BUG#15448 (group_min_max test failure): Don't bypass group-min-max optimizer entry point

if we've got a tree of type ALWAYS or MAYBE.


sql/opt_range.cc:
  BUG#15448 (group_min_max test failure): Don't buypass group-min-max optimizer entry point 
  if we've got a tree of type ALWAYS or MAYBE.
  The effect observed by this bug has been introduced by this cset: 
  ChangeSet 1.1967.5.2 2005/12/26 08:40:09 sergefp@mysql.com. The cset introduced PARAM::remove_jump_scans
  which purpose was to avoid making [always unsuccessful] attempts to construct index_merge for construct 
  index_merge for conditions like "t.key1partX=c1 OR t.key1partY=c2". The effect of this change was that some
  queries that previously had produced a SEL_TREE representing candidate for index_merge(X,Y) now started to 
  produce SEL_TREE with type=ALWAYS, and group-min-max analysis for these trees has been bypassed.
  In this fix we make SEL_TREE with type=ALWAYS to be processed in the same way as (SEL_TREE*)NULL, its equivalent.

sql/opt_range.cc

@@ -1951,9 +1951,12 @@ int SQL_SELECT::test_quick_select(THD *thd, key_map keys_to_use,
           read_time= (double) HA_POS_ERROR;
           goto free_mem;
         }
-        if (tree->type != SEL_TREE::KEY &&
-            tree->type != SEL_TREE::KEY_SMALLER)
-          goto free_mem;
+        /*
+          If the tree can't be used for range scans, proceed anyway, as we
+          can construct a group-min-max quick select
+        */
+        if (tree->type != SEL_TREE::KEY && tree->type != SEL_TREE::KEY_SMALLER)
+          tree= NULL;
       }
     }