=======================================================
RPython: Need for speed
=======================================================

aka: C and C# considered harmful
================================

:Authors: Antonio Cuni (DISI - University of Genoa), Maciej Fijalkowski (merlinux GmbH)
:Place: Europython 2007
:Date: 10 July 2007

What is RPython?
----------------

* RPython programs run unmodified on top of CPython

* Restricted subset

* Constructed from live objects - initialization
  is done in Python

* Python is the meta-programming language for RPython

Why invent new language?
------------------------

* Grown as a side effect of the PyPy project

* Python is too slow for some purposes

* Pyrex/C are ugly

* We want to play as it were Python as much as possible

Restricted, what do you mean?
-----------------------------

* Static type system

* Full type inference - you cannot mix incompatible types

* Polymorphic collections (i.e. list of ints)

* Java-like object model

* No dynamic stuff: switch class, add/remove methods, etc.

* No support for special methods (only __init__ and __del__ works)

* No introspection

Example of use case
-------------------

* Write code

* Test/debug on top of CPython

* Compile

* Graph viewer

Type inference
--------------

::

    def foo():
        return bar(1, 2) + bar(3, 4.5)

    def bar(a, b):
        return a+1

* a: Signed U Siged --> Signed

* b: Signed U Float --> Float

* result of bar: Signed + 1 --> Signed


Modules restrictions
--------------------

* Few modules work

* They have different, more static API (ie open())

* All modules might be used when initializing

RPython - why you wouldn't like to use it
-----------------------------------------

* RPython is quite new

* Wasn't meant to be a general purpose language from scratch

* Type inference usually gives bogus erros

* Compilation errors are far from perfect in general

RPython - on the other hand...
------------------------------

* it's about 60x faster than CPython

* no more pros...

Writing standalone C/.NET/JVM programs
--------------------------------------

::

    def target(*args):
        return entry_point, None
    
    def entry_point(argv):
        print 'Hello World!'
        return 0

* see also translator/goal/targetnopstandalone.py

* ./translate.py [--backend={c|cli|jvm}] targetXXX.py

Writing CPython extensions
---------------------------

* create a mixed module (there is _demo)

* compile it as CPython extension

* quite incomplete

The CLI backend
---------------

* Produces assemblies for .NET (both Mono and CLR)

* (Experimental) access to .NET libraries

* As fast as C#

* From 4 to 100 times faster than IronPython

CarbonPython
------------

* **Highly** experimental

* RPython --> .NET DLL

* DLLs can be imported by C#, IronPython, ...

* CarbonPython + IronPython = SteelPython :-)

* Explicit definition of multiple entrypoints

* Live demo

* http://codespeak.net/pypy/dist/pypy/doc/carbonpython.html


The JVM backend
---------------

* Produces classes for the JVM

* Shares a lot of code with the CLI backend

* Not as mature as gencli

* Only executables at the moment


The JS backend
--------------

* Can produce complex, standalone executables

* AJAX by method calls

* Can bind to ie. mochikit

* Demos are online at http://play1.codespeak.net

* Can be used with any possible web framework

Conclusions
-----------

* RPython has lot of rough edges

* ...but it might be useful in some contexts

* More convenient than C/C++/C#/Java

* Much faster than CPython/IronPython/Jython

* Stay tuned