~drizzle-trunk/drizzle/development : contents of drizzled/sql

~drizzle-trunk/drizzle/development : (revision 1022.2.34)
/* Copyright (C) 2000-2006 MySQL AB

   This program is free software; you can redistribute it and/or modify
   it under the terms of the GNU General Public License as published by
   the Free Software Foundation; version 2 of the License.

   This program is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   GNU General Public License for more details.

   You should have received a copy of the GNU General Public License
   along with this program; if not, write to the Free Software
   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA */

/**
  @file

  @brief
  mysql_select and join optimization


  @defgroup Query_Optimizer  Query Optimizer
  @{
*/
#include "drizzled/server_includes.h"
#include "drizzled/sql_select.h"
#include "drizzled/sj_tmp_table.h"
#include "drizzled/table_map_iterator.h"

#include "mysys/my_bit.h"
#include "drizzled/error.h"
#include "drizzled/gettext.h"
#include "drizzled/util/test.h"
#include "drizzled/name_resolution_context_state.h"
#include "drizzled/nested_join.h"
#include "drizzled/probes.h"
#include "drizzled/show.h"
#include "drizzled/item/cache.h"
#include "drizzled/item/cmpfunc.h"
#include "drizzled/item/copy_string.h"
#include "drizzled/item/uint.h"
#include "drizzled/cached_item.h"
#include "drizzled/sql_base.h"
#include "drizzled/field/blob.h"
#include "drizzled/check_stack_overrun.h"
#include "drizzled/lock.h"
#include "drizzled/item/outer_ref.h"
#include "drizzled/index_hint.h"

#include <string>
#include <iostream>

using namespace std;

const char *join_type_str[]={ "UNKNOWN","system","const","eq_ref","ref",
			      "MAYBE_REF","ALL","range","index",
			      "ref_or_null","unique_subquery","index_subquery",
                              "index_merge"
};

struct st_sargable_param;

static void optimize_keyuse(JOIN *join, DYNAMIC_ARRAY *keyuse_array);
static bool make_join_statistics(JOIN *join, TableList *leaves, COND *conds,
				 DYNAMIC_ARRAY *keyuse);
static bool update_ref_and_keys(Session *session, DYNAMIC_ARRAY *keyuse,
                                JOIN_TAB *join_tab,
                                uint32_t tables, COND *conds,
                                COND_EQUAL *cond_equal,
                                table_map table_map, Select_Lex *select_lex,
                                st_sargable_param **sargables);
static int sort_keyuse(KEYUSE *a,KEYUSE *b);
static void set_position(JOIN *join,uint32_t index,JOIN_TAB *table,KEYUSE *key);
static bool create_ref_for_key(JOIN *join, JOIN_TAB *j, KEYUSE *org_keyuse,
			       table_map used_tables);
static bool choose_plan(JOIN *join,table_map join_tables);

static void best_access_path(JOIN *join, JOIN_TAB *s, Session *session,
                             table_map remaining_tables, uint32_t idx,
                             double record_count, double read_time);
static void optimize_straight_join(JOIN *join, table_map join_tables);
static bool greedy_search(JOIN *join, table_map remaining_tables,
                             uint32_t depth, uint32_t prune_level);
static bool best_extension_by_limited_search(JOIN *join,
                                             table_map remaining_tables,
                                             uint32_t idx, double record_count,
                                             double read_time, uint32_t depth,
                                             uint32_t prune_level);
static uint32_t determine_search_depth(JOIN* join);
extern "C" int join_tab_cmp(const void* ptr1, const void* ptr2);
extern "C" int join_tab_cmp_straight(const void* ptr1, const void* ptr2);
/*
  TODO: 'find_best' is here only temporarily until 'greedy_search' is
  tested and approved.
*/
static bool find_best(JOIN *join,table_map rest_tables,uint32_t index,
		      double record_count,double read_time);
static uint32_t cache_record_length(JOIN *join,uint32_t index);
static double prev_record_reads(JOIN *join, uint32_t idx, table_map found_ref);
static bool get_best_combination(JOIN *join);
static store_key *get_store_key(Session *session,
				KEYUSE *keyuse, table_map used_tables,
				KEY_PART_INFO *key_part, unsigned char *key_buff,
				uint32_t maybe_null);
static bool make_simple_join(JOIN *join,Table *tmp_table);
static void make_outerjoin_info(JOIN *join);
static bool make_join_select(JOIN *join,SQL_SELECT *select,COND *item);
static bool make_join_readinfo(JOIN *join, uint64_t options, uint32_t no_jbuf_after);
static bool only_eq_ref_tables(JOIN *join, order_st *order, table_map tables);
static void update_depend_map(JOIN *join);
static void update_depend_map(JOIN *join, order_st *order);
static order_st *remove_constants(JOIN *join,order_st *first_order,COND *cond,
			          bool change_list, bool *simple_order);
static int return_zero_rows(JOIN *join, select_result *res,TableList *tables,
                            List<Item> &fields, bool send_row,
                            uint64_t select_options, const char *info,
                            Item *having);
static COND *build_equal_items(Session *session, COND *cond,
                               COND_EQUAL *inherited,
                               List<TableList> *join_list,
                               COND_EQUAL **cond_equal_ref);
static COND* substitute_for_best_equal_field(COND *cond,
                                             COND_EQUAL *cond_equal,
                                             void *table_join_idx);
static COND *simplify_joins(JOIN *join, List<TableList> *join_list,
                            COND *conds, bool top, bool in_sj);
static bool check_interleaving_with_nj(JOIN_TAB *last, JOIN_TAB *next);
static void restore_prev_nj_state(JOIN_TAB *last);
static void reset_nj_counters(List<TableList> *join_list);
static uint32_t build_bitmap_for_nested_joins(List<TableList> *join_list,
                                          uint32_t first_unused);

static
void advance_sj_state(const table_map remaining_tables, const JOIN_TAB *tab);
static void restore_prev_sj_state(const table_map remaining_tables,
                                  const JOIN_TAB *tab);

static COND *optimize_cond(JOIN *join, COND *conds,
                           List<TableList> *join_list,
			   Item::cond_result *cond_value);
static bool const_expression_in_where(COND *conds,Item *item, Item **comp_item);
static int do_select(JOIN *join,List<Item> *fields,Table *tmp_table);

static enum_nested_loop_state
evaluate_join_record(JOIN *join, JOIN_TAB *join_tab,
                     int error);
static enum_nested_loop_state
evaluate_null_complemented_join_record(JOIN *join, JOIN_TAB *join_tab);
static enum_nested_loop_state
flush_cached_records(JOIN *join, JOIN_TAB *join_tab, bool skip_last);
static enum_nested_loop_state
end_send(JOIN *join, JOIN_TAB *join_tab, bool end_of_records);
static enum_nested_loop_state
end_write(JOIN *join, JOIN_TAB *join_tab, bool end_of_records);
static enum_nested_loop_state
end_update(JOIN *join, JOIN_TAB *join_tab, bool end_of_records);
static enum_nested_loop_state
end_unique_update(JOIN *join, JOIN_TAB *join_tab, bool end_of_records);

static int join_read_const_table(JOIN_TAB *tab, POSITION *pos);
static int join_read_system(JOIN_TAB *tab);
static int join_read_const(JOIN_TAB *tab);
static int join_read_key(JOIN_TAB *tab);
static int join_read_always_key(JOIN_TAB *tab);
static int join_read_last_key(JOIN_TAB *tab);
static int join_no_more_records(READ_RECORD *info);
static int join_read_next(READ_RECORD *info);
static int join_read_next_different(READ_RECORD *info);
static int join_init_quick_read_record(JOIN_TAB *tab);
static int test_if_quick_select(JOIN_TAB *tab);
static int join_init_read_record(JOIN_TAB *tab);
static int join_read_first(JOIN_TAB *tab);
static int join_read_next_same(READ_RECORD *info);
static int join_read_next_same_diff(READ_RECORD *info);
static int join_read_last(JOIN_TAB *tab);
static int join_read_prev_same(READ_RECORD *info);
static int join_read_prev(READ_RECORD *info);
int join_read_always_key_or_null(JOIN_TAB *tab);
int join_read_next_same_or_null(READ_RECORD *info);
static COND *make_cond_for_table(COND *cond,table_map table,
				 table_map used_table,
                                 bool exclude_expensive_cond);
static Item* part_of_refkey(Table *form,Field *field);
static bool test_if_skip_sort_order(JOIN_TAB *tab,order_st *order,
				    ha_rows select_limit, bool no_changes,
                                    const key_map *map);
static bool list_contains_unique_index(Table *table,
                          bool (*find_func) (Field *, void *), void *data);
static bool find_field_in_item_list (Field *field, void *data);
static bool find_field_in_order_list (Field *field, void *data);
static int create_sort_index(Session *session, JOIN *join, order_st *order,
			     ha_rows filesort_limit, ha_rows select_limit,
                             bool is_order_by);
static int remove_duplicates(JOIN *join,Table *entry,List<Item> &fields,
			     Item *having);
static int remove_dup_with_compare(Session *session, Table *entry, Field **field,
				   uint32_t offset, Item *having);
static int remove_dup_with_hash_index(Session *session,Table *table,
				      uint32_t field_count, Field **first_field,
				      uint32_t key_length, Item *having);
static int join_init_cache(Session *session,JOIN_TAB *tables,uint32_t table_count);
static uint32_t used_blob_length(CACHE_FIELD **ptr);
static bool store_record_in_cache(JOIN_CACHE *cache);
static void reset_cache_read(JOIN_CACHE *cache);
static void reset_cache_write(JOIN_CACHE *cache);
static void read_cached_record(JOIN_TAB *tab);
static bool cmp_buffer_with_ref(JOIN_TAB *tab);
static order_st *create_distinct_group(Session *session, Item **ref_pointer_array,
                                    order_st *order, List<Item> &fields,
                                    List<Item> &all_fields,
				    bool *all_order_by_fields_used);
static bool test_if_subpart(order_st *a,order_st *b);
static Table *get_sort_by_table(order_st *a,order_st *b,TableList *tables);
static void calc_group_buffer(JOIN *join,order_st *group);
static bool make_group_fields(JOIN *main_join, JOIN *curr_join);
static bool alloc_group_fields(JOIN *join,order_st *group);
// Create list for using with tempory table
static bool change_to_use_tmp_fields(Session *session, Item **ref_pointer_array,
				     List<Item> &new_list1,
				     List<Item> &new_list2,
				     uint32_t elements, List<Item> &items);
// Create list for using with tempory table
static bool change_refs_to_tmp_fields(Session *session, Item **ref_pointer_array,
				      List<Item> &new_list1,
				      List<Item> &new_list2,
				      uint32_t elements, List<Item> &items);
static void init_tmptable_sum_functions(Item_sum **func);
static void update_tmptable_sum_func(Item_sum **func,Table *tmp_table);
static void copy_sum_funcs(Item_sum **func_ptr, Item_sum **end);
static bool add_ref_to_table_cond(Session *session, JOIN_TAB *join_tab);
static bool setup_sum_funcs(Session *session, Item_sum **func_ptr);
static bool init_sum_functions(Item_sum **func, Item_sum **end);
static bool update_sum_func(Item_sum **func);
void select_describe(JOIN *join, bool need_tmp_table,bool need_order,
			    bool distinct, const char *message=NULL);
static Item *remove_additional_cond(Item* conds);
static void add_group_and_distinct_keys(JOIN *join, JOIN_TAB *join_tab);
static bool test_if_ref(Item_field *left_item,Item *right_item);
static bool replace_where_subcondition(JOIN *join, Item *old_cond,
                                       Item *new_cond, bool fix_fields);

static bool eval_const_cond(COND *cond)
{
    return ((Item_func*) cond)->val_int() ? true : false;
}


/*
  This is used to mark equalities that were made from i-th IN-equality.
  We limit semi-join InsideOut optimization to handling max 64 inequalities,
  The following variable occupies 64 addresses.
*/
const char *subq_sj_cond_name=
  "0123456789ABCDEF0123456789abcdef0123456789ABCDEF0123456789abcdef-sj-cond";

static bool bitmap_covers(const table_map x, const table_map y)
{
  return !test(y & ~x);
}

/**
  This handles SELECT with and without UNION.
*/

bool handle_select(Session *session, LEX *lex, select_result *result,
                   uint64_t setup_tables_done_option)
{
  bool res;
  register Select_Lex *select_lex = &lex->select_lex;
  DRIZZLE_SELECT_START();

  if (select_lex->master_unit()->is_union() ||
      select_lex->master_unit()->fake_select_lex)
    res= mysql_union(session, lex, result, &lex->unit, setup_tables_done_option);
  else
  {
    Select_Lex_Unit *unit= &lex->unit;
    unit->set_limit(unit->global_parameters);
    session->session_marker= 0;
    /*
      'options' of mysql_select will be set in JOIN, as far as JOIN for
      every PS/SP execution new, we will not need reset this flag if
      setup_tables_done_option changed for next rexecution
    */
    res= mysql_select(session, &select_lex->ref_pointer_array,
		      (TableList*) select_lex->table_list.first,
		      select_lex->with_wild, select_lex->item_list,
		      select_lex->where,
		      select_lex->order_list.elements +
		      select_lex->group_list.elements,
		      (order_st*) select_lex->order_list.first,
		      (order_st*) select_lex->group_list.first,
		      select_lex->having,
		      select_lex->options | session->options |
                      setup_tables_done_option,
		      result, unit, select_lex);
  }
  res|= session->is_error();
  if (unlikely(res))
    result->abort();

  DRIZZLE_SELECT_END();
  return(res);
}


/*
  Fix fields referenced from inner selects.

  SYNOPSIS
    fix_inner_refs()
    session               Thread handle
    all_fields        List of all fields used in select
    select            Current select
    ref_pointer_array Array of references to Items used in current select

  DESCRIPTION
    The function serves 3 purposes - adds fields referenced from inner
    selects to the current select list, resolves which class to use
    to access referenced item (Item_ref of Item_direct_ref) and fixes
    references (Item_ref objects) to these fields.

    If a field isn't already in the select list and the ref_pointer_array
    is provided then it is added to the all_fields list and the pointer to
    it is saved in the ref_pointer_array.

    The class to access the outer field is determined by the following rules:
    1. If the outer field isn't used under an aggregate function
      then the Item_ref class should be used.
    2. If the outer field is used under an aggregate function and this
      function is aggregated in the select where the outer field was
      resolved or in some more inner select then the Item_direct_ref
      class should be used.
    The resolution is done here and not at the fix_fields() stage as
    it can be done only after sum functions are fixed and pulled up to
    selects where they are have to be aggregated.
    When the class is chosen it substitutes the original field in the
    Item_outer_ref object.

    After this we proceed with fixing references (Item_outer_ref objects) to
    this field from inner subqueries.

  RETURN
    true  an error occured
    false ok
*/

bool
fix_inner_refs(Session *session, List<Item> &all_fields, Select_Lex *select,
                 Item **ref_pointer_array)
{
  Item_outer_ref *ref;
  bool res= false;
  bool direct_ref= false;

  List_iterator<Item_outer_ref> ref_it(select->inner_refs_list);
  while ((ref= ref_it++))
  {
    Item *item= ref->outer_ref;
    Item **item_ref= ref->ref;
    Item_ref *new_ref;
    /*
      TODO: this field item already might be present in the select list.
      In this case instead of adding new field item we could use an
      existing one. The change will lead to less operations for copying fields,
      smaller temporary tables and less data passed through filesort.
    */
    if (ref_pointer_array && !ref->found_in_select_list)
    {
      int el= all_fields.elements;
      ref_pointer_array[el]= item;
      /* Add the field item to the select list of the current select. */
      all_fields.push_front(item);
      /*
        If it's needed reset each Item_ref item that refers this field with
        a new reference taken from ref_pointer_array.
      */
      item_ref= ref_pointer_array + el;
    }

    if (ref->in_sum_func)
    {
      Item_sum *sum_func;
      if (ref->in_sum_func->nest_level > select->nest_level)
        direct_ref= true;
      else
      {
        for (sum_func= ref->in_sum_func; sum_func &&
             sum_func->aggr_level >= select->nest_level;
             sum_func= sum_func->in_sum_func)
        {
          if (sum_func->aggr_level == select->nest_level)
          {
            direct_ref= true;
            break;
          }
        }
      }
    }
    new_ref= direct_ref ?
              new Item_direct_ref(ref->context, item_ref, ref->table_name,
                          ref->field_name, ref->alias_name_used) :
              new Item_ref(ref->context, item_ref, ref->table_name,
                          ref->field_name, ref->alias_name_used);
    if (!new_ref)
      return true;
    ref->outer_ref= new_ref;
    ref->ref= &ref->outer_ref;

    if (!ref->fixed && ref->fix_fields(session, 0))
      return true;
    session->used_tables|= item->used_tables();
  }
  return res;
}

/**
  Function to setup clauses without sum functions.
*/
inline int setup_without_group(Session *session, Item **ref_pointer_array,
                               TableList *tables,
                               TableList *leaves,
                               List<Item> &fields,
                               List<Item> &all_fields,
                               COND **conds,
                               order_st *order,
                               order_st *group, bool *hidden_group_fields)
{
  int res;
  nesting_map save_allow_sum_func=session->lex->allow_sum_func ;

  session->lex->allow_sum_func&= ~(1 << session->lex->current_select->nest_level);
  res= setup_conds(session, tables, leaves, conds);

  session->lex->allow_sum_func|= 1 << session->lex->current_select->nest_level;
  res= res || setup_order(session, ref_pointer_array, tables, fields, all_fields,
                          order);
  session->lex->allow_sum_func&= ~(1 << session->lex->current_select->nest_level);
  res= res || setup_group(session, ref_pointer_array, tables, fields, all_fields,
                          group, hidden_group_fields);
  session->lex->allow_sum_func= save_allow_sum_func;
  return(res);
}

/*****************************************************************************
  Check fields, find best join, do the select and output fields.
  mysql_select assumes that all tables are already opened
*****************************************************************************/

/**
  Prepare of whole select (including sub queries in future).

  @todo
    Add check of calculation of GROUP functions and fields:
    SELECT COUNT(*)+table.col1 from table1;

  @retval
    -1   on error
  @retval
    0   on success
*/
int
JOIN::prepare(Item ***rref_pointer_array,
	      TableList *tables_init,
	      uint32_t wild_num, COND *conds_init, uint32_t og_num,
	      order_st *order_init, order_st *group_init,
	      Item *having_init,
	      Select_Lex *select_lex_arg,
	      Select_Lex_Unit *unit_arg)
{
  // to prevent double initialization on EXPLAIN
  if (optimized)
    return(0);

  conds= conds_init;
  order= order_init;
  group_list= group_init;
  having= having_init;
  tables_list= tables_init;
  select_lex= select_lex_arg;
  select_lex->join= this;
  join_list= &select_lex->top_join_list;
  union_part= unit_arg->is_union();

  session->lex->current_select->is_item_list_lookup= 1;
  /*
    If we have already executed SELECT, then it have not sense to prevent
    its table from update (see unique_table())
  */
  if (session->derived_tables_processing)
    select_lex->exclude_from_table_unique_test= true;

  /* Check that all tables, fields, conds and order are ok */

  if (!(select_options & OPTION_SETUP_TABLES_DONE) &&
      setup_tables_and_check_access(session, &select_lex->context, join_list,
                                    tables_list, &select_lex->leaf_tables,
                                    false))
      return(-1);

  TableList *table_ptr;
  for (table_ptr= select_lex->leaf_tables;
       table_ptr;
       table_ptr= table_ptr->next_leaf)
    tables++;

  if (setup_wild(session, tables_list, fields_list, &all_fields, wild_num) ||
      select_lex->setup_ref_array(session, og_num) ||
      setup_fields(session, (*rref_pointer_array), fields_list, MARK_COLUMNS_READ,
		   &all_fields, 1) ||
      setup_without_group(session, (*rref_pointer_array), tables_list,
			  select_lex->leaf_tables, fields_list,
			  all_fields, &conds, order, group_list,
			  &hidden_group_fields))
    return(-1);				/* purecov: inspected */

  ref_pointer_array= *rref_pointer_array;

  if (having)
  {
    nesting_map save_allow_sum_func= session->lex->allow_sum_func;
    session->where="having clause";
    session->lex->allow_sum_func|= 1 << select_lex_arg->nest_level;
    select_lex->having_fix_field= 1;
    bool having_fix_rc= (!having->fixed &&
			 (having->fix_fields(session, &having) ||
			  having->check_cols(1)));
    select_lex->having_fix_field= 0;
    if (having_fix_rc || session->is_error())
      return(-1);				/* purecov: inspected */
    session->lex->allow_sum_func= save_allow_sum_func;
  }

  {
    Item_subselect *subselect;
    Item_in_subselect *in_subs= NULL;
    /*
      Are we in a subquery predicate?
      TODO: the block below will be executed for every PS execution without need.
    */
    if ((subselect= select_lex->master_unit()->item))
    {
      bool do_semijoin= !test(session->variables.optimizer_switch &
                              OPTIMIZER_SWITCH_NO_SEMIJOIN);
      if (subselect->substype() == Item_subselect::IN_SUBS)
        in_subs= (Item_in_subselect*)subselect;

      /*
        Check if we're in subquery that is a candidate for flattening into a
        semi-join (which is done done in flatten_subqueries()). The
        requirements are:
          1. Subquery predicate is an IN/=ANY subq predicate
          2. Subquery is a single SELECT (not a UNION)
          3. Subquery does not have GROUP BY or order_st BY
          4. Subquery does not use aggregate functions or HAVING
          5. Subquery predicate is at the AND-top-level of ON/WHERE clause
          6. No execution method was already chosen (by a prepared statement).

          (*). We are not in a subquery of a single table UPDATE/DELETE that
               doesn't have a JOIN (TODO: We should handle this at some
               point by switching to multi-table UPDATE/DELETE)

          (**). We're not in a confluent table-less subquery, like
                "SELECT 1".
      */
      if (in_subs &&                                                    // 1
          !select_lex->master_unit()->first_select()->next_select() &&  // 2
          !select_lex->group_list.elements && !order &&                 // 3
          !having && !select_lex->with_sum_func &&                      // 4
          session->session_marker &&                                            // 5
          select_lex->outer_select()->join &&                           // (*)
          select_lex->master_unit()->first_select()->leaf_tables &&     // (**)
          do_semijoin &&
          in_subs->exec_method == Item_in_subselect::NOT_TRANSFORMED)   // 6
      {
        {
          if (!in_subs->left_expr->fixed &&
               in_subs->left_expr->fix_fields(session, &in_subs->left_expr))
          {
            return(-1);
          }
          /*
            Check that the right part of the subselect contains no more than one
            column. E.g. in SELECT 1 IN (SELECT * ..) the right part is (SELECT * ...)
          */
          if (subselect->substype() == Item_subselect::IN_SUBS &&
             (select_lex->item_list.elements !=
              ((Item_in_subselect*)subselect)->left_expr->cols()))
          {
            my_error(ER_OPERAND_COLUMNS, MYF(0), ((Item_in_subselect*)subselect)->left_expr->cols());
            return(-1);
          }
        }

        /* Register the subquery for further processing */
        select_lex->outer_select()->join->sj_subselects.append(session->mem_root, in_subs);
        in_subs->expr_join_nest= (TableList*)session->session_marker;
      }
      else
      {
        bool do_materialize= !test(session->variables.optimizer_switch &
                                   OPTIMIZER_SWITCH_NO_MATERIALIZATION);
        /*
          Check if the subquery predicate can be executed via materialization.
          The required conditions are:
          1. Subquery predicate is an IN/=ANY subq predicate
          2. Subquery is a single SELECT (not a UNION)
          3. Subquery is not a table-less query. In this case there is no
             point in materializing.
          4. Subquery predicate is a top-level predicate
             (this implies it is not negated)
             TODO: this is a limitation that should be lifeted once we
             implement correct NULL semantics (WL#3830)
          5. Subquery is non-correlated
             TODO:
             This is an overly restrictive condition. It can be extended to:
             (Subquery is non-correlated ||
              Subquery is correlated to any query outer to IN predicate ||
              (Subquery is correlated to the immediate outer query &&
               Subquery !contains {GROUP BY, order_st BY [LIMIT],
               aggregate functions) && subquery predicate is not under "NOT IN"))
          6. No execution method was already chosen (by a prepared statement).

          (*) The subquery must be part of a SELECT statement. The current
               condition also excludes multi-table update statements.

          We have to determine whether we will perform subquery materialization
          before calling the IN=>EXISTS transformation, so that we know whether to
          perform the whole transformation or only that part of it which wraps
          Item_in_subselect in an Item_in_optimizer.
        */
        if (do_materialize &&
            in_subs  &&                                                   // 1
            !select_lex->master_unit()->first_select()->next_select() &&  // 2
            select_lex->master_unit()->first_select()->leaf_tables &&     // 3
            session->lex->sql_command == SQLCOM_SELECT)                       // *
        {
          if (in_subs->is_top_level_item() &&                             // 4
              !in_subs->is_correlated &&                                  // 5
              in_subs->exec_method == Item_in_subselect::NOT_TRANSFORMED) // 6
            in_subs->exec_method= Item_in_subselect::MATERIALIZATION;
        }

        Item_subselect::trans_res trans_res;
        if ((trans_res= subselect->select_transformer(this)) !=
            Item_subselect::RES_OK)
        {
          return((trans_res == Item_subselect::RES_ERROR));
        }
      }
    }
  }

  if (order)
  {
    order_st *ord;
    for (ord= order; ord; ord= ord->next)
    {
      Item *item= *ord->item;
      if (item->with_sum_func && item->type() != Item::SUM_FUNC_ITEM)
        item->split_sum_func(session, ref_pointer_array, all_fields);
    }
  }

  if (having && having->with_sum_func)
    having->split_sum_func(session, ref_pointer_array, all_fields,
                           &having, true);
  if (select_lex->inner_sum_func_list)
  {
    Item_sum *end=select_lex->inner_sum_func_list;
    Item_sum *item_sum= end;
    do
    {
      item_sum= item_sum->next;
      item_sum->split_sum_func(session, ref_pointer_array,
                               all_fields, item_sum->ref_by, false);
    } while (item_sum != end);
  }

  if (select_lex->inner_refs_list.elements &&
      fix_inner_refs(session, all_fields, select_lex, ref_pointer_array))
    return(-1);

  /*
    Check if there are references to un-aggregated columns when computing
    aggregate functions with implicit grouping (there is no GROUP BY).

    MODE_ONLY_FULL_GROUP_BY is enabled here by default
  */
  if (!group_list && select_lex->full_group_by_flag == (NON_AGG_FIELD_USED | SUM_FUNC_USED))
  {
    my_message(ER_MIX_OF_GROUP_FUNC_AND_FIELDS,
               ER(ER_MIX_OF_GROUP_FUNC_AND_FIELDS), MYF(0));
    return(-1);
  }
  {
    /* Caclulate the number of groups */
    send_group_parts= 0;
    for (order_st *group_tmp= group_list ; group_tmp ; group_tmp= group_tmp->next)
      send_group_parts++;
  }

  if (error)
    goto err;					/* purecov: inspected */

  if (result && result->prepare(fields_list, unit_arg))
    goto err;					/* purecov: inspected */

  /* Init join struct */
  count_field_types(select_lex, &tmp_table_param, all_fields, 0);
  ref_pointer_array_size= all_fields.elements*sizeof(Item*);
  this->group= group_list != 0;
  unit= unit_arg;

#ifdef RESTRICTED_GROUP
  if (sum_func_count && !group_list && (func_count || field_count))
  {
    my_message(ER_WRONG_SUM_SELECT,ER(ER_WRONG_SUM_SELECT),MYF(0));
    goto err;
  }
#endif
  if (select_lex->olap == ROLLUP_TYPE && rollup_init())
    goto err;
  if (alloc_func_list())
    goto err;

  return(0); // All OK

err:
  return(-1);				/* purecov: inspected */
}


/*
  Remove the predicates pushed down into the subquery

  SYNOPSIS
    JOIN::remove_subq_pushed_predicates()
      where   IN  Must be NULL
              OUT The remaining WHERE condition, or NULL

  DESCRIPTION
    Given that this join will be executed using (unique|index)_subquery,
    without "checking NULL", remove the predicates that were pushed down
    into the subquery.

    If the subquery compares scalar values, we can remove the condition that
    was wrapped into trig_cond (it will be checked when needed by the subquery
    engine)

    If the subquery compares row values, we need to keep the wrapped
    equalities in the WHERE clause: when the left (outer) tuple has both NULL
    and non-NULL values, we'll do a full table scan and will rely on the
    equalities corresponding to non-NULL parts of left tuple to filter out
    non-matching records.

    TODO: We can remove the equalities that will be guaranteed to be true by the
    fact that subquery engine will be using index lookup. This must be done only
    for cases where there are no conversion errors of significance, e.g. 257
    that is searched in a byte. But this requires homogenization of the return
    codes of all Field*::store() methods.
*/

void JOIN::remove_subq_pushed_predicates(Item **where)
{
  if (conds->type() == Item::FUNC_ITEM &&
      ((Item_func *)this->conds)->functype() == Item_func::EQ_FUNC &&
      ((Item_func *)conds)->arguments()[0]->type() == Item::REF_ITEM &&
      ((Item_func *)conds)->arguments()[1]->type() == Item::FIELD_ITEM &&
      test_if_ref ((Item_field *)((Item_func *)conds)->arguments()[1],
                   ((Item_func *)conds)->arguments()[0]))
  {
    *where= 0;
    return;
  }
}


/*
  Index lookup-based subquery: save some flags for EXPLAIN output

  SYNOPSIS
    save_index_subquery_explain_info()
      join_tab  Subquery's join tab (there is only one as index lookup is
                only used for subqueries that are single-table SELECTs)
      where     Subquery's WHERE clause

  DESCRIPTION
    For index lookup-based subquery (i.e. one executed with
    subselect_uniquesubquery_engine or subselect_indexsubquery_engine),
    check its EXPLAIN output row should contain
      "Using index" (TAB_INFO_FULL_SCAN_ON_NULL)
      "Using Where" (TAB_INFO_USING_WHERE)
      "Full scan on NULL key" (TAB_INFO_FULL_SCAN_ON_NULL)
    and set appropriate flags in join_tab->packed_info.
*/

static void save_index_subquery_explain_info(JOIN_TAB *join_tab, Item* where)
{
  join_tab->packed_info= TAB_INFO_HAVE_VALUE;
  if (join_tab->table->covering_keys.test(join_tab->ref.key))
    join_tab->packed_info |= TAB_INFO_USING_INDEX;
  if (where)
    join_tab->packed_info |= TAB_INFO_USING_WHERE;
  for (uint32_t i = 0; i < join_tab->ref.key_parts; i++)
  {
    if (join_tab->ref.cond_guards[i])
    {
      join_tab->packed_info |= TAB_INFO_FULL_SCAN_ON_NULL;
      break;
    }
  }
}




/*
  Check if the table's rowid is included in the temptable

  SYNOPSIS
    sj_table_is_included()
      join      The join
      join_tab  The table to be checked

  DESCRIPTION
    SemiJoinDuplicateElimination: check the table's rowid should be included
    in the temptable. This is so if

    1. The table is not embedded within some semi-join nest
    2. The has been pulled out of a semi-join nest, or

    3. The table is functionally dependent on some previous table

    [4. This is also true for constant tables that can't be
        NULL-complemented but this function is not called for such tables]

  RETURN
    true  - Include table's rowid
    false - Don't
*/

static bool sj_table_is_included(JOIN *join, JOIN_TAB *join_tab)
{
  if (join_tab->emb_sj_nest)
    return false;

  /* Check if this table is functionally dependent on the tables that
     are within the same outer join nest
  */
  TableList *embedding= join_tab->table->pos_in_table_list->embedding;
  if (join_tab->type == JT_EQ_REF)
  {
    Table_map_iterator it(join_tab->ref.depend_map & ~PSEUDO_TABLE_BITS);
    uint32_t idx;
    while ((idx= it.next_bit())!=Table_map_iterator::BITMAP_END)
    {
      JOIN_TAB *ref_tab= join->join_tab + idx;
      if (embedding == ref_tab->table->pos_in_table_list->embedding)
        return true;
    }
    /* Ok, functionally dependent */
    return false;
  }
  /* Not functionally dependent => need to include*/
  return true;
}


/*
  Setup the strategies to eliminate semi-join duplicates.

  SYNOPSIS
    setup_semijoin_dups_elimination()
      join           Join to process
      options        Join options (needed to see if join buffering will be
                     used or not)
      no_jbuf_after  Another bit of information re where join buffering will
                     be used.

  DESCRIPTION
    Setup the strategies to eliminate semi-join duplicates. ATM there are 3
    strategies:

    1. DuplicateWeedout (use of temptable to remove duplicates based on rowids
                         of row combinations)
    2. FirstMatch (pick only the 1st matching row combination of inner tables)
    3. InsideOut (scanning the sj-inner table in a way that groups duplicates
                  together and picking the 1st one)

    The join order has "duplicate-generating ranges", and every range is
    served by one strategy or a combination of FirstMatch with with some
    other strategy.

    "Duplicate-generating range" is defined as a range within the join order
    that contains all of the inner tables of a semi-join. All ranges must be
    disjoint, if tables of several semi-joins are interleaved, then the ranges
    are joined together, which is equivalent to converting
      SELECT ... WHERE oe1 IN (SELECT ie1 ...) AND oe2 IN (SELECT ie2 )
    to
      SELECT ... WHERE (oe1, oe2) IN (SELECT ie1, ie2 ... ...)
    .

    Applicability conditions are as follows:

    DuplicateWeedout strategy
    ~~~~~~~~~~~~~~~~~~~~~~~~~

      (ot|nt)*  [ it ((it|ot|nt)* (it|ot))]  (nt)*
      +------+  +=========================+  +---+
        (1)                 (2)               (3)

       (1) - Prefix of OuterTables (those that participate in
             IN-equality and/or are correlated with subquery) and outer
             Noncorrelated Tables.
       (2) - The handled range. The range starts with the first sj-inner
             table, and covers all sj-inner and outer tables
             Within the range,  Inner, Outer, outer Noncorrelated tables
             may follow in any order.
       (3) - The suffix of outer Noncorrelated tables.

    FirstMatch strategy
    ~~~~~~~~~~~~~~~~~~~

      (ot|nt)*  [ it ((it|nt)* it) ]  (nt)*
      +------+  +==================+  +---+
        (1)             (2)          (3)

      (1) - Prefix of outer and non-correlated tables
      (2) - The handled range, which may contain only inner and
            non-correlated tables.
      (3) - The suffix of outer Noncorrelated tables.

    InsideOut strategy
    ~~~~~~~~~~~~~~~~~~

     (ot|ct|nt) [ insideout_tbl (ot|nt|it)* it ]  (ot|nt)*
     +--------+   +===========+ +=============+   +------+
        (1)           (2)          (3)              (4)

      (1) - Prefix that may contain any outer tables. The prefix must contain
            all the non-trivially correlated outer tables. (non-trivially means
            that the correlation is not just through the IN-equality).

      (2) - Inner table for which the InsideOut scan is performed.

      (3) - The remainder of the duplicate-generating range. It is served by
            application of FirstMatch strategy, with the exception that
            outer IN-correlated tables are considered to be non-correlated.

      (4) - THe suffix of outer and outer non-correlated tables.

    If several strategies are applicable, their relative priorities are:
      1. InsideOut
      2. FirstMatch
      3. DuplicateWeedout

    This function walks over the join order and sets up the strategies by
    setting appropriate members in join_tab structures.

  RETURN
    false  OK
    true   Out of memory error
*/

static
int setup_semijoin_dups_elimination(JOIN *join, uint64_t options, uint32_t no_jbuf_after)
{
  table_map cur_map= join->const_table_map | PSEUDO_TABLE_BITS;
  struct {
    /*
      0 - invalid (EOF marker),
      1 - InsideOut,
      2 - Temptable (maybe confluent),
      3 - Temptable with join buffering
    */
    uint32_t strategy;
    uint32_t start_idx; /* Left range bound */
    uint32_t end_idx;   /* Right range bound */
    /*
      For Temptable strategy: Bitmap of all outer and correlated tables from
      all involved join nests.
    */
    table_map outer_tables;
  } dups_ranges [MAX_TABLES];

  TableList *emb_insideout_nest= NULL;
  table_map emb_sj_map= 0;  /* A bitmap of sj-nests (that is, their sj-inner
                               tables) whose ranges we're in */
  table_map emb_outer_tables= 0; /* sj-outer tables for those sj-nests */
  table_map range_start_map= 0; /* table_map at current range start */
  bool dealing_with_jbuf= false; /* true <=> table within cur range uses join buf */
  int cur_range= 0;
  uint32_t i;

  /*
    First pass: locate the duplicate-generating ranges and pick the strategies.
  */
  for (i=join->const_tables ; i < join->tables ; i++)
  {
    JOIN_TAB *tab=join->join_tab+i;
    Table *table=tab->table;
    cur_map |= table->map;

    if (tab->emb_sj_nest) // Encountered an sj-inner table
    {
      if (!emb_sj_map)
      {
        dups_ranges[cur_range].start_idx= i;
        range_start_map= cur_map & ~table->map;
        /*
          Remember if this is a possible start of range that is covered by
          the InsideOut strategy (the reason that it is not covered could
          be that it overlaps with anther semi-join's range. we don't
          support InsideOut for joined ranges)
        */
        if (join->best_positions[i].use_insideout_scan)
          emb_insideout_nest= tab->emb_sj_nest;
      }

      emb_sj_map |= tab->emb_sj_nest->sj_inner_tables;
      emb_outer_tables |= tab->emb_sj_nest->nested_join->sj_depends_on;

      if (tab->emb_sj_nest != emb_insideout_nest)
      {
        /*
          Two different semi-joins interleave. This cannot be handled by
          InsideOut strategy.
        */
        emb_insideout_nest= NULL;
      }
    }

    if (emb_sj_map) /* We're in duplicate-generating range */
    {
      if (i != join->const_tables && !(options & SELECT_NO_JOIN_CACHE) &&
          tab->type == JT_ALL && tab->use_quick != 2 && !tab->first_inner &&
          i <= no_jbuf_after && !dealing_with_jbuf)
      {
        /*
          This table uses join buffering, which makes use of FirstMatch or
          InsideOut strategies impossible for the current and (we assume)
          preceding duplicate-producing ranges.
          That is, for the join order:

              x x [ x  x]  x  [x x x]  x  [x x X*  x] x
                  |     |     |     |          | \
                  +-----+     +-----+          |  join buffering use
                     r1          r2         we're here

          we'll have to remove r1 and r2 and use duplicate-elimination
          strategy that spans all the tables, starting from the very 1st
          one.
        */
        dealing_with_jbuf= true;
        emb_insideout_nest= false;

        /*
          Absorb all preceding duplicate-eliminating ranges. Their strategies
          do not matter:
        */
        for (int prev_range= 0; prev_range < cur_range; prev_range++)
        {
          dups_ranges[cur_range].outer_tables |=
            dups_ranges[prev_range].outer_tables;
        }
        dups_ranges[0].start_idx= 0; /* Will need to start from the 1st table */
        dups_ranges[0].outer_tables= dups_ranges[cur_range].outer_tables;
        cur_range=  0;
      }

      /*
        Check if we are at the end of duplicate-producing range. We are if

        1. It's an InsideOut range (which presumes all correlated tables are
           in the prefix), and all inner tables are in the join order prefix,
           or
        2. It's a DuplicateElimination range (possibly covering several
           SJ-nests), and all inner, outer, and correlated tables of all
           sj-nests are in the join order prefix.
      */
      bool end_of_range= false;
      if (emb_insideout_nest &&
          bitmap_covers(cur_map, emb_insideout_nest->sj_inner_tables))
      {
        /* Save that this range is handled with InsideOut: */
        dups_ranges[cur_range].strategy= 1;
        end_of_range= true;
      }
      else if (bitmap_covers(cur_map, emb_outer_tables | emb_sj_map))
      {
        /*
          This is a complete range to be handled with either DuplicateWeedout
          or FirstMatch
        */
        dups_ranges[cur_range].strategy= dealing_with_jbuf? 3 : 2;
        /*
          This will hold tables from within the range that need to be put
          into the join buffer before we can use the FirstMatch on its tail.
        */
        dups_ranges[cur_range].outer_tables= emb_outer_tables &
                                             ~range_start_map;
        end_of_range= true;
      }

      if (end_of_range)
      {
        dups_ranges[cur_range].end_idx= i+1;
        emb_sj_map= emb_outer_tables= 0;
        emb_insideout_nest= NULL;
        dealing_with_jbuf= false;
        dups_ranges[++cur_range].strategy= 0;
      }
    }
  }

  Session *session= join->session;
  SJ_TMP_TABLE **next_sjtbl_ptr= &join->sj_tmp_tables;
  /*
    Second pass: setup the chosen strategies
  */
  for (int j= 0; j < cur_range; j++)
  {
    JOIN_TAB *tab=join->join_tab + dups_ranges[j].start_idx;
    JOIN_TAB *jump_to;
    if (dups_ranges[j].strategy == 1)  // InsideOut strategy
    {
      tab->insideout_match_tab= join->join_tab + dups_ranges[j].end_idx - 1;
      jump_to= tab++;
    }
    else // DuplicateWeedout strategy
    {
      SJ_TMP_TABLE::TAB sjtabs[MAX_TABLES];
      table_map weed_cur_map= join->const_table_map | PSEUDO_TABLE_BITS;
      uint32_t jt_rowid_offset= 0; // # tuple bytes are already occupied (w/o NULL bytes)
      uint32_t jt_null_bits= 0;    // # null bits in tuple bytes
      SJ_TMP_TABLE::TAB *last_tab= sjtabs;
      uint32_t rowid_keep_flags= JOIN_TAB::CALL_POSITION | JOIN_TAB::KEEP_ROWID;
      JOIN_TAB *last_outer_tab= tab - 1;
      /*
        Walk through the range and remember
         - tables that need their rowids to be put into temptable
         - the last outer table
      */
      for (; tab < join->join_tab + dups_ranges[j].end_idx; tab++)
      {
        if (sj_table_is_included(join, tab))
        {
          last_tab->join_tab= tab;
          last_tab->rowid_offset= jt_rowid_offset;
          jt_rowid_offset += tab->table->file->ref_length;
          if (tab->table->maybe_null)
          {
            last_tab->null_byte= jt_null_bits / 8;
            last_tab->null_bit= jt_null_bits++;
          }
          last_tab++;
          tab->table->prepare_for_position();
          tab->rowid_keep_flags= rowid_keep_flags;
        }
        weed_cur_map |= tab->table->map;
        if (!tab->emb_sj_nest && bitmap_covers(weed_cur_map,
                                               dups_ranges[j].outer_tables))
          last_outer_tab= tab;
      }

      if (jt_rowid_offset) /* Temptable has at least one rowid */
      {
        SJ_TMP_TABLE *sjtbl;
        uint32_t tabs_size= (last_tab - sjtabs) * sizeof(SJ_TMP_TABLE::TAB);
        if (!(sjtbl= (SJ_TMP_TABLE*)session->alloc(sizeof(SJ_TMP_TABLE))) ||
            !(sjtbl->tabs= (SJ_TMP_TABLE::TAB*) session->alloc(tabs_size)))
          return(true);
        memcpy(sjtbl->tabs, sjtabs, tabs_size);
        sjtbl->tabs_end= sjtbl->tabs + (last_tab - sjtabs);
        sjtbl->rowid_len= jt_rowid_offset;
        sjtbl->null_bits= jt_null_bits;
        sjtbl->null_bytes= (jt_null_bits + 7)/8;

        *next_sjtbl_ptr= sjtbl;
        next_sjtbl_ptr= &(sjtbl->next);
        sjtbl->next= NULL;

        sjtbl->tmp_table=
          create_duplicate_weedout_tmp_table(session,
                                             sjtbl->rowid_len +
                                             sjtbl->null_bytes,
                                             sjtbl);

        join->join_tab[dups_ranges[j].start_idx].flush_weedout_table= sjtbl;
        join->join_tab[dups_ranges[j].end_idx - 1].check_weed_out_table= sjtbl;
      }
      tab= last_outer_tab + 1;
      jump_to= last_outer_tab;
    }

    /* Create the FirstMatch tail */
    for (; tab < join->join_tab + dups_ranges[j].end_idx; tab++)
    {
      if (tab->emb_sj_nest)
        tab->do_firstmatch= jump_to;
      else
        jump_to= tab;
    }
  }
  return(false);
}


static void cleanup_sj_tmp_tables(JOIN *join)
{
  for (SJ_TMP_TABLE *sj_tbl= join->sj_tmp_tables; sj_tbl;
       sj_tbl= sj_tbl->next)
  {
    if (sj_tbl->tmp_table)
    {
      sj_tbl->tmp_table->free_tmp_table(join->session);
    }
  }
  join->sj_tmp_tables= NULL;
}

uint32_t make_join_orderinfo(JOIN *join);

/**
  global select optimisation.

  @note
    error code saved in field 'error'

  @retval
    0   success
  @retval
    1   error
*/

int
JOIN::optimize()
{
  // to prevent double initialization on EXPLAIN
  if (optimized)
    return(0);
  optimized= 1;

  session->set_proc_info("optimizing");
  row_limit= ((select_distinct || order || group_list) ? HA_POS_ERROR :
	      unit->select_limit_cnt);
  /* select_limit is used to decide if we are likely to scan the whole table */
  select_limit= unit->select_limit_cnt;
  if (having || (select_options & OPTION_FOUND_ROWS))
    select_limit= HA_POS_ERROR;
  do_send_rows = (unit->select_limit_cnt) ? 1 : 0;
  // Ignore errors of execution if option IGNORE present
  if (session->lex->ignore)
    session->lex->current_select->no_error= 1;

#ifdef HAVE_REF_TO_FIELDS			// Not done yet
  /* Add HAVING to WHERE if possible */
  if (having && !group_list && !sum_func_count)
  {
    if (!conds)
    {
      conds= having;
      having= 0;
    }
    else if ((conds=new Item_cond_and(conds,having)))
    {
      /*
        Item_cond_and can't be fixed after creation, so we do not check
        conds->fixed
      */
      conds->fix_fields(session, &conds);
      conds->change_ref_to_fields(session, tables_list);
      conds->top_level_item();
      having= 0;
    }
  }
#endif

  /* Convert all outer joins to inner joins if possible */
  conds= simplify_joins(this, join_list, conds, true, false);
  build_bitmap_for_nested_joins(join_list, 0);

  conds= optimize_cond(this, conds, join_list, &cond_value);
  if (session->is_error())
  {
    error= 1;
    return(1);
  }

  {
    having= optimize_cond(this, having, join_list, &having_value);
    if (session->is_error())
    {
      error= 1;
      return(1);
    }
    if (select_lex->where)
      select_lex->cond_value= cond_value;
    if (select_lex->having)
      select_lex->having_value= having_value;

    if (cond_value == Item::COND_FALSE || having_value == Item::COND_FALSE ||
        (!unit->select_limit_cnt && !(select_options & OPTION_FOUND_ROWS)))
    {						/* Impossible cond */
      zero_result_cause=  having_value == Item::COND_FALSE ?
                           "Impossible HAVING" : "Impossible WHERE";
      error= 0;
      return(0);
    }
  }

  /* Optimize count(*), cmin() and cmax() */
  if (tables_list && tmp_table_param.sum_func_count && ! group_list)
  {
    int res;
    /*
      opt_sum_query() returns HA_ERR_KEY_NOT_FOUND if no rows match
      to the WHERE conditions,
      or 1 if all items were resolved,
      or 0, or an error number HA_ERR_...
    */
    if ((res=opt_sum_query(select_lex->leaf_tables, all_fields, conds)))
    {
      if (res == HA_ERR_KEY_NOT_FOUND)
      {
	zero_result_cause= "No matching min/max row";
	error=0;
	return(0);
      }
      if (res > 1)
      {
        error= res;
        return(1);
      }
      if (res < 0)
      {
        zero_result_cause= "No matching min/max row";
        error=0;
        return(0);
      }
      zero_result_cause= "Select tables optimized away";
      tables_list= 0;				// All tables resolved
      /*
        Extract all table-independent conditions and replace the WHERE
        clause with them. All other conditions were computed by opt_sum_query
        and the MIN/MAX/COUNT function(s) have been replaced by constants,
        so there is no need to compute the whole WHERE clause again.
        Notice that make_cond_for_table() will always succeed to remove all
        computed conditions, because opt_sum_query() is applicable only to
        conjunctions.
        Preserve conditions for EXPLAIN.
      */
      if (conds && !(session->lex->describe & DESCRIBE_EXTENDED))
      {
        COND *table_independent_conds=
          make_cond_for_table(conds, PSEUDO_TABLE_BITS, 0, 0);
        conds= table_independent_conds;
      }
    }
  }
  if (!tables_list)
  {
    error= 0;
    return(0);
  }
  error= -1;					// Error is sent to client
  sort_by_table= get_sort_by_table(order, group_list, select_lex->leaf_tables);

  /* Calculate how to do the join */
  session->set_proc_info("statistics");
  if (make_join_statistics(this, select_lex->leaf_tables, conds, &keyuse) ||
      session->is_fatal_error)
  {
    return(1);
  }

  /* Remove distinct if only const tables */
  select_distinct= select_distinct && (const_tables != tables);
  session->set_proc_info("preparing");
  if (result->initialize_tables(this))
  {
    return(1);				// error == -1
  }
  if (const_table_map != found_const_table_map &&
      !(select_options & SELECT_DESCRIBE) &&
      (!conds ||
       !(conds->used_tables() & RAND_TABLE_BIT) ||
       select_lex->master_unit() == &session->lex->unit)) // upper level SELECT
  {
    zero_result_cause= "no matching row in const table";
    error= 0;
    return(0);
  }
  if (!(session->options & OPTION_BIG_SELECTS) &&
      best_read > (double) session->variables.max_join_size &&
      !(select_options & SELECT_DESCRIBE))
  {						/* purecov: inspected */
    my_message(ER_TOO_BIG_SELECT, ER(ER_TOO_BIG_SELECT), MYF(0));
    error= -1;
    return(1);
  }
  if (const_tables && !session->locked_tables &&
      !(select_options & SELECT_NO_UNLOCK))
    mysql_unlock_some_tables(session, table, const_tables);
  if (!conds && outer_join)
  {
    /* Handle the case where we have an OUTER JOIN without a WHERE */
    conds=new Item_int((int64_t) 1,1);	// Always true
  }
  select= make_select(*table, const_table_map,
                      const_table_map, conds, 1, &error);
  if (error)
  {						/* purecov: inspected */
    error= -1;					/* purecov: inspected */
    return(1);
  }

  reset_nj_counters(join_list);
  make_outerjoin_info(this);

  /*
    Among the equal fields belonging to the same multiple equality
    choose the one that is to be retrieved first and substitute
    all references to these in where condition for a reference for
    the selected field.
  */
  if (conds)
  {
    conds= substitute_for_best_equal_field(conds, cond_equal, map2table);
    conds->update_used_tables();
  }

  /*
    Permorm the the optimization on fields evaluation mentioned above
    for all on expressions.
  */
  for (JOIN_TAB *tab= join_tab + const_tables; tab < join_tab + tables ; tab++)
  {
    if (*tab->on_expr_ref)
    {
      *tab->on_expr_ref= substitute_for_best_equal_field(*tab->on_expr_ref,
                                                         tab->cond_equal,
                                                         map2table);
      (*tab->on_expr_ref)->update_used_tables();
    }
  }

  if (conds &&!outer_join && const_table_map != found_const_table_map &&
      (select_options & SELECT_DESCRIBE) &&
      select_lex->master_unit() == &session->lex->unit) // upper level SELECT
  {
    conds=new Item_int((int64_t) 0,1);	// Always false
  }
  if (make_join_select(this, select, conds))
  {
    zero_result_cause=
      "Impossible WHERE noticed after reading const tables";
    return(0);				// error == 0
  }

  error= -1;					/* if goto err */

  /* Optimize distinct away if possible */
  {
    order_st *org_order= order;
    order=remove_constants(this, order,conds,1, &simple_order);
    if (session->is_error())
    {
      error= 1;
      return(1);
    }

    /*
      If we are using order_st BY NULL or order_st BY const_expression,
      return result in any order (even if we are using a GROUP BY)
    */
    if (!order && org_order)
      skip_sort_order= 1;
  }
  /*
     Check if we can optimize away GROUP BY/DISTINCT.
     We can do that if there are no aggregate functions, the
     fields in DISTINCT clause (if present) and/or columns in GROUP BY
     (if present) contain direct references to all key parts of
     an unique index (in whatever order) and if the key parts of the
     unique index cannot contain NULLs.
     Note that the unique keys for DISTINCT and GROUP BY should not
     be the same (as long as they are unique).

