Add canonicalization for addition with negation x + (-y) -> x - y

aesara/tensor/math_opt.py

@@ -1888,6 +1888,36 @@ def local_sub_neg_to_add(fgraph, node):
                     return [new_out]
 
 
+@register_canonicalize
+@local_optimizer([add])
+def local_add_neg_to_sub(fgraph, node):
+    """
+    -x + y -> y - x
+    x + (-y) -> x - y
+    """
+
+    # Rewrite is only applicable when there are two inputs to add
+    if node.op == add and len(node.inputs) == 2:
+        node.outputs[0].dtype
+
+        # Look for pattern with either input order
+        for first, second in (node.inputs, reversed(node.inputs)):
+            if second.owner:
+                if second.owner.op == neg:
+                    pre_neg = second.owner.inputs[0]
+                    new_out = sub(first, pre_neg)
+                    return [new_out]
+
+                # Check if negation is hidden behind a DimShuffle
+                elif isinstance(second.owner.op, DimShuffle):
+                    dimshuffle_op = second.owner.op
+                    dimshuffle_input = second.owner.inputs[0]
+                    if dimshuffle_input.owner and dimshuffle_input.owner.op == neg:
+                        pre_neg = dimshuffle_input.owner.inputs[0]
+                        new_out = sub(first, dimshuffle_op(pre_neg))
+                        return [new_out]
+
+
 @local_optimizer([mul])
 def local_mul_zero(fgraph, node):
     """

tests/tensor/test_math_opt.py

@@ -4566,3 +4566,36 @@ def test_local_sub_neg_to_add(x, y, x_test, y_test):
     ]
     assert nodes == [aet.add]
     assert np.allclose(f(x_test, y_test), x_test - (-y_test))
+
+
+@pytest.mark.parametrize(
+    "x, x_test",
+    [
+        (scalar(), np.full((), 1.0, dtype=config.floatX)),
+        (vector(), np.full(1, 2.0, dtype=config.floatX)),
+        (matrix(), np.full((2, 2), 3.0, dtype=config.floatX)),
+    ],
+)
+@pytest.mark.parametrize(
+    "y, y_test",
+    [
+        (scalar(), np.full((), 1.0, dtype=config.floatX)),
+        (vector(), np.full(1, 2.0, dtype=config.floatX)),
+        (matrix(), np.full((2, 2), 3.0, dtype=config.floatX)),
+    ],
+)
+@pytest.mark.parametrize("first_negative", (True, False))
+def test_local_add_neg_to_sub(x, y, x_test, y_test, first_negative):
+    mode = Mode("py").including("canonicalize")
+
+    out = -x + y if first_negative else x + (-y)
+    f = function([x, y], out, mode=mode)
+
+    nodes = [
+        node.op
+        for node in f.maker.fgraph.toposort()
+        if not isinstance(node.op, DimShuffle)
+    ]
+    assert nodes == [aet.sub]
+    exp = -x_test + y_test if first_negative else x_test + (-y_test)
+    assert np.allclose(f(x_test, y_test), exp)