Alternative Cell.__hash__ implementation for NaN support (#4852)

* Alternative Cell.__hash__ implementation for NaN support

* add whatsnew entry

Author: bjlittle

docs/src/whatsnew/latest.rst

@@ -247,6 +247,9 @@ This document explains the changes made to Iris for this release
 
 #. `@rcomer`_ removed some now redundant testing methods. (:pull:`4838`)
 
+#. `@bjlittle`_ extended the GitHub Continuous-Integration to cover testing
+   on ``py38``, ``py39``, and ``py310``. (:pull:`4840` and :pull:`4852`)
+
 
 .. comment
     Whatsnew author names (@github name) in alphabetical order. Note that,

lib/iris/coords.py

@@ -1266,7 +1266,13 @@ def _get_2d_coord_bound_grid(bounds):
     return result
 
 
-class Cell(namedtuple("Cell", ["point", "bound"])):
+class _Hash:
+    """Mutable hash property"""
+
+    slots = ("_hash",)
+
+
+class Cell(namedtuple("Cell", ["point", "bound"]), _Hash):
     """
     An immutable representation of a single cell of a coordinate, including the
     sample point and/or boundary position.
@@ -1326,6 +1332,18 @@ def __new__(cls, point=None, bound=None):
 
         return super().__new__(cls, point, bound)
 
+    def __init__(self, *args, **kwargs):
+        # Pre-compute the hash value of this instance at creation time based
+        # on the Cell.point alone. This results in a significant performance
+        # gain, as Cell.__hash__ is reduced to a minimalist attribute lookup
+        # for each invocation.
+        try:
+            value = 0 if np.isnan(self.point) else hash((self.point,))
+        except TypeError:
+            # Passing a string to np.isnan causes this exception.
+            value = hash((self.point,))
+        self._hash = value
+
     def __mod__(self, mod):
         point = self.point
         bound = self.bound
@@ -1346,23 +1364,19 @@ def __add__(self, mod):
 
     def __hash__(self):
         # Required for >py39 and >np1.22.x due to changes in Cell behaviour for
-        # point=np.nan, as calling super().__hash__() returns a different
-        # hash and thus does not trigger the following call to Cell.__eq__
-        # to determine equality.
-        # Note that, no explicit Cell bound nan check is performed here.
+        # Cell.point=np.nan, as calling super().__hash__() returns a different
+        # hash each time and thus does not trigger the following call to
+        # Cell.__eq__ to determine equality.
+        # Note that, no explicit Cell.bound nan check is performed here.
         # That is delegated to Cell.__eq__ instead. It's imperative we keep
         # Cell.__hash__ light-weight to minimise performance degradation.
+        # Also see Cell.__init__ for Cell._hash assignment.
         # Reference:
         #   - https://bugs.python.org/issue43475
         #   - https://github.com/numpy/numpy/issues/18833
         #   - https://github.com/numpy/numpy/pull/18908
         #   - https://github.com/numpy/numpy/issues/21210
-
-        try:
-            point = "nan" if np.isnan(self.point) else self.point
-        except TypeError:
-            point = self.point
-        return hash((point,))
+        return self._hash
 
     def __eq__(self, other):
         """

lib/iris/tests/unit/coords/test_Cell.py

@@ -155,6 +155,24 @@ def test_nan_other(self):
         self.assertEqual(cell1, cell2)
 
 
+class Test___hash__(tests.IrisTest):
+    def test_nan__hash(self):
+        cell = Cell(np.nan, None)
+        self.assertEqual(cell._hash, 0)
+
+    def test_nan___hash__(self):
+        cell = Cell(np.nan, None)
+        self.assertEqual(cell.__hash__(), 0)
+
+    def test_non_nan__hash(self):
+        cell = Cell(1, None)
+        self.assertNotEqual(cell._hash, 0)
+
+    def test_non_nan___hash__(self):
+        cell = Cell("two", ("one", "three"))
+        self.assertNotEqual(cell.__hash__(), 0)
+
+
 class Test_contains_point(tests.IrisTest):
     def test_datetimelike_bounded_cell(self):
         point = object()