GH-98686: Get rid of BINARY_OP_GENERIC and COMPARE_OP_GENERIC

[brandtbucher]

This optimization made more sense back before we had exponential backoff (and the unquickened versions of these instructions were just laying around waiting to be used). Nowadays this move just complicates things, and doesn't actually seem to improve performance:

All benchmarks:
===============

Slower (16):
- regex_effbot: 3.34 ms +- 0.02 ms -> 3.69 ms +- 0.01 ms: 1.10x slower
- regex_v8: 21.2 ms +- 0.1 ms -> 22.7 ms +- 0.2 ms: 1.07x slower
- regex_dna: 202 ms +- 1 ms -> 210 ms +- 1 ms: 1.04x slower
- telco: 6.25 ms +- 0.22 ms -> 6.38 ms +- 0.16 ms: 1.02x slower
- genshi_xml: 46.9 ms +- 0.9 ms -> 47.7 ms +- 0.5 ms: 1.02x slower
- mdp: 2.51 sec +- 0.02 sec -> 2.56 sec +- 0.01 sec: 1.02x slower
- pathlib: 17.3 ms +- 0.2 ms -> 17.5 ms +- 0.4 ms: 1.01x slower
- hexiom: 6.06 ms +- 0.02 ms -> 6.14 ms +- 0.04 ms: 1.01x slower
- scimark_fft: 311 ms +- 3 ms -> 314 ms +- 3 ms: 1.01x slower
- scimark_sparse_mat_mult: 4.07 ms +- 0.07 ms -> 4.11 ms +- 0.11 ms: 1.01x slower
- mako: 9.52 ms +- 0.06 ms -> 9.59 ms +- 0.08 ms: 1.01x slower
- scimark_sor: 103 ms +- 1 ms -> 104 ms +- 1 ms: 1.01x slower
- logging_silent: 91.6 ns +- 1.1 ns -> 92.1 ns +- 1.4 ns: 1.01x slower
- nbody: 94.0 ms +- 1.1 ms -> 94.5 ms +- 1.1 ms: 1.01x slower
- gunicorn: 1.07 ms +- 0.01 ms -> 1.08 ms +- 0.01 ms: 1.00x slower
- regex_compile: 128 ms +- 2 ms -> 128 ms +- 1 ms: 1.00x slower

