numpy 2.4.1__pp311-pypy311_pp73-macosx_14_0_arm64.whl

numpy/_core/tests/test_nditer.py

@@ -0,0 +1,3533 @@
+import inspect
+import subprocess
+import sys
+import textwrap
+import warnings
+
+import pytest
+
+import numpy as np
+import numpy._core._multiarray_tests as _multiarray_tests
+import numpy._core.umath as ncu
+from numpy import all, arange, array, nditer
+from numpy.testing import (
+    HAS_REFCOUNT,
+    IS_64BIT,
+    IS_PYPY,
+    IS_WASM,
+    assert_,
+    assert_array_equal,
+    assert_equal,
+    assert_raises,
+)
+from numpy.testing._private.utils import requires_memory
+
+
+def iter_multi_index(i):
+    ret = []
+    while not i.finished:
+        ret.append(i.multi_index)
+        i.iternext()
+    return ret
+
+def iter_indices(i):
+    ret = []
+    while not i.finished:
+        ret.append(i.index)
+        i.iternext()
+    return ret
+
+def iter_iterindices(i):
+    ret = []
+    while not i.finished:
+        ret.append(i.iterindex)
+        i.iternext()
+    return ret
+
+@pytest.mark.skipif(not HAS_REFCOUNT, reason="Python lacks refcounts")
+def test_iter_refcount():
+    # Make sure the iterator doesn't leak
+
+    # Basic
+    a = arange(6)
+    dt = np.dtype('f4').newbyteorder()
+    rc_a = sys.getrefcount(a)
+    rc_dt = sys.getrefcount(dt)
+    with nditer(a, [],
+                [['readwrite', 'updateifcopy']],
+                casting='unsafe',
+                op_dtypes=[dt]) as it:
+        assert_(not it.iterationneedsapi)
+        assert_(sys.getrefcount(a) > rc_a)
+        assert_(sys.getrefcount(dt) > rc_dt)
+    # del 'it'
+    it = None
+    assert_equal(sys.getrefcount(a), rc_a)
+    assert_equal(sys.getrefcount(dt), rc_dt)
+
+    # With a copy
+    a = arange(6, dtype='f4')
+    dt = np.dtype('f4')
+    rc_a = sys.getrefcount(a)
+    rc_dt = sys.getrefcount(dt)
+    it = nditer(a, [],
+                [['readwrite']],
+                op_dtypes=[dt])
+    rc2_a = sys.getrefcount(a)
+    rc2_dt = sys.getrefcount(dt)
+    it2 = it.copy()
+    assert_(sys.getrefcount(a) > rc2_a)
+    if sys.version_info < (3, 13):
+        # np.dtype('f4') is immortal after Python 3.13
+        assert_(sys.getrefcount(dt) > rc2_dt)
+    it = None
+    assert_equal(sys.getrefcount(a), rc2_a)
+    assert_equal(sys.getrefcount(dt), rc2_dt)
+    it2 = None
+    assert_equal(sys.getrefcount(a), rc_a)
+    assert_equal(sys.getrefcount(dt), rc_dt)
+
+def test_iter_best_order():
+    # The iterator should always find the iteration order
+    # with increasing memory addresses
+
+    # Test the ordering for 1-D to 5-D shapes
+    for shape in [(5,), (3, 4), (2, 3, 4), (2, 3, 4, 3), (2, 3, 2, 2, 3)]:
+        a = arange(np.prod(shape))
+        # Test each combination of positive and negative strides
+        for dirs in range(2**len(shape)):
+            dirs_index = [slice(None)] * len(shape)
+            for bit in range(len(shape)):
+                if ((2**bit) & dirs):
+                    dirs_index[bit] = slice(None, None, -1)
+            dirs_index = tuple(dirs_index)
+
+            aview = a.reshape(shape)[dirs_index]
+            # C-order
+            i = nditer(aview, [], [['readonly']])
+            assert_equal(list(i), a)
+            # Fortran-order
+            i = nditer(aview.T, [], [['readonly']])
+            assert_equal(list(i), a)
+            # Other order
+            if len(shape) > 2:
+                i = nditer(aview.swapaxes(0, 1), [], [['readonly']])
+                assert_equal(list(i), a)
+
+def test_iter_c_order():
+    # Test forcing C order
+
+    # Test the ordering for 1-D to 5-D shapes
+    for shape in [(5,), (3, 4), (2, 3, 4), (2, 3, 4, 3), (2, 3, 2, 2, 3)]:
+        a = arange(np.prod(shape))
+        # Test each combination of positive and negative strides
+        for dirs in range(2**len(shape)):
+            dirs_index = [slice(None)] * len(shape)
+            for bit in range(len(shape)):
+                if ((2**bit) & dirs):
+                    dirs_index[bit] = slice(None, None, -1)
+            dirs_index = tuple(dirs_index)
+
+            aview = a.reshape(shape)[dirs_index]
+            # C-order
+            i = nditer(aview, order='C')
+            assert_equal(list(i), aview.ravel(order='C'))
+            # Fortran-order
+            i = nditer(aview.T, order='C')
+            assert_equal(list(i), aview.T.ravel(order='C'))
+            # Other order
+            if len(shape) > 2:
+                i = nditer(aview.swapaxes(0, 1), order='C')
+                assert_equal(list(i),
+                                    aview.swapaxes(0, 1).ravel(order='C'))
+
+def test_iter_f_order():
+    # Test forcing F order
+
+    # Test the ordering for 1-D to 5-D shapes
+    for shape in [(5,), (3, 4), (2, 3, 4), (2, 3, 4, 3), (2, 3, 2, 2, 3)]:
+        a = arange(np.prod(shape))
+        # Test each combination of positive and negative strides
+        for dirs in range(2**len(shape)):
+            dirs_index = [slice(None)] * len(shape)
+            for bit in range(len(shape)):
+                if ((2**bit) & dirs):
+                    dirs_index[bit] = slice(None, None, -1)
+            dirs_index = tuple(dirs_index)
+
+            aview = a.reshape(shape)[dirs_index]
+            # C-order
+            i = nditer(aview, order='F')
+            assert_equal(list(i), aview.ravel(order='F'))
+            # Fortran-order
+            i = nditer(aview.T, order='F')
+            assert_equal(list(i), aview.T.ravel(order='F'))
+            # Other order
+            if len(shape) > 2:
+                i = nditer(aview.swapaxes(0, 1), order='F')
+                assert_equal(list(i),
+                                    aview.swapaxes(0, 1).ravel(order='F'))
+
+def test_iter_c_or_f_order():
+    # Test forcing any contiguous (C or F) order
+
+    # Test the ordering for 1-D to 5-D shapes
+    for shape in [(5,), (3, 4), (2, 3, 4), (2, 3, 4, 3), (2, 3, 2, 2, 3)]:
+        a = arange(np.prod(shape))
+        # Test each combination of positive and negative strides
+        for dirs in range(2**len(shape)):
+            dirs_index = [slice(None)] * len(shape)
+            for bit in range(len(shape)):
+                if ((2**bit) & dirs):
+                    dirs_index[bit] = slice(None, None, -1)
+            dirs_index = tuple(dirs_index)
+
+            aview = a.reshape(shape)[dirs_index]
+            # C-order
+            i = nditer(aview, order='A')
+            assert_equal(list(i), aview.ravel(order='A'))
+            # Fortran-order
+            i = nditer(aview.T, order='A')
+            assert_equal(list(i), aview.T.ravel(order='A'))
+            # Other order
+            if len(shape) > 2:
+                i = nditer(aview.swapaxes(0, 1), order='A')
+                assert_equal(list(i),
+                                    aview.swapaxes(0, 1).ravel(order='A'))
+
+def test_nditer_multi_index_set():
+    # Test the multi_index set
+    a = np.arange(6).reshape(2, 3)
+    it = np.nditer(a, flags=['multi_index'])
+
+    # Removes the iteration on two first elements of a[0]
+    it.multi_index = (0, 2,)
+
+    assert_equal(list(it), [2, 3, 4, 5])
+
+@pytest.mark.skipif(not HAS_REFCOUNT, reason="Python lacks refcounts")
+def test_nditer_multi_index_set_refcount():
+    # Test if the reference count on index variable is decreased
+
+    index = 0
+    i = np.nditer(np.array([111, 222, 333, 444]), flags=['multi_index'])
+
+    start_count = sys.getrefcount(index)
+    i.multi_index = (index,)
+    end_count = sys.getrefcount(index)
+
+    assert_equal(start_count, end_count)
+
+def test_iter_best_order_multi_index_1d():
+    # The multi-indices should be correct with any reordering
+
+    a = arange(4)
+    # 1D order
+    i = nditer(a, ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i), [(0,), (1,), (2,), (3,)])
+    # 1D reversed order
+    i = nditer(a[::-1], ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i), [(3,), (2,), (1,), (0,)])
+
+def test_iter_best_order_multi_index_2d():
+    # The multi-indices should be correct with any reordering
+
+    a = arange(6)
+    # 2D C-order
+    i = nditer(a.reshape(2, 3), ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i), [(0, 0), (0, 1), (0, 2), (1, 0), (1, 1), (1, 2)])
+    # 2D Fortran-order
+    i = nditer(a.reshape(2, 3).copy(order='F'), ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i), [(0, 0), (1, 0), (0, 1), (1, 1), (0, 2), (1, 2)])
+    # 2D reversed C-order
+    i = nditer(a.reshape(2, 3)[::-1], ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i), [(1, 0), (1, 1), (1, 2), (0, 0), (0, 1), (0, 2)])
+    i = nditer(a.reshape(2, 3)[:, ::-1], ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i), [(0, 2), (0, 1), (0, 0), (1, 2), (1, 1), (1, 0)])
+    i = nditer(a.reshape(2, 3)[::-1, ::-1], ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i), [(1, 2), (1, 1), (1, 0), (0, 2), (0, 1), (0, 0)])
+    # 2D reversed Fortran-order
+    i = nditer(a.reshape(2, 3).copy(order='F')[::-1], ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i), [(1, 0), (0, 0), (1, 1), (0, 1), (1, 2), (0, 2)])
+    i = nditer(a.reshape(2, 3).copy(order='F')[:, ::-1],
+                                                   ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i), [(0, 2), (1, 2), (0, 1), (1, 1), (0, 0), (1, 0)])
+    i = nditer(a.reshape(2, 3).copy(order='F')[::-1, ::-1],
+                                                   ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i), [(1, 2), (0, 2), (1, 1), (0, 1), (1, 0), (0, 0)])
+
+def test_iter_best_order_multi_index_3d():
+    # The multi-indices should be correct with any reordering
+
+    a = arange(12)
+    # 3D C-order
+    i = nditer(a.reshape(2, 3, 2), ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i),
+                            [(0, 0, 0), (0, 0, 1), (0, 1, 0), (0, 1, 1), (0, 2, 0), (0, 2, 1),
+                             (1, 0, 0), (1, 0, 1), (1, 1, 0), (1, 1, 1), (1, 2, 0), (1, 2, 1)])
+    # 3D Fortran-order
+    i = nditer(a.reshape(2, 3, 2).copy(order='F'), ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i),
+                            [(0, 0, 0), (1, 0, 0), (0, 1, 0), (1, 1, 0), (0, 2, 0), (1, 2, 0),
+                             (0, 0, 1), (1, 0, 1), (0, 1, 1), (1, 1, 1), (0, 2, 1), (1, 2, 1)])
+    # 3D reversed C-order
+    i = nditer(a.reshape(2, 3, 2)[::-1], ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i),
+                            [(1, 0, 0), (1, 0, 1), (1, 1, 0), (1, 1, 1), (1, 2, 0), (1, 2, 1),
+                             (0, 0, 0), (0, 0, 1), (0, 1, 0), (0, 1, 1), (0, 2, 0), (0, 2, 1)])
+    i = nditer(a.reshape(2, 3, 2)[:, ::-1], ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i),
+                            [(0, 2, 0), (0, 2, 1), (0, 1, 0), (0, 1, 1), (0, 0, 0), (0, 0, 1),
+                             (1, 2, 0), (1, 2, 1), (1, 1, 0), (1, 1, 1), (1, 0, 0), (1, 0, 1)])
+    i = nditer(a.reshape(2, 3, 2)[:, :, ::-1], ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i),
+                            [(0, 0, 1), (0, 0, 0), (0, 1, 1), (0, 1, 0), (0, 2, 1), (0, 2, 0),
+                             (1, 0, 1), (1, 0, 0), (1, 1, 1), (1, 1, 0), (1, 2, 1), (1, 2, 0)])
+    # 3D reversed Fortran-order
+    i = nditer(a.reshape(2, 3, 2).copy(order='F')[::-1],
+                                                    ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i),
+                            [(1, 0, 0), (0, 0, 0), (1, 1, 0), (0, 1, 0), (1, 2, 0), (0, 2, 0),
+                             (1, 0, 1), (0, 0, 1), (1, 1, 1), (0, 1, 1), (1, 2, 1), (0, 2, 1)])
+    i = nditer(a.reshape(2, 3, 2).copy(order='F')[:, ::-1],
+                                                    ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i),
+                            [(0, 2, 0), (1, 2, 0), (0, 1, 0), (1, 1, 0), (0, 0, 0), (1, 0, 0),
+                             (0, 2, 1), (1, 2, 1), (0, 1, 1), (1, 1, 1), (0, 0, 1), (1, 0, 1)])
+    i = nditer(a.reshape(2, 3, 2).copy(order='F')[:, :, ::-1],
+                                                    ['multi_index'], [['readonly']])
+    assert_equal(iter_multi_index(i),
+                            [(0, 0, 1), (1, 0, 1), (0, 1, 1), (1, 1, 1), (0, 2, 1), (1, 2, 1),
+                             (0, 0, 0), (1, 0, 0), (0, 1, 0), (1, 1, 0), (0, 2, 0), (1, 2, 0)])
+
+def test_iter_best_order_c_index_1d():
+    # The C index should be correct with any reordering
+
+    a = arange(4)
+    # 1D order
+    i = nditer(a, ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i), [0, 1, 2, 3])
+    # 1D reversed order
+    i = nditer(a[::-1], ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i), [3, 2, 1, 0])
+
+def test_iter_best_order_c_index_2d():
+    # The C index should be correct with any reordering
+
+    a = arange(6)
+    # 2D C-order
+    i = nditer(a.reshape(2, 3), ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i), [0, 1, 2, 3, 4, 5])
+    # 2D Fortran-order
+    i = nditer(a.reshape(2, 3).copy(order='F'),
+                                    ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i), [0, 3, 1, 4, 2, 5])
+    # 2D reversed C-order
+    i = nditer(a.reshape(2, 3)[::-1], ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i), [3, 4, 5, 0, 1, 2])
+    i = nditer(a.reshape(2, 3)[:, ::-1], ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i), [2, 1, 0, 5, 4, 3])
+    i = nditer(a.reshape(2, 3)[::-1, ::-1], ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i), [5, 4, 3, 2, 1, 0])
+    # 2D reversed Fortran-order
+    i = nditer(a.reshape(2, 3).copy(order='F')[::-1],
+                                    ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i), [3, 0, 4, 1, 5, 2])
+    i = nditer(a.reshape(2, 3).copy(order='F')[:, ::-1],
+                                    ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i), [2, 5, 1, 4, 0, 3])
+    i = nditer(a.reshape(2, 3).copy(order='F')[::-1, ::-1],
+                                    ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i), [5, 2, 4, 1, 3, 0])
+
+def test_iter_best_order_c_index_3d():
+    # The C index should be correct with any reordering
+
+    a = arange(12)
+    # 3D C-order
+    i = nditer(a.reshape(2, 3, 2), ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11])
+    # 3D Fortran-order
+    i = nditer(a.reshape(2, 3, 2).copy(order='F'),
+                                    ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [0, 6, 2, 8, 4, 10, 1, 7, 3, 9, 5, 11])
+    # 3D reversed C-order
+    i = nditer(a.reshape(2, 3, 2)[::-1], ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5])
+    i = nditer(a.reshape(2, 3, 2)[:, ::-1], ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [4, 5, 2, 3, 0, 1, 10, 11, 8, 9, 6, 7])
+    i = nditer(a.reshape(2, 3, 2)[:, :, ::-1], ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10])
+    # 3D reversed Fortran-order
+    i = nditer(a.reshape(2, 3, 2).copy(order='F')[::-1],
+                                    ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [6, 0, 8, 2, 10, 4, 7, 1, 9, 3, 11, 5])
+    i = nditer(a.reshape(2, 3, 2).copy(order='F')[:, ::-1],
+                                    ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [4, 10, 2, 8, 0, 6, 5, 11, 3, 9, 1, 7])
+    i = nditer(a.reshape(2, 3, 2).copy(order='F')[:, :, ::-1],
+                                    ['c_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [1, 7, 3, 9, 5, 11, 0, 6, 2, 8, 4, 10])
+
+def test_iter_best_order_f_index_1d():
+    # The Fortran index should be correct with any reordering
+
+    a = arange(4)
+    # 1D order
+    i = nditer(a, ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i), [0, 1, 2, 3])
+    # 1D reversed order
+    i = nditer(a[::-1], ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i), [3, 2, 1, 0])
+
+def test_iter_best_order_f_index_2d():
+    # The Fortran index should be correct with any reordering
+
+    a = arange(6)
+    # 2D C-order
+    i = nditer(a.reshape(2, 3), ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i), [0, 2, 4, 1, 3, 5])
+    # 2D Fortran-order
+    i = nditer(a.reshape(2, 3).copy(order='F'),
+                                    ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i), [0, 1, 2, 3, 4, 5])
+    # 2D reversed C-order
+    i = nditer(a.reshape(2, 3)[::-1], ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i), [1, 3, 5, 0, 2, 4])
+    i = nditer(a.reshape(2, 3)[:, ::-1], ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i), [4, 2, 0, 5, 3, 1])
+    i = nditer(a.reshape(2, 3)[::-1, ::-1], ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i), [5, 3, 1, 4, 2, 0])
+    # 2D reversed Fortran-order
+    i = nditer(a.reshape(2, 3).copy(order='F')[::-1],
+                                    ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i), [1, 0, 3, 2, 5, 4])
+    i = nditer(a.reshape(2, 3).copy(order='F')[:, ::-1],
+                                    ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i), [4, 5, 2, 3, 0, 1])
+    i = nditer(a.reshape(2, 3).copy(order='F')[::-1, ::-1],
+                                    ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i), [5, 4, 3, 2, 1, 0])
+
+def test_iter_best_order_f_index_3d():
+    # The Fortran index should be correct with any reordering
+
+    a = arange(12)
+    # 3D C-order
+    i = nditer(a.reshape(2, 3, 2), ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [0, 6, 2, 8, 4, 10, 1, 7, 3, 9, 5, 11])
+    # 3D Fortran-order
+    i = nditer(a.reshape(2, 3, 2).copy(order='F'),
+                                    ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11])
+    # 3D reversed C-order
+    i = nditer(a.reshape(2, 3, 2)[::-1], ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [1, 7, 3, 9, 5, 11, 0, 6, 2, 8, 4, 10])
+    i = nditer(a.reshape(2, 3, 2)[:, ::-1], ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [4, 10, 2, 8, 0, 6, 5, 11, 3, 9, 1, 7])
+    i = nditer(a.reshape(2, 3, 2)[:, :, ::-1], ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [6, 0, 8, 2, 10, 4, 7, 1, 9, 3, 11, 5])
+    # 3D reversed Fortran-order
+    i = nditer(a.reshape(2, 3, 2).copy(order='F')[::-1],
+                                    ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10])
+    i = nditer(a.reshape(2, 3, 2).copy(order='F')[:, ::-1],
+                                    ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [4, 5, 2, 3, 0, 1, 10, 11, 8, 9, 6, 7])
+    i = nditer(a.reshape(2, 3, 2).copy(order='F')[:, :, ::-1],
+                                    ['f_index'], [['readonly']])
+    assert_equal(iter_indices(i),
+                            [6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5])
+
+def test_iter_no_inner_full_coalesce():
+    # Check no_inner iterators which coalesce into a single inner loop
+
+    for shape in [(5,), (3, 4), (2, 3, 4), (2, 3, 4, 3), (2, 3, 2, 2, 3)]:
