offsets.pyx

import re
import time
import warnings

from pandas.util._exceptions import find_stack_level

cimport cython
from cpython.datetime cimport (
    PyDate_Check,
    PyDateTime_Check,
    PyDelta_Check,
    date,
    datetime,
    import_datetime,
    time as dt_time,
    timedelta,
)

import warnings

import_datetime()

import numpy as np

cimport numpy as cnp
from numpy cimport (
    int64_t,
    ndarray,
)

cnp.import_array()

# TODO: formalize having _libs.properties "above" tslibs in the dependency structure

from pandas._libs.properties import cache_readonly

from pandas._libs.tslibs cimport util
from pandas._libs.tslibs.util cimport (
    is_float_object,
    is_integer_object,
)

from pandas._libs.tslibs.ccalendar import (
    MONTH_ALIASES,
    int_to_weekday,
    weekday_to_int,
)
from pandas.util._exceptions import find_stack_level

from pandas._libs.tslibs.ccalendar cimport (
    MONTH_TO_CAL_NUM,
    dayofweek,
    get_days_in_month,
    get_firstbday,
    get_lastbday,
)
from pandas._libs.tslibs.conversion cimport localize_pydatetime
from pandas._libs.tslibs.dtypes cimport (
    c_DEPR_ABBREVS,
    c_OFFSET_REMOVED_FREQSTR,
    c_REVERSE_OFFSET_REMOVED_FREQSTR,
    periods_per_day,
)
from pandas._libs.tslibs.nattype cimport (
    NPY_NAT,
    c_NaT as NaT,
)
from pandas._libs.tslibs.np_datetime cimport (
    NPY_DATETIMEUNIT,
    get_unit_from_dtype,
    import_pandas_datetime,
    npy_datetimestruct,
    npy_datetimestruct_to_datetime,
    pandas_datetime_to_datetimestruct,
    pydate_to_dtstruct,
)

import_pandas_datetime()

from .dtypes cimport PeriodDtypeCode
from .timedeltas cimport (
    _Timedelta,
    delta_to_nanoseconds,
    is_any_td_scalar,
)

from .timedeltas import Timedelta

from .timestamps cimport _Timestamp

from .timestamps import Timestamp

# ---------------------------------------------------------------------
# Misc Helpers

cdef bint is_offset_object(object obj):
    return isinstance(obj, BaseOffset)


cdef bint is_tick_object(object obj):
    return isinstance(obj, Tick)


cdef bint _is_normalized(datetime dt):
    if dt.hour != 0 or dt.minute != 0 or dt.second != 0 or dt.microsecond != 0:
        # Regardless of whether dt is datetime vs Timestamp
        return False
    if isinstance(dt, _Timestamp):
        return dt.nanosecond == 0
    return True


def apply_wraps(func):
    # Note: normally we would use `@functools.wraps(func)`, but this does
    # not play nicely with cython class methods

    def wrapper(self, other):

        if other is NaT:
            return NaT
        elif (
            isinstance(other, BaseOffset)
            or PyDelta_Check(other)
            or cnp.is_timedelta64_object(other)
        ):
            # timedelta path
            return func(self, other)
        elif cnp.is_datetime64_object(other) or PyDate_Check(other):
            # PyDate_Check includes date, datetime
            other = Timestamp(other)
        else:
            # This will end up returning NotImplemented back in __add__
            raise ApplyTypeError

        tz = other.tzinfo
        nano = other.nanosecond

        if self._adjust_dst:
            other = other.tz_localize(None)

        result = func(self, other)

        result2 = Timestamp(result).as_unit(other.unit)
        if result == result2:
            # i.e. the conversion is non-lossy, not the case for e.g.
            #  test_milliseconds_combination
            result = result2

        if self._adjust_dst:
            result = result.tz_localize(tz)

        if self.normalize:
            result = result.normalize()

        # If the offset object does not have a nanoseconds component,
        # the result's nanosecond component may be lost.
        if not self.normalize and nano != 0 and not hasattr(self, "nanoseconds"):
            if result.nanosecond != nano:
                if result.tz is not None:
                    # convert to UTC
                    res = result.tz_localize(None)
                else:
                    res = result
                value = res.as_unit("ns")._value
                result = Timestamp(value + nano)

        if tz is not None and result.tzinfo is None:
            result = result.tz_localize(tz)

        return result

    # do @functools.wraps(func) manually since it doesn't work on cdef funcs
    wrapper.__name__ = func.__name__
    wrapper.__doc__ = func.__doc__
    return wrapper


cdef _wrap_timedelta_result(result):
    """
    Tick operations dispatch to their Timedelta counterparts.  Wrap the result
    of these operations in a Tick if possible.

    Parameters
    ----------
    result : object

    Returns
    -------
    object
    """
    if PyDelta_Check(result):
        # convert Timedelta back to a Tick
        return delta_to_tick(result)

    return result

# ---------------------------------------------------------------------
# Business Helpers


cdef _get_calendar(weekmask, holidays, calendar):
    """
    Generate busdaycalendar
    """
    if isinstance(calendar, np.busdaycalendar):
        if not holidays:
            holidays = tuple(calendar.holidays)
        elif not isinstance(holidays, tuple):
            holidays = tuple(holidays)
        else:
            # trust that calendar.holidays and holidays are
            # consistent
            pass
        return calendar, holidays

    if holidays is None:
        holidays = []
    try:
        holidays = holidays + calendar.holidays().tolist()
    except AttributeError:
        pass
    holidays = [_to_dt64D(dt) for dt in holidays]
    holidays = tuple(sorted(holidays))

    kwargs = {"weekmask": weekmask}
    if holidays:
        kwargs["holidays"] = holidays

    busdaycalendar = np.busdaycalendar(**kwargs)
    return busdaycalendar, holidays


cdef _to_dt64D(dt):
    # Currently
    # > np.datetime64(dt.datetime(2013,5,1),dtype='datetime64[D]')
    # numpy.datetime64('2013-05-01T02:00:00.000000+0200')
    # Thus astype is needed to cast datetime to datetime64[D]
    if getattr(dt, "tzinfo", None) is not None:
        # Get the nanosecond timestamp,
        #  equiv `Timestamp(dt).value` or `dt.timestamp() * 10**9`
        # The `naive` must be the `dt` naive wall time
        #  instead of the naive absolute time (GH#49441)
        naive = dt.replace(tzinfo=None)
        dt = np.datetime64(naive, "D")
    else:
        dt = np.datetime64(dt)
    if dt.dtype.name != "datetime64[D]":
        dt = dt.astype("datetime64[D]")
    return dt


# ---------------------------------------------------------------------
# Validation


cdef _validate_business_time(t_input):
    if isinstance(t_input, str):
        try:
            t = time.strptime(t_input, "%H:%M")
            return dt_time(hour=t.tm_hour, minute=t.tm_min)
        except ValueError:
            raise ValueError("time data must match '%H:%M' format")
    elif isinstance(t_input, dt_time):
        if t_input.second != 0 or t_input.microsecond != 0:
            raise ValueError(
                "time data must be specified only with hour and minute")
        return t_input
    else:
        raise ValueError("time data must be string or datetime.time")


# ---------------------------------------------------------------------
# Constructor Helpers

_relativedelta_kwds = {"years", "months", "weeks", "days", "year", "month",
                       "day", "weekday", "hour", "minute", "second",
                       "microsecond", "millisecond", "nanosecond",
                       "nanoseconds", "hours", "minutes", "seconds",
                       "milliseconds", "microseconds"}


cdef _determine_offset(kwds):
    if not kwds:
        # GH 45643/45890: (historically) defaults to 1 day
        return timedelta(days=1), False

    if "millisecond" in kwds:
        raise NotImplementedError(
            "Using DateOffset to replace `millisecond` component in "
            "datetime object is not supported. Use "
            "`microsecond=timestamp.microsecond % 1000 + ms * 1000` "
            "instead."
        )

    nanos = {"nanosecond", "nanoseconds"}

    # nanos are handled by apply_wraps
    if all(k in nanos for k in kwds):
        return timedelta(days=0), False

    kwds_no_nanos = {k: v for k, v in kwds.items() if k not in nanos}

    kwds_use_relativedelta = {
        "year", "month", "day", "hour", "minute",
        "second", "microsecond", "weekday", "years", "months", "weeks", "days",
        "hours", "minutes", "seconds", "microseconds"
    }

    # "weeks" and "days" are left out despite being valid args for timedelta,
    # because (historically) timedelta is used only for sub-daily.
    kwds_use_timedelta = {
        "seconds", "microseconds", "milliseconds", "minutes", "hours",
    }

    if all(k in kwds_use_timedelta for k in kwds_no_nanos):
        # Sub-daily offset - use timedelta (tz-aware)
        # This also handles "milliseconds" (plur): see GH 49897
        return timedelta(**kwds_no_nanos), False

    # convert milliseconds to microseconds, so relativedelta can parse it
    if "milliseconds" in kwds_no_nanos:
        micro = kwds_no_nanos.pop("milliseconds") * 1000
        kwds_no_nanos["microseconds"] = kwds_no_nanos.get("microseconds", 0) + micro

    if all(k in kwds_use_relativedelta for k in kwds_no_nanos):
        from dateutil.relativedelta import relativedelta

        return relativedelta(**kwds_no_nanos), True

    raise ValueError(
        f"Invalid argument/s or bad combination of arguments: {list(kwds.keys())}"
    )

# ---------------------------------------------------------------------
# Mixins & Singletons


class ApplyTypeError(TypeError):
    # sentinel class for catching the apply error to return NotImplemented
    pass


# ---------------------------------------------------------------------
# Base Classes

cdef class BaseOffset:
    """
    Base class for DateOffset methods that are not overridden by subclasses.

    Parameters
    ----------
    n : int
        Number of multiples of the frequency.

    normalize : bool
        Whether the frequency can align with midnight.

    Examples
    --------
    >>> pd.offsets.Hour(5).n
    5
    >>> pd.offsets.Hour(5).normalize
    False
    """
    # ensure that reversed-ops with numpy scalars return NotImplemented
    __array_priority__ = 1000

    _day_opt = None
    _attributes = tuple(["n", "normalize"])
    _use_relativedelta = False
    _adjust_dst = True

    # cdef readonly:
    #    int64_t n
    #    bint normalize
    #    dict _cache

    def __init__(self, n=1, normalize=False):
        n = self._validate_n(n)
        self.n = n
        self.normalize = normalize
        self._cache = {}

    def __eq__(self, other) -> bool:
        if isinstance(other, str):
            try:
                # GH#23524 if to_offset fails, we are dealing with an
                #  incomparable type so == is False and != is True
                other = to_offset(other)
            except ValueError:
                # e.g. "infer"
                return False
        try:
            return self._params == other._params
        except AttributeError:
            # other is not a DateOffset object
            return False

    def __ne__(self, other):
        return not self == other

    def __hash__(self) -> int:
        return hash(self._params)

    @cache_readonly
    def _params(self):
        """
        Returns a tuple containing all of the attributes needed to evaluate
        equality between two DateOffset objects.
        """
        d = getattr(self, "__dict__", {})
        all_paras = d.copy()
        all_paras["n"] = self.n
        all_paras["normalize"] = self.normalize
        for attr in self._attributes:
            if hasattr(self, attr) and attr not in d:
                # cython attributes are not in __dict__
                all_paras[attr] = getattr(self, attr)

        if "holidays" in all_paras and not all_paras["holidays"]:
            all_paras.pop("holidays")
        exclude = ["kwds", "name", "calendar"]
        attrs = [(k, v) for k, v in all_paras.items()
                 if (k not in exclude) and (k[0] != "_")]
        attrs = sorted(set(attrs))
        params = tuple([str(type(self))] + attrs)
        return params

    @property
    def kwds(self) -> dict:
        """
        Return a dict of extra parameters for the offset.

        Examples
        --------
        >>> pd.DateOffset(5).kwds
        {}

        >>> pd.offsets.FY5253Quarter().kwds
        {'weekday': 0,
         'startingMonth': 1,
         'qtr_with_extra_week': 1,
         'variation': 'nearest'}
        """
        # for backwards-compatibility
        kwds = {name: getattr(self, name, None) for name in self._attributes
                if name not in ["n", "normalize"]}
        return {name: kwds[name] for name in kwds if kwds[name] is not None}

    @property
    def base(self):
        """
        Returns a copy of the calling offset object with n=1 and all other
        attributes equal.
        """
        return type(self)(n=1, normalize=self.normalize, **self.kwds)

    def __add__(self, other):
        if util.is_array(other) and other.dtype == object:
            return np.array([self + x for x in other])

        try:
            return self._apply(other)
        except ApplyTypeError:
            return NotImplemented

    def __radd__(self, other):
        return self.__add__(other)

    def __sub__(self, other):
        if PyDateTime_Check(other):
            raise TypeError("Cannot subtract datetime from offset.")
        elif type(other) is type(self):
            return type(self)(self.n - other.n, normalize=self.normalize,
                              **self.kwds)
        else:
            # e.g. PeriodIndex
            return NotImplemented

    def __rsub__(self, other):
        return (-self).__add__(other)

    def __mul__(self, other):
        if util.is_array(other):
            return np.array([self * x for x in other])
        elif is_integer_object(other):
            return type(self)(n=other * self.n, normalize=self.normalize,
                              **self.kwds)
        return NotImplemented

    def __rmul__(self, other):
        return self.__mul__(other)

    def __neg__(self):
        # Note: we are deferring directly to __mul__ instead of __rmul__, as
        # that allows us to use methods that can go in a `cdef class`
        return self * -1

    def copy(self):
        # Note: we are deferring directly to __mul__ instead of __rmul__, as
        # that allows us to use methods that can go in a `cdef class`
        """
        Return a copy of the frequency.

        Examples
        --------
        >>> freq = pd.DateOffset(1)
        >>> freq_copy = freq.copy()
        >>> freq is freq_copy
        False
        """
        return self * 1

    # ------------------------------------------------------------------
    # Name and Rendering Methods

    def __repr__(self) -> str:
        # _output_name used by B(Year|Quarter)(End|Begin) to
        #  expand "B" -> "Business"
        class_name = getattr(self, "_output_name", type(self).__name__)

        if abs(self.n) != 1:
            plural = "s"
        else:
            plural = ""

        n_str = ""
        if self.n != 1:
            n_str = f"{self.n} * "

        out = f"<{n_str}{class_name}{plural}{self._repr_attrs()}>"
        return out

    def _repr_attrs(self) -> str:
        exclude = {"n", "inc", "normalize"}
        attrs = []
        for attr in sorted(self._attributes):
            # _attributes instead of __dict__ because cython attrs are not in __dict__
            if attr.startswith("_") or attr == "kwds" or not hasattr(self, attr):
                # DateOffset may not have some of these attributes
                continue
            elif attr not in exclude:
                value = getattr(self, attr)
                attrs.append(f"{attr}={value}")

        out = ""
        if attrs:
            out += ": " + ", ".join(attrs)
        return out

    @property
    def name(self) -> str:
        """
        Return a string representing the base frequency.

        Examples
        --------
        >>> pd.offsets.Hour().name
        'h'

        >>> pd.offsets.Hour(5).name
        'h'
        """
        return self.rule_code

    @property
    def _prefix(self) -> str:
        raise NotImplementedError("Prefix not defined")

    @property
    def rule_code(self) -> str:
        return self._prefix

    @cache_readonly
    def freqstr(self) -> str:
        """
        Return a string representing the frequency.

        Examples
        --------
        >>> pd.DateOffset(5).freqstr
        '<5 * DateOffsets>'

        >>> pd.offsets.BusinessHour(2).freqstr
        '2bh'

        >>> pd.offsets.Nano().freqstr
        'ns'

        >>> pd.offsets.Nano(-3).freqstr
        '-3ns'
        """
        try:
            code = self.rule_code
        except NotImplementedError:
            return str(repr(self))

        if self.n != 1:
            fstr = f"{self.n}{code}"
        else:
            fstr = code

        try:
            if self._offset:
                fstr += self._offset_str()
        except AttributeError:
            # TODO: standardize `_offset` vs `offset` naming convention
            pass

        return fstr

    def _offset_str(self) -> str:
        return ""

    # ------------------------------------------------------------------

    def _apply(self, other):
        raise NotImplementedError("implemented by subclasses")

    def _apply_array(self, dtarr: np.ndarray) -> np.ndarray:
        # NB: _apply_array does not handle respecting `self.normalize`, the
        #  caller (DatetimeArray) handles that in post-processing.
        raise NotImplementedError(
            f"DateOffset subclass {type(self).__name__} "
            "does not have a vectorized implementation"
        )

    def rollback(self, dt) -> datetime:
        """
        Roll provided date backward to next offset only if not on offset.

        Returns
        -------
        TimeStamp
            Rolled timestamp if not on offset, otherwise unchanged timestamp.
        """
        dt = Timestamp(dt)
        if not self.is_on_offset(dt):
            dt = dt - type(self)(1, normalize=self.normalize, **self.kwds)
        return dt

    def rollforward(self, dt) -> datetime:
        """
        Roll provided date forward to next offset only if not on offset.

        Returns
        -------
        TimeStamp
            Rolled timestamp if not on offset, otherwise unchanged timestamp.
        """
        dt = Timestamp(dt)
        if not self.is_on_offset(dt):
            dt = dt + type(self)(1, normalize=self.normalize, **self.kwds)
        return dt

    def _get_offset_day(self, other: datetime) -> int:
        # subclass must implement `_day_opt`; calling from the base class
        # will implicitly assume day_opt = "business_end", see get_day_of_month.
        cdef:
            npy_datetimestruct dts
        pydate_to_dtstruct(other, &dts)
        return get_day_of_month(&dts, self._day_opt)

    def is_on_offset(self, dt: datetime) -> bool:
        """
        Return boolean whether a timestamp intersects with this frequency.

        Parameters
        ----------
        dt : datetime.datetime
            Timestamp to check intersections with frequency.

        Examples
        --------
        >>> ts = pd.Timestamp(2022, 1, 1)
        >>> freq = pd.offsets.Day(1)
        >>> freq.is_on_offset(ts)
        True

        >>> ts = pd.Timestamp(2022, 8, 6)
        >>> ts.day_name()
        'Saturday'
        >>> freq = pd.offsets.BusinessDay(1)
        >>> freq.is_on_offset(ts)
        False
        """
        if self.normalize and not _is_normalized(dt):
            return False

        # Default (slow) method for determining if some date is a member of the
        # date range generated by this offset. Subclasses may have this
        # re-implemented in a nicer way.
        a = dt
        b = (dt + self) - self
        return a == b

    # ------------------------------------------------------------------

    # Staticmethod so we can call from Tick.__init__, will be unnecessary
    #  once BaseOffset is a cdef class and is inherited by Tick
    @staticmethod
    def _validate_n(n) -> int:
        """
        Require that `n` be an integer.

