gh-91247: improve performance of list and tuple repeat (with specialization for n=1)

[eendebakpt]

Compared to current main branch there are 3 improvements

• For the list inplace repeat there is an optimized copy method
• This PR solves the inefficient pointer chasing for small (size 2 to 7) lists and tuples
• It shares code between the implementations of list repeat, tuple repeat and list inplace repeat

This PR keeps the specializations for repeats of lists and tuples with length 1. An alternative PR without specialization for n=1 (#32045) was rejected because the specializations might yield a speedup.

For performance comparisons to main and the version without specializations and more detailed discussion see issue #91247.

The inefficient pointer chasing can be seen in the following performance benchmark:

(results obtained with runner.timeit(name=f"list({n}) repeat {r}", stmt=f"x=a*{r}", setup=f'a=list(range({n}))'))

• Issue: Improve performance of list and tuple repeat methods #91247

[eendebakpt]

@sweeneyde Could you have a look at this ticket? The approach is similar to #31856 and #31999 which you reviewed.

[sweeneyde]

@eendebakpt Sorry this slipped under my radar

@vstinner, would you mind reviewing? I think the change is good in principle but I want another reviewer and to make sure the header file changes are appropriate.

[kumaraditya303]

Minor nits

[sweeneyde]

Some minor nits, but this is looking good. Thanks for your patience and persistence on this.

[sweeneyde]

I verified a benchmark on a Windows machine since this might be a bit compiler-specific.

pyperf timeit -s "a = [0,1]" "a * 10_000"
Mean +- std dev: [before] 32.9 us +- 0.3 us -> [after] 21.8 us +- 0.3 us: 1.51x faster

Looks great!

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @sweeneyde for commit 516f091 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[sweeneyde]

Note to self: this cannot overflow because list_new_prealloc uses PyMem_New, and so this multiplication already succeeded at some point.