, I found.
> Consequently, faster compile times would not alter my development practices, nor allow me to iterate any faster.
I think the web frontend space is a really good case for fast compile times. It's gotten to the point that you can make a change, save a file, the code recompiles and is sent to the browser and hot-reloaded (no page refresh) and your changes just show up.
The difference between this experience and my last time working with Ember, where we had long compile times and full page reloads, was incredibly stark.
As you mentioned, the hot build with caching definitely does a lot of heavy lifting here, but in some environments, such as a CI server, having minutes long builds can get annoying as well.
> Consequently, faster compile times would not alter my development practices, nor allow me to iterate any faster.
Maybe, maybe not, but there's no denying that faster feels nicer.
> Maybe, maybe not, but there's no denying that faster feels nicer.
Given finite developer time, spending it on improved optimization and code generation would have a much larger effect on my development. Even if builds took twice as long.
Can't agree. Iteration speed is magical when really fast.
I'm much more productive when I can see the results within 1 or 2 seconds.
> I'm much more productive when I can see the results within 1 or 2 seconds.
That's my experience today with all my Rust projects. Even though people decry the language for long compile times. As I said, hot builds, which is every build while I'm hacking, are exactly that fast already. Even on the large projects. Even on my 4 year old laptop.
On a hot build, build time is dominated by linking, not compilation. And even halving a 1s hot build will not result in any noticeable change for me.
Same here, Rust currently isn't the best tooling for UI or game development.
> Further, using Rust as an example, even a project which takes 5 minutes to build cold only takes a second or two on a hot build thanks to caching of already-built artifacts.
So optimizing compile times isn’t worthwhile because we already do things to optimize compile times? Interesting take.
What about projects for which hot builds take significantly longer than a few seconds? That’s what I assumed everyone was already talking about. It’s certainly the kind of case that I most care about when it comes to iteration speed.
> So optimizing compile times isn’t worthwhile because we already do things to optimize compile times?
That seems strange to you? If build times constituted a significant portion of my development time I might think differently. They don't. Seems the compiler developers have done an excellent job. No complaints. The pareto principle and law of diminishing returns apply.
> What about projects for which hot builds take significantly longer than a few seconds?
A hot build of Servo, one of the larger Rust projects I can think of off the top of my head, takes just a couple seconds, mostly linking. You're thinking of something larger? Which can't be broken up into smaller compilation units? That'd be an unusual project. I can think of lots of things which are probably more important than optimizing for rare projects. Can't you?
The part that seems strange to me is your evidence that multi-minute compile times are acceptable being couple-second compile times. It seems like everyone actually agrees that couple-second iteration is important.
It seems like you might have missed a few words in this comment, I'm honestly having trouble parsing it to figure out what you're trying to say.
Just for fun, I kicked off a cold build of Bevy, the largest Rust project in my working folder at the moment, which has 830 dependencies, and that took 1m 23s. A second hot build took 0.22s. Since I only have to do the cold build once, right after cloning the repository which takes just as long, that seems pretty great to me.
Are you telling me that you need faster build times than 0.22s on projects with more than 800 dependencies?
This is the context:
> > The reason to care about compile time is because it affects your iteration speed. You can iterate much faster on a program that takes 1 second to compile vs 1 minute.
> Color me skeptical. I've only got 30 years of development under the belt, but even a 1 minute compile time is dwarfed by the time it takes to write and reason about code, run tests, work with version control, etc.
If your counterexample to 1-minute builds being disruptive is a 1-second hot build, I think we’re just talking past each other. Iteration implies hot builds. A 1-minute hot build is disruptive. To answer your earlier question, I don’t experience those in my current Rust projects (where I’m usually iterating on `cargo check` anyway), but I did in C++ projects (even trivial ones that used certain pathological libraries) as well as some particularly badly-written Node ones, and build times are a serious consideration when I’m making tech decisions. (The original context seemed language-agnostic to me.)
I see. Perhaps I wasn't clear. I've never encountered a 1 minute hot build in Rust, and given my experience with large Rust codebases like Bevy I'm not even sure such a thing exists in a real Rust codebase. I was pointing out that no matter how slow a cold build is, hot builds are fast, and are what matters most for iteration. It seems we agree on that.
I too have encountered slow builds in C++. I can't think of a language with a worse tooling story. Certainly good C++ tooling exists, but is not the default, and the ecosystem suffers from decades of that situation. Thankfully modern langs do not.
Yeah, I agree. Much like how the time you spend thinking about the code massively outweighs the time you spend writing the code, the time you spend writing the code massively outweighs the time you spend compiling the code. I think the fascination with compiler performance is focusing on by far the most insignificant part of development.
This is underestimating that running the code is part of the development process.
With fast compile time, running the test suite (which implies to recompile it) is fast too.