     The FROM clause must contain a single non-constant table.
  */
  if (tables - const_tables == 1 && (group_list || select_distinct) &&
      !tmp_table_param.sum_func_count &&
      (!join_tab[const_tables].select ||
       !join_tab[const_tables].select->quick ||
       join_tab[const_tables].select->quick->get_type() !=
       QUICK_SELECT_I::QS_TYPE_GROUP_MIN_MAX))
  {
    if (group_list &&
       list_contains_unique_index(join_tab[const_tables].table,
                                 find_field_in_order_list,
                                 (void *) group_list))
    {
      /*
        We have found that grouping can be removed since groups correspond to
        only one row anyway, but we still have to guarantee correct result
        order. The line below effectively rewrites the query from GROUP BY
        <fields> to order_st BY <fields>. There are two exceptions:
        - if skip_sort_order is set (see above), then we can simply skip
          GROUP BY;
        - we can only rewrite order_st BY if the order_st BY fields are 'compatible'
          with the GROUP BY ones, i.e. either one is a prefix of another.
          We only check if the order_st BY is a prefix of GROUP BY. In this case
          test_if_subpart() copies the ASC/DESC attributes from the original
          order_st BY fields.
          If GROUP BY is a prefix of order_st BY, then it is safe to leave
          'order' as is.
       */
      if (!order || test_if_subpart(group_list, order))
          order= skip_sort_order ? 0 : group_list;
      /*
        If we have an IGNORE INDEX FOR GROUP BY(fields) clause, this must be
        rewritten to IGNORE INDEX FOR order_st BY(fields).
      */
      join_tab->table->keys_in_use_for_order_by=
        join_tab->table->keys_in_use_for_group_by;
      group_list= 0;
      group= 0;
    }
    if (select_distinct &&
       list_contains_unique_index(join_tab[const_tables].table,
                                 find_field_in_item_list,
                                 (void *) &fields_list))
    {
      select_distinct= 0;
    }
  }
  if (group_list || tmp_table_param.sum_func_count)
  {
    if (! hidden_group_fields && rollup.state == ROLLUP::STATE_NONE)
      select_distinct=0;
  }
  else if (select_distinct && tables - const_tables == 1)
  {
    /*
      We are only using one table. In this case we change DISTINCT to a
      GROUP BY query if:
      - The GROUP BY can be done through indexes (no sort) and the order_st
        BY only uses selected fields.
	(In this case we can later optimize away GROUP BY and order_st BY)
      - We are scanning the whole table without LIMIT
        This can happen if:
        - We are using CALC_FOUND_ROWS
        - We are using an order_st BY that can't be optimized away.

      We don't want to use this optimization when we are using LIMIT
      because in this case we can just create a temporary table that
      holds LIMIT rows and stop when this table is full.
    */
    JOIN_TAB *tab= &join_tab[const_tables];
    bool all_order_fields_used;
    if (order)
      skip_sort_order= test_if_skip_sort_order(tab, order, select_limit, 1,
        &tab->table->keys_in_use_for_order_by);
    if ((group_list=create_distinct_group(session, select_lex->ref_pointer_array,
                                          order, fields_list, all_fields,
				          &all_order_fields_used)))
    {
      bool skip_group= (skip_sort_order &&
        test_if_skip_sort_order(tab, group_list, select_limit, 1,
                                &tab->table->keys_in_use_for_group_by) != 0);
      count_field_types(select_lex, &tmp_table_param, all_fields, 0);
      if ((skip_group && all_order_fields_used) ||
	  select_limit == HA_POS_ERROR ||
	  (order && !skip_sort_order))
      {
	/*  Change DISTINCT to GROUP BY */
	select_distinct= 0;
	no_order= !order;
	if (all_order_fields_used)
	{
	  if (order && skip_sort_order)
	  {
	    /*
	      Force MySQL to read the table in sorted order to get result in
	      order_st BY order.
	    */
	    tmp_table_param.quick_group=0;
	  }
	  order=0;
        }
	group=1;				// For end_write_group
      }
      else
	group_list= 0;
    }
    else if (session->is_fatal_error)			// End of memory
      return(1);
  }
  simple_group= 0;
  {
    order_st *old_group_list;
    group_list= remove_constants(this, (old_group_list= group_list), conds,
                                 rollup.state == ROLLUP::STATE_NONE,
                                 &simple_group);
    if (session->is_error())
    {
      error= 1;
      return(1);
    }
    if (old_group_list && !group_list)
      select_distinct= 0;
  }
  if (!group_list && group)
  {
    order=0;					// The output has only one row
    simple_order=1;
    select_distinct= 0;                       // No need in distinct for 1 row
    group_optimized_away= 1;
  }

  calc_group_buffer(this, group_list);
  send_group_parts= tmp_table_param.group_parts; /* Save org parts */

  if (test_if_subpart(group_list, order) ||
      (!group_list && tmp_table_param.sum_func_count))
    order=0;

  // Can't use sort on head table if using row cache
  if (full_join)
  {
    if (group_list)
      simple_group=0;
    if (order)
      simple_order=0;
  }

  /*
    Check if we need to create a temporary table.
    This has to be done if all tables are not already read (const tables)
    and one of the following conditions holds:
    - We are using DISTINCT (simple distinct's are already optimized away)
    - We are using an order_st BY or GROUP BY on fields not in the first table
    - We are using different order_st BY and GROUP BY orders
    - The user wants us to buffer the result.
  */
  need_tmp= (const_tables != tables &&
	     ((select_distinct || !simple_order || !simple_group) ||
	      (group_list && order) ||
	      test(select_options & OPTION_BUFFER_RESULT)));

  uint32_t no_jbuf_after= make_join_orderinfo(this);
  uint64_t select_opts_for_readinfo=
    (select_options & (SELECT_DESCRIBE | SELECT_NO_JOIN_CACHE)) | (0);

  sj_tmp_tables= NULL;
  if (!select_lex->sj_nests.is_empty())
    setup_semijoin_dups_elimination(this, select_opts_for_readinfo,
                                    no_jbuf_after);

  // No cache for MATCH == 'Don't use join buffering when we use MATCH'.
  if (make_join_readinfo(this, select_opts_for_readinfo, no_jbuf_after))
    return(1);

  /* Create all structures needed for materialized subquery execution. */
  if (setup_subquery_materialization())
    return(1);

  /*
    is this simple IN subquery?
  */
  if (!group_list && !order &&
      unit->item && unit->item->substype() == Item_subselect::IN_SUBS &&
      tables == 1 && conds &&
      !unit->is_union())
  {
    if (!having)
    {
      Item *where= conds;
      if (join_tab[0].type == JT_EQ_REF &&
	  join_tab[0].ref.items[0]->name == in_left_expr_name)
      {
        remove_subq_pushed_predicates(&where);
        save_index_subquery_explain_info(join_tab, where);
        join_tab[0].type= JT_UNIQUE_SUBQUERY;
        error= 0;
        return(unit->item->
                    change_engine(new
                                  subselect_uniquesubquery_engine(session,
                                                                  join_tab,
                                                                  unit->item,
                                                                  where)));
      }
      else if (join_tab[0].type == JT_REF &&
	       join_tab[0].ref.items[0]->name == in_left_expr_name)
      {
	remove_subq_pushed_predicates(&where);
        save_index_subquery_explain_info(join_tab, where);
        join_tab[0].type= JT_INDEX_SUBQUERY;
        error= 0;
        return(unit->item->
                    change_engine(new
                                  subselect_indexsubquery_engine(session,
                                                                 join_tab,
                                                                 unit->item,
                                                                 where,
                                                                 NULL,
                                                                 0)));
      }
    } else if (join_tab[0].type == JT_REF_OR_NULL &&
	       join_tab[0].ref.items[0]->name == in_left_expr_name &&
               having->name == in_having_cond)
    {
      join_tab[0].type= JT_INDEX_SUBQUERY;
      error= 0;
      conds= remove_additional_cond(conds);
      save_index_subquery_explain_info(join_tab, conds);
      return(unit->item->
		  change_engine(new subselect_indexsubquery_engine(session,
								   join_tab,
								   unit->item,
								   conds,
                                                                   having,
								   1)));
    }

  }
  /*
    Need to tell handlers that to play it safe, it should fetch all
    columns of the primary key of the tables: this is because MySQL may
    build row pointers for the rows, and for all columns of the primary key
    the read set has not necessarily been set by the server code.
  */
  if (need_tmp || select_distinct || group_list || order)
  {
    for (uint32_t i = const_tables; i < tables; i++)
      join_tab[i].table->prepare_for_position();
  }

  if (const_tables != tables)
  {
    /*
      Because filesort always does a full table scan or a quick range scan
      we must add the removed reference to the select for the table.
      We only need to do this when we have a simple_order or simple_group
      as in other cases the join is done before the sort.
    */
    if ((order || group_list) &&
        (join_tab[const_tables].type != JT_ALL) &&
        (join_tab[const_tables].type != JT_REF_OR_NULL) &&
        ((order && simple_order) || (group_list && simple_group)))
    {
      if (add_ref_to_table_cond(session,&join_tab[const_tables])) {
        return(1);
      }
    }

    if (!(select_options & SELECT_BIG_RESULT) &&
        ((group_list &&
          (!simple_group ||
           !test_if_skip_sort_order(&join_tab[const_tables], group_list,
                                    unit->select_limit_cnt, 0,
                                    &join_tab[const_tables].table->
                                    keys_in_use_for_group_by))) ||
         select_distinct) &&
        tmp_table_param.quick_group)
    {
      need_tmp=1; simple_order=simple_group=0;	// Force tmp table without sort
    }
    if (order)
    {
      /*
        Force using of tmp table if sorting by a SP or UDF function due to
        their expensive and probably non-deterministic nature.
      */
      for (order_st *tmp_order= order; tmp_order ; tmp_order=tmp_order->next)
      {
        Item *item= *tmp_order->item;
        if (item->is_expensive())
        {
          /* Force tmp table without sort */
          need_tmp=1; simple_order=simple_group=0;
          break;
        }
      }
    }
  }

  tmp_having= having;
  if (select_options & SELECT_DESCRIBE)
  {
    error= 0;
    return(0);
  }
  having= 0;

  /*
    The loose index scan access method guarantees that all grouping or
    duplicate row elimination (for distinct) is already performed
    during data retrieval, and that all MIN/MAX functions are already
    computed for each group. Thus all MIN/MAX functions should be
    treated as regular functions, and there is no need to perform
    grouping in the main execution loop.
    Notice that currently loose index scan is applicable only for
    single table queries, thus it is sufficient to test only the first
    join_tab element of the plan for its access method.
  */
  if (join_tab->is_using_loose_index_scan())
    tmp_table_param.precomputed_group_by= true;

  /* Create a tmp table if distinct or if the sort is too complicated */
  if (need_tmp)
  {
    session->set_proc_info("Creating tmp table");

    init_items_ref_array();

    tmp_table_param.hidden_field_count= (all_fields.elements -
					 fields_list.elements);
    order_st *tmp_group= ((!simple_group && !(test_flags & TEST_NO_KEY_GROUP)) ? group_list :
                                                             (order_st*) 0);
    /*
      Pushing LIMIT to the temporary table creation is not applicable
      when there is order_st BY or GROUP BY or there is no GROUP BY, but
      there are aggregate functions, because in all these cases we need
      all result rows.
    */
    ha_rows tmp_rows_limit= ((order == 0 || skip_sort_order) &&
                             !tmp_group &&
                             !session->lex->current_select->with_sum_func) ?
                            select_limit : HA_POS_ERROR;

    if (!(exec_tmp_table1=
	  create_tmp_table(session, &tmp_table_param, all_fields,
                           tmp_group,
			   group_list ? 0 : select_distinct,
			   group_list && simple_group,
			   select_options,
                           tmp_rows_limit,
			   (char *) "")))
		{
      return(1);
    }

    /*
      We don't have to store rows in temp table that doesn't match HAVING if:
      - we are sorting the table and writing complete group rows to the
        temp table.
      - We are using DISTINCT without resolving the distinct as a GROUP BY
        on all columns.

      If having is not handled here, it will be checked before the row
      is sent to the client.
    */
    if (tmp_having &&
	(sort_and_group || (exec_tmp_table1->distinct && !group_list)))
      having= tmp_having;

    /* if group or order on first table, sort first */
    if (group_list && simple_group)
    {
      session->set_proc_info("Sorting for group");
      if (create_sort_index(session, this, group_list,
			    HA_POS_ERROR, HA_POS_ERROR, false) ||
	  alloc_group_fields(this, group_list) ||
          make_sum_func_list(all_fields, fields_list, 1) ||
          setup_sum_funcs(session, sum_funcs))
      {
        return(1);
      }
      group_list=0;
    }
    else
    {
      if (make_sum_func_list(all_fields, fields_list, 0) ||
          setup_sum_funcs(session, sum_funcs))
      {
        return(1);
      }

      if (!group_list && ! exec_tmp_table1->distinct && order && simple_order)
      {
        session->set_proc_info("Sorting for order");
        if (create_sort_index(session, this, order,
                              HA_POS_ERROR, HA_POS_ERROR, true))
        {
          return(1);
        }
        order=0;
      }
    }

    /*
      Optimize distinct when used on some of the tables
      SELECT DISTINCT t1.a FROM t1,t2 WHERE t1.b=t2.b
      In this case we can stop scanning t2 when we have found one t1.a
    */

    if (exec_tmp_table1->distinct)
    {
      table_map used_tables= session->used_tables;
      JOIN_TAB *last_join_tab= join_tab+tables-1;
      do
      {
	if (used_tables & last_join_tab->table->map)
	  break;
	last_join_tab->not_used_in_distinct=1;
      } while (last_join_tab-- != join_tab);
      /* Optimize "select distinct b from t1 order by key_part_1 limit #" */
      if (order && skip_sort_order)
      {
 	/* Should always succeed */
	if (test_if_skip_sort_order(&join_tab[const_tables],
				    order, unit->select_limit_cnt, 0,
                                    &join_tab[const_tables].table->
                                      keys_in_use_for_order_by))
	  order=0;
      }
    }

    /*
      If this join belongs to an uncacheable subquery save
      the original join
    */
    if (select_lex->uncacheable && !is_top_level_join() &&
        init_save_join_tab())
      return(-1);                         /* purecov: inspected */
  }

  error= 0;
  return(0);
}


/**
  Restore values in temporary join.
*/
void JOIN::restore_tmp()
{
  memcpy(tmp_join, this, (size_t) sizeof(JOIN));
}


int
JOIN::reinit()
{
  unit->offset_limit_cnt= (ha_rows)(select_lex->offset_limit ?
                                    select_lex->offset_limit->val_uint() :
                                    0UL);

  first_record= 0;

  if (exec_tmp_table1)
  {
    exec_tmp_table1->file->extra(HA_EXTRA_RESET_STATE);
    exec_tmp_table1->file->ha_delete_all_rows();
    free_io_cache(exec_tmp_table1);
    filesort_free_buffers(exec_tmp_table1,0);
  }
  if (exec_tmp_table2)
  {
    exec_tmp_table2->file->extra(HA_EXTRA_RESET_STATE);
    exec_tmp_table2->file->ha_delete_all_rows();
    free_io_cache(exec_tmp_table2);
    filesort_free_buffers(exec_tmp_table2,0);
  }
  if (items0)
    set_items_ref_array(items0);

  if (join_tab_save)
    memcpy(join_tab, join_tab_save, sizeof(JOIN_TAB) * tables);

  if (tmp_join)
    restore_tmp();

  /* Reset of sum functions */
  if (sum_funcs)
  {
    Item_sum *func, **func_ptr= sum_funcs;
    while ((func= *(func_ptr++)))
      func->clear();
  }

  return(0);
}

/**
   @brief Save the original join layout

   @details Saves the original join layout so it can be reused in
   re-execution and for EXPLAIN.

   @return Operation status
   @retval 0      success.
   @retval 1      error occurred.
*/

bool
JOIN::init_save_join_tab()
{
  if (!(tmp_join= (JOIN*)session->alloc(sizeof(JOIN))))
    return 1;                                  /* purecov: inspected */
  error= 0;				       // Ensure that tmp_join.error= 0
  restore_tmp();
  return 0;
}


bool
JOIN::save_join_tab()
{
  if (!join_tab_save && select_lex->master_unit()->uncacheable)
  {
    if (!(join_tab_save= (JOIN_TAB*)session->memdup((unsigned char*) join_tab,
						sizeof(JOIN_TAB) * tables)))
      return 1;
  }
  return 0;
}


/**
  Exec select.

  @todo
    Note, that create_sort_index calls test_if_skip_sort_order and may
    finally replace sorting with index scan if there is a LIMIT clause in
    the query.  It's never shown in EXPLAIN!

  @todo
    When can we have here session->net.report_error not zero?
*/
void JOIN::exec()
{
  List<Item> *columns_list= &fields_list;
  int      tmp_error;

  session->set_proc_info("executing");
  error= 0;

  if (!tables_list && (tables || !select_lex->with_sum_func))
  {                                           
    /* Only test of functions */
    if (select_options & SELECT_DESCRIBE)
      select_describe(this, false, false, false, (zero_result_cause?zero_result_cause:"No tables used"));
    else
    {
      result->send_fields(*columns_list, Protocol::SEND_NUM_ROWS | Protocol::SEND_EOF);
      /*
        We have to test for 'conds' here as the WHERE may not be constant
        even if we don't have any tables for prepared statements or if
        conds uses something like 'rand()'.
      */
      if (cond_value != Item::COND_FALSE &&
          (!conds || conds->val_int()) &&
          (!having || having->val_int()))
      {
        if (do_send_rows && result->send_data(fields_list))
          error= 1;
        else
        {
          error= (int) result->send_eof();
          send_records= ((select_options & OPTION_FOUND_ROWS) ? 1 : session->sent_row_count);
        }
      }
      else
      {
        error= (int) result->send_eof();
        send_records= 0;
      }
    }
    /* Single select (without union) always returns 0 or 1 row */
    session->limit_found_rows= send_records;
    session->examined_row_count= 0;
    return;
  }
  /*
    Don't reset the found rows count if there're no tables as
    FOUND_ROWS() may be called. Never reset the examined row count here.
    It must be accumulated from all join iterations of all join parts.
  */
  if (tables)
    session->limit_found_rows= 0;

  if (zero_result_cause)
  {
    (void) return_zero_rows(this, result, select_lex->leaf_tables,
                            *columns_list,
			    send_row_on_empty_set(),
			    select_options,
			    zero_result_cause,
			    having);
    return;
  }

  if ((this->select_lex->options & OPTION_SCHEMA_TABLE) && get_schema_tables_result(this, PROCESSED_BY_JOIN_EXEC))
    return;

  if (select_options & SELECT_DESCRIBE)
  {
    /*
      Check if we managed to optimize order_st BY away and don't use temporary
      table to resolve order_st BY: in that case, we only may need to do
      filesort for GROUP BY.
    */
    if (!order && !no_order && (!skip_sort_order || !need_tmp))
    {
      /* Reset 'order' to 'group_list' and reinit variables describing 'order' */
      order= group_list;
      simple_order= simple_group;
      skip_sort_order= 0;
    }
    if (order && (order != group_list || !(select_options & SELECT_BIG_RESULT)))
    {
      if (const_tables == tables 
        || ((simple_order || skip_sort_order) 
          && test_if_skip_sort_order(&join_tab[const_tables], order, select_limit, 0, &join_tab[const_tables].table->keys_in_use_for_query)))
      order= 0;
    }
    having= tmp_having;
    select_describe(this, need_tmp, order != 0 && !skip_sort_order,  select_distinct, !tables ? "No tables used" : NULL);
    return;
  }

  JOIN *curr_join= this;
  List<Item> *curr_all_fields= &all_fields;
  List<Item> *curr_fields_list= &fields_list;
  Table *curr_tmp_table= 0;
  /*
    Initialize examined rows here because the values from all join parts
    must be accumulated in examined_row_count. Hence every join
    iteration must count from zero.
  */
  curr_join->examined_rows= 0;

  /* Create a tmp table if distinct or if the sort is too complicated */
  if (need_tmp)
  {
    if (tmp_join)
    {
      /*
        We are in a non cacheable sub query. Get the saved join structure
        after optimization.
        (curr_join may have been modified during last exection and we need
        to reset it)
      */
      curr_join= tmp_join;
    }
    curr_tmp_table= exec_tmp_table1;

    /* Copy data to the temporary table */
    session->set_proc_info("Copying to tmp table");
    if (! curr_join->sort_and_group && curr_join->const_tables != curr_join->tables)
      curr_join->join_tab[curr_join->const_tables].sorted= 0;
    if ((tmp_error= do_select(curr_join, (List<Item> *) 0, curr_tmp_table)))
    {
      error= tmp_error;
      return;
    }
    curr_tmp_table->file->info(HA_STATUS_VARIABLE);

    if (curr_join->having)
      curr_join->having= curr_join->tmp_having= 0; // Allready done

    /* Change sum_fields reference to calculated fields in tmp_table */
    curr_join->all_fields= *curr_all_fields;
    if (!items1)
    {
      items1= items0 + all_fields.elements;
      if (sort_and_group || curr_tmp_table->group)
      {
        if (change_to_use_tmp_fields(session, items1,
                  tmp_fields_list1, tmp_all_fields1,
                  fields_list.elements, all_fields))
          return;
      }
      else
      {
        if (change_refs_to_tmp_fields(session, items1,
                    tmp_fields_list1, tmp_all_fields1,
                    fields_list.elements, all_fields))
          return;
      }
      curr_join->tmp_all_fields1= tmp_all_fields1;
      curr_join->tmp_fields_list1= tmp_fields_list1;
      curr_join->items1= items1;
    }
    curr_all_fields= &tmp_all_fields1;
    curr_fields_list= &tmp_fields_list1;
    curr_join->set_items_ref_array(items1);

    if (sort_and_group || curr_tmp_table->group)
    {
      curr_join->tmp_table_param.field_count+= curr_join->tmp_table_param.sum_func_count
                                             + curr_join->tmp_table_param.func_count;
      curr_join->tmp_table_param.sum_func_count= 0;
      curr_join->tmp_table_param.func_count= 0;
    }
    else
    {
      curr_join->tmp_table_param.field_count+= curr_join->tmp_table_param.func_count;
      curr_join->tmp_table_param.func_count= 0;
    }

    if (curr_tmp_table->group)
    {						// Already grouped
      if (!curr_join->order && !curr_join->no_order && !skip_sort_order)
        curr_join->order= curr_join->group_list;  /* order by group */
      curr_join->group_list= 0;
    }

    /*
      If we have different sort & group then we must sort the data by group
      and copy it to another tmp table
      This code is also used if we are using distinct something
      we haven't been able to store in the temporary table yet
      like SEC_TO_TIME(SUM(...)).
    */

    if ((curr_join->group_list && (!test_if_subpart(curr_join->group_list, curr_join->order) || curr_join->select_distinct)) 
        || (curr_join->select_distinct && curr_join->tmp_table_param.using_indirect_summary_function))
    {					/* Must copy to another table */
      /* Free first data from old join */
      curr_join->join_free();
      if (make_simple_join(curr_join, curr_tmp_table))
      	return;
      calc_group_buffer(curr_join, group_list);
      count_field_types(select_lex, &curr_join->tmp_table_param,
			curr_join->tmp_all_fields1,
			curr_join->select_distinct && !curr_join->group_list);
      curr_join->tmp_table_param.hidden_field_count= curr_join->tmp_all_fields1.elements
                                                   - curr_join->tmp_fields_list1.elements;

      if (exec_tmp_table2)
        curr_tmp_table= exec_tmp_table2;
      else
      {
        /* group data to new table */

        /*
          If the access method is loose index scan then all MIN/MAX
          functions are precomputed, and should be treated as regular
          functions. See extended comment in JOIN::exec.
        */
        if (curr_join->join_tab->is_using_loose_index_scan())
          curr_join->tmp_table_param.precomputed_group_by= true;

        if (!(curr_tmp_table=
              exec_tmp_table2= create_tmp_table(session,
                                                &curr_join->tmp_table_param,
                                                *curr_all_fields,
                                                (order_st*) 0,
                                                curr_join->select_distinct &&
                                                !curr_join->group_list,
                                                1, curr_join->select_options,
                                                HA_POS_ERROR,
                                                (char *) "")))
          return;
        curr_join->exec_tmp_table2= exec_tmp_table2;
      }
      if (curr_join->group_list)
      {
        session->set_proc_info("Creating sort index");
        if (curr_join->join_tab == join_tab && save_join_tab())
        {
          return;
        }
        if (create_sort_index(session, curr_join, curr_join->group_list,
                  HA_POS_ERROR, HA_POS_ERROR, false) ||
            make_group_fields(this, curr_join))
        {
          return;
        }
        sortorder= curr_join->sortorder;
      }

      session->set_proc_info("Copying to group table");
      tmp_error= -1;
      if (curr_join != this)
      {
        if (sum_funcs2)
        {
          curr_join->sum_funcs= sum_funcs2;
          curr_join->sum_funcs_end= sum_funcs_end2;
        }
        else
        {
          curr_join->alloc_func_list();
          sum_funcs2= curr_join->sum_funcs;
          sum_funcs_end2= curr_join->sum_funcs_end;
        }
      }
      if (curr_join->make_sum_func_list(*curr_all_fields, *curr_fields_list, 1, true))
        return;
      curr_join->group_list= 0;

      if (!curr_join->sort_and_group && (curr_join->const_tables != curr_join->tables))
        curr_join->join_tab[curr_join->const_tables].sorted= 0;
      
      if (setup_sum_funcs(curr_join->session, curr_join->sum_funcs) 
        || (tmp_error= do_select(curr_join, (List<Item> *) 0, curr_tmp_table)))
      {
        error= tmp_error;
        return;
      }
      end_read_record(&curr_join->join_tab->read_record);
      curr_join->const_tables= curr_join->tables; // Mark free for cleanup()
      curr_join->join_tab[0].table= 0;           // Table is freed

      // No sum funcs anymore
      if (!items2)
      {
        items2= items1 + all_fields.elements;
        if (change_to_use_tmp_fields(session, items2,
                  tmp_fields_list2, tmp_all_fields2,
                  fields_list.elements, tmp_all_fields1))
          return;
        curr_join->tmp_fields_list2= tmp_fields_list2;
        curr_join->tmp_all_fields2= tmp_all_fields2;
      }
      curr_fields_list= &curr_join->tmp_fields_list2;
      curr_all_fields= &curr_join->tmp_all_fields2;
      curr_join->set_items_ref_array(items2);
      curr_join->tmp_table_param.field_count+= curr_join->tmp_table_param.sum_func_count;
      curr_join->tmp_table_param.sum_func_count= 0;
    }
    if (curr_tmp_table->distinct)
      curr_join->select_distinct=0;		/* Each row is unique */

    curr_join->join_free();			/* Free quick selects */
    if (curr_join->select_distinct && ! curr_join->group_list)
    {
      session->set_proc_info("Removing duplicates");
      if (curr_join->tmp_having)
        curr_join->tmp_having->update_used_tables();

      if (remove_duplicates(curr_join, curr_tmp_table,
			    *curr_fields_list, curr_join->tmp_having))
      	return;
      
      curr_join->tmp_having=0;
      curr_join->select_distinct=0;
    }
    curr_tmp_table->reginfo.lock_type= TL_UNLOCK;
    if (make_simple_join(curr_join, curr_tmp_table))
      return;
    calc_group_buffer(curr_join, curr_join->group_list);
    count_field_types(select_lex, &curr_join->tmp_table_param, *curr_all_fields, 0);

  }

  if (curr_join->group || curr_join->tmp_table_param.sum_func_count)
  {
    if (make_group_fields(this, curr_join))
      return;

    if (! items3)
    {
      if (! items0)
        init_items_ref_array();
      items3= ref_pointer_array + (all_fields.elements*4);
      setup_copy_fields(session, &curr_join->tmp_table_param,
			items3, tmp_fields_list3, tmp_all_fields3,
			curr_fields_list->elements, *curr_all_fields);
      tmp_table_param.save_copy_funcs= curr_join->tmp_table_param.copy_funcs;
      tmp_table_param.save_copy_field= curr_join->tmp_table_param.copy_field;
      tmp_table_param.save_copy_field_end= curr_join->tmp_table_param.copy_field_end;
      curr_join->tmp_all_fields3= tmp_all_fields3;
      curr_join->tmp_fields_list3= tmp_fields_list3;
    }
    else
    {
      curr_join->tmp_table_param.copy_funcs= tmp_table_param.save_copy_funcs;
      curr_join->tmp_table_param.copy_field= tmp_table_param.save_copy_field;
      curr_join->tmp_table_param.copy_field_end= tmp_table_param.save_copy_field_end;
    }
    curr_fields_list= &tmp_fields_list3;
    curr_all_fields= &tmp_all_fields3;
    curr_join->set_items_ref_array(items3);

    if (curr_join->make_sum_func_list(*curr_all_fields, *curr_fields_list,
				      1, true) ||
        setup_sum_funcs(curr_join->session, curr_join->sum_funcs) ||
        session->is_fatal_error)
      return;
  }
  if (curr_join->group_list || curr_join->order)
  {
    session->set_proc_info("Sorting result");
    /* If we have already done the group, add HAVING to sorted table */
    if (curr_join->tmp_having && ! curr_join->group_list &&	! curr_join->sort_and_group)
    {
      // Some tables may have been const
      curr_join->tmp_having->update_used_tables();
      JOIN_TAB *curr_table= &curr_join->join_tab[curr_join->const_tables];
      table_map used_tables= (curr_join->const_table_map |
			      curr_table->table->map);

      Item* sort_table_cond= make_cond_for_table(curr_join->tmp_having, used_tables, used_tables, 0);
      if (sort_table_cond)
      {
        if (!curr_table->select)
          if (!(curr_table->select= new SQL_SELECT))
            return;
        if (!curr_table->select->cond)
          curr_table->select->cond= sort_table_cond;
        else					// This should never happen
        {
          if (!(curr_table->select->cond=
          new Item_cond_and(curr_table->select->cond,
                sort_table_cond)))
            return;
          /*
            Item_cond_and do not need fix_fields for execution, its parameters
            are fixed or do not need fix_fields, too
          */
          curr_table->select->cond->quick_fix_field();
        }
        curr_table->select_cond= curr_table->select->cond;
        curr_table->select_cond->top_level_item();
        curr_join->tmp_having= make_cond_for_table(curr_join->tmp_having,
                    ~ (table_map) 0,
                    ~used_tables, 0);
      }
    }
    {
      if (group)
        curr_join->select_limit= HA_POS_ERROR;
      else
      {
        /*
          We can abort sorting after session->select_limit rows if we there is no
          WHERE clause for any tables after the sorted one.
        */
        JOIN_TAB *curr_table= &curr_join->join_tab[curr_join->const_tables+1];
        JOIN_TAB *end_table= &curr_join->join_tab[curr_join->tables];
        for (; curr_table < end_table ; curr_table++)
        {
          /*
            table->keyuse is set in the case there was an original WHERE clause
            on the table that was optimized away.
          */
          if (curr_table->select_cond ||
              (curr_table->keyuse && !curr_table->first_inner))
          {
            /* We have to sort all rows */
            curr_join->select_limit= HA_POS_ERROR;
            break;
          }
        }
      }
      if (curr_join->join_tab == join_tab && save_join_tab())
        return;
      /*
        Here we sort rows for order_st BY/GROUP BY clause, if the optimiser
        chose FILESORT to be faster than INDEX SCAN or there is no
        suitable index present.
        Note, that create_sort_index calls test_if_skip_sort_order and may
        finally replace sorting with index scan if there is a LIMIT clause in
        the query. XXX: it's never shown in EXPLAIN!
        OPTION_FOUND_ROWS supersedes LIMIT and is taken into account.
      */
      if (create_sort_index(session, curr_join,
			    curr_join->group_list ?
			    curr_join->group_list : curr_join->order,
			    curr_join->select_limit,
			    (select_options & OPTION_FOUND_ROWS ?
			     HA_POS_ERROR : unit->select_limit_cnt),
                            curr_join->group_list ? true : false))
        return;

      sortorder= curr_join->sortorder;
      if (curr_join->const_tables != curr_join->tables &&
          !curr_join->join_tab[curr_join->const_tables].table->sort.io_cache)
      {
        /*
          If no IO cache exists for the first table then we are using an
          INDEX SCAN and no filesort. Thus we should not remove the sorted
          attribute on the INDEX SCAN.
        */
        skip_sort_order= 1;
      }
    }
  }
  /* XXX: When can we have here session->is_error() not zero? */
  if (session->is_error())
  {
    error= session->is_error();
    return;
  }
  curr_join->having= curr_join->tmp_having;
  curr_join->fields= curr_fields_list;

  session->set_proc_info("Sending data");
  result->send_fields(*curr_fields_list,
                      Protocol::SEND_NUM_ROWS | Protocol::SEND_EOF);
  error= do_select(curr_join, curr_fields_list, NULL);
  session->limit_found_rows= curr_join->send_records;

  /* Accumulate the counts from all join iterations of all join parts. */
  session->examined_row_count+= curr_join->examined_rows;

  /*
    With EXPLAIN EXTENDED we have to restore original ref_array
    for a derived table which is always materialized.
    Otherwise we would not be able to print the query  correctly.
  */
  if (items0 && (session->lex->describe & DESCRIBE_EXTENDED) && select_lex->linkage == DERIVED_TABLE_TYPE)
    set_items_ref_array(items0);

  return;
}

/**
  Clean up join.

  @return
    Return error that hold JOIN.
*/
int JOIN::destroy()
{
  select_lex->join= 0;

  if (tmp_join)
  {
    if (join_tab != tmp_join->join_tab)
    {
      JOIN_TAB *tab, *end;
      for (tab= join_tab, end= tab+tables ; tab != end ; tab++)
        tab->cleanup();
    }
    tmp_join->tmp_join= 0;
    tmp_table_param.copy_field=0;
    return(tmp_join->destroy());
  }
  cond_equal= 0;

  cleanup(1);
  if (exec_tmp_table1)
    exec_tmp_table1->free_tmp_table(session);
  if (exec_tmp_table2)
    exec_tmp_table2->free_tmp_table(session);
  delete select;
  delete_dynamic(&keyuse);
  return(error);
}

/**
  An entry point to single-unit select (a select without UNION).

  @param session                  thread handler
  @param rref_pointer_array   a reference to ref_pointer_array of
                              the top-level select_lex for this query
  @param tables               list of all tables used in this query.
                              The tables have been pre-opened.
  @param wild_num             number of wildcards used in the top level
                              select of this query.
                              For example statement
                              SELECT *, t1.*, catalog.t2.* FROM t0, t1, t2;
                              has 3 wildcards.
  @param fields               list of items in SELECT list of the top-level
                              select
                              e.g. SELECT a, b, c FROM t1 will have Item_field
                              for a, b and c in this list.
  @param conds                top level item of an expression representing
                              WHERE clause of the top level select
  @param og_num               total number of order_st BY and GROUP BY clauses
                              arguments
  @param order                linked list of order_st BY agruments
  @param group                linked list of GROUP BY arguments
  @param having               top level item of HAVING expression
  @param select_options       select options (BIG_RESULT, etc)
  @param result               an instance of result set handling class.
                              This object is responsible for send result
                              set rows to the client or inserting them
                              into a table.
  @param select_lex           the only Select_Lex of this query
  @param unit                 top-level UNIT of this query
                              UNIT is an artificial object created by the
                              parser for every SELECT clause.
                              e.g.
                              SELECT * FROM t1 WHERE a1 IN (SELECT * FROM t2)
                              has 2 unions.

  @retval
    false  success
  @retval
    true   an error
*/

bool
mysql_select(Session *session, Item ***rref_pointer_array,
	     TableList *tables, uint32_t wild_num, List<Item> &fields,
	     COND *conds, uint32_t og_num,  order_st *order, order_st *group,
	     Item *having, uint64_t select_options,
	     select_result *result, Select_Lex_Unit *unit,
	     Select_Lex *select_lex)
{
  bool err;
  bool free_join= 1;

  select_lex->context.resolve_in_select_list= true;
  JOIN *join;
  if (select_lex->join != 0)
  {
    join= select_lex->join;
    /*
      is it single SELECT in derived table, called in derived table
      creation
    */
    if (select_lex->linkage != DERIVED_TABLE_TYPE ||
	(select_options & SELECT_DESCRIBE))
    {
      if (select_lex->linkage != GLOBAL_OPTIONS_TYPE)
      {
	//here is EXPLAIN of subselect or derived table
	if (join->change_result(result))
	{
	  return(true);
	}
      }
      else
      {
        if ((err= join->prepare(rref_pointer_array, tables, wild_num,
                               conds, og_num, order, group, having, select_lex, unit)))
	{
	  goto err;
	}
      }
    }
    free_join= 0;
    join->select_options= select_options;
  }
  else
  {
    if (!(join= new JOIN(session, fields, select_options, result)))
	return(true);
    session->set_proc_info("init");
    session->used_tables=0;                         // Updated by setup_fields
    if ((err= join->prepare(rref_pointer_array, tables, wild_num,
                           conds, og_num, order, group, having,
                           select_lex, unit)) == true)
    {
      goto err;
    }
  }

  if (join->flatten_subqueries())
  {
    err= 1;
    goto err;
  }

  if ((err= join->optimize()))
  {
    goto err;					// 1
  }

  if (session->lex->describe & DESCRIBE_EXTENDED)
  {
    join->conds_history= join->conds;
    join->having_history= (join->having?join->having:join->tmp_having);
  }

  if (session->is_error())
    goto err;

  join->exec();

  if (session->lex->describe & DESCRIBE_EXTENDED)
  {
    select_lex->where= join->conds_history;
    select_lex->having= join->having_history;
  }

err:
  if (free_join)
  {
    session->set_proc_info("end");
    err|= select_lex->cleanup();
    return(err || session->is_error());
  }
  return(join->error);
}


int subq_sj_candidate_cmp(Item_in_subselect* const *el1,
                          Item_in_subselect* const *el2)
{
  return ((*el1)->sj_convert_priority < (*el2)->sj_convert_priority) ? 1 :
         ( ((*el1)->sj_convert_priority == (*el2)->sj_convert_priority)? 0 : -1);
}


inline Item * and_items(Item* cond, Item *item)
{
  return (cond? (new Item_cond_and(cond, item)) : item);
}


static TableList *alloc_join_nest(Session *session)
{
  TableList *tbl;
  if (!(tbl= (TableList*) session->calloc(ALIGN_SIZE(sizeof(TableList))+
                                       sizeof(nested_join_st))))
    return NULL;
  tbl->nested_join= (nested_join_st*) ((unsigned char*)tbl +
                                    ALIGN_SIZE(sizeof(TableList)));
  return tbl;
}


void fix_list_after_tbl_changes(Select_Lex *new_parent, List<TableList> *tlist)
{
  List_iterator<TableList> it(*tlist);
  TableList *table;
  while ((table= it++))
  {
    if (table->on_expr)
      table->on_expr->fix_after_pullout(new_parent, &table->on_expr);
    if (table->nested_join)
      fix_list_after_tbl_changes(new_parent, &table->nested_join->join_list);
  }
}


/*
  Convert a subquery predicate into a TableList semi-join nest

  SYNOPSIS
    convert_subq_to_sj()
       parent_join  Parent join, the one that has subq_pred in its WHERE/ON
                    clause
       subq_pred    Subquery predicate to be converted

  DESCRIPTION
    Convert a subquery predicate into a TableList semi-join nest. All the
    prerequisites are already checked, so the conversion is always successfull.

    Prepared Statements: the transformation is permanent:
     - Changes in TableList structures are naturally permanent
     - Item tree changes are performed on statement MEM_ROOT:
        = we activate statement MEM_ROOT
        = this function is called before the first fix_prepare_information
          call.

    This is intended because the criteria for subquery-to-sj conversion remain
    constant for the lifetime of the Prepared Statement.

  RETURN
    false  OK
    true   Out of memory error
*/

bool convert_subq_to_sj(JOIN *parent_join, Item_in_subselect *subq_pred)
{
  Select_Lex *parent_lex= parent_join->select_lex;
  TableList *emb_tbl_nest= NULL;
  List<TableList> *emb_join_list= &parent_lex->top_join_list;
  Session *session= parent_join->session;

  /*
    1. Find out where to put the predicate into.
     Note: for "t1 LEFT JOIN t2" this will be t2, a leaf.
  */
  if ((void*)subq_pred->expr_join_nest != (void*)1)
  {
    if (subq_pred->expr_join_nest->nested_join)
    {
      /*
        We're dealing with

          ... [LEFT] JOIN  ( ... ) ON (subquery AND whatever) ...

        The sj-nest will be inserted into the brackets nest.
      */
      emb_tbl_nest=  subq_pred->expr_join_nest;
      emb_join_list= &emb_tbl_nest->nested_join->join_list;
    }
    else if (!subq_pred->expr_join_nest->outer_join)
    {
      /*
        We're dealing with

          ... INNER JOIN tblX ON (subquery AND whatever) ...

        The sj-nest will be tblX's "sibling", i.e. another child of its
        parent. This is ok because tblX is joined as an inner join.
      */
      emb_tbl_nest= subq_pred->expr_join_nest->embedding;
      if (emb_tbl_nest)
        emb_join_list= &emb_tbl_nest->nested_join->join_list;
    }
    else if (!subq_pred->expr_join_nest->nested_join)
    {
      TableList *outer_tbl= subq_pred->expr_join_nest;
      TableList *wrap_nest;
      /*
        We're dealing with

          ... LEFT JOIN tbl ON (on_expr AND subq_pred) ...

        we'll need to convert it into:

          ... LEFT JOIN ( tbl SJ (subq_tables) ) ON (on_expr AND subq_pred) ...
                        |                      |
                        |<----- wrap_nest ---->|

        Q:  other subqueries may be pointing to this element. What to do?
        A1: simple solution: copy *subq_pred->expr_join_nest= *parent_nest.
            But we'll need to fix other pointers.
        A2: Another way: have TableList::next_ptr so the following
            subqueries know the table has been nested.
        A3: changes in the TableList::outer_join will make everything work
            automatically.
      */
      if (!(wrap_nest= alloc_join_nest(parent_join->session)))
      {
        return(true);
      }
      wrap_nest->embedding= outer_tbl->embedding;
      wrap_nest->join_list= outer_tbl->join_list;
      wrap_nest->alias= (char*) "(sj-wrap)";

      wrap_nest->nested_join->join_list.empty();
      wrap_nest->nested_join->join_list.push_back(outer_tbl);

      outer_tbl->embedding= wrap_nest;
      outer_tbl->join_list= &wrap_nest->nested_join->join_list;

      /*
        wrap_nest will take place of outer_tbl, so move the outer join flag
        and on_expr
      */
      wrap_nest->outer_join= outer_tbl->outer_join;
      outer_tbl->outer_join= 0;

      wrap_nest->on_expr= outer_tbl->on_expr;
      outer_tbl->on_expr= NULL;

      List_iterator<TableList> li(*wrap_nest->join_list);
      TableList *tbl;
      while ((tbl= li++))
      {
        if (tbl == outer_tbl)
        {
          li.replace(wrap_nest);
          break;
        }
      }
      /*
        Ok now wrap_nest 'contains' outer_tbl and we're ready to add the
        semi-join nest into it
      */
      emb_join_list= &wrap_nest->nested_join->join_list;
      emb_tbl_nest=  wrap_nest;
    }
  }

  TableList *sj_nest;
  nested_join_st *nested_join;
  if (!(sj_nest= alloc_join_nest(parent_join->session)))
  {
    return(true);
  }
  nested_join= sj_nest->nested_join;

  sj_nest->join_list= emb_join_list;
  sj_nest->embedding= emb_tbl_nest;
  sj_nest->alias= (char*) "(sj-nest)";
  /* Nests do not participate in those 'chains', so: */
  /* sj_nest->next_leaf= sj_nest->next_local= sj_nest->next_global == NULL*/
  emb_join_list->push_back(sj_nest);

  /*
    nested_join->used_tables and nested_join->not_null_tables are
    initialized in simplify_joins().
  */

  /*
    2. Walk through subquery's top list and set 'embedding' to point to the
       sj-nest.
  */
  Select_Lex *subq_lex= subq_pred->unit->first_select();
  nested_join->join_list.empty();
  List_iterator_fast<TableList> li(subq_lex->top_join_list);
  TableList *tl, *last_leaf;
  while ((tl= li++))
  {
    tl->embedding= sj_nest;
    tl->join_list= &nested_join->join_list;
    nested_join->join_list.push_back(tl);
  }

  /*
    Reconnect the next_leaf chain.
    TODO: Do we have to put subquery's tables at the end of the chain?
          Inserting them at the beginning would be a bit faster.
    NOTE: We actually insert them at the front! That's because the order is
          reversed in this list.
  */
  for (tl= parent_lex->leaf_tables; tl->next_leaf; tl= tl->next_leaf) {};
  tl->next_leaf= subq_lex->leaf_tables;
  last_leaf= tl;

  /*
    Same as above for next_local chain
    (a theory: a next_local chain always starts with ::leaf_tables
     because view's tables are inserted after the view)
  */
  for (tl= parent_lex->leaf_tables; tl->next_local; tl= tl->next_local) {};
  tl->next_local= subq_lex->leaf_tables;

  /* A theory: no need to re-connect the next_global chain */

  /* 3. Remove the original subquery predicate from the WHERE/ON */

  // The subqueries were replaced for Item_int(1) earlier
  subq_pred->exec_method= Item_in_subselect::SEMI_JOIN; // for subsequent executions
  /*TODO: also reset the 'with_subselect' there. */

  /* n. Adjust the parent_join->tables counter */
  uint32_t table_no= parent_join->tables;
  /* n. Walk through child's tables and adjust table->map */
  for (tl= subq_lex->leaf_tables; tl; tl= tl->next_leaf, table_no++)
  {
    tl->table->tablenr= table_no;
    tl->table->map= ((table_map)1) << table_no;
    Select_Lex *old_sl= tl->select_lex;
    tl->select_lex= parent_join->select_lex;
    for(TableList *emb= tl->embedding; emb && emb->select_lex == old_sl; emb= emb->embedding)
      emb->select_lex= parent_join->select_lex;
  }
  parent_join->tables += subq_lex->join->tables;

  /*
    Put the subquery's WHERE into semi-join's sj_on_expr
    Add the subquery-induced equalities too.
  */
  Select_Lex *save_lex= session->lex->current_select;
  session->lex->current_select=subq_lex;
  if (!subq_pred->left_expr->fixed &&
       subq_pred->left_expr->fix_fields(session, &subq_pred->left_expr))
    return(true);
  session->lex->current_select=save_lex;

  sj_nest->nested_join->sj_corr_tables= subq_pred->used_tables();
  sj_nest->nested_join->sj_depends_on=  subq_pred->used_tables() |
                                        subq_pred->left_expr->used_tables();
  sj_nest->sj_on_expr= subq_lex->where;

  /*
    Create the IN-equalities and inject them into semi-join's ON expression.
    Additionally, for InsideOut strategy
     - Record the number of IN-equalities.
     - Create list of pointers to (oe1, ..., ieN). We'll need the list to
       see which of the expressions are bound and which are not (for those
       we'll produce a distinct stream of (ie_i1,...ie_ik).

       (TODO: can we just create a list of pointers and hope the expressions
       will not substitute themselves on fix_fields()? or we need to wrap
       them into Item_direct_view_refs and store pointers to those. The
       pointers to Item_direct_view_refs are guaranteed to be stable as
       Item_direct_view_refs doesn't substitute itself with anything in
       Item_direct_view_ref::fix_fields.
  */
  sj_nest->sj_in_exprs= subq_pred->left_expr->cols();
  sj_nest->nested_join->sj_outer_expr_list.empty();

  if (subq_pred->left_expr->cols() == 1)
  {
    nested_join->sj_outer_expr_list.push_back(subq_pred->left_expr);

    Item *item_eq= new Item_func_eq(subq_pred->left_expr,
                                    subq_lex->ref_pointer_array[0]);
    item_eq->name= (char*)subq_sj_cond_name;
    sj_nest->sj_on_expr= and_items(sj_nest->sj_on_expr, item_eq);
  }
  else
  {
    for (uint32_t i= 0; i < subq_pred->left_expr->cols(); i++)
    {
      nested_join->sj_outer_expr_list.push_back(subq_pred->left_expr->
                                                element_index(i));
      Item *item_eq=
        new Item_func_eq(subq_pred->left_expr->element_index(i),
                         subq_lex->ref_pointer_array[i]);
      item_eq->name= (char*)subq_sj_cond_name + (i % 64);
      sj_nest->sj_on_expr= and_items(sj_nest->sj_on_expr, item_eq);
    }
  }
  /* Fix the created equality and AND */
  sj_nest->sj_on_expr->fix_fields(parent_join->session, &sj_nest->sj_on_expr);

  /*
    Walk through sj nest's WHERE and ON expressions and call
    item->fix_table_changes() for all items.
  */
  sj_nest->sj_on_expr->fix_after_pullout(parent_lex, &sj_nest->sj_on_expr);
  fix_list_after_tbl_changes(parent_lex, &sj_nest->nested_join->join_list);


  /* Unlink the child select_lex so it doesn't show up in EXPLAIN: */
  subq_lex->master_unit()->exclude_level();

  /* Inject sj_on_expr into the parent's WHERE or ON */
  if (emb_tbl_nest)
  {
    emb_tbl_nest->on_expr= and_items(emb_tbl_nest->on_expr,
                                     sj_nest->sj_on_expr);
    emb_tbl_nest->on_expr->fix_fields(parent_join->session, &emb_tbl_nest->on_expr);
  }
  else
  {
    /* Inject into the WHERE */
    parent_join->conds= and_items(parent_join->conds, sj_nest->sj_on_expr);
    parent_join->conds->fix_fields(parent_join->session, &parent_join->conds);
    parent_join->select_lex->where= parent_join->conds;
  }

  return(false);
}


/*
  Convert candidate subquery predicates to semi-joins

  SYNOPSIS
    JOIN::flatten_subqueries()

  DESCRIPTION
    Convert candidate subquery predicates to semi-joins.

  RETURN
    false  OK
    true   Error
*/

bool JOIN::flatten_subqueries()
{
  Item_in_subselect **in_subq;
  Item_in_subselect **in_subq_end;

  if (sj_subselects.elements() == 0)
    return(false);

  /* 1. Fix children subqueries */
  for (in_subq= sj_subselects.front(), in_subq_end= sj_subselects.back();
       in_subq != in_subq_end; in_subq++)
  {
    JOIN *child_join= (*in_subq)->unit->first_select()->join;
    child_join->outer_tables = child_join->tables;
    if (child_join->flatten_subqueries())
      return(true);
    (*in_subq)->sj_convert_priority=
      (*in_subq)->is_correlated * MAX_TABLES + child_join->outer_tables;
  }
  
  bool outer_join_disable_semi_join= false;
  /*
   * Temporary measure: disable semi-joins when they are together with outer
   * joins.
   *
   * @see LP Bug #314911
   */
  for (TableList *tbl= select_lex->leaf_tables; tbl; tbl=tbl->next_leaf)
  {
    TableList *embedding= tbl->embedding;
    if (tbl->on_expr || (tbl->embedding && !(embedding->sj_on_expr && 
                                            !embedding->embedding)))
    {
      in_subq= sj_subselects.front();
      outer_join_disable_semi_join= true;
    }
  }

  if (! outer_join_disable_semi_join)
  {
    /*
      2. Pick which subqueries to convert:
        sort the subquery array
        - prefer correlated subqueries over uncorrelated;
        - prefer subqueries that have greater number of outer tables;
    */
    sj_subselects.sort(subq_sj_candidate_cmp);
    // #tables-in-parent-query + #tables-in-subquery < MAX_TABLES
    /* Replace all subqueries to be flattened with Item_int(1) */
    for (in_subq= sj_subselects.front();
        in_subq != in_subq_end &&
        tables + ((*in_subq)->sj_convert_priority % MAX_TABLES) < MAX_TABLES;
        in_subq++)
    {
      if (replace_where_subcondition(this, *in_subq, new Item_int(1), false))
        return(true);
    }

    for (in_subq= sj_subselects.front();
        in_subq != in_subq_end &&
        tables + ((*in_subq)->sj_convert_priority % MAX_TABLES) < MAX_TABLES;
        in_subq++)
    {
      if (convert_subq_to_sj(this, *in_subq))
        return(true);
    }
  }

  /* 3. Finalize those we didn't convert */
  for (; in_subq!= in_subq_end; in_subq++)
  {
    JOIN *child_join= (*in_subq)->unit->first_select()->join;
    Item_subselect::trans_res res;
    (*in_subq)->changed= 0;
    (*in_subq)->fixed= 0;
    res= (*in_subq)->select_transformer(child_join);
    if (res == Item_subselect::RES_ERROR)
      return(true);

    (*in_subq)->changed= 1;
    (*in_subq)->fixed= 1;

    Item *substitute= (*in_subq)->substitution;
    bool do_fix_fields= !(*in_subq)->substitution->fixed;
    if (replace_where_subcondition(this, *in_subq, substitute, do_fix_fields))
      return(true);

    //if ((*in_subq)->fix_fields(session, (*in_subq)->ref_ptr))
    //  return(true);
  }
  sj_subselects.clear();
  return(false);
}

/**
  Setup for execution all subqueries of a query, for which the optimizer
  chose hash semi-join.

  @details Iterate over all subqueries of the query, and if they are under an
  IN predicate, and the optimizer chose to compute it via hash semi-join:
  - try to initialize all data structures needed for the materialized execution
    of the IN predicate,
  - if this fails, then perform the IN=>EXISTS transformation which was
    previously blocked during JOIN::prepare.

  This method is part of the "code generation" query processing phase.

  This phase must be called after substitute_for_best_equal_field() because
  that function may replace items with other items from a multiple equality,
  and we need to reference the correct items in the index access method of the
  IN predicate.

  @return Operation status
  @retval false     success.
  @retval true      error occurred.
*/
bool JOIN::setup_subquery_materialization()
{
  for (Select_Lex_Unit *un= select_lex->first_inner_unit(); un;
       un= un->next_unit())
  {
    for (Select_Lex *sl= un->first_select(); sl; sl= sl->next_select())
    {
      Item_subselect *subquery_predicate= sl->master_unit()->item;
      if (subquery_predicate &&
          subquery_predicate->substype() == Item_subselect::IN_SUBS)
      {
        Item_in_subselect *in_subs= (Item_in_subselect*) subquery_predicate;
        if (in_subs->exec_method == Item_in_subselect::MATERIALIZATION &&
            in_subs->setup_engine())
          return true;
      }
    }
  }
  return false;
}


/*
  Check if table's KEYUSE elements have an eq_ref(outer_tables) candidate

  SYNOPSIS
    find_eq_ref_candidate()
      table             Table to be checked
      sj_inner_tables   Bitmap of inner tables. eq_ref(inner_table) doesn't
                        count.

