GH-98831: Remove all remaining DISPATCH() calls from bytecodes.c

[gvanrossum]

Also mark those opcodes that have no stack effect as such (inst(FOO) -> inst(FOO, (--))).

• Issue: Generate the interpreter #98831

[gvanrossum]

@brandtbucher When you have the benchmark results can you just post them here? So Mark won't have to ask for them.

[brandtbucher]

Here you go:

Details

All benchmarks:
===============
Slower (21):
- nbody: 93.3 ms +- 1.6 ms -> 96.8 ms +- 1.9 ms: 1.04x slower
- unpickle_list: 4.91 us +- 0.06 us -> 5.04 us +- 0.20 us: 1.03x slower
- scimark_fft: 309 ms +- 4 ms -> 315 ms +- 8 ms: 1.02x slower
- sqlglot_normalize: 106 ms +- 1 ms -> 108 ms +- 1 ms: 1.02x slower
- pickle: 10.1 us +- 0.1 us -> 10.2 us +- 0.1 us: 1.02x slower
- chameleon: 6.50 ms +- 0.09 ms -> 6.60 ms +- 0.08 ms: 1.02x slower
- telco: 6.31 ms +- 0.12 ms -> 6.39 ms +- 0.16 ms: 1.01x slower
- fannkuch: 366 ms +- 3 ms -> 371 ms +- 3 ms: 1.01x slower
- crypto_pyaes: 74.5 ms +- 1.6 ms -> 75.4 ms +- 0.5 ms: 1.01x slower
- sqlglot_optimize: 51.5 ms +- 0.4 ms -> 52.0 ms +- 0.2 ms: 1.01x slower
- pycparser: 1.08 sec +- 0.02 sec -> 1.09 sec +- 0.02 sec: 1.01x slower
- json_dumps: 9.33 ms +- 0.13 ms -> 9.39 ms +- 0.14 ms: 1.01x slower
- regex_v8: 21.2 ms +- 0.2 ms -> 21.3 ms +- 0.0 ms: 1.01x slower
- scimark_sor: 104 ms +- 1 ms -> 104 ms +- 1 ms: 1.01x slower
- nqueens: 83.2 ms +- 1.0 ms -> 83.7 ms +- 1.3 ms: 1.01x slower
- sqlglot_transpile: 1.65 ms +- 0.02 ms -> 1.66 ms +- 0.02 ms: 1.01x slower
- python_startup: 8.63 ms +- 0.01 ms -> 8.67 ms +- 0.01 ms: 1.00x slower
- python_startup_no_site: 6.31 ms +- 0.01 ms -> 6.33 ms +- 0.01 ms: 1.00x slower
- coroutines: 25.4 ms +- 0.1 ms -> 25.5 ms +- 0.3 ms: 1.00x slower
- mako: 9.68 ms +- 0.05 ms -> 9.71 ms +- 0.05 ms: 1.00x slower
- sympy_integrate: 20.5 ms +- 0.1 ms -> 20.6 ms +- 0.1 ms: 1.00x slower
Faster (34):
- regex_dna: 208 ms +- 1 ms -> 199 ms +- 1 ms: 1.05x faster
- logging_simple: 6.03 us +- 0.08 us -> 5.80 us +- 0.08 us: 1.04x faster
- pickle_list: 4.19 us +- 0.03 us -> 4.04 us +- 0.04 us: 1.04x faster
- pidigits: 199 ms +- 0 ms -> 193 ms +- 0 ms: 1.03x faster
- django_template: 33.5 ms +- 1.0 ms -> 32.6 ms +- 0.4 ms: 1.03x faster
- logging_format: 6.58 us +- 0.09 us -> 6.40 us +- 0.09 us: 1.03x faster
- sqlite_synth: 2.65 us +- 0.04 us -> 2.59 us +- 0.04 us: 1.02x faster
- generators: 73.6 ms +- 0.4 ms -> 72.1 ms +- 0.3 ms: 1.02x faster
- regex_compile: 130 ms +- 2 ms -> 128 ms +- 1 ms: 1.02x faster
- richards: 43.5 ms +- 1.1 ms -> 42.8 ms +- 0.6 ms: 1.02x faster
- mdp: 2.56 sec +- 0.01 sec -> 2.52 sec +- 0.04 sec: 1.02x faster
