BUG: 2D DTA/TDA arithmetic with object-dtype (#32185)

pandas/core/arrays/datetimelike.py

@@ -42,6 +42,7 @@
 from pandas.core.algorithms import checked_add_with_arr, take, unique1d, value_counts
 from pandas.core.arrays.base import ExtensionArray, ExtensionOpsMixin
 import pandas.core.common as com
+from pandas.core.construction import array, extract_array
 from pandas.core.indexers import check_array_indexer
 from pandas.core.ops.common import unpack_zerodim_and_defer
 from pandas.core.ops.invalid import invalid_comparison, make_invalid_op
@@ -623,7 +624,7 @@ def astype(self, dtype, copy=True):
         dtype = pandas_dtype(dtype)
 
         if is_object_dtype(dtype):
-            return self._box_values(self.asi8)
+            return self._box_values(self.asi8.ravel()).reshape(self.shape)
         elif is_string_dtype(dtype) and not is_categorical_dtype(dtype):
             return self._format_native_types()
         elif is_integer_dtype(dtype):
@@ -1256,19 +1257,13 @@ def _addsub_object_array(self, other: np.ndarray, op):
             PerformanceWarning,
         )
 
-        # For EA self.astype('O') returns a numpy array, not an Index
-        left = self.astype("O")
+        # Caller is responsible for broadcasting if necessary
+        assert self.shape == other.shape, (self.shape, other.shape)
 
-        res_values = op(left, np.array(other))
-        kwargs = {}
-        if not is_period_dtype(self):
-            kwargs["freq"] = "infer"
-        try:
-            res = type(self)._from_sequence(res_values, **kwargs)
-        except ValueError:
-            # e.g. we've passed a Timestamp to TimedeltaArray
-            res = res_values
-        return res
+        res_values = op(self.astype("O"), np.array(other))
+        result = array(res_values.ravel())
+        result = extract_array(result, extract_numpy=True).reshape(self.shape)
+        return result
 
     def _time_shift(self, periods, freq=None):
         """

pandas/tests/arithmetic/test_datetime64.py

@@ -27,6 +27,7 @@
     date_range,
 )
 import pandas._testing as tm
+from pandas.core.arrays import DatetimeArray, TimedeltaArray
 from pandas.core.ops import roperator
 from pandas.tests.arithmetic.common import (
     assert_invalid_addsub_type,
@@ -956,6 +957,18 @@ def test_dt64arr_sub_NaT(self, box_with_array):
     # -------------------------------------------------------------
     # Subtraction of datetime-like array-like
 
+    def test_dt64arr_sub_dt64object_array(self, box_with_array, tz_naive_fixture):
+        dti = pd.date_range("2016-01-01", periods=3, tz=tz_naive_fixture)
+        expected = dti - dti
+
+        obj = tm.box_expected(dti, box_with_array)
+        expected = tm.box_expected(expected, box_with_array)
+
+        warn = PerformanceWarning if box_with_array is not pd.DataFrame else None
+        with tm.assert_produces_warning(warn):
+            result = obj - obj.astype(object)
+        tm.assert_equal(result, expected)
+
     def test_dt64arr_naive_sub_dt64ndarray(self, box_with_array):
         dti = pd.date_range("2016-01-01", periods=3, tz=None)
         dt64vals = dti.values
@@ -2395,3 +2408,31 @@ def test_shift_months(years, months):
     raw = [x + pd.offsets.DateOffset(years=years, months=months) for x in dti]
     expected = DatetimeIndex(raw)
     tm.assert_index_equal(actual, expected)
+
+
+def test_dt64arr_addsub_object_dtype_2d():
+    # block-wise DataFrame operations will require operating on 2D
+    #  DatetimeArray/TimedeltaArray, so check that specifically.
+    dti = pd.date_range("1994-02-13", freq="2W", periods=4)
+    dta = dti._data.reshape((4, 1))
+
+    other = np.array([[pd.offsets.Day(n)] for n in range(4)])
+    assert other.shape == dta.shape
+
+    with tm.assert_produces_warning(PerformanceWarning):
+        result = dta + other
+    with tm.assert_produces_warning(PerformanceWarning):
+        expected = (dta[:, 0] + other[:, 0]).reshape(-1, 1)
+
+    assert isinstance(result, DatetimeArray)
+    assert result.freq is None
+    tm.assert_numpy_array_equal(result._data, expected._data)
+
+    with tm.assert_produces_warning(PerformanceWarning):
+        # Case where we expect to get a TimedeltaArray back
+        result2 = dta - dta.astype(object)
+
+    assert isinstance(result2, TimedeltaArray)
+    assert result2.shape == (4, 1)
+    assert result2.freq is None
+    assert (result2.asi8 == 0).all()

pandas/tests/arithmetic/test_timedelta64.py

@@ -532,6 +532,20 @@ def test_tda_add_sub_index(self):
         expected = tdi - tdi
         tm.assert_index_equal(result, expected)
 
+    def test_tda_add_dt64_object_array(self, box_df_fail, tz_naive_fixture):
+        # Result should be cast back to DatetimeArray
+        dti = pd.date_range("2016-01-01", periods=3, tz=tz_naive_fixture)
+        dti._set_freq(None)
+        tdi = dti - dti
+
+        obj = tm.box_expected(tdi, box_df_fail)
+        other = tm.box_expected(dti, box_df_fail)
+
+        warn = PerformanceWarning if box_df_fail is not pd.DataFrame else None
+        with tm.assert_produces_warning(warn):
+            result = obj + other.astype(object)
+        tm.assert_equal(result, other)
+
     # -------------------------------------------------------------
     # Binary operations TimedeltaIndex and timedelta-like