Just want to say among many portable SIMD solutions I’ve seen recently (e.g. Fearless SIMD), this is the first that makes non-fixed vectors like SVE and RISC-V vector (RVV) easier to support. Glad to see they made this decision
How so? I imagine you'd still want to constrain the length to the maximum vector size supported by the lowest platform you want to support or you lose the portability and actually end up with code that performs much worse than the scalar alternative on some platforms.
Mojo has an even more portable simd[1] type that isn't just generic over length but also over type. In my opinion it is almost always better to specialize for each platform and use portable implementation as fallback. It's a shame that just very few languages support Zig-like comptime, because it would be excellent for specializations without introducing runtime penalties.
> constrain the length to the maximum vector size supported by the lowest platform you want to support or you lose the portability and actually end up with code that performs much worse than the scalar alternative on some platforms.
Or, put a dynamic factor into your vector size and design everything around it. Such that every platforms can plug in their own factor and _scale_ the size of vectors. This is basically what LLVM IR does for SVE and RVV: `<vscale x 4 x i32>` where vscale is the said dynamic factor. Though the exact value of vscale is only known during runtime, it doesn't matter -- we still can design compiler optimizations and lowering around it. The generated binaries can then be portable across platforms with different vscale values.
C++ is getting std::simd in the latest version and I am all aboard writing the vectorization with the least amount of intrinsic builtins I am able to. Even if not optimal, it's far better than the scalar ops.
Seconded!! This doesn’t really help the well established codebases much that are already doing this on a platform specific path but in general this is much appreciated for the future.
This feature opens many doors for optimizing low-level performance in Go projects, that are already running multicore. IIRC there aren’t a lot of languages with built-in std lib support for SIMD and variants. Love the way Go is trying new stuff lately.
Besides the usual C and C++, we have Java, .NET, D, Zig, Julia, Swift, Rust.
So yeah, also appreciate having Go in the group instead of manually having to write Assembly.
However not many languages adopt ways to manually write SIMD, because most of us have no idea how to write good SIMD code in first place, I surely don't.
Even with languages that adopt ways to manually write SIMD, it’s mostly left to library maintainers rather than application developers.
I work for a C++ timeseries database startup that leverages SIMD about as much as we possibly can, and except for some extremely rare places we just use libraries.
Yeah, that is what I have heard from some NVidia folks as well, like Bryce Adelstein, use the libraries as much as possible, and leave the kernels for experts.
However even then, it depends on how the libraries API surface looks like.
But it’s not necessary at all, the whole point is that these utility libraries bring you more elegant code that work on all platforms without having to pollute your codebase with SIMD intrinsics.
Unless this was tongue in cheek, because this is in fact a problem with AI that it degrades your codebase in these types of ways.
I did some testing with the experimental SIMD on a project I was doing to make speech-to-text and text-to-speech models run natively in Go (with CGO_ENABLED=0, so no C depenencies), and testing non-SIMD w/ SIMD.
I don't have formal benchmarks for that, but I can anecdotally say the SIMD work made a measurable improvement in the performance of the calculations vs. just plain Go. I'm very optimistic about how these improvements will help make the Go runtime an even better target for more of these types of work going forward, especially since it is cross-platform.
This is why I love Go. Nobody was asking for this, but they took the time to do it right and continue to Push go as a memory safe, high-level systems language.
It's been discussed for a long time, and the related proposals were heavily upvoted, including various older proposals.
As I understand it, part of the reason it took a while is that the core Go team was generally of the opinion that doing user-facing SIMD APIs the right way was to design a high-level, cross-platform API that would stand the test of time, and that was then punted a few times given its complexity and need to do other things.
Part of what helped the current approach take off was switching to a philosophy of designing a lower-level architecture-dependent API first (the 'simd/archsimd' package), and then later doing a higher-level portable API (the 'simd' package, which is topic of this blog post).
That two-level approach I think also gave some additional freedom for the design and implementation of the friendlier / high-level 'simd' package, including because the lower-level 'simd/archsimd' package is available for people who need or want to drop down.
This is kind of the opposite of Go. Not giving people what they are asking for.
There are pros and cons of course. You don't have 17 different ways to iterate over an array, so that's nice. But you also went 13 years without generics, despite them being one of the most requested features, because the designers didn't want that complexity inside Go.
Overall I think Go is better for this philosophy but there are times where the language is clearly written more for its maintainers than it's users.
Go shipped with generics (aka bounded parametric polymorphism), but only for built-in types: slices, arrays, and maps. That, with subtyping via interfaces, handled most demand for generics. The most common pain point was custom containers.
Go was first released in November 2009. Russ Cox posted "The Generic Dilemma" [1] in December 2009. The comments show the generics debate raging from the earliest days.
As a fun side note, I forgot I posted a comment on that post pointing to Ada's generics. I was in college, and Ada was our intro language.
> the designers didn't want that complexity inside Go.
Yes, with some nuance. Go's goal of writing server programs didn't require the type-system complexity and run-time hit of user-defined generics. [2]
> Go was intended as a language for writing server programs [...] Polymorphic programming did not seem essential [...] so was initially left out for simplicity.
>
> Generics are convenient but they come at a cost in complexity in the type system and run-time. It took a while to develop a design that we believe gives value proportionate to the complexity.
Some of the concerns around generics and why it took so long were for the users as well. One of the biggest draws to Go has always been that you get the performance of a compiled language and yet compile times are so low that it can feel like you're developing with an interpreted language. The design of generics needed to maintain the compile time advantage or else it wouldn't feel like Go any more.
That's not what "memory safe" means. "Memory safe" is a term of art meaning "not susceptible to memory corruption exploits", like stack and heap overflows, UAFs, and type confusion. Last I checked, there are essentially no non-contrived memory corruption exploits for Go programs; the best you get are people demonstrating register control on contrived programs.
The definition I'm giving is the same as the ISRG's definition at MemorySafety.org. It's the thing everybody is talking about when they talk about memory safety.
The claim being made here is "big if true", because it would imply a lot more languages than Go "aren't memory safe", despite decades without memory corruption exploits.
The exploits aren't the only issue; they just get a lot of air time.