  DESCRIPTION
    Check if table's KEYUSE elements have an eq_ref(outer_tables) candidate

  TODO
    Check again if it is feasible to factor common parts with constant table
    search

  RETURN
    true  - There exists an eq_ref(outer-tables) candidate
    false - Otherwise
*/

bool find_eq_ref_candidate(Table *table, table_map sj_inner_tables)
{
  KEYUSE *keyuse= table->reginfo.join_tab->keyuse;
  uint32_t key;

  if (keyuse)
  {
    while (1) /* For each key */
    {
      key= keyuse->key;
      KEY *keyinfo= table->key_info + key;
      key_part_map bound_parts= 0;
      if ((keyinfo->flags & HA_NOSAME) == HA_NOSAME)
      {
        do  /* For all equalities on all key parts */
        {
          /* Check if this is "t.keypart = expr(outer_tables) */
          if (!(keyuse->used_tables & sj_inner_tables) &&
              !(keyuse->optimize & KEY_OPTIMIZE_REF_OR_NULL))
          {
            bound_parts |= 1 << keyuse->keypart;
          }
          keyuse++;
        } while (keyuse->key == key && keyuse->table == table);

        if (bound_parts == PREV_BITS(uint, keyinfo->key_parts))
          return true;
        if (keyuse->table != table)
          return false;
      }
      else
      {
        do
        {
          keyuse++;
          if (keyuse->table != table)
            return false;
        }
        while (keyuse->key == key);
      }
    }
  }
  return false;
}


/*
  Pull tables out of semi-join nests, if possible

  SYNOPSIS
    pull_out_semijoin_tables()
      join  The join where to do the semi-join flattening

  DESCRIPTION
    Try to pull tables out of semi-join nests.

    PRECONDITIONS
    When this function is called, the join may have several semi-join nests
    (possibly within different semi-join nests), but it is guaranteed that
    one semi-join nest does not contain another.

    ACTION
    A table can be pulled out of the semi-join nest if
     - It is a constant table
     - It is accessed

    POSTCONDITIONS
     * Pulled out tables have JOIN_TAB::emb_sj_nest == NULL (like the outer
       tables)
     * Tables that were not pulled out have JOIN_TAB::emb_sj_nest.
     * Semi-join nests TableList::sj_inner_tables

    This operation is (and should be) performed at each PS execution since
    tables may become/cease to be constant across PS reexecutions.

  RETURN
    0 - OK
    1 - Out of memory error
*/

int pull_out_semijoin_tables(JOIN *join)
{
  TableList *sj_nest;
  List_iterator<TableList> sj_list_it(join->select_lex->sj_nests);

  /* Try pulling out of the each of the semi-joins */
  while ((sj_nest= sj_list_it++))
  {
    /* Action #1: Mark the constant tables to be pulled out */
    table_map pulled_tables= 0;

    List_iterator<TableList> child_li(sj_nest->nested_join->join_list);
    TableList *tbl;
    while ((tbl= child_li++))
    {
      if (tbl->table)
      {
        tbl->table->reginfo.join_tab->emb_sj_nest= sj_nest;
        if (tbl->table->map & join->const_table_map)
        {
          pulled_tables |= tbl->table->map;
        }
      }
    }

    /*
      Action #2: Find which tables we can pull out based on
      update_ref_and_keys() data. Note that pulling one table out can allow
      us to pull out some other tables too.
    */
    bool pulled_a_table;
    do
    {
      pulled_a_table= false;
      child_li.rewind();
      while ((tbl= child_li++))
      {
        if (tbl->table && !(pulled_tables & tbl->table->map))
        {
          if (find_eq_ref_candidate(tbl->table,
                                    sj_nest->nested_join->used_tables &
                                    ~pulled_tables))
          {
            pulled_a_table= true;
            pulled_tables |= tbl->table->map;
          }
        }
      }
    } while (pulled_a_table);

    child_li.rewind();
    if ((sj_nest)->nested_join->used_tables == pulled_tables)
    {
      (sj_nest)->sj_inner_tables= 0;
      while ((tbl= child_li++))
      {
        if (tbl->table)
          tbl->table->reginfo.join_tab->emb_sj_nest= NULL;
      }
    }
    else
    {
      /* Record the bitmap of inner tables, mark the inner tables */
      table_map inner_tables=(sj_nest)->nested_join->used_tables &
                             ~pulled_tables;
      (sj_nest)->sj_inner_tables= inner_tables;
      while ((tbl= child_li++))
      {
        if (tbl->table)
        {
          if (inner_tables & tbl->table->map)
            tbl->table->reginfo.join_tab->emb_sj_nest= (sj_nest);
          else
            tbl->table->reginfo.join_tab->emb_sj_nest= NULL;
        }
      }
    }
  }
  return(0);
}

/*****************************************************************************
  Create JOIN_TABS, make a guess about the table types,
  Approximate how many records will be used in each table
*****************************************************************************/


static ha_rows get_quick_record_count(Session *session, SQL_SELECT *select,
				      Table *table,
				      const key_map *keys,ha_rows limit)
{
  int error;
  if (check_stack_overrun(session, STACK_MIN_SIZE, NULL))
    return(0);                           // Fatal error flag is set
  if (select)
  {
    select->head=table;
    table->reginfo.impossible_range=0;
    if ((error= select->test_quick_select(session, *(key_map *)keys,(table_map) 0,
                                          limit, 0, false)) == 1)
      return(select->quick->records);
    if (error == -1)
    {
      table->reginfo.impossible_range=1;
      return(0);
    }
  }
  return(HA_POS_ERROR);			/* This shouldn't happend */
}

/*
   This structure is used to collect info on potentially sargable
   predicates in order to check whether they become sargable after
   reading const tables.
   We form a bitmap of indexes that can be used for sargable predicates.
   Only such indexes are involved in range analysis.
*/
typedef struct st_sargable_param
{
  Field *field;              /* field against which to check sargability */
  Item **arg_value;          /* values of potential keys for lookups     */
  uint32_t num_values;           /* number of values in the above array      */
} SARGABLE_PARAM;

/**
  Calculate the best possible join and initialize the join structure.

  @retval
    0	ok
  @retval
    1	Fatal error
*/

static bool
make_join_statistics(JOIN *join, TableList *tables, COND *conds,
		     DYNAMIC_ARRAY *keyuse_array)
{
  int error;
  Table *table;
  uint32_t i,table_count,const_count,key;
  table_map found_const_table_map, all_table_map, found_ref, refs;
  key_map const_ref, eq_part;
  Table **table_vector;
  JOIN_TAB *stat,*stat_end,*s,**stat_ref;
  KEYUSE *keyuse,*start_keyuse;
  table_map outer_join=0;
  SARGABLE_PARAM *sargables= 0;
  JOIN_TAB *stat_vector[MAX_TABLES+1];

  table_count=join->tables;
  stat=(JOIN_TAB*) join->session->calloc(sizeof(JOIN_TAB)*table_count);
  stat_ref=(JOIN_TAB**) join->session->alloc(sizeof(JOIN_TAB*)*MAX_TABLES);
  table_vector=(Table**) join->session->alloc(sizeof(Table*)*(table_count*2));
  if (!stat || !stat_ref || !table_vector)
    return(1);				// Eom /* purecov: inspected */

  join->best_ref=stat_vector;

  stat_end=stat+table_count;
  found_const_table_map= all_table_map=0;
  const_count=0;

  for (s= stat, i= 0;
       tables;
       s++, tables= tables->next_leaf, i++)
  {
    TableList *embedding= tables->embedding;
    stat_vector[i]=s;
    s->keys.reset();
    s->const_keys.reset();
    s->checked_keys.reset();
    s->needed_reg.reset();
    table_vector[i]=s->table=table=tables->table;
    table->pos_in_table_list= tables;
    error= table->file->info(HA_STATUS_VARIABLE | HA_STATUS_NO_LOCK);
    if(error)
    {
        table->file->print_error(error, MYF(0));
        return(1);
    }
    table->quick_keys.reset();
    table->reginfo.join_tab=s;
    table->reginfo.not_exists_optimize=0;
    memset(table->const_key_parts, 0,
           sizeof(key_part_map)*table->s->keys);
    all_table_map|= table->map;
    s->join=join;
    s->info=0;					// For describe

    s->dependent= tables->dep_tables;
    s->key_dependent= 0;
    if (tables->schema_table)
      table->file->stats.records= 2;
    table->quick_condition_rows= table->file->stats.records;

    s->on_expr_ref= &tables->on_expr;
    if (*s->on_expr_ref)
    {
      /* s is the only inner table of an outer join */
      if (!table->file->stats.records && !embedding)
      {						// Empty table
        s->dependent= 0;                        // Ignore LEFT JOIN depend.
	set_position(join,const_count++,s,(KEYUSE*) 0);
	continue;
      }
      outer_join|= table->map;
      s->embedding_map= 0;
      for (;embedding; embedding= embedding->embedding)
        s->embedding_map|= embedding->nested_join->nj_map;
      continue;
    }
    if (embedding && !(embedding->sj_on_expr && ! embedding->embedding))
    {
      /* s belongs to a nested join, maybe to several embedded joins */
      s->embedding_map= 0;
      do
      {
        nested_join_st *nested_join= embedding->nested_join;
        s->embedding_map|=nested_join->nj_map;
        s->dependent|= embedding->dep_tables;
        embedding= embedding->embedding;
        outer_join|= nested_join->used_tables;
      }
      while (embedding);
      continue;
    }
    if ((table->file->stats.records <= 1) &&
	!s->dependent &&
	(table->file->ha_table_flags() & HA_STATS_RECORDS_IS_EXACT) && !join->no_const_tables)
    {
      set_position(join,const_count++,s,(KEYUSE*) 0);
    }
  }
  stat_vector[i]=0;
  join->outer_join=outer_join;

  if (join->outer_join)
  {
    /*
       Build transitive closure for relation 'to be dependent on'.
       This will speed up the plan search for many cases with outer joins,
       as well as allow us to catch illegal cross references/
       Warshall's algorithm is used to build the transitive closure.
       As we use bitmaps to represent the relation the complexity
       of the algorithm is O((number of tables)^2).
    */
    for (i= 0, s= stat ; i < table_count ; i++, s++)
    {
      for (uint32_t j= 0 ; j < table_count ; j++)
      {
        table= stat[j].table;
        if (s->dependent & table->map)
          s->dependent |= table->reginfo.join_tab->dependent;
      }
      if (s->dependent)
        s->table->maybe_null= 1;
    }
    /* Catch illegal cross references for outer joins */
    for (i= 0, s= stat ; i < table_count ; i++, s++)
    {
      if (s->dependent & s->table->map)
      {
        join->tables=0;			// Don't use join->table
        my_message(ER_WRONG_OUTER_JOIN, ER(ER_WRONG_OUTER_JOIN), MYF(0));
        return(1);
      }
      s->key_dependent= s->dependent;
    }
  }

  if (conds || outer_join)
    if (update_ref_and_keys(join->session, keyuse_array, stat, join->tables,
                            conds, join->cond_equal,
                            ~outer_join, join->select_lex, &sargables))
      return(1);

  /* Read tables with 0 or 1 rows (system tables) */
  join->const_table_map= 0;

  for (POSITION *p_pos=join->positions, *p_end=p_pos+const_count;
       p_pos < p_end ;
       p_pos++)
  {
    int tmp;
    s= p_pos->table;
    s->type=JT_SYSTEM;
    join->const_table_map|=s->table->map;
    if ((tmp=join_read_const_table(s, p_pos)))
    {
      if (tmp > 0)
	return(1);			// Fatal error
    }
    else
      found_const_table_map|= s->table->map;
  }

  /* loop until no more const tables are found */
  int ref_changed;
  do
  {
  more_const_tables_found:
    ref_changed = 0;
    found_ref=0;

    /*
      We only have to loop from stat_vector + const_count as
      set_position() will move all const_tables first in stat_vector
    */

    for (JOIN_TAB **pos=stat_vector+const_count ; (s= *pos) ; pos++)
    {
      table=s->table;

      /*
        If equi-join condition by a key is null rejecting and after a
        substitution of a const table the key value happens to be null
        then we can state that there are no matches for this equi-join.
      */
      if ((keyuse= s->keyuse) && *s->on_expr_ref && !s->embedding_map)
      {
        /*
          When performing an outer join operation if there are no matching rows
          for the single row of the outer table all the inner tables are to be
          null complemented and thus considered as constant tables.
          Here we apply this consideration to the case of outer join operations
          with a single inner table only because the case with nested tables
          would require a more thorough analysis.
          TODO. Apply single row substitution to null complemented inner tables
          for nested outer join operations.
	*/
        while (keyuse->table == table)
        {
          if (!(keyuse->val->used_tables() & ~join->const_table_map) &&
              keyuse->val->is_null() && keyuse->null_rejecting)
          {
            s->type= JT_CONST;
            table->mark_as_null_row();
            found_const_table_map|= table->map;
	    join->const_table_map|= table->map;
	    set_position(join,const_count++,s,(KEYUSE*) 0);
            goto more_const_tables_found;
           }
	  keyuse++;
        }
      }

      if (s->dependent)				// If dependent on some table
      {
	// All dep. must be constants
	if (s->dependent & ~(found_const_table_map))
	  continue;
	if (table->file->stats.records <= 1L &&
	    (table->file->ha_table_flags() & HA_STATS_RECORDS_IS_EXACT) &&
            !table->pos_in_table_list->embedding)
	{					// system table
	  int tmp= 0;
	  s->type=JT_SYSTEM;
	  join->const_table_map|=table->map;
	  set_position(join,const_count++,s,(KEYUSE*) 0);
	  if ((tmp= join_read_const_table(s, join->positions+const_count-1)))
	  {
	    if (tmp > 0)
	      return(1);			// Fatal error
	  }
	  else
	    found_const_table_map|= table->map;
	  continue;
	}
      }
      /* check if table can be read by key or table only uses const refs */
      if ((keyuse=s->keyuse))
      {
	s->type= JT_REF;
	while (keyuse->table == table)
	{
	  start_keyuse=keyuse;
	  key=keyuse->key;
	  s->keys.set(key);               // QQ: remove this ?

	  refs=0;
          const_ref.reset();
	  eq_part.reset();
	  do
	  {
	    if (keyuse->val->type() != Item::NULL_ITEM && !keyuse->optimize)
	    {
	      if (!((~found_const_table_map) & keyuse->used_tables))
		const_ref.set(keyuse->keypart);
	      else
		refs|=keyuse->used_tables;
	      eq_part.set(keyuse->keypart);
	    }
	    keyuse++;
	  } while (keyuse->table == table && keyuse->key == key);

	  if (is_keymap_prefix(eq_part, table->key_info[key].key_parts) &&
              !table->pos_in_table_list->embedding)
	  {
            if ((table->key_info[key].flags & (HA_NOSAME))
                 == HA_NOSAME)
            {
	      if (const_ref == eq_part)
	      {					// Found everything for ref.
	        int tmp;
	        ref_changed = 1;
	        s->type= JT_CONST;
	        join->const_table_map|= table->map;
	        set_position(join,const_count++,s,start_keyuse);
	        if (create_ref_for_key(join, s, start_keyuse,
				       found_const_table_map))
		  return(1);
	        if ((tmp=join_read_const_table(s,
                                               join->positions+const_count-1)))
	        {
		  if (tmp > 0)
		    return(1);			// Fatal error
	        }
	        else
		  found_const_table_map|= table->map;
	        break;
	      }
	      else
	        found_ref|= refs;      // Table is const if all refs are const
	    }
            else if (const_ref == eq_part)
              s->const_keys.set(key);
          }
	}
      }
    }
  } while (join->const_table_map & found_ref && ref_changed);

  /*
    Update info on indexes that can be used for search lookups as
    reading const tables may has added new sargable predicates.
  */
  if (const_count && sargables)
  {
    for( ; sargables->field ; sargables++)
    {
      Field *field= sargables->field;
      JOIN_TAB *join_tab= field->table->reginfo.join_tab;
      key_map possible_keys= field->key_start;
      possible_keys&= field->table->keys_in_use_for_query;
      bool is_const= 1;
      for (uint32_t j=0; j < sargables->num_values; j++)
        is_const&= sargables->arg_value[j]->const_item();
      if (is_const)
        join_tab[0].const_keys|= possible_keys;
    }
  }

  if (pull_out_semijoin_tables(join))
    return(true);

  /* Calc how many (possible) matched records in each table */

  for (s=stat ; s < stat_end ; s++)
  {
    if (s->type == JT_SYSTEM || s->type == JT_CONST)
    {
      /* Only one matching row */
      s->found_records=s->records=s->read_time=1; s->worst_seeks=1.0;
      continue;
    }
    /* Approximate found rows and time to read them */
    s->found_records=s->records=s->table->file->stats.records;
    s->read_time=(ha_rows) s->table->file->scan_time();

    /*
      Set a max range of how many seeks we can expect when using keys
      This is can't be to high as otherwise we are likely to use
      table scan.
    */
    s->worst_seeks= cmin((double) s->found_records / 10,
			(double) s->read_time*3);
    if (s->worst_seeks < 2.0)			// Fix for small tables
      s->worst_seeks=2.0;

    /*
      Add to stat->const_keys those indexes for which all group fields or
      all select distinct fields participate in one index.
    */
    add_group_and_distinct_keys(join, s);

    if (s->const_keys.any() &&
        !s->table->pos_in_table_list->embedding)
    {
      ha_rows records;
      SQL_SELECT *select;
      select= make_select(s->table, found_const_table_map,
			  found_const_table_map,
			  *s->on_expr_ref ? *s->on_expr_ref : conds,
			  1, &error);
      if (!select)
        return(1);
      records= get_quick_record_count(join->session, select, s->table,
				      &s->const_keys, join->row_limit);
      s->quick=select->quick;
      s->needed_reg=select->needed_reg;
      select->quick=0;
      if (records == 0 && s->table->reginfo.impossible_range)
      {
        /*
          Impossible WHERE or ON expression
          In case of ON, we mark that the we match one empty NULL row.
          In case of WHERE, don't set found_const_table_map to get the
          caller to abort with a zero row result.
        */
        join->const_table_map|= s->table->map;
        set_position(join,const_count++,s,(KEYUSE*) 0);
        s->type= JT_CONST;
        if (*s->on_expr_ref)
        {
          /* Generate empty row */
          s->info= "Impossible ON condition";
          found_const_table_map|= s->table->map;
          s->type= JT_CONST;
          s->table->mark_as_null_row();		// All fields are NULL
        }
      }
      if (records != HA_POS_ERROR)
      {
        s->found_records=records;
        s->read_time= (ha_rows) (s->quick ? s->quick->read_time : 0.0);
      }
      delete select;
    }
  }

  join->join_tab=stat;
  join->map2table=stat_ref;
  join->table= join->all_tables=table_vector;
  join->const_tables=const_count;
  join->found_const_table_map=found_const_table_map;

  /* Find an optimal join order of the non-constant tables. */
  if (join->const_tables != join->tables)
  {
    optimize_keyuse(join, keyuse_array);
    if (choose_plan(join, all_table_map & ~join->const_table_map))
      return(true);
  }
  else
  {
    memcpy(join->best_positions, join->positions,
           sizeof(POSITION)*join->const_tables);
    join->best_read=1.0;
  }
  /* Generate an execution plan from the found optimal join order. */
  return(join->session->killed || get_best_combination(join));
}


/*****************************************************************************
  Check with keys are used and with tables references with tables
  Updates in stat:
	  keys	     Bitmap of all used keys
	  const_keys Bitmap of all keys with may be used with quick_select
	  keyuse     Pointer to possible keys
*****************************************************************************/

/// Used when finding key fields
typedef struct key_field_t {
  Field		*field;
  Item		*val;			///< May be empty if diff constant
  uint		level;
  uint		optimize; // KEY_OPTIMIZE_*
  bool		eq_func;
  /**
    If true, the condition this struct represents will not be satisfied
    when val IS NULL.
  */
  bool          null_rejecting;
  bool          *cond_guard; /* See KEYUSE::cond_guard */
  uint32_t          sj_pred_no; /* See KEYUSE::sj_pred_no */
} KEY_FIELD;

/**
  Merge new key definitions to old ones, remove those not used in both.

  This is called for OR between different levels.

  To be able to do 'ref_or_null' we merge a comparison of a column
  and 'column IS NULL' to one test.  This is useful for sub select queries
  that are internally transformed to something like:.

  @code
  SELECT * FROM t1 WHERE t1.key=outer_ref_field or t1.key IS NULL
  @endcode

  KEY_FIELD::null_rejecting is processed as follows: @n
  result has null_rejecting=true if it is set for both ORed references.
  for example:
  -   (t2.key = t1.field OR t2.key  =  t1.field) -> null_rejecting=true
  -   (t2.key = t1.field OR t2.key <=> t1.field) -> null_rejecting=false

  @todo
    The result of this is that we're missing some 'ref' accesses.
    OptimizerTeam: Fix this
*/

static KEY_FIELD *
merge_key_fields(KEY_FIELD *start,KEY_FIELD *new_fields,KEY_FIELD *end,
		 uint32_t and_level)
{
  if (start == new_fields)
    return start;				// Impossible or
  if (new_fields == end)
    return start;				// No new fields, skip all

  KEY_FIELD *first_free=new_fields;

  /* Mark all found fields in old array */
  for (; new_fields != end ; new_fields++)
  {
    for (KEY_FIELD *old=start ; old != first_free ; old++)
    {
      if (old->field == new_fields->field)
      {
        /*
          NOTE: below const_item() call really works as "!used_tables()", i.e.
          it can return false where it is feasible to make it return true.

          The cause is as follows: Some of the tables are already known to be
          const tables (the detection code is in make_join_statistics(),
          above the update_ref_and_keys() call), but we didn't propagate
          information about this: Table::const_table is not set to true, and
          Item::update_used_tables() hasn't been called for each item.
          The result of this is that we're missing some 'ref' accesses.
          TODO: OptimizerTeam: Fix this
        */
	if (!new_fields->val->const_item())
	{
	  /*
	    If the value matches, we can use the key reference.
	    If not, we keep it until we have examined all new values
	  */
	  if (old->val->eq(new_fields->val, old->field->binary()))
	  {
	    old->level= and_level;
	    old->optimize= ((old->optimize & new_fields->optimize &
			     KEY_OPTIMIZE_EXISTS) |
			    ((old->optimize | new_fields->optimize) &
			     KEY_OPTIMIZE_REF_OR_NULL));
            old->null_rejecting= (old->null_rejecting &&
                                  new_fields->null_rejecting);
	  }
	}
	else if (old->eq_func && new_fields->eq_func &&
                 old->val->eq_by_collation(new_fields->val,
                                           old->field->binary(),
                                           old->field->charset()))

	{
	  old->level= and_level;
	  old->optimize= ((old->optimize & new_fields->optimize &
			   KEY_OPTIMIZE_EXISTS) |
			  ((old->optimize | new_fields->optimize) &
			   KEY_OPTIMIZE_REF_OR_NULL));
          old->null_rejecting= (old->null_rejecting &&
                                new_fields->null_rejecting);
	}
	else if (old->eq_func && new_fields->eq_func &&
		 ((old->val->const_item() && old->val->is_null()) ||
                  new_fields->val->is_null()))
	{
	  /* field = expression OR field IS NULL */
	  old->level= and_level;
	  old->optimize= KEY_OPTIMIZE_REF_OR_NULL;
	  /*
            Remember the NOT NULL value unless the value does not depend
            on other tables.
          */
	  if (!old->val->used_tables() && old->val->is_null())
	    old->val= new_fields->val;
          /* The referred expression can be NULL: */
          old->null_rejecting= 0;
	}
	else
	{
	  /*
	    We are comparing two different const.  In this case we can't
	    use a key-lookup on this so it's better to remove the value
	    and let the range optimzier handle it
	  */
	  if (old == --first_free)		// If last item
	    break;
	  *old= *first_free;			// Remove old value
	  old--;				// Retry this value
	}
      }
    }
  }
  /* Remove all not used items */
  for (KEY_FIELD *old=start ; old != first_free ;)
  {
    if (old->level != and_level)
    {						// Not used in all levels
      if (old == --first_free)
	break;
      *old= *first_free;			// Remove old value
      continue;
    }
    old++;
  }
  return first_free;
}


/**
  Add a possible key to array of possible keys if it's usable as a key

    @param key_fields      Pointer to add key, if usable
    @param and_level       And level, to be stored in KEY_FIELD
    @param cond            Condition predicate
    @param field           Field used in comparision
    @param eq_func         True if we used =, <=> or IS NULL
    @param value           Value used for comparison with field
    @param usable_tables   Tables which can be used for key optimization
    @param sargables       IN/OUT Array of found sargable candidates

  @note
    If we are doing a NOT NULL comparison on a NOT NULL field in a outer join
    table, we store this to be able to do not exists optimization later.

  @returns
    *key_fields is incremented if we stored a key in the array
*/

static void
add_key_field(KEY_FIELD **key_fields,uint32_t and_level, Item_func *cond,
              Field *field, bool eq_func, Item **value, uint32_t num_values,
              table_map usable_tables, SARGABLE_PARAM **sargables)
{
  uint32_t exists_optimize= 0;
  if (!(field->flags & PART_KEY_FLAG))
  {
    // Don't remove column IS NULL on a LEFT JOIN table
    if (!eq_func || (*value)->type() != Item::NULL_ITEM ||
        !field->table->maybe_null || field->null_ptr)
      return;					// Not a key. Skip it
    exists_optimize= KEY_OPTIMIZE_EXISTS;
    assert(num_values == 1);
  }
  else
  {
    table_map used_tables=0;
    bool optimizable=0;
    for (uint32_t i=0; i<num_values; i++)
    {
      used_tables|=(value[i])->used_tables();
      if (!((value[i])->used_tables() & (field->table->map | RAND_TABLE_BIT)))
        optimizable=1;
    }
    if (!optimizable)
      return;
    if (!(usable_tables & field->table->map))
    {
      if (!eq_func || (*value)->type() != Item::NULL_ITEM ||
          !field->table->maybe_null || field->null_ptr)
	return;					// Can't use left join optimize
      exists_optimize= KEY_OPTIMIZE_EXISTS;
    }
    else
    {
      JOIN_TAB *stat=field->table->reginfo.join_tab;
      key_map possible_keys= field->key_start;
      possible_keys&= field->table->keys_in_use_for_query;
      stat[0].keys|= possible_keys;             // Add possible keys

      /*
	Save the following cases:
	Field op constant
	Field LIKE constant where constant doesn't start with a wildcard
	Field = field2 where field2 is in a different table
	Field op formula
	Field IS NULL
	Field IS NOT NULL
         Field BETWEEN ...
         Field IN ...
      */
      stat[0].key_dependent|= used_tables;

      bool is_const=1;
      for (uint32_t i=0; i<num_values; i++)
      {
        if (!(is_const&= value[i]->const_item()))
          break;
      }
      if (is_const)
        stat[0].const_keys|= possible_keys;
      else if (!eq_func)
      {
        /*
          Save info to be able check whether this predicate can be
          considered as sargable for range analisis after reading const tables.
          We do not save info about equalities as update_const_equal_items
          will take care of updating info on keys from sargable equalities.
        */
        (*sargables)--;
        (*sargables)->field= field;
        (*sargables)->arg_value= value;
        (*sargables)->num_values= num_values;
      }
      /*
	We can't always use indexes when comparing a string index to a
	number. cmp_type() is checked to allow compare of dates to numbers.
        eq_func is NEVER true when num_values > 1
       */
      if (!eq_func)
      {
        /*
          Additional optimization: if we're processing
          "t.key BETWEEN c1 AND c1" then proceed as if we were processing
          "t.key = c1".
          TODO: This is a very limited fix. A more generic fix is possible.
          There are 2 options:
          A) Make equality propagation code be able to handle BETWEEN
             (including cases like t1.key BETWEEN t2.key AND t3.key)
          B) Make range optimizer to infer additional "t.key = c" equalities
             and use them in equality propagation process (see details in
             OptimizerKBAndTodo)
        */
        if ((cond->functype() != Item_func::BETWEEN) ||
            ((Item_func_between*) cond)->negated ||
            !value[0]->eq(value[1], field->binary()))
          return;
        eq_func= true;
      }

      if (field->result_type() == STRING_RESULT)
      {
        if ((*value)->result_type() != STRING_RESULT)
        {
          if (field->cmp_type() != (*value)->result_type())
            return;
        }
        else
        {
          /*
            We can't use indexes if the effective collation
            of the operation differ from the field collation.
          */
          if (field->cmp_type() == STRING_RESULT &&
              ((Field_str*)field)->charset() != cond->compare_collation())
            return;
        }
      }
    }
  }
  /*
    For the moment eq_func is always true. This slot is reserved for future
    extensions where we want to remembers other things than just eq comparisons
  */
  assert(eq_func);
  /* Store possible eq field */
  (*key_fields)->field=		field;
  (*key_fields)->eq_func=	eq_func;
  (*key_fields)->val=		*value;
  (*key_fields)->level=		and_level;
  (*key_fields)->optimize=	exists_optimize;
  /*
    If the condition has form "tbl.keypart = othertbl.field" and
    othertbl.field can be NULL, there will be no matches if othertbl.field
    has NULL value.
    We use null_rejecting in add_not_null_conds() to add
    'othertbl.field IS NOT NULL' to tab->select_cond.
  */
  (*key_fields)->null_rejecting= ((cond->functype() == Item_func::EQ_FUNC ||
                                   cond->functype() == Item_func::MULT_EQUAL_FUNC) &&
                                  ((*value)->type() == Item::FIELD_ITEM) &&
                                  ((Item_field*)*value)->field->maybe_null());
  (*key_fields)->cond_guard= NULL;
  (*key_fields)->sj_pred_no= (cond->name >= subq_sj_cond_name &&
                              cond->name < subq_sj_cond_name + 64)?
                              cond->name - subq_sj_cond_name: UINT_MAX;
  (*key_fields)++;
}

/**
  Add possible keys to array of possible keys originated from a simple
  predicate.

    @param  key_fields     Pointer to add key, if usable
    @param  and_level      And level, to be stored in KEY_FIELD
    @param  cond           Condition predicate
    @param  field          Field used in comparision
    @param  eq_func        True if we used =, <=> or IS NULL
    @param  value          Value used for comparison with field
                           Is NULL for BETWEEN and IN
    @param  usable_tables  Tables which can be used for key optimization
    @param  sargables      IN/OUT Array of found sargable candidates

  @note
    If field items f1 and f2 belong to the same multiple equality and
    a key is added for f1, the the same key is added for f2.

  @returns
    *key_fields is incremented if we stored a key in the array
*/

static void
add_key_equal_fields(KEY_FIELD **key_fields, uint32_t and_level,
                     Item_func *cond, Item_field *field_item,
                     bool eq_func, Item **val,
                     uint32_t num_values, table_map usable_tables,
                     SARGABLE_PARAM **sargables)
{
  Field *field= field_item->field;
  add_key_field(key_fields, and_level, cond, field,
                eq_func, val, num_values, usable_tables, sargables);
  Item_equal *item_equal= field_item->item_equal;
  if (item_equal)
  {
    /*
      Add to the set of possible key values every substitution of
      the field for an equal field included into item_equal
    */
    Item_equal_iterator it(*item_equal);
    Item_field *item;
    while ((item= it++))
    {
      if (!field->eq(item->field))
      {
        add_key_field(key_fields, and_level, cond, item->field,
                      eq_func, val, num_values, usable_tables,
                      sargables);
      }
    }
  }
}

static void
add_key_fields(JOIN *join, KEY_FIELD **key_fields, uint32_t *and_level,
               COND *cond, table_map usable_tables,
               SARGABLE_PARAM **sargables)
{
  if (cond->type() == Item_func::COND_ITEM)
  {
    List_iterator_fast<Item> li(*((Item_cond*) cond)->argument_list());
    KEY_FIELD *org_key_fields= *key_fields;

    if (((Item_cond*) cond)->functype() == Item_func::COND_AND_FUNC)
    {
      Item *item;
      while ((item=li++))
        add_key_fields(join, key_fields, and_level, item, usable_tables,
                       sargables);
      for (; org_key_fields != *key_fields ; org_key_fields++)
	org_key_fields->level= *and_level;
    }
    else
    {
      (*and_level)++;
      add_key_fields(join, key_fields, and_level, li++, usable_tables,
                     sargables);
      Item *item;
      while ((item=li++))
      {
	KEY_FIELD *start_key_fields= *key_fields;
	(*and_level)++;
        add_key_fields(join, key_fields, and_level, item, usable_tables,
                       sargables);
	*key_fields=merge_key_fields(org_key_fields,start_key_fields,
				     *key_fields,++(*and_level));
      }
    }
    return;
  }

  /*
    Subquery optimization: Conditions that are pushed down into subqueries
    are wrapped into Item_func_trig_cond. We process the wrapped condition
    but need to set cond_guard for KEYUSE elements generated from it.
  */
  {
    if (cond->type() == Item::FUNC_ITEM &&
        ((Item_func*)cond)->functype() == Item_func::TRIG_COND_FUNC)
    {
      Item *cond_arg= ((Item_func*)cond)->arguments()[0];
      if (!join->group_list && !join->order &&
          join->unit->item &&
          join->unit->item->substype() == Item_subselect::IN_SUBS &&
          !join->unit->is_union())
      {
        KEY_FIELD *save= *key_fields;
        add_key_fields(join, key_fields, and_level, cond_arg, usable_tables,
                       sargables);
        // Indicate that this ref access candidate is for subquery lookup:
        for (; save != *key_fields; save++)
          save->cond_guard= ((Item_func_trig_cond*)cond)->get_trig_var();
      }
      return;
    }
  }

  /* If item is of type 'field op field/constant' add it to key_fields */
  if (cond->type() != Item::FUNC_ITEM)
    return;
  Item_func *cond_func= (Item_func*) cond;
  switch (cond_func->select_optimize()) {
  case Item_func::OPTIMIZE_NONE:
    break;
  case Item_func::OPTIMIZE_KEY:
  {
    Item **values;
    // BETWEEN, IN, NE
    if (cond_func->key_item()->real_item()->type() == Item::FIELD_ITEM &&
	!(cond_func->used_tables() & OUTER_REF_TABLE_BIT))
    {
      values= cond_func->arguments()+1;
      if (cond_func->functype() == Item_func::NE_FUNC &&
        cond_func->arguments()[1]->real_item()->type() == Item::FIELD_ITEM &&
	     !(cond_func->arguments()[0]->used_tables() & OUTER_REF_TABLE_BIT))
        values--;
      assert(cond_func->functype() != Item_func::IN_FUNC ||
                  cond_func->argument_count() != 2);
      add_key_equal_fields(key_fields, *and_level, cond_func,
                           (Item_field*) (cond_func->key_item()->real_item()),
                           0, values,
                           cond_func->argument_count()-1,
                           usable_tables, sargables);
    }
    if (cond_func->functype() == Item_func::BETWEEN)
    {
      values= cond_func->arguments();
      for (uint32_t i= 1 ; i < cond_func->argument_count() ; i++)
      {
        Item_field *field_item;
        if (cond_func->arguments()[i]->real_item()->type() == Item::FIELD_ITEM
            &&
            !(cond_func->arguments()[i]->used_tables() & OUTER_REF_TABLE_BIT))
        {
          field_item= (Item_field *) (cond_func->arguments()[i]->real_item());
          add_key_equal_fields(key_fields, *and_level, cond_func,
                               field_item, 0, values, 1, usable_tables,
                               sargables);
        }
      }
    }
    break;
  }
  case Item_func::OPTIMIZE_OP:
  {
    bool equal_func=(cond_func->functype() == Item_func::EQ_FUNC ||
		     cond_func->functype() == Item_func::EQUAL_FUNC);

    if (cond_func->arguments()[0]->real_item()->type() == Item::FIELD_ITEM &&
	!(cond_func->arguments()[0]->used_tables() & OUTER_REF_TABLE_BIT))
    {
      add_key_equal_fields(key_fields, *and_level, cond_func,
	                (Item_field*) (cond_func->arguments()[0])->real_item(),
		           equal_func,
                           cond_func->arguments()+1, 1, usable_tables,
                           sargables);
    }
    if (cond_func->arguments()[1]->real_item()->type() == Item::FIELD_ITEM &&
	cond_func->functype() != Item_func::LIKE_FUNC &&
	!(cond_func->arguments()[1]->used_tables() & OUTER_REF_TABLE_BIT))
    {
      add_key_equal_fields(key_fields, *and_level, cond_func,
                       (Item_field*) (cond_func->arguments()[1])->real_item(),
		           equal_func,
                           cond_func->arguments(),1,usable_tables,
                           sargables);
    }
    break;
  }
  case Item_func::OPTIMIZE_NULL:
    /* column_name IS [NOT] NULL */
    if (cond_func->arguments()[0]->real_item()->type() == Item::FIELD_ITEM &&
	!(cond_func->used_tables() & OUTER_REF_TABLE_BIT))
    {
      Item *tmp=new Item_null;
      if (unlikely(!tmp))                       // Should never be true
	return;
      add_key_equal_fields(key_fields, *and_level, cond_func,
		    (Item_field*) (cond_func->arguments()[0])->real_item(),
		    cond_func->functype() == Item_func::ISNULL_FUNC,
			   &tmp, 1, usable_tables, sargables);
    }
    break;
  case Item_func::OPTIMIZE_EQUAL:
    Item_equal *item_equal= (Item_equal *) cond;
    Item *const_item= item_equal->get_const();
    Item_equal_iterator it(*item_equal);
    Item_field *item;
    if (const_item)
    {
      /*
        For each field field1 from item_equal consider the equality
        field1=const_item as a condition allowing an index access of the table
        with field1 by the keys value of field1.
      */
      while ((item= it++))
      {
        add_key_field(key_fields, *and_level, cond_func, item->field,
                      true, &const_item, 1, usable_tables, sargables);
      }
    }
    else
    {
      /*
        Consider all pairs of different fields included into item_equal.
        For each of them (field1, field1) consider the equality
        field1=field2 as a condition allowing an index access of the table
        with field1 by the keys value of field2.
      */
      Item_equal_iterator fi(*item_equal);
      while ((item= fi++))
      {
        Field *field= item->field;
        while ((item= it++))
        {
          if (!field->eq(item->field))
          {
            add_key_field(key_fields, *and_level, cond_func, field,
                          true, (Item **) &item, 1, usable_tables,
                          sargables);
          }
        }
        it.rewind();
      }
    }
    break;
  }
}

/**
  Add all keys with uses 'field' for some keypart.

  If field->and_level != and_level then only mark key_part as const_part.
*/

static uint
max_part_bit(key_part_map bits)
{
  uint32_t found;
  for (found=0; bits & 1 ; found++,bits>>=1) ;
  return found;
}

static void
add_key_part(DYNAMIC_ARRAY *keyuse_array,KEY_FIELD *key_field)
{
  Field *field=key_field->field;
  Table *form= field->table;
  KEYUSE keyuse;

  if (key_field->eq_func && !(key_field->optimize & KEY_OPTIMIZE_EXISTS))
  {
    for (uint32_t key= 0 ; key < form->sizeKeys() ; key++)
    {
      if (!(form->keys_in_use_for_query.test(key)))
	continue;

      uint32_t key_parts= (uint32_t) form->key_info[key].key_parts;
      for (uint32_t part=0 ; part <  key_parts ; part++)
      {
	if (field->eq(form->key_info[key].key_part[part].field))
	{
	  keyuse.table= field->table;
	  keyuse.val =  key_field->val;
	  keyuse.key =  key;
	  keyuse.keypart=part;
	  keyuse.keypart_map= (key_part_map) 1 << part;
	  keyuse.used_tables=key_field->val->used_tables();
	  keyuse.optimize= key_field->optimize & KEY_OPTIMIZE_REF_OR_NULL;
          keyuse.null_rejecting= key_field->null_rejecting;
          keyuse.cond_guard= key_field->cond_guard;
          keyuse.sj_pred_no= key_field->sj_pred_no;
	  insert_dynamic(keyuse_array,(unsigned char*) &keyuse);
	}
      }
    }
  }
}

static int
sort_keyuse(KEYUSE *a,KEYUSE *b)
{
  int res;
  if (a->table->tablenr != b->table->tablenr)
    return (int) (a->table->tablenr - b->table->tablenr);
  if (a->key != b->key)
    return (int) (a->key - b->key);
  if (a->keypart != b->keypart)
    return (int) (a->keypart - b->keypart);
  // Place const values before other ones
  if ((res= test((a->used_tables & ~OUTER_REF_TABLE_BIT)) -
       test((b->used_tables & ~OUTER_REF_TABLE_BIT))))
    return res;
  /* Place rows that are not 'OPTIMIZE_REF_OR_NULL' first */
  return (int) ((a->optimize & KEY_OPTIMIZE_REF_OR_NULL) -
		(b->optimize & KEY_OPTIMIZE_REF_OR_NULL));
}


/*
  Add to KEY_FIELD array all 'ref' access candidates within nested join.

    This function populates KEY_FIELD array with entries generated from the
    ON condition of the given nested join, and does the same for nested joins
    contained within this nested join.

  @param[in]      nested_join_table   Nested join pseudo-table to process
  @param[in,out]  end                 End of the key field array
  @param[in,out]  and_level           And-level
  @param[in,out]  sargables           Array of found sargable candidates


  @note
    We can add accesses to the tables that are direct children of this nested
    join (1), and are not inner tables w.r.t their neighbours (2).

    Example for #1 (outer brackets pair denotes nested join this function is
    invoked for):
    @code
     ... LEFT JOIN (t1 LEFT JOIN (t2 ... ) ) ON cond
    @endcode
    Example for #2:
    @code
     ... LEFT JOIN (t1 LEFT JOIN t2 ) ON cond
    @endcode
    In examples 1-2 for condition cond, we can add 'ref' access candidates to
    t1 only.
    Example #3:
    @code
     ... LEFT JOIN (t1, t2 LEFT JOIN t3 ON inner_cond) ON cond
    @endcode
    Here we can add 'ref' access candidates for t1 and t2, but not for t3.
*/

static void add_key_fields_for_nj(JOIN *join, TableList *nested_join_table,
                                  KEY_FIELD **end, uint32_t *and_level,
                                  SARGABLE_PARAM **sargables)
{
  List_iterator<TableList> li(nested_join_table->nested_join->join_list);
  List_iterator<TableList> li2(nested_join_table->nested_join->join_list);
  bool have_another = false;
  table_map tables= 0;
  TableList *table;
  assert(nested_join_table->nested_join);

  while ((table= li++) || (have_another && (li=li2, have_another=false,
                                            (table= li++))))
  {
    if (table->nested_join)
    {
      if (!table->on_expr)
      {
        /* It's a semi-join nest. Walk into it as if it wasn't a nest */
        have_another= true;
        li2= li;
        li= List_iterator<TableList>(table->nested_join->join_list);
      }
      else
        add_key_fields_for_nj(join, table, end, and_level, sargables);
    }
    else
      if (!table->on_expr)
        tables |= table->table->map;
  }
  if (nested_join_table->on_expr)
    add_key_fields(join, end, and_level, nested_join_table->on_expr, tables,
                   sargables);
}


/**
  Update keyuse array with all possible keys we can use to fetch rows.

  @param       session
  @param[out]  keyuse         Put here ordered array of KEYUSE structures
  @param       join_tab       Array in tablenr_order
  @param       tables         Number of tables in join
  @param       cond           WHERE condition (note that the function analyzes
                              join_tab[i]->on_expr too)
  @param       normal_tables  Tables not inner w.r.t some outer join (ones
                              for which we can make ref access based the WHERE
                              clause)
  @param       select_lex     current SELECT
  @param[out]  sargables      Array of found sargable candidates

   @retval
     0  OK
   @retval
     1  Out of memory.
*/

static bool
update_ref_and_keys(Session *session, DYNAMIC_ARRAY *keyuse,JOIN_TAB *join_tab,
                    uint32_t tables, COND *cond, COND_EQUAL *,
                    table_map normal_tables, Select_Lex *select_lex,
                    SARGABLE_PARAM **sargables)
{
  uint	and_level,i,found_eq_constant;
  KEY_FIELD *key_fields, *end, *field;
  uint32_t sz;
  uint32_t m= cmax(select_lex->max_equal_elems,(uint32_t)1);

  /*
    We use the same piece of memory to store both  KEY_FIELD
    and SARGABLE_PARAM structure.
    KEY_FIELD values are placed at the beginning this memory
    while  SARGABLE_PARAM values are put at the end.
    All predicates that are used to fill arrays of KEY_FIELD
    and SARGABLE_PARAM structures have at most 2 arguments
    except BETWEEN predicates that have 3 arguments and
    IN predicates.
    This any predicate if it's not BETWEEN/IN can be used
    directly to fill at most 2 array elements, either of KEY_FIELD
    or SARGABLE_PARAM type. For a BETWEEN predicate 3 elements
    can be filled as this predicate is considered as
    saragable with respect to each of its argument.
    An IN predicate can require at most 1 element as currently
    it is considered as sargable only for its first argument.
    Multiple equality can add  elements that are filled after
    substitution of field arguments by equal fields. There
    can be not more than select_lex->max_equal_elems such
    substitutions.
  */
  sz= cmax(sizeof(KEY_FIELD),sizeof(SARGABLE_PARAM))*
      (((session->lex->current_select->cond_count+1)*2 +
	session->lex->current_select->between_count)*m+1);
  if (!(key_fields=(KEY_FIELD*)	session->alloc(sz)))
    return true; /* purecov: inspected */
  and_level= 0;
  field= end= key_fields;
  *sargables= (SARGABLE_PARAM *) key_fields +
                (sz - sizeof((*sargables)[0].field))/sizeof(SARGABLE_PARAM);
  /* set a barrier for the array of SARGABLE_PARAM */
  (*sargables)[0].field= 0;

  if (my_init_dynamic_array(keyuse,sizeof(KEYUSE),20,64))
    return true;
  if (cond)
  {
    add_key_fields(join_tab->join, &end, &and_level, cond, normal_tables,
                   sargables);
    for (; field != end ; field++)
    {
      add_key_part(keyuse,field);
      /* Mark that we can optimize LEFT JOIN */
      if (field->val->type() == Item::NULL_ITEM &&
	  !field->field->real_maybe_null())
	field->field->table->reginfo.not_exists_optimize=1;
    }
  }
  for (i=0 ; i < tables ; i++)
  {
    /*
      Block the creation of keys for inner tables of outer joins.
      Here only the outer joins that can not be converted to
      inner joins are left and all nests that can be eliminated
      are flattened.
      In the future when we introduce conditional accesses
      for inner tables in outer joins these keys will be taken
      into account as well.
    */
    if (*join_tab[i].on_expr_ref)
      add_key_fields(join_tab->join, &end, &and_level,
                     *join_tab[i].on_expr_ref,
                     join_tab[i].table->map, sargables);
  }

  /* Process ON conditions for the nested joins */
  {
    List_iterator<TableList> li(*join_tab->join->join_list);
    TableList *table;
    while ((table= li++))
    {
      if (table->nested_join)
        add_key_fields_for_nj(join_tab->join, table, &end, &and_level,
                              sargables);
    }
  }

  /* fill keyuse with found key parts */
  for ( ; field != end ; field++)
    add_key_part(keyuse,field);

  /*
    Sort the array of possible keys and remove the following key parts:
    - ref if there is a keypart which is a ref and a const.
      (e.g. if there is a key(a,b) and the clause is a=3 and b=7 and b=t2.d,
      then we skip the key part corresponding to b=t2.d)
    - keyparts without previous keyparts
      (e.g. if there is a key(a,b,c) but only b < 5 (or a=2 and c < 3) is
      used in the query, we drop the partial key parts from consideration).
  */
  if (keyuse->elements)
  {
    KEYUSE key_end,*prev,*save_pos,*use;

    my_qsort(keyuse->buffer,keyuse->elements,sizeof(KEYUSE),
	  (qsort_cmp) sort_keyuse);

    memset(&key_end, 0, sizeof(key_end)); /* Add for easy testing */
    insert_dynamic(keyuse,(unsigned char*) &key_end);

    use=save_pos=dynamic_element(keyuse,0,KEYUSE*);
    prev= &key_end;
    found_eq_constant=0;
    for (i=0 ; i < keyuse->elements-1 ; i++,use++)
    {
      if (!use->used_tables && use->optimize != KEY_OPTIMIZE_REF_OR_NULL)
	use->table->const_key_parts[use->key]|= use->keypart_map;
      {
	if (use->key == prev->key && use->table == prev->table)
	{
	  if (prev->keypart+1 < use->keypart || ((prev->keypart == use->keypart) && found_eq_constant))
	    continue;				/* remove */
	}
	else if (use->keypart != 0)		// First found must be 0
	  continue;
      }

#ifdef HAVE_purify
      /* Valgrind complains about overlapped memcpy when save_pos==use. */
      if (save_pos != use)
#endif
        *save_pos= *use;
      prev=use;
      found_eq_constant= !use->used_tables;
      /* Save ptr to first use */
      if (!use->table->reginfo.join_tab->keyuse)
	use->table->reginfo.join_tab->keyuse=save_pos;
      use->table->reginfo.join_tab->checked_keys.set(use->key);
      save_pos++;
    }
    i=(uint32_t) (save_pos-(KEYUSE*) keyuse->buffer);
    set_dynamic(keyuse,(unsigned char*) &key_end,i);
    keyuse->elements=i;
  }
  return false;
}

/**
  Update some values in keyuse for faster choose_plan() loop.
*/

static void optimize_keyuse(JOIN *join, DYNAMIC_ARRAY *keyuse_array)
{
  KEYUSE *end,*keyuse= dynamic_element(keyuse_array, 0, KEYUSE*);

  for (end= keyuse+ keyuse_array->elements ; keyuse < end ; keyuse++)
  {
    table_map map;
    /*
      If we find a ref, assume this table matches a proportional
      part of this table.
      For example 100 records matching a table with 5000 records
      gives 5000/100 = 50 records per key
      Constant tables are ignored.
      To avoid bad matches, we don't make ref_table_rows less than 100.
    */
    keyuse->ref_table_rows= ~(ha_rows) 0;	// If no ref
    if (keyuse->used_tables &
	(map= (keyuse->used_tables & ~join->const_table_map &
	       ~OUTER_REF_TABLE_BIT)))
    {
      uint32_t tablenr;
      for (tablenr=0 ; ! (map & 1) ; map>>=1, tablenr++) ;
      if (map == 1)			// Only one table
      {
	Table *tmp_table=join->all_tables[tablenr];
	keyuse->ref_table_rows= cmax(tmp_table->file->stats.records, (ha_rows)100);
      }
    }
    /*
      Outer reference (external field) is constant for single executing
      of subquery
    */
    if (keyuse->used_tables == OUTER_REF_TABLE_BIT)
      keyuse->ref_table_rows= 1;
  }
}


/**
  Discover the indexes that can be used for GROUP BY or DISTINCT queries.

  If the query has a GROUP BY clause, find all indexes that contain all
  GROUP BY fields, and add those indexes to join->const_keys.

  If the query has a DISTINCT clause, find all indexes that contain all
  SELECT fields, and add those indexes to join->const_keys.
  This allows later on such queries to be processed by a
  QUICK_GROUP_MIN_MAX_SELECT.

  @param join
  @param join_tab

  @return
    None
*/

static void
add_group_and_distinct_keys(JOIN *join, JOIN_TAB *join_tab)
{
  List<Item_field> indexed_fields;
  List_iterator<Item_field> indexed_fields_it(indexed_fields);
  order_st      *cur_group;
  Item_field *cur_item;
  key_map possible_keys(0);

  if (join->group_list)
  { /* Collect all query fields referenced in the GROUP clause. */
    for (cur_group= join->group_list; cur_group; cur_group= cur_group->next)
      (*cur_group->item)->walk(&Item::collect_item_field_processor, 0,
                               (unsigned char*) &indexed_fields);
  }
  else if (join->select_distinct)
  { /* Collect all query fields referenced in the SELECT clause. */
    List<Item> &select_items= join->fields_list;
    List_iterator<Item> select_items_it(select_items);
    Item *item;
    while ((item= select_items_it++))
      item->walk(&Item::collect_item_field_processor, 0,
                 (unsigned char*) &indexed_fields);
  }
  else
    return;

  if (indexed_fields.elements == 0)
    return;

  /* Intersect the keys of all group fields. */
  cur_item= indexed_fields_it++;
  possible_keys|= cur_item->field->part_of_key;
  while ((cur_item= indexed_fields_it++))
  {
    possible_keys&= cur_item->field->part_of_key;
  }

  if (possible_keys.any())
    join_tab->const_keys|= possible_keys;
}


/*****************************************************************************
  Go through all combinations of not marked tables and find the one
  which uses least records
*****************************************************************************/

/** Save const tables first as used tables. */

static void
set_position(JOIN *join,uint32_t idx,JOIN_TAB *table,KEYUSE *key)
{
  join->positions[idx].table= table;
  join->positions[idx].key=key;
  join->positions[idx].records_read=1.0;	/* This is a const table */
  join->positions[idx].ref_depend_map= 0;

  /* Move the const table as down as possible in best_ref */
  JOIN_TAB **pos=join->best_ref+idx+1;
  JOIN_TAB *next=join->best_ref[idx];
  for (;next != table ; pos++)
  {
    JOIN_TAB *tmp=pos[0];
    pos[0]=next;
    next=tmp;
  }
  join->best_ref[idx]=table;
}


/*
  Given a semi-join nest, find out which of the IN-equalities are bound

  SYNOPSIS
    get_bound_sj_equalities()
      sj_nest           Semi-join nest
      remaining_tables  Tables that are not yet bound

  DESCRIPTION
    Given a semi-join nest, find out which of the IN-equalities have their
    left part expression bound (i.e. the said expression doesn't refer to
    any of remaining_tables and can be evaluated).

