Title: gets filename wrong for eval/exec
Type: Stage:
Components: Library (Lib) Versions: Python 2.6
Status: closed Resolution: rejected
Dependencies: Superseder:
Assigned To: nnorwitz Nosy List: Rhamphoryncus, gvanrossum, nnorwitz, takluyver, zseil
Priority: normal Keywords: patch

Created on 2007-04-02 05:23 by Rhamphoryncus, last changed 2018-04-27 16:39 by takluyver. This issue is now closed.

File name Uploaded Description Edit
python2.6-warningfilename5.diff Rhamphoryncus, 2007-04-23 21:26 Now with more comments!
Messages (14)
msg52364 - (view) Author: Adam Olsen (Rhamphoryncus) Date: 2007-04-02 05:23
warnings.warn() gets the filename using the globals' __file__ attribute.  When using eval or exec this is often not the context in which the current line was compiled from.  The line number is correct, but will be applied to whatever file the globals are from, leading to an unrelated line being printed below the warning.

The attached patch makes it use caller.f_code.co_filename instead.  This also seems to remove the need to normalize .pyc/.pyo files, as well as not needing to use sys.argv[0] for __main__ modules.

It also cleans up warnings.warn() and adds three unit tests.
msg52365 - (view) Author: Adam Olsen (Rhamphoryncus) Date: 2007-04-03 07:00
The test_stackoverflow_message test I added causes to use sys.__warningregistry__ rather than test_warnings.__warningregistry__.  This circumvents this self-declared hack of deleting test_warnings.__warningregistry__ to allow regrtest -R to run.

This version of the patch uses a more general solution to allowing regrtest -R to run.  Probably doesn't qualify as a hack anymore...
File Added: python2.6-warningfilename2.diff
msg52366 - (view) Author: Adam Olsen (Rhamphoryncus) Date: 2007-04-04 21:28
sys._getframe(sys.maxint) produces an OverflowError on 64bit GCC.  Patch changed to use sys._getframe(10**9) instead.
File Added: python2.6-warningfilename3.diff
msg52367 - (view) Author: Adam Olsen (Rhamphoryncus) Date: 2007-04-16 01:34
I've rewritten the unit tests I added to follow the style of walter.doerwald's changes to (commited to SVN during the same period in which I posted my patch).

The guard_warnings_registry() context manager is now in test_support.
File Added: python2.6-warningfilename4.diff
msg52368 - (view) Author: Neal Norwitz (nnorwitz) * (Python committer) Date: 2007-04-23 06:15
I'm not sure why all the code was in there to begin with it was ancient history.  Hopefully Guido remembers.  Guido, do you see any particular issue with this approach (ie, removing all the gyrations in warnings and using f_code.co_filename?

There is a difference in behaviour.  If a module modifies its __file__, the old code would pick up the modified version.  This new code will find the real file.  I'm not sure if that's a bug or a feature though. :-)  Similarly with __name__.

Adam, I like adding the warning_guard using a with statement.  Although I don't think comment should be completely removed.  Could you add a comment about why the guard is necessary (ie, the repeated calls/-R part)?

Have you tested with -R and with -u all?  ie:

  ./python -E -tt ./Lib/test/ -R 4:3:
  ./python -E -tt ./Lib/test/ -u all

(two separate runs).

In the last test case you are replacing (the old spam7), why does "sys" become "no context"?
msg52369 - (view) Author: Guido van Rossum (gvanrossum) * (Python committer) Date: 2007-04-23 17:31
The main reason for using __file__ instead of the filename baked into the code object is that sometimes .pyc files are moved after being compiled; the code object is not updated, but __file__ has the correct path as it reflects how the module was found during import.  Also, code objects may have relative pathnames.
msg52370 - (view) Author: Adam Olsen (Rhamphoryncus) Date: 2007-04-23 21:26
Comment added.

regrtest -R and -u all report no problems specific to this patch.  A clean build of trunk does show problems without it, with test_structmembers, test_tcl, test_telnetlib, test_unicode_file, and test_uuid.  At this point I've seen so many weird failures that I think I've started repressing them. ;)

When no stack frame is available warn() falls back to sys as the module.  Since I'm trying to avoid claiming a warning came from a different file, I made it appear to come from <no context> instead.  Looking at it now though I wonder if I should go a step further, not having a __warningregistry__ or module at all for such warnings.

I see what you mean about stale co_filename attributes.  I'm not sure what the best fix is now.  I'm not exactly enthused at having to learn how the import mechanisms work. ;)
File Added: python2.6-warningfilename5.diff
msg52371 - (view) Author: Adam Olsen (Rhamphoryncus) Date: 2007-04-23 22:01
I "forgot", test_bsddb3 has problems too, but again they're not related to my patch.
msg52372 - (view) Author: Neal Norwitz (nnorwitz) * (Python committer) Date: 2007-04-24 05:04
Unfortunately, test_bsddb3 has lots of problems. :-(  They vary by platform and version of bsddb used.

Given the current code works, I'm not sure it's worth spending time cleaning this up.  Especially since I plan to move this code to C.  It might be worthwhile to separate out the warning guard.  I can't remember if the __warningregistry__ was mucked with in multiple places.  If it's only used in the one place, I'm not sure that's worth it either.

Learning import is a test of manhood. :-)
msg52373 - (view) Author: Adam Olsen (Rhamphoryncus) Date: 2007-04-24 06:12
test_struct is the only thing that touches __warningregistry__ besides test_warning and (in my other patch) test_py3kwarnings.  Looks like it would benefit from it, and could use some decorators too..

def check_float_coerce(format, number):
check_float_coerce = with_warning_restore(deprecated_err)

*cough*  Paraphrases of religious quotes relating to the creation of decorators aside, I think that'd be better written as a contextmanager.

By the time I'm done I'm going to end up with two dozen patch sets covering 80% of the python codebase x_x.
msg52374 - (view) Author: Ziga Seilnacht (zseil) * (Python committer) Date: 2007-04-24 10:06
Bug #1180193 has a patch with a fix for stale co_filename attributes:
msg315843 - (view) Author: Thomas Kluyver (takluyver) * Date: 2018-04-27 13:54
Time for some bug archaeology!

The reason given for not using frame.f_code.co_filename for warnings was that it can be wrong when the path of .pyc files changes. However, it looks like that was fixed in #1180193 (thanks for the pointer zseil), so I'd like this idea to be reconsidered.

Python uses frame.f_code.co_filename for tracebacks and in pdb, so it's counterintuitive that the warnings module works differently. They are all trying to do the same thing: find and show the bit of code that you're interested in.

This causes problems in IPython with warnings attributed to interactive code: warnings sees it all as part of the __main__ module, so we can't distinguish which input it's pointing to (it's usually obvious, but still...). We have integrated with linecache to make tracebacks work, and I think that if warnings used code.co_filename, it would also work as expected.
msg315845 - (view) Author: Guido van Rossum (gvanrossum) * (Python committer) Date: 2018-04-27 14:41
I recommend opening a new issue (you can link to this one for past discussion).
msg315849 - (view) Author: Thomas Kluyver (takluyver) * Date: 2018-04-27 16:39
Thanks Guido, the new issue is #33375.
Date User Action Args
2018-04-27 16:39:44takluyversetmessages: + msg315849
2018-04-27 14:41:56gvanrossumsetmessages: + msg315845
2018-04-27 13:54:50takluyversetnosy: + takluyver
messages: + msg315843
2007-04-02 05:23:18Rhamphoryncuscreate