+        size = np.prod(shape)
+        a = arange(size)
+        # Test each combination of forward and backwards indexing
+        for dirs in range(2**len(shape)):
+            dirs_index = [slice(None)] * len(shape)
+            for bit in range(len(shape)):
+                if ((2**bit) & dirs):
+                    dirs_index[bit] = slice(None, None, -1)
+            dirs_index = tuple(dirs_index)
+
+            aview = a.reshape(shape)[dirs_index]
+            # C-order
+            i = nditer(aview, ['external_loop'], [['readonly']])
+            assert_equal(i.ndim, 1)
+            assert_equal(i[0].shape, (size,))
+            # Fortran-order
+            i = nditer(aview.T, ['external_loop'], [['readonly']])
+            assert_equal(i.ndim, 1)
+            assert_equal(i[0].shape, (size,))
+            # Other order
+            if len(shape) > 2:
+                i = nditer(aview.swapaxes(0, 1),
+                                    ['external_loop'], [['readonly']])
+                assert_equal(i.ndim, 1)
+                assert_equal(i[0].shape, (size,))
+
+def test_iter_no_inner_dim_coalescing():
+    # Check no_inner iterators whose dimensions may not coalesce completely
+
+    # Skipping the last element in a dimension prevents coalescing
+    # with the next-bigger dimension
+    a = arange(24).reshape(2, 3, 4)[:, :, :-1]
+    i = nditer(a, ['external_loop'], [['readonly']])
+    assert_equal(i.ndim, 2)
+    assert_equal(i[0].shape, (3,))
+    a = arange(24).reshape(2, 3, 4)[:, :-1, :]
+    i = nditer(a, ['external_loop'], [['readonly']])
+    assert_equal(i.ndim, 2)
+    assert_equal(i[0].shape, (8,))
+    a = arange(24).reshape(2, 3, 4)[:-1, :, :]
+    i = nditer(a, ['external_loop'], [['readonly']])
+    assert_equal(i.ndim, 1)
+    assert_equal(i[0].shape, (12,))
+
+    # Even with lots of 1-sized dimensions, should still coalesce
+    a = arange(24).reshape(1, 1, 2, 1, 1, 3, 1, 1, 4, 1, 1)
+    i = nditer(a, ['external_loop'], [['readonly']])
+    assert_equal(i.ndim, 1)
+    assert_equal(i[0].shape, (24,))
+
+def test_iter_dim_coalescing():
+    # Check that the correct number of dimensions are coalesced
+
+    # Tracking a multi-index disables coalescing
+    a = arange(24).reshape(2, 3, 4)
+    i = nditer(a, ['multi_index'], [['readonly']])
+    assert_equal(i.ndim, 3)
+
+    # A tracked index can allow coalescing if it's compatible with the array
+    a3d = arange(24).reshape(2, 3, 4)
+    i = nditer(a3d, ['c_index'], [['readonly']])
+    assert_equal(i.ndim, 1)
+    i = nditer(a3d.swapaxes(0, 1), ['c_index'], [['readonly']])
+    assert_equal(i.ndim, 3)
+    i = nditer(a3d.T, ['c_index'], [['readonly']])
+    assert_equal(i.ndim, 3)
+    i = nditer(a3d.T, ['f_index'], [['readonly']])
+    assert_equal(i.ndim, 1)
+    i = nditer(a3d.T.swapaxes(0, 1), ['f_index'], [['readonly']])
+    assert_equal(i.ndim, 3)
+
+    # When C or F order is forced, coalescing may still occur
+    a3d = arange(24).reshape(2, 3, 4)
+    i = nditer(a3d, order='C')
+    assert_equal(i.ndim, 1)
+    i = nditer(a3d.T, order='C')
+    assert_equal(i.ndim, 3)
+    i = nditer(a3d, order='F')
+    assert_equal(i.ndim, 3)
+    i = nditer(a3d.T, order='F')
+    assert_equal(i.ndim, 1)
+    i = nditer(a3d, order='A')
+    assert_equal(i.ndim, 1)
+    i = nditer(a3d.T, order='A')
+    assert_equal(i.ndim, 1)
+
+def test_iter_broadcasting():
+    # Standard NumPy broadcasting rules
+
+    # 1D with scalar
+    i = nditer([arange(6), np.int32(2)], ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 6)
+    assert_equal(i.shape, (6,))
+
+    # 2D with scalar
+    i = nditer([arange(6).reshape(2, 3), np.int32(2)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 6)
+    assert_equal(i.shape, (2, 3))
+    # 2D with 1D
+    i = nditer([arange(6).reshape(2, 3), arange(3)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 6)
+    assert_equal(i.shape, (2, 3))
+    i = nditer([arange(2).reshape(2, 1), arange(3)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 6)
+    assert_equal(i.shape, (2, 3))
+    # 2D with 2D
+    i = nditer([arange(2).reshape(2, 1), arange(3).reshape(1, 3)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 6)
+    assert_equal(i.shape, (2, 3))
+
+    # 3D with scalar
+    i = nditer([np.int32(2), arange(24).reshape(4, 2, 3)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 24)
+    assert_equal(i.shape, (4, 2, 3))
+    # 3D with 1D
+    i = nditer([arange(3), arange(24).reshape(4, 2, 3)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 24)
+    assert_equal(i.shape, (4, 2, 3))
+    i = nditer([arange(3), arange(8).reshape(4, 2, 1)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 24)
+    assert_equal(i.shape, (4, 2, 3))
+    # 3D with 2D
+    i = nditer([arange(6).reshape(2, 3), arange(24).reshape(4, 2, 3)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 24)
+    assert_equal(i.shape, (4, 2, 3))
+    i = nditer([arange(2).reshape(2, 1), arange(24).reshape(4, 2, 3)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 24)
+    assert_equal(i.shape, (4, 2, 3))
+    i = nditer([arange(3).reshape(1, 3), arange(8).reshape(4, 2, 1)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 24)
+    assert_equal(i.shape, (4, 2, 3))
+    # 3D with 3D
+    i = nditer([arange(2).reshape(1, 2, 1), arange(3).reshape(1, 1, 3),
+                        arange(4).reshape(4, 1, 1)],
+                        ['multi_index'], [['readonly']] * 3)
+    assert_equal(i.itersize, 24)
+    assert_equal(i.shape, (4, 2, 3))
+    i = nditer([arange(6).reshape(1, 2, 3), arange(4).reshape(4, 1, 1)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 24)
+    assert_equal(i.shape, (4, 2, 3))
+    i = nditer([arange(24).reshape(4, 2, 3), arange(12).reshape(4, 1, 3)],
+                        ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.itersize, 24)
+    assert_equal(i.shape, (4, 2, 3))
+
+def test_iter_itershape():
+    # Check that allocated outputs work with a specified shape
+    a = np.arange(6, dtype='i2').reshape(2, 3)
+    i = nditer([a, None], [], [['readonly'], ['writeonly', 'allocate']],
+                            op_axes=[[0, 1, None], None],
+                            itershape=(-1, -1, 4))
+    assert_equal(i.operands[1].shape, (2, 3, 4))
+    assert_equal(i.operands[1].strides, (24, 8, 2))
+
+    i = nditer([a.T, None], [], [['readonly'], ['writeonly', 'allocate']],
+                            op_axes=[[0, 1, None], None],
+                            itershape=(-1, -1, 4))
+    assert_equal(i.operands[1].shape, (3, 2, 4))
+    assert_equal(i.operands[1].strides, (8, 24, 2))
+
+    i = nditer([a.T, None], [], [['readonly'], ['writeonly', 'allocate']],
+                            order='F',
+                            op_axes=[[0, 1, None], None],
+                            itershape=(-1, -1, 4))
+    assert_equal(i.operands[1].shape, (3, 2, 4))
+    assert_equal(i.operands[1].strides, (2, 6, 12))
+
+    # If we specify 1 in the itershape, it shouldn't allow broadcasting
+    # of that dimension to a bigger value
+    assert_raises(ValueError, nditer, [a, None], [],
+                            [['readonly'], ['writeonly', 'allocate']],
+                            op_axes=[[0, 1, None], None],
+                            itershape=(-1, 1, 4))
+    # Test bug that for no op_axes but itershape, they are NULLed correctly
+    i = np.nditer([np.ones(2), None, None], itershape=(2,))
+
+def test_iter_broadcasting_errors():
+    # Check that errors are thrown for bad broadcasting shapes
+
+    # 1D with 1D
+    assert_raises(ValueError, nditer, [arange(2), arange(3)],
+                    [], [['readonly']] * 2)
+    # 2D with 1D
+    assert_raises(ValueError, nditer,
+                    [arange(6).reshape(2, 3), arange(2)],
+                    [], [['readonly']] * 2)
+    # 2D with 2D
+    assert_raises(ValueError, nditer,
+                    [arange(6).reshape(2, 3), arange(9).reshape(3, 3)],
+                    [], [['readonly']] * 2)
+    assert_raises(ValueError, nditer,
+                    [arange(6).reshape(2, 3), arange(4).reshape(2, 2)],
+                    [], [['readonly']] * 2)
+    # 3D with 3D
+    assert_raises(ValueError, nditer,
+                    [arange(36).reshape(3, 3, 4), arange(24).reshape(2, 3, 4)],
+                    [], [['readonly']] * 2)
+    assert_raises(ValueError, nditer,
+                    [arange(8).reshape(2, 4, 1), arange(24).reshape(2, 3, 4)],
+                    [], [['readonly']] * 2)
+
+    # Verify that the error message mentions the right shapes
+    try:
+        nditer([arange(2).reshape(1, 2, 1),
+                arange(3).reshape(1, 3),
+                arange(6).reshape(2, 3)],
+               [],
+               [['readonly'], ['readonly'], ['writeonly', 'no_broadcast']])
+        raise AssertionError('Should have raised a broadcast error')
+    except ValueError as e:
+        msg = str(e)
+        # The message should contain the shape of the 3rd operand
+        assert_(msg.find('(2,3)') >= 0,
+                f'Message "{msg}" doesn\'t contain operand shape (2,3)')
+        # The message should contain the broadcast shape
+        assert_(msg.find('(1,2,3)') >= 0,
+                f'Message "{msg}" doesn\'t contain broadcast shape (1,2,3)')
+
+    try:
+        nditer([arange(6).reshape(2, 3), arange(2)],
+               [],
+               [['readonly'], ['readonly']],
+               op_axes=[[0, 1], [0, np.newaxis]],
+               itershape=(4, 3))
+        raise AssertionError('Should have raised a broadcast error')
+    except ValueError as e:
+        msg = str(e)
+        # The message should contain "shape->remappedshape" for each operand
+        assert_(msg.find('(2,3)->(2,3)') >= 0,
+            f'Message "{msg}" doesn\'t contain operand shape (2,3)->(2,3)')
+        assert_(msg.find('(2,)->(2,newaxis)') >= 0,
+                ('Message "%s" doesn\'t contain remapped operand shape'
+                '(2,)->(2,newaxis)') % msg)
+        # The message should contain the itershape parameter
+        assert_(msg.find('(4,3)') >= 0,
+                f'Message "{msg}" doesn\'t contain itershape parameter (4,3)')
+
+    try:
+        nditer([np.zeros((2, 1, 1)), np.zeros((2,))],
+               [],
+               [['writeonly', 'no_broadcast'], ['readonly']])
+        raise AssertionError('Should have raised a broadcast error')
+    except ValueError as e:
+        msg = str(e)
+        # The message should contain the shape of the bad operand
+        assert_(msg.find('(2,1,1)') >= 0,
+            f'Message "{msg}" doesn\'t contain operand shape (2,1,1)')
+        # The message should contain the broadcast shape
+        assert_(msg.find('(2,1,2)') >= 0,
+                f'Message "{msg}" doesn\'t contain the broadcast shape (2,1,2)')
+
+def test_iter_flags_errors():
+    # Check that bad combinations of flags produce errors
+
+    a = arange(6)
+
+    # Not enough operands
+    assert_raises(ValueError, nditer, [], [], [])
+    # Bad global flag
+    assert_raises(ValueError, nditer, [a], ['bad flag'], [['readonly']])
+    # Bad op flag
+    assert_raises(ValueError, nditer, [a], [], [['readonly', 'bad flag']])
+    # Bad order parameter
+    assert_raises(ValueError, nditer, [a], [], [['readonly']], order='G')
+    # Bad casting parameter
+    assert_raises(ValueError, nditer, [a], [], [['readonly']], casting='noon')
+    # op_flags must match ops
+    assert_raises(ValueError, nditer, [a] * 3, [], [['readonly']] * 2)
+    # Cannot track both a C and an F index
+    assert_raises(ValueError, nditer, a,
+                ['c_index', 'f_index'], [['readonly']])
+    # Inner iteration and multi-indices/indices are incompatible
+    assert_raises(ValueError, nditer, a,
+                ['external_loop', 'multi_index'], [['readonly']])
+    assert_raises(ValueError, nditer, a,
+                ['external_loop', 'c_index'], [['readonly']])
+    assert_raises(ValueError, nditer, a,
+                ['external_loop', 'f_index'], [['readonly']])
+    # Must specify exactly one of readwrite/readonly/writeonly per operand
+    assert_raises(ValueError, nditer, a, [], [[]])
+    assert_raises(ValueError, nditer, a, [], [['readonly', 'writeonly']])
+    assert_raises(ValueError, nditer, a, [], [['readonly', 'readwrite']])
+    assert_raises(ValueError, nditer, a, [], [['writeonly', 'readwrite']])
+    assert_raises(ValueError, nditer, a,
+                [], [['readonly', 'writeonly', 'readwrite']])
+    # Python scalars are always readonly
+    assert_raises(TypeError, nditer, 1.5, [], [['writeonly']])
+    assert_raises(TypeError, nditer, 1.5, [], [['readwrite']])
+    # Array scalars are always readonly
+    assert_raises(TypeError, nditer, np.int32(1), [], [['writeonly']])
+    assert_raises(TypeError, nditer, np.int32(1), [], [['readwrite']])
+    # Check readonly array
+    a.flags.writeable = False
+    assert_raises(ValueError, nditer, a, [], [['writeonly']])
+    assert_raises(ValueError, nditer, a, [], [['readwrite']])
+    a.flags.writeable = True
+    # Multi-indices available only with the multi_index flag
+    i = nditer(arange(6), [], [['readonly']])
+    assert_raises(ValueError, lambda i: i.multi_index, i)
+    # Index available only with an index flag
+    assert_raises(ValueError, lambda i: i.index, i)
+    # GotoCoords and GotoIndex incompatible with buffering or no_inner
+
+    def assign_multi_index(i):
+        i.multi_index = (0,)
+
+    def assign_index(i):
+        i.index = 0
+
+    def assign_iterindex(i):
+        i.iterindex = 0
+
+    def assign_iterrange(i):
+        i.iterrange = (0, 1)
+    i = nditer(arange(6), ['external_loop'])
+    assert_raises(ValueError, assign_multi_index, i)
+    assert_raises(ValueError, assign_index, i)
+    assert_raises(ValueError, assign_iterindex, i)
+    assert_raises(ValueError, assign_iterrange, i)
+    i = nditer(arange(6), ['buffered'])
+    assert_raises(ValueError, assign_multi_index, i)
+    assert_raises(ValueError, assign_index, i)
+    assert_raises(ValueError, assign_iterrange, i)
+    # Can't iterate if size is zero
+    assert_raises(ValueError, nditer, np.array([]))
+
+def test_iter_slice():
+    a, b, c = np.arange(3), np.arange(3), np.arange(3.)
+    i = nditer([a, b, c], [], ['readwrite'])
+    with i:
+        i[0:2] = (3, 3)
+        assert_equal(a, [3, 1, 2])
+        assert_equal(b, [3, 1, 2])
+        assert_equal(c, [0, 1, 2])
+        i[1] = 12
+        assert_equal(i[0:2], [3, 12])
+
+def test_iter_assign_mapping():
+    a = np.arange(24, dtype='f8').reshape(2, 3, 4).T
+    it = np.nditer(a, [], [['readwrite', 'updateifcopy']],
+                       casting='same_kind', op_dtypes=[np.dtype('f4')])
+    with it:
+        it.operands[0][...] = 3
+        it.operands[0][...] = 14
+    assert_equal(a, 14)
+    it = np.nditer(a, [], [['readwrite', 'updateifcopy']],
+                       casting='same_kind', op_dtypes=[np.dtype('f4')])
+    with it:
+        x = it.operands[0][-1:1]
+        x[...] = 14
+        it.operands[0][...] = -1234
+    assert_equal(a, -1234)
+    # check for no warnings on dealloc
+    x = None
+    it = None
+
+def test_iter_nbo_align_contig():
+    # Check that byte order, alignment, and contig changes work
+
+    # Byte order change by requesting a specific dtype
+    a = np.arange(6, dtype='f4')
+    au = a.byteswap()
+    au = au.view(au.dtype.newbyteorder())
+    assert_(a.dtype.byteorder != au.dtype.byteorder)
+    i = nditer(au, [], [['readwrite', 'updateifcopy']],
+                        casting='equiv',
+                        op_dtypes=[np.dtype('f4')])
+    with i:
+        # context manager triggers WRITEBACKIFCOPY on i at exit
+        assert_equal(i.dtypes[0].byteorder, a.dtype.byteorder)
+        assert_equal(i.operands[0].dtype.byteorder, a.dtype.byteorder)
+        assert_equal(i.operands[0], a)
+        i.operands[0][:] = 2
+    assert_equal(au, [2] * 6)
+    del i  # should not raise a warning
+    # Byte order change by requesting NBO
+    a = np.arange(6, dtype='f4')
+    au = a.byteswap()
+    au = au.view(au.dtype.newbyteorder())
+    assert_(a.dtype.byteorder != au.dtype.byteorder)
+    with nditer(au, [], [['readwrite', 'updateifcopy', 'nbo']],
+                        casting='equiv') as i:
+        # context manager triggers UPDATEIFCOPY on i at exit
+        assert_equal(i.dtypes[0].byteorder, a.dtype.byteorder)
+        assert_equal(i.operands[0].dtype.byteorder, a.dtype.byteorder)
+        assert_equal(i.operands[0], a)
+        i.operands[0][:] = 12345
+        i.operands[0][:] = 2
+    assert_equal(au, [2] * 6)
+
+    # Unaligned input
+    a = np.zeros((6 * 4 + 1,), dtype='i1')[1:]
+    a = a.view('f4')
+    a[:] = np.arange(6, dtype='f4')
+    assert_(not a.flags.aligned)
+    # Without 'aligned', shouldn't copy
+    i = nditer(a, [], [['readonly']])
+    assert_(not i.operands[0].flags.aligned)
+    assert_equal(i.operands[0], a)
+    # With 'aligned', should make a copy
+    with nditer(a, [], [['readwrite', 'updateifcopy', 'aligned']]) as i:
+        assert_(i.operands[0].flags.aligned)
+        # context manager triggers UPDATEIFCOPY on i at exit
+        assert_equal(i.operands[0], a)
+        i.operands[0][:] = 3
+    assert_equal(a, [3] * 6)
+
+    # Discontiguous input
+    a = arange(12)
+    # If it is contiguous, shouldn't copy
+    i = nditer(a[:6], [], [['readonly']])
+    assert_(i.operands[0].flags.contiguous)
+    assert_equal(i.operands[0], a[:6])
+    # If it isn't contiguous, should buffer
+    i = nditer(a[::2], ['buffered', 'external_loop'],
+                        [['readonly', 'contig']],
+                        buffersize=10)
+    assert_(i[0].flags.contiguous)
+    assert_equal(i[0], a[::2])
+
+def test_iter_array_cast():
+    # Check that arrays are cast as requested
+
+    # No cast 'f4' -> 'f4'
+    a = np.arange(6, dtype='f4').reshape(2, 3)
+    i = nditer(a, [], [['readwrite']], op_dtypes=[np.dtype('f4')])
+    with i:
+        assert_equal(i.operands[0], a)
+        assert_equal(i.operands[0].dtype, np.dtype('f4'))
+
+    # Byte-order cast '<f4' -> '>f4'
+    a = np.arange(6, dtype='<f4').reshape(2, 3)
+    with nditer(a, [], [['readwrite', 'updateifcopy']],
+            casting='equiv',
+            op_dtypes=[np.dtype('>f4')]) as i:
+        assert_equal(i.operands[0], a)
+        assert_equal(i.operands[0].dtype, np.dtype('>f4'))
+
+    # Safe case 'f4' -> 'f8'
+    a = np.arange(24, dtype='f4').reshape(2, 3, 4).swapaxes(1, 2)
+    i = nditer(a, [], [['readonly', 'copy']],
+            casting='safe',
+            op_dtypes=[np.dtype('f8')])
+    assert_equal(i.operands[0], a)
+    assert_equal(i.operands[0].dtype, np.dtype('f8'))
+    # The memory layout of the temporary should match a (a is (48,4,16))
+    # except negative strides get flipped to positive strides.
