homepage

I just found about the smelly build target. It checks for smelly exports.
$ make smelly
[...]
nm -p libpython2.6.a | \
 sed -n "/ [TDB] /s/.* //p" | grep -v "^_*Py" | sort -u; \
asdl_int_seq_new
asdl_seq_new
init_ast
init_codecs
initerrno
initgc
initimp
initposix
initpwd
initsignal
init_sre
init_symtable
initthread
initxxsubtype
initzipimport
nm -p libpython3.0.a | \
 sed -n "/ [TDB] /s/.* //p" | grep -v "^_*Py" | sort -u; \
_add_one_to_index_C
_add_one_to_index_F
asdl_int_seq_new
asdl_seq_new
These should be checked and fixed.

Are you suggesting that the init* functions in Python 2.6 should not
longer be exported? Why? And how?

I just suggest that somebody should look into it. I neither know how
critical the issue is nor how to change them. This bug report serves as
a reminder.

For Python 2.x you cannot change the "main" function name of C
extensions, since this is used by the Python import mechanism.

I guess you *could* change the name of extensions that become builtin
modules. However, I think it is not worth the trouble, especially since
PEP 3121 solves the problem for good, for Python 3.
I propose to simply filter out init[_a-z]+ from the set of "bad" symbols.
That leaves us with asdl_*_seq_new, and _add_one_to_index_; the latter
is already reported as issue3101.

There is no proposed patch yet, so removing 2.5.3 from the target versions.

#5591 is a duplicate of this.

In trunk:
_add_one_to_index_C
_add_one_to_index_F
asdl_int_seq_new
asdl_seq_new
init_ast
init_codecs
initerrno
initgc
initimp
initposix
initpwd
initsignal
init_sre
init_symtable
initthread
initxxsubtype
initzipimport
In py3k:
_add_one_to_index_C
_add_one_to_index_F
asdl_int_seq_new
asdl_seq_new

Re: msg #77350:
> I propose to simply filter out init[_a-z]+ from the set of "bad" 
> symbols.
I'm attaching a patch (to trunk) to Makefile.pre.in which filters out such symbols.
Relevant part of output of "make smelly" on my svn build with this patch:
nm -p libpython2.7.a | \
		sed -n "/ [TDB] /s/.* //p" | grep -v "^_*Py" | \
		grep -v -E "^init[_a-z]+" | sort -u; \
_add_one_to_index_C
_add_one_to_index_F
asdl_int_seq_new
asdl_seq_new

py3k is much cleaner than python 2, due to the change in the module API.
Here's "make smelly"'s output on a recent py3k checkout:
nm -p libpython3.2.a | \
 sed -n "/ [TDB] /s/.* //p" | grep -v "^_*Py" | sort -u; \
_add_one_to_index_C
_add_one_to_index_F
asdl_int_seq_new
asdl_seq_new

Patch to py3k which adds the "_Py" prefix to the four listed symbols.
With this, the output from "make smelly" is clean (odorless, perhaps?).
However, adding _Py does seem to go against this comment in Include/asdl.h:
/* It would be nice if the code generated by asdl_c.py was completely
 independent of Python, but it is a goal the requires too much work
 at this stage. So, for example, I'll represent identifiers as
 interned Python strings.
*/

For Python 3.3 and 3.4 make smelly still lists two exported symbols.
asdl_int_seq_new
asdl_seq_new
Do we have to check the modules, too? Some of them like _ctypes and _decimal export a lot of symbols. Here is a small shell snippet for Makefile.
 for MOD in `find $(srcdir)/build -name '*.s[ol]' ` ; do \
 EXPORT=`nm -p $$MOD | sed -n "/ [TDB] /s/.* //p" | \
 grep -E -v "^_*Py|^_init|^_fini" | sort -u`; \
 if [ -n "$$EXPORT" ]; then \
 echo $$MOD; \
 echo $$EXPORT; \
 fi \
 done

I'd like to focus this issue; it has been open long enough, and deserves to get closed once the original issue is resolved - which was that "make smelly" reports symbols. I think dmalcolm's patch is quite a good start for that.
It may well be that other modules need to be considered - but PLEASE not in this issue. For Unix with shared libraries, the extension modules certainly cause problems, in particular if Python gets embedded in a context (e.g. Apache) that also loads separate copies of the same libraries, and may cause problems if one of our functions collide with some library. However, I see *two* issues falling out of this: one to extend the "make smelly" target to include extension modules, and the other to then fix the extension modules - preferably with one issue per extension module.

This very simple patch should certainly be applied.

New changeset 142c62a490ce by Antoine Pitrou in branch 'default':
Issue #4555: All exported C symbols are now prefixed with either "Py" or "_Py".
http://hg.python.org/cpython/rev/142c62a490ce