Blog Needs a Name

Python is Actually Portable

Back in February, I put together Lua 5.4 using Cosmopolitan Libc as a quick proof-of-concept, and that led to Lua 5.4 being vendored as part of the Cosmopolitan repository on Github, along with many other interesting developments. It’s pretty exciting to try and compile well-known C projects using Cosmopolitan; the portability reward is great motivation, and I get to observe the design, coding style, and build system of high-quality codebases.

However, my initial plan with Cosmopolitan was not to compile Lua, it was to compile Python. Since February, I’ve been trying to understand how Cosmopolitan works, reading the repo code and submitting PRs occasionally, and finally I have an actually portable version of Python 2.7.181 – you can build it from the repository here. It’s not solid as Lua 5.4 because it currently passes only a third of the regression tests, but most of the parts are there. For example, here’s a GIF showing a simple Flask webapp running via the python.com APE.

webapp

It’s quite slow, but it works(tested on Debian Linux and Windows 102), and there’s a lot of room for improvement. This post describes the kinds of changes I made to get Python to work.

Getting Python to compile

Python 2.7.18 is built via the configure idiom: the tarball provides a configure shell script which runs the necessary compatibility tests for the host system/compiler. which then creates the Makefile for the actual build, and fills pyconfig.h with the necessary #defines related to the features available in the host system.

I added dummy headers like for Lua, and wrote a superconfigure script which called configure with the flags for the Cosmopolitan amalgamation, like below:

./configure --disable-shared \
    --without-threads --disable-ipv6 \
    CFLAGS="$COSMO_CFLAGS" \
    LDFLAGS="$COSMO_LDFLAGS" \
    LIBS="cosmopolitan.a"
make -j4

This started the build, but almost every source file had complaints about functions being re-declared. For example, the Makefile assumed that sinh was not available, and wrongly had #define HAVE_SINH 0, which caused it to use Python’s backup implementation.

It turns out that the CFLAGS provided for the cosmopolitan amalgamation clashed with AC_CHECK_FUNC, which is what configure scripts use to check if the host provides a particular function. At the time, the simplest fix was to edit the AC_CHECK_FUNC part of configure script to make it play well, and it just worked3.

After that, fixing the remaining errors was mostly straightforward: add a few #defines or #undefs to avoid missing functions, ensure the Makefile performed the linking correctly, find-and-replace for name clashes4, and so on.

Running the base interpreter

I tried running the python.com APE within the build directory, and it worked without a hitch. But when I moved the executable somewhere, the interpreter would start up and then exit with a failure unable to import site: Python checks for site.py in sys.path at startup, which sets up the import context for the rest of the standard library (stuff like which OS the APE is on for os.py). Running python.com -S fixed the issue5, and later I added a few relative folder paths to sys.path so that Python could search there at startup.

One annoying issue I faced with the python.com REPL on Linux was I had to press Return twice when I wanted to enter an empty line to the REPL. By running python.com -u I found out this this was related to input buffering and setvbuf: the interpreter was following Cosmopolitan’s choice of buffering IO, so I added a couple of lines to use unbuffered input by default.

When I tried python.com -S on Windows, the interpreter would start, but then every line I entered at the REPL would cause a syntax error. I thought this was also related to setvbuf, but this was because Windows cmd.exe sends CRLF on pressing Return, and the interpreter expects only LF, so the CR was read as incorrect syntax. It was straightforward to change the parser logic.

Adding standard library modules

When the Makefile finishes building python, it immediately tries to call a setup.py to build C extensions for the standard library. This was pretty stupid: setup.py tries to build and link a bunch of shared objects to the interpreter even though I’ve specified a static build (--disable-shared). I found out it was doing this after I saw a successful build on the screen, followed a deluge of (unable to link/no dlopen/Rdynamic specified but also static) errors.

There is a nice way to compile C extensions into Python statically: via Modules/Setup. This file follows a simple syntax to specify extensions and their requirements, so that you can compile extensions before setup.py is called.

*static*

# <name of extension> <source files in Modules folder> <includes> <links> -DSOME_FLAG
math mathmodule.c _math.c # -lm
array arraymodule.c

Calling make after changing Modules/Setup rebuilds the Makefile with the necessary recipes for the specified extensions. You can read about Modules/Setup here. One nasty aspect of this is if the syntax in Modules/Setup is wrong, the Makefile will be wrong, so you can’t run make even after you’ve fixed the error.

I was able to compile a lot of modules into APE using Modules/Setup: basic modules like cPickle, zlib etc., modules that required external libraries like _sqlite, _bz2, readline, _ctypes6 etc., and even external packages that depend on simple C extensions, like greenlet. It all boils down to writing the correct recipe in Modules/Setup, ensuring the necessary static libraries are available, and checking that the glue code around the imports works correctly7.