        Parameters
        ----------
        n : int

        Returns
        -------
        nint : int

        Raises
        ------
        TypeError if `int(n)` raises
        ValueError if n != int(n)
        """
        if cnp.is_timedelta64_object(n):
            raise TypeError(f"`n` argument must be an integer, got {type(n)}")
        try:
            nint = int(n)
        except (ValueError, TypeError):
            raise TypeError(f"`n` argument must be an integer, got {type(n)}")
        if n != nint:
            raise ValueError(f"`n` argument must be an integer, got {n}")
        return nint

    def __setstate__(self, state):
        """
        Reconstruct an instance from a pickled state
        """
        self.n = state.pop("n")
        self.normalize = state.pop("normalize")
        self._cache = state.pop("_cache", {})
        # At this point we expect state to be empty

    def __getstate__(self):
        """
        Return a pickleable state
        """
        state = {}
        state["n"] = self.n
        state["normalize"] = self.normalize

        # we don't want to actually pickle the calendar object
        # as its a np.busyday; we recreate on deserialization
        state.pop("calendar", None)
        if "kwds" in state:
            state["kwds"].pop("calendar", None)

        return state

    @property
    def nanos(self):
        raise ValueError(f"{self} is a non-fixed frequency")

    # ------------------------------------------------------------------

    def is_month_start(self, _Timestamp ts):
        """
        Return boolean whether a timestamp occurs on the month start.

        Parameters
        ----------
        ts : Timestamp
            The timestamp to check.

        See Also
        --------
        is_month_end : Return boolean whether a timestamp occurs on the month end.

        Examples
        --------
        >>> ts = pd.Timestamp(2022, 1, 1)
        >>> freq = pd.offsets.Hour(5)
        >>> freq.is_month_start(ts)
        True
        """
        return ts._get_start_end_field("is_month_start", self)

    def is_month_end(self, _Timestamp ts):
        """
        Return boolean whether a timestamp occurs on the month end.

        Parameters
        ----------
        ts : Timestamp
            The timestamp to check.

        See Also
        --------
        is_month_start : Return boolean whether a timestamp occurs on the month start.

        Examples
        --------
        >>> ts = pd.Timestamp(2022, 1, 1)
        >>> freq = pd.offsets.Hour(5)
        >>> freq.is_month_end(ts)
        False
        """
        return ts._get_start_end_field("is_month_end", self)

    def is_quarter_start(self, _Timestamp ts):
        """
        Return boolean whether a timestamp occurs on the quarter start.

        Parameters
        ----------
        ts : Timestamp
            The timestamp to check.

        See Also
        --------
        is_quarter_end : Return boolean whether a timestamp occurs on the quarter end.

        Examples
        --------
        >>> ts = pd.Timestamp(2022, 1, 1)
        >>> freq = pd.offsets.Hour(5)
        >>> freq.is_quarter_start(ts)
        True
        """
        return ts._get_start_end_field("is_quarter_start", self)

    def is_quarter_end(self, _Timestamp ts):
        """
        Return boolean whether a timestamp occurs on the quarter end.

        Parameters
        ----------
        ts : Timestamp
            The timestamp to check.

        See Also
        --------
        is_quarter_start : Return boolean whether a timestamp
            occurs on the quarter start.

        Examples
        --------
        >>> ts = pd.Timestamp(2022, 1, 1)
        >>> freq = pd.offsets.Hour(5)
        >>> freq.is_quarter_end(ts)
        False
        """
        return ts._get_start_end_field("is_quarter_end", self)

    def is_year_start(self, _Timestamp ts):
        """
        Return boolean whether a timestamp occurs on the year start.

        Parameters
        ----------
        ts : Timestamp
            The timestamp to check.

        See Also
        --------
        is_year_end : Return boolean whether a timestamp occurs on the year end.

        Examples
        --------
        >>> ts = pd.Timestamp(2022, 1, 1)
        >>> freq = pd.offsets.Hour(5)
        >>> freq.is_year_start(ts)
        True
        """
        return ts._get_start_end_field("is_year_start", self)

    def is_year_end(self, _Timestamp ts):
        """
        Return boolean whether a timestamp occurs on the year end.

        Parameters
        ----------
        ts : Timestamp
            The timestamp to check.

        See Also
        --------
        is_year_start : Return boolean whether a timestamp occurs on the year start.

        Examples
        --------
        >>> ts = pd.Timestamp(2022, 1, 1)
        >>> freq = pd.offsets.Hour(5)
        >>> freq.is_year_end(ts)
        False
        """
        return ts._get_start_end_field("is_year_end", self)


cdef class SingleConstructorOffset(BaseOffset):
    @classmethod
    def _from_name(cls, suffix=None):
        # default _from_name calls cls with no args
        if suffix:
            raise ValueError(f"Bad freq suffix {suffix}")
        return cls()

    def __reduce__(self):
        # This __reduce__ implementation is for all BaseOffset subclasses
        #  except for RelativeDeltaOffset
        # np.busdaycalendar objects do not pickle nicely, but we can reconstruct
        #  from attributes that do get pickled.
        tup = tuple(
            getattr(self, attr) if attr != "calendar" else None
            for attr in self._attributes
        )
        return type(self), tup


# ---------------------------------------------------------------------
# Tick Offsets

cdef class Tick(SingleConstructorOffset):
    _adjust_dst = False
    _prefix = "undefined"
    _attributes = tuple(["n", "normalize"])

    def __init__(self, n=1, normalize=False):
        n = self._validate_n(n)
        self.n = n
        self.normalize = False
        self._cache = {}
        if normalize:
            # GH#21427
            raise ValueError(
                "Tick offset with `normalize=True` are not allowed."
            )

    # Note: Without making this cpdef, we get AttributeError when calling
    #  from __mul__
    cpdef Tick _next_higher_resolution(Tick self):
        if type(self) is Day:
            return Hour(self.n * 24)
        if type(self) is Hour:
            return Minute(self.n * 60)
        if type(self) is Minute:
            return Second(self.n * 60)
        if type(self) is Second:
            return Milli(self.n * 1000)
        if type(self) is Milli:
            return Micro(self.n * 1000)
        if type(self) is Micro:
            return Nano(self.n * 1000)
        raise ValueError("Could not convert to integer offset at any resolution")

    # --------------------------------------------------------------------

    def _repr_attrs(self) -> str:
        # Since cdef classes have no __dict__, we need to override
        return ""

    @cache_readonly
    def _as_pd_timedelta(self):
        return Timedelta(self)

    @property
    def delta(self):
        warnings.warn(
            # GH#55498
            f"{type(self).__name__}.delta is deprecated and will be removed in "
            "a future version. Use pd.Timedelta(obj) instead",
            FutureWarning,
            stacklevel=find_stack_level(),
        )
        try:
            return self.n * Timedelta(self._nanos_inc)
        except OverflowError as err:
            # GH#55503 as_unit will raise a more useful OutOfBoundsTimedelta
            Timedelta(self).as_unit("ns")
            raise AssertionError("This should not be reached.")

    @property
    def nanos(self) -> int64_t:
        """
        Return an integer of the total number of nanoseconds.

        Raises
        ------
        ValueError
            If the frequency is non-fixed.

        Examples
        --------
        >>> pd.offsets.Hour(5).nanos
        18000000000000
        """
        return self.n * self._nanos_inc

    def is_on_offset(self, dt: datetime) -> bool:
        return True

    # This is identical to BaseOffset.__hash__, but has to be redefined here
    # for Python 3, because we've redefined __eq__.
    def __hash__(self) -> int:
        return hash(self._params)

    # --------------------------------------------------------------------
    # Comparison and Arithmetic Methods

    def __eq__(self, other):
        if isinstance(other, str):
            try:
                # GH#23524 if to_offset fails, we are dealing with an
                #  incomparable type so == is False and != is True
                other = to_offset(other)
            except ValueError:
                # e.g. "infer"
                return False
        return self._as_pd_timedelta == other

    def __ne__(self, other):
        return not (self == other)

    def __le__(self, other):
        return self._as_pd_timedelta.__le__(other)

    def __lt__(self, other):
        return self._as_pd_timedelta.__lt__(other)

    def __ge__(self, other):
        return self._as_pd_timedelta.__ge__(other)

    def __gt__(self, other):
        return self._as_pd_timedelta.__gt__(other)

    def __mul__(self, other):
        if is_float_object(other):
            n = other * self.n
            # If the new `n` is an integer, we can represent it using the
            #  same Tick subclass as self, otherwise we need to move up
            #  to a higher-resolution subclass
            if np.isclose(n % 1, 0):
                return type(self)(int(n))
            new_self = self._next_higher_resolution()
            return new_self * other
        return BaseOffset.__mul__(self, other)

    def __rmul__(self, other):
        return self.__mul__(other)

    def __truediv__(self, other):
        if not isinstance(self, Tick):
            # cython semantics mean the args are sometimes swapped
            result = other._as_pd_timedelta.__rtruediv__(self)
        else:
            result = self._as_pd_timedelta.__truediv__(other)
        return _wrap_timedelta_result(result)

    def __rtruediv__(self, other):
        result = self._as_pd_timedelta.__rtruediv__(other)
        return _wrap_timedelta_result(result)

    def __add__(self, other):
        if isinstance(other, Tick):
            if type(self) is type(other):
                return type(self)(self.n + other.n)
            else:
                return delta_to_tick(self._as_pd_timedelta + other._as_pd_timedelta)
        try:
            return self._apply(other)
        except ApplyTypeError:
            # Includes pd.Period
            return NotImplemented
        except OverflowError as err:
            raise OverflowError(
                f"the add operation between {self} and {other} will overflow"
            ) from err

    def __radd__(self, other):
        return self.__add__(other)

    def _apply(self, other):
        # Timestamp can handle tz and nano sec, thus no need to use apply_wraps
        if isinstance(other, _Timestamp):
            # GH#15126
            return other + self._as_pd_timedelta
        elif other is NaT:
            return NaT
        elif cnp.is_datetime64_object(other) or PyDate_Check(other):
            # PyDate_Check includes date, datetime
            return Timestamp(other) + self

        if cnp.is_timedelta64_object(other) or PyDelta_Check(other):
            return other + self._as_pd_timedelta

        raise ApplyTypeError(f"Unhandled type: {type(other).__name__}")

    # --------------------------------------------------------------------
    # Pickle Methods

    def __setstate__(self, state):
        self.n = state["n"]
        self.normalize = False


cdef class Day(Tick):
    """
    Offset ``n`` days.

    Parameters
    ----------
    n : int, default 1
        The number of days represented.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    You can use the parameter ``n`` to represent a shift of n days.

    >>> from pandas.tseries.offsets import Day
    >>> ts = pd.Timestamp(2022, 12, 9, 15)
    >>> ts
    Timestamp('2022-12-09 15:00:00')

    >>> ts + Day()
    Timestamp('2022-12-10 15:00:00')
    >>> ts - Day(4)
    Timestamp('2022-12-05 15:00:00')

    >>> ts + Day(-4)
    Timestamp('2022-12-05 15:00:00')
    """
    _nanos_inc = 24 * 3600 * 1_000_000_000
    _prefix = "D"
    _period_dtype_code = PeriodDtypeCode.D
    _creso = NPY_DATETIMEUNIT.NPY_FR_D


cdef class Hour(Tick):
    """
    Offset ``n`` hours.

    Parameters
    ----------
    n : int, default 1
        The number of hours represented.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    You can use the parameter ``n`` to represent a shift of n hours.

    >>> from pandas.tseries.offsets import Hour
    >>> ts = pd.Timestamp(2022, 12, 9, 15)
    >>> ts
    Timestamp('2022-12-09 15:00:00')

    >>> ts + Hour()
    Timestamp('2022-12-09 16:00:00')
    >>> ts - Hour(4)
    Timestamp('2022-12-09 11:00:00')

    >>> ts + Hour(-4)
    Timestamp('2022-12-09 11:00:00')
    """
    _nanos_inc = 3600 * 1_000_000_000
    _prefix = "h"
    _period_dtype_code = PeriodDtypeCode.H
    _creso = NPY_DATETIMEUNIT.NPY_FR_h


cdef class Minute(Tick):
    """
    Offset ``n`` minutes.

    Parameters
    ----------
    n : int, default 1
        The number of minutes represented.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    You can use the parameter ``n`` to represent a shift of n minutes.

    >>> from pandas.tseries.offsets import Minute
    >>> ts = pd.Timestamp(2022, 12, 9, 15)
    >>> ts
    Timestamp('2022-12-09 15:00:00')

    >>> ts + Minute(n=10)
    Timestamp('2022-12-09 15:10:00')
    >>> ts - Minute(n=10)
    Timestamp('2022-12-09 14:50:00')

    >>> ts + Minute(n=-10)
    Timestamp('2022-12-09 14:50:00')
    """
    _nanos_inc = 60 * 1_000_000_000
    _prefix = "min"
    _period_dtype_code = PeriodDtypeCode.T
    _creso = NPY_DATETIMEUNIT.NPY_FR_m


cdef class Second(Tick):
    """
    Offset ``n`` seconds.

    Parameters
    ----------
    n : int, default 1
        The number of seconds represented.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    You can use the parameter ``n`` to represent a shift of n seconds.

    >>> from pandas.tseries.offsets import Second
    >>> ts = pd.Timestamp(2022, 12, 9, 15)
    >>> ts
    Timestamp('2022-12-09 15:00:00')

    >>> ts + Second(n=10)
    Timestamp('2022-12-09 15:00:10')
    >>> ts - Second(n=10)
    Timestamp('2022-12-09 14:59:50')

    >>> ts + Second(n=-10)
    Timestamp('2022-12-09 14:59:50')
    """
    _nanos_inc = 1_000_000_000
    _prefix = "s"
    _period_dtype_code = PeriodDtypeCode.S
    _creso = NPY_DATETIMEUNIT.NPY_FR_s


cdef class Milli(Tick):
    """
    Offset ``n`` milliseconds.

    Parameters
    ----------
    n : int, default 1
        The number of milliseconds represented.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    You can use the parameter ``n`` to represent a shift of n milliseconds.

    >>> from pandas.tseries.offsets import Milli
    >>> ts = pd.Timestamp(2022, 12, 9, 15)
    >>> ts
    Timestamp('2022-12-09 15:00:00')

    >>> ts + Milli(n=10)
    Timestamp('2022-12-09 15:00:00.010000')

    >>> ts - Milli(n=10)
    Timestamp('2022-12-09 14:59:59.990000')

    >>> ts + Milli(n=-10)
    Timestamp('2022-12-09 14:59:59.990000')
    """
    _nanos_inc = 1_000_000
    _prefix = "ms"
    _period_dtype_code = PeriodDtypeCode.L
    _creso = NPY_DATETIMEUNIT.NPY_FR_ms


cdef class Micro(Tick):
    """
    Offset ``n`` microseconds.

    Parameters
    ----------
    n : int, default 1
        The number of microseconds represented.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    You can use the parameter ``n`` to represent a shift of n microseconds.

    >>> from pandas.tseries.offsets import Micro
    >>> ts = pd.Timestamp(2022, 12, 9, 15)
    >>> ts
    Timestamp('2022-12-09 15:00:00')

    >>> ts + Micro(n=1000)
    Timestamp('2022-12-09 15:00:00.001000')

    >>> ts - Micro(n=1000)
    Timestamp('2022-12-09 14:59:59.999000')

    >>> ts + Micro(n=-1000)
    Timestamp('2022-12-09 14:59:59.999000')
    """
    _nanos_inc = 1000
    _prefix = "us"
    _period_dtype_code = PeriodDtypeCode.U
    _creso = NPY_DATETIMEUNIT.NPY_FR_us


cdef class Nano(Tick):
    """
    Offset ``n`` nanoseconds.

    Parameters
    ----------
    n : int, default 1
        The number of nanoseconds represented.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    You can use the parameter ``n`` to represent a shift of n nanoseconds.

    >>> from pandas.tseries.offsets import Nano
    >>> ts = pd.Timestamp(2022, 12, 9, 15)
    >>> ts
    Timestamp('2022-12-09 15:00:00')

    >>> ts + Nano(n=1000)
    Timestamp('2022-12-09 15:00:00.000001')

    >>> ts - Nano(n=1000)
    Timestamp('2022-12-09 14:59:59.999999')

    >>> ts + Nano(n=-1000)
    Timestamp('2022-12-09 14:59:59.999999')
    """
    _nanos_inc = 1
    _prefix = "ns"
    _period_dtype_code = PeriodDtypeCode.N
    _creso = NPY_DATETIMEUNIT.NPY_FR_ns


def delta_to_tick(delta: timedelta) -> Tick:
    if delta.microseconds == 0 and getattr(delta, "nanoseconds", 0) == 0:
        # nanoseconds only for pd.Timedelta
        if delta.seconds == 0:
            return Day(delta.days)
        else:
            seconds = delta.days * 86400 + delta.seconds
            if seconds % 3600 == 0:
                return Hour(seconds / 3600)
            elif seconds % 60 == 0:
                return Minute(seconds / 60)
            else:
                return Second(seconds)
    else:
        nanos = delta_to_nanoseconds(delta)
        if nanos % 1_000_000 == 0:
            return Milli(nanos // 1_000_000)
        elif nanos % 1000 == 0:
            return Micro(nanos // 1000)
        else:  # pragma: no cover
            return Nano(nanos)


# --------------------------------------------------------------------

cdef class RelativeDeltaOffset(BaseOffset):
    """
    DateOffset subclass backed by a dateutil relativedelta object.
    """
    _attributes = tuple(["n", "normalize"] + list(_relativedelta_kwds))
    _adjust_dst = False

    def __init__(self, n=1, normalize=False, **kwds):
        BaseOffset.__init__(self, n, normalize)
        off, use_rd = _determine_offset(kwds)
        object.__setattr__(self, "_offset", off)
        object.__setattr__(self, "_use_relativedelta", use_rd)
        for key in kwds:
            val = kwds[key]
            object.__setattr__(self, key, val)

    def __getstate__(self):
        """
        Return a pickleable state
        """
        # RelativeDeltaOffset (technically DateOffset) is the only non-cdef
        #  class, so the only one with __dict__
        state = self.__dict__.copy()
        state["n"] = self.n
        state["normalize"] = self.normalize
        return state

    def __setstate__(self, state):
        """
        Reconstruct an instance from a pickled state
        """

        if "offset" in state:
            # Older (<0.22.0) versions have offset attribute instead of _offset
            if "_offset" in state:  # pragma: no cover
                raise AssertionError("Unexpected key `_offset`")
            state["_offset"] = state.pop("offset")
            state["kwds"]["offset"] = state["_offset"]

        self.n = state.pop("n")
        self.normalize = state.pop("normalize")
        self._cache = state.pop("_cache", {})

        self.__dict__.update(state)

