diff --git a/Lib/test/test_codeccallbacks.py b/Lib/test/test_codeccallbacks.py
--- a/Lib/test/test_codeccallbacks.py
+++ b/Lib/test/test_codeccallbacks.py
@@ -76,6 +76,18 @@
             b"スパモ \xe4nd eggs"
         )
 
+    def test_test_xmlcharrefreplace_with_surrogates(self):
+        strings = ['\U0001f49d', '\ud83d\udc9d', '\ud83d', '\udc9d']
+        expected = [b'💝', b'💝', b'�', b'�']
+        for encoding in ['ascii', 'latin1']:
+            for s, exp in zip(strings, expected):
+                self.assertEqual(s.encode(encoding, 'xmlcharrefreplace'),
+                                 exp)
+                self.assertEqual((s+'X').encode(encoding, 'xmlcharrefreplace'),
+                                 exp+b'X')
+
+
+
     def test_xmlcharnamereplace(self):
         # This time use a named character entity for unencodable
         # characters, if one is available.
diff --git a/Objects/unicodeobject.c b/Objects/unicodeobject.c
--- a/Objects/unicodeobject.c
+++ b/Objects/unicodeobject.c
@@ -4707,7 +4707,17 @@
                 }
                 /* generate replacement (temporarily (mis)uses p) */
                 for (p = collstart; p < collend; ++p) {
-                    str += sprintf(str, "&#%d;", (int)*p);
+                    if ((0xD800 <= *p && *p <= 0xDBFF) &&
+                        (p+1 < collend) &&
+                        (0xDC00 <= *(p+1) && *(p+1) <= 0xDFFF)) {
+                        Py_UNICODE fst = *p;
+                        Py_UCS4 ch = (((((Py_UCS4)fst & 0x03FF) << 10) |
+                                       ((Py_UCS4)(*(++p)) & 0x03FF)) + 0x10000);
+                        str += sprintf(str, "&#%d;", (int)ch);
+                    }
+                    else {
+                        str += sprintf(str, "&#%d;", (int)*p);
+                    }
                 }
                 p = collend;
                 break;