Also if the language itself is optimized towards making easy to write a fast compiler, this also makes your IDE fast.
And just if you're wondering, yes, Go is my dope.
> even a 1 minute compile time is dwarfed by the time it takes to write and reason about code, run tests, work with version control, etc.
You are far from the embedded world if you think 1 minute here or there is long. I have been involved with many projects that take hours to build, usually caused by hardware generation (fpga hdl builds) or poor cross compiling support (custom/complex toolchain requirements). These days I can keep most of the custom shenanigans in the 1hr ballpark by throwing more compute at a very heavy emulator (to fully emulate the architecture) but that's still pretty painful. One day I'll find a way to use the zig toolchain for cross compiles but it gets thrown off by some of the c macro or custom resource embedding nonsense.
Edit: missed some context on lazy first read so ignore the snark above.
> Edit: missed some context on lazy first read so ignore the snark above.
Yeah, 1 minute was the OP's number, not mine.
> fpga hdl builds
These are another thing entirely from software compilation. Placing and routing is a Hard Problem(TM) which evolutionary algorithms only find OK solutions for in reasonable time. Improvements to the algorithms for such carry broad benefits. Not just because they could be faster, but because being faster allows you to find better solutions.
Any delay between writing code and seeing the result is empty space in my mind where distractions can enter.
When you are writing and thinking about code, you are making progress. You are building your mental model and increasing understanding. Compile time is often an excuse to check hacker news while you wait for the artifact to confirm your assumptions. Of course faster compile times are important, even if dwarfed relatively by writing the code. It’s not the same kind of time.
But if those passes (or indeed any other step of the process) are inefficiently implemented, they could be improved. The compiler output would be the same, but now it would be produced more quickly, and the iteration time would be shorter.
> If you want fast compile times then use a compiler w/ no optimization passes, done. The compile times are now linear w/ respect to lines of code & there is provably no further improvement you can make on that b/c any further passes will add linear or superlinear amount of overhead depending on the complexity of the optimization.
Umm this is completely wrong. Compiling involves a lot of stuff, and the language design, as well as compiler design can make or break them. Parsing is relatively easy to make fast and linear, but the other stuff (semantic analysis) is not. Hence why we have a huge range of compile times across programming languages that are (mostly) the same.
clang -O0 is significantly slower than tcc. So the problem is clearly not (just) the optimisation passes.
Faster compiler means developers can go faster.
People are (reasonably) objecting to Clang's performance even in -O0. Still, it is a great platform for sophisticated optimizations and code analysis tools.
I think it highly depends on what you're doing.
For some work I tend to take a pen and paper and think about a solution, before I write code. For these problems compile time isn't an issue.
For UI work on the other hand, it's invaluable to have fast iteration cycles to nail the design, because it's such an artistic and creative activity
> but when it comes to emitting great assembly, it is almost unbeatable.
"Proebsting's Law: Compiler Advances Double Computing Power Every 18 Years"
The implication is that doing the easy, obvious and fast optimizations is Good Enough(tm).
Even if LLVM is God Tier(tm) at optimizing, the cost for those optimizations swings against them very quickly.
How does Zig's compilation and codegen compare to TPDE? The claim is that it's 10-20x faster than LLVM with -O0 but with limitations.
Anybody knows what exactly prevents Ghostty from being built using the self-hosted x86_64 backend?
Bugs. The Zig compiler crashes. It’s new and complex, I’m sure they’ll work it out soon.
well, most of the users are aarch64
Honestly a bit surprised that he waits 16 seconds every minor change to Ghostty while developing. That sounds painful. Glad Zig build times are getting faster though, and I understand he is still on LLVM.
I work in a large C++ codebase where incremental builds can take minutes. It's flow-interrupting, but doesn't make it impossible to be productive. You just can't rely too heavily on tight printf debugging cycles or you waste a lot of time waiting.
my only wish for ghostty is that there was openbsd support. some work needs to be done in the io layer to support openbsd kqueue to make it work.
I'm surprised that whatever IO primitive is used on Mac doesn't work on OpenBSD. Mac has select and some variant of kqueue, right? What is ghostty doing there that doesn't work on OpenBSD?
E.g., here it is kqueue-aware on FreeBSD: https://github.com/mitchellh/libxev/blob/34fa50878aec6e5fa8f...
Might not be that different to add OpenBSD. Someone would begin here: https://github.com/mitchellh/libxev/blob/main/src/backend/kq... It's about 1/3 tests and 2/3 mostly-designed-to-be-portable code. Some existing gaps for FreeBSD, but fixing those (and adding OpenBSD to some switch/enums) should get you most of the way there.
For me, Zig the language is "meh", but Zig the toolchain is very interesting. It's really ambitious, to the point I think it can't work in reality in the long term, but I want it to succeed.
Wow!