    @apply_wraps
    def _apply(self, other: datetime) -> datetime:
        other_nanos = 0
        if self._use_relativedelta:
            if isinstance(other, _Timestamp):
                other_nanos = other.nanosecond
                other = other.to_pydatetime(warn=False)

        if len(self.kwds) > 0:
            tzinfo = getattr(other, "tzinfo", None)
            if tzinfo is not None and self._use_relativedelta:
                # perform calculation in UTC
                other = other.replace(tzinfo=None)

            other = other + (self._offset * self.n)

            if hasattr(self, "nanoseconds"):
                other = self.n * Timedelta(nanoseconds=self.nanoseconds) + other
            if other_nanos != 0:
                other = Timedelta(nanoseconds=other_nanos) + other

            if tzinfo is not None and self._use_relativedelta:
                # bring tz back from UTC calculation
                other = localize_pydatetime(other, tzinfo)

            return Timestamp(other)
        else:
            return other + timedelta(self.n)

    @cache_readonly
    def _pd_timedelta(self) -> Timedelta:
        # components of _offset that can be cast to pd.Timedelta

        kwds = self.kwds
        relativedelta_fast = {
            "years",
            "months",
            "weeks",
            "days",
            "hours",
            "minutes",
            "seconds",
            "microseconds",
            "milliseconds",
        }
        # relativedelta/_offset path only valid for base DateOffset
        if self._use_relativedelta and set(kwds).issubset(relativedelta_fast):
            td_args = {
                "days",
                "hours",
                "minutes",
                "seconds",
                "microseconds",
                "milliseconds"
            }
            td_kwds = {
                key: val
                for key, val in kwds.items()
                if key in td_args
            }
            if "weeks" in kwds:
                days = td_kwds.get("days", 0)
                td_kwds["days"] = days + 7 * kwds["weeks"]

            if td_kwds:
                delta = Timedelta(**td_kwds)
                if "microseconds" in kwds:
                    delta = delta.as_unit("us")
                elif "milliseconds" in kwds:
                    delta = delta.as_unit("ms")
                else:
                    delta = delta.as_unit("s")
            else:
                delta = Timedelta(0).as_unit("s")

            return delta * self.n

        elif not self._use_relativedelta and hasattr(self, "_offset"):
            # timedelta
            num_nano = getattr(self, "nanoseconds", 0)
            if num_nano != 0:
                rem_nano = Timedelta(nanoseconds=num_nano)
                delta = Timedelta((self._offset + rem_nano) * self.n)
            else:
                delta = Timedelta(self._offset * self.n)
                if "microseconds" in kwds:
                    delta = delta.as_unit("us")
                elif "milliseconds" in kwds:
                    delta = delta.as_unit("ms")
                else:
                    delta = delta.as_unit("s")
            return delta

        else:
            # relativedelta with other keywords
            kwd = set(kwds) - relativedelta_fast
            raise NotImplementedError(
                "DateOffset with relativedelta "
                f"keyword(s) {kwd} not able to be "
                "applied vectorized"
            )

    def _apply_array(self, dtarr: np.ndarray) -> np.ndarray:
        reso = get_unit_from_dtype(dtarr.dtype)
        dt64other = np.asarray(dtarr)

        delta = self._pd_timedelta  # may raise NotImplementedError

        kwds = self.kwds
        months = (kwds.get("years", 0) * 12 + kwds.get("months", 0)) * self.n
        if months:
            shifted = shift_months(dt64other.view("i8"), months, reso=reso)
            dt64other = shifted.view(dtarr.dtype)
        return dt64other + delta

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        return True


class OffsetMeta(type):
    """
    Metaclass that allows us to pretend that all BaseOffset subclasses
    inherit from DateOffset (which is needed for backward-compatibility).
    """

    @classmethod
    def __instancecheck__(cls, obj) -> bool:
        return isinstance(obj, BaseOffset)

    @classmethod
    def __subclasscheck__(cls, obj) -> bool:
        return issubclass(obj, BaseOffset)


# TODO: figure out a way to use a metaclass with a cdef class
class DateOffset(RelativeDeltaOffset, metaclass=OffsetMeta):
    """
    Standard kind of date increment used for a date range.

    Works exactly like the keyword argument form of relativedelta.
    Note that the positional argument form of relativedelta is not
    supported. Use of the keyword n is discouraged-- you would be better
    off specifying n in the keywords you use, but regardless it is
    there for you. n is needed for DateOffset subclasses.

    DateOffset works as follows.  Each offset specify a set of dates
    that conform to the DateOffset.  For example, Bday defines this
    set to be the set of dates that are weekdays (M-F).  To test if a
    date is in the set of a DateOffset dateOffset we can use the
    is_on_offset method: dateOffset.is_on_offset(date).

    If a date is not on a valid date, the rollback and rollforward
    methods can be used to roll the date to the nearest valid date
    before/after the date.

    DateOffsets can be created to move dates forward a given number of
    valid dates.  For example, Bday(2) can be added to a date to move
    it two business days forward.  If the date does not start on a
    valid date, first it is moved to a valid date.  Thus pseudo code
    is::

        def __add__(date):
          date = rollback(date) # does nothing if date is valid
          return date + <n number of periods>

    When a date offset is created for a negative number of periods,
    the date is first rolled forward.  The pseudo code is::

        def __add__(date):
          date = rollforward(date) # does nothing if date is valid
          return date + <n number of periods>

    Zero presents a problem.  Should it roll forward or back?  We
    arbitrarily have it rollforward:

    date + BDay(0) == BDay.rollforward(date)

    Since 0 is a bit weird, we suggest avoiding its use.

    Besides, adding a DateOffsets specified by the singular form of the date
    component can be used to replace certain component of the timestamp.

    Parameters
    ----------
    n : int, default 1
        The number of time periods the offset represents.
        If specified without a temporal pattern, defaults to n days.
    normalize : bool, default False
        Whether to round the result of a DateOffset addition down to the
        previous midnight.
    weekday : int {0, 1, ..., 6}, default 0

        A specific integer for the day of the week.

        - 0 is Monday
        - 1 is Tuesday
        - 2 is Wednesday
        - 3 is Thursday
        - 4 is Friday
        - 5 is Saturday
        - 6 is Sunday

        Instead Weekday type from dateutil.relativedelta can be used.

        - MO is Monday
        - TU is Tuesday
        - WE is Wednesday
        - TH is Thursday
        - FR is Friday
        - SA is Saturday
        - SU is Sunday.

    **kwds
        Temporal parameter that add to or replace the offset value.

        Parameters that **add** to the offset (like Timedelta):

        - years
        - months
        - weeks
        - days
        - hours
        - minutes
        - seconds
        - milliseconds
        - microseconds
        - nanoseconds

        Parameters that **replace** the offset value:

        - year
        - month
        - day
        - weekday
        - hour
        - minute
        - second
        - microsecond
        - nanosecond.

    See Also
    --------
    dateutil.relativedelta.relativedelta : The relativedelta type is designed
        to be applied to an existing datetime an can replace specific components of
        that datetime, or represents an interval of time.

    Examples
    --------
    >>> from pandas.tseries.offsets import DateOffset
    >>> ts = pd.Timestamp('2017-01-01 09:10:11')
    >>> ts + DateOffset(months=3)
    Timestamp('2017-04-01 09:10:11')

    >>> ts = pd.Timestamp('2017-01-01 09:10:11')
    >>> ts + DateOffset(months=2)
    Timestamp('2017-03-01 09:10:11')
    >>> ts + DateOffset(day=31)
    Timestamp('2017-01-31 09:10:11')

    >>> ts + pd.DateOffset(hour=8)
    Timestamp('2017-01-01 08:10:11')
    """
    def __setattr__(self, name, value):
        raise AttributeError("DateOffset objects are immutable.")

# --------------------------------------------------------------------


cdef class BusinessMixin(SingleConstructorOffset):
    """
    Mixin to business types to provide related functions.
    """

    cdef readonly:
        timedelta _offset
        # Only Custom subclasses use weekmask, holiday, calendar
        object weekmask, holidays, calendar

    def __init__(self, n=1, normalize=False, offset=timedelta(0)):
        BaseOffset.__init__(self, n, normalize)
        self._offset = offset

    cpdef _init_custom(self, weekmask, holidays, calendar):
        """
        Additional __init__ for Custom subclasses.
        """
        calendar, holidays = _get_calendar(
            weekmask=weekmask, holidays=holidays, calendar=calendar
        )
        # Custom offset instances are identified by the
        # following two attributes. See DateOffset._params()
        # holidays, weekmask
        self.weekmask = weekmask
        self.holidays = holidays
        self.calendar = calendar

    @property
    def offset(self):
        """
        Alias for self._offset.
        """
        # Alias for backward compat
        return self._offset

    def _repr_attrs(self) -> str:
        if self.offset:
            attrs = [f"offset={repr(self.offset)}"]
        else:
            attrs = []
        out = ""
        if attrs:
            out += ": " + ", ".join(attrs)
        return out

    cpdef __setstate__(self, state):
        # We need to use a cdef/cpdef method to set the readonly _offset attribute
        if "_offset" in state:
            self._offset = state.pop("_offset")
        elif "offset" in state:
            # Older (<0.22.0) versions have offset attribute instead of _offset
            self._offset = state.pop("offset")

        if self._prefix.startswith(("C", "c")):
            # i.e. this is a Custom class
            weekmask = state.pop("weekmask")
            holidays = state.pop("holidays")
            calendar, holidays = _get_calendar(weekmask=weekmask,
                                               holidays=holidays,
                                               calendar=None)
            self.weekmask = weekmask
            self.calendar = calendar
            self.holidays = holidays

        BaseOffset.__setstate__(self, state)


cdef class BusinessDay(BusinessMixin):
    """
    DateOffset subclass representing possibly n business days.

    Parameters
    ----------
    n : int, default 1
        The number of days represented.
    normalize : bool, default False
        Normalize start/end dates to midnight.
    offset : timedelta, default timedelta(0)
        Time offset to apply.

    Examples
    --------
    You can use the parameter ``n`` to represent a shift of n business days.

    >>> ts = pd.Timestamp(2022, 12, 9, 15)
    >>> ts.strftime('%a %d %b %Y %H:%M')
    'Fri 09 Dec 2022 15:00'
    >>> (ts + pd.offsets.BusinessDay(n=5)).strftime('%a %d %b %Y %H:%M')
    'Fri 16 Dec 2022 15:00'

    Passing the parameter ``normalize`` equal to True, you shift the start
    of the next business day to midnight.

    >>> ts = pd.Timestamp(2022, 12, 9, 15)
    >>> ts + pd.offsets.BusinessDay(normalize=True)
    Timestamp('2022-12-12 00:00:00')
    """
    _period_dtype_code = PeriodDtypeCode.B
    _prefix = "B"
    _attributes = tuple(["n", "normalize", "offset"])

    cpdef __setstate__(self, state):
        self.n = state.pop("n")
        self.normalize = state.pop("normalize")
        if "_offset" in state:
            self._offset = state.pop("_offset")
        elif "offset" in state:
            self._offset = state.pop("offset")
        self._cache = state.pop("_cache", {})

    def _offset_str(self) -> str:
        def get_str(td):
            off_str = ""
            if td.days > 0:
                off_str += str(td.days) + "D"
            if td.seconds > 0:
                s = td.seconds
                hrs = int(s / 3600)
                if hrs != 0:
                    off_str += str(hrs) + "h"
                    s -= hrs * 3600
                mts = int(s / 60)
                if mts != 0:
                    off_str += str(mts) + "Min"
                    s -= mts * 60
                if s != 0:
                    off_str += str(s) + "s"
            if td.microseconds > 0:
                off_str += str(td.microseconds) + "us"
            return off_str

        if PyDelta_Check(self.offset):
            zero = timedelta(0, 0, 0)
            if self.offset >= zero:
                off_str = "+" + get_str(self.offset)
            else:
                off_str = "-" + get_str(-self.offset)
            return off_str
        else:
            return "+" + repr(self.offset)

    @apply_wraps
    def _apply(self, other):
        if PyDateTime_Check(other):
            n = self.n
            wday = other.weekday()

            # avoid slowness below by operating on weeks first
            weeks = n // 5
            days = self._adjust_ndays(wday, weeks)

            result = other + timedelta(days=7 * weeks + days)
            if self.offset:
                result = result + self.offset
            return result

        elif is_any_td_scalar(other):
            td = Timedelta(self.offset) + other
            return BusinessDay(
                self.n, offset=td.to_pytimedelta(), normalize=self.normalize
            )
        else:
            raise ApplyTypeError(
                "Only know how to combine business day with datetime or timedelta."
            )

    @cython.wraparound(False)
    @cython.boundscheck(False)
    cdef ndarray _shift_bdays(
        self,
        ndarray i8other,
        NPY_DATETIMEUNIT reso=NPY_DATETIMEUNIT.NPY_FR_ns,
    ):
        """
        Implementation of BusinessDay.apply_offset.

        Parameters
        ----------
        i8other : const int64_t[:]
        reso : NPY_DATETIMEUNIT, default NPY_FR_ns

        Returns
        -------
        ndarray[int64_t]
        """
        cdef:
            int periods = self.n
            Py_ssize_t i, n = i8other.size
            ndarray result = cnp.PyArray_EMPTY(
                i8other.ndim, i8other.shape, cnp.NPY_INT64, 0
            )
            int64_t val, res_val
            int wday, days
            npy_datetimestruct dts
            int64_t DAY_PERIODS = periods_per_day(reso)
            cnp.broadcast mi = cnp.PyArray_MultiIterNew2(result, i8other)

        for i in range(n):
            # Analogous to: val = i8other[i]
            val = (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 1))[0]

            if val == NPY_NAT:
                res_val = NPY_NAT
            else:
                # The rest of this is effectively a copy of BusinessDay.apply
                weeks = periods // 5
                pandas_datetime_to_datetimestruct(val, reso, &dts)
                wday = dayofweek(dts.year, dts.month, dts.day)

                days = self._adjust_ndays(wday, weeks)
                res_val = val + (7 * weeks + days) * DAY_PERIODS

            # Analogous to: out[i] = res_val
            (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 0))[0] = res_val

            cnp.PyArray_MultiIter_NEXT(mi)

        return result

    cdef int _adjust_ndays(self, int wday, int weeks):
        cdef:
            int n = self.n
            int days

        if n <= 0 and wday > 4:
            # roll forward
            n += 1

        n -= 5 * weeks

        # n is always >= 0 at this point
        if n == 0 and wday > 4:
            # roll back
            days = 4 - wday
        elif wday > 4:
            # roll forward
            days = (7 - wday) + (n - 1)
        elif wday + n <= 4:
            # shift by n days without leaving the current week
            days = n
        else:
            # shift by n days plus 2 to get past the weekend
            days = n + 2
        return days

    def _apply_array(self, dtarr: np.ndarray) -> np.ndarray:
        i8other = dtarr.view("i8")
        reso = get_unit_from_dtype(dtarr.dtype)
        res = self._shift_bdays(i8other, reso=reso)
        if self.offset:
            res = res.view(dtarr.dtype) + Timedelta(self.offset)
        return res

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        return dt.weekday() < 5


cdef class BusinessHour(BusinessMixin):
    """
    DateOffset subclass representing possibly n business hours.

    Parameters
    ----------
    n : int, default 1
        The number of hours represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    start : str, time, or list of str/time, default "09:00"
        Start time of your custom business hour in 24h format.
    end : str, time, or list of str/time, default: "17:00"
        End time of your custom business hour in 24h format.
    offset : timedelta, default timedelta(0)
        Time offset to apply.

    Examples
    --------
    You can use the parameter ``n`` to represent a shift of n hours.

    >>> ts = pd.Timestamp(2022, 12, 9, 8)
    >>> ts + pd.offsets.BusinessHour(n=5)
    Timestamp('2022-12-09 14:00:00')

    You can also change the start and the end of business hours.

    >>> ts = pd.Timestamp(2022, 8, 5, 16)
    >>> ts + pd.offsets.BusinessHour(start="11:00")
    Timestamp('2022-08-08 11:00:00')

    >>> from datetime import time as dt_time
    >>> ts = pd.Timestamp(2022, 8, 5, 22)
    >>> ts + pd.offsets.BusinessHour(end=dt_time(19, 0))
    Timestamp('2022-08-08 10:00:00')

    Passing the parameter ``normalize`` equal to True, you shift the start
    of the next business hour to midnight.

    >>> ts = pd.Timestamp(2022, 12, 9, 8)
    >>> ts + pd.offsets.BusinessHour(normalize=True)
    Timestamp('2022-12-09 00:00:00')

    You can divide your business day hours into several parts.

    >>> import datetime as dt
    >>> freq = pd.offsets.BusinessHour(start=["06:00", "10:00", "15:00"],
    ...                                end=["08:00", "12:00", "17:00"])
    >>> pd.date_range(dt.datetime(2022, 12, 9), dt.datetime(2022, 12, 13), freq=freq)
    DatetimeIndex(['2022-12-09 06:00:00', '2022-12-09 07:00:00',
                   '2022-12-09 10:00:00', '2022-12-09 11:00:00',
                   '2022-12-09 15:00:00', '2022-12-09 16:00:00',
                   '2022-12-12 06:00:00', '2022-12-12 07:00:00',
                   '2022-12-12 10:00:00', '2022-12-12 11:00:00',
                   '2022-12-12 15:00:00', '2022-12-12 16:00:00'],
                   dtype='datetime64[ns]', freq='bh')
    """

    _prefix = "bh"
    _anchor = 0
    _attributes = tuple(["n", "normalize", "start", "end", "offset"])
    _adjust_dst = False

    cdef readonly:
        tuple start, end

    def __init__(
            self, n=1, normalize=False, start="09:00", end="17:00", offset=timedelta(0)
    ):
        BusinessMixin.__init__(self, n, normalize, offset)

        # must be validated here to equality check
        if np.ndim(start) == 0:
            # i.e. not is_list_like
            start = [start]
        if not len(start):
            raise ValueError("Must include at least 1 start time")

        if np.ndim(end) == 0:
            # i.e. not is_list_like
            end = [end]
        if not len(end):
            raise ValueError("Must include at least 1 end time")

        start = np.array([_validate_business_time(x) for x in start])
        end = np.array([_validate_business_time(x) for x in end])

        # Validation of input
        if len(start) != len(end):
            raise ValueError("number of starting time and ending time must be the same")
        num_openings = len(start)

        # sort starting and ending time by starting time
        index = np.argsort(start)

        # convert to tuple so that start and end are hashable
        start = tuple(start[index])
        end = tuple(end[index])

        total_secs = 0
        for i in range(num_openings):
            total_secs += self._get_business_hours_by_sec(start[i], end[i])
            total_secs += self._get_business_hours_by_sec(
                end[i], start[(i + 1) % num_openings]
            )
        if total_secs != 24 * 60 * 60:
            raise ValueError(
                "invalid starting and ending time(s): "
                "opening hours should not touch or overlap with "
                "one another"
            )

        self.start = start
        self.end = end

    cpdef __setstate__(self, state):
        start = state.pop("start")
        start = (start,) if np.ndim(start) == 0 else tuple(start)
        end = state.pop("end")
        end = (end,) if np.ndim(end) == 0 else tuple(end)
        self.start = start
        self.end = end

        state.pop("kwds", {})
        state.pop("next_bday", None)
        BusinessMixin.__setstate__(self, state)

    def _repr_attrs(self) -> str:
        out = super()._repr_attrs()
        # Use python string formatting to be faster than strftime
        hours = ",".join(
            f"{st.hour:02d}:{st.minute:02d}-{en.hour:02d}:{en.minute:02d}"
            for st, en in zip(self.start, self.end)
        )
        attrs = [f"{self._prefix}={hours}"]
        out += ": " + ", ".join(attrs)
        return out

    def _get_business_hours_by_sec(self, start, end):
        """
        Return business hours in a day by seconds.
        """
        # create dummy datetime to calculate business hours in a day
        dtstart = datetime(2014, 4, 1, start.hour, start.minute)
        day = 1 if start < end else 2
        until = datetime(2014, 4, day, end.hour, end.minute)
        return int((until - dtstart).total_seconds())

    def _get_closing_time(self, dt: datetime) -> datetime:
        """
        Get the closing time of a business hour interval by its opening time.

