Do not install the Python test suite when configure is run with '--disable-test-suite'. This about divides by two the size of the installed standard library.

Related to issue 26852: add the '--enable-legacy-pyc-files' option to configure.

The attached listdirs.py script may be used to check that the set of sub-directories in the std library is not changed by the patch.

We have a similar patch in Buildroot (see https://git.buildroot.org/buildroot/tree/package/python3/0017-Add-an-option-to-disable-installation-of-test-module.patch) so we would be very happy to see this patch merged.

Note that we also have many more patches to disable various parts of the Python standard installation (https://git.buildroot.org/buildroot/tree/package/python3/). So either we have one option for each feature (like we have implemented), or a more general config option --disable-feature=test,this,that.

See also bug http://bugs.python.org/issue20210 which is related, and was also proposing a patch to disable the test suite.

Thanks for the review Martin. I will update the patch accordingly later.

I'm in favor of allowing to build a "subset" of Python. Not only it makes sense for embedded devices, but in many cases it's even not possible to do otherwise: "full" Python is too big for small devices.

But I would prefer to see a public discussion on python-dev before going into this direction.

The previous attempt failed because some core developers were opposed to support officially a "subset of Python". But since this previous discussion (two years ago? longer?), MicroPython became popular, and this implementation of Python has a small subset of the stdlib (but it supports asyncio ;-)) Maybe it's time for a new attempt :-) Oh, and the meanwhile we also made progress on the Android port, which is now an important platform.

In short, I would like to official support Python subsets ;-) At least, add options for that.

> But I would prefer to see a public discussion on python-dev before going into this direction.

Ah, the discussion restarted on the other issue:
http://bugs.python.org/issue20210#msg287516

> Ah, the discussion restarted on the other issue: http://bugs.python.org/issue20210#msg287516

Issue 20210 has ended up as a change in the build system eventually and not the installation process, only a tiny step towards the configuration of a smaller Python distribution :(

After browsing again the last discussion on "[Python-ideas] size of the installation of Python on mobile devices" [1], I find all the arguments in favor of such changes still relevant. I would also add the following new arguments:

* One frequent argument used against those changes is that "[packagers for mobile and embedded devices] will have to patch the code base anyway to get things working in such an environment". But this is not true, see issue 30386 (Add a build infrastructure for Android) that proposes an addition to Python that does not modify a SINGLE LINE in the existing source code. So it is not true (or at least not anymore true) that packagers for mobile and embedded devices need to patch the code base anyway.

* When Python is installed by the termux [2] Android application on an Android device, the size of the installed standard library is 34M. The termux installation script does not install the following files [3]:

    TERMUX_PKG_RM_AFTER_INSTALL="
    bin/python${_MAJOR_VERSION}m bin/idle*
    lib/python${_MAJOR_VERSION}/idlelib
    lib/python${_MAJOR_VERSION}/test
    lib/python${_MAJOR_VERSION}/tkinter
    lib/python${_MAJOR_VERSION}/turtledemo
    "

But it fails to not install also:
    lib/python${_MAJOR_VERSION}/config-${_MAJOR_VERSION}m
    lib/python${_MAJOR_VERSION}/ctypes/test
    lib/python${_MAJOR_VERSION}/distutils/tests
    lib/python${_MAJOR_VERSION}/sqlite3/test
    lib/python${_MAJOR_VERSION}/unittest/test

This is not consistent and should be fixed by a proper (endorsed by Python) installation process.

[1] https://mail.python.org/pipermail/python-ideas/2016-July/
[2] https://termux.com/
[3] https://github.com/termux/termux-packages/blob/master/packages/python/build.sh

I modified setup.py to add a new TEST_EXTENSIONS which allows to not build test extensions:

New changeset cfe172dc6bd0a02d36db31ffabcc38f9320a4510 by Victor Stinner in branch 'master':
bpo-36146: Add TEST_EXTENSIONS to setup.py (GH-12129)
https://github.com/python/cpython/commit/cfe172dc6bd0a02d36db31ffabcc38f9320a4510

setup.py should be modified manually, but it's a small step towards be able to build Python without tests ;-)

pr23886_files.diff: see https://github.com/python/cpython/pull/23886#issuecomment-752279731

New changeset 277ce3060becc120f1c877346562bd6880f4be71 by pxinwr in branch 'master':
bpo-27640: Add --disable-test-modules configure option (GH-23886)
https://github.com/python/cpython/commit/277ce3060becc120f1c877346562bd6880f4be71

Thanks all, it's now fixed in the master branch, the future Python 3.10.

Thanks Xavier de Gaye for the initial patch and reported the issue.

Thanks Peixing Xin for the final change which is now even better! 

Thanks also Thomas Petazzoni who maintained the downstream patch in buildroot.

See also discussion on PR 23886 about the technical change and the different tests that we run to ensure that the change doesn't break anything, and that it works as expected.

I'm really not comfortable at all to change the Python build system. It's very complex and fragile, I don't know it very well, so I'm not confident in what I am doing. For these reasons, I'm strongly against backporting the change to stable branches (3.8 and 3.9) and so I close the issue

I have a slightly different interest in declining to install tests... Arch Linux provides the tests as a split subpackage with currently some hacky shell globbing to delete the tests after make install, then do a second libinstall and delete everything but the tests.

It would be more convenient if we had a Makefile option to install the tests separately.

It's now possible to disable installing tests with a filthy hack: make install TEST_MODULES= TESTSUBDIRS=

However, it's not possible to similarly hack make libinstall LIBSUBDIRS=
Because it installs some toplevel stuff, then runs lib2to3.pgen2.driver on the grammar that got skipped by the LIBSUBDIRS= hack.

Thoughts on maybe splitting this up at install time? Would probably mean duplicating some of libinstall as testinstall...

> It would be more convenient if we had a Makefile option to install the tests separately.

While this use case is tighly coupled to this feature, would you mind to open a separate issue for your use case? IMO it makes sense and it sounds easy to implement it (especially after --disable-test-modules feature has been implemented).

I filed issue43282 for this.

> I filed issue43282 for this.

Thanks ;-)