GH-93657: More LOAD_ATTR specializations

[Fidget-Spinner]

Specializes for

• Mutable class attributes with descriptors (e.g. enum.Enum members)
• @property() with Python functions

[markshannon]

Specializing property accesses is clearly worthwhile.

I'm not so sure about mutable descriptors. Do they represent a large enough fraction of attributes, and is the specialized form that much faster?

[markshannon]

Is this common enough to justify its own specialization, especially as we need to do much of the work of general lookup?

[Fidget-Spinner]

According to the stats on the faster-cpython repo, this represents 5.7% of the remaining failures. In contrast, properties are 13.8%. Class method objects are quite close, at 6.2%.

Here's a microbenchmark, with no PGO nor LTO, but it's a Windows release build (similar to -O3).

python -m timeit -s "
from enum import Enum

class Colours(Enum):
 RED = 1
" "Colours.RED"

# on MAIN:
2000000 loops, best of 5: 121 nsec per loop
2000000 loops, best of 5: 120 nsec per loop
2000000 loops, best of 5: 120 nsec per loop

# on this PR:
5000000 loops, best of 5: 77.2 nsec per loop
5000000 loops, best of 5: 73.5 nsec per loop
5000000 loops, best of 5: 72.1 nsec per loop

With PGO and LTO, I expect the difference to be less dramatic. But I still expect this form to be ~=15% faster than the base version.

[markshannon]

Do you have a non-enum example?

enums members should be simple class attributes.
#93910

[Fidget-Spinner]

I don't. This opcode was specifically designed with enum in mind.

[cpython-cla-bot]

All commit authors signed the Contributor License Agreement.

[Fidget-Spinner]

I'm quite surprised that the property specialisation's speedup is less than 2x:
(no PGO, no LTO, release build)

python -m timeit -s "
class X:
 y = property(lambda s: 10)

x = X()
" "x.y"

# MAIN
5000000 loops, best of 5: 40.2 nsec per loop
5000000 loops, best of 5: 39 nsec per loop
5000000 loops, best of 5: 38.5 nsec per loop

# this PR
10000000 loops, best of 5: 27.5 nsec per loop
10000000 loops, best of 5: 27.2 nsec per loop
10000000 loops, best of 5: 27 nsec per loop

Maybe your work in #93897 shall fix that.

[markshannon]

Given that #93910 is likely to be messing up our stats, maybe put this hold for a while?
Or, if you want to make a PR with just the property specialization, we could merge that.

[Fidget-Spinner]

Thanks to #94614 and #93910 we don't need this anymore.