+    assert_equal(i.operands[0].strides, (96, 8, 32))
+    a = a[::-1, :, ::-1]
+    i = nditer(a, [], [['readonly', 'copy']],
+            casting='safe',
+            op_dtypes=[np.dtype('f8')])
+    assert_equal(i.operands[0], a)
+    assert_equal(i.operands[0].dtype, np.dtype('f8'))
+    assert_equal(i.operands[0].strides, (96, 8, 32))
+
+    # Same-kind cast 'f8' -> 'f4' -> 'f8'
+    a = np.arange(24, dtype='f8').reshape(2, 3, 4).T
+    with nditer(a, [],
+            [['readwrite', 'updateifcopy']],
+            casting='same_kind',
+            op_dtypes=[np.dtype('f4')]) as i:
+        assert_equal(i.operands[0], a)
+        assert_equal(i.operands[0].dtype, np.dtype('f4'))
+        assert_equal(i.operands[0].strides, (4, 16, 48))
+        # Check that WRITEBACKIFCOPY is activated at exit
+        i.operands[0][2, 1, 1] = -12.5
+        assert_(a[2, 1, 1] != -12.5)
+    assert_equal(a[2, 1, 1], -12.5)
+
+    a = np.arange(6, dtype='i4')[::-2]
+    with nditer(a, [],
+            [['writeonly', 'updateifcopy']],
+            casting='unsafe',
+            op_dtypes=[np.dtype('f4')]) as i:
+        assert_equal(i.operands[0].dtype, np.dtype('f4'))
+        # Even though the stride was negative in 'a', it
+        # becomes positive in the temporary
+        assert_equal(i.operands[0].strides, (4,))
+        i.operands[0][:] = [1, 2, 3]
+    assert_equal(a, [1, 2, 3])
+
+def test_iter_array_cast_errors():
+    # Check that invalid casts are caught
+
+    # Need to enable copying for casts to occur
+    assert_raises(TypeError, nditer, arange(2, dtype='f4'), [],
+                [['readonly']], op_dtypes=[np.dtype('f8')])
+    # Also need to allow casting for casts to occur
+    assert_raises(TypeError, nditer, arange(2, dtype='f4'), [],
+                [['readonly', 'copy']], casting='no',
+                op_dtypes=[np.dtype('f8')])
+    assert_raises(TypeError, nditer, arange(2, dtype='f4'), [],
+                [['readonly', 'copy']], casting='equiv',
+                op_dtypes=[np.dtype('f8')])
+    assert_raises(TypeError, nditer, arange(2, dtype='f8'), [],
+                [['writeonly', 'updateifcopy']],
+                casting='no',
+                op_dtypes=[np.dtype('f4')])
+    assert_raises(TypeError, nditer, arange(2, dtype='f8'), [],
+                [['writeonly', 'updateifcopy']],
+                casting='equiv',
+                op_dtypes=[np.dtype('f4')])
+    # '<f4' -> '>f4' should not work with casting='no'
+    assert_raises(TypeError, nditer, arange(2, dtype='<f4'), [],
+                [['readonly', 'copy']], casting='no',
+                op_dtypes=[np.dtype('>f4')])
+    # 'f4' -> 'f8' is a safe cast, but 'f8' -> 'f4' isn't
+    assert_raises(TypeError, nditer, arange(2, dtype='f4'), [],
+                [['readwrite', 'updateifcopy']],
+                casting='safe',
+                op_dtypes=[np.dtype('f8')])
+    assert_raises(TypeError, nditer, arange(2, dtype='f8'), [],
+                [['readwrite', 'updateifcopy']],
+                casting='safe',
+                op_dtypes=[np.dtype('f4')])
+    # 'f4' -> 'i4' is neither a safe nor a same-kind cast
+    assert_raises(TypeError, nditer, arange(2, dtype='f4'), [],
+                [['readonly', 'copy']],
+                casting='same_kind',
+                op_dtypes=[np.dtype('i4')])
+    assert_raises(TypeError, nditer, arange(2, dtype='i4'), [],
+                [['writeonly', 'updateifcopy']],
+                casting='same_kind',
+                op_dtypes=[np.dtype('f4')])
+
+def test_iter_scalar_cast():
+    # Check that scalars are cast as requested
+
+    # No cast 'f4' -> 'f4'
+    i = nditer(np.float32(2.5), [], [['readonly']],
+                    op_dtypes=[np.dtype('f4')])
+    assert_equal(i.dtypes[0], np.dtype('f4'))
+    assert_equal(i.value.dtype, np.dtype('f4'))
+    assert_equal(i.value, 2.5)
+    # Safe cast 'f4' -> 'f8'
+    i = nditer(np.float32(2.5), [],
+                    [['readonly', 'copy']],
+                    casting='safe',
+                    op_dtypes=[np.dtype('f8')])
+    assert_equal(i.dtypes[0], np.dtype('f8'))
+    assert_equal(i.value.dtype, np.dtype('f8'))
+    assert_equal(i.value, 2.5)
+    # Same-kind cast 'f8' -> 'f4'
+    i = nditer(np.float64(2.5), [],
+                    [['readonly', 'copy']],
+                    casting='same_kind',
+                    op_dtypes=[np.dtype('f4')])
+    assert_equal(i.dtypes[0], np.dtype('f4'))
+    assert_equal(i.value.dtype, np.dtype('f4'))
+    assert_equal(i.value, 2.5)
+    # Unsafe cast 'f8' -> 'i4'
+    i = nditer(np.float64(3.0), [],
+                    [['readonly', 'copy']],
+                    casting='unsafe',
+                    op_dtypes=[np.dtype('i4')])
+    assert_equal(i.dtypes[0], np.dtype('i4'))
+    assert_equal(i.value.dtype, np.dtype('i4'))
+    assert_equal(i.value, 3)
+    # Readonly scalars may be cast even without setting COPY or BUFFERED
+    i = nditer(3, [], [['readonly']], op_dtypes=[np.dtype('f8')])
+    assert_equal(i[0].dtype, np.dtype('f8'))
+    assert_equal(i[0], 3.)
+
+def test_iter_scalar_cast_errors():
+    # Check that invalid casts are caught
+
+    # Need to allow copying/buffering for write casts of scalars to occur
+    assert_raises(TypeError, nditer, np.float32(2), [],
+                [['readwrite']], op_dtypes=[np.dtype('f8')])
+    assert_raises(TypeError, nditer, 2.5, [],
+                [['readwrite']], op_dtypes=[np.dtype('f4')])
+    # 'f8' -> 'f4' isn't a safe cast if the value would overflow
+    assert_raises(TypeError, nditer, np.float64(1e60), [],
+                [['readonly']],
+                casting='safe',
+                op_dtypes=[np.dtype('f4')])
+    # 'f4' -> 'i4' is neither a safe nor a same-kind cast
+    assert_raises(TypeError, nditer, np.float32(2), [],
+                [['readonly']],
+                casting='same_kind',
+                op_dtypes=[np.dtype('i4')])
+
+def test_iter_object_arrays_basic():
+    # Check that object arrays work
+
+    obj = {'a': 3, 'b': 'd'}
+    a = np.array([[1, 2, 3], None, obj, None], dtype='O')
+    if HAS_REFCOUNT:
+        rc = sys.getrefcount(obj)
+
+    # Need to allow references for object arrays
+    assert_raises(TypeError, nditer, a)
+    if HAS_REFCOUNT:
+        assert_equal(sys.getrefcount(obj), rc)
+
+    i = nditer(a, ['refs_ok'], ['readonly'])
+    vals = [x_[()] for x_ in i]
+    assert_equal(np.array(vals, dtype='O'), a)
+    vals, i, x = [None] * 3
+    if HAS_REFCOUNT:
+        assert_equal(sys.getrefcount(obj), rc)
+
+    i = nditer(a.reshape(2, 2).T, ['refs_ok', 'buffered'],
+                        ['readonly'], order='C')
+    assert_(i.iterationneedsapi)
+    vals = [x_[()] for x_ in i]
+    assert_equal(np.array(vals, dtype='O'), a.reshape(2, 2).ravel(order='F'))
+    vals, i, x = [None] * 3
+    if HAS_REFCOUNT:
+        assert_equal(sys.getrefcount(obj), rc)
+
+    i = nditer(a.reshape(2, 2).T, ['refs_ok', 'buffered'],
+                        ['readwrite'], order='C')
+    with i:
+        for x in i:
+            x[...] = None
+        vals, i, x = [None] * 3
+    if HAS_REFCOUNT:
+        assert_(sys.getrefcount(obj) == rc - 1)
+    assert_equal(a, np.array([None] * 4, dtype='O'))
+
+def test_iter_object_arrays_conversions():
+    # Conversions to/from objects
+    a = np.arange(6, dtype='O')
+    i = nditer(a, ['refs_ok', 'buffered'], ['readwrite'],
+                    casting='unsafe', op_dtypes='i4')
+    with i:
+        for x in i:
+            x[...] += 1
+    assert_equal(a, np.arange(6) + 1)
+
+    a = np.arange(6, dtype='i4')
+    i = nditer(a, ['refs_ok', 'buffered'], ['readwrite'],
+                    casting='unsafe', op_dtypes='O')
+    with i:
+        for x in i:
+            x[...] += 1
+    assert_equal(a, np.arange(6) + 1)
+
+    # Non-contiguous object array
+    a = np.zeros((6,), dtype=[('p', 'i1'), ('a', 'O')])
+    a = a['a']
+    a[:] = np.arange(6)
+    i = nditer(a, ['refs_ok', 'buffered'], ['readwrite'],
+                    casting='unsafe', op_dtypes='i4')
+    with i:
+        for x in i:
+            x[...] += 1
+    assert_equal(a, np.arange(6) + 1)
+
+    # Non-contiguous value array
+    a = np.zeros((6,), dtype=[('p', 'i1'), ('a', 'i4')])
+    a = a['a']
+    a[:] = np.arange(6) + 98172488
+    i = nditer(a, ['refs_ok', 'buffered'], ['readwrite'],
+                    casting='unsafe', op_dtypes='O')
+    with i:
+        ob = i[0][()]
+        if HAS_REFCOUNT:
+            rc = sys.getrefcount(ob)
+        for x in i:
+            x[...] += 1
+        if HAS_REFCOUNT:
+            newrc = sys.getrefcount(ob)
+            assert_(newrc == rc - 1)
+    assert_equal(a, np.arange(6) + 98172489)
+
+def test_iter_common_dtype():
+    # Check that the iterator finds a common data type correctly
+    # (some checks are somewhat duplicate after adopting NEP 50)
+
+    i = nditer([array([3], dtype='f4'), array([0], dtype='f8')],
+                    ['common_dtype'],
+                    [['readonly', 'copy']] * 2,
+                    casting='safe')
+    assert_equal(i.dtypes[0], np.dtype('f8'))
+    assert_equal(i.dtypes[1], np.dtype('f8'))
+    i = nditer([array([3], dtype='i4'), array([0], dtype='f4')],
+                    ['common_dtype'],
+                    [['readonly', 'copy']] * 2,
+                    casting='safe')
+    assert_equal(i.dtypes[0], np.dtype('f8'))
+    assert_equal(i.dtypes[1], np.dtype('f8'))
+    i = nditer([array([3], dtype='f4'), array(0, dtype='f8')],
+                    ['common_dtype'],
+                    [['readonly', 'copy']] * 2,
+                    casting='same_kind')
+    assert_equal(i.dtypes[0], np.dtype('f8'))
+    assert_equal(i.dtypes[1], np.dtype('f8'))
+    i = nditer([array([3], dtype='u4'), array(0, dtype='i4')],
+                    ['common_dtype'],
+                    [['readonly', 'copy']] * 2,
+                    casting='safe')
+    assert_equal(i.dtypes[0], np.dtype('i8'))
+    assert_equal(i.dtypes[1], np.dtype('i8'))
+    i = nditer([array([3], dtype='u4'), array(-12, dtype='i4')],
+                    ['common_dtype'],
+                    [['readonly', 'copy']] * 2,
+                    casting='safe')
+    assert_equal(i.dtypes[0], np.dtype('i8'))
+    assert_equal(i.dtypes[1], np.dtype('i8'))
+    i = nditer([array([3], dtype='u4'), array(-12, dtype='i4'),
+                 array([2j], dtype='c8'), array([9], dtype='f8')],
+                    ['common_dtype'],
+                    [['readonly', 'copy']] * 4,
+                    casting='safe')
+    assert_equal(i.dtypes[0], np.dtype('c16'))
+    assert_equal(i.dtypes[1], np.dtype('c16'))
+    assert_equal(i.dtypes[2], np.dtype('c16'))
+    assert_equal(i.dtypes[3], np.dtype('c16'))
+    assert_equal(i.value, (3, -12, 2j, 9))
+
+    # When allocating outputs, other outputs aren't factored in
+    i = nditer([array([3], dtype='i4'), None, array([2j], dtype='c16')], [],
+                    [['readonly', 'copy'],
+                     ['writeonly', 'allocate'],
+                     ['writeonly']],
+                    casting='safe')
+    assert_equal(i.dtypes[0], np.dtype('i4'))
+    assert_equal(i.dtypes[1], np.dtype('i4'))
+    assert_equal(i.dtypes[2], np.dtype('c16'))
+    # But, if common data types are requested, they are
+    i = nditer([array([3], dtype='i4'), None, array([2j], dtype='c16')],
+                    ['common_dtype'],
+                    [['readonly', 'copy'],
+                     ['writeonly', 'allocate'],
+                     ['writeonly']],
+                    casting='safe')
+    assert_equal(i.dtypes[0], np.dtype('c16'))
+    assert_equal(i.dtypes[1], np.dtype('c16'))
+    assert_equal(i.dtypes[2], np.dtype('c16'))
+
+def test_iter_copy_if_overlap():
+    # Ensure the iterator makes copies on read/write overlap, if requested
+
+    # Copy not needed, 1 op
+    for flag in ['readonly', 'writeonly', 'readwrite']:
+        a = arange(10)
+        i = nditer([a], ['copy_if_overlap'], [[flag]])
+        with i:
+            assert_(i.operands[0] is a)
+
+    # Copy needed, 2 ops, read-write overlap
+    x = arange(10)
+    a = x[1:]
+    b = x[:-1]
+    with nditer([a, b], ['copy_if_overlap'], [['readonly'], ['readwrite']]) as i:
+        assert_(not np.shares_memory(*i.operands))
+
+    # Copy not needed with elementwise, 2 ops, exactly same arrays
+    x = arange(10)
+    a = x
+    b = x
+    i = nditer([a, b], ['copy_if_overlap'], [['readonly', 'overlap_assume_elementwise'],
+                                             ['readwrite', 'overlap_assume_elementwise']])
+    with i:
+        assert_(i.operands[0] is a and i.operands[1] is b)
+    with nditer([a, b], ['copy_if_overlap'], [['readonly'], ['readwrite']]) as i:
+        assert_(i.operands[0] is a and not np.shares_memory(i.operands[1], b))
+
+    # Copy not needed, 2 ops, no overlap
+    x = arange(10)
+    a = x[::2]
+    b = x[1::2]
+    i = nditer([a, b], ['copy_if_overlap'], [['readonly'], ['writeonly']])
+    assert_(i.operands[0] is a and i.operands[1] is b)
+
+    # Copy needed, 2 ops, read-write overlap
+    x = arange(4, dtype=np.int8)
+    a = x[3:]
+    b = x.view(np.int32)[:1]
+    with nditer([a, b], ['copy_if_overlap'], [['readonly'], ['writeonly']]) as i:
+        assert_(not np.shares_memory(*i.operands))
+
+    # Copy needed, 3 ops, read-write overlap
+    for flag in ['writeonly', 'readwrite']:
+        x = np.ones([10, 10])
+        a = x
+        b = x.T
+        c = x
+        with nditer([a, b, c], ['copy_if_overlap'],
+                   [['readonly'], ['readonly'], [flag]]) as i:
+            a2, b2, c2 = i.operands
+            assert_(not np.shares_memory(a2, c2))
+            assert_(not np.shares_memory(b2, c2))
+
+    # Copy not needed, 3 ops, read-only overlap
+    x = np.ones([10, 10])
+    a = x
+    b = x.T
+    c = x
+    i = nditer([a, b, c], ['copy_if_overlap'],
+               [['readonly'], ['readonly'], ['readonly']])
+    a2, b2, c2 = i.operands
+    assert_(a is a2)
+    assert_(b is b2)
+    assert_(c is c2)
+
+    # Copy not needed, 3 ops, read-only overlap
+    x = np.ones([10, 10])
+    a = x
+    b = np.ones([10, 10])
+    c = x.T
+    i = nditer([a, b, c], ['copy_if_overlap'],
+               [['readonly'], ['writeonly'], ['readonly']])
+    a2, b2, c2 = i.operands
+    assert_(a is a2)
+    assert_(b is b2)
+    assert_(c is c2)
+
+    # Copy not needed, 3 ops, write-only overlap
+    x = np.arange(7)
+    a = x[:3]
+    b = x[3:6]
+    c = x[4:7]
+    i = nditer([a, b, c], ['copy_if_overlap'],
+               [['readonly'], ['writeonly'], ['writeonly']])
+    a2, b2, c2 = i.operands
+    assert_(a is a2)
+    assert_(b is b2)
+    assert_(c is c2)
+
+def test_iter_op_axes():
+    # Check that custom axes work
+
+    # Reverse the axes
+    a = arange(6).reshape(2, 3)
+    i = nditer([a, a.T], [], [['readonly']] * 2, op_axes=[[0, 1], [1, 0]])
+    assert_(all([x == y for (x, y) in i]))
+    a = arange(24).reshape(2, 3, 4)
+    i = nditer([a.T, a], [], [['readonly']] * 2, op_axes=[[2, 1, 0], None])
+    assert_(all([x == y for (x, y) in i]))
+
+    # Broadcast 1D to any dimension
+    a = arange(1, 31).reshape(2, 3, 5)
+    b = arange(1, 3)
+    i = nditer([a, b], [], [['readonly']] * 2, op_axes=[None, [0, -1, -1]])
+    assert_equal([x * y for (x, y) in i], (a * b.reshape(2, 1, 1)).ravel())
+    b = arange(1, 4)
+    i = nditer([a, b], [], [['readonly']] * 2, op_axes=[None, [-1, 0, -1]])
