gh-82088: Improve performance of PyLong_As*() for multi-digit ints

[eendebakpt]

This is a continuation of the work in #15363 by @sir-sigurd.

Performance is improved by

• Unrolling the part of the loop where no overflow can occur
• Improving the overflow check so no temporary variable is used

Performance: main

pack n dt 460.21 [ms]
pack N dt 476.32 [ms]
pack l dt 479.76 [ms]
pack L dt 451.84 [ms]

PR

pack n dt 401.02 [ms]
pack N dt 391.06 [ms]
pack l dt 408.83 [ms]
pack L dt 402.37 [ms]

Script:

import time
from struct import Struct

if __name__ == "__main__":
    for key in 'nNlL':
        N = 1000
        pack = Struct(key*N).pack
        values = tuple(range(2**30, 2**30 + N))
            
        t0=time.perf_counter()
        for _ in range(50_000):
            pack(*values);
        dt=time.perf_counter()-t0
        
        print(f'pack {key} dt {dt*1e3:.2f} [ms]')

(performance gain is similar to earlier reported improvements)

• Issue: speed-up PyLong_As*() for large longs #82088

[eendebakpt]

We can also unroll the second digit for the other methods. It gains a bit of performance at the cost of a bit of extra code. The unrolling can be factored out to a helper function (the loop itself is not so easy because there are small differences in how to handle the overflow)

[vstinner]

I rewrote the benchmark using pyperf:

import pyperf
from struct import Struct

N = 1000

runner = pyperf.Runner()
values = tuple(range(2**30, 2**30 + N))
for key in 'nNlL':
    pack = Struct(key*N).pack
    runner.bench_func(f'pack {key}', pack, *values)

Result:

Benchmark	ref	change
pack n	7.86 us	5.80 us: 1.35x faster
pack N	7.06 us	5.77 us: 1.22x faster
pack l	9.00 us	6.31 us: 1.43x faster
pack L	7.66 us	6.00 us: 1.28x faster
Geometric mean	(ref)	1.32x faster

[vstinner]

LGTM. I have two last minor suggestions.

@serhiy-storchaka: Would you mind to double check that the integer overflow checks are correct?

Replace:

            prev = x;
            x = (x << PyLong_SHIFT) | v->long_value.ob_digit[i];
            if ((x >> PyLong_SHIFT) != prev) ...

with:

            if (x > (ULONG_MAX >> PyLong_SHIFT)) ...
            x = (x << PyLong_SHIFT) | v->long_value.ob_digit[i];

[vstinner]

Thanks @eendebakpt, I merged your PR. Let's see how it goes :-)