[Jython-dev] Re: new-style transition questions

[Samuele P. <ped...@st...>]

Brian Zimmer wrote:
> Samuele,
> 
> I have a couple of questions about how we should handle this transition 
> period for new-style classes.
> 
> I cannot pass the full test_descrtut for list because it subclasses from 
> PySequence which is an old-style class (it has an classDictInit and no 
> typeSetup).  Because of this, some methods (__iter__, __nonzero__) which 
> exist in the classDictInit call are exposed in the dict for the list 
> instance.
> 
> Until all subclasses of PySequence are converted I don't want to remove 
> the classDictInit method but I also don't want it called for instances 
> of subclasses with typeSetup methods.  It looks like I can reasonably do 
> this in PyType.fillFromClass but I'm wondering if it's worth the 
> effort.  Should I just wait until the subclasses are done and then 
> switch it over to a new-style class?
> 

PySequence should not be visible as a base class of list from Python.

list.__bases__ should be (object,)

A type_base_class directive should do that.

OTOH what we likely want is a way to instead of having some list_* 
method have corresponding seq_* methods in PySequence that can be reused 
from its subclasses typeSetups, I'm working adding the expressivity to 
specify this to the code generator.


> Also, I was looking at making all the builtin methods in __builtin__ be 
> instances of PyBuiltinFunction.  Does this sound reasonable?  

yes

> I have a 
> script which will parse __built__ and generated the appropriate 
> PyBuiltinFunction instances.  Right now they are mixture of 
> .BuiltinFunctions and ReflectedFunctions, do we care about the 
> distinction?  In CPython, they are the same type.
> 
> Which brings me to another question.  Should BuiltinMethodType in types 
> be a PyBuiltinFunction or continue to be a PyMethod?

yes, PyBuiltinFunction is meant to be much more similar to Python

types.BultinFunctionType == types.BuiltinMethodType == <type 
'builtin_function_or_method'>

ideally I would like if we were able to deprecate PyBuiltinFunctionSet.




> 
> I also looked into implementing __slots__.  How far should we take the 
> full slots approach in Jython?  Should we go all out of the space 
> savings or just make them 'feel' like CPython objects with __slots__ in 
> that they can't grow an attribute?

my plan was to have a PyObject[] array in the derived classes for the 
purpose of slots. It would be null if there are no slots. Notice that 
the precise semantics for slots vs having a dictionary are a bit tricky,
for example the slot name __dict__ is special cased

 >>> class D(object):
...   __slots__ = ('__dict__',)
...
 >>> d=D()
 >>> d.a=3

etc.

Samuele.