Shed Skin Documentation

Version 0.9.1, January 15 2012, Mark Dufour and James Coughlan

• Shed Skin Documentation
• Introduction
• Typing Restrictions
• Python Subset Restrictions
• Library Limitations
• Installation
• Compiling a Stand-Alone Program
• Generating an Extension Module
• Distributing Binaries
• Parallel Processing
• Calling C/C++ Code
• Command-line Options
• Performance Tips and Tricks
• How to help out in Development

Introduction

Shed Skin is an experimental Python-to-C++ compiler designed to speed up the execution of computation-intensive Python programs. It converts programs written in a restricted subset of Python to C++. The C++ code can be compiled to executable code, which can be run either as a standalone program or as an extension module easily imported and used in a regular Python program.

Shed Skin uses type inference techniques to determine the implicit types used in a Python program, in order to generate the explicit type declarations needed in a C++ version. Because C++ is statically typed, Shed Skin requires Python code to be written such that all variables are (implicitly!) statically typed.

Besides the typing and subset restrictions, supported programs cannot freely use the Python standard library, although 25 common modules are supported, such as random and re (see Library Limitations).

Additionally, the type inference techniques employed by Shed Skin currently do not scale very well beyond several thousand lines of code (the largest compiled program is about 4,000 lines (sloccount)). In all, this means that Shed Skin is currently mostly useful to compile smallish programs and extension modules, that do not make extensive use of dynamic Python features or the standard or external libraries. See here for a collection of more than 60 example programs.

Because Shed Skin is still in an early stage of development, it can also improve a lot. At the moment, you will probably run into some bugs when using it. Please report these, so they can be fixed!

At the moment, Shed Skin is compatible with Python versions 2.4 to 2.7, behaves like 2.6, and should work on Windows and most UNIX platforms, such as GNU/Linux and OSX. On UNIX platforms, GCC version 4.2 or higher is required to compile the resulting C++ code.

Typing Restrictions

Shed Skin translates pure, but implicitly statically typed, Python programs into C++. The static typing restriction means that variables can only ever have a single, static type. So, for example,

a = 1
a = '1' # bad

is not allowed. However, as in C++, types can be abstract, so that for example,

a = A()
a = B() # good

where A and B have a common base class, is allowed.

The typing restriction also means that the elements of some collection (list, set, etc.) cannot have different types (because their subtype must also be static). Thus:

a = ['apple', 'b', 'c'] # good
b = (1, 2, 3) # good
c = [[10.3, -2.0], [1.5, 2.3], []] # good

is allowed, but

d = [1, 2.5, 'abc'] # bad
e = [3, [1, 2]] # bad
f = (0, 'abc', [1, 2, 3]) # bad

is not allowed. Dictionary keys and values may be of different types:

g = {'a': 1, 'b': 2, 'c': 3} # good
h = {'a': 1, 'b': 'hello', 'c': [1, 2, 3]} # bad

In the current version of Shed Skin, mixed types are also permitted in tuples of length two:

a = (1, [1]) # good

In the future, mixed tuples up to a certain length will probably be allowed.

None may only be mixed with non-scalar types (i.e., not with int, float, bool or complex):

l = [1]
l = None # good

m = 1
m = None # bad

def fun(x = None): # bad: use a special value for x here, e.g. x = -1
    pass
fun(1)

Integers and floats can often be mixed, but it is better to avoid this where possible, as it may confuse Shed Skin:

a = [1.0]
a = [1] # wrong - use a float here, too

Python Subset Restrictions

Shed Skin will only ever support a subset of all Python features. The following common features are currently not supported:

• eval, getattr, hasattr, isinstance, meta classes, anything really dynamic
• arbitrary-size arithmetic (integers become 32-bit by default on most architectures, see Command-line Options)
• variable numbers of arguments and keyword arguments
• multiple inheritance
• nested functions and classes
• unicode
• inheritance from builtins (excluding Exception and object)
• overloading __iter__ and __call__
• closures

Some other features are currently only partially supported:

• class attributes must always be accessed using a class identifier:

        self.class_attr # bad
        SomeClass.class_attr # good
        SomeClass.some_static_method() # good

• function references can be passed around, but not method references, and they cannot be contained:

        var = lambda x, y: x+y # good
        var = some_func # good
        var = self.some_method # bad, method reference
        [var] # bad, contained

Library Limitations

At the moment, the following 25 modules are largely supported. Several of these, such as os.path, were compiled to C++ using Shed Skin.

