bpo-38045: Improve the performance of _decompose() in the enum.py

[hongweipeng]

The original code will access each element of _value2member_map_ and check if the value is the power of 2.

After optimization, access the elements in _value2member_map_ as needed. In the initial example:

from enum import Flag
F = Flag("F", list("abcdefghijklm"))
for idx in range(2**len(F) - 1):
    F(idx)

the speedup is more than 100x .(The same for replacing Flag with IntFlag).

https://bugs.python.org/issue38045

[corona10]

@hongweipeng Can you show me the benchmark result?

[hongweipeng]

Use the initial example:

from enum import Flag
F = Flag("F", list("abcdefghijklm"))
for idx in range(2**len(F) - 1):
    F(idx)

Before patch:

# time ./python ts.py

real	0m29.639s
user	0m29.615s
sys	0m0.020s

After patch:

# time ./python ts.py

real	0m0.135s
user	0m0.122s
sys	0m0.014s

[corona10]

@ethanfurman

For core-developers,
I didn't check this PR is the correct way.
But seems like there is performance enhancement with this PR.

I left the benchmark for the core-developers

Baseline

 ./python.exe -m pyperf timeit -s 'from enum import Flag' 'F = Flag("F", list("abcdefghijklm"))' 'for idx in range(100):' '   F(idx)'
.....................
Mean +- std dev: 24.5 ms +- 0.3 ms

PR

./python.exe -m pyperf timeit -s 'from enum import Flag' 'F = Flag("F", list("abcdefghijklm"))' 'for idx in range(100):' '   F(idx)'
.....................
Mean +- std dev: 7.30 ms +- 0.19 ms

[ethanfurman]

GitHub seem stuck on bedevere/news -- closing and reopening...