bpo-46939: Specialize calls to Python classes (POSTCALL edition)

[Fidget-Spinner]

https://bugs.python.org/issue46939

[Fidget-Spinner]

Final alternative method, if we don't want all the stack manipulation: PRECALL_PY_CLASS shall incref self, then set oparg of POST_CALL to 0 to signal it to return self. Then POST_CALL accesses into above the TOS (into the already decrefed arguments) to find arg[0] (self).

Downside is that decref during error handling is tricky. Current method of pushing things onto the stack handles that for us.

[markshannon]

Have you benchmarked this?
I'm worried that the extra 6 bytes per call, plus the extra code, is going to cost more than the specialization gains.

[Fidget-Spinner]

Have you benchmarked this? I'm worried that the extra 6 bytes per call, plus the extra code, is going to cost more than the specialization gains.

Not yet. I plan to do so tomorrow (your time). The results should be available by friday.

[Fidget-Spinner]

compare_to RETURN_VALUE hack PR -> POSTCALL:

Slower (13): - scimark_lu: 184 ms +- 1 ms -> 191 ms +- 3 ms: 1.04x slower - go: 228 ms +- 2 ms -> 235 ms +- 2 ms: 1.03x slower - pathlib: 30.1 ms +- 0.5 ms -> 30.9 ms +- 0.4 ms: 1.03x slower - pidigits: 302 ms +- 1 ms -> 309 ms +- 1 ms: 1.02x slower - scimark_sor: 184 ms +- 2 ms -> 187 ms +- 4 ms: 1.02x slower - logging_simple: 9.09 us +- 0.14 us -> 9.22 us +- 0.14 us: 1.01x slower - json_dumps: 19.2 ms +- 0.2 ms -> 19.4 ms +- 0.2 ms: 1.01x slower - hexiom: 11.0 ms +- 0.1 ms -> 11.1 ms +- 0.1 ms: 1.01x slower - richards: 76.1 ms +- 1.6 ms -> 77.1 ms +- 1.8 ms: 1.01x slower - logging_silent: 171 ns +- 4 ns -> 173 ns +- 4 ns: 1.01x slower - telco: 10.1 ms +- 0.2 ms -> 10.2 ms +- 0.1 ms: 1.01x slower - dulwich_log: 119 ms +- 1 ms -> 120 ms +- 1 ms: 1.01x slower - deltablue: 6.38 ms +- 0.06 ms -> 6.44 ms +- 0.08 ms: 1.01x slower Faster (11): - unpickle: 20.1 us +- 0.2 us -> 19.6 us +- 0.1 us: 1.02x faster - sqlite_synth: 3.49 us +- 0.13 us -> 3.41 us +- 0.07 us: 1.02x faster - regex_v8: 37.6 ms +- 0.8 ms -> 36.9 ms +- 0.3 ms: 1.02x faster - nqueens: 133 ms +- 1 ms -> 130 ms +- 1 ms: 1.02x faster - unpickle_list: 7.63 us +- 0.15 us -> 7.50 us +- 0.14 us: 1.02x faster - spectral_norm: 165 ms +- 1 ms -> 162 ms +- 4 ms: 1.02x faster - django_template: 56.8 ms +- 2.0 ms -> 56.0 ms +- 0.5 ms: 1.01x faster - unpack_sequence: 74.0 ns +- 0.7 ns -> 73.0 ns +- 0.8 ns: 1.01x faster - nbody: 147 ms +- 2 ms -> 145 ms +- 1 ms: 1.01x faster - regex_dna: 334 ms +- 1 ms -> 330 ms +- 2 ms: 1.01x faster - pickle: 16.1 us +- 0.2 us -> 15.9 us +- 0.3 us: 1.01x faster Benchmark hidden because not significant (33): 2to3, chameleon, chaos, crypto_pyaes, fannkuch, float, html5lib, json_loads, logging_format, mako, meteor_contest, pickle_dict, pickle_list, pickle_pure_python, pyflate, python_startup, python_startup_no_site, raytrace, regex_compile, regex_effbot, scimark_fft, scimark_monte_carlo, scimark_sparse_mat_mult, sympy_expand, sympy_integrate, sympy_sum, sympy_str, tornado_http, unpickle_pure_python, xml_etree_parse, xml_etree_iterparse, xml_etree_generate, xml_etree_process Geometric mean: 1.00x slower 

compare_to main->POSTCALL:

