BUG: iloc.__setitem__ with DataFrame value, multiple blocks, non-unique columns (pandas-dev#36337)

Author: jbrockmendel

pandas/core/indexing.py

@@ -1690,18 +1690,42 @@ def _setitem_with_indexer(self, indexer, value):
                     sub_indexer = list(indexer)
                     multiindex_indexer = isinstance(labels, ABCMultiIndex)
                     # TODO: we are implicitly assuming value.columns is unique
+                    unique_cols = value.columns.is_unique
+
+                    if not unique_cols and value.columns.equals(self.obj.columns):
+                        # We assume we are already aligned, see
+                        # test_iloc_setitem_frame_duplicate_columns_multiple_blocks
+                        for loc in ilocs:
+                            item = item_labels[loc]
+                            if item in value:
+                                sub_indexer[info_axis] = item
+                                v = self._align_series(
+                                    tuple(sub_indexer),
+                                    value.iloc[:, loc],
+                                    multiindex_indexer,
+                                )
+                            else:
+                                v = np.nan
 
-                    for loc in ilocs:
-                        item = item_labels[loc]
-                        if item in value:
-                            sub_indexer[info_axis] = item
-                            v = self._align_series(
-                                tuple(sub_indexer), value[item], multiindex_indexer
-                            )
-                        else:
-                            v = np.nan
+                            self._setitem_single_column(loc, v, pi)
 
-                        self._setitem_single_column(loc, v, pi)
+                    elif not unique_cols:
+                        raise ValueError(
+                            "Setting with non-unique columns is not allowed."
+                        )
+
+                    else:
+                        for loc in ilocs:
+                            item = item_labels[loc]
+                            if item in value:
+                                sub_indexer[info_axis] = item
+                                v = self._align_series(
+                                    tuple(sub_indexer), value[item], multiindex_indexer
+                                )
+                            else:
+                                v = np.nan
+
+                            self._setitem_single_column(loc, v, pi)
 
                 # we have an equal len ndarray/convertible to our labels
                 # hasattr first, to avoid coercing to ndarray without reason.

pandas/tests/indexing/test_iloc.py

@@ -369,6 +369,20 @@ def test_iloc_setitem_dups(self):
         df.iloc[[1, 0], [0, 1]] = df.iloc[[1, 0], [0, 1]].reset_index(drop=True)
         tm.assert_frame_equal(df, expected)
 
+    def test_iloc_setitem_frame_duplicate_columns_multiple_blocks(self):
+        # Same as the "assign back to self" check in test_iloc_setitem_dups
+        #  but on a DataFrame with multiple blocks
+        df = pd.DataFrame([[0, 1], [2, 3]], columns=["B", "B"])
+
+        df.iloc[:, 0] = df.iloc[:, 0].astype("f8")
+        assert len(df._mgr.blocks) == 2
+        expected = df.copy()
+
+        # assign back to self
+        df.iloc[[0, 1], [0, 1]] = df.iloc[[0, 1], [0, 1]]
+
+        tm.assert_frame_equal(df, expected)
+
     # TODO: GH#27620 this test used to compare iloc against ix; check if this
     #  is redundant with another test comparing iloc against loc
     def test_iloc_getitem_frame(self):