bpo-47012: speed up iteration of bytes and bytearray

[kumaraditya303]

Benchmark:

from pyperf import Runner, perf_counter

def bench_bytes(loops, length):
    src = b'helloworld' * length
    t0 = perf_counter()
    for _ in range(loops):
        for i in src:
            pass
    return perf_counter() - t0

def bench_bytearray(loops, length):
    src = bytearray(b'hello' * length)
    t0 = perf_counter()
    for _ in range(loops):
        for i in src:
            pass
    return perf_counter() - t0

runner = Runner()
for n in [10_000, 100_000]:
    runner.bench_time_func(f"bytes {n}", bench_bytes, n)
    runner.bench_time_func(f"bytearray {n}", bench_bytearray, n)

Results:

bytes 10000: Mean +- std dev: [base] 829 us +- 38 us -> [patch] 677 us +- 44 us: 1.23x faster
bytearray 10000: Mean +- std dev: [base] 523 us +- 34 us -> [patch] 360 us +- 19 us: 1.45x faster
bytes 100000: Mean +- std dev: [base] 8.33 ms +- 0.38 ms -> [patch] 6.89 ms +- 0.75 ms: 1.21x faster
bytearray 100000: Mean +- std dev: [base] 5.19 ms +- 0.23 ms -> [patch] 3.61 ms +- 0.23 ms: 1.44x faster

Geometric mean: 1.33x faster

https://bugs.python.org/issue47012

[sweeneyde]

I agree with @animalize that it would be safest to include a preprocessor directive, whether that's with separate #if _PY_NSMALLPOSINTS > 255/#else code, or with a #error directive.

I'm not sure why anyone would compile with fewer than 256 cached small ints. Here, @vstinner added the code

// _PyLong_GetZero() and _PyLong_GetOne() must always be available
#if _PY_NSMALLPOSINTS < 2
#  error "_PY_NSMALLPOSINTS must be greater than 1"
#endif

@vstinner, would there be any downside to requiring all of (0, 1, ..., 255) be in the small int cache?

[vstinner]

I don't think that anyone ever tuned _PY_NSMALLPOSINTS. The value should be hardcoded. But just for sanity, you can add a static_assert() in code which makes assumptions about its value, just in case if someone changes _PY_NSMALLPOSINTS in the future. For example, I added assertions to ensure that 0 and 1 singletons always exist.

[kumaraditya303]

@sweeneyde I have added the compilation guard, but FYI if _PY_NSMALLPOSINTS is changed then it would break deepfreeze and module freezing infra so it is not configurable not to mention that it is declared in internal header.