Slower (9): - pidigits: 297 ms +- 1 ms -> 309 ms +- 1 ms: 1.04x slower - fannkuch: 603 ms +- 6 ms -> 626 ms +- 5 ms: 1.04x slower - logging_simple: 8.91 us +- 0.12 us -> 9.22 us +- 0.14 us: 1.03x slower - richards: 75.4 ms +- 1.8 ms -> 77.1 ms +- 1.8 ms: 1.02x slower - logging_format: 10.1 us +- 0.2 us -> 10.2 us +- 0.2 us: 1.02x slower - pickle_dict: 47.6 us +- 0.2 us -> 48.3 us +- 0.2 us: 1.02x slower - mako: 17.7 ms +- 0.1 ms -> 18.0 ms +- 0.1 ms: 1.02x slower - go: 232 ms +- 2 ms -> 235 ms +- 2 ms: 1.01x slower - unpickle_list: 7.42 us +- 0.15 us -> 7.50 us +- 0.14 us: 1.01x slower Faster (21): - raytrace: 486 ms +- 4 ms -> 445 ms +- 3 ms: 1.09x faster - chaos: 113 ms +- 1 ms -> 105 ms +- 1 ms: 1.07x faster - float: 126 ms +- 1 ms -> 118 ms +- 1 ms: 1.07x faster - nbody: 150 ms +- 2 ms -> 145 ms +- 1 ms: 1.03x faster - chameleon: 11.3 ms +- 0.2 ms -> 10.9 ms +- 0.1 ms: 1.03x faster - unpack_sequence: 75.2 ns +- 0.9 ns -> 73.0 ns +- 0.8 ns: 1.03x faster - scimark_fft: 538 ms +- 12 ms -> 523 ms +- 3 ms: 1.03x faster - regex_effbot: 5.32 ms +- 0.06 ms -> 5.21 ms +- 0.04 ms: 1.02x faster - scimark_sor: 191 ms +- 2 ms -> 187 ms +- 4 ms: 1.02x faster - deltablue: 6.57 ms +- 0.07 ms -> 6.44 ms +- 0.08 ms: 1.02x faster - regex_dna: 336 ms +- 12 ms -> 330 ms +- 2 ms: 1.02x faster - telco: 10.4 ms +- 0.4 ms -> 10.2 ms +- 0.1 ms: 1.02x faster - dulwich_log: 123 ms +- 1 ms -> 120 ms +- 1 ms: 1.02x faster - json_loads: 42.9 us +- 0.4 us -> 42.2 us +- 0.3 us: 1.02x faster - regex_v8: 37.5 ms +- 0.8 ms -> 36.9 ms +- 0.3 ms: 1.02x faster - pyflate: 711 ms +- 9 ms -> 702 ms +- 6 ms: 1.01x faster - unpickle: 19.9 us +- 0.3 us -> 19.6 us +- 0.1 us: 1.01x faster - regex_compile: 224 ms +- 1 ms -> 221 ms +- 1 ms: 1.01x faster - spectral_norm: 164 ms +- 3 ms -> 162 ms +- 4 ms: 1.01x faster - json_dumps: 19.6 ms +- 0.3 ms -> 19.4 ms +- 0.2 ms: 1.01x faster - tornado_http: 207 ms +- 4 ms -> 205 ms +- 4 ms: 1.01x faster Benchmark hidden because not significant (27): 2to3, crypto_pyaes, django_template, hexiom, html5lib, logging_silent, meteor_contest, nqueens, pathlib, pickle, pickle_list, pickle_pure_python, python_startup, python_startup_no_site, scimark_lu, scimark_monte_carlo, scimark_sparse_mat_mult, sqlite_synth, sympy_expand, sympy_integrate, sympy_sum, sympy_str, unpickle_pure_python, xml_etree_parse, xml_etree_iterparse, xml_etree_generate, xml_etree_process Geometric mean: 1.01x faster 

[Fidget-Spinner]

@markshannon in short, POSTCALL is slightly slower than the RETURN_VALUE hack. But both approaches are still 1.01x faster than main.

Shall we go ahead with POSTCALL then? I can update this PR with main if you're satisfied.

[Fidget-Spinner]

test_dis and _bootstrap_external are core developer merge race conditions :(.

[Fidget-Spinner]

If this is accepted, I have one more specialization in mind before 3.11 beta. Namely, calls to Python classes with non-Python __init__. E.g this common code pattern:

class MyOwnException(Exception): ... ... raise MyOwnException() # Optimize this ... 

[brandtbucher]

Should this be closed in favor of #99331?