Re: [dev] JIT & VM discussion

From: Kamil Cholewiński <harry666t_AT_gmail.com>
Date: Sat, 18 Jun 2016 19:39:34 +0200

On Sat, 18 Jun 2016, Connor Lane Smith <cls_AT_lubutu.com> wrote:
> Hi all,
>
> I was wondering if others had an opinion on JIT. Suppose we don't need
> anything fancy like adaptive optimisation, but just wanted to compile
> a program at runtime. One possibility might be to generate C code and
> store that in a file and then call the C compiler and then execute the
> resulting binary... But that seems a bit unpleasant, prone to
> compilation failure, and not particularly lightweight either.
>
> One solution could be simply to produce assembly code, but then that
> is tied to a specific architecture, which is unfortunate. There's also
> LLVM, but that is a very big and complicated dependency, which
> shouldn't really be necessary if we're just jitting something quickly
> and don't mind it being a little unoptimised for the sake of
> simplicity and speed of compilation. We just want to portably generate
> machine code and then run it.
>
> An ideal might be something like an abstract instruction set together
> with a JIT for the abstract machine. To be honest a JIT might not even
> be necessary so long as it is has very little interpretation overhead,
> the instruction set is general purpose and fixed, and it plays well
> with the C memory model.
>
> Does anyone have any ideas?
>
> Thanks,
> cls

Hi Connor,

Creating a simple and general-purpose VM shouldn't be hard! It used to
be my favourite exercise for learning a new programming language.

Probably much more difficult to get real-world performance; I wouldn't
be surprised if the initial efforts resulted in a 1000x-slower-than-C
execution speed for typical programs.

With lots of test cases, tuning, benchmarks, and generally a lot of hard
work, I can imagine you could bring it to the 10-100x-slower[1] class.

[1]: https://benchmarksgame.alioth.debian.org/u64q/compare.php?lang=python3&lang2=gcc

Of course this doesn't matter that much if your purpose is mostly
scripting behavior (games), or IO-bound stuff (as in waiting for
database - things like Snabb[2] actually do need some real power).
Having good C interop via FFI can save you in many cases.

[2]: https://github.com/snabbco/snabb

Yes, JITing is inherently architecture-specific, but the bytecode can be
designed with trade-offs between interpretation and compilation speed.
These days supporting x86-64, ARM and MIPS probably covers >99% of the
devices you'll ever encounter in the wild; the rest can run a bit slower
until someone is motivated enough to write a JIT backend.

I've never had a close look at any of the Big Name VMs, as most of that
code must suck horribly. Some real-world VMs&JITs however remained
relatively simple - I think there might be a lot to learn from Dis[3]
and LuaJIT[4].

[3]: http://doc.cat-v.org/inferno/4th_edition/dis_VM_design
[4]: http://luajit.org/

If you have some concrete applications in mind, please do share. I'd
gladly give a shot at prototyping something in this area.

<3,K.
Received on Sat Jun 18 2016 - 19:39:34 CEST

This archive was generated by hypermail 2.3.0 : Sat Jun 18 2016 - 19:48:11 CEST