It's remarkably easy to segfault Go applications with data races. Any object with multiple words (so a slice that's an array pointer, a size, and a capacity. Or a fat pointer with the object pointer and the vtable pointer), can be read in an inconsistent state from two threads which can cause out of bounds reads and writes. And this comes up all the time with how heavily the language encourages concurrency.
It's just difficult to actually exploit because of other considerations that practically add a lot of runtime entropy.
They aren't the only issue in programming language theory, but they are the only issue in the ordinary context in which we discuss "memory safety", such that if someone not in a PLT forum says "is Go memory-safe" and you say "no" you will look a little batty.
The was for a time a vogue for "zero trust networking" and I'm fond of pointing out that the same thing happened there: people would come up with their own axiomatic derivation of what "zero trust" meant, but in reality it was a term of art meaning "non-Google implementations of BeyondCorp".
Terms of art are kryptonite for message board nerds.
No, it has real world implications. I worked at a heavy go shop; every time we'd have a new batch of hiring, the segfault bugs would start rolling in from these memory unsafety issues. The data race detector was good, but not perfect. Yes, those new devs would look at you batty until they saw the bug reports roll in.
And yeah, we tend to use the PLT definitions when we're talking about literal semantics of programming languages. Nothing in this thread mandates a security focused sub-definition.
And even Rob Pike described Go as "not purely memory safe", in a slide that obviously references this exact data race behavior. https://go.dev/talks/2012/splash.slide#49 They considered this a practical tradeoff for simplicity versus the major managed languages defining what happens during data races in a way that doesn't allow you to break memory safety.
Again: I am not disputing that there are correctness issues that come from having Go's concurrency and memory model, just like there are in (the large majority of) other languages with similar models. But those issues are not security problems, and security is what we're talking about when we talk about memory safety.
Java arrays are thin pointers to Array objects, which contain the length and data in the same place (on the far side of the pointer). Since thin pointers cannot tear and Array objects cannot be resized, Arrays themselves are always memory-safe in safe code.
Go slices are fat pointers to undecorated memory. The slice itself is a 3-tuple of pointer, length, and capacity. If you append to a slice that's already at capacity, the Go runtime will allocate new memory for you and return a new 3-tuple. If you assign that result to a variable that's also being accessed by another goroutine, the latter can observe the slice in an inconsistent state. It can, for example, see the old pointer but with the new length, allowing out-of-bounds access. None of this requires unsafe code.
The same issue applies to string and interface variables, which are also fat pointers.
Shared Go slices are a bad mix in concurrent code. This is a given. But its also not a fair comparison, you should instead compare java arrays to go arrays, not slices.
This goes for slices, strings and maps. Those a usually wrapped in a mutex, or used with sync primitives like sync.Map.
The exact equivalences between the two languages are not the point, and anyway, Go arrays are fixed-size and have no Java equivalent.
The point is that Java does not have this problem in the language or the standard library. Of course, you should not write racy Go code; the language provides ample ways to avoid the race, such as channels; and the race detector will generally find such racy code, provided you turn it on. However, the issue is that this race leads to memory-safety violations; it can occur especially in code written by novice Go programmers, and it's well acknowledged by the language authors [1].
> Those a usually wrapped in a mutex, or used with sync primitives like sync.Map.
The same can be said of C or C++ constructs (and many "anti-Rust" people have historically said that) -- the point is that their use is not enforced by the language and so bugs can lead to panics.
I write a fair amount of Go and Rust so I really don't think either language's flaws are fatal, but it comes off as weirdly defensive to redefine memory and data safety to be "well if you use it properly it's safe". It's totally fine to say this is a problem the Go language did not find important enough to require compile time enforcement and so solving it is done by convention and testing with the race detector (which a similar answer C and C++ give to this problem).
From what I understand a data race on a simple built in feature like an interface pointer can result in a bad address / type pair, which can cause memory safety issues on any future access. I don't think you can get the JVM itself confused about what type a pointer points to.
There is no memory safety without freedom from data races. One is a prerequisite of the other. This is why languages like C# throw exceptions on unsynchronized concurrent access to some container types, and treat all property accesses as atomic.
> I have never seen real Go code (i.e. not code written purposefully to be exploitable) that was exploitable due to a data race.
And from tptacek in that same discussion [2]:
> The fact is that Go doesn't admit memory corruption vulnerabilities, and the way you know that is the fact that there are practically zero exploits for memory corruption vulnerabilities targeting pure Go programs, despite the popularity of the language.
I am, but you’d be surprised. All of the single-threaded languages are fine, for example. Very few languages are actually low-level enough to allow data races. C# and Java go to great lengths to avoid it.
Race conditions in general are another matter, and aren’t generally considered a requirement (though you can certainly create nasty bugs).
The claim is not there is no data races in Java programs. The claim is that Java programs are memory safe, because the underlying virtual machine memory model is free of data races.
Go does not have that.
(Of course also Java may suffer from memory safety issues on system boundaries to unsafe code and due to JVM bugs.)
I suppose that if you create a map in one thread, and then access it from another thread, then that might cause segmentation faults? Because map is a type that is implemented in C.
(The gist: memory safety is a term of art coined by security practitioners. Go, Python, Rust, Java, others: memory safe. C/C++: memory unsafe. Periodically, people in different slices of industry or academia come up with new definitions of memory safety that declare Rust or Go or other languages to be memory unsafe, but that is not by the broadly accepted definition across industry.)
I define undefined behaviour as a bug in C++. Now C++ is memory-safe!
btw, it's not actually that hard to write correct code in C++, easier than in C because you have all the container types. The problem is that nothing will tell you when you write incorrect code - there's no guarantee.
UB is part of the C++ standard. Surely you can understand the difference between the C++ standard and bugs in compilers implementing the C++ standard. This is that.
> I define undefined behaviour as a bug in C++. Now C++ is memory-safe!
I mean, sure, insofar as such a thing would also imply that a) the standard would need quite a bit of cleanup/clarification work to not contradict your definition, and b) the main optimizing C++ compilers are miscompiling code, analogous to how cve-rs is a rustc miscompilation rather than an issue with Rust itself.
(Fil-C might be an interesting exception here, though IIRC its definition of memory safety is slightly different)
Since we are not at some point in the future where that correct compiler exists and there is only one official compiler, the distinction you make is practically meaningless!