+    assert_equal([x * y for (x, y) in i], (a * b.reshape(1, 3, 1)).ravel())
+    b = arange(1, 6)
+    i = nditer([a, b], [], [['readonly']] * 2,
+                            op_axes=[None, [np.newaxis, np.newaxis, 0]])
+    assert_equal([x * y for (x, y) in i], (a * b.reshape(1, 1, 5)).ravel())
+
+    # Inner product-style broadcasting
+    a = arange(24).reshape(2, 3, 4)
+    b = arange(40).reshape(5, 2, 4)
+    i = nditer([a, b], ['multi_index'], [['readonly']] * 2,
+                            op_axes=[[0, 1, -1, -1], [-1, -1, 0, 1]])
+    assert_equal(i.shape, (2, 3, 5, 2))
+
+    # Matrix product-style broadcasting
+    a = arange(12).reshape(3, 4)
+    b = arange(20).reshape(4, 5)
+    i = nditer([a, b], ['multi_index'], [['readonly']] * 2,
+                            op_axes=[[0, -1], [-1, 1]])
+    assert_equal(i.shape, (3, 5))
+
+def test_iter_op_axes_errors():
+    # Check that custom axes throws errors for bad inputs
+
+    # Wrong number of items in op_axes
+    a = arange(6).reshape(2, 3)
+    assert_raises(ValueError, nditer, [a, a], [], [['readonly']] * 2,
+                                    op_axes=[[0], [1], [0]])
+    # Out of bounds items in op_axes
+    assert_raises(ValueError, nditer, [a, a], [], [['readonly']] * 2,
+                                    op_axes=[[2, 1], [0, 1]])
+    assert_raises(ValueError, nditer, [a, a], [], [['readonly']] * 2,
+                                    op_axes=[[0, 1], [2, -1]])
+    # Duplicate items in op_axes
+    assert_raises(ValueError, nditer, [a, a], [], [['readonly']] * 2,
+                                    op_axes=[[0, 0], [0, 1]])
+    assert_raises(ValueError, nditer, [a, a], [], [['readonly']] * 2,
+                                    op_axes=[[0, 1], [1, 1]])
+
+    # Different sized arrays in op_axes
+    assert_raises(ValueError, nditer, [a, a], [], [['readonly']] * 2,
+                                    op_axes=[[0, 1], [0, 1, 0]])
+
+    # Non-broadcastable dimensions in the result
+    assert_raises(ValueError, nditer, [a, a], [], [['readonly']] * 2,
+                                    op_axes=[[0, 1], [1, 0]])
+
+def test_iter_copy():
+    # Check that copying the iterator works correctly
+    a = arange(24).reshape(2, 3, 4)
+
+    # Simple iterator
+    i = nditer(a)
+    j = i.copy()
+    assert_equal([x[()] for x in i], [x[()] for x in j])
+
+    i.iterindex = 3
+    j = i.copy()
+    assert_equal([x[()] for x in i], [x[()] for x in j])
+
+    # Buffered iterator
+    i = nditer(a, ['buffered', 'ranged'], order='F', buffersize=3)
+    j = i.copy()
+    assert_equal([x[()] for x in i], [x[()] for x in j])
+
+    i.iterindex = 3
+    j = i.copy()
+    assert_equal([x[()] for x in i], [x[()] for x in j])
+
+    i.iterrange = (3, 9)
+    j = i.copy()
+    assert_equal([x[()] for x in i], [x[()] for x in j])
+
+    i.iterrange = (2, 18)
+    next(i)
+    next(i)
+    j = i.copy()
+    assert_equal([x[()] for x in i], [x[()] for x in j])
+
+    # Casting iterator
+    with nditer(a, ['buffered'], order='F', casting='unsafe',
+                op_dtypes='f8', buffersize=5) as i:
+        j = i.copy()
+    assert_equal([x[()] for x in j], a.ravel(order='F'))
+
+    a = arange(24, dtype='<i4').reshape(2, 3, 4)
+    with nditer(a, ['buffered'], order='F', casting='unsafe',
+                op_dtypes='>f8', buffersize=5) as i:
+        j = i.copy()
+    assert_equal([x[()] for x in j], a.ravel(order='F'))
+
+
+@pytest.mark.parametrize("dtype", np.typecodes["All"])
+@pytest.mark.parametrize("loop_dtype", np.typecodes["All"])
+@pytest.mark.filterwarnings("ignore::numpy.exceptions.ComplexWarning")
+def test_iter_copy_casts(dtype, loop_dtype):
+    # Ensure the dtype is never flexible:
+    if loop_dtype.lower() == "m":
+        loop_dtype = loop_dtype + "[ms]"
+    elif np.dtype(loop_dtype).itemsize == 0:
+        loop_dtype = loop_dtype + "50"
+
+    # Make things a bit more interesting by requiring a byte-swap as well:
+    arr = np.ones(1000, dtype=np.dtype(dtype).newbyteorder())
+    try:
+        expected = arr.astype(loop_dtype)
+    except Exception:
+        # Some casts are not possible, do not worry about them
+        return
+
+    it = np.nditer((arr,), ["buffered", "external_loop", "refs_ok"],
+                   op_dtypes=[loop_dtype], casting="unsafe")
+
+    if np.issubdtype(np.dtype(loop_dtype), np.number):
+        # Casting to strings may be strange, but for simple dtypes do not rely
+        # on the cast being correct:
+        assert_array_equal(expected, np.ones(1000, dtype=loop_dtype))
+
+    it_copy = it.copy()
+    res = next(it)
+    del it
+    res_copy = next(it_copy)
+    del it_copy
+
+    assert_array_equal(res, expected)
+    assert_array_equal(res_copy, expected)
+
+
+def test_iter_copy_casts_structured():
+    # Test a complicated structured dtype for casting, as it requires
+    # both multiple steps and a more complex casting setup.
+    # Includes a structured -> unstructured (any to object), and many other
+    # casts, which cause this to require all steps in the casting machinery
+    # one level down as well as the iterator copy (which uses NpyAuxData clone)
+    in_dtype = np.dtype([("a", np.dtype("i,")),
+                         ("b", np.dtype(">i,<i,>d,S17,>d,3f,O,i1"))])
+    out_dtype = np.dtype([("a", np.dtype("O")),
+                          ("b", np.dtype(">i,>i,S17,>d,>U3,3d,i1,O"))])
+    arr = np.ones(1000, dtype=in_dtype)
+
+    it = np.nditer((arr,), ["buffered", "external_loop", "refs_ok"],
+                   op_dtypes=[out_dtype], casting="unsafe")
+    it_copy = it.copy()
+
+    res1 = next(it)
+    del it
+    res2 = next(it_copy)
+    del it_copy
+
+    expected = arr["a"].astype(out_dtype["a"])
+    assert_array_equal(res1["a"], expected)
+    assert_array_equal(res2["a"], expected)
+
+    for field in in_dtype["b"].names:
+        # Note that the .base avoids the subarray field
+        expected = arr["b"][field].astype(out_dtype["b"][field].base)
+        assert_array_equal(res1["b"][field], expected)
+        assert_array_equal(res2["b"][field], expected)
+
+
+def test_iter_copy_casts_structured2():
+    # Similar to the above, this is a fairly arcane test to cover internals
+    in_dtype = np.dtype([("a", np.dtype("O,O")),
+                         ("b", np.dtype("5O,3O,(1,)O,(1,)i,(1,)O"))])
+    out_dtype = np.dtype([("a", np.dtype("O")),
+                          ("b", np.dtype("O,3i,4O,4O,4i"))])
+
+    arr = np.ones(1, dtype=in_dtype)
+    it = np.nditer((arr,), ["buffered", "external_loop", "refs_ok"],
+                   op_dtypes=[out_dtype], casting="unsafe")
+    it_copy = it.copy()
+
+    res1 = next(it)
+    del it
+    res2 = next(it_copy)
+    del it_copy
+
+    # Array of two structured scalars:
+    for res in res1, res2:
+        # Cast to tuple by getitem, which may be weird and changeable?:
+        assert isinstance(res["a"][0], tuple)
+        assert res["a"][0] == (1, 1)
+
+    for res in res1, res2:
+        assert_array_equal(res["b"]["f0"][0], np.ones(5, dtype=object))
+        assert_array_equal(res["b"]["f1"], np.ones((1, 3), dtype="i"))
+        assert res["b"]["f2"].shape == (1, 4)
+        assert_array_equal(res["b"]["f2"][0], np.ones(4, dtype=object))
+        assert_array_equal(res["b"]["f3"][0], np.ones(4, dtype=object))
+        assert_array_equal(res["b"]["f3"][0], np.ones(4, dtype="i"))
+
+
+def test_iter_allocate_output_simple():
+    # Check that the iterator will properly allocate outputs
+
+    # Simple case
+    a = arange(6)
+    i = nditer([a, None], [], [['readonly'], ['writeonly', 'allocate']],
+                        op_dtypes=[None, np.dtype('f4')])
+    assert_equal(i.operands[1].shape, a.shape)
+    assert_equal(i.operands[1].dtype, np.dtype('f4'))
+
+def test_iter_allocate_output_buffered_readwrite():
+    # Allocated output with buffering + delay_bufalloc
+
+    a = arange(6)
+    i = nditer([a, None], ['buffered', 'delay_bufalloc'],
+                        [['readonly'], ['allocate', 'readwrite']])
+    with i:
+        i.operands[1][:] = 1
+        i.reset()
+        for x in i:
+            x[1][...] += x[0][...]
+        assert_equal(i.operands[1], a + 1)
+
+def test_iter_allocate_output_itorder():
+    # The allocated output should match the iteration order
+
+    # C-order input, best iteration order
+    a = arange(6, dtype='i4').reshape(2, 3)
+    i = nditer([a, None], [], [['readonly'], ['writeonly', 'allocate']],
+                        op_dtypes=[None, np.dtype('f4')])
+    assert_equal(i.operands[1].shape, a.shape)
+    assert_equal(i.operands[1].strides, a.strides)
+    assert_equal(i.operands[1].dtype, np.dtype('f4'))
+    # F-order input, best iteration order
+    a = arange(24, dtype='i4').reshape(2, 3, 4).T
+    i = nditer([a, None], [], [['readonly'], ['writeonly', 'allocate']],
+                        op_dtypes=[None, np.dtype('f4')])
+    assert_equal(i.operands[1].shape, a.shape)
+    assert_equal(i.operands[1].strides, a.strides)
+    assert_equal(i.operands[1].dtype, np.dtype('f4'))
+    # Non-contiguous input, C iteration order
+    a = arange(24, dtype='i4').reshape(2, 3, 4).swapaxes(0, 1)
+    i = nditer([a, None], [],
+                        [['readonly'], ['writeonly', 'allocate']],
+                        order='C',
+                        op_dtypes=[None, np.dtype('f4')])
+    assert_equal(i.operands[1].shape, a.shape)
+    assert_equal(i.operands[1].strides, (32, 16, 4))
+    assert_equal(i.operands[1].dtype, np.dtype('f4'))
+
+def test_iter_allocate_output_opaxes():
+    # Specifying op_axes should work
+
+    a = arange(24, dtype='i4').reshape(2, 3, 4)
+    i = nditer([None, a], [], [['writeonly', 'allocate'], ['readonly']],
+                        op_dtypes=[np.dtype('u4'), None],
+                        op_axes=[[1, 2, 0], None])
+    assert_equal(i.operands[0].shape, (4, 2, 3))
+    assert_equal(i.operands[0].strides, (4, 48, 16))
+    assert_equal(i.operands[0].dtype, np.dtype('u4'))
+
+def test_iter_allocate_output_types_promotion():
+    # Check type promotion of automatic outputs (this was more interesting
+    # before NEP 50...)
+
+    i = nditer([array([3], dtype='f4'), array([0], dtype='f8'), None], [],
+                    [['readonly']] * 2 + [['writeonly', 'allocate']])
+    assert_equal(i.dtypes[2], np.dtype('f8'))
+    i = nditer([array([3], dtype='i4'), array([0], dtype='f4'), None], [],
+                    [['readonly']] * 2 + [['writeonly', 'allocate']])
+    assert_equal(i.dtypes[2], np.dtype('f8'))
+    i = nditer([array([3], dtype='f4'), array(0, dtype='f8'), None], [],
+                    [['readonly']] * 2 + [['writeonly', 'allocate']])
+    assert_equal(i.dtypes[2], np.dtype('f8'))
+    i = nditer([array([3], dtype='u4'), array(0, dtype='i4'), None], [],
+                    [['readonly']] * 2 + [['writeonly', 'allocate']])
+    assert_equal(i.dtypes[2], np.dtype('i8'))
+    i = nditer([array([3], dtype='u4'), array(-12, dtype='i4'), None], [],
+                    [['readonly']] * 2 + [['writeonly', 'allocate']])
+    assert_equal(i.dtypes[2], np.dtype('i8'))
+
+def test_iter_allocate_output_types_byte_order():
+    # Verify the rules for byte order changes
+
+    # When there's just one input, the output type exactly matches
+    a = array([3], dtype='u4')
+    a = a.view(a.dtype.newbyteorder())
+    i = nditer([a, None], [],
+                    [['readonly'], ['writeonly', 'allocate']])
+    assert_equal(i.dtypes[0], i.dtypes[1])
+    # With two or more inputs, the output type is in native byte order
+    i = nditer([a, a, None], [],
+                    [['readonly'], ['readonly'], ['writeonly', 'allocate']])
+    assert_(i.dtypes[0] != i.dtypes[2])
+    assert_equal(i.dtypes[0].newbyteorder('='), i.dtypes[2])
+
+def test_iter_allocate_output_types_scalar():
+    # If the inputs are all scalars, the output should be a scalar
+
+    i = nditer([None, 1, 2.3, np.float32(12), np.complex128(3)], [],
+                [['writeonly', 'allocate']] + [['readonly']] * 4)
+    assert_equal(i.operands[0].dtype, np.dtype('complex128'))
+    assert_equal(i.operands[0].ndim, 0)
+
+def test_iter_allocate_output_subtype():
+    # Make sure that the subtype with priority wins
+    class MyNDArray(np.ndarray):
+        __array_priority__ = 15
+
+    # subclass vs ndarray
+    a = np.array([[1, 2], [3, 4]]).view(MyNDArray)
+    b = np.arange(4).reshape(2, 2).T
+    i = nditer([a, b, None], [],
+               [['readonly'], ['readonly'], ['writeonly', 'allocate']])
+    assert_equal(type(a), type(i.operands[2]))
+    assert_(type(b) is not type(i.operands[2]))
+    assert_equal(i.operands[2].shape, (2, 2))
+
+    # If subtypes are disabled, we should get back an ndarray.
+    i = nditer([a, b, None], [],
+               [['readonly'], ['readonly'],
+                ['writeonly', 'allocate', 'no_subtype']])
+    assert_equal(type(b), type(i.operands[2]))
+    assert_(type(a) is not type(i.operands[2]))
+    assert_equal(i.operands[2].shape, (2, 2))
+
+def test_iter_allocate_output_errors():
+    # Check that the iterator will throw errors for bad output allocations
+
+    # Need an input if no output data type is specified
+    a = arange(6)
+    assert_raises(TypeError, nditer, [a, None], [],
+                        [['writeonly'], ['writeonly', 'allocate']])
+    # Allocated output should be flagged for writing
+    assert_raises(ValueError, nditer, [a, None], [],
+                        [['readonly'], ['allocate', 'readonly']])
+    # Allocated output can't have buffering without delayed bufalloc
+    assert_raises(ValueError, nditer, [a, None], ['buffered'],
+                                            ['allocate', 'readwrite'])
+    # Must specify dtype if there are no inputs (cannot promote existing ones;
+    # maybe this should use the 'f4' here, but it does not historically.)
+    assert_raises(TypeError, nditer, [None, None], [],
+                        [['writeonly', 'allocate'],
+                         ['writeonly', 'allocate']],
+                        op_dtypes=[None, np.dtype('f4')])
+    # If using op_axes, must specify all the axes
+    a = arange(24, dtype='i4').reshape(2, 3, 4)
+    assert_raises(ValueError, nditer, [a, None], [],
+                        [['readonly'], ['writeonly', 'allocate']],
+                        op_dtypes=[None, np.dtype('f4')],
+                        op_axes=[None, [0, np.newaxis, 1]])
+    # If using op_axes, the axes must be within bounds
+    assert_raises(ValueError, nditer, [a, None], [],
+                        [['readonly'], ['writeonly', 'allocate']],
+                        op_dtypes=[None, np.dtype('f4')],
+                        op_axes=[None, [0, 3, 1]])
+    # If using op_axes, there can't be duplicates
+    assert_raises(ValueError, nditer, [a, None], [],
+                        [['readonly'], ['writeonly', 'allocate']],
+                        op_dtypes=[None, np.dtype('f4')],
+                        op_axes=[None, [0, 2, 1, 0]])
+    # Not all axes may be specified if a reduction. If there is a hole
+    # in op_axes, this is an error.
+    a = arange(24, dtype='i4').reshape(2, 3, 4)
+    assert_raises(ValueError, nditer, [a, None], ["reduce_ok"],
+                        [['readonly'], ['readwrite', 'allocate']],
+                        op_dtypes=[None, np.dtype('f4')],
+                        op_axes=[None, [0, np.newaxis, 2]])
+
+def test_all_allocated():
+    # When no output and no shape is given, `()` is used as shape.
