> IMO you need to implement LOAD_METHOD support for all kinds of calls, including the ones that use kwargs, to see any improvement.

Recently I played around with that idea and extended LOAD/CALL_METHOD to keyword arguments (so CALL_FUNCTION_KW is replaced). I then reapplied Pablo's patch. Here's the pyperformance results (LTO + PGO):

> pyperf compare_to 2021-05-08_03-25-general_load_method-42fcad26a487.json.gz 2021-05-08_08-50-general_load_method-dd24d58ce940.json.gz -G --min-speed=1

Slower (5):
- pathlib: 50.3 ms +- 0.8 ms -> 51.9 ms +- 1.1 ms: 1.03x slower
- unpack_sequence: 141 ns +- 2 ns -> 144 ns +- 7 ns: 1.02x slower
- telco: 16.6 ms +- 0.6 ms -> 16.9 ms +- 0.5 ms: 1.02x slower
- unpickle: 33.4 us +- 0.5 us -> 33.9 us +- 0.8 us: 1.01x slower
- nqueens: 239 ms +- 2 ms -> 242 ms +- 2 ms: 1.01x slower

Faster (23):
- logging_silent: 438 ns +- 16 ns -> 411 ns +- 18 ns: 1.07x faster
- go: 572 ms +- 55 ms -> 539 ms +- 8 ms: 1.06x faster
- pickle_pure_python: 1.10 ms +- 0.01 ms -> 1.06 ms +- 0.02 ms: 1.04x faster
- meteor_contest: 274 ms +- 4 ms -> 265 ms +- 4 ms: 1.03x faster
- logging_simple: 20.6 us +- 0.3 us -> 19.9 us +- 0.4 us: 1.03x faster
- hexiom: 23.0 ms +- 1.0 ms -> 22.2 ms +- 0.2 ms: 1.03x faster
- mako: 37.0 ms +- 0.5 ms -> 35.9 ms +- 0.4 ms: 1.03x faster
- json_dumps: 32.3 ms +- 0.4 ms -> 31.4 ms +- 0.3 ms: 1.03x faster
- scimark_sor: 496 ms +- 7 ms -> 484 ms +- 7 ms: 1.03x faster
- spectral_norm: 372 ms +- 5 ms -> 363 ms +- 4 ms: 1.02x faster
- richards: 177 ms +- 5 ms -> 173 ms +- 4 ms: 1.02x faster
- django_template: 114 ms +- 1 ms -> 112 ms +- 1 ms: 1.02x faster
- dulwich_log: 170 ms +- 2 ms -> 167 ms +- 2 ms: 1.02x faster
- pyflate: 1.65 sec +- 0.02 sec -> 1.62 sec +- 0.02 sec: 1.02x faster
- xml_etree_iterparse: 282 ms +- 5 ms -> 278 ms +- 3 ms: 1.02x faster
- unpickle_list: 12.1 us +- 0.1 us -> 11.9 us +- 0.2 us: 1.02x faster
- nbody: 343 ms +- 5 ms -> 338 ms +- 7 ms: 1.02x faster
- logging_format: 22.6 us +- 0.4 us -> 22.3 us +- 0.5 us: 1.01x faster
- sympy_str: 833 ms +- 7 ms -> 822 ms +- 7 ms: 1.01x faster
- float: 271 ms +- 3 ms -> 268 ms +- 5 ms: 1.01x faster
- scimark_fft: 1.05 sec +- 0.01 sec -> 1.03 sec +- 0.01 sec: 1.01x faster
- deltablue: 17.9 ms +- 0.3 ms -> 17.6 ms +- 0.4 ms: 1.01x faster
- scimark_lu: 403 ms +- 9 ms -> 399 ms +- 9 ms: 1.01x faster

Benchmark hidden because not significant (30): (...)
Geometric mean: 1.01x faster

Most benchmarks don't use keyword arguments. I'm going to try extend LOAD_METHOD to *args and **kwargs next.

For anyone interested, here's the diff (unstable) https://github.com/python/cpython/compare/main...Fidget-Spinner:general_load_method

I hope this helps :).