pypy with numpy

Looks like pypy now can build numpy. Well, a slightly modified numpy.

1. Get default branch of pypy. be careful cause the developers don't maintain default branch compilable. Revision 84341 (c86b42dd7613) works for me.
2. compile using 
./rpython/bin/rpython -O2 ./pypy/goal/targetpypystandalone.py --withoutmod-micronumpy
3. Create package and vitual environment. Something like this:
./pypy/tool/release/package.py --targetdir ./my_builds/build.tar.bz2 --builddir ./tmp/ --nostrip --archive-name pypy_84341
Needed to copy pypy-c and libpypy to pypy/goal beforehand.
4. Clone and follow instructions from
https://github.com/pypy/numpy/commits/cpyext-ext
Revision 3299d0d76fdb831fbcb4429a89c1f53bb36ea07f worked for me

Testing results:
----------------------------------------------------------------------
Ran 5900 tests in 78.216s

FAILED (KNOWNFAIL=3, SKIP=6, errors=218, failures=83)

scipy can be compiled if disabling submodules io/matlab and spatial. Though is still doesn't work.

Scipy in pypy

Looks like it is too early for using scipy in Pypy. There is a plenty of dependencies on C-code there.
I was able to install scipy 0.17 in pypy. I disabled all failed dependencies. Unfortunately it is completely useless. Almost everything doesn't work.

https://github.com/serge-m/scipy/blob/v0.17.0_for_pypy_bin/bin/scipy_0.17.0_for_pypy_draft.tar.gz

Building Pypy

Pypy builds faster if using -O2 option.



To build faster (according to pypy documentation) use prebuilt pypy from http://buildbot.pypy.org/nightly/trunk/

Using virtualenv to create virtual environment for it.

Build script (to be placed in pypy source directory):

#!/bin/bash
cd pypy/goal || exit 1
source <path to existing pypy environment>/bin/activate || exit 2
pypy ../../rpython/bin/rpython --batch -O2 targetpypystandalone

TIL about PyPy

Building from source root using command
pypy_src$ rpython/bin/rpython -Ojit pypy/goal/targetpypystandalone.py
produces structure with obsolete pypy-c and libpypy-c.so in
/tmp/usession-release-4.0.1-XXXX/build/pypy-nightly/bin/

Probably pypy compiler places there files integrated in the src distribution. To get fresh versions I had to use pypy-c and libpypy-c.so from sources root.

UPDATE:
Probably I was completely wrong.
pypy/tool/release/package.py has an option for (not) stripping resulting binary file: "--nostrip". By default it is enabled. Looks like it removed something unused from binaries. This operation updates timestamp of the pypy-c and libpypy-c.so. So probably that was the cause of my misunderstanding.


Script for packaging and creating virtual environment:

#!/bin/bash

rm -rf ./my_builds/ || exit 2
mkdir ./my_builds/ || exit 3

DST_NAME=$1
if [ -z "$DST_NAME" ]; then
    echo "DST_NAME is  empty"
    exit 3
fi

# runs packaging
./pypy/tool/release/package.py --builddir /home/pypy/builds/ --nostrip --archive-name $DST_NAME || exit 4

# creates a new virtual environment
virtualenv -p /home/pypy/builds/$DST_NAME/bin/pypy /home/pypy/env/$DST_NAME

# installing nose for numpy testing (optional)
source /home/pypy/env/$DST_NAME/bin/activate
pip install nose