Python patch to speed things up to 20% | Hacker News Reader

Python patch to speed things up to 20%

bugs.python.org

67 pointsvaksel22 comments

I saw this technique, namely using gcc's labels-as-values extension to efficiently implement threaded code, in gForth. The developers of gForth originally chose gcc back in 1992 because it had such a neat feature that allowed to implement threaded code efficiently. Unfortunately for gForth, somewhere along the way gcc developers changed something and gForth took massive perfomance hit (2 or 3 times slowdown) gForth developers tried to negotiate with gcc team, but they didn't seem to be interested to do additional work seemingly needed only for one application (gForth). The gForth developers couldn't afford to fork gcc either (gcc is too big; too much work)

I guess the moral of the story should be: beware of such speedups that come from nonstandard technology you cannot control.

The whole story is from my memory, as I cannot find the source right now.

seems like a candidate for inline assembly.

sure, but what if you might not want to use inline assembly i.e. for easier portability across platforms. Forths before used it all the way, but gForth didn't want to.

While I really love Python as the language I really hate its primitive implementation. 20% is nothing compared to what would be gained if Python ran on a virtual machine that was JIT'ed.

Java for example got a 10 to 20 times performance gain from moving to HotSpot (their JIT'ed VM). And similar gain's have been seen in Chrome's V8 JavaScript VM or Strongtalk.

Anyway, it's kind of sad that such a great language has such a braindead implementation :-(

And unfortunately, the same thing can be said about Ruby as well.

Python's implementation is focused on code readability. One of Guido's initial goals was that no matter how insane the input, the interpreter must never crash. His emphasis on simple and clear code (and an intuitive bytecode) gets in the way of performance.

The Ruby interpreter is quite terrible, but a lot of people are trying to make it better. Give it a couple of years.

Also, you imply that both implementations are really naive. They're not. With languages such as Python and Ruby there is much less to gain with JIT compilation than with Java. Suppose you have this Python program:

    def trivial(a, b):
        return a + b

What can the JIT optimize? The function call? Not always, because the function can be redefined during runtime (unlike with Java and .NET). Can the function be inlined? No. Can the function call frame be stack-allocated? Well, maybe, at a significant complexity cost: most scopes in Python are essentially closures so you have to treat carefully there too (upward funarg problems). Obviously the "+" function call can't be optimized much, because you don't know the types of "a" and "b". And even if you did it wouldn't help much.

Not done yet!

Even IF you optimize everything using the NBOCK (Non Braindead Omniscient Compiler Kit). Now you know that you're dealing with integers, and you know where the function is defined, and you know the function's definition is not going to change, and you know all function arguments can be declared locally and you know only integers will be passed to and returned from the function. Can you now transform it into a few MOV instructions? Alas, no.

Maybe the integers will overflow. And if they overflow you either want to throw an exception or allocate additional space and transform the integer to a heap-allocated large integer. (And you have to do all the necessary cache invalidation.)

So even if you are omniscient there is no low-hanging fruit. In the end even the logic of adding two integers is sufficiently complex that it warrants its own C function and inclining becomes only marginally useful.

When you take mundane issues such as maintainability, flexibility, and portability into account there are very few reasons left for building a JIT compiler for Python. Note that all these optimizations are viable (to an extent) with languages such as Haskell, C# and Java.

The PyPy folks are working on a JIT, apparently with great speedups. They've been detailing their work on their blog: http://morepypy.blogspot.com/

Also, are JavaScript and Python really that dissimilar? Everyone's working on tracing JITs for JavaScript now, with great results.

My impression is that Python's interpreter is pretty good, so it wouldn't benefit as much from JITing.

That being said, it would certainly help, and the same library used in TraceMonkey (nanojit) could probably be used in Python.

Take a look at the PyPy FAQ. They're using annotations, type inference assumptions, and so on. PyPy seems to be focused on a Jit-able subset of Python. It's not a project that can one day be transparently included in the new Python release.

JavaScript and Python are not that dissimilar. So we see similar results with JavaScript. People want performance, so a lot of projects are started where people attempt to JIT JavaScript - but in the 14 or so years of JS's existence JavaScript is still several orders of magnitude slower than less dynamic languages. Java and C# have never been as slow as JavaScript is today. Will JIT-ing JavaScript help? Sure. But -great- results? I wouldn't go that far.

No, PyPy is a full Python implementation that uses program analysis to identify and JIT the JITable parts of the code.

The FAQ says that PyPy is a drop-in replacement for CPython unless you depend on a CPython extension module. http://codespeak.net/pypy/dist/pypy/doc/faq.html#is-pypy-a-d...

The drop in replacement is slower than CPython. Likely to change in the future, but still.

Don't forget psyco's JIT which gives some impressive performance gains for a lot of different code. We use it while embedding Python in Counter-Strike Source and it performs admirably.

Someone needs to upgrade pyscho for 64 bit IIRC

You seem to imply that they're viable because those languages aren't as dynamic as Python. Yet Smalltalk and Lisp (edit: and Javascript) are at least this dynamic, and have implementations that go far beyond this. In fact Java's (edit: and V8's) JIT technology came from the Smalltalk world.

I've heard this claim several times, but there seems to be very little evidence to support it.

The fast native code Lisp compilers such as SBCL and Lispworks sacrifice correctness for speed. You can fully expect the VM to crash when you give incorrect hints to the compiler. This is obviously unacceptable for languages such as Ruby and Python. I don't know of any serious Lisp JIT compilers.

Smalltalk has never had decent performance as far as I know - not even with a JIT backend. Not compared to Java/C# with JIT compilation at least.

So yes, my claim is that languages such as Python and Ruby are inherently slower than languages such as C# and Java. The less information the compiler has to work with, the fewer optimizations can be made.

"Smalltalk has never had decent performance as far as I know - not even with a JIT backend. Not compared to Java/C# with JIT compilation at least."

This is largely because the team that did a JIT enabled Smalltalk dropped it and started to work on a JIT enabled VM for Java. This team got bought by Sun in the 90's and their Java VM was HotSpot, which is used in Java today. As somebody else notes, a lot of the optimizations used in JVM come from dynamic languages such as Smalltalk and Self...

Smalltalk has never had decent performance as far as I know

I don't know what you mean by "decent". Strongtalk was widely regarded as highly performant in its day, certainly far beyond what had been thought possible for a Smalltalk. Its techniques are now being applied to produce impressive gains in contemporary dynamic languages (Maglev, V8). Given that Python and Ruby are in loosely the same class of dynamicness as Smalltalk and JS, I think that's plenty of evidence. Point taken about native Lisp compilation, though.

Thanks jd - I found out a few months ago that Python JITs weren't as fast as Java JITs, but had never quite grasped that there was solid reason why.

Why is Java faster than Python? Because Sun has invested m(b)illions in optimization of Java. They for example have around 50 people+ working full-time on JVM right now.

You sound pretty knowledgeable on this stuff. Maybe use your smarts to offer solutions too? :-)

The best explanation for how the optimization works is actually in the patch itself: http://bugs.python.org/file12524/threadedceval5.patch.

Are these the 20% we lost with Python 3?

Not the "same" 20%, no: the same optimization could be applied to Python 2.x, presumably yielding a comparable relative speedup.

Reply on news.ycombinator.com