Skip to content

HTTPS clone URL

Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP

Loading…

add dirty trick for readline import on OSX #937

Merged
merged 2 commits into from

2 participants

@minrk
Owner

also made the libedit warning extremely loud, so people don't miss it.

We still get reports of people never having noticed the warning, and getting confused when readline is broken on OSX.

The libedit test is also simplified to just check libedit in readline.__doc__, which the official Python docs recommend, rather than the elaborate subprocess/otool method that was not especially reliable.

The reason for the dirty trick

pip installs to site-packages by default, but site-packages dirs always come after lib-dynload (and extras, etc.), which is where the system readline is installed.

That means that a non-setuptools install (pip or setup.py install) cannot override any package that ships with OSX, including:

  • numpy
  • readline
  • twisted
  • pyobjc

without installing to a non-standard path (not even user site-packages via --user).

Note that this path issue not unique to OSX - lib-dynload, and patform packages always come before site-packages, including --user. It's just that OSX seems to be the only platform where there's anything in those locations that you would actually want to supplant.

The method for the dirty trick

  1. remove lib-dynload from sys.path before trying to import readline the first time
  2. after import, restore lib-dynload to its place in sys.path
  3. if import failed without lib-dynload, try it one more time, to get the default module

This doesn't solve everything

It has also come to my attention that EPD and Fink Pythons both ship with broken readline (logical for EPD, given readline's license). However, neither of these report as being anything other than GNU readline, so we can't detect that they are broken.

It's possible that we could print a warning if we detect that we are in EPD and we get to step 3 without readline, but if EPD's readline ever does work, then that message would be misleading, and we would have no real way to know that the issue has been fixed.

The answer remains, in all cases (except for possibly Python.org Python and macports, only because they are untested), that the first command OSX Python users should ever do is easy_install readline. Not just the System Python. After this merge, pip does work to override readline in IPython, but not anywhere else in your Python world, so the warning message does note that pip install will likely not work, even though it does within IPython.

@minrk minrk add dirty trick for readline import on OSX
also made the libedit warning extremely loud, so people don't miss it.
We still get reports of people never having noticed the warning, and
getting confused when readline is broken on OSX.

The reason for the dirty trick:

    pip installs to site-packages by default, but site-packages dirs always
    come *after* lib-dynload (and extras, etc.), which is where the system
    readline is installed.

    That means that a non-setuptools install (pip or setup.py install) *cannot*
    override any package that ships with OSX, including:
        numpy, readline, twisted, pyobjc
    without installing to a non-standard path (not even user site-packages via `--user`).

The method for the dirty trick:

    1. remove lib-dynload from sys.path before trying to import readline the first time
    2. after import, restore lib-dynload to its place in sys.path
    3. if import failed without lib-dynload, try it one more time
387dcd6
IPython/utils/rlineimpl.py
@@ -17,6 +17,18 @@ import warnings
from subprocess import Popen, PIPE
+if sys.platform == 'darwin':
+ # dirty trick, to skip the system readline, because pip-installed readline
+ # will never be found on OSX, since lib-dynload always comes ahead of site-packages
+ from distutils import sysconfig
+ lib_dynload = sysconfig.get_config_var('DESTSHARED')
+ del sysconfig
+ try:
+ dynload_idx = sys.path.index(lib_dynload)
+ except ValueError:
+ dynload_idx = -1
@fperez Owner
fperez added a note

I'd use None instead of -1 for the sentinel, since it's a bit more obvious that a None index can never be actually correct. That will make the check below read if dynload_idx is not None, which also reads better (to me).

@minrk Owner
minrk added a note

Sure, I'll do that. It's definitely more Pythonic, I have just gotten used to c-style '-1 index means not found'.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
IPython/utils/rlineimpl.py
((43 lines not shown))
- "including some crashes on tab completion, and incorrect history navigation. "
- "It is highly recommended that you install readline, "
- "which is easy_installable with: 'easy_install readline'",
- RuntimeWarning)
- uses_libedit = True
- # cleanup names
- del dt,p,out,err
+if have_readline:
+ # Official Python docs state that 'libedit' is in the docstring for libedit readline:
+ uses_libedit = 'libedit' in _rl.__doc__
+ # Note that many non-System Pythons also do not use proper readline,
+ # but do not report libedit at all, nor are they linked dynamically against libedit.
+ # known culprits of this include: EPD, Fink
+ # There is not much we can do to detect this, until we find a specific failure
+ # case, rather than relying on the readline module to self-identify as broken.
+if uses_libedit and sys.platform == 'darwin':
@fperez Owner
fperez added a note

