REF: DataFrame._values return DTA/TDA where appropriate (pandas-dev#41531)

Author: jbrockmendel

pandas/core/frame.py

@@ -857,26 +857,37 @@ def _can_fast_transpose(self) -> bool:
         # TODO(EA2D) special case would be unnecessary with 2D EAs
         return not is_1d_only_ea_dtype(dtype)
 
+    # error: Return type "Union[ndarray, DatetimeArray, TimedeltaArray]" of
+    # "_values" incompatible with return type "ndarray" in supertype "NDFrame"
     @property
-    def _values_compat(self) -> np.ndarray | DatetimeArray | TimedeltaArray:
+    def _values(  # type: ignore[override]
+        self,
+    ) -> np.ndarray | DatetimeArray | TimedeltaArray:
         """
         Analogue to ._values that may return a 2D ExtensionArray.
         """
+        self._consolidate_inplace()
+
         mgr = self._mgr
+
         if isinstance(mgr, ArrayManager):
-            return self._values
+            if len(mgr.arrays) == 1 and not is_1d_only_ea_obj(mgr.arrays[0]):
+                # error: Item "ExtensionArray" of "Union[ndarray, ExtensionArray]"
+                # has no attribute "reshape"
+                return mgr.arrays[0].reshape(-1, 1)  # type: ignore[union-attr]
+            return self.values
 
         blocks = mgr.blocks
         if len(blocks) != 1:
-            return self._values
+            return self.values
 
         arr = blocks[0].values
         if arr.ndim == 1:
             # non-2D ExtensionArray
-            return self._values
+            return self.values
 
         # more generally, whatever we allow in NDArrayBackedExtensionBlock
-        arr = cast("DatetimeArray | TimedeltaArray", arr)
+        arr = cast("np.ndarray | DatetimeArray | TimedeltaArray", arr)
         return arr.T
 
     # ----------------------------------------------------------------------
@@ -3322,7 +3333,7 @@ def transpose(self, *args, copy: bool = False) -> DataFrame:
 
         if self._can_fast_transpose:
             # Note: tests pass without this, but this improves perf quite a bit.
-            new_vals = self._values_compat.T
+            new_vals = self._values.T
             if copy:
                 new_vals = new_vals.copy()
 
@@ -10621,11 +10632,6 @@ def values(self) -> np.ndarray:
         self._consolidate_inplace()
         return self._mgr.as_array(transpose=True)
 
-    @property
-    def _values(self) -> np.ndarray:
-        """internal implementation"""
-        return self.values
-
 
 DataFrame._add_numeric_operations()
 

pandas/core/internals/managers.py

@@ -759,7 +759,8 @@ def _slice_take_blocks_ax0(
                 blk = self.blocks[blkno]
 
                 # Otherwise, slicing along items axis is necessary.
-                if not blk._can_consolidate:
+                if not blk._can_consolidate and not blk._validate_ndim:
+                    # i.e. we dont go through here for DatetimeTZBlock
                     # A non-consolidatable block, it's easy, because there's
                     # only one item and each mgr loc is a copy of that single
                     # item.

pandas/tests/frame/methods/test_values.py

@@ -223,3 +223,54 @@ def test_values_lcd(self, mixed_float_frame, mixed_int_frame):
 
         values = mixed_int_frame[["C"]].values
         assert values.dtype == np.uint8
+
+
+class TestPrivateValues:
+    def test_private_values_dt64tz(self, using_array_manager, request):
+        if using_array_manager:
+            mark = pytest.mark.xfail(reason="doesn't share memory")
+            request.node.add_marker(mark)
+
+        dta = date_range("2000", periods=4, tz="US/Central")._data.reshape(-1, 1)
+
+        df = DataFrame(dta, columns=["A"])
+        tm.assert_equal(df._values, dta)
+
+        # we have a view
+        assert np.shares_memory(df._values._ndarray, dta._ndarray)
+
+        # TimedeltaArray
+        tda = dta - dta
+        df2 = df - df
+        tm.assert_equal(df2._values, tda)
+
+    @td.skip_array_manager_invalid_test
+    def test_private_values_dt64tz_multicol(self):
+        dta = date_range("2000", periods=8, tz="US/Central")._data.reshape(-1, 2)
+
+        df = DataFrame(dta, columns=["A", "B"])
+        tm.assert_equal(df._values, dta)
+
+        # we have a view
+        assert np.shares_memory(df._values._ndarray, dta._ndarray)
+
+        # TimedeltaArray
+        tda = dta - dta
+        df2 = df - df
+        tm.assert_equal(df2._values, tda)
+
+    def test_private_values_dt64_multiblock(self, using_array_manager, request):
+        if using_array_manager:
+            mark = pytest.mark.xfail(reason="returns ndarray")
+            request.node.add_marker(mark)
+
+        dta = date_range("2000", periods=8)._data
+
+        df = DataFrame({"A": dta[:4]}, copy=False)
+        df["B"] = dta[4:]
+
+        assert len(df._mgr.arrays) == 2
+
+        result = df._values
+        expected = dta.reshape(2, 4).T
+        tm.assert_equal(result, expected)