Port LOAD_ATTR from old opcache to PEP 659 adaptive interpreter

[markshannon]

There are many specializations for LOAD_ATTR that make sense. LOAD_ATTR in 3.10 only handles two, so we should have an easy win here.

Possible specializations are:

• Attribute is in instance dictionary
• Attribute is in a slot
• Attribute is a non-descriptor attribute of the class
• Attribute is a property
• Attribute is method
• The owner is a module

We will want dictionary keys versioning for several of the above.

[markshannon]

• Add dict keys version
• Add LOAD_ATTR_ADAPTIVE and two specializations corresponding to existing LOAD_ATTR
• Split apart specialization for instance dicts into "known keys" and "probable index"
• Add specializations for module and class attributes

Specializing for properties needs support from call specialization machinery, so will need to wait for that.

[nilansaha]

Not sure if this is the right place for this but have you guys thought about adding a workflow that automatically benchmarks any PRs opened from this fork and reports it ? Could be really interesting to see speedups as they happen no matter how small. I think the entire Python community is incredibly excited for what is going down here.

[gvanrossum]

Probably better to open a separate issue. FWIW benchmarks are noisy enough that I think this isn’t going to be all that useful.

[bluss]

@nilansaha @gvanrossum Can the existing speed.python.org infrastructure be used for this? Ideally all the merged PRs would be followed-up there. Having a bot that can run the same benchmark set on a PR before it gets merged, would also be a great flow.