gh-100239: use PyFloat_AS_DOUBLE and _PyLong_IsZero in the float / compactlong specializations

[chris-eibl]

Since we know that we deal with exact floats here.

This is just a small follow-up on #128722 and #128963.

I don't think it deserves a news entry, but I'd happily create it if deemed meaningful enough.

Microbenchmark from @iritkatriel 's PR:

Old:

>>> timeit("for i in range(10000):\n\tb = a+i", number=100000, setup="a = 1.0")
42.05178890001844
>>> timeit("for i in range(10000):\n\tb = i+a", number=100000, setup="a = 1.0")
41.31541720000678

New:

>>> timeit("for i in range(10000):\n\tb = a+i", number=100000, setup="a = 1.0")
38.316913699993165
>>> timeit("for i in range(10000):\n\tb = i+a", number=100000, setup="a = 1.0")
38.37592560000485

Fastmark results are within the noise ¹, the three arithmetic benchmarks @Fidget-Spinner mentioned are a tad faster:

• Issue: Specialize long tail of binary operations using a table. #100239

Footnotes

1. see my gist how I do them ↩

[chris-eibl]

This microbenchmark

import time


def bench():
    t = time.perf_counter()
    f = 1.0
    for i in range(60000):
        for j in range(1, 1000):
            f = f + j
            f = f - j
            f = f / j
            f = f * j
    print(time.perf_counter() - t)


if __name__ == "__main__":
    bench()

shows even better speedup:

old: 6.5 seconds
new: 5.5 seconds

[skirpichev]

I don't think it deserves a news entry, but I'd happily create it if deemed meaningful enough.

This is user-visible change. So, I believe it deserves a news entry.

[chris-eibl]

Compact ints are an implementation detail and the term medium-size integers has been chosen in news entries in the past for them.

[Eclips4]

Thank you Chris!