I mean… no? It matters whether something is a part of the language or not, because it matters if you can write code relying on this behavior. Since this is a compiler bug, you cannot - the code will stop compiling the moment the bug is fixed.
There are no known instances of this bug being encountered in the wild, and if you look into it, you will see how extremely unlikely such code is.
> Isn't ten years enough to call something a feature of the language rather than a bug?
I suppose it depends on who is doing the classifying? From the developer's standpoint I'd imagine intent is all that matters: a bug is something that does not match developer intent and that is (eventually) expected to be changed to match the intent, while a feature is something that does match developer intent regardless of how old/new it is. From a user's standpoint I'd imagine it's a combination of developer intent and the user's reliance on said behavior, but IIRC in this particular case there's no known non-demo code that has organically run into this particular bug so there's little weight in favor of calling the bug a feature despite the devs' stance.
Also as GP said I think one needs to be careful to distinguish between the compiler and the language. IIRC the devs have known for basically this entire time exactly in what manner the Rust compiler fail to implement the rules of Rust the language, but a general fix has been blocked on long-running projects that have only recently been approaching the finish line [1].
You can write unsafe code in Go (import unsafe), but then, you can do the same in Rust. Unsafe code is not the default, and in day to day Go i rarely see the use of the unsafe package.
What he probably means is data-races in go can result in memory/type unsafe accesses -- I suspect, likely due to slice types -- not sure if that is true/false.
Sure, but a data race is, IMHO not the same as memory safety. A data race, can be 100% memory safe, but just cause a logic bug in some program. I often see people mixing memory safety with racing. Go has bounds checks so you end up with a panic either way. Not UB.
As an (outside go) example, Ocaml (5) promises strong memory safety, but not to be data race free. A data race is not something we can prevent, because its usually not bound by code, but by time and the race-source rarely in source-code.
This means we have data races in http, database inserts etc. The source is usually not a concurrent task in source code-land.
Now you are pushing pixels. A data-race IS a kind of race condition.
My point is "races" happen all over. In concurrent code, databases, http and pretty much anywhere where you have some kind of timing, not scoped to a unit.
A race condition is an application invariant violation under concurrency, so it has no application-independent definition. A data race is unsynchronized access by two concurrent threads to the same memory location where at least one access is a write. Ergo, the definition of a data race has nothing to do with application logic.
I think this should make the distinction between data races and race conditions pretty clear.
A data race is by definition a class of an race condition. This is basic compsci literature, and there is no other way to put it, as even a non programmer can see the similarity.
Not sure why you would be that nitpicky for something so trivial?
I find this distinction extremely useful in practice because it explains why a static analyzer like Rust's type checker can prevent all data races but can do nothing about race conditions. (Similarly, a dynamic analyzer like ThreadSanitizer can detect data races but not race conditions, because it knows nothing about application logic.)
This is not generally considered part of the "memory safety" contract. You can not lift a nil pointer exception into a replacement for Go's "unsafe" library.
When we finally rid ourselves of C and C++ is so larded over with extensions and additions and features that we can finally plausibly say the C subset is just not in use anymore, we can perhaps consider as a community expanding what "memory safe" means, but in the meantime it has some very important meanings and we should not try to augment the term. Memory safety doesn't mean anything like "forcing exhaustiveness into sum type deconstructions" or "never has a race condition" (though it does mean said race condition shouldn't be something that allows you to escape out of an array or forcibly change the type on something in a way the language doesn't normally permit) or any of several other things that may be very nice to have indeed, but are not part of the definition of "memory safe".
Memory safe is a very old concept, and almost everything is memory safe now. But not quite, and as such the term still has use. And also zig for some reason gave it up so it won't be disappearing as soon as I'd like.'
Very neat, and comes pretty close to how Mojo handles portable SIMD.
It's great to see two of my favorite languages finally making SIMD easy to use. It's such low-hanging fruit for performance, yet somehow languages have ignored it for years. Portable SIMD, even with some performance penalty, still beats scalar computation whenever vector operations are needed. Yet language implementations always seemed to assume that hardware-specific SIMD APIs were the only way to go. That did nothing but make SIMD unusable excepting special cases where performance is absolutely critical, rather than just something anyone can use in day to day programming.
> The new simd package hides these differences by removing fixed-size vectors from the type system, and by only supporting those operations that are in the intersection of all the different platforms, and fills gaps in the intersection with efficient emulation in terms of other SIMD instructions.
The intersection would be the operations supported by all platforms and so would not have gaps.
> The interface conversion and type switch look like they should be inefficient, but the compiler-side implementation of simd specializes code and optimizes away the type switch.
I don’t understand this - how is it able to if the same go binary might run on unknown types? I’m assuming what it means is that the switch is implemented efficiently due to CPU branch prediction? I know fearless SIMD is doing cool stuff with static dispatch so that the feature set is checked just once at program start - is that what it means it’s doing under the hood? Very unclear.
It creates multiple versions of functions referencing SIMD and lifts the dispatch switching cost to their callers.
> The AST rewrite creates multiple specialized copies of functions, variables, and types that mention simd types, where simd types are replaced with references to size-specialized types in simd/internal/bridge. Each of these bridge types is defined as an archsimd type, but with a restricted set of methods. The specialized functions, variables, and types acquire a suffix of the form @simdNNN, where NNN is either a vector length (128, 256, or 512) or 0, indicating emulation. Functions that mention simd internally, but not in their signature, are converted to wrappers that switch on the SIMD level detected at program start, and call the appropriate specialized version of that function. Specialized functions call other specialized functions directly without dispatch overhead (and perhaps with inlining). This rewrite strategy was chosen as a compromise between code duplication and SIMD performance; the overhead is hoisted as high as necessary to avoid dispatch within SIMD computations, but not higher. If SIMD dispatch appears “too low” in a computation, a gratuitous mention of a simd type will move it upwards, as in this example:
Question on the multi-versioning approach: if the AST rewrite creates N specialized copies of every function that mentions the SIMD types, doesn't binary size scale with the number of SIMD-touching functions? For generic hot paths (helpers parameterized over vector widths, instantiated across many call sites) that could multiply code size noticeably. Is there dedup when two specialized copies would be identical, or has anyone measured the binary-size cost on a real codebase?
