Use explicitly declared types with branching

Author: spencerkclark

pandas/_libs/tslibs/conversion.pyx

@@ -185,6 +185,8 @@ cdef int64_t cast_from_unit(
         int64_t m
         int p
         NPY_DATETIMEUNIT in_reso
+        int64_t frac_int64
+        float64_t frac_float64
 
     if unit in ["Y", "M"]:
         if is_float_object(ts) and not ts.is_integer():
@@ -220,24 +222,29 @@ cdef int64_t cast_from_unit(
             f"cannot convert input {ts} with the unit '{unit}'"
         ) from err
 
-    # In the event that ts itself is already an integer type, cast it to a
-    # 64-bit integer to ensure that frac will also be a 64-bit integer after
-    # type promotion changes in NEP 50. If this is not done, for example,
-    # np.int32 values for ts can lead to np.int32 values for frac, which can
-    # raise unexpected overflow errors downstream (GH 56996).
+    # To maintain consistency across changes in NEP 50, handle the integer and
+    # float cases separately, using frac variables with explicitly declared
+    # types. See GH 56996 and 57984 for more discussion.
     if isinstance(ts, np.integer):
-        ts = <int64_t>ts
-
-    frac = ts - base
-    if p:
-        frac = round(frac, p)
-
-    try:
-        return <int64_t>(base * m) + <int64_t>(frac * m)
-    except OverflowError as err:
-        raise OutOfBoundsDatetime(
-            f"cannot convert input {ts} with the unit '{unit}'"
-        ) from err
+        frac_int64 = ts - base
+        if p:
+            frac_int64 = round(frac_int64, p)
+        try:
+            return <int64_t>(base * m) + <int64_t>(frac_int64 * m)
+        except OverflowError as err:
+            raise OutOfBoundsDatetime(
+                f"cannot convert input {ts} with the unit '{unit}'"
+            ) from err
+    else:
+        frac_float64 = ts - base
+        if p:
+            frac_float64 = round(frac_float64, p)
+        try:
+            return <int64_t>(base * m) + <int64_t>(frac_float64 * m)
+        except OverflowError as err:
+            raise OutOfBoundsDatetime(
+                f"cannot convert input {ts} with the unit '{unit}'"
+            ) from err
 
 
 cdef (int64_t, int) precision_from_unit(

pandas/tests/tools/test_to_timedelta.py

@@ -284,12 +284,18 @@ def test_to_timedelta_fraction(self):
         expected = pd.Timedelta("0 days 00:19:59.999999998")
         assert result == expected
 
-    @pytest.mark.parametrize("dtype", [np.int32, np.uint32])
-    def test_to_timedelta_scalar_integer(self, dtype):
+    @pytest.mark.parametrize(
+        ("dtype", "value", "expected"),
+        [
+            (np.int32, 1, "1 days 00:00:00"),
+            (np.uint32, 1, "1 days 00:00:00"),
+            (np.float32, 1.1, "1 days 02:24:00.002059940"),
+        ],
+    )
+    def test_to_timedelta_following_nep_50(self, dtype, value, expected):
         # GH 56996
-        result = to_timedelta(dtype(1), "D")
-        expected = pd.Timedelta("1 days 00:00:00")
-        assert result == expected
+        result = to_timedelta(dtype(value), "D")
+        assert result == pd.Timedelta(expected)
 
 
 def test_from_numeric_arrow_dtype(any_numeric_ea_dtype):