+typedef uint8_t v32 __attribute__ ((__vector_size__ (32)));
+void * cpy128(void * to, void * from) {
+  v32 * vto = to, * vfrom = from;
+  *vto++ = *vfrom++;
+  *vto++ = *vfrom++;
+  *vto++ = *vfrom++;
+  *vto++ = *vfrom++;
+  return to;
+}
+
 static void
 runtest(struct part *p, struct test *t)
 {
@@ -232,9 +230,13 @@ runtest(struct part *p, struct test *t)
        for (k = 0; k < passes; k++) {
                fprintf(stderr, ".");

+               uint64_t i = rounds, len = strlen(t->string) + 1;
+               char buff[len + 4096];
+               char string[len + 4096]; memcpy(string, t->string, len);
                gettimeofday(&begin, NULL);
-               for (i = 0; i < rounds; i++)
-                       go(p, t->string, t->needle);
+               do {
+                 p->f(cpy128(buff, string), t->needle);
+               } while(--i);
                gettimeofday(&end, NULL);

                time = (end.tv_sec - begin.tv_sec) * 1000.0;