gh-123745: Improve performance of deepcopy by setting keep alive key in the memo

[eendebakpt]

Benchmark:

import pyperf
runner = pyperf.Runner()

setup="""
import copy

a={'list': [1,2,3,43], 't': (1,2,3), 'str': 'hello', 'subdict': {'a': True}}

from dataclasses import dataclass

lst = [1, 's']

@dataclass
class A:
    a : int
    
dc = A(123)
list_dc = [A(1), A(2), A(3), A(4)]
"""

runner.timeit(name="deepcopy dict", stmt=f"b=copy.deepcopy(a)", setup=setup)
runner.timeit(name="deepcopy dataclass", stmt=f"b=copy.deepcopy(dc)", setup=setup)
runner.timeit(name="deepcopy small list", stmt=f"b=copy.deepcopy(lst)", setup=setup)
runner.timeit(name="deepcopy list dataclasses", stmt=f"b=copy.deepcopy(list_dc)", setup=setup) 

Results:

deepcopy dict: Mean +- std dev: [main] 11.3 us +- 0.9 us -> [pr] 10.2 us +- 0.1 us: 1.11x faster
deepcopy dataclass: Mean +- std dev: [main] 8.62 us +- 0.29 us -> [pr] 7.74 us +- 0.30 us: 1.11x faster
deepcopy small list: Mean +- std dev: [main] 2.48 us +- 0.10 us -> [pr] 2.20 us +- 0.31 us: 1.13x faster
deepcopy list dataclasses: Mean +- std dev: [main] 34.1 us +- 1.3 us -> [pr] 32.9 us +- 1.0 us: 1.03x faster

Geometric mean: 1.10x faster

See the corresponding issue for further details.

• Issue: Improve performance of deepcopy by setting keep alive key in the memo #123745

[eendebakpt]

Updated benchmarks (new machine, but roughly the same)

deepcopy int: Mean +- std dev: [main] 47.1 ns +- 0.3 ns -> [pr] 43.2 ns +- 0.2 ns: 1.09x faster
deepcopy dict: Mean +- std dev: [main] 2.58 us +- 0.02 us -> [pr] 2.38 us +- 0.02 us: 1.08x faster
deepcopy dataclass: Mean +- std dev: [main] 2.28 us +- 0.02 us -> [pr] 2.11 us +- 0.02 us: 1.08x faster
deepcopy small list: Mean +- std dev: [main] 615 ns +- 2 ns -> [pr] 494 ns +- 3 ns: 1.25x faster
deepcopy list of dataclasses: Mean +- std dev: [main] 9.19 us +- 0.09 us -> [pr] 8.89 us +- 0.07 us: 1.03x faster

Geometric mean: 1.10x faster