InnoDB row operations could be faster

While investigating MDEV-33533, I noticed that the function ha_innobase::rnd_pos() is duplicating a ut_a() assertion is also at the start of the invoked function ha_innobase::index_next(). For checking the assertion condition, two non-inlined function calls are needed. I think that we should enable this assertion only in debug builds, and retain it in a few places only, such as ha_innobase::change_active_index(). There are also other occurrences of this assertion:

@@ -9179,7 +9179,7 @@ void

 ha_innobase::try_semi_consistent_read(bool yes)

 /*===========================================*/

 {

-	ut_a(m_prebuilt->trx == thd_to_trx(ha_thd()));

+	DBUG_ASSERT(m_prebuilt->trx == thd_to_trx(ha_thd()));

 	/* Row read type is set to semi consistent read if this was

 	requested by the MySQL and either innodb_locks_unsafe_for_binlog

@@ -9349,8 +9349,7 @@ ha_innobase::index_read(

 	DBUG_ENTER("index_read");

 	DEBUG_SYNC_C("ha_innobase_index_read_begin");

-	ut_a(m_prebuilt->trx == thd_to_trx(m_user_thd));

-	ut_ad(key_len != 0 || find_flag != HA_READ_KEY_EXACT);

+	DBUG_ASSERT(m_prebuilt->trx == thd_to_trx(m_user_thd));

 	dict_index_t*	index = m_prebuilt->index;

Furthermore, the function ha_innobase::index_read() is executing two sets of tests on find_flag, while one switch (find_flag) could perform the same more efficiently. There is also some duplicated code at the end of the function, for the error handling of row_search_mvcc().