diff --git a/dask_expr/_expr.py b/dask_expr/_expr.py
index 739456a1..fdb0bd8f 100644
--- a/dask_expr/_expr.py
+++ b/dask_expr/_expr.py
@@ -3193,15 +3193,14 @@ def _fusion_pass(expr):
                     dependents[next._name] = set()
                     expr_mapping[next._name] = next
 
-            for operand in next.operands:
-                if isinstance(operand, Expr):
-                    stack.append(operand)
-                    if is_valid_blockwise_op(operand):
-                        if next._name in dependencies:
-                            dependencies[next._name].add(operand._name)
-                        dependents[operand._name].add(next._name)
-                        expr_mapping[operand._name] = operand
-                        expr_mapping[next._name] = next
+            for operand in next.dependencies():
+                stack.append(operand)
+                if is_valid_blockwise_op(operand):
+                    if next._name in dependencies:
+                        dependencies[next._name].add(operand._name)
+                    dependents[operand._name].add(next._name)
+                    expr_mapping[operand._name] = operand
+                    expr_mapping[next._name] = next
 
         # Traverse each "root" until we find a fusable sub-group.
         # Here we use root to refer to a Blockwise Expr node that
diff --git a/dask_expr/_indexing.py b/dask_expr/_indexing.py
index 2bb12a23..f1abfb9e 100644
--- a/dask_expr/_indexing.py
+++ b/dask_expr/_indexing.py
@@ -98,6 +98,8 @@ def _loc(self, iindexer, cindexer):
             elif is_series_like(iindexer) and not is_bool_dtype(iindexer.dtype):
                 return new_collection(LocList(self.obj, iindexer.values, cindexer))
             elif isinstance(iindexer, list) or is_arraylike(iindexer):
+                if len(iindexer) == 0:
+                    return new_collection(LocEmpty(self.obj._meta, cindexer))
                 return new_collection(LocList(self.obj, iindexer, cindexer))
             else:
                 # element should raise KeyError
@@ -250,6 +252,26 @@ def _layer(self) -> dict:
         return self._layer_information[0]
 
 
+class LocEmpty(LocList):
+    _parameters = ["meta", "cindexer"]
+
+    def _lower(self):
+        return None
+
+    @functools.cached_property
+    def _meta(self):
+        if self.cindexer is None:
+            return self.operand("meta")
+        else:
+            return self.operand("meta").loc[:, self.cindexer]
+
+    @functools.cached_property
+    def _layer_information(self):
+        divisions = [None, None]
+        dsk = {(self._name, 0): DataNode((self._name, 0), self._meta)}
+        return dsk, divisions
+
+
 class LocSlice(LocBase):
     @functools.cached_property
     def start(self):
diff --git a/dask_expr/tests/test_indexing.py b/dask_expr/tests/test_indexing.py
index b152fdf6..bf378945 100644
--- a/dask_expr/tests/test_indexing.py
+++ b/dask_expr/tests/test_indexing.py
@@ -35,15 +35,6 @@ def test_iloc_errors(df):
         df.iloc[(1, 2, 3)]
 
 
-def test_loc(df, pdf):
-    assert_eq(df.loc[:, "x"], pdf.loc[:, "x"])
-    assert_eq(df.loc[:, ["x"]], pdf.loc[:, ["x"]])
-    assert_eq(df.loc[:, []], pdf.loc[:, []])
-
-    assert_eq(df.loc[df.y == 20, "x"], pdf.loc[pdf.y == 20, "x"])
-    assert_eq(df.loc[df.y == 20, ["x"]], pdf.loc[pdf.y == 20, ["x"]])
-
-
 def test_loc_slice(pdf, df):
     pdf.columns = [10, 20]
     df.columns = [10, 20]
@@ -86,6 +77,12 @@ def test_columns_dtype_on_empty_slice(df, pdf, loc, update):
 
 
 def test_loc(df, pdf):
+    assert_eq(df.loc[:, "x"], pdf.loc[:, "x"])
+    assert_eq(df.loc[:, ["x"]], pdf.loc[:, ["x"]])
+    assert_eq(df.loc[:, []], pdf.loc[:, []])
+
+    assert_eq(df.loc[df.y == 20, "x"], pdf.loc[pdf.y == 20, "x"])
+    assert_eq(df.loc[df.y == 20, ["x"]], pdf.loc[pdf.y == 20, ["x"]])
     assert df.loc[3:8].divisions[0] == 3
     assert df.loc[3:8].divisions[-1] == 8