The final annoyance with the stdlib python files was that I had to copy the folder containing them alongside python.com, but this got solved because Cosmopolitan allows the APE to be used as a valid ZIP file. I just zipped up Lib/ into the APE, added python.com itself as an entry in sys.path, and Python’s own zipimport package handled the rest.

Using pip and external packages

After compiling most of the standard library packages, I still couldn’t get ensurepip to work, because it relied on threads, and the standard threading module in Python would just raise an ImportError if it could not find the _thread C extension. The fix was present within the stdlib: there is a module called _dummy_thread, which spoofs _thread so that threading doesn’t complain8. After this, I was able to run ensurepip to install pip and setuptools in Lib/site-packages.

However, running python.com -m pip install doesn’t work because the Python APE doe not currently have SSL support. I tried providing http://pypi.org/simple as the index URL but pip was adamant in not letting me get my way.

I didn’t want to figure out how to add SSL support to the APE (I expect it should be possible), so I just cheated and downloaded the packages manually. The arrangement to add packages works like this:

mkdir -p Lib/site-packages

# specify python version to pip if necessary
/usr/bin/python2 -m pip download flask -t .

# wheels are just ZIPs, use unzip if pip complains
./python.com -m pip install flask*.whl -t ./Lib/site-packages

zip -qr ./python.com ./Lib/site-packages/
rm -rf ./*.whl ./Lib/site-packages/

./python.com -m flask --version

Of course, you would have recompile the APE if you wanted to add any C extensions.

Closing Notes

It is really convenient to have the same application work on Linux/Windows/MacOS, but the techniques I have seen/used all have tradeoffs: either for the developer in terms of design, testing, and safety/consistency, or for the user in terms on application size and performance.

Cosmopolitan Libc does come with some tradeoffs as well (static compilation, C codebases, no multithreading), but I think it is amazing that I can send a python.com executable of a few megabytes as a zip file to someone, have them run it without worrying about the OS, and provide them a simple webapp with a backend that works even without an internet connection.

This successful python.com experiment unlocks many interesting directions:

  1. Did you click on this before reading the remainder of the post because you were wondering about Python 3? I know Python 2.7 reached EOL last year, but the Python 2.7 codebase was easier to read, change, and debug, so I was able to get a better idea of where I had to make changes. I put together a Python 3.6.14 APE (build it from the repo here), and it took much less time and experimentation because I knew exactly where to look. 

  2. I originally uploaded this post like twelve hours, in the excitement of seeing a Flask webapp work on the APE. Unfortunately, I forgot to test on Windows before putting the post out, and Windows, true to its nature caused several annoyances which together took a couple of hours to sort out. However, I think it ought to work now. 

  3. I originally tried the complicated solution of writing a new configure script from scratch, but I got lost in the docs for autoconf, and resorted to the simpler fix of just editing the shell script. This issue only happened because I used dummy headers with the amalgamation; if you have the Cosmopolitan repo nearby, you can use -isystem cosmopolitan/libc/isystem instead of -I dummy_headers -include cosmopolitan.h, which would avoid confusing configure

  4. I remembered this a few days back: if a name in your C file clashes with a header name somewhere, you can do something like:

    #ifdef name
    #define name __name
    #undef name
    

    and avoid the name clash instead of doing a find-and-replace. 

  5. Over the course of getting the APE to work, my muscle memory is to call python -BESsuv, which avoids a bunch of normal startup stuff and adds stderr information about imports. This is super useful when building python, but one time I was running a local script and couldn’t figure out why it was failing before I saw that I had typed python -BESsuv script.py every single time. 

  6. I was pleasantly surprised to find out that _sqlite and _bz2 could be compiled, because I didn’t expect SQLite or libbz2 would be easy to compile from source with Cosmopolitan. SQLite is now vendored in the Cosmopolitan repo.

    On the other hand, I wanted readline because I like to have the (press-Up-arrow-for-previously-typed-line) in the REPL, but I didn’t know how terminfo worked. _ctypes was also a disappointment because I had to compile libffi from source, and then I saw most of _ctypes’ use comes from having dlopen

  7. for example, greenlet contains a C extension called _greenlet which it refers to by doing things like the following in __init__.py:

    from ._greenlet import foo
    from greenlet._greenlet import bar
    

    now __init__.py expects _greenlet.so to be present in the same directory, but that is not possible because the extension is compiled into the interpreter statically. So I had to manually change this to use from _greenlet without the dot. This also means I have to change the extension name in the PyModuleDef part of the C source code: it’s not greenlet._greenlet anymore, it’s just _greenlet

  8. This works because most libraries use only the high-level threading API to handle their threads. A notable counter-example is Django: I tried getting Django to work before figuring out how _greenlet could be compiled, but Django imports the low-level _thread API to handle things.