• array
• binascii
• bisect
• collections (defaultdict, deque)
• colorsys
• ConfigParser (no SafeConfigParser)
• copy
• csv (no Dialect, Sniffer)
• datetime
• fnmatch
• getopt
• glob
• heapq
• itertools (no starmap)
• math
• mmap
• os
• os.path
• random
• re
• socket
• string
• struct (no Struct, pack_into, unpack_from)
• sys
• time

Note that any other module, such as pygame, pyqt or pickle, may be used in combination with a Shed Skin generated extension module. For examples of this, see the Shed Skin examples.

See How to help out in Development on how to help improve or add to the set of supported modules.

Installation

There are four types of downloads available: a self-extracting Windows installer, a Debian (Ubuntu) package, an RPM package, and a UNIX tarball.

Windows

To install the Windows version, simply download and start it. (If you use ActivePython or some other non-standard Python distribution, or MingW, please deinstall this first.)

Debian (Ubuntu)

To install the Debian package, simply download and install it using your package manager. Make sure the following packages are installed (at least version 4.2 of g++):

sudo apt-get install g++ libpcre3-dev libgc-dev python-dev

RPM

(Note that several RPM distributions, such as Fedora, contain Shed Skin in their package repositories.)

To install the RPM package, simply download and install it using your package manager. Make sure the following packages are installed (at least version 4.2 of gcc-g++):

sudo yum install gcc-c++ pcre-devel gc-devel python-devel

UNIX

To install the UNIX tarball on a GNU/Linux or OSX system, take the following steps:

• download and unpack it
• run sudo python setup.py install
• For OSX: install the Apple XCode development environment
• make sure you can run g++, the C++ compiler (version 4.2 or higher)
• install the Boehm garbage collector (including development files)
• install the PCRE library (including development files)
• make sure the Python development files are installed

If the Boehm garbage collector is not available via your package manager, the following is known to work. Download for example version 7.2alpha6 from the website, unpack it, and install it as follows:

./configure --prefix=/usr/local --enable-threads=posix --enable-cplusplus --enable-thread-local-alloc --enable-large-config
make
make check
sudo make install

If the PCRE library is not available via your package manager, the following is known to work. Download for example version 8.12 from the website, unpack it, and build as follows:

./configure --prefix=/usr/local
make
sudo make install

Compiling a Stand-Alone Program

Under Windows, first execute (double-click) the init.bat file in the directory where you installed Shed Skin.

To compile the following simple test program, called test.py:

print 'hello, world!'

Type:

shedskin test

This will create two C++ files, called test.cpp and test.hpp, as well as a Makefile.

To create an executable file, called test (or test.exe), type:

make

Generating an Extension Module

To compile the following program, called simple_module.py, as an extension module:

# simple_module.py

def func1(x):
    return x+1

def func2(n):
    d = dict([(i, i*i)  for i in range(n)])
    return d

if __name__ == '__main__':
    print func1(5)
    print func2(10)

Type:

shedskin -e simple_module
make

For 'make' to succeed on a non-Windows system, make sure to have the Python development files installed (under Debian, install python-dev; under Fedora, install python-devel).

Note that for type inference to be possible, the module must (indirectly) call its own functions. This is accomplished in the example by putting the function calls under the if __name__=='__main__' statement, so that they are not executed when the module is imported. Functions only have to be called indirectly, so if func2 calls func1, the call to func1 can be omitted.

The extension module can now be simply imported and used as usual:

>>> from simple_module import func1, func2
>>> func1(5)
6
>>> func2(10)
{0: 0, 1: 1, 2: 4, 3: 9, 4: 16, 5: 25, 6: 36, 7: 49, 8: 64, 9: 81}

Limitations

There are some important differences between using the compiled extension module and the original.

1. Only builtin scalar and container types (int, float, complex, bool, str, list, tuple, dict, set) as well as None and instances of user-defined classes can be passed/returned. So for instance, anonymous functions and iterators are currently not supported.
2. Builtin objects are completely converted for each call/return from Shed Skin to CPython types and back, including their contents. This means you cannot change CPython builtin objects from the Shed Skin side and vice versa, and conversion may be slow. Instances of user-defined classes can be passed/returned without any conversion, and changed from either side.
3. Global variables are converted once, at initialization time, from Shed Skin to CPython. This means that the value of the CPython version and Shed Skin version can change independently. This problem can be avoided by only using constant globals, or by adding getter/setter functions.
4. Multiple (interacting) extension modules are not supported at the moment. Also, importing and using the Python version of a module and the compiled version at the same time may not work.