It's hard to make predictions with an open source project, but my personal guess is some flavor of it will land (including it is already demonstrating good results without an enormous level of code complexity in the compiler and without overly slowing down compile speeds), but I guess we'll see.
It's being driven by an external contributor who has landed some good changes in the past to the Go compiler. (I think the autovectorization work might be part of their PhD or other academic research, but not sure.)
As a first step, it might be possible to write a linter rule that rewrites suitable numeric loops to SIMD. There are already rules to rewrite several loop types, so that should be doable.
The problem with Go isn't performance but with the C/C++ interop overhead, even with the "30% less overhead" from a few updates ago which isnt true for 99% of cases, it isnt enough
its not limited but it has overhead because of the memory model of go doesnt match the C one so there has to be some sort of rerodering being done, that's what i understood atleast, and theres also the go concurrency
Use Assembly instead of CGO, isn't that scary, back in the 8 bit days we were coding Assembly aged 10, on our Spectrum, C64, Atari, Apple, Acorn, MSX,....
> Syntax highlighting is juvenile. When I was a child, I was taught
arithmetic using colored rods
(http://en.wikipedia.org/wiki/Cuisenaire_rods). I grew up and today I
use monochromatic numerals.
Not convinced at all by these reasons. Rob Pike has long stepped down from the Go team so it shouldn't matter? This is a website not a language change.
Already using this for foreground estimation of cutouts in my project, around 30% speedup over non-SIMD, but the algorithm is probably not very optimised yet.
I personally use both, and keep using both. There are much more Golang job in the market now. Noone planning to ditch Go in my network or unhappy with it. Highload E-commerce, logistics, etc are way easier to write in Go IMHO.
Discover a very good niche for Rust - geo spatial analytics. Would not do it Go or Python. LLMs gave a huge boost to Rust too. Claude produce a very high code ... if designed right. Lot of feature complete libraries now.
I suspect you stopped reading at web services on your link, but API is indeed the correct word to describe a set of functions from a library (built-in or not). If you disagree perhaps you should share your preferred term here?
> The term API is often used to refer to web APIs, which allow communication between computers that are joined by the internet. There are also APIs for programming languages, software libraries, computer operating systems, and computer hardware.
https://imjasonh.github.io/playground/palette-swap/ swaps colors in a provided image in wasm, entirely locally in your browser, to benchmark portable SIMD vs non-portable archsimd vs non-SIMD.
Portable SIMD is ~11% slower than non-portable SIMD in this case, but both are ~5x faster than non-SIMD.
Just want to say among many portable SIMD solutions I’ve seen recently (e.g. Fearless SIMD), this is the first that makes non-fixed vectors like SVE and RISC-V vector (RVV) easier to support. Glad to see they made this decision
We pioneered this in Highway and shared some advice on the API. Great to see this decision taken :D
https://github.com/google/highway
How so? I imagine you'd still want to constrain the length to the maximum vector size supported by the lowest platform you want to support or you lose the portability and actually end up with code that performs much worse than the scalar alternative on some platforms.
Mojo has an even more portable simd[1] type that isn't just generic over length but also over type. In my opinion it is almost always better to specialize for each platform and use portable implementation as fallback. It's a shame that just very few languages support Zig-like comptime, because it would be excellent for specializations without introducing runtime penalties.
1. https://mojolang.org/docs/std/simd/SIMD/
> constrain the length to the maximum vector size supported by the lowest platform you want to support or you lose the portability and actually end up with code that performs much worse than the scalar alternative on some platforms.
Or, put a dynamic factor into your vector size and design everything around it. Such that every platforms can plug in their own factor and _scale_ the size of vectors. This is basically what LLVM IR does for SVE and RVV: `<vscale x 4 x i32>` where vscale is the said dynamic factor. Though the exact value of vscale is only known during runtime, it doesn't matter -- we still can design compiler optimizations and lowering around it. The generated binaries can then be portable across platforms with different vscale values.
C++ is getting std::simd in the latest version and I am all aboard writing the vectorization with the least amount of intrinsic builtins I am able to. Even if not optimal, it's far better than the scalar ops.
Seconded!! This doesn’t really help the well established codebases much that are already doing this on a platform specific path but in general this is much appreciated for the future.
Write it once with N errors, not N*M errors :)
This feature opens many doors for optimizing low-level performance in Go projects, that are already running multicore. IIRC there aren’t a lot of languages with built-in std lib support for SIMD and variants. Love the way Go is trying new stuff lately.
Besides the usual C and C++, we have Java, .NET, D, Zig, Julia, Swift, Rust.
So yeah, also appreciate having Go in the group instead of manually having to write Assembly.
However not many languages adopt ways to manually write SIMD, because most of us have no idea how to write good SIMD code in first place, I surely don't.
You probably weren't looking for a tutorial about SIMD, but just in case you were interested, Mitchell [0] did one recently that got on HN [1]
[0] Mitchell Hashimoto: "Everyone Should Know SIMD" https://mitchellh.com/writing/everyone-should-know-simd
[1] https://news.ycombinator.com/item?id=49010648
We recently published another: https://github.com/google/highway/blob/master/g3doc/tutorial...
Thanks, pretty much appreciated, always worth at least to skim through.
I definitely was so... thanks!
Even with languages that adopt ways to manually write SIMD, it’s mostly left to library maintainers rather than application developers.
I work for a C++ timeseries database startup that leverages SIMD about as much as we possibly can, and except for some extremely rare places we just use libraries.
Yeah, that is what I have heard from some NVidia folks as well, like Bryce Adelstein, use the libraries as much as possible, and leave the kernels for experts.
However even then, it depends on how the libraries API surface looks like.
With AI I'm pretty sure SIMD will be easier to integrate when necessary.
But it’s not necessary at all, the whole point is that these utility libraries bring you more elegant code that work on all platforms without having to pollute your codebase with SIMD intrinsics.
Unless this was tongue in cheek, because this is in fact a problem with AI that it degrades your codebase in these types of ways.
In 2026 if you are not doing A with AI you are doing it wrong /s
With AI, I expect it to eventually be good enough for us to finally have 5 GLs, so it won't really matter.
