To my surprise, it seems that it's possible to create "small" integers that should live in _PyLong_SMALL_INTS, but don't. Here are two examples I've found:

>>> import decimal
>>> i = int(decimal.Decimal(42))  # Modules/_decimal/_decimal.c:dec_as_long
>>> i
42
>>> i is 42
<stdin>:1: SyntaxWarning: "is" with a literal. Did you mean "=="?
False
>>> i = int.from_bytes(bytes([42]))  # Objects/longobject.c:_PyLong_FromByteArray
>>> i
42
>>> i is 42
<stdin>:1: SyntaxWarning: "is" with a literal. Did you mean "=="?
False

I'm not entirely sure if this is "allowed" or not, but in any case it seems beneficial to reach into the small ints here (provided it doesn't hurt performance, of course).

I'm testing out simple fixes for both of these now.

I don't *think* we currently rely on small integers being cached anywhere in the implementation (and neither do we guarantee anywhere in the docs that small integers will be cached), so as far as I can tell these omissions shouldn't lead to user-visible bugs.

I agree that these cases should be fixed, though.

Hmm. This sort of thing is a little dodgy, though (despite the comment that it's "okay"):

    PyObject *zero = _PyLong_GetZero();  // borrowed ref
    for (i = 1; i < nargs; i++) {
        /* --- 8< --- snipped code */
        if (res == zero) {
            /* Fast path: just check arguments.
               It is okay to use identity comparison here. */
            Py_DECREF(x);
            continue;
        }
        /* --- 8< --- snipped code*/
    }

And there are some similar things going on in rangeobject.c.

        if (r->step == _PyLong_GetOne()) {
            return idx;
        }

Again, technically "okay", since it's only a fast path and the slow path that follows will still do the right thing with a 1 that's not "the" 1, but it feels fragile.

The attached PR doesn't seem to have any impact on Decimal performance (non-optimized, non-debug build on a fairly quiet laptop):

main:

Convert 262,000 Decimals to "small" ints: 31.7 ms +- 5.3 ms
Convert 256,000 Decimals to 1-digit ints: 29.9 ms +- 3.1 ms
Convert 256,000 Decimals to 2-digit ints: 30.4 ms +- 2.8 ms
Convert 256,000 Decimals to 3-digit ints: 31.2 ms +- 3.1 ms

patched:

Convert 262,000 Decimals to "small" ints: 30.9 ms +- 4.0 ms
Convert 256,000 Decimals to 1-digit ints: 29.5 ms +- 3.0 ms
Convert 256,000 Decimals to 1-digit ints: 30.5 ms +- 2.5 ms
Convert 256,000 Decimals to 1-digit ints: 31.0 ms +- 2.3 ms

New changeset 5cd9a16 by Brandt Bucher in branch 'main':
bpo-46361: Fix "small" int caching (GH-30583)
5cd9a16