+    i = np.nditer([None], op_dtypes=["int64"])
+    assert i.operands[0].shape == ()
+    assert i.dtypes == (np.dtype("int64"),)
+
+    i = np.nditer([None], op_dtypes=["int64"], itershape=(2, 3, 4))
+    assert i.operands[0].shape == (2, 3, 4)
+
+def test_iter_remove_axis():
+    a = arange(24).reshape(2, 3, 4)
+
+    i = nditer(a, ['multi_index'])
+    i.remove_axis(1)
+    assert_equal(list(i), a[:, 0, :].ravel())
+
+    a = a[::-1, :, :]
+    i = nditer(a, ['multi_index'])
+    i.remove_axis(0)
+    assert_equal(list(i), a[0, :, :].ravel())
+
+def test_iter_remove_multi_index_inner_loop():
+    # Check that removing multi-index support works
+
+    a = arange(24).reshape(2, 3, 4)
+
+    i = nditer(a, ['multi_index'])
+    assert_equal(i.ndim, 3)
+    assert_equal(i.shape, (2, 3, 4))
+    assert_equal(i.itviews[0].shape, (2, 3, 4))
+
+    # Removing the multi-index tracking causes all dimensions to coalesce
+    before = list(i)
+    i.remove_multi_index()
+    after = list(i)
+
+    assert_equal(before, after)
+    assert_equal(i.ndim, 1)
+    assert_raises(ValueError, lambda i: i.shape, i)
+    assert_equal(i.itviews[0].shape, (24,))
+
+    # Removing the inner loop means there's just one iteration
+    i.reset()
+    assert_equal(i.itersize, 24)
+    assert_equal(i[0].shape, ())
+    i.enable_external_loop()
+    assert_equal(i.itersize, 24)
+    assert_equal(i[0].shape, (24,))
+    assert_equal(i.value, arange(24))
+
+def test_iter_iterindex():
+    # Make sure iterindex works
+
+    buffersize = 5
+    a = arange(24).reshape(4, 3, 2)
+    for flags in ([], ['buffered']):
+        i = nditer(a, flags, buffersize=buffersize)
+        assert_equal(iter_iterindices(i), list(range(24)))
+        i.iterindex = 2
+        assert_equal(iter_iterindices(i), list(range(2, 24)))
+
+        i = nditer(a, flags, order='F', buffersize=buffersize)
+        assert_equal(iter_iterindices(i), list(range(24)))
+        i.iterindex = 5
+        assert_equal(iter_iterindices(i), list(range(5, 24)))
+
+        i = nditer(a[::-1], flags, order='F', buffersize=buffersize)
+        assert_equal(iter_iterindices(i), list(range(24)))
+        i.iterindex = 9
+        assert_equal(iter_iterindices(i), list(range(9, 24)))
+
+        i = nditer(a[::-1, ::-1], flags, order='C', buffersize=buffersize)
+        assert_equal(iter_iterindices(i), list(range(24)))
+        i.iterindex = 13
+        assert_equal(iter_iterindices(i), list(range(13, 24)))
+
+        i = nditer(a[::1, ::-1], flags, buffersize=buffersize)
+        assert_equal(iter_iterindices(i), list(range(24)))
+        i.iterindex = 23
+        assert_equal(iter_iterindices(i), list(range(23, 24)))
+        i.reset()
+        i.iterindex = 2
+        assert_equal(iter_iterindices(i), list(range(2, 24)))
+
+def test_iter_iterrange():
+    # Make sure getting and resetting the iterrange works
+
+    buffersize = 5
+    a = arange(24, dtype='i4').reshape(4, 3, 2)
+    a_fort = a.ravel(order='F')
+
+    i = nditer(a, ['ranged'], ['readonly'], order='F',
+                buffersize=buffersize)
+    assert_equal(i.iterrange, (0, 24))
+    assert_equal([x[()] for x in i], a_fort)
+    for r in [(0, 24), (1, 2), (3, 24), (5, 5), (0, 20), (23, 24)]:
+        i.iterrange = r
+        assert_equal(i.iterrange, r)
+        assert_equal([x[()] for x in i], a_fort[r[0]:r[1]])
+
+    i = nditer(a, ['ranged', 'buffered'], ['readonly'], order='F',
+                op_dtypes='f8', buffersize=buffersize)
+    assert_equal(i.iterrange, (0, 24))
+    assert_equal([x[()] for x in i], a_fort)
+    for r in [(0, 24), (1, 2), (3, 24), (5, 5), (0, 20), (23, 24)]:
+        i.iterrange = r
+        assert_equal(i.iterrange, r)
+        assert_equal([x[()] for x in i], a_fort[r[0]:r[1]])
+
+    def get_array(i):
+        val = np.array([], dtype='f8')
+        for x in i:
+            val = np.concatenate((val, x))
+        return val
+
+    i = nditer(a, ['ranged', 'buffered', 'external_loop'],
+                ['readonly'], order='F',
+                op_dtypes='f8', buffersize=buffersize)
+    assert_equal(i.iterrange, (0, 24))
+    assert_equal(get_array(i), a_fort)
+    for r in [(0, 24), (1, 2), (3, 24), (5, 5), (0, 20), (23, 24)]:
+        i.iterrange = r
+        assert_equal(i.iterrange, r)
+        assert_equal(get_array(i), a_fort[r[0]:r[1]])
+
+def test_iter_buffering():
+    # Test buffering with several buffer sizes and types
+    arrays = []
+    # F-order swapped array
+    _tmp = np.arange(24, dtype='c16').reshape(2, 3, 4).T
+    _tmp = _tmp.view(_tmp.dtype.newbyteorder()).byteswap()
+    arrays.append(_tmp)
+    # Contiguous 1-dimensional array
+    arrays.append(np.arange(10, dtype='f4'))
+    # Unaligned array
+    a = np.zeros((4 * 16 + 1,), dtype='i1')[1:]
+    a = a.view('i4')
+    a[:] = np.arange(16, dtype='i4')
+    arrays.append(a)
+    # 4-D F-order array
+    arrays.append(np.arange(120, dtype='i4').reshape(5, 3, 2, 4).T)
+    for a in arrays:
+        for buffersize in (1, 2, 3, 5, 8, 11, 16, 1024):
+            vals = []
+            i = nditer(a, ['buffered', 'external_loop'],
+                           [['readonly', 'nbo', 'aligned']],
+                           order='C',
+                           casting='equiv',
+                           buffersize=buffersize)
+            while not i.finished:
+                assert_(i[0].size <= buffersize)
+                vals.append(i[0].copy())
+                i.iternext()
+            assert_equal(np.concatenate(vals), a.ravel(order='C'))
+
+def test_iter_write_buffering():
+    # Test that buffering of writes is working
+
+    # F-order swapped array
+    a = np.arange(24).reshape(2, 3, 4).T
+    a = a.view(a.dtype.newbyteorder()).byteswap()
+    i = nditer(a, ['buffered'],
+                   [['readwrite', 'nbo', 'aligned']],
+                   casting='equiv',
+                   order='C',
+                   buffersize=16)
+    x = 0
+    with i:
+        while not i.finished:
+            i[0] = x
+            x += 1
+            i.iternext()
+    assert_equal(a.ravel(order='C'), np.arange(24))
+
+def test_iter_buffering_delayed_alloc():
+    # Test that delaying buffer allocation works
+
+    a = np.arange(6)
+    b = np.arange(1, dtype='f4')
+    i = nditer([a, b], ['buffered', 'delay_bufalloc', 'multi_index', 'reduce_ok'],
+                    ['readwrite'],
+                    casting='unsafe',
+                    op_dtypes='f4')
+    assert_(i.has_delayed_bufalloc)
+    assert_raises(ValueError, lambda i: i.multi_index, i)
+    assert_raises(ValueError, lambda i: i[0], i)
+    assert_raises(ValueError, lambda i: i[0:2], i)
+
+    def assign_iter(i):
+        i[0] = 0
+    assert_raises(ValueError, assign_iter, i)
+
+    i.reset()
+    assert_(not i.has_delayed_bufalloc)
+    assert_equal(i.multi_index, (0,))
+    with i:
+        assert_equal(i[0], 0)
+        i[1] = 1
+        assert_equal(i[0:2], [0, 1])
+        assert_equal([[x[0][()], x[1][()]] for x in i], list(zip(range(6), [1] * 6)))
+
+def test_iter_buffered_cast_simple():
+    # Test that buffering can handle a simple cast
+
+    a = np.arange(10, dtype='f4')
+    i = nditer(a, ['buffered', 'external_loop'],
+                   [['readwrite', 'nbo', 'aligned']],
+                   casting='same_kind',
+                   op_dtypes=[np.dtype('f8')],
+                   buffersize=3)
+    with i:
+        for v in i:
+            v[...] *= 2
+
+    assert_equal(a, 2 * np.arange(10, dtype='f4'))
+
+def test_iter_buffered_cast_byteswapped():
+    # Test that buffering can handle a cast which requires swap->cast->swap
+
+    a = np.arange(10, dtype='f4')
+    a = a.view(a.dtype.newbyteorder()).byteswap()
+    i = nditer(a, ['buffered', 'external_loop'],
+                   [['readwrite', 'nbo', 'aligned']],
+                   casting='same_kind',
+                   op_dtypes=[np.dtype('f8').newbyteorder()],
+                   buffersize=3)
+    with i:
+        for v in i:
+            v[...] *= 2
+
+    assert_equal(a, 2 * np.arange(10, dtype='f4'))
+
+    with warnings.catch_warnings():
+        warnings.simplefilter('ignore', np.exceptions.ComplexWarning)
+
+        a = np.arange(10, dtype='f8')
+        a = a.view(a.dtype.newbyteorder()).byteswap()
+        i = nditer(a, ['buffered', 'external_loop'],
+                       [['readwrite', 'nbo', 'aligned']],
+                       casting='unsafe',
+                       op_dtypes=[np.dtype('c8').newbyteorder()],
+                       buffersize=3)
+        with i:
+            for v in i:
+                v[...] *= 2
+
+        assert_equal(a, 2 * np.arange(10, dtype='f8'))
+
+def test_iter_buffered_cast_byteswapped_complex():
+    # Test that buffering can handle a cast which requires swap->cast->copy
+
+    a = np.arange(10, dtype='c8')
+    a = a.view(a.dtype.newbyteorder()).byteswap()
+    a += 2j
+    i = nditer(a, ['buffered', 'external_loop'],
+                   [['readwrite', 'nbo', 'aligned']],
+                   casting='same_kind',
+                   op_dtypes=[np.dtype('c16')],
+                   buffersize=3)
+    with i:
+        for v in i:
+            v[...] *= 2
+    assert_equal(a, 2 * np.arange(10, dtype='c8') + 4j)
+
+    a = np.arange(10, dtype='c8')
+    a += 2j
+    i = nditer(a, ['buffered', 'external_loop'],
+                   [['readwrite', 'nbo', 'aligned']],
+                   casting='same_kind',
+                   op_dtypes=[np.dtype('c16').newbyteorder()],
+                   buffersize=3)
+    with i:
+        for v in i:
+            v[...] *= 2
+    assert_equal(a, 2 * np.arange(10, dtype='c8') + 4j)
+
+    a = np.arange(10, dtype=np.clongdouble)
+    a = a.view(a.dtype.newbyteorder()).byteswap()
+    a += 2j
+    i = nditer(a, ['buffered', 'external_loop'],
+                   [['readwrite', 'nbo', 'aligned']],
+                   casting='same_kind',
+                   op_dtypes=[np.dtype('c16')],
+                   buffersize=3)
+    with i:
+        for v in i:
+            v[...] *= 2
+    assert_equal(a, 2 * np.arange(10, dtype=np.clongdouble) + 4j)
+
+    a = np.arange(10, dtype=np.longdouble)
+    a = a.view(a.dtype.newbyteorder()).byteswap()
+    i = nditer(a, ['buffered', 'external_loop'],
+                   [['readwrite', 'nbo', 'aligned']],
+                   casting='same_kind',
+                   op_dtypes=[np.dtype('f4')],
+                   buffersize=7)
+    with i:
+        for v in i:
+            v[...] *= 2
+    assert_equal(a, 2 * np.arange(10, dtype=np.longdouble))
+
+def test_iter_buffered_cast_structured_type():
+    # Tests buffering of structured types
+
+    # simple -> struct type (duplicates the value)
+    sdt = [('a', 'f4'), ('b', 'i8'), ('c', 'c8', (2, 3)), ('d', 'O')]
+    a = np.arange(3, dtype='f4') + 0.5
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt)
+    vals = [np.array(x) for x in i]
+    assert_equal(vals[0]['a'], 0.5)
+    assert_equal(vals[0]['b'], 0)
+    assert_equal(vals[0]['c'], [[(0.5)] * 3] * 2)
+    assert_equal(vals[0]['d'], 0.5)
+    assert_equal(vals[1]['a'], 1.5)
+    assert_equal(vals[1]['b'], 1)
+    assert_equal(vals[1]['c'], [[(1.5)] * 3] * 2)
+    assert_equal(vals[1]['d'], 1.5)
+    assert_equal(vals[0].dtype, np.dtype(sdt))
+
+    # object -> struct type
+    sdt = [('a', 'f4'), ('b', 'i8'), ('c', 'c8', (2, 3)), ('d', 'O')]
+    a = np.zeros((3,), dtype='O')
+    a[0] = (0.5, 0.5, [[0.5, 0.5, 0.5], [0.5, 0.5, 0.5]], 0.5)
+    a[1] = (1.5, 1.5, [[1.5, 1.5, 1.5], [1.5, 1.5, 1.5]], 1.5)
+    a[2] = (2.5, 2.5, [[2.5, 2.5, 2.5], [2.5, 2.5, 2.5]], 2.5)
+    if HAS_REFCOUNT:
+        rc = sys.getrefcount(a[0])
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt)
+    vals = [x.copy() for x in i]
+    assert_equal(vals[0]['a'], 0.5)
+    assert_equal(vals[0]['b'], 0)
+    assert_equal(vals[0]['c'], [[(0.5)] * 3] * 2)
+    assert_equal(vals[0]['d'], 0.5)
+    assert_equal(vals[1]['a'], 1.5)
+    assert_equal(vals[1]['b'], 1)
+    assert_equal(vals[1]['c'], [[(1.5)] * 3] * 2)
+    assert_equal(vals[1]['d'], 1.5)
+    assert_equal(vals[0].dtype, np.dtype(sdt))
+    vals, i, x = [None] * 3
+    if HAS_REFCOUNT:
+        assert_equal(sys.getrefcount(a[0]), rc)
+
+    # single-field struct type -> simple
+    sdt = [('a', 'f4')]
+    a = np.array([(5.5,), (8,)], dtype=sdt)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes='i4')
+    assert_equal([x_[()] for x_ in i], [5, 8])
+
+    # make sure multi-field struct type -> simple doesn't work
+    sdt = [('a', 'f4'), ('b', 'i8'), ('d', 'O')]
+    a = np.array([(5.5, 7, 'test'), (8, 10, 11)], dtype=sdt)
+    assert_raises(TypeError, lambda: (
+        nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+               casting='unsafe',
+               op_dtypes='i4')))
+
+    # struct type -> struct type (field-wise copy)
+    sdt1 = [('a', 'f4'), ('b', 'i8'), ('d', 'O')]
+    sdt2 = [('d', 'u2'), ('a', 'O'), ('b', 'f8')]
+    a = np.array([(1, 2, 3), (4, 5, 6)], dtype=sdt1)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    assert_equal(i[0].dtype, np.dtype(sdt2))
+    assert_equal([np.array(x_) for x_ in i],
+                 [np.array((1, 2, 3), dtype=sdt2),
+                  np.array((4, 5, 6), dtype=sdt2)])
+
+
+def test_iter_buffered_cast_structured_type_failure_with_cleanup():
+    # make sure struct type -> struct type with different
+    # number of fields fails
+    sdt1 = [('a', 'f4'), ('b', 'i8'), ('d', 'O')]
+    sdt2 = [('b', 'O'), ('a', 'f8')]
+    a = np.array([(1, 2, 3), (4, 5, 6)], dtype=sdt1)
+
+    for intent in ["readwrite", "readonly", "writeonly"]:
+        # This test was initially designed to test an error at a different
+        # place, but will now raise earlier to to the cast not being possible:
+        # `assert np.can_cast(a.dtype, sdt2, casting="unsafe")` fails.
+        # Without a faulty DType, there is probably no reliable
+        # way to get the initial tested behaviour.
+        simple_arr = np.array([1, 2], dtype="i,i")  # requires clean up
+        with pytest.raises(TypeError):
+            nditer((simple_arr, a), ['buffered', 'refs_ok'], [intent, intent],
+                   casting='unsafe', op_dtypes=["f,f", sdt2])
+
+
+def test_buffered_cast_error_paths():
+    with pytest.raises(ValueError):
+        # The input is cast into an `S3` buffer
+        np.nditer((np.array("a", dtype="S1"),), op_dtypes=["i"],
+                  casting="unsafe", flags=["buffered"])
+
+    # The `M8[ns]` is cast into the `S3` output
+    it = np.nditer((np.array(1, dtype="i"),), op_dtypes=["S1"],
+                   op_flags=["writeonly"], casting="unsafe", flags=["buffered"])
+    with pytest.raises(ValueError):
+        with it:
+            buf = next(it)
+            buf[...] = "a"  # cannot be converted to int.
+
+@pytest.mark.skipif(IS_WASM, reason="Cannot start subprocess")
+@pytest.mark.skipif(not HAS_REFCOUNT, reason="PyPy seems to not hit this.")
+def test_buffered_cast_error_paths_unraisable():
+    # The following gives an unraisable error. Pytest sometimes captures that
+    # (depending python and/or pytest version). So with Python>=3.8 this can
+    # probably be cleaned out in the future to check for
+    # pytest.PytestUnraisableExceptionWarning:
+    code = textwrap.dedent("""
+        import numpy as np
+
+        it = np.nditer((np.array(1, dtype="i"),), op_dtypes=["S1"],
+                       op_flags=["writeonly"], casting="unsafe", flags=["buffered"])
+        buf = next(it)
+        buf[...] = "a"
+        del buf, it  # Flushing only happens during deallocate right now.
