PyPy talk¶

by Alex Gaynor

Two things go faster than C¶

psyco was JIT python
- Managing it was hard
- hardcoded for 32 bit CPUs and we are on 64 bit
- Any changes to core Python killed pysco
Years ago created a Python interpreter inside of Python
- 2000x slower than cpython
- ran on restricted Python (r-python)
- Wrote a great compiler: now 10x slower than cpython
- Added better garbage collection: now 4x slower than cpython
New JIT for Python
- Writing a JIT for Python sucks
- Writing a generator for making JITs for any language is easier
- Alex Statement: “PyPy is the only project I know of that uses SVN branches. That’s the most impressive part”
- Doing it this way made it faster? How? Wizard Magic?
Now PyPY is going fast:
- Crazy that it runs so much faster than cpython
- Hard to believe
- Python using a JIT generator to create a JIT library?
- Faster than cpython
- http://speed.pypy.org/

jitviewer
- Finding slow spots in existing code
- Looks at Python, byte code, assembler, et al

Faster than cpython
They now metric it’s speed against C, not Python
Compatibilities
- Much work with third-party integration
- lots of people are using Python instead of C extensions
- PEP in place so that new stdlib stuff has to be pure python
Quora is much more snappy

They can branch the Python 2.7 PyPy stuff to Python 3
- Make the Python 3 version work
- And since the JIT generator makes the code, both versions will just work
Can do a GIL version for single CPU or non-GIL for multicore
- JIT generated so…
- Both versions just work!!!