        Parameters
        ----------
        dt : datetime
            Opening time of a business hour interval.

        Returns
        -------
        result : datetime
            Corresponding closing time.
        """
        for i, st in enumerate(self.start):
            if st.hour == dt.hour and st.minute == dt.minute:
                return dt + timedelta(
                    seconds=self._get_business_hours_by_sec(st, self.end[i])
                )
        assert False

    @cache_readonly
    def next_bday(self):
        """
        Used for moving to next business day.
        """
        if self.n >= 0:
            nb_offset = 1
        else:
            nb_offset = -1
        if self._prefix.startswith(("c")):
            # CustomBusinessHour
            return CustomBusinessDay(
                n=nb_offset,
                weekmask=self.weekmask,
                holidays=self.holidays,
                calendar=self.calendar,
            )
        else:
            return BusinessDay(n=nb_offset)

    def _next_opening_time(self, other, sign=1):
        """
        If self.n and sign have the same sign, return the earliest opening time
        later than or equal to current time.
        Otherwise the latest opening time earlier than or equal to current
        time.

        Opening time always locates on BusinessDay.
        However, closing time may not if business hour extends over midnight.

        Parameters
        ----------
        other : datetime
            Current time.
        sign : int, default 1.
            Either 1 or -1. Going forward in time if it has the same sign as
            self.n. Going backward in time otherwise.

        Returns
        -------
        result : datetime
            Next opening time.
        """
        earliest_start = self.start[0]
        latest_start = self.start[-1]

        if self.n == 0:
            is_same_sign = sign > 0
        else:
            is_same_sign = self.n * sign >= 0

        if not self.next_bday.is_on_offset(other):
            # today is not business day
            other = other + sign * self.next_bday
            if is_same_sign:
                hour, minute = earliest_start.hour, earliest_start.minute
            else:
                hour, minute = latest_start.hour, latest_start.minute
        else:
            if is_same_sign:
                if latest_start < other.time():
                    # current time is after latest starting time in today
                    other = other + sign * self.next_bday
                    hour, minute = earliest_start.hour, earliest_start.minute
                else:
                    # find earliest starting time no earlier than current time
                    for st in self.start:
                        if other.time() <= st:
                            hour, minute = st.hour, st.minute
                            break
            else:
                if other.time() < earliest_start:
                    # current time is before earliest starting time in today
                    other = other + sign * self.next_bday
                    hour, minute = latest_start.hour, latest_start.minute
                else:
                    # find latest starting time no later than current time
                    for st in reversed(self.start):
                        if other.time() >= st:
                            hour, minute = st.hour, st.minute
                            break

        return datetime(other.year, other.month, other.day, hour, minute)

    def _prev_opening_time(self, other: datetime) -> datetime:
        """
        If n is positive, return the latest opening time earlier than or equal
        to current time.
        Otherwise the earliest opening time later than or equal to current
        time.

        Parameters
        ----------
        other : datetime
            Current time.

        Returns
        -------
        result : datetime
            Previous opening time.
        """
        return self._next_opening_time(other, sign=-1)

    @apply_wraps
    def rollback(self, dt: datetime) -> datetime:
        """
        Roll provided date backward to next offset only if not on offset.
        """
        if not self.is_on_offset(dt):
            if self.n >= 0:
                dt = self._prev_opening_time(dt)
            else:
                dt = self._next_opening_time(dt)
            return self._get_closing_time(dt)
        return dt

    @apply_wraps
    def rollforward(self, dt: datetime) -> datetime:
        """
        Roll provided date forward to next offset only if not on offset.
        """
        if not self.is_on_offset(dt):
            if self.n >= 0:
                return self._next_opening_time(dt)
            else:
                return self._prev_opening_time(dt)
        return dt

    @apply_wraps
    def _apply(self, other: datetime) -> datetime:
        # used for detecting edge condition
        nanosecond = getattr(other, "nanosecond", 0)
        # reset timezone and nanosecond
        # other may be a Timestamp, thus not use replace
        other = datetime(
            other.year,
            other.month,
            other.day,
            other.hour,
            other.minute,
            other.second,
            other.microsecond,
        )
        n = self.n

        # adjust other to reduce number of cases to handle
        if n >= 0:
            if other.time() in self.end or not self._is_on_offset(other):
                other = self._next_opening_time(other)
        else:
            if other.time() in self.start:
                # adjustment to move to previous business day
                other = other - timedelta(seconds=1)
            if not self._is_on_offset(other):
                other = self._next_opening_time(other)
                other = self._get_closing_time(other)

        # get total business hours by sec in one business day
        businesshours = sum(
            self._get_business_hours_by_sec(st, en)
            for st, en in zip(self.start, self.end)
        )

        bd, r = divmod(abs(n * 60), businesshours // 60)
        if n < 0:
            bd, r = -bd, -r

        # adjust by business days first
        if bd != 0:
            if self._prefix.startswith("c"):
                # GH#30593 this is a Custom offset
                skip_bd = CustomBusinessDay(
                    n=bd,
                    weekmask=self.weekmask,
                    holidays=self.holidays,
                    calendar=self.calendar,
                )
            else:
                skip_bd = BusinessDay(n=bd)
            # midnight business hour may not on BusinessDay
            if not self.next_bday.is_on_offset(other):
                prev_open = self._prev_opening_time(other)
                remain = other - prev_open
                other = prev_open + skip_bd + remain
            else:
                other = other + skip_bd

        # remaining business hours to adjust
        bhour_remain = timedelta(minutes=r)

        if n >= 0:
            while bhour_remain != timedelta(0):
                # business hour left in this business time interval
                bhour = (
                    self._get_closing_time(self._prev_opening_time(other)) - other
                )
                if bhour_remain < bhour:
                    # finish adjusting if possible
                    other += bhour_remain
                    bhour_remain = timedelta(0)
                else:
                    # go to next business time interval
                    bhour_remain -= bhour
                    other = self._next_opening_time(other + bhour)
        else:
            while bhour_remain != timedelta(0):
                # business hour left in this business time interval
                bhour = self._next_opening_time(other) - other
                if (
                    bhour_remain > bhour
                    or bhour_remain == bhour
                    and nanosecond != 0
                ):
                    # finish adjusting if possible
                    other += bhour_remain
                    bhour_remain = timedelta(0)
                else:
                    # go to next business time interval
                    bhour_remain -= bhour
                    other = self._get_closing_time(
                        self._next_opening_time(
                            other + bhour - timedelta(seconds=1)
                        )
                    )

        return other

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False

        if dt.tzinfo is not None:
            dt = datetime(
                dt.year, dt.month, dt.day, dt.hour, dt.minute, dt.second, dt.microsecond
            )
        # Valid bh can be on the different BusinessDay during midnight
        # Distinguish by the time spent from previous opening time
        return self._is_on_offset(dt)

    def _is_on_offset(self, dt: datetime) -> bool:
        """
        Slight speedups using calculated values.
        """
        # if self.normalize and not _is_normalized(dt):
        #     return False
        # Valid bh can be on the different BusinessDay during midnight
        # Distinguish by the time spent from previous opening time
        if self.n >= 0:
            op = self._prev_opening_time(dt)
        else:
            op = self._next_opening_time(dt)
        span = (dt - op).total_seconds()
        businesshours = 0
        for i, st in enumerate(self.start):
            if op.hour == st.hour and op.minute == st.minute:
                businesshours = self._get_business_hours_by_sec(st, self.end[i])
        if span <= businesshours:
            return True
        else:
            return False


cdef class WeekOfMonthMixin(SingleConstructorOffset):
    """
    Mixin for methods common to WeekOfMonth and LastWeekOfMonth.
    """

    cdef readonly:
        int weekday, week

    def __init__(self, n=1, normalize=False, weekday=0):
        BaseOffset.__init__(self, n, normalize)
        self.weekday = weekday

        if weekday < 0 or weekday > 6:
            raise ValueError(f"Day must be 0<=day<=6, got {weekday}")

    @apply_wraps
    def _apply(self, other: datetime) -> datetime:
        compare_day = self._get_offset_day(other)

        months = self.n
        months = roll_convention(other.day, months, compare_day)

        shifted = shift_month(other, months, "start")
        to_day = self._get_offset_day(shifted)
        return _shift_day(shifted, to_day - shifted.day)

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        return dt.day == self._get_offset_day(dt)

    @property
    def rule_code(self) -> str:
        weekday = int_to_weekday.get(self.weekday, "")
        if self.week == -1:
            # LastWeekOfMonth
            return f"{self._prefix}-{weekday}"
        return f"{self._prefix}-{self.week + 1}{weekday}"


# ----------------------------------------------------------------------
# Year-Based Offset Classes

cdef class YearOffset(SingleConstructorOffset):
    """
    DateOffset that just needs a month.
    """
    _attributes = tuple(["n", "normalize", "month"])

    # FIXME(cython#4446): python annotation here gives compile-time errors
    # _default_month: int

    cdef readonly:
        int month

    def __init__(self, n=1, normalize=False, month=None):
        BaseOffset.__init__(self, n, normalize)

        month = month if month is not None else self._default_month
        self.month = month

        if month < 1 or month > 12:
            raise ValueError("Month must go from 1 to 12")

    cpdef __setstate__(self, state):
        self.month = state.pop("month")
        self.n = state.pop("n")
        self.normalize = state.pop("normalize")
        self._cache = {}

    @classmethod
    def _from_name(cls, suffix=None):
        kwargs = {}
        if suffix:
            kwargs["month"] = MONTH_TO_CAL_NUM[suffix]
        return cls(**kwargs)

    @property
    def rule_code(self) -> str:
        month = MONTH_ALIASES[self.month]
        return f"{self._prefix}-{month}"

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        return dt.month == self.month and dt.day == self._get_offset_day(dt)

    def _get_offset_day(self, other: datetime) -> int:
        # override BaseOffset method to use self.month instead of other.month
        cdef:
            npy_datetimestruct dts
        pydate_to_dtstruct(other, &dts)
        dts.month = self.month
        return get_day_of_month(&dts, self._day_opt)

    @apply_wraps
    def _apply(self, other: datetime) -> datetime:
        years = roll_qtrday(other, self.n, self.month, self._day_opt, modby=12)
        months = years * 12 + (self.month - other.month)
        return shift_month(other, months, self._day_opt)

    def _apply_array(self, dtarr: np.ndarray) -> np.ndarray:
        reso = get_unit_from_dtype(dtarr.dtype)
        shifted = shift_quarters(
            dtarr.view("i8"), self.n, self.month, self._day_opt, modby=12, reso=reso
        )
        return shifted


cdef class BYearEnd(YearOffset):
    """
    DateOffset increments between the last business day of the year.

    Parameters
    ----------
    n : int, default 1
        The number of years represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    month : int, default 12
        A specific integer for the month of the year.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> from pandas.tseries.offsets import BYearEnd
    >>> ts = pd.Timestamp('2020-05-24 05:01:15')
    >>> ts - BYearEnd()
    Timestamp('2019-12-31 05:01:15')
    >>> ts + BYearEnd()
    Timestamp('2020-12-31 05:01:15')
    >>> ts + BYearEnd(3)
    Timestamp('2022-12-30 05:01:15')
    >>> ts + BYearEnd(-3)
    Timestamp('2017-12-29 05:01:15')
    >>> ts + BYearEnd(month=11)
    Timestamp('2020-11-30 05:01:15')
    """

    _outputName = "BusinessYearEnd"
    _default_month = 12
    _prefix = "BYE"
    _day_opt = "business_end"


cdef class BYearBegin(YearOffset):
    """
    DateOffset increments between the first business day of the year.

    Parameters
    ----------
    n : int, default 1
        The number of years represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    month : int, default 1
        A specific integer for the month of the year.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> from pandas.tseries.offsets import BYearBegin
    >>> ts = pd.Timestamp('2020-05-24 05:01:15')
    >>> ts + BYearBegin()
    Timestamp('2021-01-01 05:01:15')
    >>> ts - BYearBegin()
    Timestamp('2020-01-01 05:01:15')
    >>> ts + BYearBegin(-1)
    Timestamp('2020-01-01 05:01:15')
    >>> ts + BYearBegin(2)
    Timestamp('2022-01-03 05:01:15')
    >>> ts + BYearBegin(month=11)
    Timestamp('2020-11-02 05:01:15')
    """

    _outputName = "BusinessYearBegin"
    _default_month = 1
    _prefix = "BYS"
    _day_opt = "business_start"


cdef class YearEnd(YearOffset):
    """
    DateOffset increments between calendar year end dates.

    YearEnd goes to the next date which is the end of the year.

    Parameters
    ----------
    n : int, default 1
        The number of years represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    month : int, default 12
        A specific integer for the month of the year.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.YearEnd()
    Timestamp('2022-12-31 00:00:00')

    >>> ts = pd.Timestamp(2022, 12, 31)
    >>> ts + pd.offsets.YearEnd()
    Timestamp('2023-12-31 00:00:00')

    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.YearEnd(month=2)
    Timestamp('2022-02-28 00:00:00')

    If you want to get the end of the current year:

    >>> ts = pd.Timestamp(2022, 12, 31)
    >>> pd.offsets.YearEnd().rollforward(ts)
    Timestamp('2022-12-31 00:00:00')
    """

    _default_month = 12
    _prefix = "YE"
    _day_opt = "end"

    cdef readonly:
        int _period_dtype_code

    def __init__(self, n=1, normalize=False, month=None):
        # Because YearEnd can be the freq for a Period, define its
        #  _period_dtype_code at construction for performance
        YearOffset.__init__(self, n, normalize, month)
        self._period_dtype_code = PeriodDtypeCode.A + self.month % 12


cdef class YearBegin(YearOffset):
    """
    DateOffset increments between calendar year begin dates.

    YearBegin goes to the next date which is the start of the year.

    Parameters
    ----------
    n : int, default 1
        The number of years represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    month : int, default 1
        A specific integer for the month of the year.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 12, 1)
    >>> ts + pd.offsets.YearBegin()
    Timestamp('2023-01-01 00:00:00')

    >>> ts = pd.Timestamp(2023, 1, 1)
    >>> ts + pd.offsets.YearBegin()
    Timestamp('2024-01-01 00:00:00')

    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.YearBegin(month=2)
    Timestamp('2022-02-01 00:00:00')

    If you want to get the start of the current year:

    >>> ts = pd.Timestamp(2023, 1, 1)
    >>> pd.offsets.YearBegin().rollback(ts)
    Timestamp('2023-01-01 00:00:00')
    """

    _default_month = 1
    _prefix = "YS"
    _day_opt = "start"


# ----------------------------------------------------------------------
# Quarter-Based Offset Classes

cdef class QuarterOffset(SingleConstructorOffset):
    _attributes = tuple(["n", "normalize", "startingMonth"])
    # TODO: Consider combining QuarterOffset and YearOffset __init__ at some
    #       point.  Also apply_index, is_on_offset, rule_code if
    #       startingMonth vs month attr names are resolved

    # FIXME(cython#4446): python annotation here gives compile-time errors
    # _default_starting_month: int
    # _from_name_starting_month: int

    cdef readonly:
        int startingMonth

    def __init__(self, n=1, normalize=False, startingMonth=None):
        BaseOffset.__init__(self, n, normalize)

        if startingMonth is None:
            startingMonth = self._default_starting_month
        self.startingMonth = startingMonth

    cpdef __setstate__(self, state):
        self.startingMonth = state.pop("startingMonth")
        self.n = state.pop("n")
        self.normalize = state.pop("normalize")

    @classmethod
    def _from_name(cls, suffix=None):
        kwargs = {}
        if suffix:
            kwargs["startingMonth"] = MONTH_TO_CAL_NUM[suffix]
        else:
            if cls._from_name_starting_month is not None:
                kwargs["startingMonth"] = cls._from_name_starting_month
        return cls(**kwargs)

    @property
    def rule_code(self) -> str:
        month = MONTH_ALIASES[self.startingMonth]
        return f"{self._prefix}-{month}"

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        mod_month = (dt.month - self.startingMonth) % 3
        return mod_month == 0 and dt.day == self._get_offset_day(dt)

    @apply_wraps
    def _apply(self, other: datetime) -> datetime:
        # months_since: find the calendar quarter containing other.month,
        # e.g. if other.month == 8, the calendar quarter is [Jul, Aug, Sep].
        # Then find the month in that quarter containing an is_on_offset date for
        # self.  `months_since` is the number of months to shift other.month
        # to get to this on-offset month.
        months_since = other.month % 3 - self.startingMonth % 3
        qtrs = roll_qtrday(
            other, self.n, self.startingMonth, day_opt=self._day_opt, modby=3
        )
        months = qtrs * 3 - months_since
        return shift_month(other, months, self._day_opt)

    def _apply_array(self, dtarr: np.ndarray) -> np.ndarray:
        reso = get_unit_from_dtype(dtarr.dtype)
        shifted = shift_quarters(
            dtarr.view("i8"),
            self.n,
            self.startingMonth,
            self._day_opt,
            modby=3,
            reso=reso,
        )
        return shifted


cdef class BQuarterEnd(QuarterOffset):
    """
    DateOffset increments between the last business day of each Quarter.

    startingMonth = 1 corresponds to dates like 1/31/2007, 4/30/2007, ...
    startingMonth = 2 corresponds to dates like 2/28/2007, 5/31/2007, ...
    startingMonth = 3 corresponds to dates like 3/30/2007, 6/29/2007, ...