+        """)
+    res = subprocess.check_output([sys.executable, "-c", code],
+                                  stderr=subprocess.STDOUT, text=True)
+    assert "ValueError" in res
+
+
+def test_iter_buffered_cast_subarray():
+    # Tests buffering of subarrays
+
+    # one element -> many (copies it to all)
+    sdt1 = [('a', 'f4')]
+    sdt2 = [('a', 'f8', (3, 2, 2))]
+    a = np.zeros((6,), dtype=sdt1)
+    a['a'] = np.arange(6)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    assert_equal(i[0].dtype, np.dtype(sdt2))
+    for x, count in zip(i, list(range(6))):
+        assert_(np.all(x['a'] == count))
+
+    # one element -> many -> back (copies it to all)
+    sdt1 = [('a', 'O', (1, 1))]
+    sdt2 = [('a', 'O', (3, 2, 2))]
+    a = np.zeros((6,), dtype=sdt1)
+    a['a'][:, 0, 0] = np.arange(6)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readwrite'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    with i:
+        assert_equal(i[0].dtype, np.dtype(sdt2))
+        count = 0
+        for x in i:
+            assert_(np.all(x['a'] == count))
+            x['a'][0] += 2
+            count += 1
+    assert_equal(a['a'], np.arange(6).reshape(6, 1, 1) + 2)
+
+    # many -> one element -> back (copies just element 0)
+    sdt1 = [('a', 'O', (3, 2, 2))]
+    sdt2 = [('a', 'O', (1,))]
+    a = np.zeros((6,), dtype=sdt1)
+    a['a'][:, 0, 0, 0] = np.arange(6)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readwrite'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    with i:
+        assert_equal(i[0].dtype, np.dtype(sdt2))
+        count = 0
+        for x in i:
+            assert_equal(x['a'], count)
+            x['a'] += 2
+            count += 1
+    assert_equal(a['a'], np.arange(6).reshape(6, 1, 1, 1) * np.ones((1, 3, 2, 2)) + 2)
+
+    # many -> one element -> back (copies just element 0)
+    sdt1 = [('a', 'f8', (3, 2, 2))]
+    sdt2 = [('a', 'O', (1,))]
+    a = np.zeros((6,), dtype=sdt1)
+    a['a'][:, 0, 0, 0] = np.arange(6)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    assert_equal(i[0].dtype, np.dtype(sdt2))
+    count = 0
+    for x in i:
+        assert_equal(x['a'], count)
+        count += 1
+
+    # many -> one element (copies just element 0)
+    sdt1 = [('a', 'O', (3, 2, 2))]
+    sdt2 = [('a', 'f4', (1,))]
+    a = np.zeros((6,), dtype=sdt1)
+    a['a'][:, 0, 0, 0] = np.arange(6)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    assert_equal(i[0].dtype, np.dtype(sdt2))
+    count = 0
+    for x in i:
+        assert_equal(x['a'], count)
+        count += 1
+
+    # many -> matching shape (straightforward copy)
+    sdt1 = [('a', 'O', (3, 2, 2))]
+    sdt2 = [('a', 'f4', (3, 2, 2))]
+    a = np.zeros((6,), dtype=sdt1)
+    a['a'] = np.arange(6 * 3 * 2 * 2).reshape(6, 3, 2, 2)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    assert_equal(i[0].dtype, np.dtype(sdt2))
+    count = 0
+    for x in i:
+        assert_equal(x['a'], a[count]['a'])
+        count += 1
+
+    # vector -> smaller vector (truncates)
+    sdt1 = [('a', 'f8', (6,))]
+    sdt2 = [('a', 'f4', (2,))]
+    a = np.zeros((6,), dtype=sdt1)
+    a['a'] = np.arange(6 * 6).reshape(6, 6)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    assert_equal(i[0].dtype, np.dtype(sdt2))
+    count = 0
+    for x in i:
+        assert_equal(x['a'], a[count]['a'][:2])
+        count += 1
+
+    # vector -> bigger vector (pads with zeros)
+    sdt1 = [('a', 'f8', (2,))]
+    sdt2 = [('a', 'f4', (6,))]
+    a = np.zeros((6,), dtype=sdt1)
+    a['a'] = np.arange(6 * 2).reshape(6, 2)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    assert_equal(i[0].dtype, np.dtype(sdt2))
+    count = 0
+    for x in i:
+        assert_equal(x['a'][:2], a[count]['a'])
+        assert_equal(x['a'][2:], [0, 0, 0, 0])
+        count += 1
+
+    # vector -> matrix (broadcasts)
+    sdt1 = [('a', 'f8', (2,))]
+    sdt2 = [('a', 'f4', (2, 2))]
+    a = np.zeros((6,), dtype=sdt1)
+    a['a'] = np.arange(6 * 2).reshape(6, 2)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    assert_equal(i[0].dtype, np.dtype(sdt2))
+    count = 0
+    for x in i:
+        assert_equal(x['a'][0], a[count]['a'])
+        assert_equal(x['a'][1], a[count]['a'])
+        count += 1
+
+    # vector -> matrix (broadcasts and zero-pads)
+    sdt1 = [('a', 'f8', (2, 1))]
+    sdt2 = [('a', 'f4', (3, 2))]
+    a = np.zeros((6,), dtype=sdt1)
+    a['a'] = np.arange(6 * 2).reshape(6, 2, 1)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    assert_equal(i[0].dtype, np.dtype(sdt2))
+    count = 0
+    for x in i:
+        assert_equal(x['a'][:2, 0], a[count]['a'][:, 0])
+        assert_equal(x['a'][:2, 1], a[count]['a'][:, 0])
+        assert_equal(x['a'][2, :], [0, 0])
+        count += 1
+
+    # matrix -> matrix (truncates and zero-pads)
+    sdt1 = [('a', 'f8', (2, 3))]
+    sdt2 = [('a', 'f4', (3, 2))]
+    a = np.zeros((6,), dtype=sdt1)
+    a['a'] = np.arange(6 * 2 * 3).reshape(6, 2, 3)
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe',
+                    op_dtypes=sdt2)
+    assert_equal(i[0].dtype, np.dtype(sdt2))
+    count = 0
+    for x in i:
+        assert_equal(x['a'][:2, 0], a[count]['a'][:, 0])
+        assert_equal(x['a'][:2, 1], a[count]['a'][:, 1])
+        assert_equal(x['a'][2, :], [0, 0])
+        count += 1
+
+def test_iter_buffering_badwriteback():
+    # Writing back from a buffer cannot combine elements
+
+    # a needs write buffering, but had a broadcast dimension
+    a = np.arange(6).reshape(2, 3, 1)
+    b = np.arange(12).reshape(2, 3, 2)
+    assert_raises(ValueError, nditer, [a, b],
+                  ['buffered', 'external_loop'],
+                  [['readwrite'], ['writeonly']],
+                  order='C')
+
+    # But if a is readonly, it's fine
+    nditer([a, b], ['buffered', 'external_loop'],
+           [['readonly'], ['writeonly']],
+           order='C')
+
+    # If a has just one element, it's fine too (constant 0 stride, a reduction)
+    a = np.arange(1).reshape(1, 1, 1)
+    nditer([a, b], ['buffered', 'external_loop', 'reduce_ok'],
+           [['readwrite'], ['writeonly']],
+           order='C')
+
+    # check that it fails on other dimensions too
+    a = np.arange(6).reshape(1, 3, 2)
+    assert_raises(ValueError, nditer, [a, b],
+                  ['buffered', 'external_loop'],
+                  [['readwrite'], ['writeonly']],
+                  order='C')
+    a = np.arange(4).reshape(2, 1, 2)
+    assert_raises(ValueError, nditer, [a, b],
+                  ['buffered', 'external_loop'],
+                  [['readwrite'], ['writeonly']],
+                  order='C')
+
+def test_iter_buffering_string():
+    # Safe casting disallows shrinking strings
+    a = np.array(['abc', 'a', 'abcd'], dtype=np.bytes_)
+    assert_equal(a.dtype, np.dtype('S4'))
+    assert_raises(TypeError, nditer, a, ['buffered'], ['readonly'],
+                  op_dtypes='S2')
+    i = nditer(a, ['buffered'], ['readonly'], op_dtypes='S6')
+    assert_equal(i[0], b'abc')
+    assert_equal(i[0].dtype, np.dtype('S6'))
+
+    a = np.array(['abc', 'a', 'abcd'], dtype=np.str_)
+    assert_equal(a.dtype, np.dtype('U4'))
+    assert_raises(TypeError, nditer, a, ['buffered'], ['readonly'],
+                    op_dtypes='U2')
+    i = nditer(a, ['buffered'], ['readonly'], op_dtypes='U6')
+    assert_equal(i[0], 'abc')
+    assert_equal(i[0].dtype, np.dtype('U6'))
+
+def test_iter_buffering_growinner():
+    # Test that the inner loop grows when no buffering is needed
+    a = np.arange(30)
+    i = nditer(a, ['buffered', 'growinner', 'external_loop'],
+                           buffersize=5)
+    # Should end up with just one inner loop here
+    assert_equal(i[0].size, a.size)
+
+
+@pytest.mark.parametrize("read_or_readwrite", ["readonly", "readwrite"])
+def test_iter_contig_flag_reduce_error(read_or_readwrite):
+    # Test that a non-contiguous operand is rejected without buffering.
+    # NOTE: This is true even for a reduction, where we return a 0-stride
+    #       below!
+    with pytest.raises(TypeError, match="Iterator operand required buffering"):
+        it = np.nditer(
+            (np.zeros(()),), flags=["external_loop", "reduce_ok"],
+            op_flags=[(read_or_readwrite, "contig"),], itershape=(10,))
+
+
+@pytest.mark.parametrize("arr", [
+        lambda: np.zeros(()),
+        lambda: np.zeros((20, 1))[::20],
+        lambda: np.zeros((1, 20))[:, ::20]
+    ])
+def test_iter_contig_flag_single_operand_strides(arr):
+    """
+    Tests the strides with the contig flag for both broadcast and non-broadcast
+    operands in 3 cases where the logic is needed:
+    1. When everything has a zero stride, the broadcast op needs to repeated
+    2. When the reduce axis is the last axis (first to iterate).
+    3. When the reduce axis is the first axis (last to iterate).
+
+    NOTE: The semantics of the cast flag are not clearly defined when
+          it comes to reduction.  It is unclear that there are any users.
+    """
+    first_op = np.ones((10, 10))
+    broadcast_op = arr()
+    red_op = arr()
+    # Add a first operand to ensure no axis-reordering and the result shape.
+    iterator = np.nditer(
+        (first_op, broadcast_op, red_op),
+        flags=["external_loop", "reduce_ok", "buffered", "delay_bufalloc"],
+        op_flags=[("readonly", "contig")] * 2 + [("readwrite", "contig")])
+
+    with iterator:
+        iterator.reset()
+        for f, b, r in iterator:
+            # The first operand is contigouos, we should have a view
+            assert np.shares_memory(f, first_op)
+            # Although broadcast, the second op always has a contiguous stride
+            assert b.strides[0] == 8
+            assert not np.shares_memory(b, broadcast_op)
+            # The reduction has a contiguous stride or a 0 stride
+            if red_op.ndim == 0 or red_op.shape[-1] == 1:
+                assert r.strides[0] == 0
+            else:
+                # The stride is 8, although it was not originally:
+                assert r.strides[0] == 8
+            # If the reduce stride is 0, buffering makes no difference, but we
+            # do it anyway right now:
+            assert not np.shares_memory(r, red_op)
+
+
+@pytest.mark.xfail(reason="The contig flag was always buggy.")
+def test_iter_contig_flag_incorrect():
+    # This case does the wrong thing...
+    iterator = np.nditer(
+        (np.ones((10, 10)).T, np.ones((1, 10))),
+        flags=["external_loop", "reduce_ok", "buffered", "delay_bufalloc"],
+        op_flags=[("readonly", "contig")] * 2)
+
+    with iterator:
+        iterator.reset()
+        for a, b in iterator:
+            # Remove a and b from locals (pytest may want to format them)
+            a, b = a.strides, b.strides
+            assert a == 8
+            assert b == 8  # should be 8 but is 0 due to axis reorder
+
+
+@pytest.mark.slow
+def test_iter_buffered_reduce_reuse():
+    # large enough array for all views, including negative strides.
+    a = np.arange(2 * 3**5)[3**5:3**5 + 1]
+    flags = ['buffered', 'delay_bufalloc', 'multi_index', 'reduce_ok', 'refs_ok']
+    op_flags = [('readonly',), ('readwrite', 'allocate')]
+    op_axes_list = [[(0, 1, 2), (0, 1, -1)], [(0, 1, 2), (0, -1, -1)]]
+    # wrong dtype to force buffering
+    op_dtypes = [float, a.dtype]
+
+    def get_params():
+        for xs in range(-3**2, 3**2 + 1):
+            for ys in range(xs, 3**2 + 1):
+                for op_axes in op_axes_list:
+                    # last stride is reduced and because of that not
+                    # important for this test, as it is the inner stride.
+                    strides = (xs * a.itemsize, ys * a.itemsize, a.itemsize)
+                    arr = np.lib.stride_tricks.as_strided(a, (3, 3, 3), strides)
+
+                    for skip in [0, 1]:
+                        yield arr, op_axes, skip
+
+    for arr, op_axes, skip in get_params():
+        nditer2 = np.nditer([arr.copy(), None],
+                            op_axes=op_axes, flags=flags, op_flags=op_flags,
+                            op_dtypes=op_dtypes)
+        with nditer2:
+            nditer2.operands[-1][...] = 0
+            nditer2.reset()
+            nditer2.iterindex = skip
+
+            for (a2_in, b2_in) in nditer2:
+                b2_in += a2_in.astype(np.int_)
+
+            comp_res = nditer2.operands[-1]
+
+        for bufsize in range(3**3):
+            nditer1 = np.nditer([arr, None],
+                                op_axes=op_axes, flags=flags, op_flags=op_flags,
+                                buffersize=bufsize, op_dtypes=op_dtypes)
+            with nditer1:
+                nditer1.operands[-1][...] = 0
+                nditer1.reset()
+                nditer1.iterindex = skip
+
+                for (a1_in, b1_in) in nditer1:
+                    b1_in += a1_in.astype(np.int_)
+
+                res = nditer1.operands[-1]
+            assert_array_equal(res, comp_res)
+
+
+def test_iter_buffered_reduce_reuse_core():
+    # NumPy re-uses buffers for broadcast operands (as of writing when reading).
+    # Test this even if the offset is manually set at some point during
+    # the iteration.  (not a particularly tricky path)
+    arr = np.empty((1, 6, 4, 1)).reshape(1, 6, 4, 1)[:, ::3, ::2, :]
+    arr[...] = np.arange(arr.size).reshape(arr.shape)
+    # First and last dimension are broadcast dimensions.
