Comment by kaib | Hacker News Reader

Comment by kaib | Hacker News Reader

Parent Full thread

> So by all means, explain to me why the non-gcc implementation (6g, 8g) for Google Go not being LLVM somehow isn't classic NIH.

Disclaimer: I know a bit about the background to the choice, I wasn't involved but I've seen the consequences having worked with the 5g arm compiler.

The code generation part of 5g/5l is on the order of 10k lines of C, the whole 8g/6g/5g compiler suite is ~115k. I don't know how to break down LLVM but the whole project looks like ~270k lines of C++ so the code generation is probably more than 10k.

Now imagine you are tasked with writing a compiler for an experimental language where one of the key features is blazingly fast compilation. The plan9 compilers are fast, you wrote them yourself so you know their strengths and weaknesses. The plan9 compilers are also reasonably small and written in C which you know well.

You know nothing about LLVM. It could be good but it would take some time to find out. You also don't know too much C++ and it would take some time to get up to speed.

I would be very interested in understanding why this is NIH as opposed to reasonable management of risk?

>where one of the key features is blazingly fast compilation

Oh come on. That's one of the banal [excuses|priorities] I've ever heard. I was aware of it, I just couldn't believe someone would invoke it with a straight face.

>I would be very interested in understanding why this is NIH as opposed to reasonable management of risk?

When you ignore the prime project that has the most labor, most corporate support and funding, the most data, and the most documentation as a competitor to gcc, and it's BSD licensed, and it's perfectly available for compilation/testing to see if it's appropriate to your needs...

...that's classic NIH.

You've outlined basic FUD fueled by lack of knowledge of the alternatives. Management of risk? The code is BSD licensed, spare me! If you want to take the code and run off with it in the sunset, you're perfectly able to do so. Instead of at least re-using/borrowing code from LLVM, they decided to embark on a rather laborious and difficult aspect of language development themselves with zero input from other projects.

The cognitive dissonance required to believe that when you reach that kind of decision-chain, that what you're doing is anything other than pride/fear/NIH must be immense.

I understand that Google has a fetish for the academic/open source elite in their hiring practices, but that doesn't get them a wholesale license to chuck common sense out the window and disregard the work of all others.

ARM compatibility, gee, that does make depending on other compiler infrastructures tricky since you can't trust anyone to support anything other than x8---- wait a tick.

"An easily retargettable code generator, which currently supports X86, X86-64, PowerPC, PowerPC-64, ARM, Thumb, SPARC, Alpha, CellSPU, PIC16 MIPS, MSP430, SystemZ, and XCore."

--- http://llvm.org/Features.html

I'm not buyin' what you're selling. Was I involved in the production of Go? No, but I've been tracking it since it was publicized and unless there's some detail/issue/licensing matter/temporal rift involved that I'm made aware of, I'm just not going to swallow the notion that it was anything other than NIH muddled by incomplete information.

> If you want to take the code and run off with it in the sunset, you're perfectly able to do so. Instead of at least re-using/borrowing code from LLVM, they decided to embark on a rather laborious and difficult aspect of language development themselves with zero input from other projects.

I think there are two separate concerns, both which I'm disagreeing with you on. First, my understanding is that when ken wrote the initial compilers he wanted to focus on the frontend and re-use a backend. So he picked the plan9 compilers which he knew well (they predate LLVM by 10 years and he mostly wrote them). There was very little original work going into the code generation.

Second, I think you are grossly underestimating the initial cost of integrating with an open source project unkown to someone. The first Go compiler frontend was written in a few weeks or months. I'm guessing that having to pick up a new unknown backend (LLVM) and language (C++) would have set the project back enough to make it miss it's window of opportunity.

It seems to me that your argument is that an incorrect choice was made in what to re-use. That seems like a very different thing than NIH.

> ARM compatibility

ARM compatibility had nothing to do with the decision to re-use the plan9 compilers. I'm a compiler neophyte and when I wanted to add ARM support to Go it was an easy choice to go with the existing framework (there was a solid plan9 C compiler for ARM). If I had tried to port the system to LLVM at the same time as adding ARM support my skills would probably have failed me.

> fast compilation

One of the distinctive features of the optimizers ken wrote for plan9 C, which are used in Go, is that they are much simpler and faster than normal optimizers but produce less optimal code. The GCC Go compilers are slower and slightly less comprehensive but produce code of higher quality. LLVM would undoubtedly produce better code as well. I'm happy to go into more technical detail if you want to, it would help if you'd let me know how much practical compiler writing experience you have?

When work on Go started in late 2007, LLVM wasn't what it is today. I think that if we were doing Go again, starting today, it would make sense to use LLVM, but if you look at the kinds of things that were being done in LLVM in 2007 (llvm.org has the release notes linked), it wasn't a viable option (yet).

I for one would like to see a Go front end for LLVM. In fact, the front end used by gccgo is a separate project from gcc and BSD - not GPL - licensed (http://code.google.com/p/gofrontend) specifically to make that easier and so that gccgo and an eventual go-for-LLVM can share code.

Reply on news.ycombinator.com