  RETURN
    Bitmap of bound IN-equalities.
*/

uint64_t get_bound_sj_equalities(TableList *sj_nest,
                                  table_map remaining_tables)
{
  List_iterator<Item> li(sj_nest->nested_join->sj_outer_expr_list);
  Item *item;
  uint32_t i= 0;
  uint64_t res= 0;
  while ((item= li++))
  {
    /*
      Q: should this take into account equality propagation and how?
      A: If e->outer_side is an Item_field, walk over the equality
         class and see if there is an element that is bound?
      (this is an optional feature)
    */
    if (!(item->used_tables() & remaining_tables))
    {
      res |= 1UL < i;
    }
  }
  return res;
}


/**
  Find the best access path for an extension of a partial execution
  plan and add this path to the plan.

  The function finds the best access path to table 's' from the passed
  partial plan where an access path is the general term for any means to
  access the data in 's'. An access path may use either an index or a scan,
  whichever is cheaper. The input partial plan is passed via the array
  'join->positions' of length 'idx'. The chosen access method for 's' and its
  cost are stored in 'join->positions[idx]'.

  @param join             pointer to the structure providing all context info
                          for the query
  @param s                the table to be joined by the function
  @param session              thread for the connection that submitted the query
  @param remaining_tables set of tables not included into the partial plan yet
  @param idx              the length of the partial plan
  @param record_count     estimate for the number of records returned by the
                          partial plan
  @param read_time        the cost of the partial plan

  @return
    None
*/

static void
best_access_path(JOIN      *join,
                 JOIN_TAB  *s,
                 Session       *session,
                 table_map remaining_tables,
                 uint32_t      idx,
                 double    record_count,
                 double)
{
  KEYUSE *best_key=         0;
  uint32_t best_max_key_part=   0;
  bool found_constraint= 0;
  double best=              DBL_MAX;
  double best_time=         DBL_MAX;
  double records=           DBL_MAX;
  table_map best_ref_depends_map= 0;
  double tmp;
  ha_rows rec;
  uint32_t best_is_sj_inside_out=    0;

  if (s->keyuse)
  {                                            /* Use key if possible */
    Table *table= s->table;
    KEYUSE *keyuse,*start_key=0;
    double best_records= DBL_MAX;
    uint32_t max_key_part=0;
    uint64_t bound_sj_equalities= 0;
    bool try_sj_inside_out= false;
    /*
      Discover the bound equalites. We need to do this, if
        1. The next table is an SJ-inner table, and
        2. It is the first table from that semijoin, and
        3. We're not within a semi-join range (i.e. all semi-joins either have
           all or none of their tables in join_table_map), except
           s->emb_sj_nest (which we've just entered).
        3. All correlation references from this sj-nest are bound
    */
    if (s->emb_sj_nest &&                                                 // (1)
        s->emb_sj_nest->sj_in_exprs < 64 &&
        ((remaining_tables & s->emb_sj_nest->sj_inner_tables) ==           // (2)
         s->emb_sj_nest->sj_inner_tables) &&                               // (2)
        join->cur_emb_sj_nests == s->emb_sj_nest->sj_inner_tables &&       // (3)
        !(remaining_tables & s->emb_sj_nest->nested_join->sj_corr_tables)) // (4)
    {
      /* This table is an InsideOut scan candidate */
      bound_sj_equalities= get_bound_sj_equalities(s->emb_sj_nest,
                                                   remaining_tables);
      try_sj_inside_out= true;
    }

    /* Test how we can use keys */
    rec= s->records/MATCHING_ROWS_IN_OTHER_TABLE;  // Assumed records/key
    for (keyuse=s->keyuse ; keyuse->table == table ;)
    {
      key_part_map found_part= 0;
      table_map found_ref= 0;
      uint32_t key= keyuse->key;
      KEY *keyinfo= table->key_info+key;
      /* Bitmap of keyparts where the ref access is over 'keypart=const': */
      key_part_map const_part= 0;
      /* The or-null keypart in ref-or-null access: */
      key_part_map ref_or_null_part= 0;

      /* Calculate how many key segments of the current key we can use */
      start_key= keyuse;
      uint64_t handled_sj_equalities=0;
      key_part_map sj_insideout_map= 0;

      do /* For each keypart */
      {
        uint32_t keypart= keyuse->keypart;
        table_map best_part_found_ref= 0;
        double best_prev_record_reads= DBL_MAX;

        do /* For each way to access the keypart */
        {

          /*
            if 1. expression doesn't refer to forward tables
               2. we won't get two ref-or-null's
          */
          if (!(remaining_tables & keyuse->used_tables) &&
              !(ref_or_null_part && (keyuse->optimize &
                                     KEY_OPTIMIZE_REF_OR_NULL)))
          {
            found_part|= keyuse->keypart_map;
            if (!(keyuse->used_tables & ~join->const_table_map))
              const_part|= keyuse->keypart_map;

            double tmp2= prev_record_reads(join, idx, (found_ref |
                                                      keyuse->used_tables));
            if (tmp2 < best_prev_record_reads)
            {
              best_part_found_ref= keyuse->used_tables & ~join->const_table_map;
              best_prev_record_reads= tmp2;
            }
            if (rec > keyuse->ref_table_rows)
              rec= keyuse->ref_table_rows;
	    /*
	      If there is one 'key_column IS NULL' expression, we can
	      use this ref_or_null optimisation of this field
	    */
            if (keyuse->optimize & KEY_OPTIMIZE_REF_OR_NULL)
              ref_or_null_part |= keyuse->keypart_map;
          }

          if (try_sj_inside_out && keyuse->sj_pred_no != UINT_MAX)
          {
            if (!(remaining_tables & keyuse->used_tables))
              bound_sj_equalities |= 1UL << keyuse->sj_pred_no;
            else
            {
              handled_sj_equalities |= 1UL << keyuse->sj_pred_no;
              sj_insideout_map |= ((key_part_map)1) << keyuse->keypart;
            }
          }

          keyuse++;
        } while (keyuse->table == table && keyuse->key == key &&
                 keyuse->keypart == keypart);
	found_ref|= best_part_found_ref;
      } while (keyuse->table == table && keyuse->key == key);

      /*
        Assume that that each key matches a proportional part of table.
      */
      if (!found_part && !handled_sj_equalities)
        continue;                               // Nothing usable found

      if (rec < MATCHING_ROWS_IN_OTHER_TABLE)
        rec= MATCHING_ROWS_IN_OTHER_TABLE;      // Fix for small tables

      bool sj_inside_out_scan= false;
      {
        found_constraint= 1;
        /*
          Check if InsideOut scan is applicable:
          1. All IN-equalities are either "bound" or "handled"
          2. Index keyparts are
             ...
        */
        if (try_sj_inside_out &&
            table->covering_keys.test(key) &&
            (handled_sj_equalities | bound_sj_equalities) ==     // (1)
            PREV_BITS(uint64_t, s->emb_sj_nest->sj_in_exprs)) // (1)
        {
          uint32_t n_fixed_parts= max_part_bit(found_part);
          if (n_fixed_parts != keyinfo->key_parts &&
              (PREV_BITS(uint, n_fixed_parts) | sj_insideout_map) ==
               PREV_BITS(uint, keyinfo->key_parts))
          {
            /*
              Not all parts are fixed. Produce bitmap of remaining bits and
              check if all of them are covered.
            */
            sj_inside_out_scan= true;
            if (!n_fixed_parts)
            {
              /*
                It's a confluent ref scan.

                That is, all found KEYUSE elements refer to IN-equalities,
                and there is really no ref access because there is no
                  t.keypart0 = {bound expression}

                Calculate the cost of complete loose index scan.
              */
              records= (double)s->table->file->stats.records;

              /* The cost is entire index scan cost (divided by 2) */
              best_time= s->table->file->index_only_read_time(key, records);

              /* Now figure how many different keys we will get */
              ulong rpc;
              if ((rpc= keyinfo->rec_per_key[keyinfo->key_parts-1]))
                records= records / rpc;
              start_key= NULL;
            }
          }
        }

        /*
          Check if we found full key
        */
        if (found_part == PREV_BITS(uint,keyinfo->key_parts) &&
            !ref_or_null_part)
        {                                         /* use eq key */
          max_key_part= UINT32_MAX;
          if ((keyinfo->flags & (HA_NOSAME | HA_NULL_PART_KEY)) == HA_NOSAME)
          {
            tmp = prev_record_reads(join, idx, found_ref);
            records=1.0;
          }
          else
          {
            if (!found_ref)
            {                                     /* We found a const key */
              /*
                ReuseRangeEstimateForRef-1:
                We get here if we've found a ref(const) (c_i are constants):
                  "(keypart1=c1) AND ... AND (keypartN=cN)"   [ref_const_cond]

                If range optimizer was able to construct a "range"
                access on this index, then its condition "quick_cond" was
                eqivalent to ref_const_cond (*), and we can re-use E(#rows)
                from the range optimizer.

                Proof of (*): By properties of range and ref optimizers
                quick_cond will be equal or tighther than ref_const_cond.
                ref_const_cond already covers "smallest" possible interval -
                a singlepoint interval over all keyparts. Therefore,
                quick_cond is equivalent to ref_const_cond (if it was an
                empty interval we wouldn't have got here).
              */
              if (table->quick_keys.test(key))
                records= (double) table->quick_rows[key];
              else
              {
                /* quick_range couldn't use key! */
                records= (double) s->records/rec;
              }
            }
            else
            {
              if (!(records=keyinfo->rec_per_key[keyinfo->key_parts-1]))
              {                                   /* Prefer longer keys */
                records=
                  ((double) s->records / (double) rec *
                   (1.0 +
                    ((double) (table->s->max_key_length-keyinfo->key_length) /
                     (double) table->s->max_key_length)));
                if (records < 2.0)
                  records=2.0;               /* Can't be as good as a unique */
              }
              /*
                ReuseRangeEstimateForRef-2:  We get here if we could not reuse
                E(#rows) from range optimizer. Make another try:

                If range optimizer produced E(#rows) for a prefix of the ref
                access we're considering, and that E(#rows) is lower then our
                current estimate, make an adjustment. The criteria of when we
                can make an adjustment is a special case of the criteria used
                in ReuseRangeEstimateForRef-3.
              */
              if (table->quick_keys.test(key) &&
                  const_part & (1 << table->quick_key_parts[key]) &&
                  table->quick_n_ranges[key] == 1 &&
                  records > (double) table->quick_rows[key])
              {
                records= (double) table->quick_rows[key];
              }
            }
            /* Limit the number of matched rows */
            tmp= records;
            set_if_smaller(tmp, (double) session->variables.max_seeks_for_key);
            if (table->covering_keys.test(key))
            {
              /* we can use only index tree */
              tmp= record_count * table->file->index_only_read_time(key, tmp);
            }
            else
              tmp= record_count*cmin(tmp,s->worst_seeks);
          }
        }
        else
        {
          /*
            Use as much key-parts as possible and a uniq key is better
            than a not unique key
            Set tmp to (previous record count) * (records / combination)
          */
          if ((found_part & 1) &&
              (!(table->file->index_flags(key, 0, 0) & HA_ONLY_WHOLE_INDEX) ||
               found_part == PREV_BITS(uint,keyinfo->key_parts)))
          {
            max_key_part= max_part_bit(found_part);
            /*
              ReuseRangeEstimateForRef-3:
              We're now considering a ref[or_null] access via
              (t.keypart1=e1 AND ... AND t.keypartK=eK) [ OR
              (same-as-above but with one cond replaced
               with "t.keypart_i IS NULL")]  (**)

              Try re-using E(#rows) from "range" optimizer:
              We can do so if "range" optimizer used the same intervals as
              in (**). The intervals used by range optimizer may be not
              available at this point (as "range" access might have choosen to
              create quick select over another index), so we can't compare
              them to (**). We'll make indirect judgements instead.
              The sufficient conditions for re-use are:
              (C1) All e_i in (**) are constants, i.e. found_ref==false. (if
                   this is not satisfied we have no way to know which ranges
                   will be actually scanned by 'ref' until we execute the
                   join)
              (C2) max #key parts in 'range' access == K == max_key_part (this
                   is apparently a necessary requirement)

              We also have a property that "range optimizer produces equal or
              tighter set of scan intervals than ref(const) optimizer". Each
              of the intervals in (**) are "tightest possible" intervals when
              one limits itself to using keyparts 1..K (which we do in #2).
              From here it follows that range access used either one, or
              both of the (I1) and (I2) intervals:

               (t.keypart1=c1 AND ... AND t.keypartK=eK)  (I1)
               (same-as-above but with one cond replaced
                with "t.keypart_i IS NULL")               (I2)

              The remaining part is to exclude the situation where range
              optimizer used one interval while we're considering
              ref-or-null and looking for estimate for two intervals. This
              is done by last limitation:

              (C3) "range optimizer used (have ref_or_null?2:1) intervals"
            */
            if (table->quick_keys.test(key) && !found_ref &&          //(C1)
                table->quick_key_parts[key] == max_key_part &&          //(C2)
                table->quick_n_ranges[key] == 1+((ref_or_null_part)?1:0)) //(C3)
            {
              tmp= records= (double) table->quick_rows[key];
            }
            else
            {
              /* Check if we have statistic about the distribution */
              if ((records= keyinfo->rec_per_key[max_key_part-1]))
              {
                /*
                  Fix for the case where the index statistics is too
                  optimistic: If
                  (1) We're considering ref(const) and there is quick select
                      on the same index,
                  (2) and that quick select uses more keyparts (i.e. it will
                      scan equal/smaller interval then this ref(const))
                  (3) and E(#rows) for quick select is higher then our
                      estimate,
                  Then
                    We'll use E(#rows) from quick select.

                  Q: Why do we choose to use 'ref'? Won't quick select be
                  cheaper in some cases ?
                  TODO: figure this out and adjust the plan choice if needed.
                */
                if (!found_ref && table->quick_keys.test(key) &&    // (1)
                    table->quick_key_parts[key] > max_key_part &&     // (2)
                    records < (double)table->quick_rows[key])         // (3)
                  records= (double)table->quick_rows[key];

                tmp= records;
              }
              else
              {
                /*
                  Assume that the first key part matches 1% of the file
                  and that the whole key matches 10 (duplicates) or 1
                  (unique) records.
                  Assume also that more key matches proportionally more
                  records
                  This gives the formula:
                  records = (x * (b-a) + a*c-b)/(c-1)

                  b = records matched by whole key
                  a = records matched by first key part (1% of all records?)
                  c = number of key parts in key
                  x = used key parts (1 <= x <= c)
                */
                double rec_per_key;
                if (!(rec_per_key=(double)
                      keyinfo->rec_per_key[keyinfo->key_parts-1]))
                  rec_per_key=(double) s->records/rec+1;

                if (!s->records)
                  tmp = 0;
                else if (rec_per_key/(double) s->records >= 0.01)
                  tmp = rec_per_key;
                else
                {
                  double a=s->records*0.01;
                  if (keyinfo->key_parts > 1)
                    tmp= (max_key_part * (rec_per_key - a) +
                          a*keyinfo->key_parts - rec_per_key)/
                         (keyinfo->key_parts-1);
                  else
                    tmp= a;
                  set_if_bigger(tmp,1.0);
                }
                records = (uint32_t) tmp;
              }

              if (ref_or_null_part)
              {
                /* We need to do two key searches to find key */
                tmp *= 2.0;
                records *= 2.0;
              }

              /*
                ReuseRangeEstimateForRef-4:  We get here if we could not reuse
                E(#rows) from range optimizer. Make another try:

                If range optimizer produced E(#rows) for a prefix of the ref
                access we're considering, and that E(#rows) is lower then our
                current estimate, make the adjustment.

                The decision whether we can re-use the estimate from the range
                optimizer is the same as in ReuseRangeEstimateForRef-3,
                applied to first table->quick_key_parts[key] key parts.
              */
              if (table->quick_keys.test(key) &&
                  table->quick_key_parts[key] <= max_key_part &&
                  const_part & (1 << table->quick_key_parts[key]) &&
                  table->quick_n_ranges[key] == 1 + ((ref_or_null_part &
                                                     const_part) ? 1 : 0) &&
                  records > (double) table->quick_rows[key])
              {
                tmp= records= (double) table->quick_rows[key];
              }
            }

            /* Limit the number of matched rows */
            set_if_smaller(tmp, (double) session->variables.max_seeks_for_key);
            if (table->covering_keys.test(key))
            {
              /* we can use only index tree */
              tmp= record_count * table->file->index_only_read_time(key, tmp);
            }
            else
              tmp= record_count * cmin(tmp,s->worst_seeks);
          }
          else
            tmp= best_time;                    // Do nothing
        }

        if (sj_inside_out_scan && !start_key)
        {
          tmp= tmp/2;
          if (records)
            records= records/2;
        }

      }
      if (tmp < best_time - records/(double) TIME_FOR_COMPARE)
      {
        best_time= tmp + records/(double) TIME_FOR_COMPARE;
        best= tmp;
        best_records= records;
        best_key= start_key;
        best_max_key_part= max_key_part;
        best_ref_depends_map= found_ref;
        best_is_sj_inside_out= sj_inside_out_scan;
      }
    }
    records= best_records;
  }

  /*
    Don't test table scan if it can't be better.
    Prefer key lookup if we would use the same key for scanning.

    Don't do a table scan on InnoDB tables, if we can read the used
    parts of the row from any of the used index.
    This is because table scans uses index and we would not win
    anything by using a table scan.

    A word for word translation of the below if-statement in sergefp's
    understanding: we check if we should use table scan if:
    (1) The found 'ref' access produces more records than a table scan
        (or index scan, or quick select), or 'ref' is more expensive than
        any of them.
    (2) This doesn't hold: the best way to perform table scan is to to perform
        'range' access using index IDX, and the best way to perform 'ref'
        access is to use the same index IDX, with the same or more key parts.
        (note: it is not clear how this rule is/should be extended to
        index_merge quick selects)
    (3) See above note about InnoDB.
    (4) NOT ("FORCE INDEX(...)" is used for table and there is 'ref' access
             path, but there is no quick select)
        If the condition in the above brackets holds, then the only possible
        "table scan" access method is ALL/index (there is no quick select).
        Since we have a 'ref' access path, and FORCE INDEX instructs us to
        choose it over ALL/index, there is no need to consider a full table
        scan.
  */
  if ((records >= s->found_records || best > s->read_time) &&            // (1)
      !(s->quick && best_key && s->quick->index == best_key->key &&      // (2)
        best_max_key_part >= s->table->quick_key_parts[best_key->key]) &&// (2)
      !((s->table->file->ha_table_flags() & HA_TABLE_SCAN_ON_INDEX) &&   // (3)
        ! s->table->covering_keys.none() && best_key && !s->quick) &&// (3)
      !(s->table->force_index && best_key && !s->quick))                 // (4)
  {                                             // Check full join
    ha_rows rnd_records= s->found_records;
    /*
      If there is a filtering condition on the table (i.e. ref analyzer found
      at least one "table.keyXpartY= exprZ", where exprZ refers only to tables
      preceding this table in the join order we're now considering), then
      assume that 25% of the rows will be filtered out by this condition.

      This heuristic is supposed to force tables used in exprZ to be before
      this table in join order.
    */
    if (found_constraint)
      rnd_records-= rnd_records/4;

    /*
      If applicable, get a more accurate estimate. Don't use the two
      heuristics at once.
    */
    if (s->table->quick_condition_rows != s->found_records)
      rnd_records= s->table->quick_condition_rows;

    /*
      Range optimizer never proposes a RANGE if it isn't better
      than FULL: so if RANGE is present, it's always preferred to FULL.
      Here we estimate its cost.
    */
    if (s->quick)
    {
      /*
        For each record we:
        - read record range through 'quick'
        - skip rows which does not satisfy WHERE constraints
        TODO:
        We take into account possible use of join cache for ALL/index
        access (see first else-branch below), but we don't take it into
        account here for range/index_merge access. Find out why this is so.
      */
      tmp= record_count *
        (s->quick->read_time +
         (s->found_records - rnd_records)/(double) TIME_FOR_COMPARE);
    }
    else
    {
      /* Estimate cost of reading table. */
      tmp= s->table->file->scan_time();
      if (s->table->map & join->outer_join)     // Can't use join cache
      {
        /*
          For each record we have to:
          - read the whole table record
          - skip rows which does not satisfy join condition
        */
        tmp= record_count *
          (tmp +
           (s->records - rnd_records)/(double) TIME_FOR_COMPARE);
      }
      else
      {
        /* We read the table as many times as join buffer becomes full. */
        tmp*= (1.0 + floor((double) cache_record_length(join,idx) *
                           record_count /
                           (double) session->variables.join_buff_size));
        /*
            We don't make full cartesian product between rows in the scanned
           table and existing records because we skip all rows from the
           scanned table, which does not satisfy join condition when
           we read the table (see flush_cached_records for details). Here we
           take into account cost to read and skip these records.
        */
        tmp+= (s->records - rnd_records)/(double) TIME_FOR_COMPARE;
      }
    }

    /*
      We estimate the cost of evaluating WHERE clause for found records
      as record_count * rnd_records / TIME_FOR_COMPARE. This cost plus
      tmp give us total cost of using Table SCAN
    */
    if (best == DBL_MAX ||
        (tmp  + record_count/(double) TIME_FOR_COMPARE*rnd_records <
         best + record_count/(double) TIME_FOR_COMPARE*records))
    {
      /*
        If the table has a range (s->quick is set) make_join_select()
        will ensure that this will be used
      */
      best= tmp;
      records= rows2double(rnd_records);
      best_key= 0;
      /* range/index_merge/ALL/index access method are "independent", so: */
      best_ref_depends_map= 0;
      best_is_sj_inside_out= false;
    }
  }

  /* Update the cost information for the current partial plan */
  join->positions[idx].records_read= records;
  join->positions[idx].read_time=    best;
  join->positions[idx].key=          best_key;
  join->positions[idx].table=        s;
  join->positions[idx].ref_depend_map= best_ref_depends_map;
  join->positions[idx].use_insideout_scan= best_is_sj_inside_out;

  if (!best_key &&
      idx == join->const_tables &&
      s->table == join->sort_by_table &&
      join->unit->select_limit_cnt >= records)
    join->sort_by_table= (Table*) 1;  // Must use temporary table

  return;
}


/**
  Selects and invokes a search strategy for an optimal query plan.

  The function checks user-configurable parameters that control the search
  strategy for an optimal plan, selects the search method and then invokes
  it. Each specific optimization procedure stores the final optimal plan in
  the array 'join->best_positions', and the cost of the plan in
  'join->best_read'.

  @param join         pointer to the structure providing all context info for
                      the query
  @param join_tables  set of the tables in the query

  @todo
    'MAX_TABLES+2' denotes the old implementation of find_best before
    the greedy version. Will be removed when greedy_search is approved.

  @retval
    false       ok
  @retval
    true        Fatal error
*/

static bool
choose_plan(JOIN *join, table_map join_tables)
{
  uint32_t search_depth= join->session->variables.optimizer_search_depth;
  uint32_t prune_level=  join->session->variables.optimizer_prune_level;
  bool straight_join= test(join->select_options & SELECT_STRAIGHT_JOIN);

  join->cur_embedding_map= 0;
  reset_nj_counters(join->join_list);
  /*
    if (SELECT_STRAIGHT_JOIN option is set)
      reorder tables so dependent tables come after tables they depend
      on, otherwise keep tables in the order they were specified in the query
    else
      Apply heuristic: pre-sort all access plans with respect to the number of
      records accessed.
  */
  my_qsort(join->best_ref + join->const_tables,
           join->tables - join->const_tables, sizeof(JOIN_TAB*),
           straight_join ? join_tab_cmp_straight : join_tab_cmp);
  join->cur_emb_sj_nests= 0;
  if (straight_join)
  {
    optimize_straight_join(join, join_tables);
  }
  else
  {
    if (search_depth == MAX_TABLES+2)
    { /*
        TODO: 'MAX_TABLES+2' denotes the old implementation of find_best before
        the greedy version. Will be removed when greedy_search is approved.
      */
      join->best_read= DBL_MAX;
      if (find_best(join, join_tables, join->const_tables, 1.0, 0.0))
        return(true);
    }
    else
    {
      if (search_depth == 0)
        /* Automatically determine a reasonable value for 'search_depth' */
        search_depth= determine_search_depth(join);
      if (greedy_search(join, join_tables, search_depth, prune_level))
        return(true);
    }
  }

  /*
    Store the cost of this query into a user variable
    Don't update last_query_cost for statements that are not "flat joins" :
    i.e. they have subqueries, unions or call stored procedures.
    TODO: calculate a correct cost for a query with subqueries and UNIONs.
  */
  if (join->session->lex->is_single_level_stmt())
    join->session->status_var.last_query_cost= join->best_read;
  return(false);
}


/**
  Compare two JOIN_TAB objects based on the number of accessed records.

  @param ptr1 pointer to first JOIN_TAB object
  @param ptr2 pointer to second JOIN_TAB object

  NOTES
    The order relation implemented by join_tab_cmp() is not transitive,
    i.e. it is possible to choose such a, b and c that (a < b) && (b < c)
    but (c < a). This implies that result of a sort using the relation
    implemented by join_tab_cmp() depends on the order in which
    elements are compared, i.e. the result is implementation-specific.
    Example:
      a: dependent = 0x0 table->map = 0x1 found_records = 3 ptr = 0x907e6b0
      b: dependent = 0x0 table->map = 0x2 found_records = 3 ptr = 0x907e838
      c: dependent = 0x6 table->map = 0x10 found_records = 2 ptr = 0x907ecd0

  @retval
    1  if first is bigger
  @retval
    -1  if second is bigger
  @retval
    0  if equal
*/

int join_tab_cmp(const void* ptr1, const void* ptr2)
{
  JOIN_TAB *jt1= *(JOIN_TAB**) ptr1;
  JOIN_TAB *jt2= *(JOIN_TAB**) ptr2;

  if (jt1->dependent & jt2->table->map)
    return 1;
  if (jt2->dependent & jt1->table->map)
    return -1;
  if (jt1->found_records > jt2->found_records)
    return 1;
  if (jt1->found_records < jt2->found_records)
    return -1;
  return jt1 > jt2 ? 1 : (jt1 < jt2 ? -1 : 0);
}


/**
  Same as join_tab_cmp, but for use with SELECT_STRAIGHT_JOIN.
*/

int join_tab_cmp_straight(const void* ptr1, const void* ptr2)
{
  JOIN_TAB *jt1= *(JOIN_TAB**) ptr1;
  JOIN_TAB *jt2= *(JOIN_TAB**) ptr2;

  if (jt1->dependent & jt2->table->map)
    return 1;
  if (jt2->dependent & jt1->table->map)
    return -1;
  return jt1 > jt2 ? 1 : (jt1 < jt2 ? -1 : 0);
}

/**
  Heuristic procedure to automatically guess a reasonable degree of
  exhaustiveness for the greedy search procedure.

  The procedure estimates the optimization time and selects a search depth
  big enough to result in a near-optimal QEP, that doesn't take too long to
  find. If the number of tables in the query exceeds some constant, then
  search_depth is set to this constant.

  @param join   pointer to the structure providing all context info for
                the query

  @note
    This is an extremely simplistic implementation that serves as a stub for a
    more advanced analysis of the join. Ideally the search depth should be
    determined by learning from previous query optimizations, because it will
    depend on the CPU power (and other factors).

  @todo
    this value should be determined dynamically, based on statistics:
    uint32_t max_tables_for_exhaustive_opt= 7;

  @todo
    this value could be determined by some mapping of the form:
    depth : table_count -> [max_tables_for_exhaustive_opt..MAX_EXHAUSTIVE]

  @return
    A positive integer that specifies the search depth (and thus the
    exhaustiveness) of the depth-first search algorithm used by
    'greedy_search'.
*/

static uint
determine_search_depth(JOIN *join)
{
  uint32_t table_count=  join->tables - join->const_tables;
  uint32_t search_depth;
  /* TODO: this value should be determined dynamically, based on statistics: */
  uint32_t max_tables_for_exhaustive_opt= 7;

  if (table_count <= max_tables_for_exhaustive_opt)
    search_depth= table_count+1; // use exhaustive for small number of tables
  else
    /*
      TODO: this value could be determined by some mapping of the form:
      depth : table_count -> [max_tables_for_exhaustive_opt..MAX_EXHAUSTIVE]
    */
    search_depth= max_tables_for_exhaustive_opt; // use greedy search

  return search_depth;
}


/**
  Select the best ways to access the tables in a query without reordering them.

    Find the best access paths for each query table and compute their costs
    according to their order in the array 'join->best_ref' (thus without
    reordering the join tables). The function calls sequentially
    'best_access_path' for each table in the query to select the best table
    access method. The final optimal plan is stored in the array
    'join->best_positions', and the corresponding cost in 'join->best_read'.

  @param join          pointer to the structure providing all context info for
                       the query
  @param join_tables   set of the tables in the query

  @note
    This function can be applied to:
    - queries with STRAIGHT_JOIN
    - internally to compute the cost of an arbitrary QEP
  @par
    Thus 'optimize_straight_join' can be used at any stage of the query
    optimization process to finalize a QEP as it is.
*/

static void
optimize_straight_join(JOIN *join, table_map join_tables)
{
  JOIN_TAB *s;
  uint32_t idx= join->const_tables;
  double    record_count= 1.0;
  double    read_time=    0.0;

  for (JOIN_TAB **pos= join->best_ref + idx ; (s= *pos) ; pos++)
  {
    /* Find the best access method from 's' to the current partial plan */
    advance_sj_state(join_tables, s);
    best_access_path(join, s, join->session, join_tables, idx,
                     record_count, read_time);
    /* compute the cost of the new plan extended with 's' */
    record_count*= join->positions[idx].records_read;
    read_time+=    join->positions[idx].read_time;
    join_tables&= ~(s->table->map);
    ++idx;
  }

  read_time+= record_count / (double) TIME_FOR_COMPARE;
  if (join->sort_by_table &&
      join->sort_by_table != join->positions[join->const_tables].table->table)
    read_time+= record_count;  // We have to make a temp table
  memcpy(join->best_positions, join->positions, sizeof(POSITION)*idx);
  join->best_read= read_time;
}


/**
  Find a good, possibly optimal, query execution plan (QEP) by a greedy search.

    The search procedure uses a hybrid greedy/exhaustive search with controlled
    exhaustiveness. The search is performed in N = card(remaining_tables)
    steps. Each step evaluates how promising is each of the unoptimized tables,
    selects the most promising table, and extends the current partial QEP with
    that table.  Currenly the most 'promising' table is the one with least
    expensive extension.\

    There are two extreme cases:
    -# When (card(remaining_tables) < search_depth), the estimate finds the
    best complete continuation of the partial QEP. This continuation can be
    used directly as a result of the search.
    -# When (search_depth == 1) the 'best_extension_by_limited_search'
    consideres the extension of the current QEP with each of the remaining
    unoptimized tables.

    All other cases are in-between these two extremes. Thus the parameter
    'search_depth' controlls the exhaustiveness of the search. The higher the
    value, the longer the optimizaton time and possibly the better the
    resulting plan. The lower the value, the fewer alternative plans are
    estimated, but the more likely to get a bad QEP.

    All intermediate and final results of the procedure are stored in 'join':
    - join->positions     : modified for every partial QEP that is explored
    - join->best_positions: modified for the current best complete QEP
    - join->best_read     : modified for the current best complete QEP
    - join->best_ref      : might be partially reordered

    The final optimal plan is stored in 'join->best_positions', and its
    corresponding cost in 'join->best_read'.

  @note
    The following pseudocode describes the algorithm of 'greedy_search':

    @code
    procedure greedy_search
    input: remaining_tables
    output: pplan;
    {
      pplan = <>;
      do {
        (t, a) = best_extension(pplan, remaining_tables);
        pplan = concat(pplan, (t, a));
        remaining_tables = remaining_tables - t;
      } while (remaining_tables != {})
      return pplan;
    }

  @endcode
    where 'best_extension' is a placeholder for a procedure that selects the
    most "promising" of all tables in 'remaining_tables'.
    Currently this estimate is performed by calling
    'best_extension_by_limited_search' to evaluate all extensions of the
    current QEP of size 'search_depth', thus the complexity of 'greedy_search'
    mainly depends on that of 'best_extension_by_limited_search'.

  @par
    If 'best_extension()' == 'best_extension_by_limited_search()', then the
    worst-case complexity of this algorithm is <=
    O(N*N^search_depth/search_depth). When serch_depth >= N, then the
    complexity of greedy_search is O(N!).

  @par
    In the future, 'greedy_search' might be extended to support other
    implementations of 'best_extension', e.g. some simpler quadratic procedure.

  @param join             pointer to the structure providing all context info
                          for the query
  @param remaining_tables set of tables not included into the partial plan yet
  @param search_depth     controlls the exhaustiveness of the search
  @param prune_level      the pruning heuristics that should be applied during
                          search

  @retval
    false       ok
  @retval
    true        Fatal error
*/

static bool
greedy_search(JOIN      *join,
              table_map remaining_tables,
              uint32_t      search_depth,
              uint32_t      prune_level)
{
  double    record_count= 1.0;
  double    read_time=    0.0;
  uint32_t      idx= join->const_tables; // index into 'join->best_ref'
  uint32_t      best_idx;
  uint32_t      size_remain;    // cardinality of remaining_tables
  POSITION  best_pos;
  JOIN_TAB  *best_table; // the next plan node to be added to the curr QEP

  /* number of tables that remain to be optimized */
  size_remain= my_count_bits(remaining_tables);

  do {
    /* Find the extension of the current QEP with the lowest cost */
    join->best_read= DBL_MAX;
    if (best_extension_by_limited_search(join, remaining_tables, idx, record_count,
                                         read_time, search_depth, prune_level))
      return(true);

    if (size_remain <= search_depth)
    {
      /*
        'join->best_positions' contains a complete optimal extension of the
        current partial QEP.
      */
      return(false);
    }

    /* select the first table in the optimal extension as most promising */
    best_pos= join->best_positions[idx];
    best_table= best_pos.table;
    /*
      Each subsequent loop of 'best_extension_by_limited_search' uses
      'join->positions' for cost estimates, therefore we have to update its
      value.
    */
    join->positions[idx]= best_pos;

    /* find the position of 'best_table' in 'join->best_ref' */
    best_idx= idx;
    JOIN_TAB *pos= join->best_ref[best_idx];
    while (pos && best_table != pos)
      pos= join->best_ref[++best_idx];
    assert((pos != NULL)); // should always find 'best_table'
    /* move 'best_table' at the first free position in the array of joins */
    std::swap(join->best_ref[idx], join->best_ref[best_idx]);

    /* compute the cost of the new plan extended with 'best_table' */
    record_count*= join->positions[idx].records_read;
    read_time+=    join->positions[idx].read_time;

    remaining_tables&= ~(best_table->table->map);
    --size_remain;
    ++idx;
  } while (true);
}


/**
  Find a good, possibly optimal, query execution plan (QEP) by a possibly
  exhaustive search.

    The procedure searches for the optimal ordering of the query tables in set
    'remaining_tables' of size N, and the corresponding optimal access paths to
    each table. The choice of a table order and an access path for each table
    constitutes a query execution plan (QEP) that fully specifies how to
    execute the query.

    The maximal size of the found plan is controlled by the parameter
    'search_depth'. When search_depth == N, the resulting plan is complete and
    can be used directly as a QEP. If search_depth < N, the found plan consists
    of only some of the query tables. Such "partial" optimal plans are useful
    only as input to query optimization procedures, and cannot be used directly
    to execute a query.

    The algorithm begins with an empty partial plan stored in 'join->positions'
    and a set of N tables - 'remaining_tables'. Each step of the algorithm
    evaluates the cost of the partial plan extended by all access plans for
    each of the relations in 'remaining_tables', expands the current partial
    plan with the access plan that results in lowest cost of the expanded
    partial plan, and removes the corresponding relation from
    'remaining_tables'. The algorithm continues until it either constructs a
    complete optimal plan, or constructs an optimal plartial plan with size =
    search_depth.

    The final optimal plan is stored in 'join->best_positions'. The
    corresponding cost of the optimal plan is in 'join->best_read'.

  @note
    The procedure uses a recursive depth-first search where the depth of the
    recursion (and thus the exhaustiveness of the search) is controlled by the
    parameter 'search_depth'.

  @note
    The pseudocode below describes the algorithm of
    'best_extension_by_limited_search'. The worst-case complexity of this
    algorithm is O(N*N^search_depth/search_depth). When serch_depth >= N, then
    the complexity of greedy_search is O(N!).

    @code
    procedure best_extension_by_limited_search(
      pplan in,             // in, partial plan of tables-joined-so-far
      pplan_cost,           // in, cost of pplan
      remaining_tables,     // in, set of tables not referenced in pplan
      best_plan_so_far,     // in/out, best plan found so far
      best_plan_so_far_cost,// in/out, cost of best_plan_so_far
      search_depth)         // in, maximum size of the plans being considered
    {
      for each table T from remaining_tables
      {
        // Calculate the cost of using table T as above
        cost = complex-series-of-calculations;

        // Add the cost to the cost so far.
        pplan_cost+= cost;

        if (pplan_cost >= best_plan_so_far_cost)
          // pplan_cost already too great, stop search
          continue;

        pplan= expand pplan by best_access_method;
        remaining_tables= remaining_tables - table T;
        if (remaining_tables is not an empty set
            and
            search_depth > 1)
        {
          best_extension_by_limited_search(pplan, pplan_cost,
                                           remaining_tables,
                                           best_plan_so_far,
                                           best_plan_so_far_cost,
                                           search_depth - 1);
        }
        else
        {
          best_plan_so_far_cost= pplan_cost;
          best_plan_so_far= pplan;
        }
      }
    }
    @endcode

  @note
    When 'best_extension_by_limited_search' is called for the first time,
    'join->best_read' must be set to the largest possible value (e.g. DBL_MAX).
    The actual implementation provides a way to optionally use pruning
    heuristic (controlled by the parameter 'prune_level') to reduce the search
    space by skipping some partial plans.

  @note
    The parameter 'search_depth' provides control over the recursion
    depth, and thus the size of the resulting optimal plan.

  @param join             pointer to the structure providing all context info
                          for the query
  @param remaining_tables set of tables not included into the partial plan yet
  @param idx              length of the partial QEP in 'join->positions';
                          since a depth-first search is used, also corresponds
                          to the current depth of the search tree;
                          also an index in the array 'join->best_ref';
  @param record_count     estimate for the number of records returned by the
                          best partial plan
  @param read_time        the cost of the best partial plan
  @param search_depth     maximum depth of the recursion and thus size of the
                          found optimal plan
                          (0 < search_depth <= join->tables+1).
  @param prune_level      pruning heuristics that should be applied during
                          optimization
                          (values: 0 = EXHAUSTIVE, 1 = PRUNE_BY_TIME_OR_ROWS)

  @retval
    false       ok
  @retval
    true        Fatal error
*/

static bool
best_extension_by_limited_search(JOIN      *join,
                                 table_map remaining_tables,
                                 uint32_t      idx,
                                 double    record_count,
                                 double    read_time,
                                 uint32_t      search_depth,
                                 uint32_t      prune_level)
{
  Session *session= join->session;
  if (session->killed)  // Abort
    return(true);

  /*
     'join' is a partial plan with lower cost than the best plan so far,
     so continue expanding it further with the tables in 'remaining_tables'.
  */
  JOIN_TAB *s;
  double best_record_count= DBL_MAX;
  double best_read_time=    DBL_MAX;

  for (JOIN_TAB **pos= join->best_ref + idx ; (s= *pos) ; pos++)
  {
    table_map real_table_bit= s->table->map;
    if ((remaining_tables & real_table_bit) &&
        !(remaining_tables & s->dependent) &&
        (!idx || !check_interleaving_with_nj(join->positions[idx-1].table, s)))
    {
      double current_record_count, current_read_time;
      advance_sj_state(remaining_tables, s);

      /*
        psergey-insideout-todo:
          when best_access_path() detects it could do an InsideOut scan or
          some other scan, have it return an insideout scan and a flag that
          requests to "fork" this loop iteration. (Q: how does that behave
          when the depth is insufficient??)
      */
      /* Find the best access method from 's' to the current partial plan */
      best_access_path(join, s, session, remaining_tables, idx,
                       record_count, read_time);
      /* Compute the cost of extending the plan with 's' */
      current_record_count= record_count * join->positions[idx].records_read;
      current_read_time=    read_time + join->positions[idx].read_time;

      /* Expand only partial plans with lower cost than the best QEP so far */
      if ((current_read_time +
           current_record_count / (double) TIME_FOR_COMPARE) >= join->best_read)
      {
        restore_prev_nj_state(s);
        restore_prev_sj_state(remaining_tables, s);
        continue;
      }

      /*
        Prune some less promising partial plans. This heuristic may miss
        the optimal QEPs, thus it results in a non-exhaustive search.
      */
      if (prune_level == 1)
      {
        if (best_record_count > current_record_count ||
            best_read_time > current_read_time ||
            (idx == join->const_tables && s->table == join->sort_by_table)) // 's' is the first table in the QEP
        {
          if (best_record_count >= current_record_count &&
              best_read_time >= current_read_time &&
              /* TODO: What is the reasoning behind this condition? */
              (!(s->key_dependent & remaining_tables) ||
               join->positions[idx].records_read < 2.0))
          {
            best_record_count= current_record_count;
            best_read_time=    current_read_time;
          }
        }
        else
        {
          restore_prev_nj_state(s);
          restore_prev_sj_state(remaining_tables, s);
          continue;
        }
      }

      if ( (search_depth > 1) && (remaining_tables & ~real_table_bit) )
      { /* Recursively expand the current partial plan */
        std::swap(join->best_ref[idx], *pos);
        if (best_extension_by_limited_search(join,
                                             remaining_tables & ~real_table_bit,
                                             idx + 1,
                                             current_record_count,
                                             current_read_time,
                                             search_depth - 1,
                                             prune_level))
          return(true);
        std::swap(join->best_ref[idx], *pos);
      }
      else
      { /*
          'join' is either the best partial QEP with 'search_depth' relations,
          or the best complete QEP so far, whichever is smaller.
        */
        current_read_time+= current_record_count / (double) TIME_FOR_COMPARE;
        if (join->sort_by_table &&
            join->sort_by_table !=
            join->positions[join->const_tables].table->table)
          /* We have to make a temp table */
          current_read_time+= current_record_count;
        if ((search_depth == 1) || (current_read_time < join->best_read))
        {
          memcpy(join->best_positions, join->positions,
                 sizeof(POSITION) * (idx + 1));
          join->best_read= current_read_time - 0.001;
        }
      }
      restore_prev_nj_state(s);
      restore_prev_sj_state(remaining_tables, s);
    }
  }
  return(false);
}


/**
  @todo
  - TODO: this function is here only temporarily until 'greedy_search' is
  tested and accepted.

  RETURN VALUES
    false       ok
    true        Fatal error
*/
static bool
find_best(JOIN *join,table_map rest_tables,uint32_t idx,double record_count,
	  double read_time)
{
  Session *session= join->session;
  if (session->killed)
    return(true);
  if (!rest_tables)
  {
    read_time+=record_count/(double) TIME_FOR_COMPARE;
    if (join->sort_by_table &&
	join->sort_by_table !=
	join->positions[join->const_tables].table->table)
      read_time+=record_count;			// We have to make a temp table
    if (read_time < join->best_read)
    {
      memcpy(join->best_positions, join->positions, sizeof(POSITION)*idx);
      join->best_read= read_time - 0.001;
    }
    return(false);
  }
  if (read_time+record_count/(double) TIME_FOR_COMPARE >= join->best_read)
    return(false);					/* Found better before */

  JOIN_TAB *s;
  double best_record_count=DBL_MAX,best_read_time=DBL_MAX;
  for (JOIN_TAB **pos=join->best_ref+idx ; (s=*pos) ; pos++)
  {
    table_map real_table_bit=s->table->map;
    if ((rest_tables & real_table_bit) && !(rest_tables & s->dependent) &&
        (!idx|| !check_interleaving_with_nj(join->positions[idx-1].table, s)))
    {
      double records, best;
      advance_sj_state(rest_tables, s);
      best_access_path(join, s, session, rest_tables, idx, record_count,
                       read_time);
      records= join->positions[idx].records_read;
      best= join->positions[idx].read_time;
      /*
	Go to the next level only if there hasn't been a better key on
	this level! This will cut down the search for a lot simple cases!
      */
      double current_record_count=record_count*records;
      double current_read_time=read_time+best;
      if (best_record_count > current_record_count ||
	  best_read_time > current_read_time ||
	  (idx == join->const_tables && s->table == join->sort_by_table))
      {
	if (best_record_count >= current_record_count &&
	    best_read_time >= current_read_time &&
	    (!(s->key_dependent & rest_tables) || records < 2.0))
	{
	  best_record_count=current_record_count;
	  best_read_time=current_read_time;
	}
        std::swap(join->best_ref[idx], *pos);
	if (find_best(join,rest_tables & ~real_table_bit,idx+1,
                      current_record_count,current_read_time))
          return(true);
        std::swap(join->best_ref[idx], *pos);
      }
      restore_prev_nj_state(s);
      restore_prev_sj_state(rest_tables, s);
      if (join->select_options & SELECT_STRAIGHT_JOIN)
	break;				// Don't test all combinations
    }
  }
  return(false);
}


/**
  Find how much space the prevous read not const tables takes in cache.
*/

static void calc_used_field_length(Session *, JOIN_TAB *join_tab)
{
  uint32_t null_fields,blobs,fields,rec_length;
  Field **f_ptr,*field;

  null_fields= blobs= fields= rec_length=0;
  for (f_ptr=join_tab->table->field ; (field= *f_ptr) ; f_ptr++)
  {
    if (field->isReadSet())
    {
      uint32_t flags=field->flags;
      fields++;
      rec_length+=field->pack_length();
      if (flags & BLOB_FLAG)
	blobs++;
      if (!(flags & NOT_NULL_FLAG))
	null_fields++;
    }
  }
  if (null_fields)
    rec_length+=(join_tab->table->getNullFields() + 7)/8;
  if (join_tab->table->maybe_null)
    rec_length+=sizeof(bool);
  if (blobs)
  {
    uint32_t blob_length=(uint32_t) (join_tab->table->file->stats.mean_rec_length-
			     (join_tab->table->getRecordLength()- rec_length));
    rec_length+=(uint32_t) cmax((uint32_t)4,blob_length);
  }
  join_tab->used_fields= fields;
  join_tab->used_fieldlength= rec_length;
  join_tab->used_blobs= blobs;
}


static uint
cache_record_length(JOIN *join,uint32_t idx)
{
  uint32_t length=0;
  JOIN_TAB **pos,**end;
  Session *session=join->session;

  for (pos=join->best_ref+join->const_tables,end=join->best_ref+idx ;
       pos != end ;
       pos++)
  {
    JOIN_TAB *join_tab= *pos;
    if (!join_tab->used_fieldlength)		/* Not calced yet */
      calc_used_field_length(session, join_tab);
    length+=join_tab->used_fieldlength;
  }
  return length;
}


/*
  Get the number of different row combinations for subset of partial join

  SYNOPSIS
    prev_record_reads()
      join       The join structure
      idx        Number of tables in the partial join order (i.e. the
                 partial join order is in join->positions[0..idx-1])
      found_ref  Bitmap of tables for which we need to find # of distinct
                 row combinations.

  DESCRIPTION
    Given a partial join order (in join->positions[0..idx-1]) and a subset of
    tables within that join order (specified in found_ref), find out how many
    distinct row combinations of subset tables will be in the result of the
    partial join order.

    This is used as follows: Suppose we have a table accessed with a ref-based
    method. The ref access depends on current rows of tables in found_ref.
    We want to count # of different ref accesses. We assume two ref accesses
    will be different if at least one of access parameters is different.
    Example: consider a query

    SELECT * FROM t1, t2, t3 WHERE t1.key=c1 AND t2.key=c2 AND t3.key=t1.field

    and a join order:
      t1,  ref access on t1.key=c1
      t2,  ref access on t2.key=c2
      t3,  ref access on t3.key=t1.field

    For t1: n_ref_scans = 1, n_distinct_ref_scans = 1
    For t2: n_ref_scans = records_read(t1), n_distinct_ref_scans=1
    For t3: n_ref_scans = records_read(t1)*records_read(t2)
            n_distinct_ref_scans = #records_read(t1)

    The reason for having this function (at least the latest version of it)
    is that we need to account for buffering in join execution.

    An edge-case example: if we have a non-first table in join accessed via
    ref(const) or ref(param) where there is a small number of different
    values of param, then the access will likely hit the disk cache and will
    not require any disk seeks.

    The proper solution would be to assume an LRU disk cache of some size,
    calculate probability of cache hits, etc. For now we just count
    identical ref accesses as one.

  RETURN
    Expected number of row combinations
*/

static double
prev_record_reads(JOIN *join, uint32_t idx, table_map found_ref)
{
  double found=1.0;
  POSITION *pos_end= join->positions - 1;
  for (POSITION *pos= join->positions + idx - 1; pos != pos_end; pos--)
  {
    if (pos->table->table->map & found_ref)
    {
      found_ref|= pos->ref_depend_map;
      /*
        For the case of "t1 LEFT JOIN t2 ON ..." where t2 is a const table
        with no matching row we will get position[t2].records_read==0.
        Actually the size of output is one null-complemented row, therefore
        we will use value of 1 whenever we get records_read==0.

        Note
        - the above case can't occur if inner part of outer join has more
          than one table: table with no matches will not be marked as const.

