Improve LOAD_ATTR specializations

[sweeneyde]

I was trying to figure out some of the situations in which LOAD_ATTR specializations can fail.

Here is a list of specializations stats by benchmark, along with a closer look at of some of the lowest-hit-percentage benchmarks: https://gist.github.com/sweeneyde/a0d3abfc99308b93c1cee084684f90e0

It seems there could be a benefit in specializing something to do with the type_getattro (i.e., type.__getattr(ibute)__ in python), as many lookups of class attributes (whether from the class directly or indirectly through an instance) are not currently specialized.

[markshannon]

When you say something to do with type.__getattribute__, do you mean for attributes of the class access via an instance or on the class directly, or both?
i.e.

class C:
     a = NonDescriptor()
c = C()

c.a # This?
C.a # Or this? (or both?)

Two questions regarding the other failures:

1. Are we missing any important failure modes in our SPEC_FAIL_... list?
2. Do any of the failure modes in the SPEC_FAIL_... list not readily specialized, that could be if we split them up differently?

[sweeneyde]

attributes of the class access via an instance or on the class directly, or both?

Both:

• Sympy benchmarks have a decent amounts of:

  • Class.class_attr
  • Class.__name__ (data descriptor)
  • Class.some_static_method
  • instance.class_attribute_of_some_superclass

• Crypto_pyaes and scimark_monte_carlo:

  • failures are mostly instance.class_attribute

Regarding questions 1 and 2, right now SPEC_FAIL_OVERRIDDEN is too broad: it currently only says that type->tp_getattro != PyObject_GenericGetAttr. It would be good to distinguish (and probably specialize) the case type->tp_getattro == PyType_Type.getattro.

There's also the case for instance.class_attr where _PyDictKeys_StringLookup(keys, name); returns DKIX_EMPTY, so index != (uint16_t)index. Currently, this leads to SPEC_FAIL_OUT_OF_RANGE, which could be a bit misleading, and perhaps there could be further specializations for when the key is absent from the instance dict but present in the class dict.

[Fidget-Spinner]

For anyone interested, here's my (mostly dead) attempt at Class.class_attr specialization python/cpython#29146. I tried two different approaches (one using tp_version_tag of Class, another using tp_mro's ID). python/cpython#29146 (comment)

Some shortcomings:

1. Doesn't seem to specialize well on the standard test suite, not sure about pyperformance.
2. Uses more memory (I'm sure given enough time, someone smarter than me will have a more elegant solution which fixes this :)