+    arr = np.broadcast_to(arr, (100, 2, 2, 2))
+
+    flags = ['buffered', 'reduce_ok', 'refs_ok', 'multi_index']
+    op_flags = [('readonly',)]
+
+    buffersize = 100  # small enough to not fit the whole array
+    it = np.nditer(arr, flags=flags, op_flags=op_flags, buffersize=100)
+
+    # Iterate a bit (this will cause buffering internally)
+    expected = [next(it) for i in range(11)]
+    # Now, manually advance to inside the core (the +1)
+    it.iterindex = 10 * (2 * 2 * 2) + 1
+    result = [next(it) for i in range(10)]
+
+    assert expected[1:] == result
+
+
+def test_iter_no_broadcast():
+    # Test that the no_broadcast flag works
+    a = np.arange(24).reshape(2, 3, 4)
+    b = np.arange(6).reshape(2, 3, 1)
+    c = np.arange(12).reshape(3, 4)
+
+    nditer([a, b, c], [],
+           [['readonly', 'no_broadcast'],
+            ['readonly'], ['readonly']])
+    assert_raises(ValueError, nditer, [a, b, c], [],
+                  [['readonly'], ['readonly', 'no_broadcast'], ['readonly']])
+    assert_raises(ValueError, nditer, [a, b, c], [],
+                  [['readonly'], ['readonly'], ['readonly', 'no_broadcast']])
+
+
+class TestIterNested:
+
+    def test_basic(self):
+        # Test nested iteration basic usage
+        a = arange(12).reshape(2, 3, 2)
+
+        i, j = np.nested_iters(a, [[0], [1, 2]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 1, 2, 3, 4, 5], [6, 7, 8, 9, 10, 11]])
+
+        i, j = np.nested_iters(a, [[0, 1], [2]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 1], [2, 3], [4, 5], [6, 7], [8, 9], [10, 11]])
+
+        i, j = np.nested_iters(a, [[0, 2], [1]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 2, 4], [1, 3, 5], [6, 8, 10], [7, 9, 11]])
+
+    def test_reorder(self):
+        # Test nested iteration basic usage
+        a = arange(12).reshape(2, 3, 2)
+
+        # In 'K' order (default), it gets reordered
+        i, j = np.nested_iters(a, [[0], [2, 1]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 1, 2, 3, 4, 5], [6, 7, 8, 9, 10, 11]])
+
+        i, j = np.nested_iters(a, [[1, 0], [2]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 1], [2, 3], [4, 5], [6, 7], [8, 9], [10, 11]])
+
+        i, j = np.nested_iters(a, [[2, 0], [1]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 2, 4], [1, 3, 5], [6, 8, 10], [7, 9, 11]])
+
+        # In 'C' order, it doesn't
+        i, j = np.nested_iters(a, [[0], [2, 1]], order='C')
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 2, 4, 1, 3, 5], [6, 8, 10, 7, 9, 11]])
+
+        i, j = np.nested_iters(a, [[1, 0], [2]], order='C')
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 1], [6, 7], [2, 3], [8, 9], [4, 5], [10, 11]])
+
+        i, j = np.nested_iters(a, [[2, 0], [1]], order='C')
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 2, 4], [6, 8, 10], [1, 3, 5], [7, 9, 11]])
+
+    def test_flip_axes(self):
+        # Test nested iteration with negative axes
+        a = arange(12).reshape(2, 3, 2)[::-1, ::-1, ::-1]
+
+        # In 'K' order (default), the axes all get flipped
+        i, j = np.nested_iters(a, [[0], [1, 2]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 1, 2, 3, 4, 5], [6, 7, 8, 9, 10, 11]])
+
+        i, j = np.nested_iters(a, [[0, 1], [2]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 1], [2, 3], [4, 5], [6, 7], [8, 9], [10, 11]])
+
+        i, j = np.nested_iters(a, [[0, 2], [1]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 2, 4], [1, 3, 5], [6, 8, 10], [7, 9, 11]])
+
+        # In 'C' order, flipping axes is disabled
+        i, j = np.nested_iters(a, [[0], [1, 2]], order='C')
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[11, 10, 9, 8, 7, 6], [5, 4, 3, 2, 1, 0]])
+
+        i, j = np.nested_iters(a, [[0, 1], [2]], order='C')
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[11, 10], [9, 8], [7, 6], [5, 4], [3, 2], [1, 0]])
+
+        i, j = np.nested_iters(a, [[0, 2], [1]], order='C')
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[11, 9, 7], [10, 8, 6], [5, 3, 1], [4, 2, 0]])
+
+    def test_broadcast(self):
+        # Test nested iteration with broadcasting
+        a = arange(2).reshape(2, 1)
+        b = arange(3).reshape(1, 3)
+
+        i, j = np.nested_iters([a, b], [[0], [1]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[[0, 0], [0, 1], [0, 2]], [[1, 0], [1, 1], [1, 2]]])
+
+        i, j = np.nested_iters([a, b], [[1], [0]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[[0, 0], [1, 0]], [[0, 1], [1, 1]], [[0, 2], [1, 2]]])
+
+    def test_dtype_copy(self):
+        # Test nested iteration with a copy to change dtype
+
+        # copy
+        a = arange(6, dtype='i4').reshape(2, 3)
+        i, j = np.nested_iters(a, [[0], [1]],
+                            op_flags=['readonly', 'copy'],
+                            op_dtypes='f8')
+        assert_equal(j[0].dtype, np.dtype('f8'))
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 1, 2], [3, 4, 5]])
+        vals = None
+
+        # writebackifcopy - using context manager
+        a = arange(6, dtype='f4').reshape(2, 3)
+        i, j = np.nested_iters(a, [[0], [1]],
+                            op_flags=['readwrite', 'updateifcopy'],
+                            casting='same_kind',
+                            op_dtypes='f8')
+        with i, j:
+            assert_equal(j[0].dtype, np.dtype('f8'))
+            for x in i:
+                for y in j:
+                    y[...] += 1
+            assert_equal(a, [[0, 1, 2], [3, 4, 5]])
+        assert_equal(a, [[1, 2, 3], [4, 5, 6]])
+
+        # writebackifcopy - using close()
+        a = arange(6, dtype='f4').reshape(2, 3)
+        i, j = np.nested_iters(a, [[0], [1]],
+                            op_flags=['readwrite', 'updateifcopy'],
+                            casting='same_kind',
+                            op_dtypes='f8')
+        assert_equal(j[0].dtype, np.dtype('f8'))
+        for x in i:
+            for y in j:
+                y[...] += 1
+        assert_equal(a, [[0, 1, 2], [3, 4, 5]])
+        i.close()
+        j.close()
+        assert_equal(a, [[1, 2, 3], [4, 5, 6]])
+
+    def test_dtype_buffered(self):
+        # Test nested iteration with buffering to change dtype
+
+        a = arange(6, dtype='f4').reshape(2, 3)
+        i, j = np.nested_iters(a, [[0], [1]],
+                            flags=['buffered'],
+                            op_flags=['readwrite'],
+                            casting='same_kind',
+                            op_dtypes='f8')
+        assert_equal(j[0].dtype, np.dtype('f8'))
+        for x in i:
+            for y in j:
+                y[...] += 1
+        assert_equal(a, [[1, 2, 3], [4, 5, 6]])
+
+    def test_0d(self):
+        a = np.arange(12).reshape(2, 3, 2)
+        i, j = np.nested_iters(a, [[], [1, 0, 2]])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]])
+
+        i, j = np.nested_iters(a, [[1, 0, 2], []])
+        vals = [list(j) for _ in i]
+        assert_equal(vals, [[0], [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11]])
+
+        i, j, k = np.nested_iters(a, [[2, 0], [], [1]])
+        vals = []
+        for x in i:
+            for y in j:
+                vals.append(list(k))
+        assert_equal(vals, [[0, 2, 4], [1, 3, 5], [6, 8, 10], [7, 9, 11]])
+
+    def test_iter_nested_iters_dtype_buffered(self):
+        # Test nested iteration with buffering to change dtype
+
+        a = arange(6, dtype='f4').reshape(2, 3)
+        i, j = np.nested_iters(a, [[0], [1]],
+                            flags=['buffered'],
+                            op_flags=['readwrite'],
+                            casting='same_kind',
+                            op_dtypes='f8')
+        with i, j:
+            assert_equal(j[0].dtype, np.dtype('f8'))
+            for x in i:
+                for y in j:
+                    y[...] += 1
+        assert_equal(a, [[1, 2, 3], [4, 5, 6]])
+
+def test_iter_reduction_error():
+
+    a = np.arange(6)
+    assert_raises(ValueError, nditer, [a, None], [],
+                    [['readonly'], ['readwrite', 'allocate']],
+                    op_axes=[[0], [-1]])
+
+    a = np.arange(6).reshape(2, 3)
+    assert_raises(ValueError, nditer, [a, None], ['external_loop'],
+                    [['readonly'], ['readwrite', 'allocate']],
+                    op_axes=[[0, 1], [-1, -1]])
+
+def test_iter_reduction():
+    # Test doing reductions with the iterator
+
+    a = np.arange(6)
+    i = nditer([a, None], ['reduce_ok'],
+                    [['readonly'], ['readwrite', 'allocate']],
+                    op_axes=[[0], [-1]])
+    # Need to initialize the output operand to the addition unit
+    with i:
+        i.operands[1][...] = 0
+        # Do the reduction
+        for x, y in i:
+            y[...] += x
+        # Since no axes were specified, should have allocated a scalar
+        assert_equal(i.operands[1].ndim, 0)
+        assert_equal(i.operands[1], np.sum(a))
+
+    a = np.arange(6).reshape(2, 3)
+    i = nditer([a, None], ['reduce_ok', 'external_loop'],
+                    [['readonly'], ['readwrite', 'allocate']],
+                    op_axes=[[0, 1], [-1, -1]])
+    # Need to initialize the output operand to the addition unit
+    with i:
+        i.operands[1][...] = 0
+        # Reduction shape/strides for the output
+        assert_equal(i[1].shape, (6,))
+        assert_equal(i[1].strides, (0,))
+        # Do the reduction
+        for x, y in i:
+            # Use a for loop instead of ``y[...] += x``
+            # (equivalent to ``y[...] = y[...].copy() + x``),
+            # because y has zero strides we use for the reduction
+            for j in range(len(y)):
+                y[j] += x[j]
+        # Since no axes were specified, should have allocated a scalar
+        assert_equal(i.operands[1].ndim, 0)
+        assert_equal(i.operands[1], np.sum(a))
+
+    # This is a tricky reduction case for the buffering double loop
+    # to handle
+    a = np.ones((2, 3, 5))
+    it1 = nditer([a, None], ['reduce_ok', 'external_loop'],
+                    [['readonly'], ['readwrite', 'allocate']],
+                    op_axes=[None, [0, -1, 1]])
+    it2 = nditer([a, None], ['reduce_ok', 'external_loop',
+                            'buffered', 'delay_bufalloc'],
+                    [['readonly'], ['readwrite', 'allocate']],
+                    op_axes=[None, [0, -1, 1]], buffersize=10)
+    with it1, it2:
+        it1.operands[1].fill(0)
+        it2.operands[1].fill(0)
+        it2.reset()
+        for x in it1:
+            x[1][...] += x[0]
+        for x in it2:
+            x[1][...] += x[0]
+        assert_equal(it1.operands[1], it2.operands[1])
+        assert_equal(it2.operands[1].sum(), a.size)
+
+def test_iter_buffering_reduction():
+    # Test doing buffered reductions with the iterator
+
+    a = np.arange(6)
+    b = np.array(0., dtype='f8').byteswap()
+    b = b.view(b.dtype.newbyteorder())
+    i = nditer([a, b], ['reduce_ok', 'buffered'],
+                    [['readonly'], ['readwrite', 'nbo']],
+                    op_axes=[[0], [-1]])
+    with i:
+        assert_equal(i[1].dtype, np.dtype('f8'))
+        assert_(i[1].dtype != b.dtype)
+        # Do the reduction
+        for x, y in i:
+            y[...] += x
+    # Since no axes were specified, should have allocated a scalar
+    assert_equal(b, np.sum(a))
+
+    a = np.arange(6).reshape(2, 3)
+    b = np.array([0, 0], dtype='f8').byteswap()
+    b = b.view(b.dtype.newbyteorder())
+    i = nditer([a, b], ['reduce_ok', 'external_loop', 'buffered'],
+                    [['readonly'], ['readwrite', 'nbo']],
+                    op_axes=[[0, 1], [0, -1]])
+    # Reduction shape/strides for the output
+    with i:
+        assert_equal(i[1].shape, (3,))
+        assert_equal(i[1].strides, (0,))
+        # Do the reduction
+        for x, y in i:
+            # Use a for loop instead of ``y[...] += x``
+            # (equivalent to ``y[...] = y[...].copy() + x``),
+            # because y has zero strides we use for the reduction
+            for j in range(len(y)):
+                y[j] += x[j]
+    assert_equal(b, np.sum(a, axis=1))
+
+    # Iterator inner double loop was wrong on this one
+    p = np.arange(2) + 1
+    it = np.nditer([p, None],
+            ['delay_bufalloc', 'reduce_ok', 'buffered', 'external_loop'],
+            [['readonly'], ['readwrite', 'allocate']],
+            op_axes=[[-1, 0], [-1, -1]],
+            itershape=(2, 2))
+    with it:
+        it.operands[1].fill(0)
+        it.reset()
+        assert_equal(it[0], [1, 2, 1, 2])
+
+    # Iterator inner loop should take argument contiguity into account
+    x = np.ones((7, 13, 8), np.int8)[4:6, 1:11:6, 1:5].transpose(1, 2, 0)
+    x[...] = np.arange(x.size).reshape(x.shape)
+    y_base = np.arange(4 * 4, dtype=np.int8).reshape(4, 4)
+    y_base_copy = y_base.copy()
+    y = y_base[::2, :, None]
+
+    it = np.nditer([y, x],
+                   ['buffered', 'external_loop', 'reduce_ok'],
+                   [['readwrite'], ['readonly']])
+    with it:
+        for a, b in it:
+            a.fill(2)
+
+    assert_equal(y_base[1::2], y_base_copy[1::2])
+    assert_equal(y_base[::2], 2)
+
+def test_iter_buffering_reduction_reuse_reduce_loops():
+    # There was a bug triggering reuse of the reduce loop inappropriately,
+    # which caused processing to happen in unnecessarily small chunks
+    # and overran the buffer.
+
+    a = np.zeros((2, 7))
+    b = np.zeros((1, 7))
+    it = np.nditer([a, b], flags=['reduce_ok', 'external_loop', 'buffered'],
+                    op_flags=[['readonly'], ['readwrite']],
+                    buffersize=5)
+
+    with it:
+        bufsizes = [x.shape[0] for x, y in it]
+    assert_equal(bufsizes, [5, 2, 5, 2])
+    assert_equal(sum(bufsizes), a.size)
+
+def test_iter_writemasked_badinput():
+    a = np.zeros((2, 3))
+    b = np.zeros((3,))
+    m = np.array([[True, True, False], [False, True, False]])
+    m2 = np.array([True, True, False])
+    m3 = np.array([0, 1, 1], dtype='u1')
+    mbad1 = np.array([0, 1, 1], dtype='i1')
+    mbad2 = np.array([0, 1, 1], dtype='f4')
+
+    # Need an 'arraymask' if any operand is 'writemasked'
+    assert_raises(ValueError, nditer, [a, m], [],
+                    [['readwrite', 'writemasked'], ['readonly']])
+
+    # A 'writemasked' operand must not be readonly
+    assert_raises(ValueError, nditer, [a, m], [],
+                    [['readonly', 'writemasked'], ['readonly', 'arraymask']])
+
+    # 'writemasked' and 'arraymask' may not be used together
+    assert_raises(ValueError, nditer, [a, m], [],
+                    [['readonly'], ['readwrite', 'arraymask', 'writemasked']])
+
+    # 'arraymask' may only be specified once
+    assert_raises(ValueError, nditer, [a, m, m2], [],
+                    [['readwrite', 'writemasked'],
+                     ['readonly', 'arraymask'],
+                     ['readonly', 'arraymask']])
+
+    # An 'arraymask' with nothing 'writemasked' also doesn't make sense
+    assert_raises(ValueError, nditer, [a, m], [],
+                    [['readwrite'], ['readonly', 'arraymask']])
+
+    # A writemasked reduction requires a similarly smaller mask
+    assert_raises(ValueError, nditer, [a, b, m], ['reduce_ok'],
+                    [['readonly'],
+                     ['readwrite', 'writemasked'],
+                     ['readonly', 'arraymask']])
+    # But this should work with a smaller/equal mask to the reduction operand
+    np.nditer([a, b, m2], ['reduce_ok'],
+                    [['readonly'],
+                     ['readwrite', 'writemasked'],
+                     ['readonly', 'arraymask']])
+    # The arraymask itself cannot be a reduction
+    assert_raises(ValueError, nditer, [a, b, m2], ['reduce_ok'],
+                    [['readonly'],
+                     ['readwrite', 'writemasked'],
+                     ['readwrite', 'arraymask']])
+
+    # A uint8 mask is ok too
+    np.nditer([a, m3], ['buffered'],
+                    [['readwrite', 'writemasked'],
+                     ['readonly', 'arraymask']],
+                    op_dtypes=['f4', None],
+                    casting='same_kind')
+    # An int8 mask isn't ok
+    assert_raises(TypeError, np.nditer, [a, mbad1], ['buffered'],
+                    [['readwrite', 'writemasked'],
+                     ['readonly', 'arraymask']],
+                    op_dtypes=['f4', None],
+                    casting='same_kind')
+    # A float32 mask isn't ok
+    assert_raises(TypeError, np.nditer, [a, mbad2], ['buffered'],
+                    [['readwrite', 'writemasked'],
+                     ['readonly', 'arraymask']],
+                    op_dtypes=['f4', None],
+                    casting='same_kind')
+
+
+def _is_buffered(iterator):
+    try:
+        iterator.itviews
+    except ValueError:
+        return True
+    return False
+
+@pytest.mark.parametrize("arrs",
+        [np.zeros((3,), dtype='f8'),
+         np.zeros((9876, 3 * 5), dtype='f8')[::2, :],
+         np.zeros((4, 312, 124, 3), dtype='f8')[::2, :, ::2, :],
+         # Also test with the last dimension strided (so it does not fit if
+         # there is repeated access)
+         np.zeros((9,), dtype='f8')[::3],
+         np.zeros((9876, 3 * 10), dtype='f8')[::2, ::5],
+         np.zeros((4, 312, 124, 3), dtype='f8')[::2, :, ::2, ::-1]])
+def test_iter_writemasked(arrs):
+    # Note, the slicing above is to ensure that nditer cannot combine multiple
+    # axes into one.  The repetition is just to make things a bit more
+    # interesting.
+    a = arrs.copy()
+    shape = a.shape
+    reps = shape[-1] // 3
+    msk = np.empty(shape, dtype=bool)
+    msk[...] = [True, True, False] * reps
+
+    # When buffering is unused, 'writemasked' effectively does nothing.
+    # It's up to the user of the iterator to obey the requested semantics.
+    it = np.nditer([a, msk], [],
+                [['readwrite', 'writemasked'],
+                 ['readonly', 'arraymask']])
+    with it:
+        for x, m in it:
+            x[...] = 1
+    # Because we violated the semantics, all the values became 1
+    assert_equal(a, np.broadcast_to([1, 1, 1] * reps, shape))
+
+    # Even if buffering is enabled, we still may be accessing the array
+    # directly.
+    it = np.nditer([a, msk], ['buffered'],
+                [['readwrite', 'writemasked'],
+                 ['readonly', 'arraymask']])
+    # @seberg: I honestly don't currently understand why a "buffered" iterator
+    # would end up not using a buffer for the small array here at least when
+    # "writemasked" is used, that seems confusing...  Check by testing for
+    # actual memory overlap!
+    is_buffered = True
+    with it:
+        for x, m in it:
+            x[...] = 2.5
+            if np.may_share_memory(x, a):
+                is_buffered = False
+
+    if not is_buffered:
+        # Because we violated the semantics, all the values became 2.5
+        assert_equal(a, np.broadcast_to([2.5, 2.5, 2.5] * reps, shape))
+    else:
+        # For large sizes, the iterator may be buffered:
+        assert_equal(a, np.broadcast_to([2.5, 2.5, 1] * reps, shape))
+        a[...] = 2.5
+
+    # If buffering will definitely happening, for instance because of
+    # a cast, only the items selected by the mask will be copied back from
+    # the buffer.
+    it = np.nditer([a, msk], ['buffered'],
+                [['readwrite', 'writemasked'],
+                 ['readonly', 'arraymask']],
+                op_dtypes=['i8', None],
+                casting='unsafe')
+    with it:
+        for x, m in it:
+            x[...] = 3
+    # Even though we violated the semantics, only the selected values
+    # were copied back
+    assert_equal(a, np.broadcast_to([3, 3, 2.5] * reps, shape))
+
+
+@pytest.mark.parametrize(["mask", "mask_axes"], [
+        # Allocated operand (only broadcasts with -1)
+        (None, [-1, 0]),
+        # Reduction along the first dimension (with and without op_axes)
+        (np.zeros((1, 4), dtype="bool"), [0, 1]),
+        (np.zeros((1, 4), dtype="bool"), None),
+        # Test 0-D and -1 op_axes
+        (np.zeros(4, dtype="bool"), [-1, 0]),
+        (np.zeros((), dtype="bool"), [-1, -1]),
+        (np.zeros((), dtype="bool"), None)])
+def test_iter_writemasked_broadcast_error(mask, mask_axes):
+    # This assumes that a readwrite mask makes sense. This is likely not the
+    # case and should simply be deprecated.
+    arr = np.zeros((3, 4))
+    itflags = ["reduce_ok"]
+    mask_flags = ["arraymask", "readwrite", "allocate"]
+    a_flags = ["writeonly", "writemasked"]
+    if mask_axes is None:
+        op_axes = None
+    else:
+        op_axes = [mask_axes, [0, 1]]
+
+    with assert_raises(ValueError):
+        np.nditer((mask, arr), flags=itflags, op_flags=[mask_flags, a_flags],
+                  op_axes=op_axes)
+
+
+def test_iter_writemasked_decref():
+    # force casting (to make it interesting) by using a structured dtype.
+    arr = np.arange(10000).astype(">i,O")
+    original = arr.copy()
+    mask = np.random.randint(0, 2, size=10000).astype(bool)
+
+    it = np.nditer([arr, mask], ['buffered', "refs_ok"],
+                   [['readwrite', 'writemasked'],
+                    ['readonly', 'arraymask']],
+                   op_dtypes=["<i,O", "?"])
+    singleton = object()
+    if HAS_REFCOUNT:
+        count = sys.getrefcount(singleton)
+    for buf, mask_buf in it:
+        buf[...] = (3, singleton)
+
+    del buf, mask_buf, it   # delete everything to ensure correct cleanup
+
+    if HAS_REFCOUNT:
+        # The buffer would have included additional items, they must be
+        # cleared correctly:
+        assert sys.getrefcount(singleton) - count == np.count_nonzero(mask)
+
+    assert_array_equal(arr[~mask], original[~mask])
+    assert (arr[mask] == np.array((3, singleton), arr.dtype)).all()
+    del arr
+
+    if HAS_REFCOUNT:
+        assert sys.getrefcount(singleton) == count
+
+
+def test_iter_non_writable_attribute_deletion():
+    it = np.nditer(np.ones(2))
+    attr = ["value", "shape", "operands", "itviews", "has_delayed_bufalloc",
+            "iterationneedsapi", "has_multi_index", "has_index", "dtypes",
+            "ndim", "nop", "itersize", "finished"]
+
+    for s in attr:
+        assert_raises(AttributeError, delattr, it, s)
+
+
+def test_iter_writable_attribute_deletion():
+    it = np.nditer(np.ones(2))
+    attr = ["multi_index", "index", "iterrange", "iterindex"]
+    for s in attr:
+        assert_raises(AttributeError, delattr, it, s)
+
+
+def test_iter_element_deletion():
+    it = np.nditer(np.ones(3))
+    try:
+        del it[1]
+        del it[1:2]
+    except TypeError:
+        pass
+    except Exception:
+        raise AssertionError
+
+def test_iter_allocated_array_dtypes():
+    # If the dtype of an allocated output has a shape, the shape gets
+    # tacked onto the end of the result.
+    it = np.nditer(([1, 3, 20], None), op_dtypes=[None, ('i4', (2,))])
+    for a, b in it:
+        b[0] = a - 1
+        b[1] = a + 1
+    assert_equal(it.operands[1], [[0, 2], [2, 4], [19, 21]])
+
+    # Check the same (less sensitive) thing when `op_axes` with -1 is given.