        - Ideally we should add 1 to records_read for every possible null-
          complemented row. We're not doing it because: 1. it will require
          non-trivial code and add overhead. 2. The value of records_read
          is an inprecise estimate and adding 1 (or, in the worst case,
          #max_nested_outer_joins=64-1) will not make it any more precise.
      */
      if (pos->records_read > DBL_EPSILON)
        found*= pos->records_read;
    }
  }
  return found;
}


/**
  Set up join struct according to best position.
*/

static bool
get_best_combination(JOIN *join)
{
  uint32_t i,tablenr;
  table_map used_tables;
  JOIN_TAB *join_tab,*j;
  KEYUSE *keyuse;
  uint32_t table_count;
  Session *session=join->session;

  table_count=join->tables;
  if (!(join->join_tab=join_tab=
	(JOIN_TAB*) session->alloc(sizeof(JOIN_TAB)*table_count)))
    return(true);

  join->full_join=0;

  used_tables= OUTER_REF_TABLE_BIT;		// Outer row is already read
  for (j=join_tab, tablenr=0 ; tablenr < table_count ; tablenr++,j++)
  {
    Table *form;
    *j= *join->best_positions[tablenr].table;
    form=join->table[tablenr]=j->table;
    used_tables|= form->map;
    form->reginfo.join_tab=j;
    if (!*j->on_expr_ref)
      form->reginfo.not_exists_optimize=0;	// Only with LEFT JOIN
    if (j->type == JT_CONST)
      continue;					// Handled in make_join_stat..

    j->ref.key = -1;
    j->ref.key_parts=0;

    if (j->type == JT_SYSTEM)
      continue;
    if (j->keys.none() || !(keyuse= join->best_positions[tablenr].key))
    {
      j->type=JT_ALL;
      if (tablenr != join->const_tables)
	join->full_join=1;
    }
    else if (create_ref_for_key(join, j, keyuse, used_tables))
      return(true);                        // Something went wrong
  }

  for (i=0 ; i < table_count ; i++)
    join->map2table[join->join_tab[i].table->tablenr]=join->join_tab+i;
  update_depend_map(join);
  return(0);
}


static bool create_ref_for_key(JOIN *join, JOIN_TAB *j, KEYUSE *org_keyuse,
			       table_map used_tables)
{
  KEYUSE *keyuse=org_keyuse;
  Session  *session= join->session;
  uint32_t keyparts,length,key;
  Table *table;
  KEY *keyinfo;

  /*  Use best key from find_best */
  table=j->table;
  key=keyuse->key;
  keyinfo=table->key_info+key;

  {
    keyparts=length=0;
    uint32_t found_part_ref_or_null= 0;
    /*
      Calculate length for the used key
      Stop if there is a missing key part or when we find second key_part
      with KEY_OPTIMIZE_REF_OR_NULL
    */
    do
    {
      if (!(~used_tables & keyuse->used_tables))
      {
	if (keyparts == keyuse->keypart &&
	    !(found_part_ref_or_null & keyuse->optimize))
	{
	  keyparts++;
	  length+= keyinfo->key_part[keyuse->keypart].store_length;
	  found_part_ref_or_null|= keyuse->optimize;
	}
      }
      keyuse++;
    } while (keyuse->table == table && keyuse->key == key);
  }

  /* set up fieldref */
  keyinfo=table->key_info+key;
  j->ref.key_parts=keyparts;
  j->ref.key_length=length;
  j->ref.key=(int) key;
  if (!(j->ref.key_buff= (unsigned char*) session->calloc(ALIGN_SIZE(length)*2)) ||
      !(j->ref.key_copy= (store_key**) session->alloc((sizeof(store_key*) *
						   (keyparts+1)))) ||
      !(j->ref.items=    (Item**) session->alloc(sizeof(Item*)*keyparts)) ||
      !(j->ref.cond_guards= (bool**) session->alloc(sizeof(uint*)*keyparts)))
  {
    return(true);
  }
  j->ref.key_buff2=j->ref.key_buff+ALIGN_SIZE(length);
  j->ref.key_err=1;
  j->ref.null_rejecting= 0;
  j->ref.disable_cache= false;
  keyuse=org_keyuse;

  store_key **ref_key= j->ref.key_copy;
  unsigned char *key_buff=j->ref.key_buff, *null_ref_key= 0;
  bool keyuse_uses_no_tables= true;
  {
    uint32_t i;
    for (i=0 ; i < keyparts ; keyuse++,i++)
    {
      while (keyuse->keypart != i ||
	     ((~used_tables) & keyuse->used_tables))
	keyuse++;				/* Skip other parts */

      uint32_t maybe_null= test(keyinfo->key_part[i].null_bit);
      j->ref.items[i]=keyuse->val;		// Save for cond removal
      j->ref.cond_guards[i]= keyuse->cond_guard;
      if (keyuse->null_rejecting)
        j->ref.null_rejecting |= 1 << i;
      keyuse_uses_no_tables= keyuse_uses_no_tables && !keyuse->used_tables;
      if (!keyuse->used_tables &&
	  !(join->select_options & SELECT_DESCRIBE))
      {					// Compare against constant
	store_key_item tmp(session, keyinfo->key_part[i].field,
                           key_buff + maybe_null,
                           maybe_null ?  key_buff : 0,
                           keyinfo->key_part[i].length, keyuse->val);
	if (session->is_fatal_error)
	  return(true);
	tmp.copy();
      }
      else
	*ref_key++= get_store_key(session,
				  keyuse,join->const_table_map,
				  &keyinfo->key_part[i],
				  key_buff, maybe_null);
      /*
	Remember if we are going to use REF_OR_NULL
	But only if field _really_ can be null i.e. we force JT_REF
	instead of JT_REF_OR_NULL in case if field can't be null
      */
      if ((keyuse->optimize & KEY_OPTIMIZE_REF_OR_NULL) && maybe_null)
	null_ref_key= key_buff;
      key_buff+=keyinfo->key_part[i].store_length;
    }
  }
  *ref_key=0;				// end_marker
  if (j->type == JT_CONST)
    j->table->const_table= 1;
  else if (((keyinfo->flags & (HA_NOSAME | HA_NULL_PART_KEY)) != HA_NOSAME) ||
           keyparts != keyinfo->key_parts || null_ref_key)
  {
    /* Must read with repeat */
    j->type= null_ref_key ? JT_REF_OR_NULL : JT_REF;
    j->ref.null_ref_key= null_ref_key;
  }
  else if (keyuse_uses_no_tables)
  {
    /*
      This happen if we are using a constant expression in the ON part
      of an LEFT JOIN.
      SELECT * FROM a LEFT JOIN b ON b.key=30
      Here we should not mark the table as a 'const' as a field may
      have a 'normal' value or a NULL value.
    */
    j->type=JT_CONST;
  }
  else
    j->type=JT_EQ_REF;
  return(0);
}



static store_key *
get_store_key(Session *session, KEYUSE *keyuse, table_map used_tables,
	      KEY_PART_INFO *key_part, unsigned char *key_buff, uint32_t maybe_null)
{
  if (!((~used_tables) & keyuse->used_tables))		// if const item
  {
    return new store_key_const_item(session,
				    key_part->field,
				    key_buff + maybe_null,
				    maybe_null ? key_buff : 0,
				    key_part->length,
				    keyuse->val);
  }
  else if (keyuse->val->type() == Item::FIELD_ITEM ||
           (keyuse->val->type() == Item::REF_ITEM &&
            ((Item_ref*)keyuse->val)->ref_type() == Item_ref::OUTER_REF &&
            (*(Item_ref**)((Item_ref*)keyuse->val)->ref)->ref_type() ==
             Item_ref::DIRECT_REF &&
            keyuse->val->real_item()->type() == Item::FIELD_ITEM))
    return new store_key_field(session,
			       key_part->field,
			       key_buff + maybe_null,
			       maybe_null ? key_buff : 0,
			       key_part->length,
			       ((Item_field*) keyuse->val->real_item())->field,
			       keyuse->val->full_name());
  return new store_key_item(session,
			    key_part->field,
			    key_buff + maybe_null,
			    maybe_null ? key_buff : 0,
			    key_part->length,
			    keyuse->val);
}

/**
  This function is only called for const items on fields which are keys.

  @return
    returns 1 if there was some conversion made when the field was stored.
*/

bool
store_val_in_field(Field *field, Item *item, enum_check_fields check_flag)
{
  bool error;
  Table *table= field->table;
  Session *session= table->in_use;
  ha_rows cuted_fields=session->cuted_fields;

  /*
    we should restore old value of count_cuted_fields because
    store_val_in_field can be called from mysql_insert
    with select_insert, which make count_cuted_fields= 1
   */
  enum_check_fields old_count_cuted_fields= session->count_cuted_fields;
  session->count_cuted_fields= check_flag;
  error= item->save_in_field(field, 1);
  session->count_cuted_fields= old_count_cuted_fields;
  return error || cuted_fields != session->cuted_fields;
}


static bool
make_simple_join(JOIN *join,Table *tmp_table)
{
  Table **tableptr;
  JOIN_TAB *join_tab;

  /*
    Reuse Table * and JOIN_TAB if already allocated by a previous call
    to this function through JOIN::exec (may happen for sub-queries).
  */
  if (!join->table_reexec)
  {
    if (!(join->table_reexec= (Table**) join->session->alloc(sizeof(Table*))))
      return(true);                        /* purecov: inspected */
    if (join->tmp_join)
      join->tmp_join->table_reexec= join->table_reexec;
  }
  if (!join->join_tab_reexec)
  {
    if (!(join->join_tab_reexec=
          (JOIN_TAB*) join->session->alloc(sizeof(JOIN_TAB))))
      return(true);                        /* purecov: inspected */
    if (join->tmp_join)
      join->tmp_join->join_tab_reexec= join->join_tab_reexec;
  }
  tableptr= join->table_reexec;
  join_tab= join->join_tab_reexec;

  join->join_tab=join_tab;
  join->table=tableptr; tableptr[0]=tmp_table;
  join->tables=1;
  join->const_tables=0;
  join->const_table_map=0;
  join->tmp_table_param.field_count= join->tmp_table_param.sum_func_count=
    join->tmp_table_param.func_count=0;
  join->tmp_table_param.copy_field=join->tmp_table_param.copy_field_end=0;
  join->first_record=join->sort_and_group=0;
  join->send_records=(ha_rows) 0;
  join->group=0;
  join->row_limit=join->unit->select_limit_cnt;
  join->do_send_rows = (join->row_limit) ? 1 : 0;

  join_tab->cache.buff=0;			/* No caching */
  join_tab->table=tmp_table;
  join_tab->select=0;
  join_tab->select_cond=0;
  join_tab->quick=0;
  join_tab->type= JT_ALL;			/* Map through all records */
  join_tab->keys.set();                     /* test everything in quick */
  join_tab->info=0;
  join_tab->on_expr_ref=0;
  join_tab->last_inner= 0;
  join_tab->first_unmatched= 0;
  join_tab->ref.key = -1;
  join_tab->not_used_in_distinct=0;
  join_tab->read_first_record= join_init_read_record;
  join_tab->join=join;
  join_tab->ref.key_parts= 0;
  join_tab->flush_weedout_table= join_tab->check_weed_out_table= NULL;
  join_tab->do_firstmatch= NULL;
  memset(&join_tab->read_record, 0, sizeof(join_tab->read_record));
  tmp_table->status=0;
  tmp_table->null_row=0;
  return(false);
}


inline void add_cond_and_fix(Item **e1, Item *e2)
{
  if (*e1)
  {
    Item *res;
    if ((res= new Item_cond_and(*e1, e2)))
    {
      *e1= res;
      res->quick_fix_field();
    }
  }
  else
    *e1= e2;
}


/**
  Add to join_tab->select_cond[i] "table.field IS NOT NULL" conditions
  we've inferred from ref/eq_ref access performed.

    This function is a part of "Early NULL-values filtering for ref access"
    optimization.

    Example of this optimization:
    For query SELECT * FROM t1,t2 WHERE t2.key=t1.field @n
    and plan " any-access(t1), ref(t2.key=t1.field) " @n
    add "t1.field IS NOT NULL" to t1's table condition. @n

    Description of the optimization:

      We look through equalities choosen to perform ref/eq_ref access,
      pick equalities that have form "tbl.part_of_key = othertbl.field"
      (where othertbl is a non-const table and othertbl.field may be NULL)
      and add them to conditions on correspoding tables (othertbl in this
      example).

      Exception from that is the case when referred_tab->join != join.
      I.e. don't add NOT NULL constraints from any embedded subquery.
      Consider this query:
      @code
      SELECT A.f2 FROM t1 LEFT JOIN t2 A ON A.f2 = f1
      WHERE A.f3=(SELECT MIN(f3) FROM  t2 C WHERE A.f4 = C.f4) OR A.f3 IS NULL;
      @endocde
      Here condition A.f3 IS NOT NULL is going to be added to the WHERE
      condition of the embedding query.
      Another example:
      SELECT * FROM t10, t11 WHERE (t10.a < 10 OR t10.a IS NULL)
      AND t11.b <=> t10.b AND (t11.a = (SELECT MAX(a) FROM t12
      WHERE t12.b = t10.a ));
      Here condition t10.a IS NOT NULL is going to be added.
      In both cases addition of NOT NULL condition will erroneously reject
      some rows of the result set.
      referred_tab->join != join constraint would disallow such additions.

      This optimization doesn't affect the choices that ref, range, or join
      optimizer make. This was intentional because this was added after 4.1
      was GA.

    Implementation overview
      1. update_ref_and_keys() accumulates info about null-rejecting
         predicates in in KEY_FIELD::null_rejecting
      1.1 add_key_part saves these to KEYUSE.
      2. create_ref_for_key copies them to TABLE_REF.
      3. add_not_null_conds adds "x IS NOT NULL" to join_tab->select_cond of
         appropiate JOIN_TAB members.
*/

static void add_not_null_conds(JOIN *join)
{
  for (uint32_t i=join->const_tables ; i < join->tables ; i++)
  {
    JOIN_TAB *tab=join->join_tab+i;
    if ((tab->type == JT_REF || tab->type == JT_EQ_REF ||
         tab->type == JT_REF_OR_NULL) &&
        !tab->table->maybe_null)
    {
      for (uint32_t keypart= 0; keypart < tab->ref.key_parts; keypart++)
      {
        if (tab->ref.null_rejecting & (1 << keypart))
        {
          Item *item= tab->ref.items[keypart];
          Item *notnull;
          assert(item->type() == Item::FIELD_ITEM);
          Item_field *not_null_item= (Item_field*)item;
          JOIN_TAB *referred_tab= not_null_item->field->table->reginfo.join_tab;
          /*
            For UPDATE queries such as:
            UPDATE t1 SET t1.f2=(SELECT MAX(t2.f4) FROM t2 WHERE t2.f3=t1.f1);
            not_null_item is the t1.f1, but it's referred_tab is 0.
          */
          if (!referred_tab || referred_tab->join != join)
            continue;
          if (!(notnull= new Item_func_isnotnull(not_null_item)))
            return;
          /*
            We need to do full fix_fields() call here in order to have correct
            notnull->const_item(). This is needed e.g. by test_quick_select
            when it is called from make_join_select after this function is
            called.
          */
          if (notnull->fix_fields(join->session, &notnull))
            return;
          add_cond_and_fix(&referred_tab->select_cond, notnull);
        }
      }
    }
  }
  return;
}

/**
  Build a predicate guarded by match variables for embedding outer joins.
  The function recursively adds guards for predicate cond
  assending from tab to the first inner table  next embedding
  nested outer join and so on until it reaches root_tab
  (root_tab can be 0).

  @param tab       the first inner table for most nested outer join
  @param cond      the predicate to be guarded (must be set)
  @param root_tab  the first inner table to stop

  @return
    -  pointer to the guarded predicate, if success
    -  0, otherwise
*/

static COND*
add_found_match_trig_cond(JOIN_TAB *tab, COND *cond, JOIN_TAB *root_tab)
{
  COND *tmp;
  assert(cond != 0);
  if (tab == root_tab)
    return cond;
  if ((tmp= add_found_match_trig_cond(tab->first_upper, cond, root_tab)))
    tmp= new Item_func_trig_cond(tmp, &tab->found);
  if (tmp)
  {
    tmp->quick_fix_field();
    tmp->update_used_tables();
  }
  return tmp;
}


/**
  Fill in outer join related info for the execution plan structure.

    For each outer join operation left after simplification of the
    original query the function set up the following pointers in the linear
    structure join->join_tab representing the selected execution plan.
    The first inner table t0 for the operation is set to refer to the last
    inner table tk through the field t0->last_inner.
    Any inner table ti for the operation are set to refer to the first
    inner table ti->first_inner.
    The first inner table t0 for the operation is set to refer to the
    first inner table of the embedding outer join operation, if there is any,
    through the field t0->first_upper.
    The on expression for the outer join operation is attached to the
    corresponding first inner table through the field t0->on_expr_ref.
    Here ti are structures of the JOIN_TAB type.

  EXAMPLE. For the query:
  @code
        SELECT * FROM t1
                      LEFT JOIN
                      (t2, t3 LEFT JOIN t4 ON t3.a=t4.a)
                      ON (t1.a=t2.a AND t1.b=t3.b)
          WHERE t1.c > 5,
  @endcode

    given the execution plan with the table order t1,t2,t3,t4
    is selected, the following references will be set;
    t4->last_inner=[t4], t4->first_inner=[t4], t4->first_upper=[t2]
    t2->last_inner=[t4], t2->first_inner=t3->first_inner=[t2],
    on expression (t1.a=t2.a AND t1.b=t3.b) will be attached to
    *t2->on_expr_ref, while t3.a=t4.a will be attached to *t4->on_expr_ref.

  @param join   reference to the info fully describing the query

  @note
    The function assumes that the simplification procedure has been
    already applied to the join query (see simplify_joins).
    This function can be called only after the execution plan
    has been chosen.
*/

static void
make_outerjoin_info(JOIN *join)
{
  for (uint32_t i=join->const_tables ; i < join->tables ; i++)
  {
    JOIN_TAB *tab=join->join_tab+i;
    Table *table=tab->table;
    TableList *tbl= table->pos_in_table_list;
    TableList *embedding= tbl->embedding;

    if (tbl->outer_join)
    {
      /*
        Table tab is the only one inner table for outer join.
        (Like table t4 for the table reference t3 LEFT JOIN t4 ON t3.a=t4.a
        is in the query above.)
      */
      tab->last_inner= tab->first_inner= tab;
      tab->on_expr_ref= &tbl->on_expr;
      tab->cond_equal= tbl->cond_equal;
      if (embedding)
        tab->first_upper= embedding->nested_join->first_nested;
    }
    for ( ; embedding ; embedding= embedding->embedding)
    {
      /* Ignore sj-nests: */
      if (!embedding->on_expr)
        continue;
      nested_join_st *nested_join= embedding->nested_join;
      if (!nested_join->counter_)
      {
        /*
          Table tab is the first inner table for nested_join.
          Save reference to it in the nested join structure.
        */
        nested_join->first_nested= tab;
        tab->on_expr_ref= &embedding->on_expr;
        tab->cond_equal= tbl->cond_equal;
        if (embedding->embedding)
          tab->first_upper= embedding->embedding->nested_join->first_nested;
      }
      if (!tab->first_inner)
        tab->first_inner= nested_join->first_nested;
      if (++nested_join->counter_ < nested_join->join_list.elements)
        break;
      /* Table tab is the last inner table for nested join. */
      nested_join->first_nested->last_inner= tab;
    }
  }
  return;
}


static bool
make_join_select(JOIN *join,SQL_SELECT *select,COND *cond)
{
  Session *session= join->session;
  if (select)
  {
    add_not_null_conds(join);
    table_map used_tables;
    if (cond)                /* Because of QUICK_GROUP_MIN_MAX_SELECT */
    {                        /* there may be a select without a cond. */
      if (join->tables > 1)
        cond->update_used_tables();		// Tablenr may have changed
      if (join->const_tables == join->tables &&
	  session->lex->current_select->master_unit() ==
	  &session->lex->unit)		// not upper level SELECT
        join->const_table_map|=RAND_TABLE_BIT;
      {						// Check const tables
        COND *const_cond=
	  make_cond_for_table(cond,
                              join->const_table_map,
                              (table_map) 0, 1);
        for (JOIN_TAB *tab= join->join_tab+join->const_tables;
             tab < join->join_tab+join->tables ; tab++)
        {
          if (*tab->on_expr_ref)
          {
            JOIN_TAB *cond_tab= tab->first_inner;
            COND *tmp= make_cond_for_table(*tab->on_expr_ref,
                                           join->const_table_map,
                                           (  table_map) 0, 0);
            if (!tmp)
              continue;
            tmp= new Item_func_trig_cond(tmp, &cond_tab->not_null_compl);
            if (!tmp)
              return(1);
            tmp->quick_fix_field();
            cond_tab->select_cond= !cond_tab->select_cond ? tmp :
	                            new Item_cond_and(cond_tab->select_cond,
                                                      tmp);
            if (!cond_tab->select_cond)
	      return(1);
            cond_tab->select_cond->quick_fix_field();
          }
        }
        if (const_cond && !const_cond->val_int())
        {
	  return(1);	 // Impossible const condition
        }
      }
    }
    used_tables=((select->const_tables=join->const_table_map) |
		 OUTER_REF_TABLE_BIT | RAND_TABLE_BIT);
    for (uint32_t i=join->const_tables ; i < join->tables ; i++)
    {
      JOIN_TAB *tab=join->join_tab+i;
      /*
        first_inner is the X in queries like:
        SELECT * FROM t1 LEFT OUTER JOIN (t2 JOIN t3) ON X
      */
      JOIN_TAB *first_inner_tab= tab->first_inner;
      table_map current_map= tab->table->map;
      bool use_quick_range=0;
      COND *tmp;

      /*
	Following force including random expression in last table condition.
	It solve problem with select like SELECT * FROM t1 WHERE rand() > 0.5
      */
      if (i == join->tables-1)
	current_map|= OUTER_REF_TABLE_BIT | RAND_TABLE_BIT;
      used_tables|=current_map;

      if (tab->type == JT_REF && tab->quick &&
	  (uint32_t) tab->ref.key == tab->quick->index &&
	  tab->ref.key_length < tab->quick->max_used_key_length)
      {
	/* Range uses longer key;  Use this instead of ref on key */
	tab->type=JT_ALL;
	use_quick_range=1;
	tab->use_quick=1;
        tab->ref.key= -1;
	tab->ref.key_parts=0;		// Don't use ref key.
	join->best_positions[i].records_read= rows2double(tab->quick->records);
        /*
          We will use join cache here : prevent sorting of the first
          table only and sort at the end.
        */
        if (i != join->const_tables && join->tables > join->const_tables + 1)
          join->full_join= 1;
      }

      tmp= NULL;
      if (cond)
        tmp= make_cond_for_table(cond,used_tables,current_map, 0);
      if (cond && !tmp && tab->quick)
      {						// Outer join
        if (tab->type != JT_ALL)
        {
          /*
            Don't use the quick method
            We come here in the case where we have 'key=constant' and
            the test is removed by make_cond_for_table()
          */
          delete tab->quick;
          tab->quick= 0;
        }
        else
        {
          /*
            Hack to handle the case where we only refer to a table
            in the ON part of an OUTER JOIN. In this case we want the code
            below to check if we should use 'quick' instead.
          */
          tmp= new Item_int((int64_t) 1,1);	// Always true
        }

      }
      if (tmp || !cond || tab->type == JT_REF || tab->type == JT_REF_OR_NULL ||
          tab->type == JT_EQ_REF)
      {
	SQL_SELECT *sel= tab->select= ((SQL_SELECT*)
                                       session->memdup((unsigned char*) select,
                                                   sizeof(*select)));
	if (!sel)
	  return(1);			// End of memory
        /*
          If tab is an inner table of an outer join operation,
          add a match guard to the pushed down predicate.
          The guard will turn the predicate on only after
          the first match for outer tables is encountered.
	*/
        if (cond && tmp)
        {
          /*
            Because of QUICK_GROUP_MIN_MAX_SELECT there may be a select without
            a cond, so neutralize the hack above.
          */
          if (!(tmp= add_found_match_trig_cond(first_inner_tab, tmp, 0)))
            return(1);
          tab->select_cond=sel->cond=tmp;
          /* Push condition to storage engine if this is enabled
             and the condition is not guarded */
          tab->table->file->pushed_cond= NULL;
	  if (session->variables.engine_condition_pushdown)
          {
            COND *push_cond=
              make_cond_for_table(tmp, current_map, current_map, 0);
            if (push_cond)
            {
              /* Push condition to handler */
              if (!tab->table->file->cond_push(push_cond))
                tab->table->file->pushed_cond= push_cond;
            }
          }
        }
        else
          tab->select_cond= sel->cond= NULL;

	sel->head=tab->table;
	if (tab->quick)
	{
	  /* Use quick key read if it's a constant and it's not used
	     with key reading */
	  if (tab->needed_reg.none() && tab->type != JT_EQ_REF
	      && (tab->type != JT_REF || (uint32_t) tab->ref.key == tab->quick->index))
	  {
	    sel->quick=tab->quick;		// Use value from get_quick_...
	    sel->quick_keys.reset();
	    sel->needed_reg.reset();
	  }
	  else
	  {
	    delete tab->quick;
	  }
	  tab->quick=0;
	}
	uint32_t ref_key=(uint32_t) sel->head->reginfo.join_tab->ref.key+1;
	if (i == join->const_tables && ref_key)
	{
	  if (tab->const_keys.any() &&
              tab->table->reginfo.impossible_range)
	    return(1);
	}
	else if (tab->type == JT_ALL && ! use_quick_range)
	{
	  if (tab->const_keys.any() &&
	      tab->table->reginfo.impossible_range)
	    return(1);				// Impossible range
	  /*
	    We plan to scan all rows.
	    Check again if we should use an index.
	    We could have used an column from a previous table in
	    the index if we are using limit and this is the first table
	  */

	  if ((cond && (!((tab->keys & tab->const_keys) == tab->keys) && i > 0)) ||
	      (!tab->const_keys.none() && (i == join->const_tables) && (join->unit->select_limit_cnt < join->best_positions[i].records_read) && ((join->select_options & OPTION_FOUND_ROWS) == false)))
	  {
	    /* Join with outer join condition */
	    COND *orig_cond=sel->cond;
	    sel->cond= and_conds(sel->cond, *tab->on_expr_ref);

	    /*
              We can't call sel->cond->fix_fields,
              as it will break tab->on_expr if it's AND condition
              (fix_fields currently removes extra AND/OR levels).
              Yet attributes of the just built condition are not needed.
              Thus we call sel->cond->quick_fix_field for safety.
	    */
	    if (sel->cond && !sel->cond->fixed)
	      sel->cond->quick_fix_field();

	    if (sel->test_quick_select(session, tab->keys,
				       used_tables & ~ current_map,
				       (join->select_options &
					OPTION_FOUND_ROWS ?
					HA_POS_ERROR :
					join->unit->select_limit_cnt), 0,
                                        false) < 0)
            {
	      /*
		Before reporting "Impossible WHERE" for the whole query
		we have to check isn't it only "impossible ON" instead
	      */
              sel->cond=orig_cond;
              if (!*tab->on_expr_ref ||
                  sel->test_quick_select(session, tab->keys,
                                         used_tables & ~ current_map,
                                         (join->select_options &
                                          OPTION_FOUND_ROWS ?
                                          HA_POS_ERROR :
                                          join->unit->select_limit_cnt),0,
                                          false) < 0)
		return(1);			// Impossible WHERE
            }
            else
	      sel->cond=orig_cond;

	    /* Fix for EXPLAIN */
	    if (sel->quick)
	      join->best_positions[i].records_read= (double)sel->quick->records;
	  }
	  else
	  {
	    sel->needed_reg=tab->needed_reg;
	    sel->quick_keys.reset();
	  }
          if (!((tab->checked_keys & sel->quick_keys) == sel->quick_keys) ||
              !((tab->checked_keys & sel->needed_reg) == sel->needed_reg))
	  {
	    tab->keys= sel->quick_keys;
            tab->keys|= sel->needed_reg;
	    tab->use_quick= (!sel->needed_reg.none() &&
			     (select->quick_keys.none() ||
			      (select->quick &&
			       (select->quick->records >= 100L)))) ?
	      2 : 1;
	    sel->read_tables= used_tables & ~current_map;
	  }
	  if (i != join->const_tables && tab->use_quick != 2)
	  {					/* Read with cache */
	    if (cond &&
                (tmp=make_cond_for_table(cond,
					 join->const_table_map |
					 current_map,
					 current_map, 0)))
	    {
	      tab->cache.select=(SQL_SELECT*)
		session->memdup((unsigned char*) sel, sizeof(SQL_SELECT));
	      tab->cache.select->cond=tmp;
	      tab->cache.select->read_tables=join->const_table_map;
	    }
	  }
	}
      }

      /*
        Push down conditions from all on expressions.
        Each of these conditions are guarded by a variable
        that turns if off just before null complemented row for
        outer joins is formed. Thus, the condition from an
        'on expression' are guaranteed not to be checked for
        the null complemented row.
      */

      /* First push down constant conditions from on expressions */
      for (JOIN_TAB *join_tab= join->join_tab+join->const_tables;
           join_tab < join->join_tab+join->tables ; join_tab++)
      {
        if (*join_tab->on_expr_ref)
        {
          JOIN_TAB *cond_tab= join_tab->first_inner;
          tmp= make_cond_for_table(*join_tab->on_expr_ref,
                                   join->const_table_map,
                                   (table_map) 0, 0);
          if (!tmp)
            continue;
          tmp= new Item_func_trig_cond(tmp, &cond_tab->not_null_compl);
          if (!tmp)
            return(1);
          tmp->quick_fix_field();
          cond_tab->select_cond= !cond_tab->select_cond ? tmp :
	                            new Item_cond_and(cond_tab->select_cond,tmp);
          if (!cond_tab->select_cond)
	    return(1);
          cond_tab->select_cond->quick_fix_field();
        }
      }

      /* Push down non-constant conditions from on expressions */
      JOIN_TAB *last_tab= tab;
      while (first_inner_tab && first_inner_tab->last_inner == last_tab)
      {
        /*
          Table tab is the last inner table of an outer join.
          An on expression is always attached to it.
	*/
        COND *on_expr= *first_inner_tab->on_expr_ref;

        table_map used_tables2= (join->const_table_map |
                                 OUTER_REF_TABLE_BIT | RAND_TABLE_BIT);
	for (tab= join->join_tab+join->const_tables; tab <= last_tab ; tab++)
        {
          current_map= tab->table->map;
          used_tables2|= current_map;
          COND *tmp_cond= make_cond_for_table(on_expr, used_tables2,
                                              current_map, 0);
          if (tmp_cond)
          {
            JOIN_TAB *cond_tab= tab < first_inner_tab ? first_inner_tab : tab;
            /*
              First add the guards for match variables of
              all embedding outer join operations.
	    */
            if (!(tmp_cond= add_found_match_trig_cond(cond_tab->first_inner,
                                                     tmp_cond,
                                                     first_inner_tab)))
              return(1);
            /*
              Now add the guard turning the predicate off for
              the null complemented row.
	    */
            tmp_cond= new Item_func_trig_cond(tmp_cond,
                                              &first_inner_tab->
                                              not_null_compl);
            if (tmp_cond)
              tmp_cond->quick_fix_field();
	    /* Add the predicate to other pushed down predicates */
            cond_tab->select_cond= !cond_tab->select_cond ? tmp_cond :
	                          new Item_cond_and(cond_tab->select_cond,
                                                    tmp_cond);
            if (!cond_tab->select_cond)
	      return(1);
            cond_tab->select_cond->quick_fix_field();
          }
        }
        first_inner_tab= first_inner_tab->first_upper;
      }
    }
  }
  return(0);
}


/*
  Check if given expression uses only table fields covered by the given index

  SYNOPSIS
    uses_index_fields_only()
      item           Expression to check
      tbl            The table having the index
      keyno          The index number
      other_tbls_ok  true <=> Fields of other non-const tables are allowed

  DESCRIPTION
    Check if given expression only uses fields covered by index #keyno in the
    table tbl. The expression can use any fields in any other tables.

    The expression is guaranteed not to be AND or OR - those constructs are
    handled outside of this function.

  RETURN
    true   Yes
    false  No
*/

bool uses_index_fields_only(Item *item, Table *tbl, uint32_t keyno,
                            bool other_tbls_ok)
{
  if (item->const_item())
    return true;

  /*
    Don't push down the triggered conditions. Nested outer joins execution
    code may need to evaluate a condition several times (both triggered and
    untriggered), and there is no way to put thi
    TODO: Consider cloning the triggered condition and using the copies for:
      1. push the first copy down, to have most restrictive index condition
         possible
      2. Put the second copy into tab->select_cond.
  */
  if (item->type() == Item::FUNC_ITEM &&
      ((Item_func*)item)->functype() == Item_func::TRIG_COND_FUNC)
    return false;

  if (!(item->used_tables() & tbl->map))
    return other_tbls_ok;

  Item::Type item_type= item->type();
  switch (item_type) {
  case Item::FUNC_ITEM:
    {
      /* This is a function, apply condition recursively to arguments */
      Item_func *item_func= (Item_func*)item;
      Item **child;
      Item **item_end= (item_func->arguments()) + item_func->argument_count();
      for (child= item_func->arguments(); child != item_end; child++)
      {
        if (!uses_index_fields_only(*child, tbl, keyno, other_tbls_ok))
          return false;
      }
      return true;
    }
  case Item::COND_ITEM:
    {
      /* This is a function, apply condition recursively to arguments */
      List_iterator<Item> li(*((Item_cond*)item)->argument_list());
      Item *list_item;
      while ((list_item=li++))
      {
        if (!uses_index_fields_only(item, tbl, keyno, other_tbls_ok))
          return false;
      }
      return true;
    }
  case Item::FIELD_ITEM:
    {
      Item_field *item_field= (Item_field*)item;
      if (item_field->field->table != tbl)
        return true;
      return item_field->field->part_of_key.test(keyno);
    }
  case Item::REF_ITEM:
    return uses_index_fields_only(item->real_item(), tbl, keyno,
                                  other_tbls_ok);
  default:
    return false; /* Play it safe, don't push unknown non-const items */
  }
}


#define ICP_COND_USES_INDEX_ONLY 10

/*
  Get a part of the condition that can be checked using only index fields

  SYNOPSIS
    make_cond_for_index()
      cond           The source condition
      table          The table that is partially available
      keyno          The index in the above table. Only fields covered by the index
                     are available
      other_tbls_ok  true <=> Fields of other non-const tables are allowed

  DESCRIPTION
    Get a part of the condition that can be checked when for the given table
    we have values only of fields covered by some index. The condition may
    refer to other tables, it is assumed that we have values of all of their
    fields.

    Example:
      make_cond_for_index(
         "cond(t1.field) AND cond(t2.key1) AND cond(t2.non_key) AND cond(t2.key2)",
          t2, keyno(t2.key1))
      will return
        "cond(t1.field) AND cond(t2.key2)"

  RETURN
    Index condition, or NULL if no condition could be inferred.
*/

Item *make_cond_for_index(Item *cond, Table *table, uint32_t keyno,
                          bool other_tbls_ok)
{
  if (!cond)
    return NULL;
  if (cond->type() == Item::COND_ITEM)
  {
    uint32_t n_marked= 0;
    if (((Item_cond*) cond)->functype() == Item_func::COND_AND_FUNC)
    {
      Item_cond_and *new_cond=new Item_cond_and;
      if (!new_cond)
	return (COND*) 0;
      List_iterator<Item> li(*((Item_cond*) cond)->argument_list());
      Item *item;
      while ((item=li++))
      {
	Item *fix= make_cond_for_index(item, table, keyno, other_tbls_ok);
	if (fix)
	  new_cond->argument_list()->push_back(fix);
        n_marked += test(item->marker == ICP_COND_USES_INDEX_ONLY);
      }
      if (n_marked ==((Item_cond*)cond)->argument_list()->elements)
        cond->marker= ICP_COND_USES_INDEX_ONLY;
      switch (new_cond->argument_list()->elements) {
      case 0:
	return (COND*) 0;
      case 1:
	return new_cond->argument_list()->head();
      default:
	new_cond->quick_fix_field();
	return new_cond;
      }
    }
    else /* It's OR */
    {
      Item_cond_or *new_cond=new Item_cond_or;
      if (!new_cond)
	return (COND*) 0;
      List_iterator<Item> li(*((Item_cond*) cond)->argument_list());
      Item *item;
      while ((item=li++))
      {
	Item *fix= make_cond_for_index(item, table, keyno, other_tbls_ok);
	if (!fix)
	  return (COND*) 0;
	new_cond->argument_list()->push_back(fix);
        n_marked += test(item->marker == ICP_COND_USES_INDEX_ONLY);
      }
      if (n_marked ==((Item_cond*)cond)->argument_list()->elements)
        cond->marker= ICP_COND_USES_INDEX_ONLY;
      new_cond->quick_fix_field();
      new_cond->top_level_item();
      return new_cond;
    }
  }

  if (!uses_index_fields_only(cond, table, keyno, other_tbls_ok))
    return (COND*) 0;
  cond->marker= ICP_COND_USES_INDEX_ONLY;
  return cond;
}


Item *make_cond_remainder(Item *cond, bool exclude_index)
{
  if (exclude_index && cond->marker == ICP_COND_USES_INDEX_ONLY)
    return 0; /* Already checked */

  if (cond->type() == Item::COND_ITEM)
  {
    table_map tbl_map= 0;
    if (((Item_cond*) cond)->functype() == Item_func::COND_AND_FUNC)
    {
      /* Create new top level AND item */
      Item_cond_and *new_cond=new Item_cond_and;
      if (!new_cond)
	return (COND*) 0;
      List_iterator<Item> li(*((Item_cond*) cond)->argument_list());
      Item *item;
      while ((item=li++))
      {
	Item *fix= make_cond_remainder(item, exclude_index);
	if (fix)
        {
	  new_cond->argument_list()->push_back(fix);
          tbl_map |= fix->used_tables();
        }
      }
      switch (new_cond->argument_list()->elements) {
      case 0:
	return (COND*) 0;
      case 1:
	return new_cond->argument_list()->head();
      default:
	new_cond->quick_fix_field();
        ((Item_cond*)new_cond)->used_tables_cache= tbl_map;
	return new_cond;
      }
    }
    else /* It's OR */
    {
      Item_cond_or *new_cond=new Item_cond_or;
      if (!new_cond)
	return (COND*) 0;
      List_iterator<Item> li(*((Item_cond*) cond)->argument_list());
      Item *item;
      while ((item=li++))
      {
	Item *fix= make_cond_remainder(item, false);
	if (!fix)
	  return (COND*) 0;
	new_cond->argument_list()->push_back(fix);
        tbl_map |= fix->used_tables();
      }
      new_cond->quick_fix_field();
      ((Item_cond*)new_cond)->used_tables_cache= tbl_map;
      new_cond->top_level_item();
      return new_cond;
    }
  }
  return cond;
}


/*
  Try to extract and push the index condition

  SYNOPSIS
    push_index_cond()
      tab            A join tab that has tab->table->file and its condition
                     in tab->select_cond
      keyno          Index for which extract and push the condition
      other_tbls_ok  true <=> Fields of other non-const tables are allowed

  DESCRIPTION
    Try to extract and push the index condition down to table handler
*/

static void push_index_cond(JOIN_TAB *tab, uint32_t keyno, bool other_tbls_ok)
{
  Item *idx_cond;
  if (tab->table->file->index_flags(keyno, 0, 1) & HA_DO_INDEX_COND_PUSHDOWN &&
      tab->join->session->variables.engine_condition_pushdown)
  {
    idx_cond= make_cond_for_index(tab->select_cond, tab->table, keyno,
                                  other_tbls_ok);

    if (idx_cond)
    {
      tab->pre_idx_push_select_cond= tab->select_cond;
      Item *idx_remainder_cond=
        tab->table->file->idx_cond_push(keyno, idx_cond);

      /*
        Disable eq_ref's "lookup cache" if we've pushed down an index
        condition.
        TODO: This check happens to work on current ICP implementations, but
        there may exist a compliant implementation that will not work
        correctly with it. Sort this out when we stabilize the condition
        pushdown APIs.
      */
      if (idx_remainder_cond != idx_cond)
        tab->ref.disable_cache= true;

      Item *row_cond= make_cond_remainder(tab->select_cond, true);

      if (row_cond)
      {
        if (!idx_remainder_cond)
          tab->select_cond= row_cond;
        else
        {
          tab->select_cond= new Item_cond_and(row_cond, idx_remainder_cond);
	  tab->select_cond->quick_fix_field();
          ((Item_cond_and*)tab->select_cond)->used_tables_cache=
            row_cond->used_tables() | idx_remainder_cond->used_tables();
        }
      }
      else
        tab->select_cond= idx_remainder_cond;
      if (tab->select)
      {
        tab->select->cond= tab->select_cond;
      }
    }
  }
  return;
}



    /*
      Determine if the set is already ordered for order_st BY, so it can
      disable join cache because it will change the ordering of the results.
      Code handles sort table that is at any location (not only first after
      the const tables) despite the fact that it's currently prohibited.
      We must disable join cache if the first non-const table alone is
      ordered. If there is a temp table the ordering is done as a last
      operation and doesn't prevent join cache usage.
    */
uint32_t make_join_orderinfo(JOIN *join)
{
  uint32_t i;
  if (join->need_tmp)
    return join->tables;

  for (i=join->const_tables ; i < join->tables ; i++)
  {
    JOIN_TAB *tab=join->join_tab+i;
    Table *table=tab->table;
    if ((table == join->sort_by_table &&
         (!join->order || join->skip_sort_order)) ||
        (join->sort_by_table == (Table *) 1 && i != join->const_tables))
    {
      break;
    }
  }
  return i;
}


/*
  Plan refinement stage: do various set ups for the executioner

  SYNOPSIS
    make_join_readinfo()
      join           Join being processed
      options        Join's options (checking for SELECT_DESCRIBE,
                     SELECT_NO_JOIN_CACHE)
      no_jbuf_after  Don't use join buffering after table with this number.

  DESCRIPTION
    Plan refinement stage: do various set ups for the executioner
      - set up use of join buffering
      - push index conditions
      - increment counters
      - etc

  RETURN
    false - OK
    true  - Out of memory
*/

static bool
make_join_readinfo(JOIN *join, uint64_t options, uint32_t no_jbuf_after)
{
  uint32_t i;
  bool statistics= test(!(join->select_options & SELECT_DESCRIBE));
  bool sorted= 1;

  for (i=join->const_tables ; i < join->tables ; i++)
  {
    JOIN_TAB *tab=join->join_tab+i;
    Table *table=tab->table;
    bool using_join_cache;
    tab->read_record.table= table;
    tab->read_record.file=table->file;
    tab->next_select=sub_select;		/* normal select */
    /*
      TODO: don't always instruct first table's ref/range access method to
      produce sorted output.
    */
    tab->sorted= sorted;
    sorted= 0;                                  // only first must be sorted
    if (tab->insideout_match_tab)
    {
      if (!(tab->insideout_buf= (unsigned char*)join->session->alloc(tab->table->key_info
                                                         [tab->index].
                                                         key_length)))
        return true;
    }
    switch (tab->type) {
    case JT_SYSTEM:				// Only happens with left join
      table->status=STATUS_NO_RECORD;
      tab->read_first_record= join_read_system;
      tab->read_record.read_record= join_no_more_records;
      break;
    case JT_CONST:				// Only happens with left join
      table->status=STATUS_NO_RECORD;
      tab->read_first_record= join_read_const;
      tab->read_record.read_record= join_no_more_records;
      if (table->covering_keys.test(tab->ref.key) &&
          !table->no_keyread)
      {
        table->key_read=1;
        table->file->extra(HA_EXTRA_KEYREAD);
      }
      break;
    case JT_EQ_REF:
      table->status=STATUS_NO_RECORD;
      if (tab->select)
      {
	delete tab->select->quick;
	tab->select->quick=0;
      }
      delete tab->quick;
      tab->quick=0;
      tab->read_first_record= join_read_key;
      tab->read_record.read_record= join_no_more_records;
      if (table->covering_keys.test(tab->ref.key) &&
	  !table->no_keyread)
      {
	table->key_read=1;
	table->file->extra(HA_EXTRA_KEYREAD);
      }
      else
        push_index_cond(tab, tab->ref.key, true);
      break;
    case JT_REF_OR_NULL:
    case JT_REF:
      table->status=STATUS_NO_RECORD;
      if (tab->select)
      {
	delete tab->select->quick;
	tab->select->quick=0;
      }
      delete tab->quick;
      tab->quick=0;
      if (table->covering_keys.test(tab->ref.key) &&
	  !table->no_keyread)
      {
	table->key_read=1;
	table->file->extra(HA_EXTRA_KEYREAD);
      }
      else
        push_index_cond(tab, tab->ref.key, true);
      if (tab->type == JT_REF)
      {
	tab->read_first_record= join_read_always_key;
	tab->read_record.read_record= tab->insideout_match_tab?
           join_read_next_same_diff : join_read_next_same;
      }
      else
      {
	tab->read_first_record= join_read_always_key_or_null;
	tab->read_record.read_record= join_read_next_same_or_null;
      }
      break;
    case JT_ALL:
      /*
	If previous table use cache
        If the incoming data set is already sorted don't use cache.
      */
      table->status=STATUS_NO_RECORD;
      using_join_cache= false;
      if (i != join->const_tables && !(options & SELECT_NO_JOIN_CACHE) &&
          tab->use_quick != 2 && !tab->first_inner && i <= no_jbuf_after &&
          !tab->insideout_match_tab)
      {
	if ((options & SELECT_DESCRIBE) ||
	    !join_init_cache(join->session,join->join_tab+join->const_tables,
			     i-join->const_tables))
	{
          using_join_cache= true;
	  tab[-1].next_select=sub_select_cache; /* Patch previous */
	}
      }
      /* These init changes read_record */
      if (tab->use_quick == 2)
      {
	join->session->server_status|=SERVER_QUERY_NO_GOOD_INDEX_USED;
	tab->read_first_record= join_init_quick_read_record;
	if (statistics)
	  status_var_increment(join->session->status_var.select_range_check_count);
      }
      else
      {
	tab->read_first_record= join_init_read_record;
	if (i == join->const_tables)
	{
	  if (tab->select && tab->select->quick)
	  {
	    if (statistics)
	      status_var_increment(join->session->status_var.select_range_count);
	  }
	  else
	  {
	    join->session->server_status|=SERVER_QUERY_NO_INDEX_USED;
	    if (statistics)
	      status_var_increment(join->session->status_var.select_scan_count);
	  }
	}
	else
	{
	  if (tab->select && tab->select->quick)
	  {
	    if (statistics)
	      status_var_increment(join->session->status_var.select_full_range_join_count);
	  }
	  else
	  {
	    join->session->server_status|=SERVER_QUERY_NO_INDEX_USED;
	    if (statistics)
	      status_var_increment(join->session->status_var.select_full_join_count);
	  }
	}
	if (!table->no_keyread)
	{
	  if (tab->select && tab->select->quick &&
              tab->select->quick->index != MAX_KEY && //not index_merge
	      table->covering_keys.test(tab->select->quick->index))
	  {
	    table->key_read=1;
	    table->file->extra(HA_EXTRA_KEYREAD);
	  }
	  else if (!table->covering_keys.none() &&
		   !(tab->select && tab->select->quick))
	  {					// Only read index tree
            if (!tab->insideout_match_tab)
            {
              /*
                See bug #26447: "Using the clustered index for a table scan
                is always faster than using a secondary index".
              */
              if (table->s->primary_key != MAX_KEY &&
                  table->file->primary_key_is_clustered())
                tab->index= table->s->primary_key;
              else
                tab->index= table->find_shortest_key(&table->covering_keys);
            }
	    tab->read_first_record= join_read_first;
	    tab->type=JT_NEXT;		// Read with index_first / index_next
	  }
	}
        if (tab->select && tab->select->quick &&
            tab->select->quick->index != MAX_KEY && ! tab->table->key_read)
          push_index_cond(tab, tab->select->quick->index, !using_join_cache);
      }
      break;
    default:
      break;					/* purecov: deadcode */
    case JT_UNKNOWN:
    case JT_MAYBE_REF:
      abort();					/* purecov: deadcode */
    }
  }
  join->join_tab[join->tables-1].next_select=0; /* Set by do_select */
  return(false);
}


/**
  Give error if we some tables are done with a full join.

  This is used by multi_table_update and multi_table_delete when running
  in safe mode.

  @param join		Join condition

  @retval
    0	ok
  @retval
    1	Error (full join used)
*/

bool error_if_full_join(JOIN *join)
{
  for (JOIN_TAB *tab=join->join_tab, *end=join->join_tab+join->tables;
       tab < end;
       tab++)
  {
    if (tab->type == JT_ALL && (!tab->select || !tab->select->quick))
    {
      my_message(ER_UPDATE_WITHOUT_KEY_IN_SAFE_MODE,
                 ER(ER_UPDATE_WITHOUT_KEY_IN_SAFE_MODE), MYF(0));
      return(1);
    }
  }
  return(0);
}


/**
  cleanup JOIN_TAB.
*/

void JOIN_TAB::cleanup()
{
  delete select;
  select= 0;
  delete quick;
  quick= 0;
  if (cache.buff)
    free(cache.buff);
  cache.buff= 0;
  limit= 0;
  if (table)
  {
    if (table->key_read)
    {
      table->key_read= 0;
      table->file->extra(HA_EXTRA_NO_KEYREAD);
    }
    table->file->ha_index_or_rnd_end();
    /*
      We need to reset this for next select
      (Tested in part_of_refkey)
    */
    table->reginfo.join_tab= 0;
  }
  end_read_record(&read_record);
}


/**
  Partially cleanup JOIN after it has executed: close index or rnd read
  (table cursors), free quick selects.

    This function is called in the end of execution of a JOIN, before the used
    tables are unlocked and closed.

    For a join that is resolved using a temporary table, the first sweep is
    performed against actual tables and an intermediate result is inserted
    into the temprorary table.
    The last sweep is performed against the temporary table. Therefore,
    the base tables and associated buffers used to fill the temporary table
    are no longer needed, and this function is called to free them.

    For a join that is performed without a temporary table, this function
    is called after all rows are sent, but before EOF packet is sent.

    For a simple SELECT with no subqueries this function performs a full
    cleanup of the JOIN and calls mysql_unlock_read_tables to free used base
    tables.

    If a JOIN is executed for a subquery or if it has a subquery, we can't
    do the full cleanup and need to do a partial cleanup only.
    - If a JOIN is not the top level join, we must not unlock the tables
    because the outer select may not have been evaluated yet, and we
    can't unlock only selected tables of a query.
    - Additionally, if this JOIN corresponds to a correlated subquery, we
    should not free quick selects and join buffers because they will be
    needed for the next execution of the correlated subquery.
    - However, if this is a JOIN for a [sub]select, which is not
    a correlated subquery itself, but has subqueries, we can free it
    fully and also free JOINs of all its subqueries. The exception
    is a subquery in SELECT list, e.g: @n
    SELECT a, (select cmax(b) from t1) group by c @n
    This subquery will not be evaluated at first sweep and its value will
    not be inserted into the temporary table. Instead, it's evaluated
    when selecting from the temporary table. Therefore, it can't be freed
    here even though it's not correlated.

  @todo
    Unlock tables even if the join isn't top level select in the tree
*/

void JOIN::join_free()
{
  Select_Lex_Unit *tmp_unit;
  Select_Lex *sl;
  /*
    Optimization: if not EXPLAIN and we are done with the JOIN,
    free all tables.
  */
  bool full= (!select_lex->uncacheable && !session->lex->describe);
  bool can_unlock= full;

  cleanup(full);

  for (tmp_unit= select_lex->first_inner_unit();
       tmp_unit;
       tmp_unit= tmp_unit->next_unit())
    for (sl= tmp_unit->first_select(); sl; sl= sl->next_select())
    {
      Item_subselect *subselect= sl->master_unit()->item;
      bool full_local= full && (!subselect || subselect->is_evaluated());
      /*
        If this join is evaluated, we can fully clean it up and clean up all
        its underlying joins even if they are correlated -- they will not be
        used any more anyway.
        If this join is not yet evaluated, we still must clean it up to
        close its table cursors -- it may never get evaluated, as in case of
        ... HAVING false OR a IN (SELECT ...))
        but all table cursors must be closed before the unlock.
      */
      sl->cleanup_all_joins(full_local);
      /* Can't unlock if at least one JOIN is still needed */
      can_unlock= can_unlock && full_local;
    }

  /*
    We are not using tables anymore
    Unlock all tables. We may be in an INSERT .... SELECT statement.
  */
  if (can_unlock && lock && session->lock &&
      !(select_options & SELECT_NO_UNLOCK) &&
      !select_lex->subquery_in_having &&
      (select_lex == (session->lex->unit.fake_select_lex ?
                      session->lex->unit.fake_select_lex : &session->lex->select_lex)))
  {
    /*
      TODO: unlock tables even if the join isn't top level select in the
      tree.
    */
    mysql_unlock_read_tables(session, lock);           // Don't free join->lock
    lock= 0;
  }

  return;
}


/**
  Free resources of given join.

  @param fill   true if we should free all resources, call with full==1
                should be last, before it this function can be called with
                full==0

  @note
    With subquery this function definitely will be called several times,
    but even for simple query it can be called several times.
*/

void JOIN::cleanup(bool full)
{
  if (table)
  {
    JOIN_TAB *tab,*end;
    /*
      Only a sorted table may be cached.  This sorted table is always the
      first non const table in join->table
    */
    if (tables > const_tables) // Test for not-const tables
    {
      free_io_cache(table[const_tables]);
      filesort_free_buffers(table[const_tables],full);
    }

    if (full)
    {
      for (tab= join_tab, end= tab+tables; tab != end; tab++)
	tab->cleanup();
      table= 0;
    }
    else
    {
      for (tab= join_tab, end= tab+tables; tab != end; tab++)
      {
	if (tab->table)
          tab->table->file->ha_index_or_rnd_end();
      }
    }
    cleanup_sj_tmp_tables(this);//
  }
  /*
    We are not using tables anymore
    Unlock all tables. We may be in an INSERT .... SELECT statement.
  */
  if (full)
  {
    if (tmp_join)
      tmp_table_param.copy_field= 0;
    group_fields.delete_elements();
    /*
      We can't call delete_elements() on copy_funcs as this will cause
      problems in free_elements() as some of the elements are then deleted.
    */
    tmp_table_param.copy_funcs.empty();
    /*
      If we have tmp_join and 'this' JOIN is not tmp_join and
      tmp_table_param.copy_field's  of them are equal then we have to remove
      pointer to  tmp_table_param.copy_field from tmp_join, because it qill
      be removed in tmp_table_param.cleanup().
    */
    if (tmp_join &&
        tmp_join != this &&
        tmp_join->tmp_table_param.copy_field ==
        tmp_table_param.copy_field)
    {
      tmp_join->tmp_table_param.copy_field=
        tmp_join->tmp_table_param.save_copy_field= 0;
    }
    tmp_table_param.cleanup();
  }
  return;
}


/**
  Remove the following expressions from order_st BY and GROUP BY:
  Constant expressions @n
  Expression that only uses tables that are of type EQ_REF and the reference
  is in the order_st list or if all refereed tables are of the above type.