    Parameters
    ----------
    n : int, default 1
        The number of quarters represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    startingMonth : int, default 3
        A specific integer for the month of the year from which we start quarters.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> from pandas.tseries.offsets import BQuarterEnd
    >>> ts = pd.Timestamp('2020-05-24 05:01:15')
    >>> ts + BQuarterEnd()
    Timestamp('2020-06-30 05:01:15')
    >>> ts + BQuarterEnd(2)
    Timestamp('2020-09-30 05:01:15')
    >>> ts + BQuarterEnd(1, startingMonth=2)
    Timestamp('2020-05-29 05:01:15')
    >>> ts + BQuarterEnd(startingMonth=2)
    Timestamp('2020-05-29 05:01:15')
    """
    _output_name = "BusinessQuarterEnd"
    _default_starting_month = 3
    _from_name_starting_month = 12
    _prefix = "BQE"
    _day_opt = "business_end"


cdef class BQuarterBegin(QuarterOffset):
    """
    DateOffset increments between the first business day of each Quarter.

    startingMonth = 1 corresponds to dates like 1/01/2007, 4/01/2007, ...
    startingMonth = 2 corresponds to dates like 2/01/2007, 5/01/2007, ...
    startingMonth = 3 corresponds to dates like 3/01/2007, 6/01/2007, ...

    Parameters
    ----------
    n : int, default 1
        The number of quarters represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    startingMonth : int, default 3
        A specific integer for the month of the year from which we start quarters.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> from pandas.tseries.offsets import BQuarterBegin
    >>> ts = pd.Timestamp('2020-05-24 05:01:15')
    >>> ts + BQuarterBegin()
    Timestamp('2020-06-01 05:01:15')
    >>> ts + BQuarterBegin(2)
    Timestamp('2020-09-01 05:01:15')
    >>> ts + BQuarterBegin(startingMonth=2)
    Timestamp('2020-08-03 05:01:15')
    >>> ts + BQuarterBegin(-1)
    Timestamp('2020-03-02 05:01:15')
    """
    _output_name = "BusinessQuarterBegin"
    _default_starting_month = 3
    _from_name_starting_month = 1
    _prefix = "BQS"
    _day_opt = "business_start"


cdef class QuarterEnd(QuarterOffset):
    """
    DateOffset increments between Quarter end dates.

    startingMonth = 1 corresponds to dates like 1/31/2007, 4/30/2007, ...
    startingMonth = 2 corresponds to dates like 2/28/2007, 5/31/2007, ...
    startingMonth = 3 corresponds to dates like 3/31/2007, 6/30/2007, ...

    Parameters
    ----------
    n : int, default 1
        The number of quarters represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    startingMonth : int, default 3
        A specific integer for the month of the year from which we start quarters.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.QuarterEnd()
    Timestamp('2022-03-31 00:00:00')
    """
    _default_starting_month = 3
    _prefix = "QE"
    _day_opt = "end"

    cdef readonly:
        int _period_dtype_code

    def __init__(self, n=1, normalize=False, startingMonth=None):
        # Because QuarterEnd can be the freq for a Period, define its
        #  _period_dtype_code at construction for performance
        QuarterOffset.__init__(self, n, normalize, startingMonth)
        self._period_dtype_code = PeriodDtypeCode.Q_DEC + self.startingMonth % 12


cdef class QuarterBegin(QuarterOffset):
    """
    DateOffset increments between Quarter start dates.

    startingMonth = 1 corresponds to dates like 1/01/2007, 4/01/2007, ...
    startingMonth = 2 corresponds to dates like 2/01/2007, 5/01/2007, ...
    startingMonth = 3 corresponds to dates like 3/01/2007, 6/01/2007, ...

    Parameters
    ----------
    n : int, default 1
        The number of quarters represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    startingMonth : int, default 3
        A specific integer for the month of the year from which we start quarters.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.QuarterBegin()
    Timestamp('2022-03-01 00:00:00')
    """
    _default_starting_month = 3
    _from_name_starting_month = 1
    _prefix = "QS"
    _day_opt = "start"


# ----------------------------------------------------------------------
# Month-Based Offset Classes

cdef class MonthOffset(SingleConstructorOffset):
    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        return dt.day == self._get_offset_day(dt)

    @apply_wraps
    def _apply(self, other: datetime) -> datetime:
        compare_day = self._get_offset_day(other)
        n = roll_convention(other.day, self.n, compare_day)
        return shift_month(other, n, self._day_opt)

    def _apply_array(self, dtarr: np.ndarray) -> np.ndarray:
        reso = get_unit_from_dtype(dtarr.dtype)
        shifted = shift_months(dtarr.view("i8"), self.n, self._day_opt, reso=reso)
        return shifted

    cpdef __setstate__(self, state):
        state.pop("_use_relativedelta", False)
        state.pop("offset", None)
        state.pop("_offset", None)
        state.pop("kwds", {})

        BaseOffset.__setstate__(self, state)


cdef class MonthEnd(MonthOffset):
    """
    DateOffset of one month end.

    MonthEnd goes to the next date which is an end of the month.

    Parameters
    ----------
    n : int, default 1
        The number of months represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 1, 30)
    >>> ts + pd.offsets.MonthEnd()
    Timestamp('2022-01-31 00:00:00')

    >>> ts = pd.Timestamp(2022, 1, 31)
    >>> ts + pd.offsets.MonthEnd()
    Timestamp('2022-02-28 00:00:00')

    If you want to get the end of the current month:

    >>> ts = pd.Timestamp(2022, 1, 31)
    >>> pd.offsets.MonthEnd().rollforward(ts)
    Timestamp('2022-01-31 00:00:00')
    """
    _period_dtype_code = PeriodDtypeCode.M
    _prefix = "ME"
    _day_opt = "end"


cdef class MonthBegin(MonthOffset):
    """
    DateOffset of one month at beginning.

    MonthBegin goes to the next date which is a start of the month.

    Parameters
    ----------
    n : int, default 1
        The number of months represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 11, 30)
    >>> ts + pd.offsets.MonthBegin()
    Timestamp('2022-12-01 00:00:00')

    >>> ts = pd.Timestamp(2022, 12, 1)
    >>> ts + pd.offsets.MonthBegin()
    Timestamp('2023-01-01 00:00:00')

    If you want to get the start of the current month:

    >>> ts = pd.Timestamp(2022, 12, 1)
    >>> pd.offsets.MonthBegin().rollback(ts)
    Timestamp('2022-12-01 00:00:00')
    """
    _prefix = "MS"
    _day_opt = "start"


cdef class BusinessMonthEnd(MonthOffset):
    """
    DateOffset increments between the last business day of the month.

    BusinessMonthEnd goes to the next date which is the last business day of the month.

    Parameters
    ----------
    n : int, default 1
        The number of months represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 11, 29)
    >>> ts + pd.offsets.BMonthEnd()
    Timestamp('2022-11-30 00:00:00')

    >>> ts = pd.Timestamp(2022, 11, 30)
    >>> ts + pd.offsets.BMonthEnd()
    Timestamp('2022-12-30 00:00:00')

    If you want to get the end of the current business month:

    >>> ts = pd.Timestamp(2022, 11, 30)
    >>> pd.offsets.BMonthEnd().rollforward(ts)
    Timestamp('2022-11-30 00:00:00')
    """
    _prefix = "BME"
    _day_opt = "business_end"


cdef class BusinessMonthBegin(MonthOffset):
    """
    DateOffset of one month at the first business day.

    BusinessMonthBegin goes to the next date which is the first business day
    of the month.

    Parameters
    ----------
    n : int, default 1
        The number of months represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 11, 30)
    >>> ts + pd.offsets.BMonthBegin()
    Timestamp('2022-12-01 00:00:00')

    >>> ts = pd.Timestamp(2022, 12, 1)
    >>> ts + pd.offsets.BMonthBegin()
    Timestamp('2023-01-02 00:00:00')

    If you want to get the start of the current business month:

    >>> ts = pd.Timestamp(2022, 12, 1)
    >>> pd.offsets.BMonthBegin().rollback(ts)
    Timestamp('2022-12-01 00:00:00')
    """
    _prefix = "BMS"
    _day_opt = "business_start"


# ---------------------------------------------------------------------
# Semi-Month Based Offsets

cdef class SemiMonthOffset(SingleConstructorOffset):
    _default_day_of_month = 15
    _min_day_of_month = 2
    _attributes = tuple(["n", "normalize", "day_of_month"])

    cdef readonly:
        int day_of_month

    def __init__(self, n=1, normalize=False, day_of_month=None):
        BaseOffset.__init__(self, n, normalize)

        if day_of_month is None:
            day_of_month = self._default_day_of_month

        self.day_of_month = int(day_of_month)
        if not self._min_day_of_month <= self.day_of_month <= 27:
            raise ValueError(
                "day_of_month must be "
                f"{self._min_day_of_month}<=day_of_month<=27, "
                f"got {self.day_of_month}"
            )

    cpdef __setstate__(self, state):
        self.n = state.pop("n")
        self.normalize = state.pop("normalize")
        self.day_of_month = state.pop("day_of_month")

    @classmethod
    def _from_name(cls, suffix=None):
        return cls(day_of_month=suffix)

    @property
    def rule_code(self) -> str:
        suffix = f"-{self.day_of_month}"
        return self._prefix + suffix

    @apply_wraps
    def _apply(self, other: datetime) -> datetime:
        is_start = isinstance(self, SemiMonthBegin)

        # shift `other` to self.day_of_month, incrementing `n` if necessary
        n = roll_convention(other.day, self.n, self.day_of_month)

        days_in_month = get_days_in_month(other.year, other.month)
        # For SemiMonthBegin on other.day == 1 and
        # SemiMonthEnd on other.day == days_in_month,
        # shifting `other` to `self.day_of_month` _always_ requires
        # incrementing/decrementing `n`, regardless of whether it is
        # initially positive.
        if is_start and (self.n <= 0 and other.day == 1):
            n -= 1
        elif (not is_start) and (self.n > 0 and other.day == days_in_month):
            n += 1

        if is_start:
            months = n // 2 + n % 2
            to_day = 1 if n % 2 else self.day_of_month
        else:
            months = n // 2
            to_day = 31 if n % 2 else self.day_of_month

        return shift_month(other, months, to_day)

    @cython.wraparound(False)
    @cython.boundscheck(False)
    def _apply_array(self, dtarr: np.ndarray) -> np.ndarray:
        cdef:
            ndarray i8other = dtarr.view("i8")
            Py_ssize_t i, count = dtarr.size
            int64_t val, res_val
            ndarray out = cnp.PyArray_EMPTY(
                i8other.ndim, i8other.shape, cnp.NPY_INT64, 0
            )
            npy_datetimestruct dts
            int months, to_day, nadj, n = self.n
            int days_in_month, day, anchor_dom = self.day_of_month
            bint is_start = isinstance(self, SemiMonthBegin)
            NPY_DATETIMEUNIT reso = get_unit_from_dtype(dtarr.dtype)
            cnp.broadcast mi = cnp.PyArray_MultiIterNew2(out, i8other)

        with nogil:
            for i in range(count):
                # Analogous to: val = i8other[i]
                val = (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 1))[0]

                if val == NPY_NAT:
                    res_val = NPY_NAT

                else:
                    pandas_datetime_to_datetimestruct(val, reso, &dts)
                    day = dts.day

                    # Adjust so that we are always looking at self.day_of_month,
                    #  incrementing/decrementing n if necessary.
                    nadj = roll_convention(day, n, anchor_dom)

                    days_in_month = get_days_in_month(dts.year, dts.month)
                    # For SemiMonthBegin on other.day == 1 and
                    #  SemiMonthEnd on other.day == days_in_month,
                    #  shifting `other` to `self.day_of_month` _always_ requires
                    #  incrementing/decrementing `n`, regardless of whether it is
                    #  initially positive.
                    if is_start and (n <= 0 and day == 1):
                        nadj -= 1
                    elif (not is_start) and (n > 0 and day == days_in_month):
                        nadj += 1

                    if is_start:
                        # See also: SemiMonthBegin._apply
                        months = nadj // 2 + nadj % 2
                        to_day = 1 if nadj % 2 else anchor_dom

                    else:
                        # See also: SemiMonthEnd._apply
                        months = nadj // 2
                        to_day = 31 if nadj % 2 else anchor_dom

                    dts.year = year_add_months(dts, months)
                    dts.month = month_add_months(dts, months)
                    days_in_month = get_days_in_month(dts.year, dts.month)
                    dts.day = min(to_day, days_in_month)

                    res_val = npy_datetimestruct_to_datetime(reso, &dts)

                # Analogous to: out[i] = res_val
                (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 0))[0] = res_val

                cnp.PyArray_MultiIter_NEXT(mi)

        return out


cdef class SemiMonthEnd(SemiMonthOffset):
    """
    Two DateOffset's per month repeating on the last day of the month & day_of_month.

    Parameters
    ----------
    n : int, default 1
        The number of months represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    day_of_month : int, {1, 3,...,27}, default 15
        A specific integer for the day of the month.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 1, 14)
    >>> ts + pd.offsets.SemiMonthEnd()
    Timestamp('2022-01-15 00:00:00')

    >>> ts = pd.Timestamp(2022, 1, 15)
    >>> ts + pd.offsets.SemiMonthEnd()
    Timestamp('2022-01-31 00:00:00')

    >>> ts = pd.Timestamp(2022, 1, 31)
    >>> ts + pd.offsets.SemiMonthEnd()
    Timestamp('2022-02-15 00:00:00')

    If you want to get the result for the current month:

    >>> ts = pd.Timestamp(2022, 1, 15)
    >>> pd.offsets.SemiMonthEnd().rollforward(ts)
    Timestamp('2022-01-15 00:00:00')
    """
    _prefix = "SME"
    _min_day_of_month = 1

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        days_in_month = get_days_in_month(dt.year, dt.month)
        return dt.day in (self.day_of_month, days_in_month)


cdef class SemiMonthBegin(SemiMonthOffset):
    """
    Two DateOffset's per month repeating on the first day of the month & day_of_month.

    Parameters
    ----------
    n : int, default 1
        The number of months represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    day_of_month : int, {1, 3,...,27}, default 15
        A specific integer for the day of the month.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.SemiMonthBegin()
    Timestamp('2022-01-15 00:00:00')
    """

    _prefix = "SMS"

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        return dt.day in (1, self.day_of_month)


# ---------------------------------------------------------------------
# Week-Based Offset Classes


cdef class Week(SingleConstructorOffset):
    """
    Weekly offset.

    Parameters
    ----------
    n : int, default 1
        The number of weeks represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    weekday : int or None, default None
        Always generate specific day of week.
        0 for Monday and 6 for Sunday.

    See Also
    --------
    pd.tseries.offsets.WeekOfMonth :
     Describes monthly dates like, the Tuesday of the
     2nd week of each month.

    Examples
    --------

    >>> date_object = pd.Timestamp("2023-01-13")
    >>> date_object
    Timestamp('2023-01-13 00:00:00')

    >>> date_plus_one_week = date_object + pd.tseries.offsets.Week(n=1)
    >>> date_plus_one_week
    Timestamp('2023-01-20 00:00:00')

    >>> date_next_monday = date_object + pd.tseries.offsets.Week(weekday=0)
    >>> date_next_monday
    Timestamp('2023-01-16 00:00:00')

    >>> date_next_sunday = date_object + pd.tseries.offsets.Week(weekday=6)
    >>> date_next_sunday
    Timestamp('2023-01-15 00:00:00')
    """

    _inc = timedelta(weeks=1)
    _prefix = "W"
    _attributes = tuple(["n", "normalize", "weekday"])

    cdef readonly:
        object weekday  # int or None
        int _period_dtype_code

    def __init__(self, n=1, normalize=False, weekday=None):
        BaseOffset.__init__(self, n, normalize)
        self.weekday = weekday

        if self.weekday is not None:
            if self.weekday < 0 or self.weekday > 6:
                raise ValueError(f"Day must be 0<=day<=6, got {self.weekday}")

            self._period_dtype_code = PeriodDtypeCode.W_SUN + (weekday + 1) % 7

    cpdef __setstate__(self, state):
        self.n = state.pop("n")
        self.normalize = state.pop("normalize")
        self.weekday = state.pop("weekday")
        self._cache = state.pop("_cache", {})

    @apply_wraps
    def _apply(self, other):
        if self.weekday is None:
            return other + self.n * self._inc

        if not PyDateTime_Check(other):
            raise TypeError(
                f"Cannot add {type(other).__name__} to {type(self).__name__}"
            )

        k = self.n
        otherDay = other.weekday()
        if otherDay != self.weekday:
            other = other + timedelta((self.weekday - otherDay) % 7)
            if k > 0:
                k -= 1

        return other + timedelta(weeks=k)

    def _apply_array(self, dtarr: np.ndarray) -> np.ndarray:
        if self.weekday is None:
            td = timedelta(days=7 * self.n)
            unit = np.datetime_data(dtarr.dtype)[0]
            td64 = np.timedelta64(td, unit)
            return dtarr + td64
        else:
            reso = get_unit_from_dtype(dtarr.dtype)
            i8other = dtarr.view("i8")
            return self._end_apply_index(i8other, reso=reso)

    @cython.wraparound(False)
    @cython.boundscheck(False)
    cdef ndarray _end_apply_index(self, ndarray i8other, NPY_DATETIMEUNIT reso):
        """
        Add self to the given DatetimeIndex, specialized for case where
        self.weekday is non-null.

        Parameters
        ----------
        i8other : const int64_t[:]
        reso : NPY_DATETIMEUNIT

        Returns
        -------
        ndarray[int64_t]
        """
        cdef:
            Py_ssize_t i, count = i8other.size
            int64_t val, res_val
            ndarray out = cnp.PyArray_EMPTY(
                i8other.ndim, i8other.shape, cnp.NPY_INT64, 0
            )
            npy_datetimestruct dts
            int wday, days, weeks, n = self.n
            int anchor_weekday = self.weekday
            int64_t DAY_PERIODS = periods_per_day(reso)
            cnp.broadcast mi = cnp.PyArray_MultiIterNew2(out, i8other)

        with nogil:
            for i in range(count):
                # Analogous to: val = i8other[i]
                val = (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 1))[0]

                if val == NPY_NAT:
                    res_val = NPY_NAT
                else:
                    pandas_datetime_to_datetimestruct(val, reso, &dts)
                    wday = dayofweek(dts.year, dts.month, dts.day)

                    days = 0
                    weeks = n
                    if wday != anchor_weekday:
                        days = (anchor_weekday - wday) % 7
                        if weeks > 0:
                            weeks -= 1

                    res_val = val + (7 * weeks + days) * DAY_PERIODS

                # Analogous to: out[i] = res_val
                (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 0))[0] = res_val

                cnp.PyArray_MultiIter_NEXT(mi)

        return out

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        elif self.weekday is None:
            return True
        return dt.weekday() == self.weekday

    @property
    def rule_code(self) -> str:
        suffix = ""
        if self.weekday is not None:
            weekday = int_to_weekday[self.weekday]
            suffix = f"-{weekday}"
        return self._prefix + suffix

    @classmethod
    def _from_name(cls, suffix=None):
        if not suffix:
            weekday = None
        else:
            weekday = weekday_to_int[suffix]
        return cls(weekday=weekday)


cdef class WeekOfMonth(WeekOfMonthMixin):
    """
    Describes monthly dates like "the Tuesday of the 2nd week of each month".

