Maybe, it be easier to define tmp variables to avoid dealing the indexing here:

date_pattern = r'(?P<year>\d{4})-(?P<month>\d{2})-(?P<day>\d{2})'
time_pattern = r'(?P<hour>\d{2}):(?P<minute>\d{2}):(?P<second>\d{2})(?P<microsecond>\.\d+)?(?P<tzinfo>Z|[+-]\d{2}:\d{2})?'
datetime_pattern = date_pattern + r'[T ]' + time_pattern
date_re = re_compile(date_pattern + '$', ASCII)
time_re = re_compile(time_pattern + '$', ASCII | IGNORECASE)
datetime_re = re_compile(datetime_pattern + '$', ASCII | IGNORECASE)

I'll note that this won't match all the outputs of datetime.isoformat, if that's a goal of this version of datetime.fromisoformat, since the sep parameter does accept non-ASCII components.

That said, it seems like _date_re is requiring a full specification of YYYY-MM-DD, in which case you can just split the string as:

date_str = dt_str[0:10]
time_str = dt_str[11:]
dt_sep = dt_str[10:11]

It's then kinda trivial to enforce whatever restrictions you want on dt_sep, and continue to use the ASCII flag for date_re and time_re.

Both the time and datetime classes require microsecond to be strictly less than 1 million, and it looks like you don’t handle rolling over seconds, minutes, etc. Test case:

datetime.fromisoformat('2017-12-31 23:59:59.9999995') -> datetime(2018, 1, 1, 0, 0, 0)

For datetime, I might do the rolling over by adding a timedelta. For time, maybe it is not worth doing any rounding.

Also, float is approximate, e.g. for me float(".000_001_499_999_999_999_999_99") rounds up over 1.5 µs, which round would round up to 2 µs. I wonder if it is worth doing the rounding without float; then you could claim in the documentation that rounding is always half-to-even. Untested code:

_time_re = ... r'(?:(?P<microsecond>\.\d{1,6})(?P<us_frac>\d*))?' ...
...
us = int(float(us) * 1e6)
frac = kw.pop('us_frac').rstrip("0")
# Round halfway up to even rather than down to odd
us += frac > "5" or frac == "5" and us % 2

Convention is to use imperative form ("Construct"). This comment applies in other places.

Any reason not to have this at the top level ?

You've used only one blank line in the other places (which is IMHO more beautiful).

Also, do you reckon it would make sense to have a single function like:

static PyObject *
strptime_fromisoformat(PyObject *cls, PyObject *args, const char *method_name)
{
    static PyObject *module = NULL;
    PyObject *string;

    if (!PyArg_ParseTuple(args, "U:fromisoformat", &string))
        return NULL;

    if (module == NULL) {
        module = PyImport_ImportModule("_strptime");
        if (module == NULL)
            return NULL;
    }

    return PyObject_CallMethod(module, method_name, "OO", cls, string);
}

static PyObject *
date_fromisoformat(PyObject *cls, PyObject *args)
{
    return strptime_fromisoformat(cls, args, "_parse_isodate");
}

static PyObject *
time_fromisoformat(PyObject *cls, PyObject *args)
{
    return strptime_fromisoformat(cls, args, "_parse_isotime");
}

static PyObject *
datetime_fromisoformat(PyObject *cls, PyObject *args)
{
    return strptime_fromisoformat(cls, args, "_parse_isodatetime");
}