"CGO 2022 Keynote: Compiler 2.0"
https://www.youtube.com/watch?v=w_sX9aZoZxg
Vectorizing computations has been Matlabs secret sauce.
Does matlab these days do stuff like JIT operator fusing to avoid memory roundtrips and take advantage of FMAs?
Yes: https://www.mathworks.com/help/fixedpoint/ref/half.fma.html
I'm grateful that Go a non-proprietary language offers these features.
Julia does that too.
I did some testing with the experimental SIMD on a project I was doing to make speech-to-text and text-to-speech models run natively in Go (with CGO_ENABLED=0, so no C depenencies), and testing non-SIMD w/ SIMD.
I don't have formal benchmarks for that, but I can anecdotally say the SIMD work made a measurable improvement in the performance of the calculations vs. just plain Go. I'm very optimistic about how these improvements will help make the Go runtime an even better target for more of these types of work going forward, especially since it is cross-platform.
This is why I love Go. Nobody was asking for this, but they took the time to do it right and continue to Push go as a memory safe, high-level systems language.
People were definitely asking for it.
It's been discussed for a long time, and the related proposals were heavily upvoted, including various older proposals.
As I understand it, part of the reason it took a while is that the core Go team was generally of the opinion that doing user-facing SIMD APIs the right way was to design a high-level, cross-platform API that would stand the test of time, and that was then punted a few times given its complexity and need to do other things.
Part of what helped the current approach take off was switching to a philosophy of designing a lower-level architecture-dependent API first (the 'simd/archsimd' package), and then later doing a higher-level portable API (the 'simd' package, which is topic of this blog post).
That two-level approach I think also gave some additional freedom for the design and implementation of the friendlier / high-level 'simd' package, including because the lower-level 'simd/archsimd' package is available for people who need or want to drop down.
It's a nice design.
Layered API design is a great way to resolve ergonomics/performance tradeoffs.
This is kind of the opposite of Go. Not giving people what they are asking for.
There are pros and cons of course. You don't have 17 different ways to iterate over an array, so that's nice. But you also went 13 years without generics, despite them being one of the most requested features, because the designers didn't want that complexity inside Go.
Overall I think Go is better for this philosophy but there are times where the language is clearly written more for its maintainers than it's users.
> But you also went 13 years without generics
Go shipped with generics (aka bounded parametric polymorphism), but only for built-in types: slices, arrays, and maps. That, with subtyping via interfaces, handled most demand for generics. The most common pain point was custom containers.
Go was first released in November 2009. Russ Cox posted "The Generic Dilemma" [1] in December 2009. The comments show the generics debate raging from the earliest days.
As a fun side note, I forgot I posted a comment on that post pointing to Ada's generics. I was in college, and Ada was our intro language.
[1]: https://research.swtch.com/generic
> the designers didn't want that complexity inside Go.
Yes, with some nuance. Go's goal of writing server programs didn't require the type-system complexity and run-time hit of user-defined generics. [2]
> Go was intended as a language for writing server programs [...] Polymorphic programming did not seem essential [...] so was initially left out for simplicity. > > Generics are convenient but they come at a cost in complexity in the type system and run-time. It took a while to develop a design that we believe gives value proportionate to the complexity.
[2]: https://go.dev/doc/faq#beginning_generics
Out of curiosity, I collected all proposals for Go's journey to generics. https://gist.github.com/jschaf/eaa7aff1af14ea7276a18a1b7370d...
Some of the concerns around generics and why it took so long were for the users as well. One of the biggest draws to Go has always been that you get the performance of a compiled language and yet compile times are so low that it can feel like you're developing with an interpreted language. The design of generics needed to maintain the compile time advantage or else it wouldn't feel like Go any more.
Mostly safe, contrary to other safer languages, Go memory model doesn't prevent data tearing.
go data races aren't memory safe
That's not what "memory safe" means. "Memory safe" is a term of art meaning "not susceptible to memory corruption exploits", like stack and heap overflows, UAFs, and type confusion. Last I checked, there are essentially no non-contrived memory corruption exploits for Go programs; the best you get are people demonstrating register control on contrived programs.
The definition I'm giving is the same as the ISRG's definition at MemorySafety.org. It's the thing everybody is talking about when they talk about memory safety.
The claim being made here is "big if true", because it would imply a lot more languages than Go "aren't memory safe", despite decades without memory corruption exploits.
The exploits aren't the only issue; they just get a lot of air time.
It's remarkably easy to segfault Go applications with data races. Any object with multiple words (so a slice that's an array pointer, a size, and a capacity. Or a fat pointer with the object pointer and the vtable pointer), can be read in an inconsistent state from two threads which can cause out of bounds reads and writes. And this comes up all the time with how heavily the language encourages concurrency.
It's just difficult to actually exploit because of other considerations that practically add a lot of runtime entropy.
They aren't the only issue in programming language theory, but they are the only issue in the ordinary context in which we discuss "memory safety", such that if someone not in a PLT forum says "is Go memory-safe" and you say "no" you will look a little batty.
The was for a time a vogue for "zero trust networking" and I'm fond of pointing out that the same thing happened there: people would come up with their own axiomatic derivation of what "zero trust" meant, but in reality it was a term of art meaning "non-Google implementations of BeyondCorp".
Terms of art are kryptonite for message board nerds.
No, it has real world implications. I worked at a heavy go shop; every time we'd have a new batch of hiring, the segfault bugs would start rolling in from these memory unsafety issues. The data race detector was good, but not perfect. Yes, those new devs would look at you batty until they saw the bug reports roll in.
And yeah, we tend to use the PLT definitions when we're talking about literal semantics of programming languages. Nothing in this thread mandates a security focused sub-definition.
And even Rob Pike described Go as "not purely memory safe", in a slide that obviously references this exact data race behavior. https://go.dev/talks/2012/splash.slide#49 They considered this a practical tradeoff for simplicity versus the major managed languages defining what happens during data races in a way that doesn't allow you to break memory safety.
Again: I am not disputing that there are correctness issues that come from having Go's concurrency and memory model, just like there are in (the large majority of) other languages with similar models. But those issues are not security problems, and security is what we're talking about when we talk about memory safety.
Yeah, we get it, you performatively hate go.