- chaos: 68.0 ms +- 0.6 ms -> 66.9 ms +- 0.7 ms: 1.02x faster
- json: 4.64 ms +- 0.15 ms -> 4.57 ms +- 0.11 ms: 1.02x faster
- pickle_pure_python: 287 us +- 4 us -> 283 us +- 5 us: 1.01x faster
- deepcopy_reduce: 3.01 us +- 0.04 us -> 2.97 us +- 0.04 us: 1.01x faster
- pyflate: 405 ms +- 8 ms -> 400 ms +- 4 ms: 1.01x faster
- deepcopy: 334 us +- 5 us -> 329 us +- 3 us: 1.01x faster
- xml_etree_generate: 77.0 ms +- 0.8 ms -> 76.1 ms +- 0.8 ms: 1.01x faster
- pprint_pformat: 1.42 sec +- 0.02 sec -> 1.41 sec +- 0.02 sec: 1.01x faster
- xml_etree_process: 53.0 ms +- 0.6 ms -> 52.4 ms +- 0.6 ms: 1.01x faster
- pprint_safe_repr: 692 ms +- 7 ms -> 684 ms +- 8 ms: 1.01x faster
- sympy_expand: 460 ms +- 7 ms -> 456 ms +- 6 ms: 1.01x faster
- coverage: 96.5 ms +- 1.2 ms -> 95.6 ms +- 1.8 ms: 1.01x faster
- genshi_xml: 47.4 ms +- 0.8 ms -> 47.1 ms +- 0.5 ms: 1.01x faster
- deepcopy_memo: 34.4 us +- 0.6 us -> 34.2 us +- 0.3 us: 1.01x faster
- raytrace: 287 ms +- 3 ms -> 284 ms +- 7 ms: 1.01x faster
- deltablue: 3.34 ms +- 0.05 ms -> 3.32 ms +- 0.05 ms: 1.01x faster
- genshi_text: 20.7 ms +- 0.3 ms -> 20.6 ms +- 0.3 ms: 1.01x faster
- sympy_str: 285 ms +- 4 ms -> 283 ms +- 3 ms: 1.01x faster
- go: 138 ms +- 1 ms -> 137 ms +- 1 ms: 1.01x faster
- unpickle_pure_python: 205 us +- 2 us -> 204 us +- 2 us: 1.01x faster
- xml_etree_parse: 148 ms +- 2 ms -> 147 ms +- 1 ms: 1.01x faster
- json_loads: 24.0 us +- 0.1 us -> 23.9 us +- 0.2 us: 1.00x faster
- pickle_dict: 30.9 us +- 0.1 us -> 30.8 us +- 0.1 us: 1.00x faster
Benchmark hidden because not significant (27): 2to3, aiohttp, async_tree_none, async_tree_cpu_io_mixed, async_tree_io, async_tree_memoization, dulwich_log, float, gunicorn, hexiom, html5lib, logging_silent, meteor_contest, mypy, pathlib, regex_effbot, scimark_lu, scimark_monte_carlo, scimark_sparse_mat_mult, spectral_norm, sqlglot_parse, sympy_sum, thrift, tornado_http, unpack_sequence, unpickle, xml_etree_iterparse
Geometric mean: 1.00x faster

[markshannon]

I feel that this ought to fail, as we should have NULL checks in debug mode.

[gvanrossum]

Yeah, but we never had those. In debug mode we check for stack overflow/underflow but not for NULL-ness. (Of course the NULL will wreak havoc if it gets popped by the wrong opcode, bu that's also nothing new.) The code generator still checks for stack over/underflow by calling STACK_SHRINK()/GROW(), so nothing is really changed.

[brandtbucher]

Once we have typed stack effects, we can auto-generate those sorts of assertions. That would be really nice.

[brandtbucher]

:(

[gvanrossum]

Yeah, I'm really counting on the compiler to optimize this for us.