Improve performance of find_max_char

[ruema]

Feature or enhancement

Proposal:

find_max_char is called each time a string is created.
By reducing the amount of tests, the performance can be improved considerably.

python -m pyperf timeit -s "b=('a' *1000+ '\u019f'*2)" "b[:-1]"

+-----------+--------+----------------------+
| Benchmark | ref    | patch                |
+===========+========+======================+
| timeit    | 442 ns | 388 ns: 1.14x faster |
+-----------+--------+----------------------+

Has this already been discussed elsewhere?

This is a minor feature, which does not need previous discussion elsewhere

Links to previous discussion of this feature:

No response

Linked PRs

• gh-122901: Improve performance of find_max_char #122902

[terryjreedy]

@vstinner Should you or someone else look at this? Should OP say more about idea of patch? Is the above the proper speed test?

[picnixz]

I would also look at the arguments in #120212.

Personally, I would include in the benchmarks:

• ASCII data only
• lots of non-ASCII data
• a bit of non-ASCII data
• short, medium and very long strings

By the way, is your build a release or a PGO (maybe with LTO) build?

[vstinner]

@ruema: Would you mind to run more benchmarks? See previous @picnixz comment about that.

[rhpvorderman]

This is basically the same as #120212 but with larger chunks. Given that this approach unifies the find_max_char functions for all
char sizes, from a maintainability standpoint it is better than #120212.

I would add benchmarks for line by line reading. for line in file_descriptor is a very common pattern.

[vstinner]

Would it be possible to have a few more benchmarks on this change to have an idea if it's worth it or not?