BUG: For compatibility with 1.5, revert to permitting limited broadca…

…sting of the assignment output

This change got Travis's -10 veto for 1.6.

An unfortunate consequence of reverting this is that some of the broadcasting
error messages get worse, but they're still no worse than in 1.5.

doc/source/reference/c-api.array.rst

@@ -1901,7 +1901,7 @@ Array Functions
 
 .. cfunction:: PyObject* PyArray_EinsteinSum(char* subscripts, npy_intp nop, PyArrayObject** op_in, PyArray_Descr* dtype, NPY_ORDER order, NPY_CASTING casting, PyArrayObject* out)
 
-   .. versionadded:: 1.6
+    .. versionadded:: 1.6
 
     Applies the einstein summation convention to the array operands
     provided, returning a new array or placing the result in *out*.

numpy/core/src/multiarray/ctors.c

@@ -2702,7 +2702,13 @@ PyArray_CopyInto(PyArrayObject *dst, PyArrayObject *src)
 
         op[0] = dst;
         op[1] = src;
-        op_flags[0] = NPY_ITER_WRITEONLY|NPY_ITER_NO_BROADCAST;
+        /*
+         * TODO: In NumPy 2.0, renable NPY_ITER_NO_BROADCAST. This
+         *       was removed during NumPy 1.6 testing for compatibility
+         *       with NumPy 1.5, as per Travis's -10 veto power.
+         */
+        /*op_flags[0] = NPY_ITER_WRITEONLY|NPY_ITER_NO_BROADCAST;*/
+        op_flags[0] = NPY_ITER_WRITEONLY;
         op_flags[1] = NPY_ITER_READONLY;
 
         iter = NpyIter_MultiNew(2, op,

numpy/core/tests/test_multiarray.py

@@ -112,6 +112,30 @@ def test_fill(self):
         x.fill(x[0])
         assert_equal(x['f1'][1], x['f1'][0])
 
+class TestAssignment(TestCase):
+    def test_assignment_broadcasting(self):
+        a = np.arange(6).reshape(2,3)
+
+        # Broadcasting the input to the output
+        a[...] = np.arange(3)
+        assert_equal(a, [[0,1,2],[0,1,2]])
+        a[...] = np.arange(2).reshape(2,1)
+        assert_equal(a, [[0,0,0],[1,1,1]])
+
+        # For compatibility with <= 1.5, a limited version of broadcasting
+        # the output to the input.
+        #
+        # This behavior is inconsistent with NumPy broadcasting
+        # in general, because it only uses one of the two broadcasting
+        # rules (adding a new "1" dimension to the left of the shape),
+        # applied to the output instead of an input. In NumPy 2.0, this kind
+        # of broadcasting assignment will likely be disallowed.
+        a[...] = np.arange(6)[::-1].reshape(1,2,3)
+        assert_equal(a, [[5,4,3],[2,1,0]])
+        # The other type of broadcasting would require a reduction operation.
+        def assign(a,b):
+            a[...] = b
+        assert_raises(ValueError, assign, a, np.arange(12).reshape(2,2,3))
 
 class TestDtypedescr(TestCase):
     def test_construction(self):