gh-145481: Optimize frozendict lookup in free threading

[vstinner]

• Issue: Optimize frozendict lookup on Free-Threaded build #145481

[vstinner]

According the benchmark below, frozendict[key] lookup scales way better with this change: up to 33x faster.

Benchmark results:

Benchmark	nogil	nogil_optim
str keys	334 ms	10.0 ms: 33.21x faster
Geometric mean	(ref)	5.81x faster

Benchmark hidden because not significant (1): int keys

Benchmark results compared to the regular build (ref):

Benchmark	ref	nogil	nogil_optim
int keys	8.36 ms	7.40 ms: 1.13x faster	7.27 ms: 1.15x faster
str keys	231 ms	334 ms: 1.45x slower	10.0 ms: 22.94x faster
Geometric mean	(ref)	1.13x slower	5.14x faster

Python built in release mode, gcc -O3 (without PGO, LTO, or CPU isolation).

Benchmark run on my Fedora 43 laptop which has 12 threads (6 CPU cores).

[colesbury]

The string benchmark has really bad scaling because the strings aren't interned so we have reference count contention, plus some other things means that we also fall back to the locking code path.

Most dicts (and presumably frozendicts) use string literals as keys, which are interned and immortalized and avoid this problem.

You can see this if you update the benchmark:

frozendict_str_keys = frozendict({'0': 0, '1': 1, '2': 2, '3': 3, '4': 4,
                                  '5': 5, '6': 6, '7': 7, '8': 8, '9': 9})

On my machine, I'm not seeing an improvement with this PR for string literals keys or ints. In fact, it's slightly worse, which is surprising. (Maybe something weird with branch predictor behavior?)

main (671a953)

...........
int keys: Mean +- std dev: 6.24 ms +- 0.20 ms
...........
str keys: Mean +- std dev: 6.52 ms +- 0.24 ms

PR + frozendict_str_keys literals

...........
int keys: Mean +- std dev: 6.59 ms +- 0.33 ms
...........
str keys: Mean +- std dev: 6.74 ms +- 0.26 ms