diff -r 9304172ee5eb Objects/unicodeobject.c
--- a/Objects/unicodeobject.c	Thu Mar 29 09:18:14 2012 +0200
+++ b/Objects/unicodeobject.c	Fri Mar 30 11:42:18 2012 +0300
@@ -5393,9 +5393,11 @@
 #if (SIZEOF_LONG == 8)
 # define FAST_CHAR_MASK         0x8000800080008000L
 # define SWAPPED_FAST_CHAR_MASK 0x0080008000800080L
+# define STRIPPED_MASK          0x00FF00FF00FF00FFL
 #elif (SIZEOF_LONG == 4)
 # define FAST_CHAR_MASK         0x80008000L
 # define SWAPPED_FAST_CHAR_MASK 0x00800080L
+# define STRIPPED_MASK          0x00FF00FFL
 #else
 # error C 'long' size should be either 4 or 8!
 #endif
@@ -5497,7 +5499,6 @@
             void *data = PyUnicode_DATA(unicode);
             while (_q < aligned_end) {
                 unsigned long block = * (unsigned long *) _q;
-                unsigned short *pblock = (unsigned short*)&block;
                 Py_UCS4 maxch;
                 if (native_ordering) {
                     /* Can use buffer directly */
@@ -5506,23 +5507,17 @@
                 }
                 else {
                     /* Need to byte-swap */
-                    unsigned char *_p = (unsigned char*)pblock;
                     if (block & SWAPPED_FAST_CHAR_MASK)
                         break;
-                    _p[0] = _q[1];
-                    _p[1] = _q[0];
-                    _p[2] = _q[3];
-                    _p[3] = _q[2];
-#if (SIZEOF_LONG == 8)
-                    _p[4] = _q[5];
-                    _p[5] = _q[4];
-                    _p[6] = _q[7];
-                    _p[7] = _q[6];
-#endif
-                }
-                maxch = Py_MAX(pblock[0], pblock[1]);
+                    block = ((block >> 8) & STRIPPED_MASK) |
+                            ((block & STRIPPED_MASK) << 8);
+                }
+                maxch = Py_MAX((Py_UCS2)(block & 0xFFFFu),
+                               (Py_UCS2)((block >> 16) & 0xFFFFu));
 #if SIZEOF_LONG == 8
-                maxch = Py_MAX(maxch, Py_MAX(pblock[2], pblock[3]));
+                maxch = Py_MAX(maxch,
+                        Py_MAX((Py_UCS2)((block >> 32) & 0xFFFFu),
+                               (Py_UCS2)((block >> 48) & 0xFFFFu)));
 #endif
                 if (maxch > PyUnicode_MAX_CHAR_VALUE(unicode)) {
                     if (unicode_widen(&unicode, maxch) < 0)
@@ -5530,11 +5525,20 @@
                     kind = PyUnicode_KIND(unicode);
                     data = PyUnicode_DATA(unicode);
                 }
-                PyUnicode_WRITE(kind, data, outpos++, pblock[0]);
-                PyUnicode_WRITE(kind, data, outpos++, pblock[1]);
+#ifdef BYTEORDER_IS_LITTLE_ENDIAN
+                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)(block & 0xFFFFu));
+                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 16) & 0xFFFFu));
 #if SIZEOF_LONG == 8
-                PyUnicode_WRITE(kind, data, outpos++, pblock[2]);
-                PyUnicode_WRITE(kind, data, outpos++, pblock[3]);
+                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 32) & 0xFFFFu));
+                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 48) & 0xFFFFu));
+#endif
+#else
+#if SIZEOF_LONG == 8
+                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 48) & 0xFFFFu));
+                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 32) & 0xFFFFu));
+#endif
+                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 16) & 0xFFFFu));
+                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)(block & 0xFFFFu));
 #endif
                 _q += SIZEOF_LONG;
             }