BUG: Do not rely on undefined behaviour to cast from float to datetime

A cast from float to signed integer is only valid for the range of
integer, which means that a cast of a NaN to int is undefined.  On x86
cvttd2si conveniently seems to put the mantissa into the long (I
haven't done an exhaustive check to see if that is what it does, but
seems that way), which works for conversion from NaN to NaT.  On
aarch64 however, the corresponding fcvtzs sets the value to 0 for all
undefined ranges, thus breaking the cast.

This explicitly checks for NaN for conversions from floating point
types to DATETIME and TIMEDELTA and sets the target value as NaT,
which should make the behaviour well defined for NaN to NaT
translations.

2 files changed, 44 insertions, 2 deletions

diff --git a/numpy/core/src/multiarray/arraytypes.c.src b/numpy/core/src/multiarray/arraytypes.c.src
index 5d9e990e8..9639646e2 100644
--- a/numpy/core/src/multiarray/arraytypes.c.src
+++ b/numpy/core/src/multiarray/arraytypes.c.src
@@ -1081,6 +1081,7 @@ TIMEDELTA_setitem(PyObject *op, void *ov, void *vap)
  *           npy_long, npy_ulong, npy_longlong, npy_ulonglong,
  *           npy_float, npy_double, npy_longdouble,
  *           npy_datetime, npy_timedelta#
+ * #supports_nat = 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1#
  */
 
 /**begin repeat1
@@ -1092,6 +1093,7 @@ TIMEDELTA_setitem(PyObject *op, void *ov, void *vap)
  *             npy_long, npy_ulong, npy_longlong, npy_ulonglong,
  *             npy_float, npy_double, npy_longdouble,
  *             npy_datetime, npy_timedelta#
+ * #floatingpoint = 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0#
  */
 static void
 @FROMTYPE@_to_@TOTYPE@(void *input, void *output, npy_intp n,
@@ -1101,7 +1103,15 @@ static void
     @totype@ *op = output;
 
     while (n--) {
-        *op++ = (@totype@)*ip++;
+        @fromtype@ f = *ip++;
+        @totype@ t = (@totype@)f;
+#if @supports_nat@ && @floatingpoint@
+        /* Avoid undefined behaviour for NaN -> NaT */
+        if (npy_isnan(f)) {
+            t = (@totype@)NPY_DATETIME_NAT;
+        }
+#endif
+        *op++ = t;
     }
 }
 /**end repeat1**/
@@ -1119,7 +1129,15 @@ static void
     @totype@ *op = output;
 
     while (n--) {
-        *op++ = (@totype@)*ip;
+        @fromtype@ f = *ip;
+        @totype@ t = (@totype@)f;
+#if @supports_nat@
+        /* Avoid undefined behaviour for NaN -> NaT */
+        if (npy_isnan(f)) {
+            t = (@totype@)NPY_DATETIME_NAT;
+        }
+#endif
+        *op++ = t;
         ip += 2;
     }
 }
diff --git a/numpy/core/tests/test_datetime.py b/numpy/core/tests/test_datetime.py
index f99c0f72b..c60b80edc 100644
--- a/numpy/core/tests/test_datetime.py
+++ b/numpy/core/tests/test_datetime.py
@@ -483,6 +483,30 @@ class TestDateTime(object):
         assert_equal(np.datetime64(a, '[Y]'), np.datetime64('NaT', '[Y]'))
         assert_equal(np.datetime64(a, '[W]'), np.datetime64('NaT', '[W]'))
 
+        # NaN -> NaT
+        nan = np.array([np.nan] * 8)
+        fnan = nan.astype('f')
+        lnan = nan.astype('g')
+        cnan = nan.astype('D')
+        cfnan = nan.astype('F')
+        clnan = nan.astype('G')
+
+        nat = np.array([np.datetime64('NaT')] * 8)
+        assert_equal(nan.astype('M8[ns]'), nat)
+        assert_equal(fnan.astype('M8[ns]'), nat)
+        assert_equal(lnan.astype('M8[ns]'), nat)
+        assert_equal(cnan.astype('M8[ns]'), nat)
+        assert_equal(cfnan.astype('M8[ns]'), nat)
+        assert_equal(clnan.astype('M8[ns]'), nat)
+
+        nat = np.array([np.timedelta64('NaT')] * 8)
+        assert_equal(nan.astype('timedelta64[ns]'), nat)
+        assert_equal(fnan.astype('timedelta64[ns]'), nat)
+        assert_equal(lnan.astype('timedelta64[ns]'), nat)
+        assert_equal(cnan.astype('timedelta64[ns]'), nat)
+        assert_equal(cfnan.astype('timedelta64[ns]'), nat)
+        assert_equal(clnan.astype('timedelta64[ns]'), nat)
+
     def test_days_creation(self):
         assert_equal(np.array('1599', dtype='M8[D]').astype('i8'),
                 (1600-1970)*365 - (1972-1600)/4 + 3 - 365)