Yes, but in practice they are extremely hard to exploit. It has been discussed extensively here on HN and in other forums.
That does not make sense to me. Go is memory-safe, but it does not guarantee data-race freedom.
So whats your point here? Haskell?
Probably Rust, that's always Rust with this kind of comments...
Or basic software engineering and understanding of what memory safe means.
Memory safety != Data race free
Yes, but Rust gets you both.
> That does not make sense to me.
You said that because you assumed Go is memory-safe in all conditions.
> Go is memory-safe
Yes, but only if there's no data race.
Go is not like Java. Java doesn't guarantee no data race, but when it happens, it's still memory-safe.
Neither Java or Go guarantees data-race freedom. A data race does not by itself make ordinary Java or Go code memory-unsafe in the C/C++ sense.
I fail to see how a racy Java program is more memory safe than a racy Go program?
Java arrays are thin pointers to Array objects, which contain the length and data in the same place (on the far side of the pointer). Since thin pointers cannot tear and Array objects cannot be resized, Arrays themselves are always memory-safe in safe code.
Go slices are fat pointers to undecorated memory. The slice itself is a 3-tuple of pointer, length, and capacity. If you append to a slice that's already at capacity, the Go runtime will allocate new memory for you and return a new 3-tuple. If you assign that result to a variable that's also being accessed by another goroutine, the latter can observe the slice in an inconsistent state. It can, for example, see the old pointer but with the new length, allowing out-of-bounds access. None of this requires unsafe code.
The same issue applies to string and interface variables, which are also fat pointers.
Indeed.
Shared Go slices are a bad mix in concurrent code. This is a given. But its also not a fair comparison, you should instead compare java arrays to go arrays, not slices.
This goes for slices, strings and maps. Those a usually wrapped in a mutex, or used with sync primitives like sync.Map.
The exact equivalences between the two languages are not the point, and anyway, Go arrays are fixed-size and have no Java equivalent.
The point is that Java does not have this problem in the language or the standard library. Of course, you should not write racy Go code; the language provides ample ways to avoid the race, such as channels; and the race detector will generally find such racy code, provided you turn it on. However, the issue is that this race leads to memory-safety violations; it can occur especially in code written by novice Go programmers, and it's well acknowledged by the language authors [1].
[1]: https://research.swtch.com/gorace
> Those a usually wrapped in a mutex, or used with sync primitives like sync.Map.
The same can be said of C or C++ constructs (and many "anti-Rust" people have historically said that) -- the point is that their use is not enforced by the language and so bugs can lead to panics.
I write a fair amount of Go and Rust so I really don't think either language's flaws are fatal, but it comes off as weirdly defensive to redefine memory and data safety to be "well if you use it properly it's safe". It's totally fine to say this is a problem the Go language did not find important enough to require compile time enforcement and so solving it is done by convention and testing with the race detector (which a similar answer C and C++ give to this problem).
From what I understand a data race on a simple built in feature like an interface pointer can result in a bad address / type pair, which can cause memory safety issues on any future access. I don't think you can get the JVM itself confused about what type a pointer points to.
There is no memory safety without freedom from data races. One is a prerequisite of the other. This is why languages like C# throw exceptions on unsynchronized concurrent access to some container types, and treat all property accesses as atomic.
From the former head of the Go security team [1]:
> I have never seen real Go code (i.e. not code written purposefully to be exploitable) that was exploitable due to a data race.
And from tptacek in that same discussion [2]:
> The fact is that Go doesn't admit memory corruption vulnerabilities, and the way you know that is the fact that there are practically zero exploits for memory corruption vulnerabilities targeting pure Go programs, despite the popularity of the language.
[1] https://news.ycombinator.com/item?id=44672003
[2] https://news.ycombinator.com/item?id=44672371
Memory safety isn’t really about vulnerabilities. This thread is about Go, so I won’t go further into it here.
Are you saying any language that does not promise data-race freedom is memory unsafe? That would rule out almost every programming language.
I am, but you’d be surprised. All of the single-threaded languages are fine, for example. Very few languages are actually low-level enough to allow data races. C# and Java go to great lengths to avoid it.
Race conditions in general are another matter, and aren’t generally considered a requirement (though you can certainly create nasty bugs).
I have not written a line of Java in 15 years. But im pretty sure Java has threads? Once you have threads, you pretty much have data races.
The claim is not there is no data races in Java programs. The claim is that Java programs are memory safe, because the underlying virtual machine memory model is free of data races.
Go does not have that.
(Of course also Java may suffer from memory safety issues on system boundaries to unsafe code and due to JVM bugs.)
I don’t know Java or the JVM, but if it’s anything like the CLR/.NET, this would either be compiled to atomic operations, or throw an exception.
I suppose that if you create a map in one thread, and then access it from another thread, then that might cause segmentation faults? Because map is a type that is implemented in C.
(removed)
Go is broadly considered to be a memory safe language.
See for example comments from tptacek like:
https://news.ycombinator.com/item?id=43335748
https://news.ycombinator.com/item?id=46028232
https://news.ycombinator.com/item?id=44672371
(The gist: memory safety is a term of art coined by security practitioners. Go, Python, Rust, Java, others: memory safe. C/C++: memory unsafe. Periodically, people in different slices of industry or academia come up with new definitions of memory safety that declare Rust or Go or other languages to be memory unsafe, but that is not by the broadly accepted definition across industry.)
Rust does allow you to overflow buffers, confuse types, and duplicate mutable pointers in safe code. See cve-rs.
No, Rust does not allow that. The current Rust compiler does, but that’s a bug that is being fixed.
At some point in the future, a fully backwards compatible Rust compiler will report an error when you try to compile cve-rs.
I define undefined behaviour as a bug in C++. Now C++ is memory-safe!
btw, it's not actually that hard to write correct code in C++, easier than in C because you have all the container types. The problem is that nothing will tell you when you write incorrect code - there's no guarantee.
UB is part of the C++ standard. Surely you can understand the difference between the C++ standard and bugs in compilers implementing the C++ standard. This is that.
> I define undefined behaviour as a bug in C++. Now C++ is memory-safe!
