Specialize FOR_ITER via GET_ITER

[sweeneyde]

see faster-cpython/ideas#350

[sweeneyde]

Microbenchmarks show very little change to range iteration, but a decent (~1.13x) speedup to list iteration.

sween@DESKTOP-3GGN2FL:~/before_getiter2$ ./python -m pyperf timeit -s "R = range(10**5)" "for i in R: pass"
.....................
Mean +- std dev: 1.16 ms +- 0.03 ms
sween@DESKTOP-3GGN2FL:~/before_getiter2$ ./python -m pyperf timeit -s "R = list(range(10**5))" "for i in R: pass"
.....................
Mean +- std dev: 562 us +- 7 us


sween@DESKTOP-3GGN2FL:~/get_iter2$ ./python -m pyperf timeit -s "R = range(10**5)" "for i in R: pass"
.....................
Mean +- std dev: 1.15 ms +- 0.03 ms
sween@DESKTOP-3GGN2FL:~/get_iter2$ ./python -m pyperf timeit -s "R = list(range(10**5))" "for i in R: pass"
.....................
Mean +- std dev: 495 us +- 7 us

[sweeneyde]

pyperformance results (Geometric mean: 1.00x faster):

Slower (18):
- scimark_sparse_mat_mult: 5.17 ms +- 0.28 ms -> 5.55 ms +- 0.20 ms: 1.07x slower
- logging_format: 7.20 us +- 0.13 us -> 7.69 us +- 0.57 us: 1.07x slower
- logging_simple: 6.49 us +- 0.14 us -> 6.82 us +- 0.27 us: 1.05x slower
- mako: 10.1 ms +- 0.1 ms -> 10.6 ms +- 0.7 ms: 1.05x slower
- fannkuch: 396 ms +- 11 ms -> 412 ms +- 10 ms: 1.04x slower
- pickle_list: 4.44 us +- 0.14 us -> 4.62 us +- 0.17 us: 1.04x slower
- unpack_sequence: 43.0 ns +- 0.7 ns -> 44.5 ns +- 2.1 ns: 1.03x slower
- scimark_monte_carlo: 67.9 ms +- 1.9 ms -> 70.0 ms +- 5.6 ms: 1.03x slower
- nbody: 92.9 ms +- 3.6 ms -> 95.6 ms +- 3.2 ms: 1.03x slower
- pickle_dict: 28.7 us +- 0.8 us -> 29.5 us +- 0.6 us: 1.03x slower
- pidigits: 180 ms +- 2 ms -> 185 ms +- 2 ms: 1.03x slower
- telco: 6.52 ms +- 0.12 ms -> 6.65 ms +- 0.16 ms: 1.02x slower
- unpickle_pure_python: 233 us +- 8 us -> 237 us +- 8 us: 1.02x slower
- raytrace: 312 ms +- 5 ms -> 317 ms +- 15 ms: 1.02x slower
- pathlib: 19.2 ms +- 0.2 ms -> 19.5 ms +- 0.4 ms: 1.01x slower
- pickle_pure_python: 315 us +- 11 us -> 318 us +- 7 us: 1.01x slower
- meteor_contest: 109 ms +- 2 ms -> 109 ms +- 2 ms: 1.01x slower
- scimark_lu: 106 ms +- 2 ms -> 107 ms +- 2 ms: 1.01x slower

Faster (16):
- json_loads: 29.8 us +- 0.8 us -> 27.0 us +- 1.4 us: 1.11x faster
- richards: 55.8 ms +- 3.8 ms -> 52.5 ms +- 3.8 ms: 1.06x faster
- django_template: 40.4 ms +- 1.6 ms -> 38.1 ms +- 1.7 ms: 1.06x faster
- python_startup_no_site: 6.81 ms +- 0.08 ms -> 6.51 ms +- 0.05 ms: 1.05x faster
- pyflate: 456 ms +- 33 ms -> 439 ms +- 11 ms: 1.04x faster
- unpickle_list: 4.97 us +- 0.16 us -> 4.80 us +- 0.15 us: 1.04x faster
- go: 147 ms +- 4 ms -> 142 ms +- 4 ms: 1.04x faster
- regex_v8: 24.5 ms +- 1.8 ms -> 23.9 ms +- 1.3 ms: 1.03x faster
- regex_dna: 218 ms +- 3 ms -> 212 ms +- 10 ms: 1.02x faster
- regex_compile: 143 ms +- 2 ms -> 140 ms +- 4 ms: 1.02x faster
- deltablue: 3.91 ms +- 0.13 ms -> 3.83 ms +- 0.12 ms: 1.02x faster
- html5lib: 67.9 ms +- 3.0 ms -> 66.5 ms +- 2.7 ms: 1.02x faster
- chaos: 75.0 ms +- 2.3 ms -> 74.1 ms +- 1.8 ms: 1.01x faster
- scimark_fft: 357 ms +- 10 ms -> 354 ms +- 6 ms: 1.01x faster
- xml_etree_iterparse: 105 ms +- 1 ms -> 104 ms +- 2 ms: 1.01x faster
- sympy_expand: 522 ms +- 9 ms -> 518 ms +- 7 ms: 1.01x faster

Benchmark hidden because not significant (25): 2to3, chameleon, crypto_pyaes, dulwich_log, float, hexiom, json_dumps, logging_silent, nqueens, pickle, python_startup, regex_effbot, scimark_sor, spectral_norm, sqlalchemy_declarative, sqlalchemy_imperative, sqlite_synth, sympy_integrate, sympy_sum, sympy_str, tornado_http, unpickle, xml_etree_parse, xml_etree_generate, xml_etree_process

Geometric mean: 1.00x faster

[Fidget-Spinner]

Can I review this or is it still WIP?

[sweeneyde]

Feedback/reviews/discussions are appreciated! I'm just not yet sure if this is a good approach at a high level.

[Fidget-Spinner]

Just a random thought:

I noticed that for the two instructions, they are similar to the tp_iternext functions of their respective types. Would it make more sense/save more code to just call them here and let PGO inline them? I'm not sure about the perf impact.

[sweeneyde]

Closing this, as it is more complex and conceptually different from other specializations (type checks are cheap anyway).