    Parameters
    ----------
    n : int, default 1
        The number of months represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    week : int {0, 1, 2, 3, ...}, default 0
        A specific integer for the week of the month.
        e.g. 0 is 1st week of month, 1 is the 2nd week, etc.
    weekday : int {0, 1, ..., 6}, default 0
        A specific integer for the day of the week.

        - 0 is Monday
        - 1 is Tuesday
        - 2 is Wednesday
        - 3 is Thursday
        - 4 is Friday
        - 5 is Saturday
        - 6 is Sunday.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.WeekOfMonth()
    Timestamp('2022-01-03 00:00:00')
    """

    _prefix = "WOM"
    _attributes = tuple(["n", "normalize", "week", "weekday"])

    def __init__(self, n=1, normalize=False, week=0, weekday=0):
        WeekOfMonthMixin.__init__(self, n, normalize, weekday)
        self.week = week

        if self.week < 0 or self.week > 3:
            raise ValueError(f"Week must be 0<=week<=3, got {self.week}")

    cpdef __setstate__(self, state):
        self.n = state.pop("n")
        self.normalize = state.pop("normalize")
        self.weekday = state.pop("weekday")
        self.week = state.pop("week")

    def _get_offset_day(self, other: datetime) -> int:
        """
        Find the day in the same month as other that has the same
        weekday as self.weekday and is the self.week'th such day in the month.

        Parameters
        ----------
        other : datetime

        Returns
        -------
        day : int
        """
        mstart = datetime(other.year, other.month, 1)
        wday = mstart.weekday()
        shift_days = (self.weekday - wday) % 7
        return 1 + shift_days + self.week * 7

    @classmethod
    def _from_name(cls, suffix=None):
        if not suffix:
            raise ValueError(f"Prefix {repr(cls._prefix)} requires a suffix.")
        # only one digit weeks (1 --> week 0, 2 --> week 1, etc.)
        week = int(suffix[0]) - 1
        weekday = weekday_to_int[suffix[1:]]
        return cls(week=week, weekday=weekday)


cdef class LastWeekOfMonth(WeekOfMonthMixin):
    """
    Describes monthly dates in last week of month.

    For example "the last Tuesday of each month".

    Parameters
    ----------
    n : int, default 1
        The number of months represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    weekday : int {0, 1, ..., 6}, default 0
        A specific integer for the day of the week.

        - 0 is Monday
        - 1 is Tuesday
        - 2 is Wednesday
        - 3 is Thursday
        - 4 is Friday
        - 5 is Saturday
        - 6 is Sunday.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.LastWeekOfMonth()
    Timestamp('2022-01-31 00:00:00')
    """

    _prefix = "LWOM"
    _attributes = tuple(["n", "normalize", "weekday"])

    def __init__(self, n=1, normalize=False, weekday=0):
        WeekOfMonthMixin.__init__(self, n, normalize, weekday)
        self.week = -1

        if self.n == 0:
            raise ValueError("N cannot be 0")

    cpdef __setstate__(self, state):
        self.n = state.pop("n")
        self.normalize = state.pop("normalize")
        self.weekday = state.pop("weekday")
        self.week = -1

    def _get_offset_day(self, other: datetime) -> int:
        """
        Find the day in the same month as other that has the same
        weekday as self.weekday and is the last such day in the month.

        Parameters
        ----------
        other: datetime

        Returns
        -------
        day: int
        """
        dim = get_days_in_month(other.year, other.month)
        mend = datetime(other.year, other.month, dim)
        wday = mend.weekday()
        shift_days = (wday - self.weekday) % 7
        return dim - shift_days

    @classmethod
    def _from_name(cls, suffix=None):
        if not suffix:
            raise ValueError(f"Prefix {repr(cls._prefix)} requires a suffix.")
        weekday = weekday_to_int[suffix]
        return cls(weekday=weekday)


# ---------------------------------------------------------------------
# Special Offset Classes

cdef class FY5253Mixin(SingleConstructorOffset):
    cdef readonly:
        int startingMonth
        int weekday
        str variation

    def __init__(
        self, n=1, normalize=False, weekday=0, startingMonth=1, variation="nearest"
    ):
        BaseOffset.__init__(self, n, normalize)
        self.startingMonth = startingMonth
        self.weekday = weekday
        self.variation = variation

        if self.n == 0:
            raise ValueError("N cannot be 0")

        if self.variation not in ["nearest", "last"]:
            raise ValueError(f"{self.variation} is not a valid variation")

    cpdef __setstate__(self, state):
        self.n = state.pop("n")
        self.normalize = state.pop("normalize")
        self.weekday = state.pop("weekday")
        self.variation = state.pop("variation")

    # --------------------------------------------------------------------
    # Name-related methods

    @property
    def rule_code(self) -> str:
        prefix = self._prefix
        suffix = self.get_rule_code_suffix()
        return f"{prefix}-{suffix}"

    def _get_suffix_prefix(self) -> str:
        if self.variation == "nearest":
            return "N"
        else:
            return "L"

    def get_rule_code_suffix(self) -> str:
        prefix = self._get_suffix_prefix()
        month = MONTH_ALIASES[self.startingMonth]
        weekday = int_to_weekday[self.weekday]
        return f"{prefix}-{month}-{weekday}"


cdef class FY5253(FY5253Mixin):
    """
    Describes 52-53 week fiscal year. This is also known as a 4-4-5 calendar.

    It is used by companies that desire that their
    fiscal year always end on the same day of the week.

    It is a method of managing accounting periods.
    It is a common calendar structure for some industries,
    such as retail, manufacturing and parking industry.

    For more information see:
    https://en.wikipedia.org/wiki/4-4-5_calendar

    The year may either:

    - end on the last X day of the Y month.
    - end on the last X day closest to the last day of the Y month.

    X is a specific day of the week.
    Y is a certain month of the year

    Parameters
    ----------
    n : int
        The number of fiscal years represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    weekday : int {0, 1, ..., 6}, default 0
        A specific integer for the day of the week.

        - 0 is Monday
        - 1 is Tuesday
        - 2 is Wednesday
        - 3 is Thursday
        - 4 is Friday
        - 5 is Saturday
        - 6 is Sunday.

    startingMonth : int {1, 2, ... 12}, default 1
        The month in which the fiscal year ends.

    variation : str, default "nearest"
        Method of employing 4-4-5 calendar.

        There are two options:

        - "nearest" means year end is **weekday** closest to last day of month in year.
        - "last" means year end is final **weekday** of the final month in fiscal year.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    In the example below the default parameters give the next 52-53 week fiscal year.

    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.FY5253()
    Timestamp('2022-01-31 00:00:00')

    By the parameter ``startingMonth`` we can specify
    the month in which fiscal years end.

    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.FY5253(startingMonth=3)
    Timestamp('2022-03-28 00:00:00')

    52-53 week fiscal year can be specified by
    ``weekday`` and ``variation`` parameters.

    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.FY5253(weekday=5, startingMonth=12, variation="last")
    Timestamp('2022-12-31 00:00:00')
    """

    _prefix = "RE"
    _attributes = tuple(["n", "normalize", "weekday", "startingMonth", "variation"])

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        dt = datetime(dt.year, dt.month, dt.day)
        year_end = self.get_year_end(dt)

        if self.variation == "nearest":
            # We have to check the year end of "this" cal year AND the previous
            return year_end == dt or self.get_year_end(shift_month(dt, -1, None)) == dt
        else:
            return year_end == dt

    @apply_wraps
    def _apply(self, other: datetime) -> datetime:
        norm = Timestamp(other).normalize()

        n = self.n
        prev_year = self.get_year_end(datetime(other.year - 1, self.startingMonth, 1))
        cur_year = self.get_year_end(datetime(other.year, self.startingMonth, 1))
        next_year = self.get_year_end(datetime(other.year + 1, self.startingMonth, 1))

        prev_year = localize_pydatetime(prev_year, other.tzinfo)
        cur_year = localize_pydatetime(cur_year, other.tzinfo)
        next_year = localize_pydatetime(next_year, other.tzinfo)

        # Note: next_year.year == other.year + 1, so we will always
        # have other < next_year
        if norm == prev_year:
            n -= 1
        elif norm == cur_year:
            pass
        elif n > 0:
            if norm < prev_year:
                n -= 2
            elif prev_year < norm < cur_year:
                n -= 1
            elif cur_year < norm < next_year:
                pass
        else:
            if cur_year < norm < next_year:
                n += 1
            elif prev_year < norm < cur_year:
                pass
            elif (
                norm.year == prev_year.year
                and norm < prev_year
                and prev_year - norm <= timedelta(6)
            ):
                # GH#14774, error when next_year.year == cur_year.year
                # e.g. prev_year == datetime(2004, 1, 3),
                # other == datetime(2004, 1, 1)
                n -= 1
            else:
                assert False

        shifted = datetime(other.year + n, self.startingMonth, 1)
        result = self.get_year_end(shifted)
        result = datetime(
            result.year,
            result.month,
            result.day,
            other.hour,
            other.minute,
            other.second,
            other.microsecond,
        )
        return result

    def get_year_end(self, dt: datetime) -> datetime:
        assert dt.tzinfo is None

        dim = get_days_in_month(dt.year, self.startingMonth)
        target_date = datetime(dt.year, self.startingMonth, dim)
        wkday_diff = self.weekday - target_date.weekday()
        if wkday_diff == 0:
            # year_end is the same for "last" and "nearest" cases
            return target_date

        if self.variation == "last":
            days_forward = (wkday_diff % 7) - 7

            # days_forward is always negative, so we always end up
            # in the same year as dt
            return target_date + timedelta(days=days_forward)
        else:
            # variation == "nearest":
            days_forward = wkday_diff % 7
            if days_forward <= 3:
                # The upcoming self.weekday is closer than the previous one
                return target_date + timedelta(days_forward)
            else:
                # The previous self.weekday is closer than the upcoming one
                return target_date + timedelta(days_forward - 7)

    @classmethod
    def _parse_suffix(cls, varion_code, startingMonth_code, weekday_code):
        if varion_code == "N":
            variation = "nearest"
        elif varion_code == "L":
            variation = "last"
        else:
            raise ValueError(f"Unable to parse varion_code: {varion_code}")

        startingMonth = MONTH_TO_CAL_NUM[startingMonth_code]
        weekday = weekday_to_int[weekday_code]

        return {
            "weekday": weekday,
            "startingMonth": startingMonth,
            "variation": variation,
        }

    @classmethod
    def _from_name(cls, *args):
        return cls(**cls._parse_suffix(*args))


cdef class FY5253Quarter(FY5253Mixin):
    """
    DateOffset increments between business quarter dates for 52-53 week fiscal year.

    Also known as a 4-4-5 calendar.

    It is used by companies that desire that their
    fiscal year always end on the same day of the week.

    It is a method of managing accounting periods.
    It is a common calendar structure for some industries,
    such as retail, manufacturing and parking industry.

    For more information see:
    https://en.wikipedia.org/wiki/4-4-5_calendar

    The year may either:

    - end on the last X day of the Y month.
    - end on the last X day closest to the last day of the Y month.

    X is a specific day of the week.
    Y is a certain month of the year

    startingMonth = 1 corresponds to dates like 1/31/2007, 4/30/2007, ...
    startingMonth = 2 corresponds to dates like 2/28/2007, 5/31/2007, ...
    startingMonth = 3 corresponds to dates like 3/30/2007, 6/29/2007, ...

    Parameters
    ----------
    n : int
        The number of business quarters represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    weekday : int {0, 1, ..., 6}, default 0
        A specific integer for the day of the week.

        - 0 is Monday
        - 1 is Tuesday
        - 2 is Wednesday
        - 3 is Thursday
        - 4 is Friday
        - 5 is Saturday
        - 6 is Sunday.

    startingMonth : int {1, 2, ..., 12}, default 1
        The month in which fiscal years end.

    qtr_with_extra_week : int {1, 2, 3, 4}, default 1
        The quarter number that has the leap or 14 week when needed.

    variation : str, default "nearest"
        Method of employing 4-4-5 calendar.

        There are two options:

        - "nearest" means year end is **weekday** closest to last day of month in year.
        - "last" means year end is final **weekday** of the final month in fiscal year.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    In the example below the default parameters give
    the next business quarter for 52-53 week fiscal year.

    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.FY5253Quarter()
    Timestamp('2022-01-31 00:00:00')

    By the parameter ``startingMonth`` we can specify
    the month in which fiscal years end.

    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.FY5253Quarter(startingMonth=3)
    Timestamp('2022-03-28 00:00:00')

    Business quarters for 52-53 week fiscal year can be specified by
    ``weekday`` and ``variation`` parameters.

    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.FY5253Quarter(weekday=5, startingMonth=12, variation="last")
    Timestamp('2022-04-02 00:00:00')
    """

    _prefix = "REQ"
    _attributes = tuple(
        [
            "n",
            "normalize",
            "weekday",
            "startingMonth",
            "qtr_with_extra_week",
            "variation",
        ]
    )

    cdef readonly:
        int qtr_with_extra_week

    def __init__(
        self,
        n=1,
        normalize=False,
        weekday=0,
        startingMonth=1,
        qtr_with_extra_week=1,
        variation="nearest",
    ):
        FY5253Mixin.__init__(
            self, n, normalize, weekday, startingMonth, variation
        )
        self.qtr_with_extra_week = qtr_with_extra_week

    cpdef __setstate__(self, state):
        FY5253Mixin.__setstate__(self, state)
        self.qtr_with_extra_week = state.pop("qtr_with_extra_week")

    @cache_readonly
    def _offset(self):
        return FY5253(
            startingMonth=self.startingMonth,
            weekday=self.weekday,
            variation=self.variation,
        )

    def _rollback_to_year(self, other: datetime):
        """
        Roll `other` back to the most recent date that was on a fiscal year
        end.

        Return the date of that year-end, the number of full quarters
        elapsed between that year-end and other, and the remaining Timedelta
        since the most recent quarter-end.

        Parameters
        ----------
        other : datetime or Timestamp

        Returns
        -------
        tuple of
        prev_year_end : Timestamp giving most recent fiscal year end
        num_qtrs : int
        tdelta : Timedelta
        """
        num_qtrs = 0

        norm = Timestamp(other).tz_localize(None)
        start = self._offset.rollback(norm)
        # Note: start <= norm and self._offset.is_on_offset(start)

        if start < norm:
            # roll adjustment
            qtr_lens = self.get_weeks(norm)

            # check that qtr_lens is consistent with self._offset addition
            end = _shift_day(start, days=7 * sum(qtr_lens))
            assert self._offset.is_on_offset(end), (start, end, qtr_lens)

            tdelta = norm - start
            for qlen in qtr_lens:
                if qlen * 7 <= tdelta.days:
                    num_qtrs += 1
                    tdelta -= (
                        <_Timedelta>Timedelta(days=qlen * 7)
                    )._as_creso(norm._creso)
                else:
                    break
        else:
            tdelta = Timedelta(0)

        # Note: we always have tdelta._value>= 0
        return start, num_qtrs, tdelta

    @apply_wraps
    def _apply(self, other: datetime) -> datetime:
        # Note: self.n == 0 is not allowed.

        n = self.n

        prev_year_end, num_qtrs, tdelta = self._rollback_to_year(other)
        res = prev_year_end
        n += num_qtrs
        if self.n <= 0 and tdelta._value > 0:
            n += 1

        # Possible speedup by handling years first.
        years = n // 4
        if years:
            res += self._offset * years
            n -= years * 4

        # Add an extra day to make *sure* we are getting the quarter lengths
        # for the upcoming year, not the previous year
        qtr_lens = self.get_weeks(res + Timedelta(days=1))

        # Note: we always have 0 <= n < 4
        weeks = sum(qtr_lens[:n])
        if weeks:
            res = _shift_day(res, days=weeks * 7)

        return res

    def get_weeks(self, dt: datetime):
        ret = [13] * 4

        year_has_extra_week = self.year_has_extra_week(dt)

        if year_has_extra_week:
            ret[self.qtr_with_extra_week - 1] = 14

        return ret

    def year_has_extra_week(self, dt: datetime) -> bool:
        # Avoid round-down errors --> normalize to get
        # e.g. '370D' instead of '360D23H'
        norm = Timestamp(dt).normalize().tz_localize(None)

        next_year_end = self._offset.rollforward(norm)
        prev_year_end = norm - self._offset
        weeks_in_year = (next_year_end - prev_year_end).days / 7
        assert weeks_in_year in [52, 53], weeks_in_year
        return weeks_in_year == 53

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        if self._offset.is_on_offset(dt):
            return True

        next_year_end = dt - self._offset

        qtr_lens = self.get_weeks(dt)

        current = next_year_end
        for qtr_len in qtr_lens:
            current = _shift_day(current, days=qtr_len * 7)
            if dt == current:
                return True
        return False

    @property
    def rule_code(self) -> str:
        suffix = FY5253Mixin.rule_code.__get__(self)
        qtr = self.qtr_with_extra_week
        return f"{suffix}-{qtr}"

    @classmethod
    def _from_name(cls, *args):
        return cls(
            **dict(FY5253._parse_suffix(*args[:-1]), qtr_with_extra_week=int(args[-1]))
        )


cdef class Easter(SingleConstructorOffset):
    """
    DateOffset for the Easter holiday using logic defined in dateutil.

    Right now uses the revised method which is valid in years 1583-4099.

    Parameters
    ----------
    n : int, default 1
        The number of years represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    >>> ts = pd.Timestamp(2022, 1, 1)
    >>> ts + pd.offsets.Easter()
    Timestamp('2022-04-17 00:00:00')
    """

    cpdef __setstate__(self, state):
        self.n = state.pop("n")
        self.normalize = state.pop("normalize")

    @apply_wraps
    def _apply(self, other: datetime) -> datetime:
        from dateutil.easter import easter

        current_easter = easter(other.year)
        current_easter = datetime(
            current_easter.year, current_easter.month, current_easter.day
        )
        current_easter = localize_pydatetime(current_easter, other.tzinfo)

        n = self.n
        if n >= 0 and other < current_easter:
            n -= 1
        elif n < 0 and other > current_easter:
            n += 1
        # TODO: Why does this handle the 0 case the opposite of others?

