Ruby+OMR JIT Compiler: What’s Next?(developer.ibm.com) |
Ruby+OMR JIT Compiler: What’s Next?(developer.ibm.com) |
Actually unladen_swallow and cperl are doing this. Compile the whole runtime to a lib<rt>.bc (trivial make rule), load this bitcode file (a single call), and add the expensive optimizations, esp. the inliner and IPO in llvm. Just the type-checks need to be done in the jit. This inliner does a lot of good magic esp. for the small RT functions. >10x faster.
The compiler is expensive though (i.e. very slow), and LLVM changed it's jit API 3 times already. <=3.4 jit, then mcjit and >= 3.6 ocrjit, all 3 of them still having major quirks.
With mcjit you cannot selectively add jit code to a module. So every single body needs to be a new module. So module != package/class/namespace. So the jitcache is a bit complicated. With the latest ocrjit you got problems in finding the symbols in the bc. The C API doesn't support the name resolver at all, and it's a major quirks with C++. LLVM's C API is really behind, but at least they don't change it that often as the C++ API.
unladen_swallow has a huge JIT overhead library, which nobody really needs.
For the cperl jit I just used 4 days so far and doesn't even link yet. I used the C API, not the better C++ API. No C++ for cperl. But it's very simple. https://github.com/perl11/cperl/blob/feature/gh220-llvmjit/j...
Practically usable are only LLVM and dynasm.
Now the orm guys did something special I appreciate a lot. Basically they re-invented parrot, with a proper, fast object system with compile-time class extensions in C++, plus still supporting run-time inheritance via vtables. And on top of their existing stable jit engine for java, they started targeting ruby and python (unofficially).
Much better than parrot (which I'm maintaining on life-support), and probably also better than rpython. We will see how this will turn out, as they already have truffle, which don't like that much. orm looks much better to me.
Both are not working well. unladen swallow, because of incredible bloat and overhead. cperl, because it's not done yet. Just 4 days so far :) And supporting the 3 LLVM jits is awkward. I'll better decide on only one and add the llvm version as fixed requirement. MCJIT sounds best so far, when storing the jit'ed bodies for all subs in a module in extra files does not kill all the performance. v8 went without that cache, java with. With a jit cache you can easily pre-compile/warmup your stdlib and do more expensive optimizations beforehand. LLVM is really slow.
And no, just because Ruby is highly dynamic - that's not an excuse for it being slow.
To quote Mike:
" tl;dr: The reason why X is slow, is because X's implementation is slow, unoptimized or untuned. Language design just influences how hard it is to make up for it. There are no excuses. "
[1] https://www.reddit.com/r/programming/comments/19gv4c/why_pyt...
I'm working on techniques to optimise through the core library in Ruby, and it's requiring novel research techniques that clearly weren't needed in LuaJIT, as it was done without them, so there must be something extra in Ruby.
So it can be done.
And to quote Mike Pall again
"" tl;dr: The reason why X is slow, is because X's implementation is slow, unoptimized or untuned. Language design just influences how hard it is to make up for it. There are no excuses. "
Knowing Ruby only superficially I fail to see why the former three managed to have AOT and JIT compilers, while Ruby still requires novel research techniques as you say.
Thinking of Smalltalk, become: is probably one of the best ways to kill anything a JIT knows about an object.
As much as I reach for more performant languages, I still need to work with Ruby on a daily basis. I really appreciate the ongoing work done by you and others to keep the ecosystem alive and prosperous.
https://pdfs.semanticscholar.org/d1fc/e50f5476088671adc3910d...
Can you create new native languages using Eclipse OMR ?
It's been done before with the technology underlying the OMR compiler component, however there is some work that would need to be done to support this.
In principle though, there's no reason it couldn't be done.
I was thinking, OMR providing the native platform semantics that will enable multiple language syntaxes.
Ruby+OMR is the proof-of-concept project that aims to show that OMR is indeed language-agnostic and VM-agnostic and can be used as drop-in parts not only for newly-designed VMs but also easily integrated into existing VMs. The Java 8 version of IBM's J9 JDK is already built on top of OMR, but you could consider that cheating, since that's the very VM OMR came out of; Ruby+OMR shows that you can also do that with languages other than Java and VMs that have a different design than J9.
I believe there is also a Python+OMR project which does similar things with the CPython VM, but that project is not (yet) public, AFAIK.
