Golang on ARM
github.com
github.com
Contrast that with C/C++ - sure you can cross-compile, and I have in the past, but it's a huge pain to set up and who ever remembers the command line options. Those "target triples" aren't even documented anywhere. Plus, good luck building a statically linked C++ program.
Lua might be a decent option too.
Also, FWIW, the reason why Go can even be cross compiled in the first place is because it has platform specific C and assembly code from cgo's runtime so it can support the different architectures/platforms.
Of course, if Go was just an LLVM frontend instead of using gcc then it wouldn't even need the platform specific C code... And it would support a shitton more platforms like asm.js. But I guess Google just hates Apple that much?
This is silly. Google contributes a huge amount to LLVM. They're very invested in the project.
(NB: I'm not defending, and don't agree with, the choice to avoid LLVM in Go.)
"Most importantly, it has been something we understand completely and is quite easy to adapt as needed. The standard ABIs and toolchains are for C, and the decisions made there may be completely inappropriate for languages with actual runtimes.
For example, no standard ABIs and toolchains supported segmented stacks; we had to build that, so it was going to be incompatible from day one. If step one had been "learn the GCC or LLVM toolchains well enough to add segmented stacks", I'm not sure we'd have gotten to step two."
"... the custom toolchain is one of the key reasons we've accomplished so much in so little time. I think the author doesn't fully appreciate all the reasons that C tools don't work well for languages that do so much more than C. Most other managed languages are doing JIT compilation and don't go through standard linkers either."
LLVM now supports both segmented stacks [1] and precise garbage collection with patchpoints [2].
Tweaking support for segmented stacks is quite easy—I added the support for it on x86 Mac. You simply adjust the prolog emission code in X86FrameLowering.cpp and likewise for the other targets.
[1]: http://llvm.org/docs/SegmentedStacks.html
[2]: http://llvm.org/docs/StackMaps.html http://llvm.org/docs/Statepoints.html
I'm not sure what you mean by "cgo's runtime", but go runtime is in go since 1.5. Some bits are still written in assembly but that's mainly for performance.
I don't think targeting every arch (and virtual archs like asm.js you mention [though there is a 3rd party GopherJS]) and OS out there was ever their purpose.
In case you're interested, there is llgo which has recently been merged into LLVM. It's still not practical however.
> But I guess Google just hates Apple that much?
I don't understand what this is supposed to mean since LLVM does not belong to Apple, and Google is a heavy user and contributer thereof.
Many bits in math and crypto have fast assembly variants, but the assembly code in the runtime and sync packages is not for performance, but because assembly is the only way to put data in the right registers, issue memory barriers, implement the right atomic operations and so on. Not performance.
It does not have platform-specific C code, as there is no more C code anymore. It sure does have platform-specific Go code and platform-specific assembly code though.
> from cgo's runtime so it can support the different architectures/platforms.
There is platform-specific code in the runtime, and there is platform-specific code elsewhere in the standard library, e.g. in the syscall package, but it's not from "cgo's runtime" (whatever that means). Go does not require cgo to interact with the target system, even on platforms where the interaction is done through shared libraries, not system calls, like Solaris and Windows.
I'm not aware of anything that even comes close to tooling out cross compilation so easily.
build.sh cross compiles one codebase for 57(+)[0] architectures. Even for a "native" build, the tooling is the same abstraction that builds for a completely foreign target, so it's continuously well-exercised.
darwin/arm, darwin/arm64, plan9/386, plan9/amd64 and solaris/amd64 don't compile for any of the 11 binaries (e.g. plan9 doesn't work due to logrus not having support for it).
This leaves me with the list at http://www.robustperception.io/prometheus-nightly-binaries/ That's still 20 architectures I can compile almost everything to automatically with an ansible+jenkins setup.
The logrus library we use for logging doesn't support Solaris.
https://github.com/Sirupsen/logrus/blob/master/terminal_sola...
Isn't the best course usually to build a solid foundation, then branch out?
That said, the counterpoint is that cross-compilation is not something you can easily bolt on later -- if you aren't doing, and testing it along the way, you will never be able to add it back in. It needs to be there from the beginning and stay a first-class support goal, otherwise mistakes, good intentions and laziness will all yield a product that just can't easily make the leap back to solid cross compilation with no hassle.
I'd also like to see Go on MIPS64, though probably less important than ARM.
It seems they accepted this sacrifice for some other unimportant feature*
*edit: Looks as though the feature they sacrificed performance for was writing Go in Go.
1) Perf for a product should be marked against some shipping benchmark and intermediate regressions don't matter as long as we come below the benchmark at ship time.
2) Perf improvements should be "locked in" and not regressed, unless strenuously justified.
I think the parent is of group (2), where it doesn't necessarily matter that it's better than the benchmark of "against other compilers" but only that Go regressed from their earlier perf metrics for what they believe to be inadequate justification.
I think both groups have valid points. I think (2) is especially good at preventing "perf creep" wherein you continually justify small perf regressions for new features, explaining it away that competitors are doing the same. Eventually, you get to the point where everyone's just slow, but they have no impetus to improve because they're still better than the competition.
If the developers take a specific principled stance against $really_useful_feature on account that it would introduce complexity and thus degrade performance, then compromise this same principle in order to accomplish $entirely_unnecessary_port, it put the obvious lie to their avowed principles.
I don't mind that go compiles more slowly. To Thomas's point, it's still really fast. What I mind is that a good portion of the go community has accepted the lack of generics on the basis of principles that they haphazardly violated in order to scratch the useless but fun intellectual itch of bootstrapping the compiler.