+    it = np.nditer(([[1, 3, 20]], None), op_dtypes=[None, ('i4', (2,))],
+                   flags=["reduce_ok"], op_axes=[None, (-1, 0)])
+    for a, b in it:
+        b[0] = a - 1
+        b[1] = a + 1
+    assert_equal(it.operands[1], [[0, 2], [2, 4], [19, 21]])
+
+    # Make sure this works for scalars too
+    it = np.nditer((10, 2, None), op_dtypes=[None, None, ('i4', (2, 2))])
+    for a, b, c in it:
+        c[0, 0] = a - b
+        c[0, 1] = a + b
+        c[1, 0] = a * b
+        c[1, 1] = a / b
+    assert_equal(it.operands[2], [[8, 12], [20, 5]])
+
+
+def test_0d_iter():
+    # Basic test for iteration of 0-d arrays:
+    i = nditer([2, 3], ['multi_index'], [['readonly']] * 2)
+    assert_equal(i.ndim, 0)
+    assert_equal(next(i), (2, 3))
+    assert_equal(i.multi_index, ())
+    assert_equal(i.iterindex, 0)
+    assert_raises(StopIteration, next, i)
+    # test reset:
+    i.reset()
+    assert_equal(next(i), (2, 3))
+    assert_raises(StopIteration, next, i)
+
+    # test forcing to 0-d
+    i = nditer(np.arange(5), ['multi_index'], [['readonly']], op_axes=[()])
+    assert_equal(i.ndim, 0)
+    assert_equal(len(i), 1)
+
+    i = nditer(np.arange(5), ['multi_index'], [['readonly']],
+               op_axes=[()], itershape=())
+    assert_equal(i.ndim, 0)
+    assert_equal(len(i), 1)
+
+    # passing an itershape alone is not enough, the op_axes are also needed
+    with assert_raises(ValueError):
+        nditer(np.arange(5), ['multi_index'], [['readonly']], itershape=())
+
+    # Test a more complex buffered casting case (same as another test above)
+    sdt = [('a', 'f4'), ('b', 'i8'), ('c', 'c8', (2, 3)), ('d', 'O')]
+    a = np.array(0.5, dtype='f4')
+    i = nditer(a, ['buffered', 'refs_ok'], ['readonly'],
+                    casting='unsafe', op_dtypes=sdt)
+    vals = next(i)
+    assert_equal(vals['a'], 0.5)
+    assert_equal(vals['b'], 0)
+    assert_equal(vals['c'], [[(0.5)] * 3] * 2)
+    assert_equal(vals['d'], 0.5)
+
+def test_object_iter_cleanup():
+    # see gh-18450
+    # object arrays can raise a python exception in ufunc inner loops using
+    # nditer, which should cause iteration to stop & cleanup. There were bugs
+    # in the nditer cleanup when decref'ing object arrays.
+    # This test would trigger valgrind "uninitialized read" before the bugfix.
+    assert_raises(TypeError, lambda: np.zeros((17000, 2), dtype='f4') * None)
+
+    # this more explicit code also triggers the invalid access
+    arr = np.arange(ncu.BUFSIZE * 10).reshape(10, -1).astype(str)
+    oarr = arr.astype(object)
+    oarr[:, -1] = None
+    assert_raises(TypeError, lambda: np.add(oarr[:, ::-1], arr[:, ::-1]))
+
+    # followup: this tests for a bug introduced in the first pass of gh-18450,
+    # caused by an incorrect fallthrough of the TypeError
+    class T:
+        def __bool__(self):
+            raise TypeError("Ambiguous")
+    assert_raises(TypeError, np.logical_or.reduce,
+                             np.array([T(), T()], dtype='O'))
+
+def test_object_iter_cleanup_reduce():
+    # Similar as above, but a complex reduction case that was previously
+    # missed (see gh-18810).
+    # The following array is special in that it cannot be flattened:
+    arr = np.array([[None, 1], [-1, -1], [None, 2], [-1, -1]])[::2]
+    with pytest.raises(TypeError):
+        np.sum(arr)
+
+@pytest.mark.parametrize("arr", [
+        np.ones((8000, 4, 2), dtype=object)[:, ::2, :],
+        np.ones((8000, 4, 2), dtype=object, order="F")[:, ::2, :],
+        np.ones((8000, 4, 2), dtype=object)[:, ::2, :].copy("F")])
+def test_object_iter_cleanup_large_reduce(arr):
+    # More complicated calls are possible for large arrays:
+    out = np.ones(8000, dtype=np.intp)
+    # force casting with `dtype=object`
+    res = np.sum(arr, axis=(1, 2), dtype=object, out=out)
+    assert_array_equal(res, np.full(8000, 4, dtype=object))
+
+def test_iter_too_large():
+    # The total size of the iterator must not exceed the maximum intp due
+    # to broadcasting. Dividing by 1024 will keep it small enough to
+    # give a legal array.
+    size = np.iinfo(np.intp).max // 1024
+    arr = np.lib.stride_tricks.as_strided(np.zeros(1), (size,), (0,))
+    assert_raises(ValueError, nditer, (arr, arr[:, None]))
+    # test the same for multiindex. That may get more interesting when
+    # removing 0 dimensional axis is allowed (since an iterator can grow then)
+    assert_raises(ValueError, nditer,
+                  (arr, arr[:, None]), flags=['multi_index'])
+
+
+def test_iter_too_large_with_multiindex():
+    # When a multi index is being tracked, the error is delayed this
+    # checks the delayed error messages and getting below that by
+    # removing an axis.
+    base_size = 2**10
+    num = 1
+    while base_size**num < np.iinfo(np.intp).max:
+        num += 1
+
+    shape_template = [1, 1] * num
+    arrays = []
+    for i in range(num):
+        shape = shape_template[:]
+        shape[i * 2] = 2**10
+        arrays.append(np.empty(shape))
+    arrays = tuple(arrays)
+
+    # arrays are now too large to be broadcast. The different modes test
+    # different nditer functionality with or without GIL.
+    for mode in range(6):
+        with assert_raises(ValueError):
+            _multiarray_tests.test_nditer_too_large(arrays, -1, mode)
+    # but if we do nothing with the nditer, it can be constructed:
+    _multiarray_tests.test_nditer_too_large(arrays, -1, 7)
+
+    # When an axis is removed, things should work again (half the time):
+    for i in range(num):
+        for mode in range(6):
+            # an axis with size 1024 is removed:
+            _multiarray_tests.test_nditer_too_large(arrays, i * 2, mode)
+            # an axis with size 1 is removed:
+            with assert_raises(ValueError):
+                _multiarray_tests.test_nditer_too_large(arrays, i * 2 + 1, mode)
+
+
+def test_invalid_call_of_enable_external_loop():
+    with pytest.raises(ValueError,
+                       match='Iterator flag EXTERNAL_LOOP cannot be used'):
+        np.nditer(([[1], [2]], [3, 4]), ['multi_index']).enable_external_loop()
+
+
+def test_writebacks():
+    a = np.arange(6, dtype='f4')
+    au = a.byteswap()
+    au = au.view(au.dtype.newbyteorder())
+    assert_(a.dtype.byteorder != au.dtype.byteorder)
+    it = nditer(au, [], [['readwrite', 'updateifcopy']],
+                        casting='equiv', op_dtypes=[np.dtype('f4')])
+    with it:
+        it.operands[0][:] = 100
+    assert_equal(au, 100)
+    # do it again, this time raise an error,
+    it = nditer(au, [], [['readwrite', 'updateifcopy']],
+                        casting='equiv', op_dtypes=[np.dtype('f4')])
+    try:
+        with it:
+            assert_equal(au.flags.writeable, False)
+            it.operands[0][:] = 0
+            raise ValueError('exit context manager on exception')
+    except Exception:
+        pass
+    assert_equal(au, 0)
+    assert_equal(au.flags.writeable, True)
+    # cannot reuse i outside context manager
+    assert_raises(ValueError, getattr, it, 'operands')
+
+    it = nditer(au, [], [['readwrite', 'updateifcopy']],
+                        casting='equiv', op_dtypes=[np.dtype('f4')])
+    with it:
+        x = it.operands[0]
+        x[:] = 6
+        assert_(x.flags.writebackifcopy)
+    assert_equal(au, 6)
+    assert_(not x.flags.writebackifcopy)
+    x[:] = 123  # x.data still valid
+    assert_equal(au, 6)  # but not connected to au
+
+    it = nditer(au, [],
+                 [['readwrite', 'updateifcopy']],
+                 casting='equiv', op_dtypes=[np.dtype('f4')])
+    # reentering works
+    with it:
+        with it:
+            for x in it:
+                x[...] = 123
+
+    it = nditer(au, [],
+                 [['readwrite', 'updateifcopy']],
+                 casting='equiv', op_dtypes=[np.dtype('f4')])
+    # make sure exiting the inner context manager closes the iterator
+    with it:
+        with it:
+            for x in it:
+                x[...] = 123
+        assert_raises(ValueError, getattr, it, 'operands')
+    # do not crash if original data array is decrefed
+    it = nditer(au, [],
+                 [['readwrite', 'updateifcopy']],
+                 casting='equiv', op_dtypes=[np.dtype('f4')])
+    del au
+    with it:
+        for x in it:
+            x[...] = 123
+    # make sure we cannot reenter the closed iterator
+    enter = it.__enter__
+    assert_raises(RuntimeError, enter)
+
+def test_close_equivalent():
+    ''' using a context amanger and using nditer.close are equivalent
+    '''
+    def add_close(x, y, out=None):
+        addop = np.add
+        it = np.nditer([x, y, out], [],
+                    [['readonly'], ['readonly'], ['writeonly', 'allocate']])
+        for (a, b, c) in it:
+            addop(a, b, out=c)
+        ret = it.operands[2]
+        it.close()
+        return ret
+
+    def add_context(x, y, out=None):
+        addop = np.add
+        it = np.nditer([x, y, out], [],
+                    [['readonly'], ['readonly'], ['writeonly', 'allocate']])
+        with it:
+            for (a, b, c) in it:
+                addop(a, b, out=c)
+            return it.operands[2]
+    z = add_close(range(5), range(5))
+    assert_equal(z, range(0, 10, 2))
+    z = add_context(range(5), range(5))
+    assert_equal(z, range(0, 10, 2))
+
+def test_close_raises():
+    it = np.nditer(np.arange(3))
+    assert_equal(next(it), 0)
+    it.close()
+    assert_raises(StopIteration, next, it)
+    assert_raises(ValueError, getattr, it, 'operands')
+
+def test_close_parameters():
+    it = np.nditer(np.arange(3))
+    assert_raises(TypeError, it.close, 1)
+
+@pytest.mark.skipif(not HAS_REFCOUNT, reason="Python lacks refcounts")
+def test_warn_noclose():
+    a = np.arange(6, dtype='f4')
+    au = a.byteswap()
+    au = au.view(au.dtype.newbyteorder())
+    with pytest.warns(RuntimeWarning):
+        it = np.nditer(au, [], [['readwrite', 'updateifcopy']],
+                       casting='equiv', op_dtypes=[np.dtype('f4')])
+        del it
+
+@pytest.mark.parametrize(["in_dtype", "buf_dtype"],
+        [("i", "O"), ("O", "i"),  # most simple cases
+         ("i,O", "O,O"),  # structured partially only copying O
+         ("O,i", "i,O"),  # structured casting to and from O
+         ])
+@pytest.mark.parametrize("steps", [1, 2, 3])
+def test_partial_iteration_cleanup(in_dtype, buf_dtype, steps):
+    """
+    Checks for reference counting leaks during cleanup.  Using explicit
+    reference counts lead to occasional false positives (at least in parallel
+    test setups).  This test now should still test leaks correctly when
+    run e.g. with pytest-valgrind or pytest-leaks
+    """
+    value = 2**30 + 1  # just a random value that Python won't intern
+    arr = np.full(int(ncu.BUFSIZE * 2.5), value).astype(in_dtype)
+
+    it = np.nditer(arr, op_dtypes=[np.dtype(buf_dtype)],
+            flags=["buffered", "external_loop", "refs_ok"], casting="unsafe")
+    for step in range(steps):
+        # The iteration finishes in 3 steps, the first two are partial
+        next(it)
+
+    del it  # not necessary, but we test the cleanup
+
+    # Repeat the test with `iternext`
+    it = np.nditer(arr, op_dtypes=[np.dtype(buf_dtype)],
+                   flags=["buffered", "external_loop", "refs_ok"], casting="unsafe")
+    for step in range(steps):
+        it.iternext()
+
+    del it  # not necessary, but we test the cleanup
+
+@pytest.mark.skipif(not HAS_REFCOUNT, reason="Python lacks refcounts")
+@pytest.mark.parametrize(["in_dtype", "buf_dtype"],
+         [("O", "i"),  # most simple cases
+          ("O,i", "i,O"),  # structured casting to and from O
+          ])
+def test_partial_iteration_error(in_dtype, buf_dtype):
+    value = 123  # relies on python cache (leak-check will still find it)
+    arr = np.full(int(ncu.BUFSIZE * 2.5), value).astype(in_dtype)
+    if in_dtype == "O":
+        arr[int(ncu.BUFSIZE * 1.5)] = None
+    else:
+        arr[int(ncu.BUFSIZE * 1.5)]["f0"] = None
+
+    count = sys.getrefcount(value)
+
+    it = np.nditer(arr, op_dtypes=[np.dtype(buf_dtype)],
+            flags=["buffered", "external_loop", "refs_ok"], casting="unsafe")
+    with pytest.raises(TypeError):
+        # pytest.raises seems to have issues with the error originating
+        # in the for loop, so manually unravel:
+        next(it)
+        next(it)  # raises TypeError
+
+    # Repeat the test with `iternext` after resetting, the buffers should
+    # already be cleared from any references, so resetting is sufficient.
+    it.reset()
+    with pytest.raises(TypeError):
+        it.iternext()
+        it.iternext()
+
+    assert count == sys.getrefcount(value)
+
+
+def test_arbitrary_number_of_ops():
+    # 2*16 + 1 is still just a few kiB, so should be fast and easy to deal with
+    # but larger than any small custom integer.
+    ops = [np.arange(10) for a in range(2**16 + 1)]
+
+    it = np.nditer(ops)
+    for i, vals in enumerate(it):
+        assert all(v == i for v in vals)
+
+
+def test_arbitrary_number_of_ops_nested():
+    # 2*16 + 1 is still just a few kiB, so should be fast and easy to deal with
+    # but larger than any small custom integer.
+    ops = [np.arange(10) for a in range(2**16 + 1)]
+
+    it = np.nested_iters(ops, [[0], []])
+    for i, vals in enumerate(it):
+        assert all(v == i for v in vals)
+
+
+@pytest.mark.slow
+@pytest.mark.skipif(not IS_64BIT, reason="test requires 64-bit system")
+@requires_memory(9 * np.iinfo(np.intc).max)
+@pytest.mark.thread_unsafe(reason="crashes with low memory")
+def test_arbitrary_number_of_ops_error():
+    # A different error may happen for more than integer operands, but that
+    # is too large to test nicely.
+    a = np.ones(1)
+    args = [a] * (np.iinfo(np.intc).max + 1)
+    with pytest.raises(ValueError, match="Too many operands to nditer"):
+        np.nditer(args)
+
+    with pytest.raises(ValueError, match="Too many operands to nditer"):
+        np.nested_iters(args, [[0], []])
+
+
+@pytest.mark.thread_unsafe(reason="capfd is thread-unsafe")
+def test_debug_print(capfd):
+    """
+    Matches the expected output of a debug print with the actual output.
+    Note that the iterator dump should not be considered stable API,
+    this test is mainly to ensure the print does not crash.
+
+    Currently uses a subprocess to avoid dealing with the C level `printf`s.
+    """
+    # the expected output with all addresses and sizes stripped (they vary
+    # and/or are platform dependent).
+    expected = """
+    ------ BEGIN ITERATOR DUMP ------
+    | Iterator Address:
+    | ItFlags: BUFFER REDUCE
+    | NDim: 2
+    | NOp: 2
+    | IterSize: 50
+    | IterStart: 0
+    | IterEnd: 50
+    | IterIndex: 0
+    | Iterator SizeOf:
+    | BufferData SizeOf:
+    | AxisData SizeOf:
+    |
+    | Perm: 0 1
+    | DTypes:
+    | DTypes: dtype('float64') dtype('int32')
+    | InitDataPtrs:
+    | BaseOffsets: 0 0
+    | Ptrs:
+    | User/buffer ptrs:
+    | Operands:
+    | Operand DTypes: dtype('int64') dtype('float64')
+    | OpItFlags:
+    |   Flags[0]: READ CAST
+    |   Flags[1]: READ WRITE CAST REDUCE
+    |
+    | BufferData:
+    |   BufferSize: 50
+    |   Size: 5
+    |   BufIterEnd: 5
+    |   BUFFER CoreSize: 5
+    |   REDUCE Pos: 0
+    |   REDUCE OuterSize: 10
+    |   REDUCE OuterDim: 1
+    |   Strides: 8 4
+    |   REDUCE Outer Strides: 40 0
+    |   REDUCE Outer Ptrs:
+    |   ReadTransferFn:
+    |   ReadTransferData:
+    |   WriteTransferFn:
+    |   WriteTransferData:
+    |   Buffers:
+    |
+    | AxisData[0]:
+    |   Shape: 5
+    |   Index: 0
+    |   Strides: 16 8
+    | AxisData[1]:
+    |   Shape: 10
+    |   Index: 0
+    |   Strides: 80 0
+    ------- END ITERATOR DUMP -------
+    """.strip().splitlines()
+
+    arr1 = np.arange(100, dtype=np.int64).reshape(10, 10)[:, ::2]
+    arr2 = np.arange(5.)
+    it = np.nditer((arr1, arr2), op_dtypes=["d", "i4"], casting="unsafe",
+                   flags=["reduce_ok", "buffered"],
+                   op_flags=[["readonly"], ["readwrite"]])
+    it.debug_print()
+    res = capfd.readouterr().out
+    res = res.strip().splitlines()
+
+    assert len(res) == len(expected)
+    for res_line, expected_line in zip(res, expected):
+        # The actual output may have additional pointers listed that are
+        # stripped from the example output:
+        assert res_line.startswith(expected_line.strip())
+
+
+@pytest.mark.skipif(sys.flags.optimize == 2, reason="Python running -OO")
+@pytest.mark.xfail(IS_PYPY, reason="PyPy does not modify tp_doc")
+def test_signature_constructor():
+    sig = inspect.signature(np.nditer)
+
+    assert sig.parameters
+    assert "self" not in sig.parameters
+    assert "args" not in sig.parameters
+    assert "kwargs" not in sig.parameters
+
+
+@pytest.mark.skipif(sys.flags.optimize == 2, reason="Python running -OO")
+@pytest.mark.xfail(IS_PYPY, reason="PyPy does not modify tp_doc")
+@pytest.mark.parametrize(
+    "method",
+    [fn for name, fn in vars(np.nditer).items() if callable(fn) and name[0] != "_"],
+)
+def test_signature_methods(method):
+    sig = inspect.signature(method)
+
+    assert "self" in sig.parameters
+    assert sig.parameters["self"].kind is inspect.Parameter.POSITIONAL_ONLY