Numpy Integration

Shed Skin does not currently come with direct support for Numpy. It is possible however to pass a Numpy array to a Shed Skin compiled extension module as a list, using its tolist method. Note that this is very inefficient (see above), so it is only useful if a relatively large amount of time is spent inside the extension module. Consider the following example:

# simple_module2.py

def my_sum(a):
    """ compute sum of elements in list of lists (matrix) """
    h = len(a) # number of rows in matrix
    w = len(a[0]) # number of columns
    s = 0.0
    for i in range(h):
        for j in range(w):
            s += a[i][j]
    return s

if __name__ == '__main__':
    print my_sum([[1.0, 2.0], [3.0, 4.0]]) 

After compiling this module as an extension module with Shed Skin, we can pass in a Numpy array as follows:

>>> import numpy
>>> import simple_module2
>>> a = numpy.array(([1.0, 2.0], [3.0, 4.0]))
>>> simple_module2.my_sum(a.tolist())
10.0

Distributing Binaries

Windows

To use a generated Windows binary on another system, or to start it without having to double-click init.bat, place the following files into the same directory as the binary:

shedskin-0.9\shedskin\gc.dll
shedskin-0.9\shedskin-libpcre-0.dll
shedskin-0.9\bin\libgcc_s_dw-1.dll
shedskin-0.9\bin\libstdc++.dll

UNIX

To use a generated binary on another system, make sure libgc and libpcre3 are installed there. If they are not, and you cannot install them globally, you can place copies of these libraries into the same directory as the binary, using the following approach:

$ ldd test
libgc.so.1 => /usr/lib/libgc.so.1
libpcre.so.3 => /lib/x86_64-linux-gnu/libpcre.so.3
$ cp /usr/lib/libgc.so.1 .
$ cp /lib/x86_64-linux-gnu/libpcre.so.3 .
$ LD_LIBRARY_PATH=. ./test

Note that both systems have to be 32- or 64-bit for this to work. If not, Shed Skin must be installed on the other system, to recompile the binary.

Parallel Processing

Suppose we have defined the following function in a file, called meuk.py:

def part_sum(start, end):
    """ calculate partial sum """
    sum = 0
    for x in xrange(start, end):
        if x % 2 == 0:
            sum -= 1.0 / x
        else:
            sum += 1.0 / x
    return sum

if __name__ == ’__main__’:
    part_sum(1, 10)

To compile this into an extension module, type:

    shedskin -e meuk
    make

To use the generated extension module with the multiprocessing standard library module, simply add a pure-Python wrapper:

from multiprocessing import Pool

def part_sum((start, end)):
    import meuk
    return meuk.part_sum(start, end)

pool = Pool(processes=2)
print sum(pool.map(part_sum, [(1,10000000), (10000001, 20000000)]))

Calling C/C++ Code

To call manually written C/C++ code, follow these steps:

• Provide Shed Skin with enough information to perform type inference, by providing it with a type model of the C/C++ code. Suppose we wish to call a simple function that returns a list with the n smallest prime numbers larger than some number. The following type model, contained in a file called stuff.py, is sufficient for Shed Skin to perform type inference:

#stuff.py

def more_primes(n, nr=10):
    return [1]

• To actually perform type inference, create a test program, called test.py, that uses the type model, and compile it:

#test.py

import stuff
print stuff.more_primes(100)

shedskin test

• Besides test.py, this also compiles stuff.py to C++. Now you can fill in manual C/C++ code in stuff.cpp. To avoid that it is overwritten the next time test.py is compiled, move stuff.* to the Shed Skin lib/ dir.

Standard Library

By moving stuff.* to lib/, we have in fact added support for an arbitrary library module to Shed Skin. Other programs compiled by Shed Skin can now import stuff and use more_primes. In fact, in the lib/ directory, you can find type models and implementations for all supported modules. As you may notice, some have been partially converted to C++ using Shed Skin.

Shed Skin Types

Shed Skin reimplements the Python builtins with its own set of C++ classes. These have a similar interface to their Python counterparts, so they should be easy to use (provided you have some basic C++ knowledge.) See the class definitions in lib/builtin.hpp for details. If in doubt, convert some equivalent Python code to C++, and have a look at the result!

