Type-specialized Py_DECREF

[sweeneyde]

BPO	46509
Nosy	@sweeneyde
PRs	gh-90667: Add specializations of Py_DECREF when types are known #30872

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = None
created_at = <Date 2022-01-25.05:27:08.416>
labels = ['interpreter-core', '3.11', 'performance']
title = 'Type-specialized Py_DECREF'
updated_at = <Date 2022-01-25.05:33:34.410>
user = 'https://github.com/sweeneyde'

bugs.python.org fields:

activity = <Date 2022-01-25.05:33:34.410>
actor = 'Dennis Sweeney'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['Interpreter Core']
creation = <Date 2022-01-25.05:27:08.416>
creator = 'Dennis Sweeney'
dependencies = []
files = []
hgrepos = []
issue_num = 46509
keywords = ['patch']
message_count = 2.0
messages = ['411553', '411554']
nosy_count = 1.0
nosy_names = ['Dennis Sweeney']
pr_nums = ['30872']
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue46509'
versions = ['Python 3.11']

[sweeneyde]

GCC --enable-optimizations --with-lto on WSL:

Slower (7):

• json_dumps: 12.8 ms +- 0.2 ms -> 13.1 ms +- 0.3 ms: 1.02x slower
• meteor_contest: 106 ms +- 2 ms -> 109 ms +- 3 ms: 1.02x slower
• telco: 6.50 ms +- 0.17 ms -> 6.58 ms +- 0.18 ms: 1.01x slower
• fannkuch: 383 ms +- 5 ms -> 388 ms +- 8 ms: 1.01x slower
• regex_compile: 143 ms +- 2 ms -> 145 ms +- 4 ms: 1.01x slower
• mako: 11.1 ms +- 0.1 ms -> 11.2 ms +- 0.2 ms: 1.01x slower
• chameleon: 7.08 ms +- 0.07 ms -> 7.12 ms +- 0.10 ms: 1.01x slower

Faster (27):

• unpack_sequence: 45.9 ns +- 1.1 ns -> 41.6 ns +- 1.0 ns: 1.10x faster
• logging_silent: 108 ns +- 11 ns -> 101 ns +- 3 ns: 1.06x faster
• nbody: 95.6 ms +- 3.2 ms -> 90.2 ms +- 1.9 ms: 1.06x faster
• spectral_norm: 98.3 ms +- 2.3 ms -> 92.8 ms +- 1.6 ms: 1.06x faster
• regex_dna: 202 ms +- 3 ms -> 194 ms +- 3 ms: 1.04x faster
• scimark_fft: 342 ms +- 12 ms -> 331 ms +- 7 ms: 1.03x faster
• crypto_pyaes: 89.6 ms +- 1.7 ms -> 86.8 ms +- 1.1 ms: 1.03x faster
• json_loads: 27.4 us +- 0.9 us -> 26.5 us +- 1.3 us: 1.03x faster
• scimark_monte_carlo: 69.3 ms +- 1.5 ms -> 67.3 ms +- 1.2 ms: 1.03x faster
• pickle_list: 4.62 us +- 0.21 us -> 4.51 us +- 0.15 us: 1.02x faster
• scimark_sparse_mat_mult: 5.14 ms +- 0.21 ms -> 5.02 ms +- 0.18 ms: 1.02x faster
• xml_etree_parse: 161 ms +- 5 ms -> 157 ms +- 6 ms: 1.02x faster
• regex_effbot: 3.07 ms +- 0.05 ms -> 3.00 ms +- 0.07 ms: 1.02x faster
• deltablue: 4.36 ms +- 0.14 ms -> 4.27 ms +- 0.14 ms: 1.02x faster
• pickle_pure_python: 343 us +- 6 us -> 335 us +- 8 us: 1.02x faster
• sqlite_synth: 2.60 us +- 0.06 us -> 2.55 us +- 0.04 us: 1.02x faster
• xml_etree_iterparse: 110 ms +- 2 ms -> 108 ms +- 2 ms: 1.02x faster
• go: 146 ms +- 2 ms -> 143 ms +- 3 ms: 1.02x faster
• pathlib: 20.2 ms +- 0.5 ms -> 19.8 ms +- 0.3 ms: 1.02x faster
• scimark_sor: 117 ms +- 3 ms -> 115 ms +- 2 ms: 1.02x faster
• dulwich_log: 80.9 ms +- 2.0 ms -> 79.6 ms +- 1.7 ms: 1.02x faster
• nqueens: 84.4 ms +- 1.7 ms -> 83.1 ms +- 2.0 ms: 1.02x faster
• python_startup: 8.84 ms +- 0.07 ms -> 8.76 ms +- 0.07 ms: 1.01x faster
• 2to3: 269 ms +- 4 ms -> 266 ms +- 3 ms: 1.01x faster
• float: 77.0 ms +- 1.2 ms -> 76.5 ms +- 1.5 ms: 1.01x faster
• sympy_integrate: 22.7 ms +- 0.3 ms -> 22.5 ms +- 0.2 ms: 1.01x faster
• xml_etree_process: 55.7 ms +- 0.7 ms -> 55.4 ms +- 0.6 ms: 1.01x faster

Benchmark hidden because not significant (24): chaos, django_template, hexiom, logging_format, logging_simple, pickle, pickle_dict, pidigits, pyflate, python_startup_no_site, raytrace, regex_v8, richards, scimark_lu, sqlalchemy_declarative, sqlalchemy_imperative, sympy_expand, sympy_sum, sympy_str, tornado_http, unpickle, unpickle_list, unpickle_pure_python, xml_etree_generate

Geometric mean: 1.01x faster

[sweeneyde]

This attempts to avoid the dispatch dance of

Py_DECREF(op) :: Py_TYPE(op)->tp_dealloc(op) :: Py_TYPE(op)->tp_free((PyObject *)op);

I suspect this earns the most speedup from floats, where freelist manipulation can be inlined. This might make a single-digit-int freelist more impactful.