REF: return ndarray from to_perioddelta; use it less

pandas/_libs/tslibs/offsets.pyx

@@ -36,7 +36,7 @@ from pandas._libs.tslibs.base cimport ABCTimestamp
 from pandas._libs.tslibs.ccalendar import (
     MONTH_ALIASES, MONTH_TO_CAL_NUM, weekday_to_int, int_to_weekday,
 )
-from pandas._libs.tslibs.ccalendar cimport get_days_in_month, dayofweek
+from pandas._libs.tslibs.ccalendar cimport get_days_in_month, dayofweek, DAY_NANOS
 from pandas._libs.tslibs.conversion cimport (
     convert_datetime_to_tsobject,
     localize_pydatetime,
@@ -1046,6 +1046,8 @@ cdef class RelativeDeltaOffset(BaseOffset):
         -------
         DatetimeIndex
         """
+        from pandas._libs.tslibs.timedeltas import Timedelta
+
         kwds = self.kwds
         relativedelta_fast = {
             "years",
@@ -1067,19 +1069,14 @@ cdef class RelativeDeltaOffset(BaseOffset):
 
             weeks = kwds.get("weeks", 0) * self.n
             if weeks:
-                # integer addition on PeriodIndex is deprecated,
-                #   so we directly use _time_shift instead
-                asper = index.to_period("W")
-                shifted = asper._time_shift(weeks)
-                index = shifted.to_timestamp() + index.to_perioddelta("W")
+                index = index + Timedelta(days=7*weeks)
 
             timedelta_kwds = {
                 k: v
                 for k, v in kwds.items()
                 if k in ["days", "hours", "minutes", "seconds", "microseconds"]
             }
             if timedelta_kwds:
-                from .timedeltas import Timedelta
                 delta = Timedelta(**timedelta_kwds)
                 index = index + (self.n * delta)
             return index
@@ -1383,13 +1380,15 @@ cdef class BusinessDay(BusinessMixin):
 
     @apply_index_wraps
     def apply_index(self, dtindex):
-        time = dtindex.to_perioddelta("D")
+        i8other = dtindex.asi8
+        time = (i8other % DAY_NANOS).view("timedelta64[ns]")
+
         # to_period rolls forward to next BDay; track and
         # reduce n where it does when rolling forward
         asper = dtindex.to_period("B")
 
         if self.n > 0:
-            shifted = (dtindex.to_perioddelta("B") - time).asi8 != 0
+            shifted = (dtindex.to_perioddelta("B") - time).view("i8") != 0
 
             # Integer-array addition is deprecated, so we use
             # _time_shift directly
@@ -2275,8 +2274,9 @@ cdef class SemiMonthOffset(SingleConstructorOffset):
         # determine how many days away from the 1st of the month we are
         from pandas import Timedelta
 
+        i8other = dtindex.asi8
         dti = dtindex
-        days_from_start = dtindex.to_perioddelta("M").asi8
+        days_from_start = dtindex.to_perioddelta("M").view("i8")
         delta = Timedelta(days=self.day_of_month - 1).value
 
         # get boolean array for each element before the day_of_month
@@ -2289,7 +2289,7 @@ cdef class SemiMonthOffset(SingleConstructorOffset):
         roll = self._get_roll(dtindex, before_day_of_month, after_day_of_month)
 
         # isolate the time since it will be striped away one the next line
-        time = dtindex.to_perioddelta("D")
+        time = (i8other % DAY_NANOS).view("timedelta64[ns]")
 
         # apply the correct number of months
 
@@ -2504,12 +2504,10 @@ cdef class Week(SingleConstructorOffset):
     @apply_index_wraps
     def apply_index(self, dtindex):
         if self.weekday is None:
-            # integer addition on PeriodIndex is deprecated,
-            #  so we use _time_shift directly
-            asper = dtindex.to_period("W")
+            from pandas._libs.tslibs.timedeltas import Timedelta
 
-            shifted = asper._time_shift(self.n)
-            return shifted.to_timestamp() + dtindex.to_perioddelta("W")
+            result = np.asarray(dtindex) + Timedelta(days=7 * self.n)
+            return type(dtindex)._simple_new(result)
         else:
             return self._end_apply_index(dtindex)
 
@@ -2529,7 +2527,8 @@ cdef class Week(SingleConstructorOffset):
         from pandas import Timedelta
         from .frequencies import get_freq_code  # TODO: avoid circular import
 
-        off = dtindex.to_perioddelta("D")
+        i8other = dtindex.asi8
+        off = (i8other % DAY_NANOS).view("timedelta64[ns]")
 
         base, mult = get_freq_code(self.freqstr)
         base_period = dtindex.to_period(base)

pandas/core/arrays/datetimes.py

@@ -1122,14 +1122,13 @@ def to_perioddelta(self, freq):
 
         Returns
         -------
-        TimedeltaArray/Index
+        ndarray[timedelta64[ns]]
         """
         # TODO: consider privatizing (discussion in GH#23113)
-        from pandas.core.arrays.timedeltas import TimedeltaArray
 
         i8delta = self.asi8 - self.to_period(freq).to_timestamp().asi8
         m8delta = i8delta.view("m8[ns]")
-        return TimedeltaArray(m8delta)
+        return m8delta
 
     # -----------------------------------------------------------------
     # Properties - Vectorized Timestamp Properties/Methods

pandas/tests/arrays/test_datetimelike.py

@@ -513,7 +513,7 @@ def test_to_perioddelta(self, datetime_index, freqstr):
 
         expected = dti.to_perioddelta(freq=freqstr)
         result = arr.to_perioddelta(freq=freqstr)
-        assert isinstance(result, TimedeltaArray)
+        assert isinstance(result, np.ndarray)
 
         # placeholder until these become actual EA subclasses and we can use
         #  an EA-specific tm.assert_ function