[0] https://github.com/Valloric/YouCompleteMe#semantic-completio...
No, seriously, if your heart beats for Ruby, you will fall in love with Elixir. Give it a try and you won't look back.
It'd be closer to say: > Ruby 3x3 is here already. It's called Crystal ;)
Definitely check out Elixir and Crystal though if you need more performance.
1) It borders on proselytizing and it's annoying to see people coming over to preach the good word on every other Ruby article. If your language of choice has merits, it doesn't need hype men, people will use it naturally.
2) Elixir is not Ruby, stop trying to sell it as such. They're two completely different languages. That one has heavily influenced the other's syntax choices is merely a sign that the language author liked what he found in Ruby before creating Elixir.
I like Ruby, I like Elixir, but the latter for basically none of the same reasons as the former.
While it doesn't invalidate the reasoning behind the 'Pitfalls' paper, we think that we're addressing the problem in a fairly different manner: Instead of providing a fully featured JIT compiler that you need to wedge the language into, we've instead chosen to create a more modular system that lets the JIT compiler be more customized to the target language and VM.
I strongly disagree (and always have) with your statement that repurposing a statically typed JIT is limiting. The JIT projects you mention get lots of their performance from profile data which essentially presents type guesses so the compiler can speculatively optimize and recover when that info is wrong. Guess what? That's what statically typed Java JIT compilers do too, and the OMR compiler is no different, though we haven't sunk all that capability to OMR from Testarossa just yet (it's coming!).
I would be a lot happier with you repeating, or at least not propagating specifically the opposite conclusion, this line from your own paper:
On page 16, under Conclusion: "Our point, however, is not to argue against the repurposing of JIT compilers, but to define guiding principles and to promote techniques to construct an effective RJIT compiler."
Since this article was written, the entire Testarossa code base has been substantially refactored and parts of it rewritten (by the production Testarossa team) and streamlined to make it more amenable to dynamic optimization of other languages. We haven't finished that process (there's still more stuff we need to pull into the OMR compiler technology), but I really think your provocative statement above promotes a very broad conclusion from a research project that unfairly tries to close the door on what I still believe is a very reasonable technology evolution path. Especially since the article's conclusion is based on one research project for one language and one runtime where IMO the most reasonable conclusion is that just adding a JIT to an existing runtime is not enough to compete with a project that rewrote the entire runtime. That doesn't mean existing runtimes have nothing but bespoke paths forward or that reusable JIT technology cannot play a substantial part in that runtime's future evolution.
Finally, note that the OMR technology is being used (commits pulled in on an hourly basis) to build the production J9 JVM, which is also slated to be open sourced this year. There are also projects to use OMR for Ruby (see original post :) ), Lua, Rosie, Smalltalk (SOM++), Javascript, Python (currently inactive), and a runtime for workshops called Base9. I can provide links for the ones that have been open sourced already if people can't find them on github.
For those kind enough to have read to the end of this response (thank you!), please know that we have proposed several GSoC projects [1] to port to other languages and we're pretty open minded and positive about the potential and possibilities of reusing the Eclipse OMR compiler and other runtime components. Feel free to suggest alternative projects if you have a particular area of interest since we have lots of interested mentors available!
[1] https://wiki.eclipse.org/Google_Summer_of_Code_2017_Ideas#Ec...
The problem is turning Ruby into a low level program in the first place.
I use a Ruby library called psd.rb as an example. It's for handling Photoshop files. It represents pixels as a hash (map) of r, g and b values, and it parameterises the kind of filter in your image by making reflective method calls with dynamically created strings instead of using a function object of some kind.
So to compile that to tight machine code that looks like something a C compiler would output, you are going to have to completely optimise away the pixel hash and the reflective method call.
You could probably write that kind of code in the languages you mentioned, but people usually don't. In Ruby, that's how a lot of the code is written, starting with the standard library itself.
Maybe I could restate it by saying that optimising typical idiomatic Ruby code is more difficult than optimising typical idiomatic Smalltalk, Lisp, Lua etc.
More here: http://chrisseaton.com/rubytruffle/pushing-pixels/. Look half-way down at 'Acid Test'. My implementation of Ruby can compile that whole program to a constant.
I remembered you mentioned that example in some Graal talks I've watched.
It also needs a special mcjit resolver: https://github.com/cloudera/Impala/blob/cdh5-trunk/be/src/co...