gh-129275: avoid temporary buffer in dec_as_long()

[skirpichev]

According to the documentation: "If rdata is non-NULL, it MUST be allocated by one of libmpdec’s allocation functions and rlen MUST be correct. If necessary, the function will resize rdata. Resizing is slow and should not occur if rlen has been obtained by a call to mpd_sizeinbase."

So, possible resizing in mpd_qexport_u32/16() is for guarding against broken log10() implementations (log10 is used in the mpd_sizeinbase()).

• Issue: Avoid temporary allocation in dec_as_long() #129275

---

Temporary memory allocation slowdown conversion for small integers (~2 digits):

Benchmark	ref	patch
int(Decimal(1<<7))	499 ns	495 ns: 1.01x faster
int(Decimal(1<<300))	2.16 us	2.06 us: 1.05x faster
Geometric mean	(ref)	1.02x faster

Benchmark hidden because not significant (2): int(Decimal(1<<38)), int(Decimal(1<<3000))

Details

# bench_Decimal-to-int.py

import pyperf
from decimal import Decimal

values = ['1<<7', '1<<38', '1<<300', '1<<3000']

runner = pyperf.Runner()
for v in values:
    d = Decimal(eval(v))
    bn = 'int(Decimal('+v+'))'
    runner.bench_func(bn, int, d)

[skirpichev]

Edit: For a context, Serhiy's concerns about using mpd_sizeinbase() from old pr:

• gh-127937: convert decimal module to use import API for ints (PEP 757) #127925 (comment)
• gh-127937: convert decimal module to use import API for ints (PEP 757) #127925 (review)

In our case (base is a power of 2) we can avoid using mpd_sizeinbase() and implement more accurate version. Here is the mod_sizeinbase code:

cpython/Modules/_decimal/libmpdec/mpdecimal.c

Lines 8084 to 8113 in 65ae3d5

	size_t
	mpd_sizeinbase(const mpd_t *a, uint32_t base)
	{
	double x;
	size_t digits;
	double upper_bound;
	assert(mpd_isinteger(a));
	assert(base >= 2);
	if (mpd_iszero(a)) {
	return 1;
	}
	digits = a->digits+a->exp;
	#ifdef CONFIG_64
	/* ceil(2711437152599294 / log10(2)) + 4 == 2**53 */
	if (digits > 2711437152599294ULL) {
	return SIZE_MAX;
	}
	upper_bound = (double)((1ULL<<53)-1);
	#else
	upper_bound = (double)(SIZE_MAX-1);
	#endif
	x = (double)digits / log10(base);
	return (x > upper_bound) ? SIZE_MAX : (size_t)x + 1;
	}

Essentially, it does ndigits * log2(10)/shift. This should be also a correct bound:

(size_t)(3.321928094887363*((ndigits + shift - 1)/shift))

For shift=30 and ndigits ~ 1<<53 (upper_bound for typical case) - it will overestimate size in just 1 digit.

[erlend-aasland]

Based on the previous PR and discussions, this looks correct to me. I'd like Serhiy's opinion on it, though, since he expressed concerns earlier.

[skirpichev]

@erlend-aasland, It seems that, that Serhiy is still against this change, see the issue thread.

So, I'm going to close this in few days, unless someone else can argue further.

[skirpichev]

Thanks for review to all.