Faster (35):
- unpack_sequence: 46.1 ns +- 4.1 ns -> 43.3 ns +- 0.6 ns: 1.06x faster
- genshi_text: 21.1 ms +- 0.2 ms -> 20.5 ms +- 0.3 ms: 1.03x faster
- pidigits: 199 ms +- 0 ms -> 193 ms +- 0 ms: 1.03x faster
- deepcopy_reduce: 2.94 us +- 0.04 us -> 2.86 us +- 0.04 us: 1.03x faster
- spectral_norm: 95.9 ms +- 3.8 ms -> 93.8 ms +- 2.1 ms: 1.02x faster
- unpickle_pure_python: 204 us +- 2 us -> 200 us +- 2 us: 1.02x faster
- sqlglot_parse: 1.34 ms +- 0.01 ms -> 1.31 ms +- 0.01 ms: 1.02x faster
- pickle_list: 4.03 us +- 0.03 us -> 3.96 us +- 0.05 us: 1.02x faster
- generators: 78.7 ms +- 0.5 ms -> 77.4 ms +- 0.5 ms: 1.02x faster
- json_loads: 24.2 us +- 0.4 us -> 23.9 us +- 0.2 us: 1.02x faster
- chameleon: 6.50 ms +- 0.08 ms -> 6.41 ms +- 0.09 ms: 1.01x faster
- sqlglot_transpile: 1.63 ms +- 0.02 ms -> 1.60 ms +- 0.02 ms: 1.01x faster
- deepcopy_memo: 34.1 us +- 0.5 us -> 33.6 us +- 0.9 us: 1.01x faster
- raytrace: 281 ms +- 5 ms -> 278 ms +- 2 ms: 1.01x faster
- coverage: 98.1 ms +- 1.2 ms -> 96.8 ms +- 1.1 ms: 1.01x faster
- xml_etree_iterparse: 104 ms +- 2 ms -> 103 ms +- 2 ms: 1.01x faster
- deepcopy: 328 us +- 3 us -> 324 us +- 3 us: 1.01x faster
- xml_etree_generate: 76.9 ms +- 0.7 ms -> 76.1 ms +- 0.6 ms: 1.01x faster
- xml_etree_process: 53.2 ms +- 0.7 ms -> 52.6 ms +- 0.6 ms: 1.01x faster
- sympy_str: 283 ms +- 4 ms -> 280 ms +- 4 ms: 1.01x faster
- crypto_pyaes: 74.1 ms +- 0.7 ms -> 73.4 ms +- 0.8 ms: 1.01x faster
- sympy_sum: 163 ms +- 1 ms -> 162 ms +- 2 ms: 1.01x faster
- pyflate: 394 ms +- 4 ms -> 391 ms +- 4 ms: 1.01x faster
- sqlglot_normalize: 106 ms +- 1 ms -> 105 ms +- 1 ms: 1.01x faster
- sympy_expand: 457 ms +- 3 ms -> 453 ms +- 5 ms: 1.01x faster
- django_template: 32.6 ms +- 0.5 ms -> 32.4 ms +- 0.5 ms: 1.01x faster
- sqlglot_optimize: 50.9 ms +- 0.2 ms -> 50.6 ms +- 0.3 ms: 1.01x faster
- float: 71.5 ms +- 0.8 ms -> 71.1 ms +- 0.9 ms: 1.01x faster
- sympy_integrate: 20.3 ms +- 0.1 ms -> 20.2 ms +- 0.1 ms: 1.01x faster
- chaos: 67.7 ms +- 1.0 ms -> 67.3 ms +- 0.7 ms: 1.01x faster
- pickle_pure_python: 282 us +- 2 us -> 281 us +- 3 us: 1.00x faster
- pickle_dict: 31.3 us +- 0.1 us -> 31.3 us +- 0.1 us: 1.00x faster
- python_startup: 8.60 ms +- 0.01 ms -> 8.58 ms +- 0.01 ms: 1.00x faster
- python_startup_no_site: 6.27 ms +- 0.01 ms -> 6.26 ms +- 0.01 ms: 1.00x faster
- 2to3: 246 ms +- 1 ms -> 246 ms +- 1 ms: 1.00x faster

Benchmark hidden because not significant (31): aiohttp, async_tree_none, async_tree_cpu_io_mixed, async_tree_io, async_tree_memoization, coroutines, deltablue, dulwich_log, fannkuch, go, html5lib, json, json_dumps, logging_format, logging_simple, meteor_contest, mypy, nqueens, pickle, pprint_safe_repr, pprint_pformat, pycparser, richards, scimark_lu, scimark_monte_carlo, sqlite_synth, thrift, tornado_http, unpickle, unpickle_list, xml_etree_parse

Geometric mean: 1.00x faster

• Issue: Quicken everything #98686

[gvanrossum]

Thanks for the merge conflict. :-/

Don't you need to change the magic number? All the opcodes are different.

[brandtbucher]

Thanks for the merge conflict. :-/

I'll wait until after #99313 is merged to land this. Anything else in flight that I should wait for?

Don't you need to change the magic number? All the opcodes are different.

We don't need to bump the magic number when adding/removing/changing specialized instructions or superinstructions, since those never get included in pycs! It's a pretty nice design.

[gvanrossum]

In that case I welcome your review of #99313.

[gvanrossum]

You have the baton. Have fun fixing the merge conflicts. :-)

[gvanrossum]

IIUC this means that un-specializable binary ops and comparisons will now be a teensy bit slower because they always have to check the counter. But it seems clear that happens rarely enough.