        # NOTE: easter returns a datetime.date so we have to convert to type of
        # other
        new = easter(other.year + n)
        new = datetime(
            new.year,
            new.month,
            new.day,
            other.hour,
            other.minute,
            other.second,
            other.microsecond,
        )
        return new

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False

        from dateutil.easter import easter

        return date(dt.year, dt.month, dt.day) == easter(dt.year)


# ----------------------------------------------------------------------
# Custom Offset classes


cdef class CustomBusinessDay(BusinessDay):
    """
    DateOffset subclass representing possibly n custom business days.

    In CustomBusinessDay we can use custom weekmask, holidays, and calendar.

    Parameters
    ----------
    n : int, default 1
        The number of days represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    weekmask : str, Default 'Mon Tue Wed Thu Fri'
        Weekmask of valid business days, passed to ``numpy.busdaycalendar``.
    holidays : list
        List/array of dates to exclude from the set of valid business days,
        passed to ``numpy.busdaycalendar``.
    calendar : np.busdaycalendar
        Calendar to integrate.
    offset : timedelta, default timedelta(0)
        Time offset to apply.

    Examples
    --------
    In the example below the default parameters give the next business day.

    >>> ts = pd.Timestamp(2022, 8, 5, 16)
    >>> ts + pd.offsets.CustomBusinessDay()
    Timestamp('2022-08-08 16:00:00')

    Business days can be specified by ``weekmask`` parameter. To convert
    the returned datetime object to its string representation
    the function strftime() is used in the next example.

    >>> import datetime as dt
    >>> freq = pd.offsets.CustomBusinessDay(weekmask="Mon Wed Fri")
    >>> pd.date_range(dt.datetime(2022, 12, 10), dt.datetime(2022, 12, 21),
    ...               freq=freq).strftime('%a %d %b %Y %H:%M')
    Index(['Mon 12 Dec 2022 00:00', 'Wed 14 Dec 2022 00:00',
           'Fri 16 Dec 2022 00:00', 'Mon 19 Dec 2022 00:00',
           'Wed 21 Dec 2022 00:00'],
           dtype='object')

    Using NumPy business day calendar you can define custom holidays.

    >>> import datetime as dt
    >>> bdc = np.busdaycalendar(holidays=['2022-12-12', '2022-12-14'])
    >>> freq = pd.offsets.CustomBusinessDay(calendar=bdc)
    >>> pd.date_range(dt.datetime(2022, 12, 10), dt.datetime(2022, 12, 25), freq=freq)
    DatetimeIndex(['2022-12-13', '2022-12-15', '2022-12-16', '2022-12-19',
                   '2022-12-20', '2022-12-21', '2022-12-22', '2022-12-23'],
                   dtype='datetime64[ns]', freq='C')

    If you want to shift the result on n day you can use the parameter ``offset``.

    >>> pd.Timestamp(2022, 8, 5, 16) + pd.offsets.CustomBusinessDay(1)
    Timestamp('2022-08-08 16:00:00')

    >>> import datetime as dt
    >>> ts = pd.Timestamp(2022, 8, 5, 16)
    >>> ts + pd.offsets.CustomBusinessDay(1, offset=dt.timedelta(days=1))
    Timestamp('2022-08-09 16:00:00')
    """

    _prefix = "C"
    _attributes = tuple(
        ["n", "normalize", "weekmask", "holidays", "calendar", "offset"]
    )

    @property
    def _period_dtype_code(self):
        # GH#52534
        raise ValueError(f"{self.base} is not supported as period frequency")

    _apply_array = BaseOffset._apply_array

    def __init__(
        self,
        n=1,
        normalize=False,
        weekmask="Mon Tue Wed Thu Fri",
        holidays=None,
        calendar=None,
        offset=timedelta(0),
    ):
        BusinessDay.__init__(self, n, normalize, offset)
        self._init_custom(weekmask, holidays, calendar)

    cpdef __setstate__(self, state):
        self.holidays = state.pop("holidays")
        self.weekmask = state.pop("weekmask")
        BusinessDay.__setstate__(self, state)

    @apply_wraps
    def _apply(self, other):
        if self.n <= 0:
            roll = "forward"
        else:
            roll = "backward"

        if PyDateTime_Check(other):
            date_in = other
            np_dt = np.datetime64(date_in.date())

            np_incr_dt = np.busday_offset(
                np_dt, self.n, roll=roll, busdaycal=self.calendar
            )

            dt_date = np_incr_dt.astype(datetime)
            result = datetime.combine(dt_date, date_in.time())

            if self.offset:
                result = result + self.offset
            return result

        elif is_any_td_scalar(other):
            td = Timedelta(self.offset) + other
            return BDay(self.n, offset=td.to_pytimedelta(), normalize=self.normalize)
        else:
            raise ApplyTypeError(
                "Only know how to combine trading day with "
                "datetime, datetime64 or timedelta."
            )

    def is_on_offset(self, dt: datetime) -> bool:
        if self.normalize and not _is_normalized(dt):
            return False
        day64 = _to_dt64D(dt)
        return np.is_busday(day64, busdaycal=self.calendar)


cdef class CustomBusinessHour(BusinessHour):
    """
    DateOffset subclass representing possibly n custom business days.

    In CustomBusinessHour we can use custom weekmask, holidays, and calendar.

    Parameters
    ----------
    n : int, default 1
        The number of hours represented.
    normalize : bool, default False
        Normalize start/end dates to midnight before generating date range.
    weekmask : str, Default 'Mon Tue Wed Thu Fri'
        Weekmask of valid business days, passed to ``numpy.busdaycalendar``.
    holidays : list
        List/array of dates to exclude from the set of valid business days,
        passed to ``numpy.busdaycalendar``.
    calendar : np.busdaycalendar
        Calendar to integrate.
    start : str, time, or list of str/time, default "09:00"
        Start time of your custom business hour in 24h format.
    end : str, time, or list of str/time, default: "17:00"
        End time of your custom business hour in 24h format.
    offset : timedelta, default timedelta(0)
        Time offset to apply.

    Examples
    --------
    In the example below the default parameters give the next business hour.

    >>> ts = pd.Timestamp(2022, 8, 5, 16)
    >>> ts + pd.offsets.CustomBusinessHour()
    Timestamp('2022-08-08 09:00:00')

    We can also change the start and the end of business hours.

    >>> ts = pd.Timestamp(2022, 8, 5, 16)
    >>> ts + pd.offsets.CustomBusinessHour(start="11:00")
    Timestamp('2022-08-08 11:00:00')

    >>> from datetime import time as dt_time
    >>> ts = pd.Timestamp(2022, 8, 5, 16)
    >>> ts + pd.offsets.CustomBusinessHour(end=dt_time(19, 0))
    Timestamp('2022-08-05 17:00:00')

    >>> ts = pd.Timestamp(2022, 8, 5, 22)
    >>> ts + pd.offsets.CustomBusinessHour(end=dt_time(19, 0))
    Timestamp('2022-08-08 10:00:00')

    You can divide your business day hours into several parts.

    >>> import datetime as dt
    >>> freq = pd.offsets.CustomBusinessHour(start=["06:00", "10:00", "15:00"],
    ...                                      end=["08:00", "12:00", "17:00"])
    >>> pd.date_range(dt.datetime(2022, 12, 9), dt.datetime(2022, 12, 13), freq=freq)
    DatetimeIndex(['2022-12-09 06:00:00', '2022-12-09 07:00:00',
                   '2022-12-09 10:00:00', '2022-12-09 11:00:00',
                   '2022-12-09 15:00:00', '2022-12-09 16:00:00',
                   '2022-12-12 06:00:00', '2022-12-12 07:00:00',
                   '2022-12-12 10:00:00', '2022-12-12 11:00:00',
                   '2022-12-12 15:00:00', '2022-12-12 16:00:00'],
                   dtype='datetime64[ns]', freq='cbh')

    Business days can be specified by ``weekmask`` parameter. To convert
    the returned datetime object to its string representation
    the function strftime() is used in the next example.

    >>> import datetime as dt
    >>> freq = pd.offsets.CustomBusinessHour(weekmask="Mon Wed Fri",
    ...                                      start="10:00", end="13:00")
    >>> pd.date_range(dt.datetime(2022, 12, 10), dt.datetime(2022, 12, 18),
    ...               freq=freq).strftime('%a %d %b %Y %H:%M')
    Index(['Mon 12 Dec 2022 10:00', 'Mon 12 Dec 2022 11:00',
           'Mon 12 Dec 2022 12:00', 'Wed 14 Dec 2022 10:00',
           'Wed 14 Dec 2022 11:00', 'Wed 14 Dec 2022 12:00',
           'Fri 16 Dec 2022 10:00', 'Fri 16 Dec 2022 11:00',
           'Fri 16 Dec 2022 12:00'],
           dtype='object')

    Using NumPy business day calendar you can define custom holidays.

    >>> import datetime as dt
    >>> bdc = np.busdaycalendar(holidays=['2022-12-12', '2022-12-14'])
    >>> freq = pd.offsets.CustomBusinessHour(calendar=bdc, start="10:00", end="13:00")
    >>> pd.date_range(dt.datetime(2022, 12, 10), dt.datetime(2022, 12, 18), freq=freq)
    DatetimeIndex(['2022-12-13 10:00:00', '2022-12-13 11:00:00',
                   '2022-12-13 12:00:00', '2022-12-15 10:00:00',
                   '2022-12-15 11:00:00', '2022-12-15 12:00:00',
                   '2022-12-16 10:00:00', '2022-12-16 11:00:00',
                   '2022-12-16 12:00:00'],
                   dtype='datetime64[ns]', freq='cbh')
    """

    _prefix = "cbh"
    _anchor = 0
    _attributes = tuple(
        ["n", "normalize", "weekmask", "holidays", "calendar", "start", "end", "offset"]
    )

    def __init__(
        self,
        n=1,
        normalize=False,
        weekmask="Mon Tue Wed Thu Fri",
        holidays=None,
        calendar=None,
        start="09:00",
        end="17:00",
        offset=timedelta(0),
    ):
        BusinessHour.__init__(self, n, normalize, start=start, end=end, offset=offset)
        self._init_custom(weekmask, holidays, calendar)


cdef class _CustomBusinessMonth(BusinessMixin):
    _attributes = tuple(
        ["n", "normalize", "weekmask", "holidays", "calendar", "offset"]
    )

    def __init__(
        self,
        n=1,
        normalize=False,
        weekmask="Mon Tue Wed Thu Fri",
        holidays=None,
        calendar=None,
        offset=timedelta(0),
    ):
        BusinessMixin.__init__(self, n, normalize, offset)
        self._init_custom(weekmask, holidays, calendar)

    @cache_readonly
    def cbday_roll(self):
        """
        Define default roll function to be called in apply method.
        """
        cbday_kwds = self.kwds.copy()
        cbday_kwds["offset"] = timedelta(0)

        cbday = CustomBusinessDay(n=1, normalize=False, **cbday_kwds)

        if self._prefix.endswith("S"):
            # MonthBegin
            roll_func = cbday.rollforward
        else:
            # MonthEnd
            roll_func = cbday.rollback
        return roll_func

    @cache_readonly
    def m_offset(self):
        if self._prefix.endswith("S"):
            # MonthBegin
            moff = MonthBegin(n=1, normalize=False)
        else:
            # MonthEnd
            moff = MonthEnd(n=1, normalize=False)
        return moff

    @cache_readonly
    def month_roll(self):
        """
        Define default roll function to be called in apply method.
        """
        if self._prefix.endswith("S"):
            # MonthBegin
            roll_func = self.m_offset.rollback
        else:
            # MonthEnd
            roll_func = self.m_offset.rollforward
        return roll_func

    @apply_wraps
    def _apply(self, other: datetime) -> datetime:
        # First move to month offset
        cur_month_offset_date = self.month_roll(other)

        # Find this custom month offset
        compare_date = self.cbday_roll(cur_month_offset_date)
        n = roll_convention(other.day, self.n, compare_date.day)

        new = cur_month_offset_date + n * self.m_offset
        result = self.cbday_roll(new)

        if self.offset:
            result = result + self.offset
        return result


cdef class CustomBusinessMonthEnd(_CustomBusinessMonth):
    """
    DateOffset subclass representing custom business month(s).

    Increments between end of month dates.

    Parameters
    ----------
    n : int, default 1
        The number of months represented.
    normalize : bool, default False
        Normalize end dates to midnight before generating date range.
    weekmask : str, Default 'Mon Tue Wed Thu Fri'
        Weekmask of valid business days, passed to ``numpy.busdaycalendar``.
    holidays : list
        List/array of dates to exclude from the set of valid business days,
        passed to ``numpy.busdaycalendar``.
    calendar : np.busdaycalendar
        Calendar to integrate.
    offset : timedelta, default timedelta(0)
        Time offset to apply.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    In the example below we use the default parameters.

    >>> ts = pd.Timestamp(2022, 8, 5)
    >>> ts + pd.offsets.CustomBusinessMonthEnd()
    Timestamp('2022-08-31 00:00:00')

    Custom business month end can be specified by ``weekmask`` parameter.
    To convert the returned datetime object to its string representation
    the function strftime() is used in the next example.

    >>> import datetime as dt
    >>> freq = pd.offsets.CustomBusinessMonthEnd(weekmask="Wed Thu")
    >>> pd.date_range(dt.datetime(2022, 7, 10), dt.datetime(2022, 12, 18),
    ...               freq=freq).strftime('%a %d %b %Y %H:%M')
    Index(['Thu 28 Jul 2022 00:00', 'Wed 31 Aug 2022 00:00',
           'Thu 29 Sep 2022 00:00', 'Thu 27 Oct 2022 00:00',
           'Wed 30 Nov 2022 00:00'],
           dtype='object')

    Using NumPy business day calendar you can define custom holidays.

    >>> import datetime as dt
    >>> bdc = np.busdaycalendar(holidays=['2022-08-01', '2022-09-30',
    ...                                   '2022-10-31', '2022-11-01'])
    >>> freq = pd.offsets.CustomBusinessMonthEnd(calendar=bdc)
    >>> pd.date_range(dt.datetime(2022, 7, 10), dt.datetime(2022, 11, 10), freq=freq)
    DatetimeIndex(['2022-07-29', '2022-08-31', '2022-09-29', '2022-10-28'],
                   dtype='datetime64[ns]', freq='CBME')
    """

    _prefix = "CBME"


cdef class CustomBusinessMonthBegin(_CustomBusinessMonth):
    """
    DateOffset subclass representing custom business month(s).

    Increments between beginning of month dates.

    Parameters
    ----------
    n : int, default 1
        The number of months represented.
    normalize : bool, default False
        Normalize start dates to midnight before generating date range.
    weekmask : str, Default 'Mon Tue Wed Thu Fri'
        Weekmask of valid business days, passed to ``numpy.busdaycalendar``.
    holidays : list
        List/array of dates to exclude from the set of valid business days,
        passed to ``numpy.busdaycalendar``.
    calendar : np.busdaycalendar
        Calendar to integrate.
    offset : timedelta, default timedelta(0)
        Time offset to apply.

    See Also
    --------
    :class:`~pandas.tseries.offsets.DateOffset` : Standard kind of date increment.

    Examples
    --------
    In the example below we use the default parameters.

    >>> ts = pd.Timestamp(2022, 8, 5)
    >>> ts + pd.offsets.CustomBusinessMonthBegin()
    Timestamp('2022-09-01 00:00:00')

    Custom business month start can be specified by ``weekmask`` parameter.
    To convert the returned datetime object to its string representation
    the function strftime() is used in the next example.

    >>> import datetime as dt
    >>> freq = pd.offsets.CustomBusinessMonthBegin(weekmask="Wed Thu")
    >>> pd.date_range(dt.datetime(2022, 7, 10), dt.datetime(2022, 12, 18),
    ...               freq=freq).strftime('%a %d %b %Y %H:%M')
    Index(['Wed 03 Aug 2022 00:00', 'Thu 01 Sep 2022 00:00',
           'Wed 05 Oct 2022 00:00', 'Wed 02 Nov 2022 00:00',
           'Thu 01 Dec 2022 00:00'],
           dtype='object')

    Using NumPy business day calendar you can define custom holidays.

    >>> import datetime as dt
    >>> bdc = np.busdaycalendar(holidays=['2022-08-01', '2022-09-30',
    ...                                   '2022-10-31', '2022-11-01'])
    >>> freq = pd.offsets.CustomBusinessMonthBegin(calendar=bdc)
    >>> pd.date_range(dt.datetime(2022, 7, 10), dt.datetime(2022, 11, 10), freq=freq)
    DatetimeIndex(['2022-08-02', '2022-09-01', '2022-10-03', '2022-11-02'],
                   dtype='datetime64[ns]', freq='CBMS')
    """

    _prefix = "CBMS"


BDay = BusinessDay
BMonthEnd = BusinessMonthEnd
BMonthBegin = BusinessMonthBegin
CBMonthEnd = CustomBusinessMonthEnd
CBMonthBegin = CustomBusinessMonthBegin
CDay = CustomBusinessDay

# ----------------------------------------------------------------------
# to_offset helpers

prefix_mapping = {
    offset._prefix: offset
    for offset in [
        YearBegin,  # 'YS'
        YearEnd,  # 'YE'
        BYearBegin,  # 'BYS'
        BYearEnd,  # 'BYE'
        BusinessDay,  # 'B'
        BusinessMonthBegin,  # 'BMS'
        BusinessMonthEnd,  # 'BME'
        BQuarterEnd,  # 'BQE'
        BQuarterBegin,  # 'BQS'
        BusinessHour,  # 'bh'
        CustomBusinessDay,  # 'C'
        CustomBusinessMonthEnd,  # 'CBME'
        CustomBusinessMonthBegin,  # 'CBMS'
        CustomBusinessHour,  # 'cbh'
        MonthEnd,  # 'ME'
        MonthBegin,  # 'MS'
        Nano,  # 'ns'
        SemiMonthEnd,  # 'SME'
        SemiMonthBegin,  # 'SMS'
        Week,  # 'W'
        Second,  # 's'
        Minute,  # 'min'
        Micro,  # 'us'
        QuarterEnd,  # 'QE'
        QuarterBegin,  # 'QS'
        Milli,  # 'ms'
        Hour,  # 'h'
        Day,  # 'D'
        WeekOfMonth,  # 'WOM'
        FY5253,
        FY5253Quarter,
    ]
}

# hack to handle WOM-1MON
opattern = re.compile(
    r"([+\-]?\d*|[+\-]?\d*\.\d*)\s*([A-Za-z]+([\-][\dA-Za-z\-]+)?)"
)

_lite_rule_alias = {
    "W": "W-SUN",
    "QE": "QE-DEC",

    "YE": "YE-DEC",      # YearEnd(month=12),
    "YS": "YS-JAN",    # YearBegin(month=1),
    "BYE": "BYE-DEC",    # BYearEnd(month=12),
    "BYS": "BYS-JAN",  # BYearBegin(month=1),

    "Min": "min",
    "min": "min",
    "ms": "ms",
    "us": "us",
    "ns": "ns",
}

_dont_uppercase = {"h", "bh", "cbh", "MS", "ms", "s"}


INVALID_FREQ_ERR_MSG = "Invalid frequency: {0}"

# TODO: still needed?
# cache of previously seen offsets
_offset_map = {}


# TODO: better name?
def _get_offset(name: str) -> BaseOffset:
    """
    Return DateOffset object associated with rule name.