  In the following, the X field can be removed:
  @code
  SELECT * FROM t1,t2 WHERE t1.a=t2.a order_st BY t1.a,t2.X
  SELECT * FROM t1,t2,t3 WHERE t1.a=t2.a AND t2.b=t3.b order_st BY t1.a,t3.X
  @endcode

  These can't be optimized:
  @code
  SELECT * FROM t1,t2 WHERE t1.a=t2.a order_st BY t2.X,t1.a
  SELECT * FROM t1,t2 WHERE t1.a=t2.a AND t1.b=t2.b order_st BY t1.a,t2.c
  SELECT * FROM t1,t2 WHERE t1.a=t2.a order_st BY t2.b,t1.a
  @endcode
*/

static bool
eq_ref_table(JOIN *join, order_st *start_order, JOIN_TAB *tab)
{
  if (tab->cached_eq_ref_table)			// If cached
    return tab->eq_ref_table;
  tab->cached_eq_ref_table=1;
  /* We can skip const tables only if not an outer table */
  if (tab->type == JT_CONST && !tab->first_inner)
    return (tab->eq_ref_table=1);		/* purecov: inspected */
  if (tab->type != JT_EQ_REF || tab->table->maybe_null)
    return (tab->eq_ref_table=0);		// We must use this
  Item **ref_item=tab->ref.items;
  Item **end=ref_item+tab->ref.key_parts;
  uint32_t found=0;
  table_map map=tab->table->map;

  for (; ref_item != end ; ref_item++)
  {
    if (! (*ref_item)->const_item())
    {						// Not a const ref
      order_st *order;
      for (order=start_order ; order ; order=order->next)
      {
	if ((*ref_item)->eq(order->item[0],0))
	  break;
      }
      if (order)
      {
	found++;
	assert(!(order->used & map));
	order->used|=map;
	continue;				// Used in order_st BY
      }
      if (!only_eq_ref_tables(join,start_order, (*ref_item)->used_tables()))
	return (tab->eq_ref_table=0);
    }
  }
  /* Check that there was no reference to table before sort order */
  for (; found && start_order ; start_order=start_order->next)
  {
    if (start_order->used & map)
    {
      found--;
      continue;
    }
    if (start_order->depend_map & map)
      return (tab->eq_ref_table=0);
  }
  return tab->eq_ref_table=1;
}


static bool
only_eq_ref_tables(JOIN *join,order_st *order,table_map tables)
{
  for (JOIN_TAB **tab=join->map2table ; tables ; tab++, tables>>=1)
  {
    if (tables & 1 && !eq_ref_table(join, order, *tab))
      return 0;
  }
  return 1;
}


/** Update the dependency map for the tables. */

static void update_depend_map(JOIN *join)
{
  JOIN_TAB *join_tab=join->join_tab, *end=join_tab+join->tables;

  for (; join_tab != end ; join_tab++)
  {
    TABLE_REF *ref= &join_tab->ref;
    table_map depend_map=0;
    Item **item=ref->items;
    uint32_t i;
    for (i=0 ; i < ref->key_parts ; i++,item++)
      depend_map|=(*item)->used_tables();
    ref->depend_map=depend_map & ~OUTER_REF_TABLE_BIT;
    depend_map&= ~OUTER_REF_TABLE_BIT;
    for (JOIN_TAB **tab=join->map2table;
	 depend_map ;
	 tab++,depend_map>>=1 )
    {
      if (depend_map & 1)
	ref->depend_map|=(*tab)->ref.depend_map;
    }
  }
}


/** Update the dependency map for the sort order. */

static void update_depend_map(JOIN *join, order_st *order)
{
  for (; order ; order=order->next)
  {
    table_map depend_map;
    order->item[0]->update_used_tables();
    order->depend_map=depend_map=order->item[0]->used_tables();
    // Not item_sum(), RAND() and no reference to table outside of sub select
    if (!(order->depend_map & (OUTER_REF_TABLE_BIT | RAND_TABLE_BIT))
        && !order->item[0]->with_sum_func)
    {
      for (JOIN_TAB **tab=join->map2table;
	   depend_map ;
	   tab++, depend_map>>=1)
      {
	if (depend_map & 1)
	  order->depend_map|=(*tab)->ref.depend_map;
      }
    }
  }
}


/**
  Remove all constants and check if order_st only contains simple
  expressions.

  simple_order is set to 1 if sort_order only uses fields from head table
  and the head table is not a LEFT JOIN table.

  @param join			Join handler
  @param first_order		List of SORT or GROUP order
  @param cond			WHERE statement
  @param change_list		Set to 1 if we should remove things from list.
                               If this is not set, then only simple_order is
                               calculated.
  @param simple_order		Set to 1 if we are only using simple expressions

  @return
    Returns new sort order
*/

static order_st *
remove_constants(JOIN *join,order_st *first_order, COND *cond,
                 bool change_list, bool *simple_order)
{
  if (join->tables == join->const_tables)
    return change_list ? 0 : first_order;		// No need to sort

  order_st *order,**prev_ptr;
  table_map first_table= join->join_tab[join->const_tables].table->map;
  table_map not_const_tables= ~join->const_table_map;
  table_map ref;

  prev_ptr= &first_order;
  *simple_order= *join->join_tab[join->const_tables].on_expr_ref ? 0 : 1;

  /* NOTE: A variable of not_const_tables ^ first_table; breaks gcc 2.7 */

  update_depend_map(join, first_order);
  for (order=first_order; order ; order=order->next)
  {
    table_map order_tables=order->item[0]->used_tables();
    if (order->item[0]->with_sum_func)
      *simple_order=0;				// Must do a temp table to sort
    else if (!(order_tables & not_const_tables))
    {
      if (order->item[0]->with_subselect)
        order->item[0]->val_str(&order->item[0]->str_value);
      continue;					// skip const item
    }
    else
    {
      if (order_tables & (RAND_TABLE_BIT | OUTER_REF_TABLE_BIT))
	*simple_order=0;
      else
      {
	Item *comp_item=0;
	if (cond && const_expression_in_where(cond,order->item[0], &comp_item))
	{
	  continue;
	}
	if ((ref=order_tables & (not_const_tables ^ first_table)))
	{
	  if (!(order_tables & first_table) &&
              only_eq_ref_tables(join,first_order, ref))
	  {
	    continue;
	  }
	  *simple_order=0;			// Must do a temp table to sort
	}
      }
    }
    if (change_list)
      *prev_ptr= order;				// use this entry
    prev_ptr= &order->next;
  }
  if (change_list)
    *prev_ptr=0;
  if (prev_ptr == &first_order)			// Nothing to sort/group
    *simple_order=1;
  return(first_order);
}


static int
return_zero_rows(JOIN *join, select_result *result,TableList *tables,
		 List<Item> &fields, bool send_row, uint64_t select_options,
		 const char *info, Item *having)
{
  if (select_options & SELECT_DESCRIBE)
  {
    select_describe(join, false, false, false, info);
    return(0);
  }

  join->join_free();

  if (send_row)
  {
    for (TableList *table= tables; table; table= table->next_leaf)
      table->table->mark_as_null_row();		// All fields are NULL
    if (having && having->val_int() == 0)
      send_row=0;
  }
  if (!(result->send_fields(fields,
                              Protocol::SEND_NUM_ROWS | Protocol::SEND_EOF)))
  {
    if (send_row)
    {
      List_iterator_fast<Item> it(fields);
      Item *item;
      while ((item= it++))
	item->no_rows_in_result();
      result->send_data(fields);
    }
    result->send_eof();				// Should be safe
  }
  /* Update results for FOUND_ROWS */
  join->session->limit_found_rows= join->session->examined_row_count= 0;
  return(0);
}

/*
  used only in JOIN::clear
*/
static void clear_tables(JOIN *join)
{
  /*
    must clear only the non-const tables, as const tables
    are not re-calculated.
  */
  for (uint32_t i=join->const_tables ; i < join->tables ; i++)
    join->table[i]->mark_as_null_row();		// All fields are NULL
}

/*****************************************************************************
  Make som simple condition optimization:
  If there is a test 'field = const' change all refs to 'field' to 'const'
  Remove all dummy tests 'item = item', 'const op const'.
  Remove all 'item is NULL', when item can never be null!
  item->marker should be 0 for all items on entry
  Return in cond_value false if condition is impossible (1 = 2)
*****************************************************************************/

class COND_CMP :public ilink {
public:
  static void *operator new(size_t size)
  {
    return (void*) sql_alloc((uint32_t) size);
  }
  static void operator delete(void *, size_t)
  { TRASH(ptr, size); }

  Item *and_level;
  Item_func *cmp_func;
  COND_CMP(Item *a,Item_func *b) :and_level(a),cmp_func(b) {}
};

#ifdef HAVE_EXPLICIT_TEMPLATE_INSTANTIATION
template class I_List<COND_CMP>;
template class I_List_iterator<COND_CMP>;
#endif


/**
  Find the multiple equality predicate containing a field.

  The function retrieves the multiple equalities accessed through
  the con_equal structure from current level and up looking for
  an equality containing field. It stops retrieval as soon as the equality
  is found and set up inherited_fl to true if it's found on upper levels.

  @param cond_equal          multiple equalities to search in
  @param field               field to look for
  @param[out] inherited_fl   set up to true if multiple equality is found
                             on upper levels (not on current level of
                             cond_equal)

  @return
    - Item_equal for the found multiple equality predicate if a success;
    - NULL otherwise.
*/

Item_equal *find_item_equal(COND_EQUAL *cond_equal, Field *field,
                            bool *inherited_fl)
{
  Item_equal *item= 0;
  bool in_upper_level= false;
  while (cond_equal)
  {
    List_iterator_fast<Item_equal> li(cond_equal->current_level);
    while ((item= li++))
    {
      if (item->contains(field))
        goto finish;
    }
    in_upper_level= true;
    cond_equal= cond_equal->upper_levels;
  }
  in_upper_level= false;
finish:
  *inherited_fl= in_upper_level;
  return item;
}


/**
  Check whether an equality can be used to build multiple equalities.

    This function first checks whether the equality (left_item=right_item)
    is a simple equality i.e. the one that equates a field with another field
    or a constant (field=field_item or field=const_item).
    If this is the case the function looks for a multiple equality
    in the lists referenced directly or indirectly by cond_equal inferring
    the given simple equality. If it doesn't find any, it builds a multiple
    equality that covers the predicate, i.e. the predicate can be inferred
    from this multiple equality.
    The built multiple equality could be obtained in such a way:
    create a binary  multiple equality equivalent to the predicate, then
    merge it, if possible, with one of old multiple equalities.
    This guarantees that the set of multiple equalities covering equality
    predicates will be minimal.

  EXAMPLE:
    For the where condition
    @code
      WHERE a=b AND b=c AND
            (b=2 OR f=e)
    @endcode
    the check_equality will be called for the following equality
    predicates a=b, b=c, b=2 and f=e.
    - For a=b it will be called with *cond_equal=(0,[]) and will transform
      *cond_equal into (0,[Item_equal(a,b)]).
    - For b=c it will be called with *cond_equal=(0,[Item_equal(a,b)])
      and will transform *cond_equal into CE=(0,[Item_equal(a,b,c)]).
    - For b=2 it will be called with *cond_equal=(ptr(CE),[])
      and will transform *cond_equal into (ptr(CE),[Item_equal(2,a,b,c)]).
    - For f=e it will be called with *cond_equal=(ptr(CE), [])
      and will transform *cond_equal into (ptr(CE),[Item_equal(f,e)]).

  @note
    Now only fields that have the same type definitions (verified by
    the Field::eq_def method) are placed to the same multiple equalities.
    Because of this some equality predicates are not eliminated and
    can be used in the constant propagation procedure.
    We could weeken the equlity test as soon as at least one of the
    equal fields is to be equal to a constant. It would require a
    more complicated implementation: we would have to store, in
    general case, its own constant for each fields from the multiple
    equality. But at the same time it would allow us to get rid
    of constant propagation completely: it would be done by the call
    to build_equal_items_for_cond.


    The implementation does not follow exactly the above rules to
    build a new multiple equality for the equality predicate.
    If it processes the equality of the form field1=field2, it
    looks for multiple equalities me1 containig field1 and me2 containing
    field2. If only one of them is found the fuction expands it with
    the lacking field. If multiple equalities for both fields are
    found they are merged. If both searches fail a new multiple equality
    containing just field1 and field2 is added to the existing
    multiple equalities.
    If the function processes the predicate of the form field1=const,
    it looks for a multiple equality containing field1. If found, the
    function checks the constant of the multiple equality. If the value
    is unknown, it is setup to const. Otherwise the value is compared with
    const and the evaluation of the equality predicate is performed.
    When expanding/merging equality predicates from the upper levels
    the function first copies them for the current level. It looks
    acceptable, as this happens rarely. The implementation without
    copying would be much more complicated.

  @param left_item   left term of the quality to be checked
  @param right_item  right term of the equality to be checked
  @param item        equality item if the equality originates from a condition
                     predicate, 0 if the equality is the result of row
                     elimination
  @param cond_equal  multiple equalities that must hold together with the
                     equality

  @retval
    true    if the predicate is a simple equality predicate to be used
    for building multiple equalities
  @retval
    false   otherwise
*/

static bool check_simple_equality(Item *left_item, Item *right_item,
                                  Item *item, COND_EQUAL *cond_equal)
{
  if (left_item->type() == Item::FIELD_ITEM &&
      right_item->type() == Item::FIELD_ITEM &&
      !((Item_field*)left_item)->depended_from &&
      !((Item_field*)right_item)->depended_from)
  {
    /* The predicate the form field1=field2 is processed */

    Field *left_field= ((Item_field*) left_item)->field;
    Field *right_field= ((Item_field*) right_item)->field;

    if (!left_field->eq_def(right_field))
      return false;

    /* Search for multiple equalities containing field1 and/or field2 */
    bool left_copyfl, right_copyfl;
    Item_equal *left_item_equal=
               find_item_equal(cond_equal, left_field, &left_copyfl);
    Item_equal *right_item_equal=
               find_item_equal(cond_equal, right_field, &right_copyfl);

    /* As (NULL=NULL) != true we can't just remove the predicate f=f */
    if (left_field->eq(right_field)) /* f = f */
      return (!(left_field->maybe_null() && !left_item_equal));

    if (left_item_equal && left_item_equal == right_item_equal)
    {
      /*
        The equality predicate is inference of one of the existing
        multiple equalities, i.e the condition is already covered
        by upper level equalities
      */
       return true;
    }

    bool copy_item_name= test(item && item->name >= subq_sj_cond_name &&
                              item->name < subq_sj_cond_name + 64);
    /* Copy the found multiple equalities at the current level if needed */
    if (left_copyfl)
    {
      /* left_item_equal of an upper level contains left_item */
      left_item_equal= new Item_equal(left_item_equal);
      cond_equal->current_level.push_back(left_item_equal);
      if (copy_item_name)
        left_item_equal->name = item->name;
    }
    if (right_copyfl)
    {
      /* right_item_equal of an upper level contains right_item */
      right_item_equal= new Item_equal(right_item_equal);
      cond_equal->current_level.push_back(right_item_equal);
      if (copy_item_name)
        right_item_equal->name = item->name;
    }

    if (left_item_equal)
    {
      /* left item was found in the current or one of the upper levels */
      if (! right_item_equal)
        left_item_equal->add((Item_field *) right_item);
      else
      {
        /* Merge two multiple equalities forming a new one */
        left_item_equal->merge(right_item_equal);
        /* Remove the merged multiple equality from the list */
        List_iterator<Item_equal> li(cond_equal->current_level);
        while ((li++) != right_item_equal) {};
        li.remove();
      }
    }
    else
    {
      /* left item was not found neither the current nor in upper levels  */
      if (right_item_equal)
      {
        right_item_equal->add((Item_field *) left_item);
        if (copy_item_name)
          right_item_equal->name = item->name;
      }
      else
      {
        /* None of the fields was found in multiple equalities */
        Item_equal *item_equal= new Item_equal((Item_field *) left_item,
                                               (Item_field *) right_item);
        cond_equal->current_level.push_back(item_equal);
        if (copy_item_name)
          item_equal->name = item->name;
      }
    }
    return true;
  }

  {
    /* The predicate of the form field=const/const=field is processed */
    Item *const_item= 0;
    Item_field *field_item= 0;
    if (left_item->type() == Item::FIELD_ITEM &&
        !((Item_field*)left_item)->depended_from &&
        right_item->const_item())
    {
      field_item= (Item_field*) left_item;
      const_item= right_item;
    }
    else if (right_item->type() == Item::FIELD_ITEM &&
             !((Item_field*)right_item)->depended_from &&
             left_item->const_item())
    {
      field_item= (Item_field*) right_item;
      const_item= left_item;
    }

    if (const_item &&
        field_item->result_type() == const_item->result_type())
    {
      bool copyfl;

      if (field_item->result_type() == STRING_RESULT)
      {
        const CHARSET_INFO * const cs= ((Field_str*) field_item->field)->charset();
        if (!item)
        {
          Item_func_eq *eq_item;
          if ((eq_item= new Item_func_eq(left_item, right_item)))
            return false;
          eq_item->set_cmp_func();
          eq_item->quick_fix_field();
          item= eq_item;
        }
        if ((cs != ((Item_func *) item)->compare_collation()) ||
            !cs->coll->propagate(cs, 0, 0))
          return false;
      }

      Item_equal *item_equal = find_item_equal(cond_equal,
                                               field_item->field, &copyfl);
      if (copyfl)
      {
        item_equal= new Item_equal(item_equal);
        cond_equal->current_level.push_back(item_equal);
      }
      if (item_equal)
      {
        /*
          The flag cond_false will be set to 1 after this, if item_equal
          already contains a constant and its value is  not equal to
          the value of const_item.
        */
        item_equal->add(const_item);
      }
      else
      {
        item_equal= new Item_equal(const_item, field_item);
        cond_equal->current_level.push_back(item_equal);
      }
      return true;
    }
  }
  return false;
}


/**
  Convert row equalities into a conjunction of regular equalities.

    The function converts a row equality of the form (E1,...,En)=(E'1,...,E'n)
    into a list of equalities E1=E'1,...,En=E'n. For each of these equalities
    Ei=E'i the function checks whether it is a simple equality or a row
    equality. If it is a simple equality it is used to expand multiple
    equalities of cond_equal. If it is a row equality it converted to a
    sequence of equalities between row elements. If Ei=E'i is neither a
    simple equality nor a row equality the item for this predicate is added
    to eq_list.

  @param session        thread handle
  @param left_row   left term of the row equality to be processed
  @param right_row  right term of the row equality to be processed
  @param cond_equal multiple equalities that must hold together with the
                    predicate
  @param eq_list    results of conversions of row equalities that are not
                    simple enough to form multiple equalities

  @retval
    true    if conversion has succeeded (no fatal error)
  @retval
    false   otherwise
*/

static bool check_row_equality(Session *session, Item *left_row, Item_row *right_row,
                               COND_EQUAL *cond_equal, List<Item>* eq_list)
{
  uint32_t n= left_row->cols();
  for (uint32_t i= 0 ; i < n; i++)
  {
    bool is_converted;
    Item *left_item= left_row->element_index(i);
    Item *right_item= right_row->element_index(i);
    if (left_item->type() == Item::ROW_ITEM &&
        right_item->type() == Item::ROW_ITEM)
    {
      is_converted= check_row_equality(session,
                                       (Item_row *) left_item,
                                       (Item_row *) right_item,
			               cond_equal, eq_list);
      if (!is_converted)
        session->lex->current_select->cond_count++;
    }
    else
    {
      is_converted= check_simple_equality(left_item, right_item, 0, cond_equal);
      session->lex->current_select->cond_count++;
    }

    if (!is_converted)
    {
      Item_func_eq *eq_item;
      if (!(eq_item= new Item_func_eq(left_item, right_item)))
        return false;
      eq_item->set_cmp_func();
      eq_item->quick_fix_field();
      eq_list->push_back(eq_item);
    }
  }
  return true;
}


/**
  Eliminate row equalities and form multiple equalities predicates.

    This function checks whether the item is a simple equality
    i.e. the one that equates a field with another field or a constant
    (field=field_item or field=constant_item), or, a row equality.
    For a simple equality the function looks for a multiple equality
    in the lists referenced directly or indirectly by cond_equal inferring
    the given simple equality. If it doesn't find any, it builds/expands
    multiple equality that covers the predicate.
    Row equalities are eliminated substituted for conjunctive regular
    equalities which are treated in the same way as original equality
    predicates.

  @param session        thread handle
  @param item       predicate to process
  @param cond_equal multiple equalities that must hold together with the
                    predicate
  @param eq_list    results of conversions of row equalities that are not
                    simple enough to form multiple equalities

  @retval
    true   if re-writing rules have been applied
  @retval
    false  otherwise, i.e.
           if the predicate is not an equality,
           or, if the equality is neither a simple one nor a row equality,
           or, if the procedure fails by a fatal error.
*/

static bool check_equality(Session *session, Item *item, COND_EQUAL *cond_equal,
                           List<Item> *eq_list)
{
  if (item->type() == Item::FUNC_ITEM &&
         ((Item_func*) item)->functype() == Item_func::EQ_FUNC)
  {
    Item *left_item= ((Item_func*) item)->arguments()[0];
    Item *right_item= ((Item_func*) item)->arguments()[1];

    if (left_item->type() == Item::ROW_ITEM &&
        right_item->type() == Item::ROW_ITEM)
    {
      session->lex->current_select->cond_count--;
      return check_row_equality(session,
                                (Item_row *) left_item,
                                (Item_row *) right_item,
                                cond_equal, eq_list);
    }
    else
      return check_simple_equality(left_item, right_item, item, cond_equal);
  }
  return false;
}


/**
  Replace all equality predicates in a condition by multiple equality items.

    At each 'and' level the function detects items for equality predicates
    and replaced them by a set of multiple equality items of class Item_equal,
    taking into account inherited equalities from upper levels.
    If an equality predicate is used not in a conjunction it's just
    replaced by a multiple equality predicate.
    For each 'and' level the function set a pointer to the inherited
    multiple equalities in the cond_equal field of the associated
    object of the type Item_cond_and.
    The function also traverses the cond tree and and for each field reference
    sets a pointer to the multiple equality item containing the field, if there
    is any. If this multiple equality equates fields to a constant the
    function replaces the field reference by the constant in the cases
    when the field is not of a string type or when the field reference is
    just an argument of a comparison predicate.
    The function also determines the maximum number of members in
    equality lists of each Item_cond_and object assigning it to
    session->lex->current_select->max_equal_elems.

  @note
    Multiple equality predicate =(f1,..fn) is equivalent to the conjuction of
    f1=f2, .., fn-1=fn. It substitutes any inference from these
    equality predicates that is equivalent to the conjunction.
    Thus, =(a1,a2,a3) can substitute for ((a1=a3) AND (a2=a3) AND (a2=a1)) as
    it is equivalent to ((a1=a2) AND (a2=a3)).
    The function always makes a substitution of all equality predicates occured
    in a conjuction for a minimal set of multiple equality predicates.
    This set can be considered as a canonical representation of the
    sub-conjunction of the equality predicates.
    E.g. (t1.a=t2.b AND t2.b>5 AND t1.a=t3.c) is replaced by
    (=(t1.a,t2.b,t3.c) AND t2.b>5), not by
    (=(t1.a,t2.b) AND =(t1.a,t3.c) AND t2.b>5);
    while (t1.a=t2.b AND t2.b>5 AND t3.c=t4.d) is replaced by
    (=(t1.a,t2.b) AND =(t3.c=t4.d) AND t2.b>5),
    but if additionally =(t4.d,t2.b) is inherited, it
    will be replaced by (=(t1.a,t2.b,t3.c,t4.d) AND t2.b>5)

    The function performs the substitution in a recursive descent by
    the condtion tree, passing to the next AND level a chain of multiple
    equality predicates which have been built at the upper levels.
    The Item_equal items built at the level are attached to other
    non-equality conjucts as a sublist. The pointer to the inherited
    multiple equalities is saved in the and condition object (Item_cond_and).
    This chain allows us for any field reference occurence easyly to find a
    multiple equality that must be held for this occurence.
    For each AND level we do the following:
    - scan it for all equality predicate (=) items
    - join them into disjoint Item_equal() groups
    - process the included OR conditions recursively to do the same for
      lower AND levels.

    We need to do things in this order as lower AND levels need to know about
    all possible Item_equal objects in upper levels.

  @param session        thread handle
  @param cond       condition(expression) where to make replacement
  @param inherited  path to all inherited multiple equality items

  @return
    pointer to the transformed condition
*/

static COND *build_equal_items_for_cond(Session *session, COND *cond,
                                        COND_EQUAL *inherited)
{
  Item_equal *item_equal;
  COND_EQUAL cond_equal;
  cond_equal.upper_levels= inherited;

  if (cond->type() == Item::COND_ITEM)
  {
    List<Item> eq_list;
    bool and_level= ((Item_cond*) cond)->functype() ==
      Item_func::COND_AND_FUNC;
    List<Item> *args= ((Item_cond*) cond)->argument_list();

    List_iterator<Item> li(*args);
    Item *item;

    if (and_level)
    {
      /*
         Retrieve all conjucts of this level detecting the equality
         that are subject to substitution by multiple equality items and
         removing each such predicate from the conjunction after having
         found/created a multiple equality whose inference the predicate is.
     */
      while ((item= li++))
      {
        /*
          PS/SP note: we can safely remove a node from AND-OR
          structure here because it's restored before each
          re-execution of any prepared statement/stored procedure.
        */
        if (check_equality(session, item, &cond_equal, &eq_list))
          li.remove();
      }

      List_iterator_fast<Item_equal> it(cond_equal.current_level);
      while ((item_equal= it++))
      {
        item_equal->fix_length_and_dec();
        item_equal->update_used_tables();
        set_if_bigger(session->lex->current_select->max_equal_elems,
                      item_equal->members());
      }

      ((Item_cond_and*)cond)->cond_equal= cond_equal;
      inherited= &(((Item_cond_and*)cond)->cond_equal);
    }
    /*
       Make replacement of equality predicates for lower levels
       of the condition expression.
    */
    li.rewind();
    while ((item= li++))
    {
      Item *new_item;
      if ((new_item= build_equal_items_for_cond(session, item, inherited)) != item)
      {
        /* This replacement happens only for standalone equalities */
        /*
          This is ok with PS/SP as the replacement is done for
          arguments of an AND/OR item, which are restored for each
          execution of PS/SP.
        */
        li.replace(new_item);
      }
    }
    if (and_level)
    {
      args->concat(&eq_list);
      args->concat((List<Item> *)&cond_equal.current_level);
    }
  }
  else if (cond->type() == Item::FUNC_ITEM)
  {
    List<Item> eq_list;
    /*
      If an equality predicate forms the whole and level,
      we call it standalone equality and it's processed here.
      E.g. in the following where condition
      WHERE a=5 AND (b=5 or a=c)
      (b=5) and (a=c) are standalone equalities.
      In general we can't leave alone standalone eqalities:
      for WHERE a=b AND c=d AND (b=c OR d=5)
      b=c is replaced by =(a,b,c,d).
     */
    if (check_equality(session, cond, &cond_equal, &eq_list))
    {
      int n= cond_equal.current_level.elements + eq_list.elements;
      if (n == 0)
        return new Item_int((int64_t) 1,1);
      else if (n == 1)
      {
        if ((item_equal= cond_equal.current_level.pop()))
        {
          item_equal->fix_length_and_dec();
          item_equal->update_used_tables();
	}
        else
          item_equal= (Item_equal *) eq_list.pop();
        set_if_bigger(session->lex->current_select->max_equal_elems,
                      item_equal->members());
        return item_equal;
      }
      else
      {
        /*
          Here a new AND level must be created. It can happen only
          when a row equality is processed as a standalone predicate.
	*/
        Item_cond_and *and_cond= new Item_cond_and(eq_list);
        and_cond->quick_fix_field();
        List<Item> *args= and_cond->argument_list();
        List_iterator_fast<Item_equal> it(cond_equal.current_level);
        while ((item_equal= it++))
        {
          item_equal->fix_length_and_dec();
          item_equal->update_used_tables();
          set_if_bigger(session->lex->current_select->max_equal_elems,
                        item_equal->members());
        }
        and_cond->cond_equal= cond_equal;
        args->concat((List<Item> *)&cond_equal.current_level);

        return and_cond;
      }
    }
    /*
      For each field reference in cond, not from equal item predicates,
      set a pointer to the multiple equality it belongs to (if there is any)
      as soon the field is not of a string type or the field reference is
      an argument of a comparison predicate.
    */
    unsigned char *is_subst_valid= (unsigned char *) 1;
    cond= cond->compile(&Item::subst_argument_checker,
                        &is_subst_valid,
                        &Item::equal_fields_propagator,
                        (unsigned char *) inherited);
    cond->update_used_tables();
  }
  return cond;
}


/**
  Build multiple equalities for a condition and all on expressions that
  inherit these multiple equalities.

    The function first applies the build_equal_items_for_cond function
    to build all multiple equalities for condition cond utilizing equalities
    referred through the parameter inherited. The extended set of
    equalities is returned in the structure referred by the cond_equal_ref
    parameter. After this the function calls itself recursively for
    all on expressions whose direct references can be found in join_list
    and who inherit directly the multiple equalities just having built.

  @note
    The on expression used in an outer join operation inherits all equalities
    from the on expression of the embedding join, if there is any, or
    otherwise - from the where condition.
    This fact is not obvious, but presumably can be proved.
    Consider the following query:
    @code
      SELECT * FROM (t1,t2) LEFT JOIN (t3,t4) ON t1.a=t3.a AND t2.a=t4.a
        WHERE t1.a=t2.a;
    @endcode
    If the on expression in the query inherits =(t1.a,t2.a), then we
    can build the multiple equality =(t1.a,t2.a,t3.a,t4.a) that infers
    the equality t3.a=t4.a. Although the on expression
    t1.a=t3.a AND t2.a=t4.a AND t3.a=t4.a is not equivalent to the one
    in the query the latter can be replaced by the former: the new query
    will return the same result set as the original one.

    Interesting that multiple equality =(t1.a,t2.a,t3.a,t4.a) allows us
    to use t1.a=t3.a AND t3.a=t4.a under the on condition:
    @code
      SELECT * FROM (t1,t2) LEFT JOIN (t3,t4) ON t1.a=t3.a AND t3.a=t4.a
        WHERE t1.a=t2.a
    @endcode
    This query equivalent to:
    @code
      SELECT * FROM (t1 LEFT JOIN (t3,t4) ON t1.a=t3.a AND t3.a=t4.a),t2
        WHERE t1.a=t2.a
    @endcode
    Similarly the original query can be rewritten to the query:
    @code
      SELECT * FROM (t1,t2) LEFT JOIN (t3,t4) ON t2.a=t4.a AND t3.a=t4.a
        WHERE t1.a=t2.a
    @endcode
    that is equivalent to:
    @code
      SELECT * FROM (t2 LEFT JOIN (t3,t4)ON t2.a=t4.a AND t3.a=t4.a), t1
        WHERE t1.a=t2.a
    @endcode
    Thus, applying equalities from the where condition we basically
    can get more freedom in performing join operations.
    Althogh we don't use this property now, it probably makes sense to use
    it in the future.
  @param session		      Thread handler
  @param cond                condition to build the multiple equalities for
  @param inherited           path to all inherited multiple equality items
  @param join_list           list of join tables to which the condition
                             refers to
  @param[out] cond_equal_ref pointer to the structure to place built
                             equalities in

  @return
    pointer to the transformed condition containing multiple equalities
*/

static COND *build_equal_items(Session *session, COND *cond,
                               COND_EQUAL *inherited,
                               List<TableList> *join_list,
                               COND_EQUAL **cond_equal_ref)
{
  COND_EQUAL *cond_equal= 0;

  if (cond)
  {
    cond= build_equal_items_for_cond(session, cond, inherited);
    cond->update_used_tables();
    if (cond->type() == Item::COND_ITEM &&
        ((Item_cond*) cond)->functype() == Item_func::COND_AND_FUNC)
      cond_equal= &((Item_cond_and*) cond)->cond_equal;
    else if (cond->type() == Item::FUNC_ITEM &&
             ((Item_cond*) cond)->functype() == Item_func::MULT_EQUAL_FUNC)
    {
      cond_equal= new COND_EQUAL;
      cond_equal->current_level.push_back((Item_equal *) cond);
    }
  }
  if (cond_equal)
  {
    cond_equal->upper_levels= inherited;
    inherited= cond_equal;
  }
  *cond_equal_ref= cond_equal;

  if (join_list)
  {
    TableList *table;
    List_iterator<TableList> li(*join_list);

    while ((table= li++))
    {
      if (table->on_expr)
      {
        List<TableList> *nested_join_list= table->nested_join ?
          &table->nested_join->join_list : NULL;
        /*
          We can modify table->on_expr because its old value will
          be restored before re-execution of PS/SP.
        */
        table->on_expr= build_equal_items(session, table->on_expr, inherited,
                                          nested_join_list,
                                          &table->cond_equal);
      }
    }
  }

  return cond;
}


/**
  Compare field items by table order in the execution plan.

    field1 considered as better than field2 if the table containing
    field1 is accessed earlier than the table containing field2.
    The function finds out what of two fields is better according
    this criteria.

  @param field1          first field item to compare
  @param field2          second field item to compare
  @param table_join_idx  index to tables determining table order

  @retval
    1  if field1 is better than field2
  @retval
    -1  if field2 is better than field1
  @retval
    0  otherwise
*/

static int compare_fields_by_table_order(Item_field *field1,
                                  Item_field *field2,
                                  void *table_join_idx)
{
  int cmp= 0;
  bool outer_ref= 0;
  if (field2->used_tables() & OUTER_REF_TABLE_BIT)
  {
    outer_ref= 1;
    cmp= -1;
  }
  if (field2->used_tables() & OUTER_REF_TABLE_BIT)
  {
    outer_ref= 1;
    cmp++;
  }
  if (outer_ref)
    return cmp;
  JOIN_TAB **idx= (JOIN_TAB **) table_join_idx;
  cmp= idx[field2->field->table->tablenr]-idx[field1->field->table->tablenr];
  return cmp < 0 ? -1 : (cmp ? 1 : 0);
}


/**
  Generate minimal set of simple equalities equivalent to a multiple equality.

    The function retrieves the fields of the multiple equality item
    item_equal and  for each field f:
    - if item_equal contains const it generates the equality f=const_item;
    - otherwise, if f is not the first field, generates the equality
      f=item_equal->get_first().
    All generated equality are added to the cond conjunction.

  @param cond            condition to add the generated equality to
  @param upper_levels    structure to access multiple equality of upper levels
  @param item_equal      multiple equality to generate simple equality from

  @note
    Before generating an equality function checks that it has not
    been generated for multiple equalities of the upper levels.
    E.g. for the following where condition
    WHERE a=5 AND ((a=b AND b=c) OR  c>4)
    the upper level AND condition will contain =(5,a),
    while the lower level AND condition will contain =(5,a,b,c).
    When splitting =(5,a,b,c) into a separate equality predicates
    we should omit 5=a, as we have it already in the upper level.
    The following where condition gives us a more complicated case:
    WHERE t1.a=t2.b AND t3.c=t4.d AND (t2.b=t3.c OR t4.e>5 ...) AND ...
    Given the tables are accessed in the order t1->t2->t3->t4 for
    the selected query execution plan the lower level multiple
    equality =(t1.a,t2.b,t3.c,t4.d) formally  should be converted to
    t1.a=t2.b AND t1.a=t3.c AND t1.a=t4.d. But t1.a=t2.a will be
    generated for the upper level. Also t3.c=t4.d will be generated there.
    So only t1.a=t3.c should be left in the lower level.
    If cond is equal to 0, then not more then one equality is generated
    and a pointer to it is returned as the result of the function.

  @return
    - The condition with generated simple equalities or
    a pointer to the simple generated equality, if success.
    - 0, otherwise.
*/

static Item *eliminate_item_equal(COND *cond, COND_EQUAL *upper_levels,
                                  Item_equal *item_equal)
{
  List<Item> eq_list;
  Item_func_eq *eq_item= 0;
  if (((Item *) item_equal)->const_item() && !item_equal->val_int())
    return new Item_int((int64_t) 0,1);
  Item *item_const= item_equal->get_const();
  Item_equal_iterator it(*item_equal);
  Item *head;
  if (item_const)
    head= item_const;
  else
  {
    head= item_equal->get_first();
    it++;
  }
  Item_field *item_field;
  while ((item_field= it++))
  {
    Item_equal *upper= item_field->find_item_equal(upper_levels);
    Item_field *item= item_field;
    if (upper)
    {
      if (item_const && upper->get_const())
        item= 0;
      else
      {
        Item_equal_iterator li(*item_equal);
        while ((item= li++) != item_field)
        {
          if (item->find_item_equal(upper_levels) == upper)
            break;
        }
      }
    }
    if (item == item_field)
    {
      if (eq_item)
        eq_list.push_back(eq_item);
      eq_item= new Item_func_eq(item_field, head);
      if (!eq_item)
        return 0;
      eq_item->set_cmp_func();
      eq_item->quick_fix_field();
   }
  }

  if (!cond && !eq_list.head())
  {
    if (!eq_item)
      return new Item_int((int64_t) 1,1);
    return eq_item;
  }

  if (eq_item)
    eq_list.push_back(eq_item);
  if (!cond)
    cond= new Item_cond_and(eq_list);
  else
  {
    assert(cond->type() == Item::COND_ITEM);
    ((Item_cond *) cond)->add_at_head(&eq_list);
  }

  cond->quick_fix_field();
  cond->update_used_tables();

  return cond;
}


/**
  Substitute every field reference in a condition by the best equal field
  and eliminate all multiple equality predicates.

    The function retrieves the cond condition and for each encountered
    multiple equality predicate it sorts the field references in it
    according to the order of tables specified by the table_join_idx
    parameter. Then it eliminates the multiple equality predicate it
    replacing it by the conjunction of simple equality predicates
    equating every field from the multiple equality to the first
    field in it, or to the constant, if there is any.
    After this the function retrieves all other conjuncted
    predicates substitute every field reference by the field reference
    to the first equal field or equal constant if there are any.
  @param cond            condition to process
  @param cond_equal      multiple equalities to take into consideration
  @param table_join_idx  index to tables determining field preference

  @note
    At the first glance full sort of fields in multiple equality
    seems to be an overkill. Yet it's not the case due to possible
    new fields in multiple equality item of lower levels. We want
    the order in them to comply with the order of upper levels.

  @return
    The transformed condition
*/

static COND* substitute_for_best_equal_field(COND *cond,
                                             COND_EQUAL *cond_equal,
                                             void *table_join_idx)
{
  Item_equal *item_equal;

  if (cond->type() == Item::COND_ITEM)
  {
    List<Item> *cond_list= ((Item_cond*) cond)->argument_list();

    bool and_level= ((Item_cond*) cond)->functype() ==
                      Item_func::COND_AND_FUNC;
    if (and_level)
    {
      cond_equal= &((Item_cond_and *) cond)->cond_equal;
      cond_list->disjoin((List<Item> *) &cond_equal->current_level);

      List_iterator_fast<Item_equal> it(cond_equal->current_level);
      while ((item_equal= it++))
      {
        item_equal->sort(&compare_fields_by_table_order, table_join_idx);
      }
    }

    List_iterator<Item> li(*cond_list);
    Item *item;
    while ((item= li++))
    {
      Item *new_item =substitute_for_best_equal_field(item, cond_equal,
                                                      table_join_idx);
      /*
        This works OK with PS/SP re-execution as changes are made to
        the arguments of AND/OR items only
      */
      if (new_item != item)
        li.replace(new_item);
    }

    if (and_level)
    {
      List_iterator_fast<Item_equal> it(cond_equal->current_level);
      while ((item_equal= it++))
      {
        cond= eliminate_item_equal(cond, cond_equal->upper_levels, item_equal);
        // This occurs when eliminate_item_equal() founds that cond is
        // always false and substitutes it with Item_int 0.
        // Due to this, value of item_equal will be 0, so just return it.
        if (cond->type() != Item::COND_ITEM)
          break;
      }
    }
    if (cond->type() == Item::COND_ITEM &&
        !((Item_cond*)cond)->argument_list()->elements)
      cond= new Item_int((int32_t)cond->val_bool());

  }
  else if (cond->type() == Item::FUNC_ITEM &&
           ((Item_cond*) cond)->functype() == Item_func::MULT_EQUAL_FUNC)
  {
    item_equal= (Item_equal *) cond;
    item_equal->sort(&compare_fields_by_table_order, table_join_idx);
    if (cond_equal && cond_equal->current_level.head() == item_equal)
      cond_equal= 0;
    return eliminate_item_equal(0, cond_equal, item_equal);
  }
  else
    cond->transform(&Item::replace_equal_field, 0);
  return cond;
}


/**
  Check appearance of new constant items in multiple equalities
  of a condition after reading a constant table.

    The function retrieves the cond condition and for each encountered
    multiple equality checks whether new constants have appeared after
    reading the constant (single row) table tab. If so it adjusts
    the multiple equality appropriately.

  @param cond       condition whose multiple equalities are to be checked
  @param table      constant table that has been read
*/

static void update_const_equal_items(COND *cond, JOIN_TAB *tab)
{
  if (!(cond->used_tables() & tab->table->map))
    return;

  if (cond->type() == Item::COND_ITEM)
  {
    List<Item> *cond_list= ((Item_cond*) cond)->argument_list();
    List_iterator_fast<Item> li(*cond_list);
    Item *item;
    while ((item= li++))
      update_const_equal_items(item, tab);
  }
  else if (cond->type() == Item::FUNC_ITEM &&
           ((Item_cond*) cond)->functype() == Item_func::MULT_EQUAL_FUNC)
  {
    Item_equal *item_equal= (Item_equal *) cond;
    bool contained_const= item_equal->get_const() != NULL;
    item_equal->update_const();
    if (!contained_const && item_equal->get_const())
    {
      /* Update keys for range analysis */
      Item_equal_iterator it(*item_equal);
      Item_field *item_field;
      while ((item_field= it++))
      {
        Field *field= item_field->field;
        JOIN_TAB *stat= field->table->reginfo.join_tab;
        key_map possible_keys= field->key_start;
        possible_keys&= field->table->keys_in_use_for_query;
        stat[0].const_keys|= possible_keys;

        /*
          For each field in the multiple equality (for which we know that it
          is a constant) we have to find its corresponding key part, and set
          that key part in const_key_parts.
        */
        if (possible_keys.any())
        {
          Table *field_tab= field->table;
          KEYUSE *use;
          for (use= stat->keyuse; use && use->table == field_tab; use++)
            if (possible_keys.test(use->key) &&
                field_tab->key_info[use->key].key_part[use->keypart].field ==
                field)
              field_tab->const_key_parts[use->key]|= use->keypart_map;
        }
      }
    }
  }
}


/*
  change field = field to field = const for each found field = const in the
  and_level
*/

static void
change_cond_ref_to_const(Session *session, I_List<COND_CMP> *save_list,
                         Item *and_father, Item *cond,
                         Item *field, Item *value)
{
  if (cond->type() == Item::COND_ITEM)
  {
    bool and_level= ((Item_cond*) cond)->functype() ==
      Item_func::COND_AND_FUNC;
    List_iterator<Item> li(*((Item_cond*) cond)->argument_list());
    Item *item;
    while ((item=li++))
      change_cond_ref_to_const(session, save_list,and_level ? cond : item, item,
			       field, value);
    return;
  }
  if (cond->eq_cmp_result() == Item::COND_OK)
    return;					// Not a boolean function

  Item_bool_func2 *func=  (Item_bool_func2*) cond;
  Item **args= func->arguments();
  Item *left_item=  args[0];
  Item *right_item= args[1];
  Item_func::Functype functype=  func->functype();

  if (right_item->eq(field,0) && left_item != value &&
      right_item->cmp_context == field->cmp_context &&
      (left_item->result_type() != STRING_RESULT ||
       value->result_type() != STRING_RESULT ||
       left_item->collation.collation == value->collation.collation))
  {
    Item *tmp=value->clone_item();
    tmp->collation.set(right_item->collation);

    if (tmp)
    {
      session->change_item_tree(args + 1, tmp);
      func->update_used_tables();
      if ((functype == Item_func::EQ_FUNC || functype == Item_func::EQUAL_FUNC)
	  && and_father != cond && !left_item->const_item())
      {
	cond->marker=1;
	COND_CMP *tmp2;
	if ((tmp2=new COND_CMP(and_father,func)))
	  save_list->push_back(tmp2);
      }
      func->set_cmp_func();
    }
  }
  else if (left_item->eq(field,0) && right_item != value &&
           left_item->cmp_context == field->cmp_context &&
           (right_item->result_type() != STRING_RESULT ||
            value->result_type() != STRING_RESULT ||
            right_item->collation.collation == value->collation.collation))
  {
    Item *tmp= value->clone_item();
    tmp->collation.set(left_item->collation);

    if (tmp)
    {
      session->change_item_tree(args, tmp);
      value= tmp;
      func->update_used_tables();
      if ((functype == Item_func::EQ_FUNC || functype == Item_func::EQUAL_FUNC)
	  && and_father != cond && !right_item->const_item())
      {
        args[0]= args[1];                       // For easy check
        session->change_item_tree(args + 1, value);
	cond->marker=1;
	COND_CMP *tmp2;
	if ((tmp2=new COND_CMP(and_father,func)))
	  save_list->push_back(tmp2);
      }
      func->set_cmp_func();
    }
  }
}

/**
  Remove additional condition inserted by IN/ALL/ANY transformation.

  @param conds   condition for processing

  @return
    new conditions
*/

static Item *remove_additional_cond(Item* conds)
{
  if (conds->name == in_additional_cond)
    return 0;
  if (conds->type() == Item::COND_ITEM)
  {
    Item_cond *cnd= (Item_cond*) conds;
    List_iterator<Item> li(*(cnd->argument_list()));
    Item *item;
    while ((item= li++))
    {
      if (item->name == in_additional_cond)
      {
	li.remove();
	if (cnd->argument_list()->elements == 1)
	  return cnd->argument_list()->head();
	return conds;
      }
    }
  }
  return conds;
}

static void
propagate_cond_constants(Session *session, I_List<COND_CMP> *save_list,
                         COND *and_father, COND *cond)
{
  if (cond->type() == Item::COND_ITEM)
  {
    bool and_level= ((Item_cond*) cond)->functype() ==
      Item_func::COND_AND_FUNC;
    List_iterator_fast<Item> li(*((Item_cond*) cond)->argument_list());
    Item *item;
    I_List<COND_CMP> save;
    while ((item=li++))
    {
      propagate_cond_constants(session, &save,and_level ? cond : item, item);
    }
    if (and_level)
    {						// Handle other found items
      I_List_iterator<COND_CMP> cond_itr(save);
      COND_CMP *cond_cmp;
      while ((cond_cmp=cond_itr++))
      {
        Item **args= cond_cmp->cmp_func->arguments();
        if (!args[0]->const_item())
          change_cond_ref_to_const(session, &save,cond_cmp->and_level,
                                   cond_cmp->and_level, args[0], args[1]);
      }
    }
  }
  else if (and_father != cond && !cond->marker)		// In a AND group
  {
    if (cond->type() == Item::FUNC_ITEM &&
	(((Item_func*) cond)->functype() == Item_func::EQ_FUNC ||
	 ((Item_func*) cond)->functype() == Item_func::EQUAL_FUNC))
    {
      Item_func_eq *func=(Item_func_eq*) cond;
      Item **args= func->arguments();
      bool left_const= args[0]->const_item();
      bool right_const= args[1]->const_item();
      if (!(left_const && right_const) &&
          args[0]->result_type() == args[1]->result_type())
      {
	if (right_const)
	{
          resolve_const_item(session, &args[1], args[0]);
	  func->update_used_tables();
          change_cond_ref_to_const(session, save_list, and_father, and_father,
                                   args[0], args[1]);
	}
	else if (left_const)
	{
          resolve_const_item(session, &args[0], args[1]);
	  func->update_used_tables();
          change_cond_ref_to_const(session, save_list, and_father, and_father,
                                   args[1], args[0]);
	}
      }
    }
  }
}


/**
  Simplify joins replacing outer joins by inner joins whenever it's
  possible.

    The function, during a retrieval of join_list,  eliminates those
    outer joins that can be converted into inner join, possibly nested.
    It also moves the on expressions for the converted outer joins
    and from inner joins to conds.
    The function also calculates some attributes for nested joins:
    - used_tables
    - not_null_tables
    - dep_tables.
    - on_expr_dep_tables
    The first two attributes are used to test whether an outer join can
    be substituted for an inner join. The third attribute represents the
    relation 'to be dependent on' for tables. If table t2 is dependent
    on table t1, then in any evaluated execution plan table access to
    table t2 must precede access to table t2. This relation is used also
    to check whether the query contains  invalid cross-references.
    The forth attribute is an auxiliary one and is used to calculate
    dep_tables.
    As the attribute dep_tables qualifies possibles orders of tables in the
    execution plan, the dependencies required by the straight join
    modifiers are reflected in this attribute as well.
    The function also removes all braces that can be removed from the join
    expression without changing its meaning.

  @note
    An outer join can be replaced by an inner join if the where condition
    or the on expression for an embedding nested join contains a conjunctive
    predicate rejecting null values for some attribute of the inner tables.

    E.g. in the query:
    @code
      SELECT * FROM t1 LEFT JOIN t2 ON t2.a=t1.a WHERE t2.b < 5
    @endcode
    the predicate t2.b < 5 rejects nulls.
    The query is converted first to:
    @code
      SELECT * FROM t1 INNER JOIN t2 ON t2.a=t1.a WHERE t2.b < 5
    @endcode
    then to the equivalent form:
    @code
      SELECT * FROM t1, t2 ON t2.a=t1.a WHERE t2.b < 5 AND t2.a=t1.a
    @endcode


    Similarly the following query:
    @code
      SELECT * from t1 LEFT JOIN (t2, t3) ON t2.a=t1.a t3.b=t1.b
        WHERE t2.c < 5
    @endcode
    is converted to:
    @code
      SELECT * FROM t1, (t2, t3) WHERE t2.c < 5 AND t2.a=t1.a t3.b=t1.b

    @endcode

    One conversion might trigger another:
    @code
      SELECT * FROM t1 LEFT JOIN t2 ON t2.a=t1.a
                       LEFT JOIN t3 ON t3.b=t2.b
        WHERE t3 IS NOT NULL =>
      SELECT * FROM t1 LEFT JOIN t2 ON t2.a=t1.a, t3
        WHERE t3 IS NOT NULL AND t3.b=t2.b =>
      SELECT * FROM t1, t2, t3
        WHERE t3 IS NOT NULL AND t3.b=t2.b AND t2.a=t1.a
  @endcode

    The function removes all unnecessary braces from the expression
    produced by the conversions.
    E.g.
    @code
      SELECT * FROM t1, (t2, t3) WHERE t2.c < 5 AND t2.a=t1.a AND t3.b=t1.b
    @endcode
    finally is converted to:
    @code
      SELECT * FROM t1, t2, t3 WHERE t2.c < 5 AND t2.a=t1.a AND t3.b=t1.b

    @endcode


    It also will remove braces from the following queries:
    @code
      SELECT * from (t1 LEFT JOIN t2 ON t2.a=t1.a) LEFT JOIN t3 ON t3.b=t2.b
      SELECT * from (t1, (t2,t3)) WHERE t1.a=t2.a AND t2.b=t3.b.
    @endcode

    The benefit of this simplification procedure is that it might return
    a query for which the optimizer can evaluate execution plan with more
    join orders. With a left join operation the optimizer does not
    consider any plan where one of the inner tables is before some of outer
    tables.

  IMPLEMENTATION
    The function is implemented by a recursive procedure.  On the recursive
    ascent all attributes are calculated, all outer joins that can be
    converted are replaced and then all unnecessary braces are removed.
    As join list contains join tables in the reverse order sequential
    elimination of outer joins does not require extra recursive calls.

  SEMI-JOIN NOTES
    Remove all semi-joins that have are within another semi-join (i.e. have
    an "ancestor" semi-join nest)

  EXAMPLES
    Here is an example of a join query with invalid cross references:
    @code
      SELECT * FROM t1 LEFT JOIN t2 ON t2.a=t3.a LEFT JOIN t3 ON t3.b=t1.b
    @endcode

  @param join        reference to the query info
  @param join_list   list representation of the join to be converted
  @param conds       conditions to add on expressions for converted joins
  @param top         true <=> conds is the where condition

  @return
    - The new condition, if success
    - 0, otherwise
*/

static COND *
simplify_joins(JOIN *join, List<TableList> *join_list, COND *conds, bool top,
               bool in_sj)
{
  TableList *table;
  nested_join_st *nested_join;
  TableList *prev_table= 0;
  List_iterator<TableList> li(*join_list);

  /*
    Try to simplify join operations from join_list.
    The most outer join operation is checked for conversion first.
  */
  while ((table= li++))
  {
    table_map used_tables;
    table_map not_null_tables= (table_map) 0;

    if ((nested_join= table->nested_join))
    {
      /*
         If the element of join_list is a nested join apply
         the procedure to its nested join list first.
      */
      if (table->on_expr)
      {
        Item *expr= table->on_expr;
        /*
           If an on expression E is attached to the table,
           check all null rejected predicates in this expression.
           If such a predicate over an attribute belonging to
           an inner table of an embedded outer join is found,
           the outer join is converted to an inner join and
           the corresponding on expression is added to E.
	*/
        expr= simplify_joins(join, &nested_join->join_list,
                             expr, false, in_sj || table->sj_on_expr);

        if (!table->prep_on_expr || expr != table->on_expr)
        {
          assert(expr);

          table->on_expr= expr;
          table->prep_on_expr= expr->copy_andor_structure(join->session);
        }
      }
      nested_join->used_tables= (table_map) 0;
      nested_join->not_null_tables=(table_map) 0;
      conds= simplify_joins(join, &nested_join->join_list, conds, top,
                            in_sj || table->sj_on_expr);
      used_tables= nested_join->used_tables;
      not_null_tables= nested_join->not_null_tables;
    }
    else
    {
      if (!table->prep_on_expr)
        table->prep_on_expr= table->on_expr;
      used_tables= table->table->map;
      if (conds)
        not_null_tables= conds->not_null_tables();
    }

    if (table->embedding)
    {
      table->embedding->nested_join->used_tables|= used_tables;
      table->embedding->nested_join->not_null_tables|= not_null_tables;
    }

    if (!table->outer_join || (used_tables & not_null_tables))
    {
      /*
        For some of the inner tables there are conjunctive predicates
        that reject nulls => the outer join can be replaced by an inner join.
      */
      table->outer_join= 0;
      if (table->on_expr)
      {
        /* Add ON expression to the WHERE or upper-level ON condition. */
        if (conds)
        {
          conds= and_conds(conds, table->on_expr);
          conds->top_level_item();
          /* conds is always a new item as both cond and on_expr existed */
          assert(!conds->fixed);
          conds->fix_fields(join->session, &conds);
        }
        else
          conds= table->on_expr;
        table->prep_on_expr= table->on_expr= 0;
      }
    }

    if (!top)
      continue;

    /*
      Only inner tables of non-convertible outer joins
      remain with on_expr.
    */
    if (table->on_expr)
    {
      table->dep_tables|= table->on_expr->used_tables();
      if (table->embedding)
      {
        table->dep_tables&= ~table->embedding->nested_join->used_tables;
        /*
           Embedding table depends on tables used
           in embedded on expressions.
        */
        table->embedding->on_expr_dep_tables|= table->on_expr->used_tables();
      }
      else
        table->dep_tables&= ~table->table->map;
    }

    if (prev_table)
    {
      /* The order of tables is reverse: prev_table follows table */
      if (prev_table->straight)
        prev_table->dep_tables|= used_tables;
      if (prev_table->on_expr)
      {
        prev_table->dep_tables|= table->on_expr_dep_tables;
        table_map prev_used_tables= prev_table->nested_join ?
	                            prev_table->nested_join->used_tables :
	                            prev_table->table->map;
        /*
          If on expression contains only references to inner tables
          we still make the inner tables dependent on the outer tables.
          It would be enough to set dependency only on one outer table
          for them. Yet this is really a rare case.
	*/
        if (!(prev_table->on_expr->used_tables() & ~prev_used_tables))
          prev_table->dep_tables|= used_tables;
      }
    }
    prev_table= table;
  }

  /*
    Flatten nested joins that can be flattened.
    no ON expression and not a semi-join => can be flattened.
  */
  li.rewind();
  while ((table= li++))
  {
    nested_join= table->nested_join;
    if (table->sj_on_expr && !in_sj)
    {
       /*
         If this is a semi-join that is not contained within another semi-join,
         leave it intact (otherwise it is flattened)
       */
      join->select_lex->sj_nests.push_back(table);
    }
    else if (nested_join && !table->on_expr)
    {
      TableList *tbl;
      List_iterator<TableList> it(nested_join->join_list);
      while ((tbl= it++))
      {
        tbl->embedding= table->embedding;
        tbl->join_list= table->join_list;
      }
      li.replace(nested_join->join_list);
    }
  }
  return(conds);
}


/**
  Assign each nested join structure a bit in nested_join_map.