Now that they've done this, they have revealed themselves to be entirely dishonest about their complexity/performance principles, and now the one legitimate objection that has kept generics out of the language has nothing to stand on.
Again, I don't care about performance per se, I care about dishonesty and hypocrisy in the platform's objectives.
Absolutely nobody in the universe promised you generics in Go. You may never, ever get them; the longer the language thrives without them, the less urgent they seem.
> Generics are convenient but they come at a cost in complexity in the type system and run-time. We haven't yet found a design that gives value proportionate to the complexity, although we continue to think about it.
The word performance doesn't appear in that paragraph.
Writing Go in Go is a huge feature, not an unimportant feature.
Writing the runtime in Go means we write the runtime in a much safer language. We found and fixed lots of bugs just by rewriting the runtime. Everybody, every user benefits from this tremendously.
Writing the runtime in Go means you can have a precise garbage collector. Every user of Go benefits from this tremendously.
Writing the runtime in Go means the runtime implementation becomes in many respects simpler, because there are no longer complex C-Go interactions (the Plan 9-derived C compiler used a slightly different calling convention!). This again reduces bugs and makes debugging easier. Everybody wins.
Writing the runtime in Go means there's one less custom compiler to maintain (no more Plan 9-derived C compiler!). Less maintenance means more time available for more important things. Also this simplifies new ports. Again, everybody wins.
Writing the compiler in Go means you can have a compiler that's not written in a language that sucks for writing compilers. Because writing compilers is now easier and consumes less time, there's more time available for more important things. Everybody wins.
Writing the compiler in Go means you benefit from all the tooling built for Go. This includes source code transformation tools. Automatic refactoring is very easy to do in Go because you have Go parsers in the standard library. This is not merely hypothetical, it's something that's actually used. The end result is that changing the compiler is faster, again allowing more work to be done.
Writing the compiler in Go also means the compiler can eventually become an idiomatic Go program that can be understood by Go programmers instead of being a ken-idiomatic C program that can only be understood by a few people. More people can work on the compiler. More work is being done as we speak, including work that will make generated code much faster. Everybody wins.
Writing the compiler in Go means you don't depend on million of lines of C++ toolchains in order to exist. As a Go compiler writer this has simplified my life a lot. I can compile the Go distribution in under a minute. On my phone. Which is 5 years old. Last time I tried to compile LLVM it died with a compile error after 6 (six!) hours.
Writing the compiler in Go means I can have fun while writing compilers for fun, so you benefit because I write it. Everybody wins.
I've written three compiler targets by now, one in C and two in Go. I don't plan to stop because I like what I am doing, but I'm not sure I would continue to do it if it were still written in C. People benefit from this not only by having more toys to play with, but new ports mean more comprehensive testing and new bugs found. Different platforms exercise different bugs differently, bugs which are present on all platforms.
Also the actual work done on the ports stimulates further refactoring efforts that will make Go even easier to port in the future. Again everybody wins! And it's only because this "unimportant feature" that is not even visible to most users of Go, coupled with the powerful refactoring technology which makes this possible.
Does make "go test" complete faster, throws baby out with bathwater though.
Also, work is in progress to improve build times, although I don't know if that's an active focus, rather than just something that's being fixed as time permits. A few improvements have already gone back to the current development gate for a variety of cases.
I think it's important to remember that development can be done in parallel, just because you see work being done on an area that isn't important to you doesn't mean that there isn't work being done on the area that's important to you as well. Some of these changes take a long time.
Yes, and I think that is what they are doing. First build something stable. Then target a lot of platforms. Then optimize the language itself and also the build times.
Btw, in case it's not clear, Go has worked on ARM for the last 6 years, this is not a new thing.
It is the job of the compiler and runtime writer to ensure this model on whatever target platform. It doesn't matter if the target platform has a strong or weak memory model. Correct Go programs written for the Go memory model should work correctly on any target.
Of course that it's possible to write Go programs that violate the memory model but work on amd64, and not on some other platforms. But those are incorrect Go programs (and the race detector has a very good chance of detecting these problems). I will note that it's significantly harder to write these programs accidentally in Go though, than it is to write them in C.
That would be my hunch. I think it makes a lot of sense. Also, note that there is that project which exists : http://llvm.org/viewvc/llvm-project/llgo/
There can be a compile time hit, but you can use FastISel, which is not bad. The compile time hit is mitigated by the fact that you get a far, far more mature series of optimization passes. Since the Go compiler is itself written in Go, more optimizations would make the Go compiler itself faster, further reducing the performance impact.
> Flexibility over the backend ?
Writing to LLVM in no way restricts you to a backend. You simply need a language-level IR that you can retarget to different backends as you would like. This is what Swift, for example, is doing with SIL.
It still ignores toolchain flexibility. Especially given the specific needs of Go's managed runtime.
The other aspect is that the Go compiler is intended to be very fast. That's not an explicit design goal of llvm - using llvm would slow the compiler down although it would have other advantages.
In the meantime, Go still supports more hardware ISAs than LLVM. Apparently reinventing the wheel is much cheaper than adding wheels to the old car.
I realize that standard towers may be supplanted by their SoC brethren, but so what? Will we all be using iMac-like computers as desktops in the next 36-72 months? Will there even be any differentiation between "regular" desktops with discrete CPUs, memory, etc and the system-on-a-chip variant?
http://techcrunch.com/2015/11/17/google-and-asus-launch-the-...