    Examples
    --------
    _get_offset('EOM') --> BMonthEnd(1)
    """
    if (
        name not in _lite_rule_alias
        and (name.upper() in _lite_rule_alias)
        and name != "ms"
    ):
        warnings.warn(
            f"\'{name}\' is deprecated and will be removed "
            f"in a future version, please use \'{name.upper()}\' instead.",
            FutureWarning,
            stacklevel=find_stack_level(),
        )
    elif (
        name not in _lite_rule_alias
        and (name.lower() in _lite_rule_alias)
        and name != "MS"
    ):
        warnings.warn(
            f"\'{name}\' is deprecated and will be removed "
            f"in a future version, please use \'{name.lower()}\' instead.",
            FutureWarning,
            stacklevel=find_stack_level(),
        )
    if name not in _dont_uppercase:
        name = name.upper()
        name = _lite_rule_alias.get(name, name)
        name = _lite_rule_alias.get(name.lower(), name)
    else:
        name = _lite_rule_alias.get(name, name)

    if name not in _offset_map:
        try:
            split = name.split("-")
            klass = prefix_mapping[split[0]]
            # handles case where there's no suffix (and will TypeError if too
            # many '-')
            offset = klass._from_name(*split[1:])
        except (ValueError, TypeError, KeyError) as err:
            # bad prefix or suffix
            raise ValueError(INVALID_FREQ_ERR_MSG.format(
                f"{name}, failed to parse with error message: {repr(err)}")
            )
        # cache
        _offset_map[name] = offset

    return _offset_map[name]


cpdef to_offset(freq, bint is_period=False):
    """
    Return DateOffset object from string or datetime.timedelta object.

    Parameters
    ----------
    freq : str, datetime.timedelta, BaseOffset or None
        The frequency represented.
    is_period : bool, default False
        Convert string denoting period frequency to corresponding offsets
        frequency if is_period=True.

    Returns
    -------
    BaseOffset subclass or None

    Raises
    ------
    ValueError
        If freq is an invalid frequency

    See Also
    --------
    BaseOffset : Standard kind of date increment used for a date range.

    Examples
    --------
    >>> from pandas.tseries.frequencies import to_offset
    >>> to_offset("5min")
    <5 * Minutes>

    >>> to_offset("1D1h")
    <25 * Hours>

    >>> to_offset("2W")
    <2 * Weeks: weekday=6>

    >>> to_offset("2B")
    <2 * BusinessDays>

    >>> to_offset(pd.Timedelta(days=1))
    <Day>

    >>> to_offset(pd.offsets.Hour())
    <Hour>

    Passing the parameter ``is_period`` equal to True, you can use a string
    denoting period frequency:

    >>> freq = to_offset(freq="ME", is_period=False)
    >>> freq.rule_code
    'ME'

    >>> freq = to_offset(freq="M", is_period=True)
    >>> freq.rule_code
    'ME'
    """
    if freq is None:
        return None

    if isinstance(freq, tuple):
        raise TypeError(
            f"to_offset does not support tuples {freq}, pass as a string instead"
        )

    if isinstance(freq, BaseOffset):
        result = freq

    elif PyDelta_Check(freq):
        result = delta_to_tick(freq)

    elif isinstance(freq, str):
        result = None
        stride_sign = None

        try:
            split = opattern.split(freq)
            if split[-1] != "" and not split[-1].isspace():
                # the last element must be blank
                raise ValueError("last element must be blank")

            tups = zip(split[0::4], split[1::4], split[2::4])
            for n, (sep, stride, name) in enumerate(tups):
                if not is_period and name.upper() in c_OFFSET_REMOVED_FREQSTR:
                    raise ValueError(
                        f"\'{name}\' is no longer supported for offsets."
                    )
                if (not is_period and
                        name != name.upper() and
                        name.lower() not in {"s", "ms", "us", "ns"} and
                        name.upper().split("-")[0].endswith(("S", "E"))):
                    raise ValueError(
                        f"\'{name}\' is no longer supported for offsets."
                    )
                if is_period and name.upper() in c_REVERSE_OFFSET_REMOVED_FREQSTR:
                    if name.upper().startswith("Y"):
                        raise ValueError(
                            f"for Period, please use \'Y{name.upper()[2:]}\' "
                            f"instead of \'{name}\'"
                        )
                    if (name.upper().startswith("B") or
                            name.upper().startswith("S") or
                            name.upper().startswith("C")):
                        raise ValueError(INVALID_FREQ_ERR_MSG.format(name))
                    else:
                        raise ValueError(
                            f"for Period, please use "
                            f"\'{c_REVERSE_OFFSET_REMOVED_FREQSTR.get(name.upper())}\' "
                            f"instead of \'{name}\'"
                        )
                elif is_period and name.upper() in c_OFFSET_REMOVED_FREQSTR:
                    if name.upper() != name:
                        warnings.warn(
                            f"\'{name}\' is deprecated and will be removed in "
                            f"a future version, please use \'{name.upper()}\' "
                            f"instead.",
                            FutureWarning,
                            stacklevel=find_stack_level(),
                        )
                    name = c_OFFSET_REMOVED_FREQSTR.get(name.upper())

                if sep != "" and not sep.isspace():
                    raise ValueError("separator must be spaces")
                prefix = _lite_rule_alias.get(name) or name
                if stride_sign is None:
                    stride_sign = -1 if stride.startswith("-") else 1
                if not stride:
                    stride = 1

                if prefix in c_DEPR_ABBREVS:
                    warnings.warn(
                        f"\'{prefix}\' is deprecated and will be removed "
                        f"in a future version, please use "
                        f"\'{c_DEPR_ABBREVS.get(prefix)}\' instead.",
                        FutureWarning,
                        stacklevel=find_stack_level(),
                    )
                    prefix = c_DEPR_ABBREVS[prefix]

                if prefix in {"D", "h", "min", "s", "ms", "us", "ns"}:
                    # For these prefixes, we have something like "3h" or
                    #  "2.5min", so we can construct a Timedelta with the
                    #  matching unit and get our offset from delta_to_tick
                    td = Timedelta(1, unit=prefix)
                    off = delta_to_tick(td)
                    offset = off * float(stride)
                    if n != 0:
                        # If n==0, then stride_sign is already incorporated
                        #  into the offset
                        offset *= stride_sign
                else:
                    stride = int(stride)
                    offset = _get_offset(prefix)
                    offset = offset * int(np.fabs(stride) * stride_sign)

                if result is None:
                    result = offset
                else:
                    result = result + offset
        except (ValueError, TypeError) as err:
            raise ValueError(INVALID_FREQ_ERR_MSG.format(
                f"{freq}, failed to parse with error message: {repr(err)}")
            )
    else:
        result = None

    if result is None:
        raise ValueError(INVALID_FREQ_ERR_MSG.format(freq))

    if is_period and not hasattr(result, "_period_dtype_code"):
        if isinstance(freq, str):
            raise ValueError(f"{result.name} is not supported as period frequency")
        else:
            raise ValueError(f"{freq} is not supported as period frequency")

    return result


# ----------------------------------------------------------------------
# RelativeDelta Arithmetic

cdef datetime _shift_day(datetime other, int days):
    """
    Increment the datetime `other` by the given number of days, retaining
    the time-portion of the datetime.  For tz-naive datetimes this is
    equivalent to adding a timedelta.  For tz-aware datetimes it is similar to
    dateutil's relativedelta.__add__, but handles pytz tzinfo objects.

    Parameters
    ----------
    other : datetime or Timestamp
    days : int

    Returns
    -------
    shifted: datetime or Timestamp
    """
    if other.tzinfo is None:
        return other + timedelta(days=days)

    tz = other.tzinfo
    naive = other.replace(tzinfo=None)
    shifted = naive + timedelta(days=days)
    return localize_pydatetime(shifted, tz)


cdef int year_add_months(npy_datetimestruct dts, int months) noexcept nogil:
    """
    New year number after shifting npy_datetimestruct number of months.
    """
    return dts.year + (dts.month + months - 1) // 12


cdef int month_add_months(npy_datetimestruct dts, int months) noexcept nogil:
    """
    New month number after shifting npy_datetimestruct
    number of months.
    """
    cdef:
        int new_month = (dts.month + months) % 12
    return 12 if new_month == 0 else new_month


@cython.wraparound(False)
@cython.boundscheck(False)
cdef ndarray shift_quarters(
    ndarray dtindex,
    int quarters,
    int q1start_month,
    str day_opt,
    int modby=3,
    NPY_DATETIMEUNIT reso=NPY_DATETIMEUNIT.NPY_FR_ns,
):
    """
    Given an int64 array representing nanosecond timestamps, shift all elements
    by the specified number of quarters using DateOffset semantics.

    Parameters
    ----------
    dtindex : int64_t[:] timestamps for input dates
    quarters : int number of quarters to shift
    q1start_month : int month in which Q1 begins by convention
    day_opt : {'start', 'end', 'business_start', 'business_end'}
    modby : int (3 for quarters, 12 for years)
    reso : NPY_DATETIMEUNIT, default NPY_FR_ns

    Returns
    -------
    out : ndarray[int64_t]
    """
    if day_opt not in ["start", "end", "business_start", "business_end"]:
        raise ValueError("day must be None, 'start', 'end', "
                         "'business_start', or 'business_end'")

    cdef:
        Py_ssize_t count = dtindex.size
        ndarray out = cnp.PyArray_EMPTY(dtindex.ndim, dtindex.shape, cnp.NPY_INT64, 0)
        Py_ssize_t i
        int64_t val, res_val
        int months_since, n
        npy_datetimestruct dts
        cnp.broadcast mi = cnp.PyArray_MultiIterNew2(out, dtindex)

    with nogil:
        for i in range(count):
            # Analogous to: val = dtindex[i]
            val = (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 1))[0]

            if val == NPY_NAT:
                res_val = NPY_NAT
            else:
                pandas_datetime_to_datetimestruct(val, reso, &dts)
                n = quarters

                months_since = (dts.month - q1start_month) % modby
                n = _roll_qtrday(&dts, n, months_since, day_opt)

                dts.year = year_add_months(dts, modby * n - months_since)
                dts.month = month_add_months(dts, modby * n - months_since)
                dts.day = get_day_of_month(&dts, day_opt)

                res_val = npy_datetimestruct_to_datetime(reso, &dts)

            # Analogous to: out[i] = res_val
            (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 0))[0] = res_val

            cnp.PyArray_MultiIter_NEXT(mi)

    return out


@cython.wraparound(False)
@cython.boundscheck(False)
def shift_months(
    ndarray dtindex,  # int64_t, arbitrary ndim
    int months,
    str day_opt=None,
    NPY_DATETIMEUNIT reso=NPY_DATETIMEUNIT.NPY_FR_ns,
):
    """
    Given an int64-based datetime index, shift all elements
    specified number of months using DateOffset semantics

    day_opt: {None, 'start', 'end', 'business_start', 'business_end'}
       * None: day of month
       * 'start' 1st day of month
       * 'end' last day of month
    """
    cdef:
        Py_ssize_t i
        npy_datetimestruct dts
        int count = dtindex.size
        ndarray out = cnp.PyArray_EMPTY(dtindex.ndim, dtindex.shape, cnp.NPY_INT64, 0)
        int months_to_roll
        int64_t val, res_val

        cnp.broadcast mi = cnp.PyArray_MultiIterNew2(out, dtindex)

    if day_opt is not None and day_opt not in {
            "start", "end", "business_start", "business_end"
    }:
        raise ValueError("day must be None, 'start', 'end', "
                         "'business_start', or 'business_end'")

    if day_opt is None:
        # TODO: can we combine this with the non-None case?
        with nogil:
            for i in range(count):
                # Analogous to: val = i8other[i]
                val = (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 1))[0]

                if val == NPY_NAT:
                    res_val = NPY_NAT
                else:
                    pandas_datetime_to_datetimestruct(val, reso, &dts)
                    dts.year = year_add_months(dts, months)
                    dts.month = month_add_months(dts, months)

                    dts.day = min(dts.day, get_days_in_month(dts.year, dts.month))
                    res_val = npy_datetimestruct_to_datetime(reso, &dts)

                # Analogous to: out[i] = res_val
                (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 0))[0] = res_val

                cnp.PyArray_MultiIter_NEXT(mi)

    else:
        with nogil:
            for i in range(count):

                # Analogous to: val = i8other[i]
                val = (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 1))[0]

                if val == NPY_NAT:
                    res_val = NPY_NAT
                else:
                    pandas_datetime_to_datetimestruct(val, reso, &dts)
                    months_to_roll = months

                    months_to_roll = _roll_qtrday(&dts, months_to_roll, 0, day_opt)

                    dts.year = year_add_months(dts, months_to_roll)
                    dts.month = month_add_months(dts, months_to_roll)
                    dts.day = get_day_of_month(&dts, day_opt)

                    res_val = npy_datetimestruct_to_datetime(reso, &dts)

                # Analogous to: out[i] = res_val
                (<int64_t*>cnp.PyArray_MultiIter_DATA(mi, 0))[0] = res_val

                cnp.PyArray_MultiIter_NEXT(mi)

    return out


def shift_month(stamp: datetime, months: int, day_opt: object = None) -> datetime:
    """
    Given a datetime (or Timestamp) `stamp`, an integer `months` and an
    option `day_opt`, return a new datetimelike that many months later,
    with day determined by `day_opt` using relativedelta semantics.

    Scalar analogue of shift_months.

    Parameters
    ----------
    stamp : datetime or Timestamp
    months : int
    day_opt : None, 'start', 'end', 'business_start', 'business_end', or int
        None: returned datetimelike has the same day as the input, or the
              last day of the month if the new month is too short
        'start': returned datetimelike has day=1
        'end': returned datetimelike has day on the last day of the month
        'business_start': returned datetimelike has day on the first
            business day of the month
        'business_end': returned datetimelike has day on the last
            business day of the month
        int: returned datetimelike has day equal to day_opt

    Returns
    -------
    shifted : datetime or Timestamp (same as input `stamp`)
    """
    cdef:
        int year, month, day
        int days_in_month, dy

    dy = (stamp.month + months) // 12
    month = (stamp.month + months) % 12

    if month == 0:
        month = 12
        dy -= 1
    year = stamp.year + dy

    if day_opt is None:
        days_in_month = get_days_in_month(year, month)
        day = min(stamp.day, days_in_month)
    elif day_opt == "start":
        day = 1
    elif day_opt == "end":
        day = get_days_in_month(year, month)
    elif day_opt == "business_start":
        # first business day of month
        day = get_firstbday(year, month)
    elif day_opt == "business_end":
        # last business day of month
        day = get_lastbday(year, month)
    elif is_integer_object(day_opt):
        days_in_month = get_days_in_month(year, month)
        day = min(day_opt, days_in_month)
    else:
        raise ValueError(day_opt)
    return stamp.replace(year=year, month=month, day=day)


cdef int get_day_of_month(npy_datetimestruct* dts, str day_opt) noexcept nogil:
    """
    Find the day in `other`'s month that satisfies a DateOffset's is_on_offset
    policy, as described by the `day_opt` argument.

    Parameters
    ----------
    dts : npy_datetimestruct*
    day_opt : {'start', 'end', 'business_start', 'business_end'}
        'start': returns 1
        'end': returns last day of the month
        'business_start': returns the first business day of the month
        'business_end': returns the last business day of the month

    Returns
    -------
    day_of_month : int

    Examples
    -------
    >>> other = datetime(2017, 11, 14)
    >>> get_day_of_month(other, 'start')
    1
    >>> get_day_of_month(other, 'end')
    30

    Notes
    -----
    Caller is responsible for ensuring one of the four accepted day_opt values
    is passed.
    """

    if day_opt == "start":
        return 1
    elif day_opt == "end":
        return get_days_in_month(dts.year, dts.month)
    elif day_opt == "business_start":
        # first business day of month
        return get_firstbday(dts.year, dts.month)
    else:
        # i.e. day_opt == "business_end":
        # last business day of month
        return get_lastbday(dts.year, dts.month)


cpdef int roll_convention(int other, int n, int compare) noexcept nogil:
    """
    Possibly increment or decrement the number of periods to shift
    based on rollforward/rollbackward conventions.

    Parameters
    ----------
    other : int, generally the day component of a datetime
    n : number of periods to increment, before adjusting for rolling
    compare : int, generally the day component of a datetime, in the same
              month as the datetime form which `other` was taken.

    Returns
    -------
    n : int number of periods to increment
    """
    if n > 0 and other < compare:
        n -= 1
    elif n <= 0 and other > compare:
        # as if rolled forward already
        n += 1
    return n


def roll_qtrday(other: datetime, n: int, month: int,
                day_opt: str, modby: int) -> int:
    """
    Possibly increment or decrement the number of periods to shift
    based on rollforward/rollbackward conventions.

    Parameters
    ----------
    other : datetime or Timestamp
    n : number of periods to increment, before adjusting for rolling
    month : int reference month giving the first month of the year
    day_opt : {'start', 'end', 'business_start', 'business_end'}
        The convention to use in finding the day in a given month against
        which to compare for rollforward/rollbackward decisions.
    modby : int 3 for quarters, 12 for years

    Returns
    -------
    n : int number of periods to increment

    See Also
    --------
    get_day_of_month : Find the day in a month provided an offset.
    """
    cdef:
        int months_since
        npy_datetimestruct dts

    if day_opt not in ["start", "end", "business_start", "business_end"]:
        raise ValueError(day_opt)

    pydate_to_dtstruct(other, &dts)

    if modby == 12:
        # We care about the month-of-year, not month-of-quarter, so skip mod
        months_since = other.month - month
    else:
        months_since = other.month % modby - month % modby

    return _roll_qtrday(&dts, n, months_since, day_opt)


cdef int _roll_qtrday(npy_datetimestruct* dts,
                      int n,
                      int months_since,
                      str day_opt) except? -1 nogil:
    """
    See roll_qtrday.__doc__
    """

    if n > 0:
        if months_since < 0 or (months_since == 0 and
                                dts.day < get_day_of_month(dts, day_opt)):
            # pretend to roll back if on same month but
            # before compare_day
            n -= 1
    else:
        if months_since > 0 or (months_since == 0 and
                                dts.day > get_day_of_month(dts, day_opt)):
            # make sure to roll forward, so negate
            n += 1
    return n