Finish up LOAD_ATTR specialisation

[Fidget-Spinner]

We should target the following specialisation failures:

• has managed dict
• not in dict (how does this even happen??)

With these two we can hit >90% specialisation successes.
If we're feeling really ambitious, we could aim for "not managed dict" failure too. But we don't need that to achieve >90% successes.

I'm doing the first one. Is anyone interested in investigating the second specialisation failure?

Linked PRs

• gh-100288: Specialise LOAD_ATTR_METHOD for managed dictionaries #100289
• Revert "gh-100288: Specialise LOAD_ATTR_METHOD for managed dictionaries (GH-100289)" #100468
• gh-100288: Specialise LOAD_ATTR_METHOD for managed dictionaries #100492
• GH-100288: Remove LOAD_ATTR_METHOD_WITH_DICT instruction. #100753
• GH-100288: Skip extra work when failing to specialize LOAD_ATTR #101354
• GH-100288: Specialize class attribute loads #101379
• GH-100288: Specialize LOAD_ATTR for simple class attributes. #105990
• GH-100288: regen cases after #105990 #106589

[corona10]

Oh I am interested in if you don't mind?

[Fidget-Spinner]

Sure! Go ahead!

[Fidget-Spinner]

CC @markshannon just these two added specialisations should net us 10% all specialisation failures/successes, bringing us to over 90% specialisation successes.

[Fidget-Spinner]

@corona10 seems like the stats were misleading, not in dict is actually the following:

class C: 
    i = 0

c = C()
c.i # this fails to specialise

So it's a class attribute lookup, which makes more sense.

[corona10]

Thanks I read the PR too: #100295

[markshannon]

Regarding "not managed dict" failures, I think the approach should be to reduce the number of objects that don't have managed dicts, not specialize for them.

We can tweak the management of cached keys, and do some static analysis in the compiler to do that.