I'd add a line of whitespace above 94 to separate these two big if blocks for readability

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
@fperez
Owner

Other than the minor edits above, I have nothing against the approach. We should probably ping Enthought to ask them to make their readline identifiable in some manner though, I'm sure they could just add a flag to their module to allow us to recognize it.

I haven't run the code myself, but given your OSX expertise and how much you've dealt with this issue, I'm OK if you want to merge once the above tiny things are fixed. Though if you want a second pair of OSX-specific eyes, that's OK too.

@minrk
Owner

I do worry about mucking with the path, but I was helping someone out today, explaining why pip doesn't work, and they asked if I could just instruct Python to ignore the system one, allowing the site-packages to be searched for readline, and I realized it actually is possible. I tested it, and it certainly does work - pip-installed readline will be found by IPython, at least with System Python and EPD on my machine.

I honestly don't know why EPD and/or fink readline are broken - we should ask Ilan, or someone. Just this week, I have had people with EPD at py4science and Fink (in #915) report weird readline behavior (not the same), and both found that their respective weirdnesses were resolved by easy_install readline. If it is that they are linked against libedit (fink is not, but EPD might be) and just not reporting it, that would make sense, and it is a bug that their readline.__doc__ is wrong. But I don't know what the default behavior is, when you build Python and readline is not present.

@minrk minrk STY: rlineimpl review from @fperez
* add some line breaks
* use None to indicate dynload not found, instead of -1
d86e772
@minrk
Owner

style comments addressed

@fperez
Owner

I'd say go for this one, but a ping to Ilan on enthought-dev would be probably a good idea. Sorry to dump that one on you, but I simply don't know enough about osx to say anything intelligent on such a discussion...

@fperez
Owner

Awesome, merging now!

@fperez fperez merged commit 90b0139 into ipython:master
@fperez fperez referenced this pull request from a commit
Commit has since been removed from the repository and is no longer available.
@minrk minrk referenced this pull request in ludwigschwardt/python-gnureadline
Merged

put readline.so in python_readline.readline #26

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Commits on Oct 28, 2011
  1. @minrk

    add dirty trick for readline import on OSX

    minrk authored
    also made the libedit warning extremely loud, so people don't miss it.
    We still get reports of people never having noticed the warning, and
    getting confused when readline is broken on OSX.
    
    The reason for the dirty trick:
    
        pip installs to site-packages by default, but site-packages dirs always
        come *after* lib-dynload (and extras, etc.), which is where the system
        readline is installed.
    
        That means that a non-setuptools install (pip or setup.py install) *cannot*
        override any package that ships with OSX, including:
            numpy, readline, twisted, pyobjc
        without installing to a non-standard path (not even user site-packages via `--user`).
    
    The method for the dirty trick:
    
        1. remove lib-dynload from sys.path before trying to import readline the first time
        2. after import, restore lib-dynload to its place in sys.path
        3. if import failed without lib-dynload, try it one more time
  2. @minrk

