Improve performance of startswith by eliminating double work in tailmatch

eendebakpt · eendebakpt · commit 035b3e23c7c3 · 2024-04-11T21:40:12.000+02:00
diff --git a/Objects/unicodeobject.c b/Objects/unicodeobject.c
@@ -9272,24 +9272,23 @@ tailmatch(PyObject *self,
     else
         offset = start;
 
-    if (PyUnicode_READ(kind_self, data_self, offset) ==
-        PyUnicode_READ(kind_sub, data_sub, 0) &&
-        PyUnicode_READ(kind_self, data_self, offset + end_sub) ==
-        PyUnicode_READ(kind_sub, data_sub, end_sub)) {
+    int last_character_matches = PyUnicode_READ(kind_self, data_self, offset + end_sub) ==
+        PyUnicode_READ(kind_sub, data_sub, end_sub);
+
+    if (last_character_matches) {
+        if (end_sub==0)
+            return 1;
         /* If both are of the same kind, memcmp is sufficient */
         if (kind_self == kind_sub) {
-            return ! memcmp((char *)data_self +
-                                (offset * PyUnicode_KIND(substring)),
-                            data_sub,
-                            PyUnicode_GET_LENGTH(substring) *
-                                PyUnicode_KIND(substring));
+            return ! memcmp((char *)data_self + (offset * kind_sub),
+                            data_sub, end_sub * kind_sub);
         }
         /* otherwise we have to compare each character by first accessing it */
         else {
             /* We do not need to compare 0 and len(substring)-1 because
                the if statement above ensured already that they are equal
                when we end up here. */
-            for (i = 1; i < end_sub; ++i) {
+            for (i = 0; i < end_sub; ++i) {
                 if (PyUnicode_READ(kind_self, data_self, offset + i) !=
                     PyUnicode_READ(kind_sub, data_sub, i))
                     return 0;
