BUG: Fix out-of-bound minimum offset for in1d table method

We could always use ar2.dtype, but we want intp anyway, so that
is nicer.  But if the minimum value doesn't fit intp, NumPy would
complain (it may even guaranteed to work to force-cast,
but...)

So just try to use intp, but if that doesn't work use ar2.dtype
which must work, but means we'll have to cast once more (even if
that cast is probably just a view in principle).

Closes numpygh-26922

Author: seberg

numpy/lib/_arraysetops_impl.py

@@ -915,8 +915,25 @@ def _in1d(ar1, ar2, assume_unique=False, invert=False, *, kind=None):
 
             # Mask out elements we know won't work
             basic_mask = (ar1 <= ar2_max) & (ar1 >= ar2_min)
+            in_range_ar1 = ar1[basic_mask]
+            if in_range_ar1.size == 0:
+                # Nothing more to do, since all values are out of range.
+                return outgoing_array
+
+            # Unfortunately, ar2_min can be out of range for `intp` even
+            # if the calculation result must fit in range (and be positive).
+            # In that case, use ar2.dtype which must work for all unmasked
+            # values.
+            try:
+                ar2_min = np.array(ar2_min, dtype=np.intp)
+                dtype = np.intp
+            except OverflowError:
+                dtype = ar2.dtype
+
+            out = np.empty_like(in_range_ar1, dtype=np.intp)
             outgoing_array[basic_mask] = isin_helper_ar[
-                    np.subtract(ar1[basic_mask], ar2_min, dtype=np.intp)]
+                    np.subtract(in_range_ar1, ar2_min, dtype=dtype,
+                                out=out, casting="unsafe")]
 
             return outgoing_array
         elif kind == 'table':  # not range_safe_from_overflow

numpy/lib/tests/test_arraysetops.py

@@ -440,6 +440,22 @@ def test_isin_mixed_dtype(self, dtype1, dtype2, kind):
         else:
             assert_array_equal(isin(ar1, ar2, kind=kind), expected)
 
+    @pytest.mark.parametrize("data", [
+        np.array([2**63, 2**63+1], dtype=np.uint64),
+        np.array([-2**62, -2**62-1], dtype=np.int64),
+    ])
+    @pytest.mark.parametrize("kind", [None, "sort", "table"])
+    def test_isin_mixed_huge_vals(self, kind, data):
+        """Test values outside intp range (negative ones if 32bit system)"""
+        query = data[1]
+        res = np.isin(data, query, kind=kind)
+        assert_array_equal(res, [False, True])
+        # Also check that nothing weird happens for values can't possibly
+        # in range.
+        data = data.astype(np.int32)  # clearly different values
+        res = np.isin(data, query, kind=kind)
+        assert_array_equal(res, [False, False])
+
     @pytest.mark.parametrize("kind", [None, "sort", "table"])
     def test_isin_mixed_boolean(self, kind):
         """Test that isin works as expected for bool/int input."""