perf(ir): don't recreate nodes in replace if their children haven't…

… changed

ibis/backends/sql/rewrites.py

@@ -370,7 +370,7 @@ def sqlize(
 
     # lower the expression graph to a SQL-like relational algebra
     context = {"params": params}
-    sqlized = node.replace(
+    result = node.replace(
         replace_parameter
         | project_to_select
         | filter_to_select
@@ -385,24 +385,23 @@ def sqlize(
 
     # squash subsequent Select nodes into one
     if fuse_selects:
-        simplified = sqlized.replace(merge_select_select)
-    else:
-        simplified = sqlized
+        result = result.replace(merge_select_select)
 
     if post_rewrites:
-        simplified = simplified.replace(reduce(operator.or_, post_rewrites))
+        result = result.replace(reduce(operator.or_, post_rewrites))
 
     # extract common table expressions while wrapping them in a CTE node
-    ctes = extract_ctes(simplified)
+    ctes = extract_ctes(result)
 
-    def wrap(node, _, **kwargs):
-        new = node.__recreate__(kwargs)
-        return CTE(new) if node in ctes else new
+    if ctes:
 
-    result = simplified.replace(wrap)
-    ctes = [cte.parent for cte in result.find(CTE, ordered=True)]
+        def apply_ctes(node, kwargs):
+            new = node.__recreate__(kwargs) if kwargs else node
+            return CTE(new) if node in ctes else new
 
-    return result, ctes
+        result = result.replace(apply_ctes)
+        return result, [cte.parent for cte in result.find(CTE, ordered=True)]
+    return result, []
 
 
 # supplemental rewrites selectively used on a per-backend basis

ibis/backends/tests/test_numeric.py

@@ -1382,16 +1382,14 @@ def test_histogram(con, alltypes):
     hist = con.execute(alltypes.int_col.histogram(n).name("hist"))
     vc = hist.value_counts().sort_index()
     vc_np, _bin_edges = np.histogram(alltypes.int_col.execute(), bins=n)
-    assert vc.tolist() == vc_np.tolist()
-    assert (
-        con.execute(
-            ibis.memtable({"value": range(100)})
-            .select(bin=_.value.histogram(10))
-            .value_counts()
-            .bin_count.nunique()
-        )
-        == 1
+    expr = (
+        ibis.memtable({"value": range(100)})
+        .select(bin=_.value.histogram(10))
+        .value_counts()
+        .bin_count.nunique()
     )
+    assert vc.tolist() == vc_np.tolist()
+    assert con.execute(expr) == 1
 
 
 @pytest.mark.parametrize("const", ["pi", "e"])

ibis/common/graph.py

@@ -22,7 +22,7 @@
 Finder = Callable[["Node"], bool]
 FinderLike = Union[Finder, Pattern, _ClassInfo]
 
-Replacer = Callable[["Node", dict["Node", Any]], "Node"]
+Replacer = Callable[["Node", dict["Node", Any] | None], "Node"]
 ReplacerLike = Union[Replacer, Pattern, Mapping]
 
 
@@ -127,6 +127,47 @@ def _recursive_lookup(obj: Any, dct: dict) -> Any:
         return obj
 
 
+def _apply_replacements(obj: Any, replacements: dict) -> tuple[Any, bool]:
+    """Replace nodes in a possibly nested object.
+
+    Parameters
+    ----------
+    obj
+        The object to traverse.
+    replacements
+        A mapping of replacement values.
+
+    Returns
+    -------
+    tuple[Any, bool]
+        A tuple of the replaced object and whether any replacements were made.
+    """
+    if isinstance(obj, Node):
+        val = replacements.get(obj)
+        return (obj, False) if val is None else (val, True)
+    typ = type(obj)
+    if typ in (tuple, frozenset, list):
+        changed = False
+        items = []
+        for i in obj:
+            i, ichanged = _apply_replacements(i, replacements)
+            changed |= ichanged
+            items.append(i)
+        return typ(items), changed
+    elif isinstance(obj, dict):
+        changed = False
+        items = {}
+        for k, v in obj.items():
+            k, kchanged = _apply_replacements(k, replacements)
+            v, vchanged = _apply_replacements(v, replacements)
+            changed |= kchanged
+            changed |= vchanged
+            items[k] = v
+        return items, changed
+    else:
+        return obj, False
+
+
 def _coerce_finder(obj: FinderLike, context: Optional[dict] = None) -> Finder:
     """Coerce an object into a callable finder function.
 
@@ -165,8 +206,7 @@ def _coerce_replacer(obj: ReplacerLike, context: Optional[dict] = None) -> Repla
     Parameters
     ----------
     obj
-        A Pattern, a Mapping or a callable which can be fed to `node.map()`
-        to replace nodes.
+        A Pattern, Mapping, or Callable.
     context
         Optional context to use if the replacer is a pattern.
 