I mean, sure, insofar as such a thing would also imply that a) the standard would need quite a bit of cleanup/clarification work to not contradict your definition, and b) the main optimizing C++ compilers are miscompiling code, analogous to how cve-rs is a rustc miscompilation rather than an issue with Rust itself.
(Fil-C might be an interesting exception here, though IIRC its definition of memory safety is slightly different)
Since we are not at some point in the future where that correct compiler exists and there is only one official compiler, the distinction you make is practically meaningless!
I mean… no? It matters whether something is a part of the language or not, because it matters if you can write code relying on this behavior. Since this is a compiler bug, you cannot - the code will stop compiling the moment the bug is fixed.
There are no known instances of this bug being encountered in the wild, and if you look into it, you will see how extremely unlikely such code is.
Isn't one of the bugs around ten years old, now? Isn't ten years enough to call something a feature of the language rather than a bug?
I like Rust, but with this bug existing for so long, I personally no longer think of it as memory-safe.
> Isn't ten years enough to call something a feature of the language rather than a bug?
I suppose it depends on who is doing the classifying? From the developer's standpoint I'd imagine intent is all that matters: a bug is something that does not match developer intent and that is (eventually) expected to be changed to match the intent, while a feature is something that does match developer intent regardless of how old/new it is. From a user's standpoint I'd imagine it's a combination of developer intent and the user's reliance on said behavior, but IIRC in this particular case there's no known non-demo code that has organically run into this particular bug so there's little weight in favor of calling the bug a feature despite the devs' stance.
Also as GP said I think one needs to be careful to distinguish between the compiler and the language. IIRC the devs have known for basically this entire time exactly in what manner the Rust compiler fail to implement the rules of Rust the language, but a general fix has been blocked on long-running projects that have only recently been approaching the finish line [1].
[0]: https://news.ycombinator.com/item?id=40431444
[1]: https://blog.rust-lang.org/2026/08/21/enabling-next-solver-o...
And yet, there's a C compiler (fil-c) that doesn't allow that.
He’s very wrong about this. Just because ‘tptacek posts a lot and did security once upon a time does not make him “broad consideration”.
You can write unsafe code in Go (import unsafe), but then, you can do the same in Rust. Unsafe code is not the default, and in day to day Go i rarely see the use of the unsafe package.
What he probably means is data-races in go can result in memory/type unsafe accesses -- I suspect, likely due to slice types -- not sure if that is true/false.
Sure, but a data race is, IMHO not the same as memory safety. A data race, can be 100% memory safe, but just cause a logic bug in some program. I often see people mixing memory safety with racing. Go has bounds checks so you end up with a panic either way. Not UB.
As an (outside go) example, Ocaml (5) promises strong memory safety, but not to be data race free. A data race is not something we can prevent, because its usually not bound by code, but by time and the race-source rarely in source-code.
This means we have data races in http, database inserts etc. The source is usually not a concurrent task in source code-land.
Russ Cox points out that races are the one place in Go besides unsafe where Go lacks memory safety: https://research.swtch.com/gorace
I do think the nitpicking about this is mostly from people that want to say “my favorite language is safer than Go” which stupid and annoying.
Races can be memory safe (they are in Java and Ocaml), but in Go they can indeed cause UB.
https://go.dev/ref/mem#restrictions:~:text=such%20races,corr...
https://www.ralfj.de/blog/2025/07/24/memory-safety.html (I don't agree with everything here, but it's a very thorough explanation)
That’s a race condition, not a data race.
Now you are pushing pixels. A data-race IS a kind of race condition.
My point is "races" happen all over. In concurrent code, databases, http and pretty much anywhere where you have some kind of timing, not scoped to a unit.
A race condition is an application invariant violation under concurrency, so it has no application-independent definition. A data race is unsynchronized access by two concurrent threads to the same memory location where at least one access is a write. Ergo, the definition of a data race has nothing to do with application logic.
I think this should make the distinction between data races and race conditions pretty clear.
A data race is by definition a class of an race condition. This is basic compsci literature, and there is no other way to put it, as even a non programmer can see the similarity.
Not sure why you would be that nitpicky for something so trivial?
I find this distinction extremely useful in practice because it explains why a static analyzer like Rust's type checker can prevent all data races but can do nothing about race conditions. (Similarly, a dynamic analyzer like ThreadSanitizer can detect data races but not race conditions, because it knows nothing about application logic.)
No idea why you're getting downvoted for true statement. Without a ? like in C# you're always at risk of a nil pointer being dereferenced
> like in C# you're always at risk of a nil pointer being dereferenced
That throws a NullReferenceException
Cool, very clever. But at least the compiler warns me of a potential exception, whereas in Go no such op even exists.
This is not generally considered part of the "memory safety" contract. You can not lift a nil pointer exception into a replacement for Go's "unsafe" library.
When we finally rid ourselves of C and C++ is so larded over with extensions and additions and features that we can finally plausibly say the C subset is just not in use anymore, we can perhaps consider as a community expanding what "memory safe" means, but in the meantime it has some very important meanings and we should not try to augment the term. Memory safety doesn't mean anything like "forcing exhaustiveness into sum type deconstructions" or "never has a race condition" (though it does mean said race condition shouldn't be something that allows you to escape out of an array or forcibly change the type on something in a way the language doesn't normally permit) or any of several other things that may be very nice to have indeed, but are not part of the definition of "memory safe".
Memory safe is a very old concept, and almost everything is memory safe now. But not quite, and as such the term still has use. And also zig for some reason gave it up so it won't be disappearing as soon as I'd like.'
you can dereference nil in Go and it panics
And then use Go's pseudo exception handling and recover.
This will welcome more database/warehouses to be written in Go.
Personally I will implement it in https://github.com/viggy28/streambed
They could have used third party packages or Assembly directly.
This naturally is an easier way.
You're right. I could have but this encourages me to seriously consider it.
Very neat, and comes pretty close to how Mojo handles portable SIMD.
It's great to see two of my favorite languages finally making SIMD easy to use. It's such low-hanging fruit for performance, yet somehow languages have ignored it for years. Portable SIMD, even with some performance penalty, still beats scalar computation whenever vector operations are needed. Yet language implementations always seemed to assume that hardware-specific SIMD APIs were the only way to go. That did nothing but make SIMD unusable excepting special cases where performance is absolutely critical, rather than just something anyone can use in day to day programming.
