Re: [Jython-dev] Exposed types not in CoreExposed.includes

[Jeff A. <ja...@fa...>]

Good. There has been so much in PyType I thought I understood but broke. 
This could easily, in my mind, have been another. I've corrected as many 
types as I've come across, and it passes tests, but I don't quite know 
what I'm doing here. (I had notes from the first time I learned it, but 
they were on the Wiki.)

I've just pushed recent work to 
https://bitbucket.org/tournesol/jython-fgtype . I still have a few more 
ideas for things I can break ...

Jeff


On 26/08/2017 16:51, Jim Baker wrote:
> Jeff,
>
> Agreed that these all must be an oversight. I specifically contributed 
> PyXRangeIter per the blame info, and I know I would have added this to 
> CoreExposed.includes. Not being exposed can impact some aspects of 
> Python integration, eg right now the __doc__ string is not being 
> exposed properly, but most likely the key thing that is impacted with 
> these specific classes is possibly higher calling overhead due to 
> reflection. So that's why we haven't caught it until now.
>
> Thanks for systematically digging into this!
>
> I also plan to respond to your earlier analysis about using 
> ClassValue. I was busy, then on vacation, then busy again. Summer!
>
> Lastly I was hopeful that exposing this iterator would see at least 
> some impact in a tight loop, but it's not visible in a trivial 
> benchmark like
>
> jython -m timeit -n 100000 'def f():' ' c = 0' ' for i in 
> xrange(10000):' '  c += 1' 'f()'
>
> It might be worthwhile trying https://github.com/zware/codespeed, but 
> obviously a secondary concern. Any performance tuning with respect to 
> function calling — and the use of fun stuff like InvokeDynamic and 
> SwitchPoint — should go into Jython 3.
>
> - Jim
>
> On Sat, Aug 26, 2017 at 6:46 AM, Jeff Allen <ja...@fa... 
> <mailto:ja...@fa...>> wrote:
>
>     Tinkering with the logic of PyType, I have run across cases where
>     a class is annotated @ExposedType but does produce a TypeBuilder
>     when initialised. This is because they are not listed in
>     CoreExposed.includes, are not processed by the exposer, and
>     therefore are represented by a  PyJavaType object.
>
>     I am experimenting with using the annotation, rather than static
>     initialisation and looking for a builder, as my way to distinguish
>     the two cases, and which reveals this discrepancy whe addBuilder
>     is not called.
>
>     So far I found: PyNullImporter, org.python.modules.itertools.imap,
>     org.python.antlr.ast.SetComp, org.python.antlr.ast.SetComp,
>     PyXRangeIter.
>
>     Such cases are just an oversight, not some subtle choice I haven't
>     understood, right?
>
>     Jeff
>
>