Index: Lib/test/test_re.py
===================================================================
--- Lib/test/test_re.py	(révision 76112)
+++ Lib/test/test_re.py	(copie de travail)
@@ -716,6 +716,17 @@
         self.assertRaises(ValueError, re.compile, '(?u)\w', re.ASCII)
         self.assertRaises(ValueError, re.compile, '(?a)\w', re.UNICODE)
         self.assertRaises(ValueError, re.compile, '(?au)\w')
+    
+    def test_bug_6509(self):
+        # Replacement strings of both types must parse properly.
+        pat = re.compile('a(\w)')
+        self.assertEqual(pat.sub('b\\1', 'ac'), 'bc')
+        pat = re.compile('a(.)')
+        self.assertEqual(pat.sub('b\\1', 'a\u1234'), 'b\u1234')
+        pat = re.compile(b'a(\w)')
+        self.assertEqual(pat.sub(b'b\\1', b'ac'), b'bc')
+        pat = re.compile(b'a(.)')
+        self.assertEqual(pat.sub(b'b\\1', b'a\xCD'), b'b\xCD')
 
 
 def run_re_tests():
Index: Lib/sre_parse.py
===================================================================
--- Lib/sre_parse.py	(révision 76112)
+++ Lib/sre_parse.py	(copie de travail)
@@ -786,12 +786,18 @@
     groups = []
     groupsappend = groups.append
     literals = [None] * len(p)
+    if isinstance(source, str):
+        encode = lambda x: x
+    else:
+        # The tokenizer implicitly decodes bytes objects as latin-1, we must
+        # therefore re-encode the final representation.
+        encode = lambda x: x.encode('latin1')
     for c, s in p:
         if c is MARK:
             groupsappend((i, s))
             # literal[i] is already None
         else:
-            literals[i] = s
+            literals[i] = encode(s)
         i = i + 1
     return groups, literals