Curious how much of the emulation ends up in hot paths before SVE and the feature variants land.
> The new simd package hides these differences by removing fixed-size vectors from the type system, and by only supporting those operations that are in the intersection of all the different platforms, and fills gaps in the intersection with efficient emulation in terms of other SIMD instructions.
The intersection would be the operations supported by all platforms and so would not have gaps.
I am using Go assembly for SIMD very heavily in https://github.com/rcarmo/go-pherence, this is just icing on the cake.
> The interface conversion and type switch look like they should be inefficient, but the compiler-side implementation of simd specializes code and optimizes away the type switch.
I don’t understand this - how is it able to if the same go binary might run on unknown types? I’m assuming what it means is that the switch is implemented efficiently due to CPU branch prediction? I know fearless SIMD is doing cool stuff with static dispatch so that the feature set is checked just once at program start - is that what it means it’s doing under the hood? Very unclear.
It creates multiple versions of functions referencing SIMD and lifts the dispatch switching cost to their callers.
> The AST rewrite creates multiple specialized copies of functions, variables, and types that mention simd types, where simd types are replaced with references to size-specialized types in simd/internal/bridge. Each of these bridge types is defined as an archsimd type, but with a restricted set of methods. The specialized functions, variables, and types acquire a suffix of the form @simdNNN, where NNN is either a vector length (128, 256, or 512) or 0, indicating emulation. Functions that mention simd internally, but not in their signature, are converted to wrappers that switch on the SIMD level detected at program start, and call the appropriate specialized version of that function. Specialized functions call other specialized functions directly without dispatch overhead (and perhaps with inlining). This rewrite strategy was chosen as a compromise between code duplication and SIMD performance; the overhead is hoisted as high as necessary to avoid dispatch within SIMD computations, but not higher. If SIMD dispatch appears “too low” in a computation, a gratuitous mention of a simd type will move it upwards, as in this example:
Question on the multi-versioning approach: if the AST rewrite creates N specialized copies of every function that mentions the SIMD types, doesn't binary size scale with the number of SIMD-touching functions? For generic hot paths (helpers parameterized over vector widths, instantiated across many call sites) that could multiply code size noticeably. Is there dedup when two specialized copies would be identical, or has anyone measured the binary-size cost on a real codebase?
Oh this is great, it was one of my biggest bugbears about Go since you almost always have to link C/C++ code to get the appropriate performance.
The one negative I'd say is that often autovectorisation is 'good enough' and this doesn't really tackle that gap.
FWIW, there is some pretty substantial autovectorization work that is already in-flight for the Go compiler.
There's a CL stack here:
https://go.dev/cl/791740
It's hard to make predictions with an open source project, but my personal guess is some flavor of it will land (including it is already demonstrating good results without an enormous level of code complexity in the compiler and without overly slowing down compile speeds), but I guess we'll see.
It's being driven by an external contributor who has landed some good changes in the past to the Go compiler. (I think the autovectorization work might be part of their PhD or other academic research, but not sure.)
As a first step, it might be possible to write a linter rule that rewrites suitable numeric loops to SIMD. There are already rules to rewrite several loop types, so that should be doable.
The poor Assembler and the unsafe package forgotten in the corner.
While reaching out to CGO is the easier way, it doesn't mean it is the only tool available in Go.
The problem with Go isn't performance but with the C/C++ interop overhead, even with the "30% less overhead" from a few updates ago which isnt true for 99% of cases, it isnt enough
Why is that the case? I don’t know low level programming so why is Go limited in interop with C?
its not limited but it has overhead because of the memory model of go doesnt match the C one so there has to be some sort of rerodering being done, that's what i understood atleast, and theres also the go concurrency
Use Assembly instead of CGO, isn't that scary, back in the 8 bit days we were coding Assembly aged 10, on our Spectrum, C64, Atari, Apple, Acorn, MSX,....
For God sake, add a syntax highlighting on the official page! Otherwise this is awesome
Rob Pike seems to dislike syntax highlighting:
> Syntax highlighting is juvenile. When I was a child, I was taught arithmetic using colored rods (http://en.wikipedia.org/wiki/Cuisenaire_rods). I grew up and today I use monochromatic numerals.
https://groups.google.com/g/golang-nuts/c/hJHCAaiL0so/m/kG3B...
He is definitely not alone in thinking this, for example (but for different reasons): https://www.linusakesson.net/programming/syntaxhighlighting/
I like Linus's work, but that post smells of rotting strawmen. All the examples look manipulated into throwing the opposition under the bus.
Not convinced at all by these reasons. Rob Pike has long stepped down from the Go team so it shouldn't matter? This is a website not a language change.
I hope portable simd will be stabilized some time in rust :/
Already using this for foreground estimation of cutouts in my project, around 30% speedup over non-SIMD, but the algorithm is probably not very optimised yet.
How many "functions" compile to movd?
Rust kind of seems to have overtaken Go in momentum recently. I wonder if Go will do well in, say, two years from now on.
I personally use both, and keep using both. There are much more Golang job in the market now. Noone planning to ditch Go in my network or unhappy with it. Highload E-commerce, logistics, etc are way easier to write in Go IMHO.
Discover a very good niche for Rust - geo spatial analytics. Would not do it Go or Python. LLMs gave a huge boost to Rust too. Claude produce a very high code ... if designed right. Lot of feature complete libraries now.
Both will do fine
> Go 1.26 and 1.27 include experimental APIs for Single Instruction Multiple Data (SIMD) operations.
You'd think these people would know the meaning of API, no?
One wonders what overly-narrow definition of API you're stuck on.
One wonders why you wonder.
https://en.wikipedia.org/wiki/API
I suspect you stopped reading at web services on your link, but API is indeed the correct word to describe a set of functions from a library (built-in or not). If you disagree perhaps you should share your preferred term here?
> The term API is often used to refer to web APIs, which allow communication between computers that are joined by the internet. There are also APIs for programming languages, software libraries, computer operating systems, and computer hardware.
> API is indeed the correct word to describe a set of functions from a library (built-in or not).
Uncorroborated by WP, note.
> If you disagree perhaps you should share your preferred term here?
Library functions.