The previous correction of the cost estimate to access a joined table

in the function best_access_path revealed another bug: currently 
table scans on NULL keys used for NOT IN subqueries cannot work 
together with employment of join caches for inner tables of these 
subqueries. Otherwise the result can be wrong as it could be seen 
with the result of the test case constructed for bug #37894 
in the file subselect3_jcl6.result.

mysql-test/r/subselect3_jcl6.result

@@ -822,8 +822,6 @@ LEFT JOIN t1 AS INNR ON ( INNR2.int_key = INNR.int_key )
 WHERE INNR.varchar_key > 'n{'
 );
 varchar_nokey
-NULL
-p
 DROP TABLE t1;
 CREATE TABLE t1 (a INT);
 INSERT INTO t1 VALUES (1), (2), (11);

sql/sql_select.cc

@@ -9081,6 +9081,11 @@ uint check_join_cache_usage(JOIN_TAB *tab,
   case JT_EQ_REF:
     if (cache_level <=2 || (no_hashed_cache && no_bka_cache))
       goto no_join_cache;
+    for (uint i= 0; i < tab->ref.key_parts; i++)
+    {
+      if (tab->ref.cond_guards[i])
+        goto no_join_cache;
+    }
     if (!tab->is_ref_for_hash_join())
     {
       flags= HA_MRR_NO_NULL_ENDPOINTS | HA_MRR_SINGLE_POINT;