    Assign each nested join structure (except "confluent" ones - those that
    embed only one element) a bit in nested_join_map.

  @param join          Join being processed
  @param join_list     List of tables
  @param first_unused  Number of first unused bit in nested_join_map before the
                       call

  @note
    This function is called after simplify_joins(), when there are no
    redundant nested joins, #non_confluent_nested_joins <= #tables_in_join so
    we will not run out of bits in nested_join_map.

  @return
    First unused bit in nested_join_map after the call.
*/

static uint32_t build_bitmap_for_nested_joins(List<TableList> *join_list,
                                          uint32_t first_unused)
{
  List_iterator<TableList> li(*join_list);
  TableList *table;
  while ((table= li++))
  {
    nested_join_st *nested_join;
    if ((nested_join= table->nested_join))
    {
      /*
        It is guaranteed by simplify_joins() function that a nested join
        that has only one child is either
         - a single-table view (the child is the underlying table), or
         - a single-table semi-join nest

        We don't assign bits to such sj-nests because
        1. it is redundant (a "sequence" of one table cannot be interleaved
            with anything)
        2. we could run out bits in nested_join_map otherwise.
      */
      if (nested_join->join_list.elements != 1)
      {
        /* Don't assign bits to sj-nests */
        if (table->on_expr)
          nested_join->nj_map= (nested_join_map) 1 << first_unused++;
        first_unused= build_bitmap_for_nested_joins(&nested_join->join_list,
                                                    first_unused);
      }
    }
  }
  return(first_unused);
}


/**
  Set nested_join_st::counter=0 in all nested joins in passed list.

    Recursively set nested_join_st::counter=0 for all nested joins contained in
    the passed join_list.

  @param join_list  List of nested joins to process. It may also contain base
                    tables which will be ignored.
*/

static void reset_nj_counters(List<TableList> *join_list)
{
  List_iterator<TableList> li(*join_list);
  TableList *table;
  while ((table= li++))
  {
    nested_join_st *nested_join;
    if ((nested_join= table->nested_join))
    {
      nested_join->counter_= 0;
      reset_nj_counters(&nested_join->join_list);
    }
  }
  return;
}


/**
  Check interleaving with an inner tables of an outer join for
  extension table.

    Check if table next_tab can be added to current partial join order, and
    if yes, record that it has been added.

    The function assumes that both current partial join order and its
    extension with next_tab are valid wrt table dependencies.

  @verbatim
     IMPLEMENTATION
       LIMITATIONS ON JOIN order_st
         The nested [outer] joins executioner algorithm imposes these limitations
         on join order:
         1. "Outer tables first" -  any "outer" table must be before any
             corresponding "inner" table.
         2. "No interleaving" - tables inside a nested join must form a continuous
            sequence in join order (i.e. the sequence must not be interrupted by
            tables that are outside of this nested join).

         #1 is checked elsewhere, this function checks #2 provided that #1 has
         been already checked.

       WHY NEED NON-INTERLEAVING
         Consider an example:

           select * from t0 join t1 left join (t2 join t3) on cond1

         The join order "t1 t2 t0 t3" is invalid:

         table t0 is outside of the nested join, so WHERE condition for t0 is
         attached directly to t0 (without triggers, and it may be used to access
         t0). Applying WHERE(t0) to (t2,t0,t3) record is invalid as we may miss
         combinations of (t1, t2, t3) that satisfy condition cond1, and produce a
         null-complemented (t1, t2.NULLs, t3.NULLs) row, which should not have
         been produced.

         If table t0 is not between t2 and t3, the problem doesn't exist:
          If t0 is located after (t2,t3), WHERE(t0) is applied after nested join
           processing has finished.
          If t0 is located before (t2,t3), predicates like WHERE_cond(t0, t2) are
           wrapped into condition triggers, which takes care of correct nested
           join processing.

       HOW IT IS IMPLEMENTED
         The limitations on join order can be rephrased as follows: for valid
         join order one must be able to:
           1. write down the used tables in the join order on one line.
           2. for each nested join, put one '(' and one ')' on the said line
           3. write "LEFT JOIN" and "ON (...)" where appropriate
           4. get a query equivalent to the query we're trying to execute.

         Calls to check_interleaving_with_nj() are equivalent to writing the
         above described line from left to right.
         A single check_interleaving_with_nj(A,B) call is equivalent to writing
         table B and appropriate brackets on condition that table A and
         appropriate brackets is the last what was written. Graphically the
         transition is as follows:

                              +---- current position
                              |
             ... last_tab ))) | ( next_tab )  )..) | ...
                                X          Y   Z   |
                                                   +- need to move to this
                                                      position.

         Notes about the position:
           The caller guarantees that there is no more then one X-bracket by
           checking "!(remaining_tables & s->dependent)" before calling this
           function. X-bracket may have a pair in Y-bracket.

         When "writing" we store/update this auxilary info about the current
         position:
          1. join->cur_embedding_map - bitmap of pairs of brackets (aka nested
             joins) we've opened but didn't close.
          2. {each nested_join_st structure not simplified away}->counter - number
             of this nested join's children that have already been added to to
             the partial join order.
  @endverbatim

  @param join       Join being processed
  @param last_tab   Last table in current partial join order (this function is
                    not called for empty partial join orders)
  @param next_tab   Table we're going to extend the current partial join with

  @retval
    false  Join order extended, nested joins info about current join
    order (see NOTE section) updated.
  @retval
    true   Requested join order extension not allowed.
*/

static bool check_interleaving_with_nj(JOIN_TAB *last_tab, JOIN_TAB *next_tab)
{
  TableList *next_emb= next_tab->table->pos_in_table_list->embedding;
  JOIN *join= last_tab->join;

  if (join->cur_embedding_map & ~next_tab->embedding_map)
  {
    /*
      next_tab is outside of the "pair of brackets" we're currently in.
      Cannot add it.
    */
    return true;
  }

  /*
    Do update counters for "pairs of brackets" that we've left (marked as
    X,Y,Z in the above picture)
  */
  for (;next_emb; next_emb= next_emb->embedding)
  {
    next_emb->nested_join->counter_++;
    if (next_emb->nested_join->counter_ == 1)
    {
      /*
        next_emb is the first table inside a nested join we've "entered". In
        the picture above, we're looking at the 'X' bracket. Don't exit yet as
        X bracket might have Y pair bracket.
      */
      join->cur_embedding_map |= next_emb->nested_join->nj_map;
    }

    if (next_emb->nested_join->join_list.elements !=
        next_emb->nested_join->counter_)
      break;

    /*
      We're currently at Y or Z-bracket as depicted in the above picture.
      Mark that we've left it and continue walking up the brackets hierarchy.
    */
    join->cur_embedding_map &= ~next_emb->nested_join->nj_map;
  }
  return false;
}


/**
  Nested joins perspective: Remove the last table from the join order.

    Remove the last table from the partial join order and update the nested
    joins counters and join->cur_embedding_map. It is ok to call this
    function for the first table in join order (for which
    check_interleaving_with_nj has not been called)

  @param last  join table to remove, it is assumed to be the last in current
               partial join order.
*/

static void restore_prev_nj_state(JOIN_TAB *last)
{
  TableList *last_emb= last->table->pos_in_table_list->embedding;
  JOIN *join= last->join;
  while (last_emb)
  {
    if (last_emb->on_expr)
    {
      if (!(--last_emb->nested_join->counter_))
        join->cur_embedding_map&= ~last_emb->nested_join->nj_map;
      else if (last_emb->nested_join->join_list.elements-1 ==
               last_emb->nested_join->counter_)
        join->cur_embedding_map|= last_emb->nested_join->nj_map;
      else
        break;
    }
    last_emb= last_emb->embedding;
  }
}



static
void advance_sj_state(const table_map remaining_tables, const JOIN_TAB *tab)
{
  TableList *emb_sj_nest;
  if ((emb_sj_nest= tab->emb_sj_nest))
  {
    tab->join->cur_emb_sj_nests |= emb_sj_nest->sj_inner_tables;
    /* Remove the sj_nest if all of its SJ-inner tables are in cur_table_map */
    if (!(remaining_tables & emb_sj_nest->sj_inner_tables))
      tab->join->cur_emb_sj_nests &= ~emb_sj_nest->sj_inner_tables;
  }
}


/*
  we assume remaining_tables doesnt contain @tab.
*/

static void restore_prev_sj_state(const table_map remaining_tables,
                                  const JOIN_TAB *tab)
{
  TableList *emb_sj_nest;
  if ((emb_sj_nest= tab->emb_sj_nest))
  {
    /* If we're removing the last SJ-inner table, remove the sj-nest */
    if ((remaining_tables & emb_sj_nest->sj_inner_tables) ==
        (emb_sj_nest->sj_inner_tables & ~tab->table->map))
    {
      tab->join->cur_emb_sj_nests &= ~emb_sj_nest->sj_inner_tables;
    }
  }
}


static COND *
optimize_cond(JOIN *join, COND *conds, List<TableList> *join_list,
              Item::cond_result *cond_value)
{
  Session *session= join->session;

  if (!conds)
    *cond_value= Item::COND_TRUE;
  else
  {
    /*
      Build all multiple equality predicates and eliminate equality
      predicates that can be inferred from these multiple equalities.
      For each reference of a field included into a multiple equality
      that occurs in a function set a pointer to the multiple equality
      predicate. Substitute a constant instead of this field if the
      multiple equality contains a constant.
    */
    conds= build_equal_items(join->session, conds, NULL, join_list,
                             &join->cond_equal);

    /* change field = field to field = const for each found field = const */
    propagate_cond_constants(session, (I_List<COND_CMP> *) 0, conds, conds);
    /*
      Remove all instances of item == item
      Remove all and-levels where CONST item != CONST item
    */
    conds= remove_eq_conds(session, conds, cond_value) ;
  }
  return(conds);
}


/**
  Remove const and eq items.

  @return
    Return new item, or NULL if no condition @n
    cond_value is set to according:
    - COND_OK     : query is possible (field = constant)
    - COND_TRUE   : always true	( 1 = 1 )
    - COND_FALSE  : always false	( 1 = 2 )
*/

COND *
remove_eq_conds(Session *session, COND *cond, Item::cond_result *cond_value)
{
  if (cond->type() == Item::COND_ITEM)
  {
    bool and_level= (((Item_cond*) cond)->functype() == Item_func::COND_AND_FUNC);

    List_iterator<Item> li(*((Item_cond*) cond)->argument_list());
    Item::cond_result tmp_cond_value;
    bool should_fix_fields= false;

    *cond_value= Item::COND_UNDEF;
    Item *item;
    while ((item= li++))
    {
      Item *new_item= remove_eq_conds(session, item, &tmp_cond_value);
      if (! new_item)
	      li.remove();
      else if (item != new_item)
      {
        li.replace(new_item);
        should_fix_fields= true;
      }
      if (*cond_value == Item::COND_UNDEF)
	      *cond_value= tmp_cond_value;

      switch (tmp_cond_value) 
      {
        case Item::COND_OK:			/* Not true or false */
          if (and_level || (*cond_value == Item::COND_FALSE))
            *cond_value= tmp_cond_value;
          break;
        case Item::COND_FALSE:
          if (and_level)
          {
            *cond_value= tmp_cond_value;
            return (COND *) NULL;			/* Always false */
          }
          break;
        case Item::COND_TRUE:
          if (! and_level)
          {
            *cond_value= tmp_cond_value;
            return (COND *) NULL;			/* Always true */
          }
          break;
        case Item::COND_UNDEF:			/* Impossible */
          break; /* purecov: deadcode */
      }
    }

    if (should_fix_fields)
      cond->update_used_tables();

    if (! ((Item_cond*) cond)->argument_list()->elements || *cond_value != Item::COND_OK)
      return (COND*) NULL;

    if (((Item_cond*) cond)->argument_list()->elements == 1)
    {						
      /* Argument list contains only one element, so reduce it so a single item, then remove list */
      item= ((Item_cond*) cond)->argument_list()->head();
      ((Item_cond*) cond)->argument_list()->empty();
      return item;
    }
  }
  else if (cond->type() == Item::FUNC_ITEM && ((Item_func*) cond)->functype() == Item_func::ISNULL_FUNC)
  {
    /*
      Handles this special case for some ODBC applications:
      The are requesting the row that was just updated with a auto_increment
      value with this construct:

      SELECT * from table_name where auto_increment_column IS NULL
      This will be changed to:
      SELECT * from table_name where auto_increment_column = LAST_INSERT_ID
    */

    Item_func_isnull *func= (Item_func_isnull*) cond;
    Item **args= func->arguments();
    if (args[0]->type() == Item::FIELD_ITEM)
    {
      Field *field= ((Item_field*) args[0])->field;
      if (field->flags & AUTO_INCREMENT_FLAG 
          && ! field->table->maybe_null 
          && session->options & OPTION_AUTO_IS_NULL
          && (
            session->first_successful_insert_id_in_prev_stmt > 0 
            && session->substitute_null_with_insert_id
            )
          )
      {
        COND *new_cond;
        if ((new_cond= new Item_func_eq(args[0], new Item_int("last_insert_id()",
                                                          session->read_first_successful_insert_id_in_prev_stmt(),
                                                          MY_INT64_NUM_DECIMAL_DIGITS))))
        {
          cond= new_cond;
          /*
            Item_func_eq can't be fixed after creation so we do not check
            cond->fixed, also it do not need tables so we use 0 as second
            argument.
          */
          cond->fix_fields(session, &cond);
        }
        /*
          IS NULL should be mapped to LAST_INSERT_ID only for first row, so
          clear for next row
        */
        session->substitute_null_with_insert_id= false;
      }
#ifdef NOTDEFINED
      /* fix to replace 'NULL' dates with '0' (shreeve@uci.edu) */
      else if (
          ((field->type() == DRIZZLE_TYPE_DATE) || (field->type() == DRIZZLE_TYPE_DATETIME)) 
          && (field->flags & NOT_NULL_FLAG) 
          && ! field->table->maybe_null)
      {
        COND *new_cond;
        if ((new_cond= new Item_func_eq(args[0],new Item_int("0", 0, 2))))
        {
          cond= new_cond;
          /*
            Item_func_eq can't be fixed after creation so we do not check
            cond->fixed, also it do not need tables so we use 0 as second
            argument.
          */
          cond->fix_fields(session, &cond);
        }
      }
#endif /* NOTDEFINED */
    }
    if (cond->const_item())
    {
      *cond_value= eval_const_cond(cond) ? Item::COND_TRUE : Item::COND_FALSE;
      return (COND *) NULL;
    }
  }
  else if (cond->const_item() && !cond->is_expensive())
  /*
    TODO:
    Excluding all expensive functions is too restritive we should exclude only
    materialized IN because it is created later than this phase, and cannot be
    evaluated at this point.
    The condition should be something as (need to fix member access):
      !(cond->type() == Item::FUNC_ITEM &&
        ((Item_func*)cond)->func_name() == "<in_optimizer>" &&
        ((Item_in_optimizer*)cond)->is_expensive()))
  */
  {
    *cond_value= eval_const_cond(cond) ? Item::COND_TRUE : Item::COND_FALSE;
    return (COND *) NULL;
  }
  else if ((*cond_value= cond->eq_cmp_result()) != Item::COND_OK)
  {						
    /* boolan compare function */
    Item *left_item=	((Item_func*) cond)->arguments()[0];
    Item *right_item= ((Item_func*) cond)->arguments()[1];
    if (left_item->eq(right_item,1))
    {
      if (!left_item->maybe_null || ((Item_func*) cond)->functype() == Item_func::EQUAL_FUNC)
	      return (COND*) NULL;			/* Comparison of identical items */
    }
  }
  *cond_value= Item::COND_OK;
  return cond;					/* Point at next and return into recursion */
}

/*
  Check if equality can be used in removing components of GROUP BY/DISTINCT

  SYNOPSIS
    test_if_equality_guarantees_uniqueness()
      l          the left comparison argument (a field if any)
      r          the right comparison argument (a const of any)

  DESCRIPTION
    Checks if an equality predicate can be used to take away
    DISTINCT/GROUP BY because it is known to be true for exactly one
    distinct value (e.g. <expr> == <const>).
    Arguments must be of the same type because e.g.
    <string_field> = <int_const> may match more than 1 distinct value from
    the column.
    We must take into consideration and the optimization done for various
    string constants when compared to dates etc (see Item_int_with_ref) as
    well as the collation of the arguments.

  RETURN VALUE
    true    can be used
    false   cannot be used
*/
static bool
test_if_equality_guarantees_uniqueness(Item *l, Item *r)
{
  return r->const_item() &&
    /* elements must be compared as dates */
     (Arg_comparator::can_compare_as_dates(l, r, 0) ||
      /* or of the same result type */
      (r->result_type() == l->result_type() &&
       /* and must have the same collation if compared as strings */
       (l->result_type() != STRING_RESULT ||
        l->collation.collation == r->collation.collation)));
}

/**
  Return true if the item is a const value in all the WHERE clause.
*/

static bool
const_expression_in_where(COND *cond, Item *comp_item, Item **const_item)
{
  if (cond->type() == Item::COND_ITEM)
  {
    bool and_level= (((Item_cond*) cond)->functype()
		     == Item_func::COND_AND_FUNC);
    List_iterator_fast<Item> li(*((Item_cond*) cond)->argument_list());
    Item *item;
    while ((item=li++))
    {
      bool res=const_expression_in_where(item, comp_item, const_item);
      if (res)					// Is a const value
      {
	if (and_level)
	  return 1;
      }
      else if (!and_level)
	return 0;
    }
    return and_level ? 0 : 1;
  }
  else if (cond->eq_cmp_result() != Item::COND_OK)
  {						// boolan compare function
    Item_func* func= (Item_func*) cond;
    if (func->functype() != Item_func::EQUAL_FUNC &&
	func->functype() != Item_func::EQ_FUNC)
      return 0;
    Item *left_item=	((Item_func*) cond)->arguments()[0];
    Item *right_item= ((Item_func*) cond)->arguments()[1];
    if (left_item->eq(comp_item,1))
    {
      if (test_if_equality_guarantees_uniqueness (left_item, right_item))
      {
	if (*const_item)
	  return right_item->eq(*const_item, 1);
	*const_item=right_item;
	return 1;
      }
    }
    else if (right_item->eq(comp_item,1))
    {
      if (test_if_equality_guarantees_uniqueness (right_item, left_item))
      {
	if (*const_item)
	  return left_item->eq(*const_item, 1);
	*const_item=left_item;
	return 1;
      }
    }
  }
  return 0;
}


/**
  @details
  Rows produced by a join sweep may end up in a temporary table or be sent
  to a client. Setup the function of the nested loop join algorithm which
  handles final fully constructed and matched records.

  @param join   join to setup the function for.

  @return
    end_select function to use. This function can't fail.
*/

Next_select_func setup_end_select_func(JOIN *join)
{
  Table *table= join->tmp_table;
  Tmp_Table_Param *tmp_tbl= &join->tmp_table_param;
  Next_select_func end_select;

  /* Set up select_end */
  if (table)
  {
    if (table->group && tmp_tbl->sum_func_count &&
        !tmp_tbl->precomputed_group_by)
    {
      if (table->s->keys)
      {
	end_select=end_update;
      }
      else
      {
	end_select=end_unique_update;
      }
    }
    else if (join->sort_and_group && !tmp_tbl->precomputed_group_by)
    {
      end_select=end_write_group;
    }
    else
    {
      end_select=end_write;
      if (tmp_tbl->precomputed_group_by)
      {
        /*
          A preceding call to create_tmp_table in the case when loose
          index scan is used guarantees that
          Tmp_Table_Param::items_to_copy has enough space for the group
          by functions. It is OK here to use memcpy since we copy
          Item_sum pointers into an array of Item pointers.
        */
        memcpy(tmp_tbl->items_to_copy + tmp_tbl->func_count,
               join->sum_funcs,
               sizeof(Item*)*tmp_tbl->sum_func_count);
        tmp_tbl->items_to_copy[tmp_tbl->func_count+tmp_tbl->sum_func_count]= 0;
      }
    }
  }
  else
  {
    if ((join->sort_and_group) &&
        !tmp_tbl->precomputed_group_by)
      end_select= end_send_group;
    else
      end_select= end_send;
  }
  return end_select;
}


/**
  Make a join of all tables and write it on socket or to table.

  @retval
    0  if ok
  @retval
    1  if error is sent
  @retval
    -1  if error should be sent
*/

static int
do_select(JOIN *join,List<Item> *fields,Table *table)
{
  int rc= 0;
  enum_nested_loop_state error= NESTED_LOOP_OK;
  JOIN_TAB *join_tab= NULL;

  join->tmp_table= table;			/* Save for easy recursion */
  join->fields= fields;

  if (table)
  {
    table->file->extra(HA_EXTRA_WRITE_CACHE);
    table->emptyRecord();
    if (table->group && join->tmp_table_param.sum_func_count &&
        table->s->keys && !table->file->inited)
      table->file->ha_index_init(0, 0);
  }
  /* Set up select_end */
  Next_select_func end_select= setup_end_select_func(join);
  if (join->tables)
  {
    join->join_tab[join->tables-1].next_select= end_select;

    join_tab=join->join_tab+join->const_tables;
  }
  join->send_records=0;
  if (join->tables == join->const_tables)
  {
    /*
      HAVING will be checked after processing aggregate functions,
      But WHERE should checkd here (we alredy have read tables)
    */
    if (!join->conds || join->conds->val_int())
    {
      error= (*end_select)(join, 0, 0);
      if (error == NESTED_LOOP_OK || error == NESTED_LOOP_QUERY_LIMIT)
	error= (*end_select)(join, 0, 1);

      /*
        If we don't go through evaluate_join_record(), do the counting
        here.  join->send_records is increased on success in end_send(),
        so we don't touch it here.
      */
      join->examined_rows++;
      join->session->row_count++;
      assert(join->examined_rows <= 1);
    }
    else if (join->send_row_on_empty_set())
    {
      List<Item> *columns_list= fields;
      rc= join->result->send_data(*columns_list);
    }
  }
  else
  {
    assert(join->tables);
    error= sub_select(join,join_tab,0);
    if (error == NESTED_LOOP_OK || error == NESTED_LOOP_NO_MORE_ROWS)
      error= sub_select(join,join_tab,1);
    if (error == NESTED_LOOP_QUERY_LIMIT)
      error= NESTED_LOOP_OK;                    /* select_limit used */
  }
  if (error == NESTED_LOOP_NO_MORE_ROWS)
    error= NESTED_LOOP_OK;

  if (error == NESTED_LOOP_OK)
  {
    /*
      Sic: this branch works even if rc != 0, e.g. when
      send_data above returns an error.
    */
    if (!table)					// If sending data to client
    {
      /*
	The following will unlock all cursors if the command wasn't an
	update command
      */
      join->join_free();			// Unlock all cursors
      if (join->result->send_eof())
	rc= 1;                                  // Don't send error
    }
  }
  else
    rc= -1;
  if (table)
  {
    int tmp, new_errno= 0;
    if ((tmp=table->file->extra(HA_EXTRA_NO_CACHE)))
    {
      new_errno= tmp;
    }
    if ((tmp=table->file->ha_index_or_rnd_end()))
    {
      new_errno= tmp;
    }
    if (new_errno)
      table->file->print_error(new_errno,MYF(0));
  }
  return(join->session->is_error() ? -1 : rc);
}


enum_nested_loop_state
sub_select_cache(JOIN *join,JOIN_TAB *join_tab,bool end_of_records)
{
  enum_nested_loop_state rc;

  if (end_of_records)
  {
    rc= flush_cached_records(join,join_tab,false);
    if (rc == NESTED_LOOP_OK || rc == NESTED_LOOP_NO_MORE_ROWS)
      rc= sub_select(join,join_tab,end_of_records);
    return rc;
  }
  if (join->session->killed)		// If aborted by user
  {
    join->session->send_kill_message();
    return NESTED_LOOP_KILLED;                   /* purecov: inspected */
  }
  if (join_tab->use_quick != 2 || test_if_quick_select(join_tab) <= 0)
  {
    if (!store_record_in_cache(&join_tab->cache))
      return NESTED_LOOP_OK;                     // There is more room in cache
    return flush_cached_records(join,join_tab,false);
  }
  rc= flush_cached_records(join, join_tab, true);
  if (rc == NESTED_LOOP_OK || rc == NESTED_LOOP_NO_MORE_ROWS)
    rc= sub_select(join, join_tab, end_of_records);
  return rc;
}

/**
  Retrieve records ends with a given beginning from the result of a join.

    For a given partial join record consisting of records from the tables
    preceding the table join_tab in the execution plan, the function
    retrieves all matching full records from the result set and
    send them to the result set stream.

  @note
    The function effectively implements the  final (n-k) nested loops
    of nested loops join algorithm, where k is the ordinal number of
    the join_tab table and n is the total number of tables in the join query.
    It performs nested loops joins with all conjunctive predicates from
    the where condition pushed as low to the tables as possible.
    E.g. for the query
    @code
      SELECT * FROM t1,t2,t3
      WHERE t1.a=t2.a AND t2.b=t3.b AND t1.a BETWEEN 5 AND 9
    @endcode
    the predicate (t1.a BETWEEN 5 AND 9) will be pushed to table t1,
    given the selected plan prescribes to nest retrievals of the
    joined tables in the following order: t1,t2,t3.
    A pushed down predicate are attached to the table which it pushed to,
    at the field join_tab->select_cond.
    When executing a nested loop of level k the function runs through
    the rows of 'join_tab' and for each row checks the pushed condition
    attached to the table.
    If it is false the function moves to the next row of the
    table. If the condition is true the function recursively executes (n-k-1)
    remaining embedded nested loops.
    The situation becomes more complicated if outer joins are involved in
    the execution plan. In this case the pushed down predicates can be
    checked only at certain conditions.
    Suppose for the query
    @code
      SELECT * FROM t1 LEFT JOIN (t2,t3) ON t3.a=t1.a
      WHERE t1>2 AND (t2.b>5 OR t2.b IS NULL)
    @endcode
    the optimizer has chosen a plan with the table order t1,t2,t3.
    The predicate P1=t1>2 will be pushed down to the table t1, while the
    predicate P2=(t2.b>5 OR t2.b IS NULL) will be attached to the table
    t2. But the second predicate can not be unconditionally tested right
    after a row from t2 has been read. This can be done only after the
    first row with t3.a=t1.a has been encountered.
    Thus, the second predicate P2 is supplied with a guarded value that are
    stored in the field 'found' of the first inner table for the outer join
    (table t2). When the first row with t3.a=t1.a for the  current row
    of table t1  appears, the value becomes true. For now on the predicate
    is evaluated immediately after the row of table t2 has been read.
    When the first row with t3.a=t1.a has been encountered all
    conditions attached to the inner tables t2,t3 must be evaluated.
    Only when all of them are true the row is sent to the output stream.
    If not, the function returns to the lowest nest level that has a false
    attached condition.
    The predicates from on expressions are also pushed down. If in the
    the above example the on expression were (t3.a=t1.a AND t2.a=t1.a),
    then t1.a=t2.a would be pushed down to table t2, and without any
    guard.
    If after the run through all rows of table t2, the first inner table
    for the outer join operation, it turns out that no matches are
    found for the current row of t1, then current row from table t1
    is complemented by nulls  for t2 and t3. Then the pushed down predicates
    are checked for the composed row almost in the same way as it had
    been done for the first row with a match. The only difference is
    the predicates from on expressions are not checked.

  @par
  @b IMPLEMENTATION
  @par
    The function forms output rows for a current partial join of k
    tables tables recursively.
    For each partial join record ending with a certain row from
    join_tab it calls sub_select that builds all possible matching
    tails from the result set.
    To be able  check predicates conditionally items of the class
    Item_func_trig_cond are employed.
    An object of  this class is constructed from an item of class COND
    and a pointer to a guarding boolean variable.
    When the value of the guard variable is true the value of the object
    is the same as the value of the predicate, otherwise it's just returns
    true.
    To carry out a return to a nested loop level of join table t the pointer
    to t is remembered in the field 'return_tab' of the join structure.
    Consider the following query:
    @code
        SELECT * FROM t1,
                      LEFT JOIN
                      (t2, t3 LEFT JOIN (t4,t5) ON t5.a=t3.a)
                      ON t4.a=t2.a
           WHERE (t2.b=5 OR t2.b IS NULL) AND (t4.b=2 OR t4.b IS NULL)
    @endcode
    Suppose the chosen execution plan dictates the order t1,t2,t3,t4,t5
    and suppose for a given joined rows from tables t1,t2,t3 there are
    no rows in the result set yet.
    When first row from t5 that satisfies the on condition
    t5.a=t3.a is found, the pushed down predicate t4.b=2 OR t4.b IS NULL
    becomes 'activated', as well the predicate t4.a=t2.a. But
    the predicate (t2.b=5 OR t2.b IS NULL) can not be checked until
    t4.a=t2.a becomes true.
    In order not to re-evaluate the predicates that were already evaluated
    as attached pushed down predicates, a pointer to the the first
    most inner unmatched table is maintained in join_tab->first_unmatched.
    Thus, when the first row from t5 with t5.a=t3.a is found
    this pointer for t5 is changed from t4 to t2.

    @par
    @b STRUCTURE @b NOTES
    @par
    join_tab->first_unmatched points always backwards to the first inner
    table of the embedding nested join, if any.

  @param join      pointer to the structure providing all context info for
                   the query
  @param join_tab  the first next table of the execution plan to be retrieved
  @param end_records  true when we need to perform final steps of retrival

  @return
    return one of enum_nested_loop_state, except NESTED_LOOP_NO_MORE_ROWS.
*/
int do_sj_reset(SJ_TMP_TABLE *sj_tbl);

enum_nested_loop_state
sub_select(JOIN *join,JOIN_TAB *join_tab,bool end_of_records)
{
  join_tab->table->null_row=0;
  if (end_of_records)
    return (*join_tab->next_select)(join,join_tab+1,end_of_records);

  int error;
  enum_nested_loop_state rc;
  READ_RECORD *info= &join_tab->read_record;

  if (join_tab->flush_weedout_table)
  {
    do_sj_reset(join_tab->flush_weedout_table);
  }

  if (join->resume_nested_loop)
  {
    /* If not the last table, plunge down the nested loop */
    if (join_tab < join->join_tab + join->tables - 1)
      rc= (*join_tab->next_select)(join, join_tab + 1, 0);
    else
    {
      join->resume_nested_loop= false;
      rc= NESTED_LOOP_OK;
    }
  }
  else
  {
    join->return_tab= join_tab;

    if (join_tab->last_inner)
    {
      /* join_tab is the first inner table for an outer join operation. */

      /* Set initial state of guard variables for this table.*/
      join_tab->found=0;
      join_tab->not_null_compl= 1;

      /* Set first_unmatched for the last inner table of this group */
      join_tab->last_inner->first_unmatched= join_tab;
    }
    join->session->row_count= 0;

    error= (*join_tab->read_first_record)(join_tab);
    rc= evaluate_join_record(join, join_tab, error);
  }

  /*
    Note: psergey has added the 2nd part of the following condition; the
    change should probably be made in 5.1, too.
  */
  while (rc == NESTED_LOOP_OK && join->return_tab >= join_tab)
  {
    error= info->read_record(info);
    rc= evaluate_join_record(join, join_tab, error);
  }

  if (rc == NESTED_LOOP_NO_MORE_ROWS &&
      join_tab->last_inner && !join_tab->found)
    rc= evaluate_null_complemented_join_record(join, join_tab);

  if (rc == NESTED_LOOP_NO_MORE_ROWS)
    rc= NESTED_LOOP_OK;
  return rc;
}




/*
  SemiJoinDuplicateElimination: Weed out duplicate row combinations

  SYNPOSIS
    do_sj_dups_weedout()

  RETURN
    -1  Error
    1   The row combination is a duplicate (discard it)
    0   The row combination is not a duplicate (continue)
*/

int do_sj_dups_weedout(Session *session, SJ_TMP_TABLE *sjtbl)
{
  int error;
  SJ_TMP_TABLE::TAB *tab= sjtbl->tabs;
  SJ_TMP_TABLE::TAB *tab_end= sjtbl->tabs_end;
  unsigned char *ptr= sjtbl->tmp_table->record[0] + 1;
  unsigned char *nulls_ptr= ptr;

  /* Put the the rowids tuple into table->record[0]: */

  // 1. Store the length
  if (((Field_varstring*)(sjtbl->tmp_table->field[0]))->length_bytes == 1)
  {
    *ptr= (unsigned char)(sjtbl->rowid_len + sjtbl->null_bytes);
    ptr++;
  }
  else
  {
    int2store(ptr, sjtbl->rowid_len + sjtbl->null_bytes);
    ptr += 2;
  }

  // 2. Zero the null bytes
  if (sjtbl->null_bytes)
  {
    memset(ptr, 0, sjtbl->null_bytes);
    ptr += sjtbl->null_bytes;
  }

  // 3. Put the rowids
  for (uint32_t i=0; tab != tab_end; tab++, i++)
  {
    handler *h= tab->join_tab->table->file;
    if (tab->join_tab->table->maybe_null && tab->join_tab->table->null_row)
    {
      /* It's a NULL-complemented row */
      *(nulls_ptr + tab->null_byte) |= tab->null_bit;
      memset(ptr + tab->rowid_offset, 0, h->ref_length);
    }
    else
    {
      /* Copy the rowid value */
      if (tab->join_tab->rowid_keep_flags & JOIN_TAB::CALL_POSITION)
        h->position(tab->join_tab->table->record[0]);
      memcpy(ptr + tab->rowid_offset, h->ref, h->ref_length);
    }
  }

  error= sjtbl->tmp_table->file->ha_write_row(sjtbl->tmp_table->record[0]);
  if (error)
  {
    /* create_myisam_from_heap will generate error if needed */
    if (sjtbl->tmp_table->file->is_fatal_error(error, HA_CHECK_DUP) &&
        create_myisam_from_heap(session, sjtbl->tmp_table, sjtbl->start_recinfo,
                                &sjtbl->recinfo, error, 1))
      return -1;
    //return (error == HA_ERR_FOUND_DUPP_KEY || error== HA_ERR_FOUND_DUPP_UNIQUE) ? 1: -1;
    return 1;
  }
  return 0;
}


/*
  SemiJoinDuplicateElimination: Reset the temporary table
*/

int do_sj_reset(SJ_TMP_TABLE *sj_tbl)
{
  if (sj_tbl->tmp_table)
    return sj_tbl->tmp_table->file->ha_delete_all_rows();
  return 0;
}

/*
  Process one record of the nested loop join.

    This function will evaluate parts of WHERE/ON clauses that are
    applicable to the partial record on hand and in case of success
    submit this record to the next level of the nested loop.
*/

static enum_nested_loop_state
evaluate_join_record(JOIN *join, JOIN_TAB *join_tab,
                     int error)
{
  bool not_used_in_distinct=join_tab->not_used_in_distinct;
  ha_rows found_records=join->found_records;
  COND *select_cond= join_tab->select_cond;

  if (error > 0 || (join->session->is_error()))     // Fatal error
    return NESTED_LOOP_ERROR;
  if (error < 0)
    return NESTED_LOOP_NO_MORE_ROWS;
  if (join->session->killed)			// Aborted by user
  {
    join->session->send_kill_message();
    return NESTED_LOOP_KILLED;               /* purecov: inspected */
  }
  if (!select_cond || select_cond->val_int())
  {
    /*
      There is no select condition or the attached pushed down
      condition is true => a match is found.
    */
    bool found= 1;
    while (join_tab->first_unmatched && found)
    {
      /*
        The while condition is always false if join_tab is not
        the last inner join table of an outer join operation.
      */
      JOIN_TAB *first_unmatched= join_tab->first_unmatched;
      /*
        Mark that a match for current outer table is found.
        This activates push down conditional predicates attached
        to the all inner tables of the outer join.
      */
      first_unmatched->found= 1;
      for (JOIN_TAB *tab= first_unmatched; tab <= join_tab; tab++)
      {
        if (tab->table->reginfo.not_exists_optimize)
          return NESTED_LOOP_NO_MORE_ROWS;
        /* Check all predicates that has just been activated. */
        /*
          Actually all predicates non-guarded by first_unmatched->found
          will be re-evaluated again. It could be fixed, but, probably,
          it's not worth doing now.
        */
        if (tab->select_cond && !tab->select_cond->val_int())
        {
          /* The condition attached to table tab is false */
          if (tab == join_tab)
            found= 0;
          else
          {
            /*
              Set a return point if rejected predicate is attached
              not to the last table of the current nest level.
            */
            join->return_tab= tab;
            return NESTED_LOOP_OK;
          }
        }
      }
      /*
        Check whether join_tab is not the last inner table
        for another embedding outer join.
      */
      if ((first_unmatched= first_unmatched->first_upper) &&
          first_unmatched->last_inner != join_tab)
        first_unmatched= 0;
      join_tab->first_unmatched= first_unmatched;
    }

    JOIN_TAB *return_tab= join->return_tab;
    join_tab->found_match= true;
    if (join_tab->check_weed_out_table)
    {
      int res= do_sj_dups_weedout(join->session, join_tab->check_weed_out_table);
      if (res == -1)
        return NESTED_LOOP_ERROR;
      if (res == 1)
        return NESTED_LOOP_OK;
    }
    else if (join_tab->do_firstmatch)
    {
      /*
        We should return to the join_tab->do_firstmatch after we have
        enumerated all the suffixes for current prefix row combination
      */
      return_tab= join_tab->do_firstmatch;
    }

    /*
      It was not just a return to lower loop level when one
      of the newly activated predicates is evaluated as false
      (See above join->return_tab= tab).
    */
    join->examined_rows++;
    join->session->row_count++;

    if (found)
    {
      enum enum_nested_loop_state rc;
      /* A match from join_tab is found for the current partial join. */
      rc= (*join_tab->next_select)(join, join_tab+1, 0);
      if (rc != NESTED_LOOP_OK && rc != NESTED_LOOP_NO_MORE_ROWS)
        return rc;
      if (return_tab < join->return_tab)
        join->return_tab= return_tab;

      if (join->return_tab < join_tab)
        return NESTED_LOOP_OK;
      /*
        Test if this was a SELECT DISTINCT query on a table that
        was not in the field list;  In this case we can abort if
        we found a row, as no new rows can be added to the result.
      */
      if (not_used_in_distinct && found_records != join->found_records)
        return NESTED_LOOP_NO_MORE_ROWS;
    }
    else
      join_tab->read_record.file->unlock_row();
  }
  else
  {
    /*
      The condition pushed down to the table join_tab rejects all rows
      with the beginning coinciding with the current partial join.
    */
    join->examined_rows++;
    join->session->row_count++;
    join_tab->read_record.file->unlock_row();
  }
  return NESTED_LOOP_OK;
}


/**

  @details
    Construct a NULL complimented partial join record and feed it to the next
    level of the nested loop. This function is used in case we have
    an OUTER join and no matching record was found.
*/

static enum_nested_loop_state
evaluate_null_complemented_join_record(JOIN *join, JOIN_TAB *join_tab)
{
  /*
    The table join_tab is the first inner table of a outer join operation
    and no matches has been found for the current outer row.
  */
  JOIN_TAB *last_inner_tab= join_tab->last_inner;
  /* Cache variables for faster loop */
  COND *select_cond;
  for ( ; join_tab <= last_inner_tab ; join_tab++)
  {
    /* Change the the values of guard predicate variables. */
    join_tab->found= 1;
    join_tab->not_null_compl= 0;
    /* The outer row is complemented by nulls for each inner tables */
    join_tab->table->restoreRecordAsDefault();  // Make empty record
    join_tab->table->mark_as_null_row();       // For group by without error
    select_cond= join_tab->select_cond;
    /* Check all attached conditions for inner table rows. */
    if (select_cond && !select_cond->val_int())
      return NESTED_LOOP_OK;
  }
  join_tab--;
  /*
    The row complemented by nulls might be the first row
    of embedding outer joins.
    If so, perform the same actions as in the code
    for the first regular outer join row above.
  */
  for ( ; ; )
  {
    JOIN_TAB *first_unmatched= join_tab->first_unmatched;
    if ((first_unmatched= first_unmatched->first_upper) &&
        first_unmatched->last_inner != join_tab)
      first_unmatched= 0;
    join_tab->first_unmatched= first_unmatched;
    if (!first_unmatched)
      break;
    first_unmatched->found= 1;
    for (JOIN_TAB *tab= first_unmatched; tab <= join_tab; tab++)
    {
      if (tab->select_cond && !tab->select_cond->val_int())
      {
        join->return_tab= tab;
        return NESTED_LOOP_OK;
      }
    }
  }
  /*
    The row complemented by nulls satisfies all conditions
    attached to inner tables.
    Send the row complemented by nulls to be joined with the
    remaining tables.
  */
  return (*join_tab->next_select)(join, join_tab+1, 0);
}


static enum_nested_loop_state
flush_cached_records(JOIN *join,JOIN_TAB *join_tab,bool skip_last)
{
  enum_nested_loop_state rc= NESTED_LOOP_OK;
  int error;
  READ_RECORD *info;

  join_tab->table->null_row= 0;
  if (!join_tab->cache.records)
    return NESTED_LOOP_OK;                      /* Nothing to do */
  if (skip_last)
    (void) store_record_in_cache(&join_tab->cache); // Must save this for later
  if (join_tab->use_quick == 2)
  {
    if (join_tab->select->quick)
    {					/* Used quick select last. reset it */
      delete join_tab->select->quick;
      join_tab->select->quick=0;
    }
  }
 /* read through all records */
  if ((error=join_init_read_record(join_tab)))
  {
    reset_cache_write(&join_tab->cache);
    return error < 0 ? NESTED_LOOP_NO_MORE_ROWS: NESTED_LOOP_ERROR;
  }

  for (JOIN_TAB *tmp=join->join_tab; tmp != join_tab ; tmp++)
  {
    tmp->status=tmp->table->status;
    tmp->table->status=0;
  }

  info= &join_tab->read_record;
  do
  {
    if (join->session->killed)
    {
      join->session->send_kill_message();
      return NESTED_LOOP_KILLED; // Aborted by user /* purecov: inspected */
    }
    SQL_SELECT *select=join_tab->select;
    if (rc == NESTED_LOOP_OK &&
        (!join_tab->cache.select || !join_tab->cache.select->skip_record()))
    {
      uint32_t i;
      reset_cache_read(&join_tab->cache);
      for (i=(join_tab->cache.records- (skip_last ? 1 : 0)) ; i-- > 0 ;)
      {
	read_cached_record(join_tab);
	if (!select || !select->skip_record())
        {
          int res= 0;
          if (!join_tab->check_weed_out_table ||
              !(res= do_sj_dups_weedout(join->session, join_tab->check_weed_out_table)))
          {
            rc= (join_tab->next_select)(join,join_tab+1,0);
            if (rc != NESTED_LOOP_OK && rc != NESTED_LOOP_NO_MORE_ROWS)
            {
              reset_cache_write(&join_tab->cache);
              return rc;
            }
          }
          if (res == -1)
            return NESTED_LOOP_ERROR;
        }
      }
    }
  } while (!(error=info->read_record(info)));

  if (skip_last)
    read_cached_record(join_tab);		// Restore current record
  reset_cache_write(&join_tab->cache);
  if (error > 0)				// Fatal error
    return NESTED_LOOP_ERROR;                   /* purecov: inspected */
  for (JOIN_TAB *tmp2=join->join_tab; tmp2 != join_tab ; tmp2++)
    tmp2->table->status=tmp2->status;
  return NESTED_LOOP_OK;
}

int safe_index_read(JOIN_TAB *tab)
{
  int error;
  Table *table= tab->table;
  if ((error=table->file->index_read_map(table->record[0],
                                         tab->ref.key_buff,
                                         make_prev_keypart_map(tab->ref.key_parts),
                                         HA_READ_KEY_EXACT)))
    return table->report_error(error);
  return 0;
}


static int
join_read_const_table(JOIN_TAB *tab, POSITION *pos)
{
  int error;
  Table *table=tab->table;
  table->const_table=1;
  table->null_row=0;
  table->status=STATUS_NO_RECORD;

  if (tab->type == JT_SYSTEM)
  {
    if ((error=join_read_system(tab)))
    {						// Info for DESCRIBE
      tab->info="const row not found";
      /* Mark for EXPLAIN that the row was not found */
      pos->records_read=0.0;
      pos->ref_depend_map= 0;
      if (!table->maybe_null || error > 0)
	return(error);
    }
  }
  else
  {
    if (!table->key_read && table->covering_keys.test(tab->ref.key) &&
	!table->no_keyread &&
        (int) table->reginfo.lock_type <= (int) TL_READ_WITH_SHARED_LOCKS)
    {
      table->key_read=1;
      table->file->extra(HA_EXTRA_KEYREAD);
      tab->index= tab->ref.key;
    }
    error=join_read_const(tab);
    if (table->key_read)
    {
      table->key_read=0;
      table->file->extra(HA_EXTRA_NO_KEYREAD);
    }
    if (error)
    {
      tab->info="unique row not found";
      /* Mark for EXPLAIN that the row was not found */
      pos->records_read=0.0;
      pos->ref_depend_map= 0;
      if (!table->maybe_null || error > 0)
	return(error);
    }
  }
  if (*tab->on_expr_ref && !table->null_row)
  {
    if ((table->null_row= test((*tab->on_expr_ref)->val_int() == 0)))
      table->mark_as_null_row();
  }
  if (!table->null_row)
    table->maybe_null=0;

  /* Check appearance of new constant items in Item_equal objects */
  JOIN *join= tab->join;
  if (join->conds)
    update_const_equal_items(join->conds, tab);
  TableList *tbl;
  for (tbl= join->select_lex->leaf_tables; tbl; tbl= tbl->next_leaf)
  {
    TableList *embedded;
    TableList *embedding= tbl;
    do
    {
      embedded= embedding;
      if (embedded->on_expr)
         update_const_equal_items(embedded->on_expr, tab);
      embedding= embedded->embedding;
    }
    while (embedding &&
           embedding->nested_join->join_list.head() == embedded);
  }

  return(0);
}


static int
join_read_system(JOIN_TAB *tab)
{
  Table *table= tab->table;
  int error;
  if (table->status & STATUS_GARBAGE)		// If first read
  {
    if ((error=table->file->read_first_row(table->record[0],
					   table->s->primary_key)))
    {
      if (error != HA_ERR_END_OF_FILE)
	return table->report_error(error);
      tab->table->mark_as_null_row();
      table->emptyRecord();			// Make empty record
      return -1;
    }
    table->storeRecord();
  }
  else if (!table->status)			// Only happens with left join
    table->restoreRecord();			// restore old record
  table->null_row=0;
  return table->status ? -1 : 0;
}


/**
  Read a (constant) table when there is at most one matching row.

  @param tab			Table to read

  @retval
    0	Row was found
  @retval
    -1   Row was not found
  @retval
    1   Got an error (other than row not found) during read
*/

static int
join_read_const(JOIN_TAB *tab)
{
  int error;
  Table *table= tab->table;
  if (table->status & STATUS_GARBAGE)		// If first read
  {
    table->status= 0;
    if (cp_buffer_from_ref(tab->join->session, &tab->ref))
      error=HA_ERR_KEY_NOT_FOUND;
    else
    {
      error=table->file->index_read_idx_map(table->record[0],tab->ref.key,
                                            (unsigned char*) tab->ref.key_buff,
                                            make_prev_keypart_map(tab->ref.key_parts),
                                            HA_READ_KEY_EXACT);
    }
    if (error)
    {
      table->status= STATUS_NOT_FOUND;
      tab->table->mark_as_null_row();
      table->emptyRecord();
      if (error != HA_ERR_KEY_NOT_FOUND && error != HA_ERR_END_OF_FILE)
	return table->report_error(error);
      return -1;
    }
    table->storeRecord();
  }
  else if (!(table->status & ~STATUS_NULL_ROW))	// Only happens with left join
  {
    table->status=0;
    table->restoreRecord();			// restore old record
  }
  table->null_row=0;
  return table->status ? -1 : 0;
}


/*
  eq_ref access method implementation: "read_first" function

  SYNOPSIS
    join_read_key()
      tab  JOIN_TAB of the accessed table

  DESCRIPTION
    This is "read_fist" function for the "ref" access method. The difference
    from "ref" is that it has a one-element "cache" (see cmp_buffer_with_ref)

  RETURN
    0  - Ok
   -1  - Row not found
    1  - Error
*/

static int
join_read_key(JOIN_TAB *tab)
{
  int error;
  Table *table= tab->table;

  if (!table->file->inited)
  {
    table->file->ha_index_init(tab->ref.key, tab->sorted);
  }

  /* TODO: Why don't we do "Late NULLs Filtering" here? */
  if (cmp_buffer_with_ref(tab) ||
      (table->status & (STATUS_GARBAGE | STATUS_NO_PARENT | STATUS_NULL_ROW)))
  {
    if (tab->ref.key_err)
    {
      table->status=STATUS_NOT_FOUND;
      return -1;
    }
    error=table->file->index_read_map(table->record[0],
                                      tab->ref.key_buff,
                                      make_prev_keypart_map(tab->ref.key_parts),
                                      HA_READ_KEY_EXACT);
    if (error && error != HA_ERR_KEY_NOT_FOUND && error != HA_ERR_END_OF_FILE)
      return table->report_error(error);
  }
  table->null_row=0;
  return table->status ? -1 : 0;
}


/*
  ref access method implementation: "read_first" function

  SYNOPSIS
    join_read_always_key()
      tab  JOIN_TAB of the accessed table

  DESCRIPTION
    This is "read_fist" function for the "ref" access method.

    The functon must leave the index initialized when it returns.
    ref_or_null access implementation depends on that.

