Poole The Jekyll Butler

RPython experiment

There is this project called pypy which implements a python interpreter in using the RPython translation toolchain which compiles it down into C code with a JIT - A Just In Time tracing compiler that tries to identify hot loops and then watches the interpreter executing them so that it can try to specialize the traces into more efficient compiled versions. You can use the same tools to turn any bytecode interpreter into a JIT based interpreter (I’ve heard that the tools work much better on bytecode than an AST based interpreter).

You can get a lot of really good technicaly information about how this is done from the following two links The Architecture of Open Source Applications, Fast Enough VMs in Fast Enough Time. There’s a good tutorial for beginners on the main pypy blog here: Writing an Interpreter with PyPy and some more nice example code on bitbucket Kermit.

I had heard this ‘generating a jit compiler from an interpreter’ thing was based on the 3rd futamura projection but im a little skeptical about that, I think that a tracing JIT specializer is a slightly different thing. I am probably wrong or being too specific about terms but it’s not too important: pypy is about 6x faster than cpython - that’s what matters.

So I wanted to try this stuff out - I thought it would be really interesting to apply it to the UM virtual machine from ICFP 2006. My plain C implementation takes 20.8 seconds to run (when compiled with -O3, at first I did the comparison without -O3 since I forgot and the results were incredible.. you have to be careful when doing “science”), the RPython version takes 16.9 seconds. It is faster but not by too much. I think an RPython expert could probably make this a lot faster I don’t really know how to.

I got some further insight into why this isn’t giving me a good speedup. The toolchain here is not targeted at low level bytecode interpreters, but instead its for high level bytecode interpreters. That would mean something like python which you compile to bytecode and then interpret: In this case you will have a lot of common patterns which can be specialized well. You need to set up the JIT stuff so that it’s able to recognize these patterns and take advantage of them. Further the dynamicness of the UM vm makes things very hard: You can edit the instructions as its running as well as build new arrays and jump into them.

Random extra links: