On OS X 10.11.6, python 2.7 builds without readline since configure fail to find rl_resize_terminal in -lreadline, so I could not test it.

That's a bit weird. I think it should build fine since all of the usages of rl_resize_terminal covered by the HAVE_RL_RESIZE_TERMINAL check. See this for an example. Could you post a little bit more information here?

The issue is this check in setup.py:

if host_platform == 'darwin':
    os_release = int(os.uname()[2].split('.')[0])
    dep_target = sysconfig.get_config_var('MACOSX_DEPLOYMENT_TARGET')
    if (dep_target and
            (tuple(int(n) for n in dep_target.split('.')[0:2])
                < (10, 5) ) ): 
        os_release = 8
    if os_release < 9: 
        # MacOSX 10.4 has a broken readline. Don't try to build
        # the readline module unless the user has installed a fixed
        # readline package
        if find_file('readline/rlconf.h', inc_dirs, []) is None:
            do_readline = False```

When configuring 2.7, we MACOSX_DEPLOYMENT_TARGET is "10.4", and readline is
disabled. When configuring master/3.6/3.5, we get "10.11", so readline is not disabled.

MACOSX_DEPLOYMENT_TARGET Identifies the earliest OS X version the product is to
run on. So with 2.7 we try to build a version that runs on any OS X version down to 10.4,
and in master/3.x we build only for the current version. I did not find why this happens.

This is also reproducible on travis OS X build, see:
https://travis-ci.org/python/cpython/jobs/251594769

Python build finished, but the necessary bits to build these modules were not found:
_bsddb             dl                 imageop         
linuxaudiodev      ossaudiodev        readline        
spwd               sunaudiodev                        

I think we need to open a separate issue for this, this is not related to this fix, since readline
is not built by default on OS X.

I think we need to open a separate issue for this, this is not related to this fix, since readline
is not built by default on OS X.

Agreed, could you create a new issue on bugs.p.o?

@Haypo, the title is hopefully better now.

[2.7] bpo-29854: Backport fix for 2.7 (without the test) #2637

Sorry, but this title is still not good. I would prefer "[2.7] bpo-29854: Fix segfault in call_readline()".

Why not backporting the unit test?

@Haypo, the test depends on run_pty(), not available in 2.7, and it depends on the selectors module. Too much work for one test.

@berkerpeksag: What do you think?

@berkerpeksag: What do you think?

Agreed. I just merged the PR. Thank you all!