gh-91247: improve performance of list and tuple repeat

[eendebakpt]

Improve the performance of the list repeat methods by reducing the number of reference count operations and copying data using memcpy. The approach to reduce the number of memcpy invocations is similar to #31999, but due to the handling of reference counts in the list and tuple repeat methods the code is slightly different.

Note: the specialization for the case of 1 item in a list or tuple can be removed with (almost) no loss of performance. More details and a comparison against the version with specializations (#91482) are the issue #91247.

https://bugs.python.org/issue47091

[eendebakpt]

Microbenchmark (main against commit 8b2cc9c):

list(3) repeat 1: Mean +- std dev: [base] 142 ns +- 4 ns -> [pr2] 136 ns +- 2 ns: 1.04x faster
list(1) repeat 1: Mean +- std dev: [base] 59.0 ns +- 0.7 ns -> [pr2] 55.9 ns +- 2.1 ns: 1.05x faster
list(3) repeat inplace 1: Mean +- std dev: [base] 74.5 ns +- 0.5 ns -> [pr2] 72.4 ns +- 0.8 ns: 1.03x faster
tuple(4) repeat 1: Mean +- std dev: [base] 46.6 ns +- 1.1 ns -> [pr2] 48.0 ns +- 0.8 ns: 1.03x slower
list(100) repeat 2: Mean +- std dev: [base] 1.28 us +- 0.00 us -> [pr2] 1.26 us +- 0.01 us: 1.02x faster
list(3) repeat 2: Mean +- std dev: [base] 156 ns +- 3 ns -> [pr2] 146 ns +- 3 ns: 1.07x faster
list(3) repeat inplace 2: Mean +- std dev: [base] 106 ns +- 1 ns -> [pr2] 105 ns +- 3 ns: 1.01x faster
tuple(4) repeat 2: Mean +- std dev: [base] 146 ns +- 0 ns -> [pr2] 141 ns +- 2 ns: 1.03x faster
list(100) repeat 10: Mean +- std dev: [base] 4.35 us +- 0.05 us -> [pr2] 4.23 us +- 0.02 us: 1.03x faster
list(3) repeat 10: Mean +- std dev: [base] 267 ns +- 3 ns -> [pr2] 192 ns +- 5 ns: 1.39x faster
list(1) repeat 10: Mean +- std dev: [base] 68.8 ns +- 2.4 ns -> [pr2] 75.7 ns +- 1.8 ns: 1.10x slower
list(3) repeat inplace 10: Mean +- std dev: [base] 143 ns +- 1 ns -> [pr2] 129 ns +- 3 ns: 1.10x faster
tuple(4) repeat 10: Mean +- std dev: [base] 240 ns +- 3 ns -> [pr2] 269 ns +- 3 ns: 1.12x slower
list(100) repeat 1000: Mean +- std dev: [base] 409 us +- 1 us -> [pr2] 421 us +- 1 us: 1.03x slower
list(3) repeat 1000: Mean +- std dev: [base] 18.8 us +- 0.2 us -> [pr2] 5.42 us +- 0.28 us: 3.48x faster
list(1) repeat 1000: Mean +- std dev: [base] 2.01 us +- 0.02 us -> [pr2] 1.96 us +- 0.02 us: 1.03x faster
list(3) repeat inplace 1000: Mean +- std dev: [base] 4.93 us +- 0.16 us -> [pr2] 2.60 us +- 0.02 us: 1.90x faster
tuple(4) repeat 1000: Mean +- std dev: [base] 9.80 us +- 0.29 us -> [pr2] 6.98 us +- 0.47 us: 1.40x faster

Benchmark hidden because not significant (3): list(100) repeat 1, list(1) repeat 2, [control] list pop+append

Geometric mean: 1.14x faster

Code for benchmark

``` import pyperf runner = pyperf.Runner()

setup='a=[1,2,3]; a1=[1,]; t=(1,2,3,4)'

for n in [1,2,10,1000]:
runner.timeit(name=f"list(100) repeat {n}",
stmt=f"x=a*{n}; y=a*{n}",
setup=f'a=[1.]*100')

runner.timeit(name=f"list(3) repeat {n}",
          stmt=f"x=a*{n}; y=a*{n}",
          setup=setup)

runner.timeit(name=f"list(1) repeat {n}",
          stmt=f"x=a1*{n};",
          setup=setup)

runner.timeit(name=f"list(3) repeat inplace {n}",
          stmt=f"a=[1,2,None]; a*={n}",
          setup=setup)
          
runner.timeit(name=f"tuple(4) repeat {n}",
          stmt=f"x=t*{n}; y=t*{n}",
          setup=setup)

runner.timeit(name=f"[control] list pop+append",
stmt=f"x=a.pop(); a.append(1)",
setup=setup)

</details>

[ghost]

Commit authors are required to sign the Contributor License Agreement.

[MaxwellDupre]

414 tests OK.
1 test failed:
test_embed
I don't think this is related, hence with most test passing looks ok.

[eendebakpt]

@MaxwellDupre Thanks for the approval. Note that this PR was marked draft. There are two PRs to resolve #91247, this one and #91482.

I have a small preference for #91482 (but this PR is also an improvement)

[eendebakpt]

Closing as #91482 seems a better approach