Ben.Young@risk.sungard.com writes:

Hi All,

As you probably know by now, I enjoy following the progress of the PyPy project, and generally I can understand roughly what is going on, but I have to admit I am stumpt by the JIT.

I'm pretty lost too :)

Is there a document that describes what is going on anywhere? What's the goal of the current bit of work? I can't actually work out if the current JIT for the toy language is actaully jitt'ed, or if you are currently just working on the data structures that will produce the JIT? I.e will there need to be a machine code backend for the JIT or does it somehow magically produce fast code without a compiler, using some other kind of dispatch loop?

This bit I can answer to some extent. Currently the JIT code produces L2 graphs, which are the low level graphs that can be translated to C. This is fine for the moment, but not in the long run, for two reasons: one is that RPython programs cannot manipulate L2 graphs (and eventually the JIT needs to be RPython so that it runs at a reasonable speed) and the other is that we don't want our JIT to have to compile it's generated code to C (or use llinterp...). So, we need something else. One option is indeed to generate machine code, but first we'll probably target another representation called L3 graphs (low low level graphs, if you like) which are simple enough to be manipulated and executed by RPython programs. The l3 interpreter/evaluator is partially written at this point, although some deep issues remain, particularly the interaction with garbage collection. Cheers, mwh -- Hiro dicks about with his computer, naturally. Being stranded on a life raft in the Pacific is a perfect venue for a hacker. -- Snow Crash, Neal Stephenson

Ben.Young@risk.sungard.com writes:

Hi All,

As you probably know by now, I enjoy following the progress of the PyPy project, and generally I can understand roughly what is going on, but I have to admit I am stumpt by the JIT.

I'm pretty lost too :)

Is there a document that describes what is going on anywhere? What's the goal of the current bit of work? I can't actually work out if the current JIT for the toy language is actaully jitt'ed, or if you are currently just working on the data structures that will produce the JIT? I.e will there need to be a machine code backend for the JIT or does it somehow magically produce fast code without a compiler, using some other kind of dispatch loop?

This bit I can answer to some extent. Currently the JIT code produces L2 graphs, which are the low level graphs that can be translated to C. This is fine for the moment, but not in the long run, for two reasons: one is that RPython programs cannot manipulate L2 graphs (and eventually the JIT needs to be RPython so that it runs at a reasonable speed) and the other is that we don't want our JIT to have to compile it's generated code to C (or use llinterp...). So, we need something else. One option is indeed to generate machine code, but first we'll probably target another representation called L3 graphs (low low level graphs, if you like) which are simple enough to be manipulated and executed by RPython programs. The l3 interpreter/evaluator is partially written at this point, although some deep issues remain, particularly the interaction with garbage collection. Cheers, mwh -- Hiro dicks about with his computer, naturally. Being stranded on a life raft in the Pacific is a perfect venue for a hacker. -- Snow Crash, Neal Stephenson