@@ -177,26 +217,26 @@ def _coerce_replacer(obj: ReplacerLike, context: Optional[dict] = None) -> Repla
     """
     if isinstance(obj, Pattern):
 
-        def fn(node, _, **kwargs):
+        def fn(node, kwargs):
             ctx = context or {}
             # need to first reconstruct the node from the possible rewritten
             # children, so we can match on the new node containing the rewritten
             # child arguments, this way we can propagate the rewritten nodes
-            # upward in the hierarchy, using a specialized __recreate__ method
-            # improves the performance by 17% compared node.__class__(**kwargs)
-            recreated = node.__recreate__(kwargs)
+            # upward in the hierarchy
+            recreated = node.__recreate__(kwargs) if kwargs else node
             if (result := obj.match(recreated, ctx)) is NoMatch:
                 return recreated
-            else:
-                return result
+            return result
 
     elif isinstance(obj, Mapping):
 
-        def fn(node, _, **kwargs):
+        def fn(node, kwargs):
+            # For a mapping we want to lookup the original node first, and
+            # return a recreated one from the children if it's not present
             try:
                 return obj[node]
             except KeyError:
-                return node.__recreate__(kwargs)
+                return node.__recreate__(kwargs) if kwargs else node
     elif callable(obj):
         fn = obj
     else:
@@ -313,7 +353,7 @@ def map_clear(self, fn: Callable, filter: Optional[Finder] = None) -> Any:
                 if not dependents[dependency]:
                     del results[dependency]
 
-        return results[self]
+        return results.get(self, self)
 
     @experimental
     def map_nodes(self, fn: Callable, filter: Optional[Finder] = None) -> Any:
@@ -451,8 +491,9 @@ def replace(
         Parameters
         ----------
         replacer
-            A `Pattern`, a `Mapping` or a callable which can be fed to
-            `node.map()` directly to replace nodes.
+            A `Pattern`, `Mapping` or Callable taking the original unrewritten
+            node, and a mapping of attribute name to value of its rewritten
+            children (or None if no children were rewritten).
         filter
             A type, tuple of types, a pattern or a callable to filter out nodes
             from the traversal. The traversal will only visit nodes that match
@@ -465,9 +506,28 @@ def replace(
         The root node of the graph with the replaced nodes.
 
         """
-        replacer = _coerce_replacer(replacer, context)
-        results = self.map(replacer, filter=filter)
-        return results.get(self, self)
+        replacements: dict[Node, Any] = {}
+
+        fn = _coerce_replacer(replacer, context)
+
+        graph, _ = Graph.from_bfs(self, filter=filter).toposort()
+        for node in graph:
+            kwargs = {}
+            # Apply already rewritten nodes to the children of the node
+            changed = False
+            for k, v in zip(node.__argnames__, node.__args__):
+                v, vchanged = _apply_replacements(v, replacements)
+                changed |= vchanged
+                kwargs[k] = v
+
+            # Call the replacer on the node with any rewritten nodes (or None
+            # if unchanged).
+            result = fn(node, kwargs if changed else None)
+            if result is not node:
+                # The node is changed, store it in the mapping of replacements
+                replacements[node] = result
+
+        return replacements.get(self, self)
 
 
 class Graph(dict[Node, Sequence[Node]]):

ibis/common/tests/test_graph.py

@@ -19,7 +19,7 @@
     traverse,
 )
 from ibis.common.grounds import Annotable, Concrete
-from ibis.common.patterns import Eq, If, InstanceOf, Object, TupleOf, _
+from ibis.common.patterns import Eq, If, InstanceOf, Object, TupleOf, _, pattern
 
 
 class MyNode(Node):
@@ -170,6 +170,36 @@ def test_replace_with_mapping():
     assert result == new_A
 
 
+@pytest.mark.parametrize("kind", ["pattern", "mapping", "function"])
+def test_replace_doesnt_recreate_unchanged_nodes(kind):
+    A1 = MyNode(name="A1", children=[])
+    A2 = MyNode(name="A2", children=[A1])
+    B1 = MyNode(name="B1", children=[])
+    B2 = MyNode(name="B2", children=[B1])
+    C = MyNode(name="C", children=[A2, B2])
+
+    B3 = MyNode(name="B3", children=[])
+
+    if kind == "pattern":
+        replacer = pattern(MyNode)(name="B2") >> B3
+    elif kind == "mapping":
+        replacer = {B2: B3}
+    else:
+
+        def replacer(node, children):
+            if node is B2:
+                return B3
+            return node.__recreate__(children) if children else node
+
+    res = C.replace(replacer)
+
+    assert res is not C
+    assert res.name == "C"
+    assert len(res.children) == 2
+    assert res.children[0] is A2
+    assert res.children[1] is B3
+
+
 def test_example():
     class Example(Annotable, Node):
         def __hash__(self):
@@ -343,17 +373,18 @@ def test_coerce_finder():
 
 
 def test_coerce_replacer():
-    r = _coerce_replacer(lambda x, _, **kwargs: D)
-    assert r(C, {}) == D
+    r = _coerce_replacer(lambda x, children: D if children else C)
+    assert r(C, {"children": []}) is D
+    assert r(C, None) is C
 
     r = _coerce_replacer({C: D, D: E})
     assert r(C, {}) == D
     assert r(D, {}) == E
-    assert r(A, {}, name="A", children=[B, C]) == A
+    assert r(A, {"name": "A", "children": [B, C]}) == A
 
     r = _coerce_replacer(InstanceOf(MyNode) >> _.copy(name=_.name.lower()))
-    assert r(C, {}, name="C", children=[]) == MyNode(name="c", children=[])
-    assert r(D, {}, name="D", children=[]) == MyNode(name="d", children=[])
+    assert r(C, {"name": "C", "children": []}) == MyNode(name="c", children=[])
+    assert r(D, {"name": "D", "children": []}) == MyNode(name="d", children=[])
 
 
 def test_node_find_using_type():