GH-91432: Add more FOR_ITER specializations

[sweeneyde]

#91432

This does:

• FOR_ITER(tuple)
• FOR_ITER(dict_items) + UNPACK_SEQUENCE(2)
  • Eliminates handling the tuple, one dispatch
• FOR_ITER(enumerate) + UNPACK_SEQUENCE(2) + STORE_FAST
  • Eliminates handling the tuple, two dispatches, allocating the PyLongObject

I'm not sure whether all of these are worth it, but I want to see how this moves stats and micro- and macro- benchmarks.

[sweeneyde]

In microbenchmarks, it seems adding these extra opcodes bumped some things around and made FOR_ITER_RANGE and FOR_ITER_LIST a bit slower. Maybe the compiler decided not to inline _PyLong_AssignValue after this change, since this code uses it in two places.

But dict items, enumerate, and tuple did speed up, as expected.

microbenchmark script

frompyperfimportRunner, perf_counterfromitertoolsimportrepeatdeffor_range(loops, length): repetitions=repeat(None, loops) R=range(length) t0=perf_counter() for_inrepetitions: forxinR: passt1=perf_counter() returnt1-t0deffor_list(loops, length): repetitions=repeat(None, loops) L=list(map(float, range(length))) t0=perf_counter() for_inrepetitions: forxinL: passt1=perf_counter() returnt1-t0deffor_tuple(loops, length): repetitions=repeat(None, loops) T=tuple(map(float, range(length))) t0=perf_counter() for_inrepetitions: forxinT: passt1=perf_counter() returnt1-t0deffor_dict(loops, length): repetitions=repeat(None, loops) D=dict.fromkeys(map(float, range(length))) t0=perf_counter() for_inrepetitions: forx, yinD.items(): passt1=perf_counter() returnt1-t0deffor_enumerate(loops, length): repetitions=repeat(None, loops) L= [None] *lengtht0=perf_counter() for_inrepetitions: fori, xinenumerate(L): passt1=perf_counter() returnt1-t0deffor_map(loops, length): repetitions=repeat(None, loops) L= [()] *lengtht0=perf_counter() for_inrepetitions: forxinmap(len, L): passt1=perf_counter() returnt1-t0deffor_string(loops, length): repetitions=repeat(None, loops) S="a"*lengtht0=perf_counter() for_inrepetitions: forxinS: passt1=perf_counter() returnt1-t0deffor_set(loops, length): repetitions=repeat(None, loops) S={f"a{i}"foriinrange(length)} t0=perf_counter() for_inrepetitions: forxinS: passt1=perf_counter() returnt1-t0bench=Runner().bench_time_funcfornin [20, 200, 2_000, 20_000]: bench(f"for_range {n:_}", for_range, n, inner_loops=n) bench(f"for_list {n:_}", for_list, n, inner_loops=n) bench(f"for_tuple {n:_}", for_tuple, n, inner_loops=n) bench(f"for_dict {n:_}", for_dict, n, inner_loops=n) bench(f"for_enumerate {n:_}", for_enumerate, n, inner_loops=n) bench(f"for_map {n:_}", for_map, n, inner_loops=n) bench(f"for_string {n:_}", for_string, n, inner_loops=n) bench(f"for_set {n:_}", for_set, n, inner_loops=n)

Slower (7): - for_range 2_000: 4.01 ns +- 0.02 ns -> 5.58 ns +- 0.14 ns: 1.39x slower - for_range 200: 4.47 ns +- 0.06 ns -> 6.04 ns +- 0.43 ns: 1.35x slower - for_range 20_000: 4.19 ns +- 0.03 ns -> 5.57 ns +- 0.35 ns: 1.33x slower - for_range 20: 6.03 ns +- 0.32 ns -> 7.58 ns +- 0.27 ns: 1.26x slower - for_list 20: 5.95 ns +- 0.44 ns -> 6.19 ns +- 0.85 ns: 1.04x slower - for_list 20_000: 4.56 ns +- 0.17 ns -> 4.73 ns +- 0.38 ns: 1.04x slower - for_list 200: 4.69 ns +- 0.26 ns -> 4.84 ns +- 0.36 ns: 1.03x slower Faster (24): - for_enumerate 20_000: 22.2 ns +- 0.4 ns -> 9.52 ns +- 0.39 ns: 2.34x faster - for_enumerate 2_000: 21.2 ns +- 0.4 ns -> 9.58 ns +- 0.37 ns: 2.21x faster - for_dict 20_000: 13.5 ns +- 0.1 ns -> 7.52 ns +- 0.05 ns: 1.80x faster - for_dict 2_000: 13.4 ns +- 0.1 ns -> 7.52 ns +- 0.13 ns: 1.79x faster - for_dict 200: 13.8 ns +- 0.2 ns -> 7.80 ns +- 0.13 ns: 1.77x faster - for_dict 20: 17.5 ns +- 0.2 ns -> 11.6 ns +- 0.1 ns: 1.51x faster - for_enumerate 200: 13.5 ns +- 0.3 ns -> 9.71 ns +- 0.41 ns: 1.39x faster - for_tuple 20_000: 6.49 ns +- 0.05 ns -> 4.97 ns +- 0.04 ns: 1.31x faster - for_enumerate 20: 17.5 ns +- 0.5 ns -> 13.4 ns +- 0.4 ns: 1.30x faster - for_tuple 20: 8.20 ns +- 0.06 ns -> 6.31 ns +- 0.23 ns: 1.30x faster - for_tuple 2_000: 6.49 ns +- 0.05 ns -> 5.01 ns +- 0.22 ns: 1.29x faster - for_tuple 200: 6.65 ns +- 0.07 ns -> 5.16 ns +- 0.14 ns: 1.29x faster - for_string 2_000: 6.75 ns +- 0.06 ns -> 5.85 ns +- 0.21 ns: 1.16x faster - for_string 20_000: 6.74 ns +- 0.04 ns -> 5.87 ns +- 0.30 ns: 1.15x faster - for_string 200: 6.90 ns +- 0.09 ns -> 6.16 ns +- 0.39 ns: 1.12x faster - for_string 20: 8.59 ns +- 0.08 ns -> 7.70 ns +- 0.26 ns: 1.12x faster - for_set 200: 8.50 ns +- 0.10 ns -> 7.71 ns +- 0.10 ns: 1.10x faster - for_set 20: 12.4 ns +- 0.1 ns -> 11.4 ns +- 0.2 ns: 1.09x faster - for_map 20_000: 17.8 ns +- 0.3 ns -> 16.4 ns +- 0.3 ns: 1.08x faster - for_map 2_000: 17.9 ns +- 0.4 ns -> 16.5 ns +- 0.5 ns: 1.08x faster - for_map 200: 18.5 ns +- 0.3 ns -> 17.2 ns +- 0.2 ns: 1.07x faster - for_map 20: 22.0 ns +- 0.3 ns -> 20.9 ns +- 0.4 ns: 1.05x faster - for_set 20_000: 22.4 ns +- 0.2 ns -> 21.4 ns +- 0.5 ns: 1.05x faster - for_set 2_000: 12.8 ns +- 0.1 ns -> 12.3 ns +- 0.3 ns: 1.04x faster Benchmark hidden because not significant (1): for_list 2_000 Geometric mean: 1.18x faster 

[markshannon]

Before adding any more specializations for builtin iterators, I'd like to try implementing faster-cpython/ideas#392 and add specialization for generators.