Command-line Options

The shedskin command can be given the following options:

    -a --ann               Output annotated source code (.ss.py)
    -b --nobounds          Disable bounds checking
    -e --extmod            Generate extension module
    -f --flags             Provide alternate Makefile flags
    -l --long              Use long long ("64-bit") integers
    -m --makefile          Specify alternate Makefile name
    -n --silent            Silent mode, only show warnings
    -o --noassert          Disable assert statements
    -r --random            Use fast random number generator (rand())
    -s --strhash           Use fast string hashing algorithm (murmur)
    -w --nowrap            Disable wrap-around checking
    -x --traceback         Print traceback for uncaught exceptions
    -L --lib               Add a library directory

For example, to compile the file test.py as an extension module, type shedskin –e test or shedskin ––extmod test.

Using -b or --nobounds is also very common, as it disables out-of-bounds exceptions (IndexError), which can have a large impact on performance.

    a = [1, 2, 3]
    print a[5] # invalid index: out of bounds

Performance Tips and Tricks

Performance Tips

• Allocating many small objects (e.g. tuples) typically does not slow down Python programs by much. However, after compilation to C++, it can quickly become a bottleneck. The key to getting excellent performance is to allocate as few small objects as possible.

• But note that for the idiomatic for a, b in enumerate(..), for a, b in enumerate(..) and for a, b in somedict.iteritems(), the intermediate small objects are optimized away, and that 1-length strings are cached.

• Several Python features (that may slow down generated code) are not always necessary, and can be turned off. See the section Command-line Options for details. Turning off bounds checking is usually a very safe optimization, and can help a lot for indexing-heavy code.

• Attribute access is faster in the generated code than indexing. For example, v.x * v.y * v.z is faster than v[0] * v[1] * v[2].

• Shed Skin takes the flags it sends to the C++ compiler from the FLAGS* files in the Shed Skin installation directory. These flags can be modified, or overruled by creating a local file named FLAGS.

• When doing float-heavy calculations, it is not always necessary to follow exact IEEE floating-point specifications. Avoiding this by adding -ffast-math can sometimes greatly improve performance.

• Profile-guided optimization can help to squeeze out even more performance. For a recent version of GCC, first compile and run the generated code with -fprofile-generate, then with fprofile-use.

• For best results, configure a recent version of the Boehm GC using CPPFLAGS="-O3 -march=native" ./configure --enable-cplusplus --enable-threads=pthreads --enable-thread-local-alloc --enable-large-config --enable-parallel-mark. The last option allows the GC to take advantage of having multiple cores.

• When optimizing, it is extremely useful to know exactly how much time is spent in each part of your program. The program Gprof2Dot can be used to generate beautiful graphs for a stand-alone program, as well as the original Python code. The program OProfile can be used to profile an extension module.

To use Gprof2dot, download gprof2dot.py from the website, and install Graphviz. Then:

shedskin program
make program_prof
./program_prof
gprof program_prof | gprof2dot.py | dot -Tpng -ooutput.png

To use OProfile, install it and use it as follows.

shedskin -e extmod
make
sudo opcontrol --start
python main_program_that_imports_extmod
sudo opcontrol --shutdown
opreport -l extmod.so

Tricks

• The following two code fragments work the same, but only the second one is supported:

statistics = {'nodes': 28, 'solutions': set()}

class statistics: pass
s = statistics(); s.nodes = 28; s.solutions = set()

• The evaluation order of arguments to a function or print changes with translation to C++, so it's better not to depend on this:

print 'hoei', raw_input() # raw_input is called before printing 'hoei'!

• Tuples with different types of elements and length > 2 are currently not supported. It can however be useful to 'simulate' them:

class mytuple:
    def __init__(self, a, b, c):
        self.a, self.b, self.c = a, b, c

• Block comments surrounded by #{ and #} are ignored by Shed Skin. This can be used to comment out code that cannot be compiled. For example, the following will only produce a plot when run using CPython:

print "x =", x
print "y =", y
#{
import pylab as pl
pl.plot(x, y)
pl.show()
#}

How to help out in Development

Open source projects thrive on feedback. Please send in bug reports, patches or other code, or suggestions about this document; or join the mailing list and start or participate in discussions. There is also a page with suggestions for possible tasks to start out with.

If you are a student, you might want to consider applying for the yearly Google Summer of Code or GHOP projects. Shed Skin has so far successfully participated in one Summer of Code and one GHOP.