[eternalban]

I greped for "core", "memory manager", "thread", "threading", and "cache", in the pdf [1]. Am I missing something? I'll probably get flak for this, but this programmer looks for a modern systems language that directly addresses these concerns.

"The fundamental problems of a systems-programming task or environment are hard limits on computational resources and a lack of safety protections. Systems programs have to deal with hardware-imposed limitations on their CPU time, registers, memory space, storage space on I/O devices, and I/O bandwidth. They also often have to deal with a lack of the safety features normally given to most programming environments: garbage-collection of memory, validation of memory accesses, synchronization primitives, abstracted access to I/O devices, and transparent multitasking. In fact, the point of systems programming is usually to create such abstractions and program such protections."

Let's take Go, as an example. (Or D). Per above definition, neither is a "systems programming" language. I know for a fact [2] that Go team would disagree.

So what is the accepted definition of a "modern" systems programming language?

[1]: http://code.google.com/p/decac/downloads/detail?name=Deca%20...

[2]: http://golang.org/doc/go_faq.html#What_is_the_purpose_of_the...

[smosher]

I've always taken "systems language" to mean something you might write an OS in while the Go team takes it to mean something you'd replace Java with. Whichever way you want to look at it one thing the Go team can't say is that writing an OS is a "non-systems" problem.

Ideally these would be disjoint sets (in my opinion) but consider this: ASM -> C -> C++ -> Java -> ... looks like a continuum if you're writing any kind of Java code, but there's a sharp discontinuity in there if you're writing (say) a boot loader.

[eternalban]

> Whichever way you want to look at it one thing the Go team can't say is that writing an OS is a "non-systems" problem.

I do agree with that. (Perhaps it would be helpful to apply the high/low qualifier to systems languages?)

> Ideally these would be disjoint sets (in my opinion) but consider this: ASM -> C -> C++ -> Java -> ... looks like a continuum if you're writing any kind of Java code, but there's a sharp discontinuity in there if you're writing (say) a boot loader.

Food for thought. I am not arguing for needless complexity (in a language) as far as the boot up phase is concerned. As far as I understand it, the loader has its critical, but short lived, life-cycle role to play and then it is out of the way. Arguably, it is distinct from the OS. Can we not continue to write them in ASM/C and load operating systems that are written in a more "modern" language?

[smosher]

Personally I'd sooner map the somewhat fuzzier concept of "applications language" onto that supposed continuum instead. It would put Go in a clearer category, since it seems to compete with Java and Python better than it does with C.

I still don't imagine an OS being written in a non-systems application language, since you've got drivers and performance critical subsystems to write and things like GC to worry about from the apps-only languages.

[pjmlp]

Have a look at Spin or Native Oberon, just to name two operating systems written in GC enabled languages.

Before C existed, there were already operating systems written in PL/I and ALGOL, which provide better type safety and memory management as C or C++.

Unix and C success has regressed what meant to be a proper systems programming language.

There is even a paper from Adele Goldberg about this phenomena, unfortunately I cannot recall the title now.

[wbhart]

One of the examples in the repository is an implementation of malloc.

I tend to think of Go as an attempt to redefine what a systems programming language is (hence the "modern").

I don't think that there can be an established definition of what a modern systems language is, otherwise it would no longer be termed modern. Almost by definition theorists will differ on what the essential characteristics must be.

[eternalban]

> One of the examples in the repository is an implementation of malloc.

Right. And (imho) malloc is quite possibly an "old fashioned" way of looking at things:

   function malloc(num_bytes: nat): @byte

A reference to a byte block obtained from specifying the number of bytes! I would like to see the modern memory manger be type aware, have a very rich memory model, and allow for the propagation of application level semantics to OS (so they can cooperate in dealing with the current issue: memory access latencies and concurrency).

> I don't think that there can be an established definition of [a] modern systems language

Key word is "modern" so here is an attempt to box that a bit: There are 2 possible dimensions to the notion of modern:

1 - it addresses new hardware realities

2 - it enables "progress" in dealing with matter rendered difficult by existing paradigms of system programming.

The realities of a modern system surely include: multi-core SMP; strong likelihood to be employed as distributed nodes; virtualization. So "modern" platforms and deployment patterns are not addressed.

Second, Deca, I gather, would delegate dealing with all this to some higher level (user level) framework built on top of the language. I don't see any progress in that critical front, either.

[rdtsc]

> I would like to see the modern memory manger be type aware

Because these are system languages, that would happen when hardware is type aware.

[nponeccop]

It is not necessary. With software replacement for MMU possible (see Singularity OS), hardly any hardware safety support is a requirement nowadays. TAL is the way to go - AFAIK it is possible to make plain x86 assembly dependently typable by attaching proof witnesses to EXE.

[eli_gottlieb]

Deca does deal with SMP and concurrency in one way: module-level variables are tagged thread-local in the bitcode by default, at least if they're mutable.

Once the "newsig" branch (it's a rewrite from scratch after the original code had troubles with multi-module compilation) with the new signature system (region, effect, type, mutability) can compile some example code again, I'm merging it back into the main branch.