diff -r 7b420953418c Objects/bytesobject.c
--- a/Objects/bytesobject.c	Sun Jul 27 00:27:04 2008 +0200
+++ b/Objects/bytesobject.c	Mon Jul 28 20:48:53 2008 +0200
@@ -1433,10 +1433,10 @@ string_join(PyObject *self, PyObject *or
 	const Py_ssize_t seplen = PyBytes_GET_SIZE(self);
 	PyObject *res = NULL;
 	char *p;
-	Py_ssize_t seqlen = 0;
+	Py_ssize_t seqlen = 0, nb_nonempty = 0;
 	size_t sz = 0;
 	Py_ssize_t i;
-	PyObject *seq, *item;
+	PyObject *seq, *item, *nonempty = NULL;
 
 	seq = PySequence_Fast(orig, "");
 	if (seq == NULL) {
@@ -1465,6 +1465,7 @@ string_join(PyObject *self, PyObject *or
 	/* XXX Shouldn't we use _getbuffer() on these items instead? */
 	for (i = 0; i < seqlen; i++) {
 		const size_t old_sz = sz;
+		size_t item_sz;
 		item = PySequence_Fast_GET_ITEM(seq, i);
 		if (!PyBytes_Check(item) && !PyByteArray_Check(item)) {
 			PyErr_Format(PyExc_TypeError,
@@ -1474,7 +1475,12 @@ string_join(PyObject *self, PyObject *or
 			Py_DECREF(seq);
 			return NULL;
 		}
-		sz += Py_SIZE(item);
+		item_sz = Py_SIZE(item);
+		if (item_sz > 0) {
+			nonempty = item;
+			nb_nonempty++;
+			sz += item_sz;
+		}
 		if (i != 0)
 			sz += seplen;
 		if (sz < old_sz || sz > PY_SSIZE_T_MAX) {
@@ -1482,6 +1488,20 @@ string_join(PyObject *self, PyObject *or
 			    "join() result is too long for bytes");
 			Py_DECREF(seq);
 			return NULL;
+		}
+	}
+
+	/* Shortcut when the separator is empty and there is only zero or one
+       non-empty item */
+	if (seplen == 0) {
+		if (nb_nonempty == 0) {
+			Py_DECREF(seq);
+			return PyBytes_FromString("");
+		}
+		if (nb_nonempty == 1 && PyBytes_CheckExact(nonempty)) {
+			Py_INCREF(nonempty);
+			Py_DECREF(seq);
+			return nonempty;
 		}
 	}