    STY: rlineimpl review from @fperez

    minrk authored
    * add some line breaks
    * use None to indicate dynload not found, instead of -1
This page is out of date. Refresh to see the latest.
Showing with 56 additions and 44 deletions.
  1. +56 −44 IPython/utils/rlineimpl.py
View
100 IPython/utils/rlineimpl.py
@@ -17,6 +17,18 @@
from subprocess import Popen, PIPE
+if sys.platform == 'darwin':
+ # dirty trick, to skip the system readline, because pip-installed readline
+ # will never be found on OSX, since lib-dynload always comes ahead of site-packages
+ from distutils import sysconfig
+ lib_dynload = sysconfig.get_config_var('DESTSHARED')
+ del sysconfig
+ try:
+ dynload_idx = sys.path.index(lib_dynload)
+ except ValueError:
+ dynload_idx = None
+ else:
+ sys.path.pop(dynload_idx)
try:
from readline import *
import readline as _rl
@@ -29,6 +41,26 @@
except ImportError:
have_readline = False
+if sys.platform == 'darwin':
+ # dirty trick, part II:
+ if dynload_idx is not None:
+ # restore path
+ sys.path.insert(dynload_idx, lib_dynload)
+ if not have_readline:
+ # *only* have system readline, try import again
+ try:
+ from readline import *
+ import readline as _rl
+ have_readline = True
+ except ImportError:
+ have_readline = False
+ else:
+ # if we want to warn about EPD / Fink having bad readline
+ # we would do it here
+ pass
+ # cleanup dirty trick vars
+ del dynload_idx, lib_dynload
+
if have_readline and hasattr(_rl, 'rlmain'):
# patch add_history to allow for strings in pyreadline <= 1.5:
# fix copied from pyreadline 1.6
@@ -51,52 +83,32 @@ def add_history(line):
# Test to see if libedit is being used instead of GNU readline.
# Thanks to Boyd Waters for the original patch.
uses_libedit = False
-if sys.platform == 'darwin' and have_readline:
- # Previously this used commands.getstatusoutput, which uses os.popen.
- # Switching to subprocess.Popen, and exponential falloff for EINTR
- # seems to make this better behaved in environments such as PyQt and gdb
- dt = 1e-3
- while dt < 1:
- try:
- p = Popen(['otool', '-L', _rl.__file__], stdout=PIPE, stderr=PIPE)
- except OSError:
- try:
- # otool not available (no XCode), use lsof instead.
- # This *could* have a false positive
- # if another package that uses libedit explicitly
- # has been imported prior to this test.
- p = Popen(['lsof', '-p', str(os.getpid())], stdout=PIPE, stderr=PIPE)
- except OSError:
- # This is highly unlikely, but let's be sure
- # we don't crash IPython just because we can't find lsof
- p = out = err = None
- warnings.warn("libedit detection failed")
- break
-
- out,err = p.communicate()
- if p.returncode == 4:
- # EINTR
- time.sleep(dt)
- dt *= 2
- continue
- elif p is None or p.returncode:
- warnings.warn("libedit detection failed: %s"%err)
- break
- else:
- break
+if have_readline:
+ # Official Python docs state that 'libedit' is in the docstring for libedit readline:
+ uses_libedit = 'libedit' in _rl.__doc__
+ # Note that many non-System Pythons also do not use proper readline,
+ # but do not report libedit at all, nor are they linked dynamically against libedit.
+ # known culprits of this include: EPD, Fink
+ # There is not much we can do to detect this, until we find a specific failure
+ # case, rather than relying on the readline module to self-identify as broken.
- if p is not None and p.returncode == 0 and re.search(br'/libedit[\.\d+]*\.dylib\s', out):
- # we are bound to libedit - new in Leopard
- _rl.parse_and_bind("bind ^I rl_complete")
- warnings.warn("Leopard libedit detected - readline will not be well behaved "
- "including some crashes on tab completion, and incorrect history navigation. "
- "It is highly recommended that you install readline, "
- "which is easy_installable with: 'easy_install readline'",
- RuntimeWarning)
- uses_libedit = True
- # cleanup names
- del dt,p,out,err
+if uses_libedit and sys.platform == 'darwin':
+ _rl.parse_and_bind("bind ^I rl_complete")
+ warnings.warn('\n'.join(['', "*"*78,
+ "libedit detected - readline will not be well behaved, including but not limited to:",
+ " * crashes on tab completion",
+ " * incorrect history navigation",
+ " * corrupting long-lines",
+ " * failure to wrap or indent lines properly",
+ "It is highly recommended that you install readline, which is easy_installable:",
+ " easy_install readline",
+ "Note that `pip install readline` generally DOES NOT WORK, because",
+ "it installs to site-packages, which come *after* lib-dynload in sys.path,",
+ "where readline is located. It must be `easy_install readline`, or to a custom",
+ "location on your PYTHONPATH (even --user comes after lib-dyload).",
+ "*"*78]),
+ RuntimeWarning)
# the clear_history() function was only introduced in Python 2.4 and is
# actually optional in the readline API, so we must explicitly check for its
Something went wrong with that request. Please try again.