Nested class name mangling can be wrong in case of double nesting

[hivert]

In the following class tree:

class Bla(UniqueRepresentation):
    class Bla1(UniqueRepresentation):
        class Bla11:
            Pass
    class Bla2:
        class Bla21:
            Pass

The names are set to

        sage: Bla.Bla1.__name__
        'Bla.Bla1'
        sage: Bla.Bla2.__name__
        'Bla.Bla2'
        sage: Bla.Bla2.Bla21.__name__
        'Bla.Bla2.Bla21'

But

        sage: Bla.Bla1.Bla11.__name__
        'Bla1.Bla11'

whereas one would expect 'Bla.Bla1.Bla11'
This breaks a lot of doc in categories and in particular in functorial constructions.

Apply

• attachment: trac9107_nesting_nested_classes.patch
• attachment: trac_9107_fix_cross_reference.patch

Depends on #12808

CC: @simon-king-jena @zabrocki

Component: categories

Author: Simon King, Nicolas M. Thiéry

Branch: 8b14e05

Reviewer: Volker Braun, Florent Hivert, Travis Scrimshaw

Issue created by migration from https://trac.sagemath.org/ticket/9107

[simon-king-jena]

Dependencies: #12808

[simon-king-jena]

comment:2

I think we should make this depend on #12808, because it cythonises nested classes.

Here is my analysis:

1. In sage.misc.nested_class.modify_for_nested_pickling, only those attributes of a class are (recursively) renamed that are instances of type or of ClassType. However, ironically, instances of NestedClassMetaclass are ignored.
2. It is verified that the name of the to-be-changed class is identical with its name as an attribute of the calling class. But the renaming breaks the identity.

[simon-king-jena]

comment:3

I think the attached patch solves the problem. I get:

sage: class Bla(UniqueRepresentation):
....:     class Bla1(UniqueRepresentation):
....:         class Bla11:
....:             pass
....:     class Bla2:                   
....:         class Bla21:   
....:             pass
....:         
sage: Bla.Bla1.Bla11
<class __main__.Bla.Bla1.Bla11 at 0x46e7808>

The change is in modify_for_nested_pickle, which is called recursively. The idea is that the function should have an extra argument first_run, that is True by default. If the extra argument is False, then it is assumed that it is not applied for the first time.

Here: Since Bla.Bla1 is an instance of NestedClassMetaclass, modify_for_nested_pickle is called on Bla.Bla1.Bla11, resulting in Bla.Bla1.Bla11.__name__=='Bla1.Bla11'. However, since Bla is an instance of NestedClassMetaclass as well, the function is applied to Bla.Bla1 and thus recursively to Bla.Bla1.Bla11 another time.

Now, without my patch, in the second run, modify_for_nested_pickle would be confused by the fact that Bla.Bla1.__dict__ lists Bla.Bla1.Bla11 under the name Bla11, but Bla11.__name__=='Bla1.Bla11'. With my patch, modify_for_nested_pickle expects exactly that naming scheme, and is thus changing Bla.Bla1.Bla11.__name__ into "Bla.Bla1.Bla11".

Much BlaBla, but I think it works...

Potential problems

sage: module = sys.modules['__main__']
sage: getattr(module, 'Bla1.Bla11')                      
<class __main__.Bla.Bla1.Bla11 at 0x46e7808>
sage: getattr(module, 'Bla.Bla1.Bla11')
<class __main__.Bla.Bla1.Bla11 at 0x46e7808>

Hence, Bla.Bla1.Bla11 is listed in the module under two different names. If you think it is bad, then one could probably modify the function when first_run is false, such that the name given in the first run is erased from the module.

Moreover, the reviewer will likely find a speed regression, when excessively creating nested unique representations. But that's hardly realistic...

[simon-king-jena]

Author: Simon King

[simon-king-jena]

comment:4

Another problem: Source inspection does not work yet in the following example.

sage: cython_code = [
... "from sage.structure.unique_representation import UniqueRepresentation",
... "class A1(UniqueRepresentation):",
... "    class B1(UniqueRepresentation):",
... "        class C1: pass",
... "    class B2:",
... "        class C2: pass"]
sage: import os
sage: cython(os.linesep.join(cython_code))
sage: A1.B1.C1??          
Error getting source: class A1.B1.C1 has no attribute '__class__'
Type:		classobj
String Form:	_mnt_local_king__sage_temp_mpc622_6475_tmp_0_spyx_0.A1.B1.C1
Namespace:	Interactive
Loaded File:	/mnt/local/king/.sage/temp/mpc622/6475/spyx/_mnt_local_king__sage_temp_mpc622_6475_tmp_0_spyx/_mnt_local_king__sage_temp_mpc622_6475_tmp_0_spyx_0.so
Source File:	/mnt/local/king/.sage/temp/mpc622/6475/spyx/_mnt_local_king__sage_temp_mpc622_6475_tmp_0_spyx/_mnt_local_king__sage_temp_mpc622_6475_tmp_0_spyx_0.so

Even #11768 does not solve the problem.

Shall that be dealt with on a different ticket? Or in one go?

Probably on a different ticket, since I just find that even source inspection for A1 (which has a usual name) does not work...

[simon-king-jena]

comment:5

For the record: If #11791 is applied after this ticket, source inspection in the example above works (and is doctested).

[simon-king-jena]

comment:6

Is there a reviewer to fix name mangling of nested classes (needed in the category framework)?

[saliola]

comment:7

This patch also fixes an issue that came up in #8899 regarding documentation of nested classes not appearing in the reference manual.

See here for a description of the issue, see the thread on sage-combinat-devel.

See here for the confirmation that this works: #8899 comment:31

[vbraun]

comment:8

LGTM!