  RETURN
    0  - Ok
   -1  - Row not found
    1  - Error
*/

static int
join_read_always_key(JOIN_TAB *tab)
{
  int error;
  Table *table= tab->table;

  /* Initialize the index first */
  if (!table->file->inited)
    table->file->ha_index_init(tab->ref.key, tab->sorted);

  /* Perform "Late NULLs Filtering" (see internals manual for explanations) */
  for (uint32_t i= 0 ; i < tab->ref.key_parts ; i++)
  {
    if ((tab->ref.null_rejecting & 1 << i) && tab->ref.items[i]->is_null())
        return -1;
  }

  if (cp_buffer_from_ref(tab->join->session, &tab->ref))
    return -1;
  if ((error=table->file->index_read_map(table->record[0],
                                         tab->ref.key_buff,
                                         make_prev_keypart_map(tab->ref.key_parts),
                                         HA_READ_KEY_EXACT)))
  {
    if (error != HA_ERR_KEY_NOT_FOUND && error != HA_ERR_END_OF_FILE)
      return table->report_error(error);
    return -1; /* purecov: inspected */
  }

  return 0;
}


/**
  This function is used when optimizing away order_st BY in
  SELECT * FROM t1 WHERE a=1 order_st BY a DESC,b DESC.
*/

static int
join_read_last_key(JOIN_TAB *tab)
{
  int error;
  Table *table= tab->table;

  if (!table->file->inited)
    table->file->ha_index_init(tab->ref.key, tab->sorted);
  if (cp_buffer_from_ref(tab->join->session, &tab->ref))
    return -1;
  if ((error=table->file->index_read_last_map(table->record[0],
                                              tab->ref.key_buff,
                                              make_prev_keypart_map(tab->ref.key_parts))))
  {
    if (error != HA_ERR_KEY_NOT_FOUND && error != HA_ERR_END_OF_FILE)
      return table->report_error(error);
    return -1; /* purecov: inspected */
  }
  return 0;
}


	/* ARGSUSED */
static int
join_no_more_records(READ_RECORD *)
{
  return -1;
}

static int
join_read_next_same_diff(READ_RECORD *info)
{
  Table *table= info->table;
  JOIN_TAB *tab=table->reginfo.join_tab;
  if (tab->insideout_match_tab->found_match)
  {
    KEY *key= tab->table->key_info + tab->index;
    do
    {
      int error;
      /* Save index tuple from record to the buffer */
      key_copy(tab->insideout_buf, info->record, key, 0);

      if ((error=table->file->index_next_same(table->record[0],
                                              tab->ref.key_buff,
                                              tab->ref.key_length)))
      {
        if (error != HA_ERR_END_OF_FILE)
          return table->report_error(error);
        table->status= STATUS_GARBAGE;
        return -1;
      }
    } while (!key_cmp(tab->table->key_info[tab->index].key_part,
                      tab->insideout_buf, key->key_length));
    tab->insideout_match_tab->found_match= 0;
    return 0;
  }
  else
    return join_read_next_same(info);
}

static int
join_read_next_same(READ_RECORD *info)
{
  int error;
  Table *table= info->table;
  JOIN_TAB *tab=table->reginfo.join_tab;

  if ((error=table->file->index_next_same(table->record[0],
					  tab->ref.key_buff,
					  tab->ref.key_length)))
  {
    if (error != HA_ERR_END_OF_FILE)
      return table->report_error(error);
    table->status= STATUS_GARBAGE;
    return -1;
  }

  return 0;
}


static int
join_read_prev_same(READ_RECORD *info)
{
  int error;
  Table *table= info->table;
  JOIN_TAB *tab=table->reginfo.join_tab;

  if ((error=table->file->index_prev(table->record[0])))
    return table->report_error(error);
  if (key_cmp_if_same(table, tab->ref.key_buff, tab->ref.key,
                      tab->ref.key_length))
  {
    table->status=STATUS_NOT_FOUND;
    error= -1;
  }

  return error;
}


static int
join_init_quick_read_record(JOIN_TAB *tab)
{
  if (test_if_quick_select(tab) == -1)
    return -1;					/* No possible records */
  return join_init_read_record(tab);
}


int rr_sequential(READ_RECORD *info);
int init_read_record_seq(JOIN_TAB *tab)
{
  tab->read_record.read_record= rr_sequential;
  if (tab->read_record.file->ha_rnd_init(1))
    return 1;
  return (*tab->read_record.read_record)(&tab->read_record);
}

static int
test_if_quick_select(JOIN_TAB *tab)
{
  delete tab->select->quick;
  tab->select->quick=0;
  return tab->select->test_quick_select(tab->join->session, tab->keys,
					(table_map) 0, HA_POS_ERROR, 0,
                                        false);
}


static int
join_init_read_record(JOIN_TAB *tab)
{
  if (tab->select && tab->select->quick && tab->select->quick->reset())
    return 1;
  init_read_record(&tab->read_record, tab->join->session, tab->table,
		   tab->select,1,1);
  return (*tab->read_record.read_record)(&tab->read_record);
}


static int
join_read_first(JOIN_TAB *tab)
{
  int error;
  Table *table=tab->table;
  if (!table->key_read && table->covering_keys.test(tab->index) &&
      !table->no_keyread)
  {
    table->key_read=1;
    table->file->extra(HA_EXTRA_KEYREAD);
  }
  tab->table->status=0;
  tab->read_record.table=table;
  tab->read_record.file=table->file;
  tab->read_record.index=tab->index;
  tab->read_record.record=table->record[0];
  if (tab->insideout_match_tab)
  {
    tab->read_record.do_insideout_scan= tab;
    tab->read_record.read_record=join_read_next_different;
    tab->insideout_match_tab->found_match= 0;
  }
  else
  {
    tab->read_record.read_record=join_read_next;
    tab->read_record.do_insideout_scan= 0;
  }

  if (!table->file->inited)
    table->file->ha_index_init(tab->index, tab->sorted);
  if ((error=tab->table->file->index_first(tab->table->record[0])))
  {
    if (error != HA_ERR_KEY_NOT_FOUND && error != HA_ERR_END_OF_FILE)
      table->report_error(error);
    return -1;
  }

  return 0;
}


static int
join_read_next_different(READ_RECORD *info)
{
  JOIN_TAB *tab= info->do_insideout_scan;
  if (tab->insideout_match_tab->found_match)
  {
    KEY *key= tab->table->key_info + tab->index;
    do
    {
      int error;
      /* Save index tuple from record to the buffer */
      key_copy(tab->insideout_buf, info->record, key, 0);

      if ((error=info->file->index_next(info->record)))
        return info->table->report_error(error);
    } while (!key_cmp(tab->table->key_info[tab->index].key_part,
                      tab->insideout_buf, key->key_length));
    tab->insideout_match_tab->found_match= 0;
    return 0;
  }
  else
    return join_read_next(info);
}


static int
join_read_next(READ_RECORD *info)
{
  int error;
  if ((error=info->file->index_next(info->record)))
    return info->table->report_error(error);
  return 0;
}


static int
join_read_last(JOIN_TAB *tab)
{
  Table *table=tab->table;
  int error;
  if (!table->key_read && table->covering_keys.test(tab->index) &&
      !table->no_keyread)
  {
    table->key_read=1;
    table->file->extra(HA_EXTRA_KEYREAD);
  }
  tab->table->status=0;
  tab->read_record.read_record=join_read_prev;
  tab->read_record.table=table;
  tab->read_record.file=table->file;
  tab->read_record.index=tab->index;
  tab->read_record.record=table->record[0];
  if (!table->file->inited)
    table->file->ha_index_init(tab->index, 1);
  if ((error= tab->table->file->index_last(tab->table->record[0])))
    return table->report_error(error);

  return 0;
}


static int
join_read_prev(READ_RECORD *info)
{
  int error;
  if ((error= info->file->index_prev(info->record)))
    return info->table->report_error(error);

  return 0;
}

/**
  Reading of key with key reference and one part that may be NULL.
*/

int
join_read_always_key_or_null(JOIN_TAB *tab)
{
  int res;

  /* First read according to key which is NOT NULL */
  *tab->ref.null_ref_key= 0;			// Clear null byte
  if ((res= join_read_always_key(tab)) >= 0)
    return res;

  /* Then read key with null value */
  *tab->ref.null_ref_key= 1;			// Set null byte
  return safe_index_read(tab);
}


int
join_read_next_same_or_null(READ_RECORD *info)
{
  int error;
  if ((error= join_read_next_same(info)) >= 0)
    return error;
  JOIN_TAB *tab= info->table->reginfo.join_tab;

  /* Test if we have already done a read after null key */
  if (*tab->ref.null_ref_key)
    return -1;					// All keys read
  *tab->ref.null_ref_key= 1;			// Set null byte
  return safe_index_read(tab);			// then read null keys
}


/*****************************************************************************
  DESCRIPTION
    Functions that end one nested loop iteration. Different functions
    are used to support GROUP BY clause and to redirect records
    to a table (e.g. in case of SELECT into a temporary table) or to the
    network client.

  RETURN VALUES
    NESTED_LOOP_OK           - the record has been successfully handled
    NESTED_LOOP_ERROR        - a fatal error (like table corruption)
                               was detected
    NESTED_LOOP_KILLED       - thread shutdown was requested while processing
                               the record
    NESTED_LOOP_QUERY_LIMIT  - the record has been successfully handled;
                               additionally, the nested loop produced the
                               number of rows specified in the LIMIT clause
                               for the query
    NESTED_LOOP_CURSOR_LIMIT - the record has been successfully handled;
                               additionally, there is a cursor and the nested
                               loop algorithm produced the number of rows
                               that is specified for current cursor fetch
                               operation.
   All return values except NESTED_LOOP_OK abort the nested loop.
*****************************************************************************/

/* ARGSUSED */
static enum_nested_loop_state
end_send(JOIN *join, JOIN_TAB *,
         bool end_of_records)
{
  if (!end_of_records)
  {
    int error;
    if (join->having && join->having->val_int() == 0)
      return(NESTED_LOOP_OK);               // Didn't match having
    error=0;
    if (join->do_send_rows)
      error=join->result->send_data(*join->fields);
    if (error)
      return(NESTED_LOOP_ERROR); /* purecov: inspected */
    if (++join->send_records >= join->unit->select_limit_cnt &&
	join->do_send_rows)
    {
      if (join->select_options & OPTION_FOUND_ROWS)
      {
	JOIN_TAB *jt=join->join_tab;
	if ((join->tables == 1) && !join->tmp_table && !join->sort_and_group
	    && !join->send_group_parts && !join->having && !jt->select_cond &&
	    !(jt->select && jt->select->quick) &&
	    (jt->table->file->ha_table_flags() & HA_STATS_RECORDS_IS_EXACT) &&
            (jt->ref.key < 0))
	{
	  /* Join over all rows in table;  Return number of found rows */
	  Table *table=jt->table;

	  join->select_options ^= OPTION_FOUND_ROWS;
	  if (table->sort.record_pointers ||
	      (table->sort.io_cache && my_b_inited(table->sort.io_cache)))
	  {
	    /* Using filesort */
	    join->send_records= table->sort.found_records;
	  }
	  else
	  {
	    table->file->info(HA_STATUS_VARIABLE);
	    join->send_records= table->file->stats.records;
	  }
	}
	else
	{
	  join->do_send_rows= 0;
	  if (join->unit->fake_select_lex)
	    join->unit->fake_select_lex->select_limit= 0;
	  return(NESTED_LOOP_OK);
	}
      }
      return(NESTED_LOOP_QUERY_LIMIT);      // Abort nicely
    }
    else if (join->send_records >= join->fetch_limit)
    {
      /*
        There is a server side cursor and all rows for
        this fetch request are sent.
      */
      return(NESTED_LOOP_CURSOR_LIMIT);
    }
  }

  return(NESTED_LOOP_OK);
}


/* ARGSUSED */
enum_nested_loop_state
end_send_group(JOIN *join, JOIN_TAB *, bool end_of_records)
{
  int idx= -1;
  enum_nested_loop_state ok_code= NESTED_LOOP_OK;

  if (!join->first_record || end_of_records ||
      (idx=test_if_item_cache_changed(join->group_fields)) >= 0)
  {
    if (join->first_record ||
        (end_of_records && !join->group && !join->group_optimized_away))
    {
      if (idx < (int) join->send_group_parts)
      {
	int error=0;
	{
	  if (!join->first_record)
	  {
            List_iterator_fast<Item> it(*join->fields);
            Item *item;
	    /* No matching rows for group function */
	    join->clear();

            while ((item= it++))
              item->no_rows_in_result();
	  }
	  if (join->having && join->having->val_int() == 0)
	    error= -1;				// Didn't satisfy having
	  else
	  {
	    if (join->do_send_rows)
	      error=join->result->send_data(*join->fields) ? 1 : 0;
	    join->send_records++;
	  }
	  if (join->rollup.state != ROLLUP::STATE_NONE && error <= 0)
	  {
	    if (join->rollup_send_data((uint32_t) (idx+1)))
	      error= 1;
	  }
	}
	if (error > 0)
          return(NESTED_LOOP_ERROR);        /* purecov: inspected */
	if (end_of_records)
	  return(NESTED_LOOP_OK);
	if (join->send_records >= join->unit->select_limit_cnt &&
	    join->do_send_rows)
	{
	  if (!(join->select_options & OPTION_FOUND_ROWS))
	    return(NESTED_LOOP_QUERY_LIMIT); // Abort nicely
	  join->do_send_rows=0;
	  join->unit->select_limit_cnt = HA_POS_ERROR;
        }
        else if (join->send_records >= join->fetch_limit)
        {
          /*
            There is a server side cursor and all rows
            for this fetch request are sent.
          */
          /*
            Preventing code duplication. When finished with the group reset
            the group functions and copy_fields. We fall through. bug #11904
          */
          ok_code= NESTED_LOOP_CURSOR_LIMIT;
        }
      }
    }
    else
    {
      if (end_of_records)
	return(NESTED_LOOP_OK);
      join->first_record=1;
      test_if_item_cache_changed(join->group_fields);
    }
    if (idx < (int) join->send_group_parts)
    {
      /*
        This branch is executed also for cursors which have finished their
        fetch limit - the reason for ok_code.
      */
      copy_fields(&join->tmp_table_param);
      if (init_sum_functions(join->sum_funcs, join->sum_funcs_end[idx+1]))
	return(NESTED_LOOP_ERROR);
      return(ok_code);
    }
  }
  if (update_sum_func(join->sum_funcs))
    return(NESTED_LOOP_ERROR);
  return(NESTED_LOOP_OK);
}


/* ARGSUSED */
enum_nested_loop_state
end_write(JOIN *join, JOIN_TAB *,
          bool end_of_records)
{
  Table *table=join->tmp_table;

  if (join->session->killed)			// Aborted by user
  {
    join->session->send_kill_message();
    return(NESTED_LOOP_KILLED);             /* purecov: inspected */
  }
  if (!end_of_records)
  {
    copy_fields(&join->tmp_table_param);
    copy_funcs(join->tmp_table_param.items_to_copy);
    if (!join->having || join->having->val_int())
    {
      int error;
      join->found_records++;
      if ((error=table->file->ha_write_row(table->record[0])))
      {
        if (!table->file->is_fatal_error(error, HA_CHECK_DUP))
	  goto end;
	if (create_myisam_from_heap(join->session, table,
                                    join->tmp_table_param.start_recinfo,
                                    &join->tmp_table_param.recinfo,
				    error, 1))
	  return(NESTED_LOOP_ERROR);        // Not a table_is_full error
	table->s->uniques=0;			// To ensure rows are the same
      }
      if (++join->send_records >= join->tmp_table_param.end_write_records &&
	  join->do_send_rows)
      {
	if (!(join->select_options & OPTION_FOUND_ROWS))
	  return(NESTED_LOOP_QUERY_LIMIT);
	join->do_send_rows=0;
	join->unit->select_limit_cnt = HA_POS_ERROR;
	return(NESTED_LOOP_OK);
      }
    }
  }
end:
  return(NESTED_LOOP_OK);
}

/* ARGSUSED */
/** Group by searching after group record and updating it if possible. */

static enum_nested_loop_state
end_update(JOIN *join, JOIN_TAB *,
           bool end_of_records)
{
  Table *table=join->tmp_table;
  order_st   *group;
  int	  error;

  if (end_of_records)
    return(NESTED_LOOP_OK);
  if (join->session->killed)			// Aborted by user
  {
    join->session->send_kill_message();
    return(NESTED_LOOP_KILLED);             /* purecov: inspected */
  }

  join->found_records++;
  copy_fields(&join->tmp_table_param);		// Groups are copied twice.
  /* Make a key of group index */
  for (group=table->group ; group ; group=group->next)
  {
    Item *item= *group->item;
    item->save_org_in_field(group->field);
    /* Store in the used key if the field was 0 */
    if (item->maybe_null)
      group->buff[-1]= (char) group->field->is_null();
  }
  if (!table->file->index_read_map(table->record[1],
                                   join->tmp_table_param.group_buff,
                                   HA_WHOLE_KEY,
                                   HA_READ_KEY_EXACT))
  {						/* Update old record */
    table->restoreRecord();
    update_tmptable_sum_func(join->sum_funcs,table);
    if ((error=table->file->ha_update_row(table->record[1],
                                          table->record[0])))
    {
      table->file->print_error(error,MYF(0));	/* purecov: inspected */
      return(NESTED_LOOP_ERROR);            /* purecov: inspected */
    }
    return(NESTED_LOOP_OK);
  }

  /*
    Copy null bits from group key to table
    We can't copy all data as the key may have different format
    as the row data (for example as with VARCHAR keys)
  */
  KEY_PART_INFO *key_part;
  for (group=table->group,key_part=table->key_info[0].key_part;
       group ;
       group=group->next,key_part++)
  {
    if (key_part->null_bit)
      memcpy(table->record[0]+key_part->offset, group->buff, 1);
  }
  init_tmptable_sum_functions(join->sum_funcs);
  copy_funcs(join->tmp_table_param.items_to_copy);
  if ((error=table->file->ha_write_row(table->record[0])))
  {
    if (create_myisam_from_heap(join->session, table,
                                join->tmp_table_param.start_recinfo,
                                &join->tmp_table_param.recinfo,
				error, 0))
      return(NESTED_LOOP_ERROR);            // Not a table_is_full error
    /* Change method to update rows */
    table->file->ha_index_init(0, 0);
    join->join_tab[join->tables-1].next_select=end_unique_update;
  }
  join->send_records++;
  return(NESTED_LOOP_OK);
}


/** Like end_update, but this is done with unique constraints instead of keys.  */

static enum_nested_loop_state
end_unique_update(JOIN *join, JOIN_TAB *,
                  bool end_of_records)
{
  Table *table=join->tmp_table;
  int	  error;

  if (end_of_records)
    return(NESTED_LOOP_OK);
  if (join->session->killed)			// Aborted by user
  {
    join->session->send_kill_message();
    return(NESTED_LOOP_KILLED);             /* purecov: inspected */
  }

  init_tmptable_sum_functions(join->sum_funcs);
  copy_fields(&join->tmp_table_param);		// Groups are copied twice.
  copy_funcs(join->tmp_table_param.items_to_copy);

  if (!(error=table->file->ha_write_row(table->record[0])))
    join->send_records++;			// New group
  else
  {
    if ((int) table->file->get_dup_key(error) < 0)
    {
      table->file->print_error(error,MYF(0));	/* purecov: inspected */
      return(NESTED_LOOP_ERROR);            /* purecov: inspected */
    }
    if (table->file->rnd_pos(table->record[1],table->file->dup_ref))
    {
      table->file->print_error(error,MYF(0));	/* purecov: inspected */
      return(NESTED_LOOP_ERROR);            /* purecov: inspected */
    }
    table->restoreRecord();
    update_tmptable_sum_func(join->sum_funcs,table);
    if ((error=table->file->ha_update_row(table->record[1],
                                          table->record[0])))
    {
      table->file->print_error(error,MYF(0));	/* purecov: inspected */
      return(NESTED_LOOP_ERROR);            /* purecov: inspected */
    }
  }
  return(NESTED_LOOP_OK);
}


/* ARGSUSED */
enum_nested_loop_state
end_write_group(JOIN *join, JOIN_TAB *,
                bool end_of_records)
{
  Table *table=join->tmp_table;
  int	  idx= -1;

  if (join->session->killed)
  {						// Aborted by user
    join->session->send_kill_message();
    return(NESTED_LOOP_KILLED);             /* purecov: inspected */
  }
  if (!join->first_record || end_of_records ||
      (idx=test_if_item_cache_changed(join->group_fields)) >= 0)
  {
    if (join->first_record || (end_of_records && !join->group))
    {
      int send_group_parts= join->send_group_parts;
      if (idx < send_group_parts)
      {
	if (!join->first_record)
	{
	  /* No matching rows for group function */
	  join->clear();
	}
        copy_sum_funcs(join->sum_funcs,
                       join->sum_funcs_end[send_group_parts]);
	if (!join->having || join->having->val_int())
	{
          int error= table->file->ha_write_row(table->record[0]);
          if (error && create_myisam_from_heap(join->session, table,
                                               join->tmp_table_param.start_recinfo,
                                                &join->tmp_table_param.recinfo,
                                               error, 0))
	    return(NESTED_LOOP_ERROR);
        }
        if (join->rollup.state != ROLLUP::STATE_NONE)
	{
	  if (join->rollup_write_data((uint32_t) (idx+1), table))
	    return(NESTED_LOOP_ERROR);
	}
	if (end_of_records)
	  return(NESTED_LOOP_OK);
      }
    }
    else
    {
      if (end_of_records)
	return(NESTED_LOOP_OK);
      join->first_record=1;
      test_if_item_cache_changed(join->group_fields);
    }
    if (idx < (int) join->send_group_parts)
    {
      copy_fields(&join->tmp_table_param);
      copy_funcs(join->tmp_table_param.items_to_copy);
      if (init_sum_functions(join->sum_funcs, join->sum_funcs_end[idx+1]))
	return(NESTED_LOOP_ERROR);
      return(NESTED_LOOP_OK);
    }
  }
  if (update_sum_func(join->sum_funcs))
    return(NESTED_LOOP_ERROR);
  return(NESTED_LOOP_OK);
}


/*****************************************************************************
  Remove calculation with tables that aren't yet read. Remove also tests
  against fields that are read through key where the table is not a
  outer join table.
  We can't remove tests that are made against columns which are stored
  in sorted order.
*****************************************************************************/

/**
  @return
    1 if right_item is used removable reference key on left_item
*/

static bool test_if_ref(Item_field *left_item,Item *right_item)
{
  Field *field=left_item->field;
  // No need to change const test. We also have to keep tests on LEFT JOIN
  if (!field->table->const_table && !field->table->maybe_null)
  {
    Item *ref_item=part_of_refkey(field->table,field);
    if (ref_item && ref_item->eq(right_item,1))
    {
      right_item= right_item->real_item();
      if (right_item->type() == Item::FIELD_ITEM)
	return (field->eq_def(((Item_field *) right_item)->field));
      /* remove equalities injected by IN->EXISTS transformation */
      else if (right_item->type() == Item::CACHE_ITEM)
        return ((Item_cache *)right_item)->eq_def (field);
      if (right_item->const_item() && !(right_item->is_null()))
      {
	/*
	  We can remove binary fields and numerical fields except float,
	  as float comparison isn't 100 % secure
	  We have to keep normal strings to be able to check for end spaces

          sergefp: the above seems to be too restrictive. Counterexample:
            create table t100 (v varchar(10), key(v)) default charset=latin1;
            insert into t100 values ('a'),('a ');
            explain select * from t100 where v='a';
          The EXPLAIN shows 'using Where'. Running the query returns both
          rows, so it seems there are no problems with endspace in the most
          frequent case?
	*/
	if (field->binary() &&
	    field->real_type() != DRIZZLE_TYPE_VARCHAR &&
	    field->decimals() == 0)
	{
	  return !store_val_in_field(field, right_item, CHECK_FIELD_WARN);
	}
      }
    }
  }
  return 0;					// keep test
}

/**
   @brief Replaces an expression destructively inside the expression tree of
   the WHERE clase.

   @note Because of current requirements for semijoin flattening, we do not
   need to recurse here, hence this function will only examine the top-level
   AND conditions. (see JOIN::prepare, comment above the line
   'if (do_materialize)'

   @param join The top-level query.
   @param old_cond The expression to be replaced.
   @param new_cond The expression to be substituted.
   @param do_fix_fields If true, Item::fix_fields(Session*, Item**) is called for
   the new expression.
   @return <code>true</code> if there was an error, <code>false</code> if
   successful.
*/
static bool replace_where_subcondition(JOIN *join, Item *old_cond,
                                       Item *new_cond, bool do_fix_fields)
{
  if (join->conds == old_cond) {
    join->conds= new_cond;
    if (do_fix_fields)
      new_cond->fix_fields(join->session, &join->conds);
    return false;
  }

  if (join->conds->type() == Item::COND_ITEM) {
    List_iterator<Item> li(*((Item_cond*)join->conds)->argument_list());
    Item *item;
    while ((item= li++))
      if (item == old_cond)
      {
        li.replace(new_cond);
        if (do_fix_fields)
          new_cond->fix_fields(join->session, li.ref());
        return false;
      }
  }

  return true;
}

/*
  Extract a condition that can be checked after reading given table

  SYNOPSIS
    make_cond_for_table()
      cond         Condition to analyze
      tables       Tables for which "current field values" are available
      used_table   Table that we're extracting the condition for (may
                   also include PSEUDO_TABLE_BITS

  DESCRIPTION
    Extract the condition that can be checked after reading the table
    specified in 'used_table', given that current-field values for tables
    specified in 'tables' bitmap are available.

    The function assumes that
      - Constant parts of the condition has already been checked.
      - Condition that could be checked for tables in 'tables' has already
        been checked.

    The function takes into account that some parts of the condition are
    guaranteed to be true by employed 'ref' access methods (the code that
    does this is located at the end, search down for "EQ_FUNC").


  SEE ALSO
    make_cond_for_info_schema uses similar algorithm

  RETURN
    Extracted condition
*/

static COND *
make_cond_for_table(COND *cond, table_map tables, table_map used_table,
                    bool exclude_expensive_cond)
{
  if (used_table && !(cond->used_tables() & used_table) &&
      /*
        Exclude constant conditions not checked at optimization time if
        the table we are pushing conditions to is the first one.
        As a result, such conditions are not considered as already checked
        and will be checked at execution time, attached to the first table.
      */
      !((used_table & 1) && cond->is_expensive()))
    return (COND*) 0;				// Already checked
  if (cond->type() == Item::COND_ITEM)
  {
    if (((Item_cond*) cond)->functype() == Item_func::COND_AND_FUNC)
    {
      /* Create new top level AND item */
      Item_cond_and *new_cond=new Item_cond_and;
      if (!new_cond)
	return (COND*) 0;			// OOM /* purecov: inspected */
      List_iterator<Item> li(*((Item_cond*) cond)->argument_list());
      Item *item;
      while ((item=li++))
      {
	Item *fix=make_cond_for_table(item,tables,used_table,
                                      exclude_expensive_cond);
	if (fix)
	  new_cond->argument_list()->push_back(fix);
      }
      switch (new_cond->argument_list()->elements) {
      case 0:
	return (COND*) 0;			// Always true
      case 1:
	return new_cond->argument_list()->head();
      default:
	/*
	  Item_cond_and do not need fix_fields for execution, its parameters
	  are fixed or do not need fix_fields, too
	*/
	new_cond->quick_fix_field();
	new_cond->used_tables_cache=
	  ((Item_cond_and*) cond)->used_tables_cache &
	  tables;
	return new_cond;
      }
    }
    else
    {						// Or list
      Item_cond_or *new_cond=new Item_cond_or;
      if (!new_cond)
	return (COND*) 0;			// OOM /* purecov: inspected */
      List_iterator<Item> li(*((Item_cond*) cond)->argument_list());
      Item *item;
      while ((item=li++))
      {
	Item *fix=make_cond_for_table(item,tables,0L, exclude_expensive_cond);
	if (!fix)
	  return (COND*) 0;			// Always true
	new_cond->argument_list()->push_back(fix);
      }
      /*
	Item_cond_and do not need fix_fields for execution, its parameters
	are fixed or do not need fix_fields, too
      */
      new_cond->quick_fix_field();
      new_cond->used_tables_cache= ((Item_cond_or*) cond)->used_tables_cache;
      new_cond->top_level_item();
      return new_cond;
    }
  }

  /*
    Because the following test takes a while and it can be done
    table_count times, we mark each item that we have examined with the result
    of the test
  */

  if (cond->marker == 3 || (cond->used_tables() & ~tables) ||
      /*
        When extracting constant conditions, treat expensive conditions as
        non-constant, so that they are not evaluated at optimization time.
      */
      (!used_table && exclude_expensive_cond && cond->is_expensive()))
    return (COND*) 0;				// Can't check this yet
  if (cond->marker == 2 || cond->eq_cmp_result() == Item::COND_OK)
    return cond;				// Not boolean op

  /*
    Remove equalities that are guaranteed to be true by use of 'ref' access
    method
  */
  if (((Item_func*) cond)->functype() == Item_func::EQ_FUNC)
  {
    Item *left_item=	((Item_func*) cond)->arguments()[0];
    Item *right_item= ((Item_func*) cond)->arguments()[1];
    if (left_item->type() == Item::FIELD_ITEM &&
	test_if_ref((Item_field*) left_item,right_item))
    {
      cond->marker=3;			// Checked when read
      return (COND*) 0;
    }
    if (right_item->type() == Item::FIELD_ITEM &&
	test_if_ref((Item_field*) right_item,left_item))
    {
      cond->marker=3;			// Checked when read
      return (COND*) 0;
    }
  }
  cond->marker=2;
  return cond;
}


static Item *
part_of_refkey(Table *table,Field *field)
{
  if (!table->reginfo.join_tab)
    return (Item*) 0;             // field from outer non-select (UPDATE,...)

  uint32_t ref_parts=table->reginfo.join_tab->ref.key_parts;
  if (ref_parts)
  {
    KEY_PART_INFO *key_part=
      table->key_info[table->reginfo.join_tab->ref.key].key_part;
    uint32_t part;

    for (part=0 ; part < ref_parts ; part++)
    {
      if (table->reginfo.join_tab->ref.cond_guards[part])
        return 0;
    }

    for (part=0 ; part < ref_parts ; part++,key_part++)
      if (field->eq(key_part->field) &&
	  !(key_part->key_part_flag & HA_PART_KEY_SEG))
	return table->reginfo.join_tab->ref.items[part];
  }
  return (Item*) 0;
}


/**
  Test if one can use the key to resolve order_st BY.

  @param order                 Sort order
  @param table                 Table to sort
  @param idx                   Index to check
  @param used_key_parts        Return value for used key parts.


  @note
    used_key_parts is set to correct key parts used if return value != 0
    (On other cases, used_key_part may be changed)

  @retval
    1   key is ok.
  @retval
    0   Key can't be used
  @retval
    -1   Reverse key can be used
*/

static int test_if_order_by_key(order_st *order, Table *table, uint32_t idx,
				uint32_t *used_key_parts)
{
  KEY_PART_INFO *key_part,*key_part_end;
  key_part=table->key_info[idx].key_part;
  key_part_end=key_part+table->key_info[idx].key_parts;
  key_part_map const_key_parts=table->const_key_parts[idx];
  int reverse=0;
  bool on_primary_key= false;

  for (; order ; order=order->next, const_key_parts>>=1)
  {
    Field *field=((Item_field*) (*order->item)->real_item())->field;
    int flag;

    /*
      Skip key parts that are constants in the WHERE clause.
      These are already skipped in the order_st BY by const_expression_in_where()
    */
    for (; const_key_parts & 1 ; const_key_parts>>= 1)
      key_part++;

    if (key_part == key_part_end)
    {
      /*
        We are at the end of the key. Check if the engine has the primary
        key as a suffix to the secondary keys. If it has continue to check
        the primary key as a suffix.
      */
      if (!on_primary_key &&
          (table->file->ha_table_flags() & HA_PRIMARY_KEY_IN_READ_INDEX) &&
          table->s->primary_key != MAX_KEY)
      {
        on_primary_key= true;
        key_part= table->key_info[table->s->primary_key].key_part;
        key_part_end=key_part+table->key_info[table->s->primary_key].key_parts;
        const_key_parts=table->const_key_parts[table->s->primary_key];

        for (; const_key_parts & 1 ; const_key_parts>>= 1)
          key_part++;
        /*
         The primary and secondary key parts were all const (i.e. there's
         one row).  The sorting doesn't matter.
        */
        if (key_part == key_part_end && reverse == 0)
          return(1);
      }
      else
        return(0);
    }

    if (key_part->field != field)
      return(0);

    /* set flag to 1 if we can use read-next on key, else to -1 */
    flag= ((order->asc == !(key_part->key_part_flag & HA_REVERSE_SORT)) ?
           1 : -1);
    if (reverse && flag != reverse)
      return(0);
    reverse=flag;				// Remember if reverse
    key_part++;
  }
  *used_key_parts= on_primary_key ? table->key_info[idx].key_parts :
    (uint32_t) (key_part - table->key_info[idx].key_part);
  if (reverse == -1 && !(table->file->index_flags(idx, *used_key_parts-1, 1) &
                         HA_READ_PREV))
    reverse= 0;                                 // Index can't be used
  return(reverse);
}


/**
  Test if a second key is the subkey of the first one.

  @param key_part              First key parts
  @param ref_key_part          Second key parts
  @param ref_key_part_end      Last+1 part of the second key

  @note
    Second key MUST be shorter than the first one.

  @retval
    1	is a subkey
  @retval
    0	no sub key
*/

inline bool
is_subkey(KEY_PART_INFO *key_part, KEY_PART_INFO *ref_key_part,
	  KEY_PART_INFO *ref_key_part_end)
{
  for (; ref_key_part < ref_key_part_end; key_part++, ref_key_part++)
    if (!key_part->field->eq(ref_key_part->field))
      return 0;
  return 1;
}

/**
  Test if we can use one of the 'usable_keys' instead of 'ref' key
  for sorting.

  @param ref			Number of key, used for WHERE clause
  @param usable_keys		Keys for testing

  @return
    - MAX_KEY			If we can't use other key
    - the number of found key	Otherwise
*/

static uint
test_if_subkey(order_st *order, Table *table, uint32_t ref, uint32_t ref_key_parts,
	       const key_map *usable_keys)
{
  uint32_t nr;
  uint32_t min_length= UINT32_MAX;
  uint32_t best= MAX_KEY;
  uint32_t not_used;
  KEY_PART_INFO *ref_key_part= table->key_info[ref].key_part;
  KEY_PART_INFO *ref_key_part_end= ref_key_part + ref_key_parts;

  for (nr= 0 ; nr < table->s->keys ; nr++)
  {
    if (usable_keys->test(nr) &&
	table->key_info[nr].key_length < min_length &&
	table->key_info[nr].key_parts >= ref_key_parts &&
	is_subkey(table->key_info[nr].key_part, ref_key_part,
		  ref_key_part_end) &&
	test_if_order_by_key(order, table, nr, &not_used))
    {
      min_length= table->key_info[nr].key_length;
      best= nr;
    }
  }
  return best;
}


/**
  Check if GROUP BY/DISTINCT can be optimized away because the set is
  already known to be distinct.

  Used in removing the GROUP BY/DISTINCT of the following types of
  statements:
  @code
    SELECT [DISTINCT] <unique_key_cols>... FROM <single_table_ref>
      [GROUP BY <unique_key_cols>,...]
  @endcode

    If (a,b,c is distinct)
    then <any combination of a,b,c>,{whatever} is also distinct

    This function checks if all the key parts of any of the unique keys
    of the table are referenced by a list : either the select list
    through find_field_in_item_list or GROUP BY list through
    find_field_in_order_list.
    If the above holds and the key parts cannot contain NULLs then we
    can safely remove the GROUP BY/DISTINCT,
    as no result set can be more distinct than an unique key.

  @param table                The table to operate on.
  @param find_func            function to iterate over the list and search
                              for a field

  @retval
    1                    found
  @retval
    0                    not found.
*/

static bool
list_contains_unique_index(Table *table,
                          bool (*find_func) (Field *, void *), void *data)
{
  for (uint32_t keynr= 0; keynr < table->s->keys; keynr++)
  {
    if (keynr == table->s->primary_key ||
         (table->key_info[keynr].flags & HA_NOSAME))
    {
      KEY *keyinfo= table->key_info + keynr;
      KEY_PART_INFO *key_part, *key_part_end;

      for (key_part=keyinfo->key_part,
           key_part_end=key_part+ keyinfo->key_parts;
           key_part < key_part_end;
           key_part++)
      {
        if (key_part->field->maybe_null() ||
            !find_func(key_part->field, data))
          break;
      }
      if (key_part == key_part_end)
        return 1;
    }
  }
  return 0;
}


/**
  Helper function for list_contains_unique_index.
  Find a field reference in a list of order_st structures.
  Finds a direct reference of the Field in the list.

  @param field                The field to search for.
  @param data                 order_st *.The list to search in

  @retval
    1                    found
  @retval
    0                    not found.
*/

static bool
find_field_in_order_list (Field *field, void *data)
{
  order_st *group= (order_st *) data;
  bool part_found= 0;
  for (order_st *tmp_group= group; tmp_group; tmp_group=tmp_group->next)
  {
    Item *item= (*tmp_group->item)->real_item();
    if (item->type() == Item::FIELD_ITEM &&
        ((Item_field*) item)->field->eq(field))
    {
      part_found= 1;
      break;
    }
  }
  return part_found;
}


/**
  Helper function for list_contains_unique_index.
  Find a field reference in a dynamic list of Items.
  Finds a direct reference of the Field in the list.

  @param[in] field             The field to search for.
  @param[in] data              List<Item> *.The list to search in

  @retval
    1                    found
  @retval
    0                    not found.
*/

static bool
find_field_in_item_list (Field *field, void *data)
{
  List<Item> *fields= (List<Item> *) data;
  bool part_found= 0;
  List_iterator<Item> li(*fields);
  Item *item;

  while ((item= li++))
  {
    if (item->type() == Item::FIELD_ITEM &&
        ((Item_field*) item)->field->eq(field))
    {
      part_found= 1;
      break;
    }
  }
  return part_found;
}


/**
  Test if we can skip the order_st BY by using an index.

  SYNOPSIS
    test_if_skip_sort_order()
      tab
      order
      select_limit
      no_changes
      map

  If we can use an index, the JOIN_TAB / tab->select struct
  is changed to use the index.

  The index must cover all fields in <order>, or it will not be considered.

  @todo
    - sergeyp: Results of all index merge selects actually are ordered
    by clustered PK values.

  @retval
    0    We have to use filesort to do the sorting
  @retval
    1    We can use an index.
*/

static bool
test_if_skip_sort_order(JOIN_TAB *tab,order_st *order,ha_rows select_limit,
			bool no_changes, const key_map *map)
{
  int32_t ref_key;
  uint32_t ref_key_parts;
  int order_direction;
  uint32_t used_key_parts;
  Table *table=tab->table;
  SQL_SELECT *select=tab->select;
  key_map usable_keys;
  QUICK_SELECT_I *save_quick= 0;

  /*
    Keys disabled by ALTER Table ... DISABLE KEYS should have already
    been taken into account.
  */
  usable_keys= *map;

  for (order_st *tmp_order=order; tmp_order ; tmp_order=tmp_order->next)
  {
    Item *item= (*tmp_order->item)->real_item();
    if (item->type() != Item::FIELD_ITEM)
    {
      usable_keys.reset();
      return(0);
    }
    usable_keys&= ((Item_field*) item)->field->part_of_sortkey;
    if (usable_keys.none())
      return(0);					// No usable keys
  }

  ref_key= -1;
  /* Test if constant range in WHERE */
  if (tab->ref.key >= 0 && tab->ref.key_parts)
  {
    ref_key=	   tab->ref.key;
    ref_key_parts= tab->ref.key_parts;
    if (tab->type == JT_REF_OR_NULL)
      return(0);
  }
  else if (select && select->quick)		// Range found by opt_range
  {
    int quick_type= select->quick->get_type();
    save_quick= select->quick;
    /*
      assume results are not ordered when index merge is used
      TODO: sergeyp: Results of all index merge selects actually are ordered
      by clustered PK values.
    */

    if (quick_type == QUICK_SELECT_I::QS_TYPE_INDEX_MERGE ||
        quick_type == QUICK_SELECT_I::QS_TYPE_ROR_UNION ||
        quick_type == QUICK_SELECT_I::QS_TYPE_ROR_INTERSECT)
      return(0);
    ref_key=	   select->quick->index;
    ref_key_parts= select->quick->used_key_parts;
  }

  if (ref_key >= 0)
  {
    /*
      We come here when there is a REF key.
    */
    if (! usable_keys.test(ref_key))
    {
      /*
	We come here when ref_key is not among usable_keys
      */
      uint32_t new_ref_key;
      /*
	If using index only read, only consider other possible index only
	keys
      */
      if (table->covering_keys.test(ref_key))
	usable_keys&= table->covering_keys;
      if (tab->pre_idx_push_select_cond)
        tab->select_cond= tab->select->cond= tab->pre_idx_push_select_cond;
      if ((new_ref_key= test_if_subkey(order, table, ref_key, ref_key_parts,
				       &usable_keys)) < MAX_KEY)
      {
	/* Found key that can be used to retrieve data in sorted order */
	if (tab->ref.key >= 0)
	{
          /*
            We'll use ref access method on key new_ref_key. In general case
            the index search tuple for new_ref_key will be different (e.g.
            when one index is defined as (part1, part2, ...) and another as
            (part1, part2(N), ...) and the WHERE clause contains
            "part1 = const1 AND part2=const2".
            So we build tab->ref from scratch here.
          */
          KEYUSE *keyuse= tab->keyuse;
          while (keyuse->key != new_ref_key && keyuse->table == tab->table)
            keyuse++;

          if (create_ref_for_key(tab->join, tab, keyuse,
                                 tab->join->const_table_map))
            return(0);
	}
	else
	{
          /*
            The range optimizer constructed QUICK_RANGE for ref_key, and
            we want to use instead new_ref_key as the index. We can't
            just change the index of the quick select, because this may
            result in an incosistent QUICK_SELECT object. Below we
            create a new QUICK_SELECT from scratch so that all its
            parameres are set correctly by the range optimizer.
           */
          key_map new_ref_key_map;
          new_ref_key_map.reset();  // Force the creation of quick select
          new_ref_key_map.set(new_ref_key); // only for new_ref_key.

          if (select->test_quick_select(tab->join->session, new_ref_key_map, 0,
                                        (tab->join->select_options &
                                         OPTION_FOUND_ROWS) ?
                                        HA_POS_ERROR :
                                        tab->join->unit->select_limit_cnt,0,
                                        true) <=
              0)
            return(0);
	}
        ref_key= new_ref_key;
      }
    }
    /* Check if we get the rows in requested sorted order by using the key */
    if (usable_keys.test(ref_key) &&
        (order_direction= test_if_order_by_key(order,table,ref_key,
					       &used_key_parts)))
      goto check_reverse_order;
  }
  {
    /*
      Check whether there is an index compatible with the given order
      usage of which is cheaper than usage of the ref_key index (ref_key>=0)
      or a table scan.
      It may be the case if order_st/GROUP BY is used with LIMIT.
    */
    uint32_t nr;
    key_map keys;
    uint32_t best_key_parts= 0;
    int best_key_direction= 0;
    ha_rows best_records= 0;
    double read_time;
    int best_key= -1;
    bool is_best_covering= false;
    double fanout= 1;
    JOIN *join= tab->join;
    uint32_t tablenr= tab - join->join_tab;
    ha_rows table_records= table->file->stats.records;
    bool group= join->group && order == join->group_list;

    /*
      If not used with LIMIT, only use keys if the whole query can be
      resolved with a key;  This is because filesort() is usually faster than
      retrieving all rows through an index.
    */
    if (select_limit >= table_records)
    {
      /*
        filesort() and join cache are usually faster than reading in
        index order and not using join cache
        */
      if (tab->type == JT_ALL && tab->join->tables > tab->join->const_tables + 1)
        return(0);
      keys= *table->file->keys_to_use_for_scanning();
      keys|= table->covering_keys;

      /*
	We are adding here also the index specified in FORCE INDEX clause,
	if any.
        This is to allow users to use index in order_st BY.
      */
      if (table->force_index)
	keys|= (group ? table->keys_in_use_for_group_by :
                           table->keys_in_use_for_order_by);
      keys&= usable_keys;
    }
    else
      keys= usable_keys;

    read_time= join->best_positions[tablenr].read_time;
    for (uint32_t i= tablenr+1; i < join->tables; i++)
      fanout*= join->best_positions[i].records_read; // fanout is always >= 1

    for (nr=0; nr < table->s->keys ; nr++)
    {
      int direction;
      if (keys.test(nr) &&
          (direction= test_if_order_by_key(order, table, nr, &used_key_parts)))
      {
        bool is_covering= table->covering_keys.test(nr) || (nr == table->s->primary_key && table->file->primary_key_is_clustered());

        /*
          Don't use an index scan with order_st BY without limit.
          For GROUP BY without limit always use index scan
          if there is a suitable index.
          Why we hold to this asymmetry hardly can be explained
          rationally. It's easy to demonstrate that using
          temporary table + filesort could be cheaper for grouping
          queries too.
	*/
        if (is_covering ||
            select_limit != HA_POS_ERROR ||
            (ref_key < 0 && (group || table->force_index)))
        {
          double rec_per_key;
          double index_scan_time;
          KEY *keyinfo= tab->table->key_info+nr;
          if (select_limit == HA_POS_ERROR)
            select_limit= table_records;
          if (group)
          {
            rec_per_key= keyinfo->rec_per_key[used_key_parts-1];
            set_if_bigger(rec_per_key, 1.0);
            /*
              With a grouping query each group containing on average
              rec_per_key records produces only one row that will
              be included into the result set.
	    */
            if (select_limit > table_records/rec_per_key)
                select_limit= table_records;
            else
              select_limit= (ha_rows) (select_limit*rec_per_key);
          }
          /*
            If tab=tk is not the last joined table tn then to get first
            L records from the result set we can expect to retrieve
            only L/fanout(tk,tn) where fanout(tk,tn) says how many
            rows in the record set on average will match each row tk.
            Usually our estimates for fanouts are too pessimistic.
            So the estimate for L/fanout(tk,tn) will be too optimistic
            and as result we'll choose an index scan when using ref/range
            access + filesort will be cheaper.
	  */
          select_limit= (ha_rows) (select_limit < fanout ?
                                   1 : select_limit/fanout);
          /*
            We assume that each of the tested indexes is not correlated
            with ref_key. Thus, to select first N records we have to scan
            N/selectivity(ref_key) index entries.
            selectivity(ref_key) = #scanned_records/#table_records =
            table->quick_condition_rows/table_records.
            In any case we can't select more than #table_records.
            N/(table->quick_condition_rows/table_records) > table_records
            <=> N > table->quick_condition_rows.
          */
          if (select_limit > table->quick_condition_rows)
            select_limit= table_records;
          else
            select_limit= (ha_rows) (select_limit *
                                     (double) table_records /
                                      table->quick_condition_rows);
          rec_per_key= keyinfo->rec_per_key[keyinfo->key_parts-1];
          set_if_bigger(rec_per_key, 1.0);
          /*
            Here we take into account the fact that rows are
            accessed in sequences rec_per_key records in each.
            Rows in such a sequence are supposed to be ordered
            by rowid/primary key. When reading the data
            in a sequence we'll touch not more pages than the
            table file contains.
            TODO. Use the formula for a disk sweep sequential access
            to calculate the cost of accessing data rows for one
            index entry.
	  */
          index_scan_time= select_limit/rec_per_key *
	                   cmin(rec_per_key, table->file->scan_time());
          if (is_covering || (ref_key < 0 && (group || table->force_index)) ||
              index_scan_time < read_time)
          {
            ha_rows quick_records= table_records;
            if (is_best_covering && !is_covering)
              continue;
            if (table->quick_keys.test(nr))
              quick_records= table->quick_rows[nr];
            if (best_key < 0 ||
                (select_limit <= cmin(quick_records,best_records) ?
                 keyinfo->key_parts < best_key_parts :
                 quick_records < best_records))
            {
              best_key= nr;
              best_key_parts= keyinfo->key_parts;
              best_records= quick_records;
              is_best_covering= is_covering;
              best_key_direction= direction;
            }
          }
	}
      }
    }
    if (best_key >= 0)
    {
      bool quick_created= false;
      if (table->quick_keys.test(best_key) && best_key != ref_key)
      {
        key_map test_map;
        test_map.reset();       // Force the creation of quick select
        test_map.set(best_key); // only best_key.
        quick_created=
          select->test_quick_select(join->session, test_map, 0,
                                    join->select_options & OPTION_FOUND_ROWS ?
                                    HA_POS_ERROR :
                                    join->unit->select_limit_cnt,
                                    true, false) > 0;
      }
      if (!no_changes)
      {
        if (!quick_created)
	{
          tab->index= best_key;
          tab->read_first_record= best_key_direction > 0 ?
                                  join_read_first:join_read_last;
          tab->type=JT_NEXT;           // Read with index_first(), index_next()
          if (select && select->quick)
          {
            delete select->quick;
            select->quick= 0;
          }
          if (table->covering_keys.test(best_key))
          {
            table->key_read=1;
            table->file->extra(HA_EXTRA_KEYREAD);
          }
          table->file->ha_index_or_rnd_end();
          if (join->select_options & SELECT_DESCRIBE)
          {
            tab->ref.key= -1;
            tab->ref.key_parts= 0;
            if (select_limit < table_records)
              tab->limit= select_limit;
          }
        }
        else if (tab->type != JT_ALL)
        {
          /*
            We're about to use a quick access to the table.
            We need to change the access method so as the quick access
            method is actually used.
          */
          assert(tab->select->quick);
          tab->type=JT_ALL;
          tab->use_quick=1;
          tab->ref.key= -1;
          tab->ref.key_parts=0;		// Don't use ref key.
          tab->read_first_record= join_init_read_record;
          /*
            TODO: update the number of records in join->best_positions[tablenr]
          */
        }
      }
      used_key_parts= best_key_parts;
      order_direction= best_key_direction;
    }
    else
      return(0);
  }

check_reverse_order:
  if (order_direction == -1)		// If order_st BY ... DESC
  {
    if (select && select->quick)
    {
      /*
	Don't reverse the sort order, if it's already done.
        (In some cases test_if_order_by_key() can be called multiple times
      */
      if (!select->quick->reverse_sorted())
      {
        QUICK_SELECT_DESC *tmp;
        bool error= false;
        int quick_type= select->quick->get_type();
        if (quick_type == QUICK_SELECT_I::QS_TYPE_INDEX_MERGE ||
            quick_type == QUICK_SELECT_I::QS_TYPE_ROR_INTERSECT ||
            quick_type == QUICK_SELECT_I::QS_TYPE_ROR_UNION ||
            quick_type == QUICK_SELECT_I::QS_TYPE_GROUP_MIN_MAX)
        {
          tab->limit= 0;
          select->quick= save_quick;
          return(0);                   // Use filesort
        }

        /* order_st BY range_key DESC */
	tmp= new QUICK_SELECT_DESC((QUICK_RANGE_SELECT*)(select->quick),
                                    used_key_parts, &error);
	if (!tmp || error)
	{
	  delete tmp;
          select->quick= save_quick;
          tab->limit= 0;
	  return(0);		// Reverse sort not supported
	}
	select->quick=tmp;
      }
    }
    else if (tab->type != JT_NEXT &&
             tab->ref.key >= 0 && tab->ref.key_parts <= used_key_parts)
    {
      /*
	SELECT * FROM t1 WHERE a=1 order_st BY a DESC,b DESC

	Use a traversal function that starts by reading the last row
	with key part (A) and then traverse the index backwards.
      */
      tab->read_first_record= join_read_last_key;
      tab->read_record.read_record= join_read_prev_same;
    }
  }
  else if (select && select->quick)
    select->quick->sorted= 1;
  return(1);
}


/*
  If not selecting by given key, create an index how records should be read

  SYNOPSIS
   create_sort_index()
     session		Thread handler
     tab		Table to sort (in join structure)
     order		How table should be sorted
     filesort_limit	Max number of rows that needs to be sorted
     select_limit	Max number of rows in final output
		        Used to decide if we should use index or not
     is_order_by        true if we are sorting on order_st BY, false if GROUP BY
                        Used to decide if we should use index or not


  IMPLEMENTATION
   - If there is an index that can be used, 'tab' is modified to use
     this index.
   - If no index, create with filesort() an index file that can be used to
     retrieve rows in order (should be done with 'read_record').
     The sorted data is stored in tab->table and will be freed when calling
     free_io_cache(tab->table).

  RETURN VALUES
    0		ok
    -1		Some fatal error
    1		No records
*/

static int
create_sort_index(Session *session, JOIN *join, order_st *order,
		  ha_rows filesort_limit, ha_rows select_limit,
                  bool is_order_by)
{
  uint32_t length= 0;
  ha_rows examined_rows;
  Table *table;
  SQL_SELECT *select;
  JOIN_TAB *tab;

  if (join->tables == join->const_tables)
    return(0);				// One row, no need to sort
  tab=    join->join_tab + join->const_tables;
  table=  tab->table;
  select= tab->select;

  /*
    When there is SQL_BIG_RESULT do not sort using index for GROUP BY,
    and thus force sorting on disk unless a group min-max optimization
    is going to be used as it is applied now only for one table queries
    with covering indexes.
  */
  if ((order != join->group_list ||
       !(join->select_options & SELECT_BIG_RESULT) ||
       (select && select->quick && (select->quick->get_type() == QUICK_SELECT_I::QS_TYPE_GROUP_MIN_MAX))) &&
      test_if_skip_sort_order(tab,order,select_limit,0,
                              is_order_by ?  &table->keys_in_use_for_order_by :
                              &table->keys_in_use_for_group_by))
    return(0);
  for (order_st *ord= join->order; ord; ord= ord->next)
    length++;
  if (!(join->sortorder=
        make_unireg_sortorder(order, &length, join->sortorder)))
    goto err;				/* purecov: inspected */

  table->sort.io_cache= new IO_CACHE;
  memset(table->sort.io_cache, 0, sizeof(IO_CACHE));
  table->status=0;				// May be wrong if quick_select

  // If table has a range, move it to select
  if (select && !select->quick && tab->ref.key >= 0)
  {
    if (tab->quick)
    {
      select->quick=tab->quick;
      tab->quick=0;
      /*
        We can only use 'Only index' if quick key is same as ref_key
        and in index_merge 'Only index' cannot be used
      */
      if (table->key_read && ((uint32_t) tab->ref.key != select->quick->index))
      {
	table->key_read=0;
	table->file->extra(HA_EXTRA_NO_KEYREAD);
      }
    }
    else
    {
      /*
	We have a ref on a const;  Change this to a range that filesort
	can use.
	For impossible ranges (like when doing a lookup on NULL on a NOT NULL
	field, quick will contain an empty record set.
      */
      if (!(select->quick= (get_quick_select_for_ref(session, table, &tab->ref,
                                                     tab->found_records))))
	goto err;
    }
  }

  /* Fill schema tables with data before filesort if it's necessary */
  if ((join->select_lex->options & OPTION_SCHEMA_TABLE) &&
      get_schema_tables_result(join, PROCESSED_BY_CREATE_SORT_INDEX))
    goto err;

  if (table->s->tmp_table)
    table->file->info(HA_STATUS_VARIABLE);	// Get record count
  table->sort.found_records=filesort(session, table,join->sortorder, length,
                                     select, filesort_limit, 0,
                                     &examined_rows);
  tab->records= table->sort.found_records;	// For SQL_CALC_ROWS
  if (select)
  {
    select->cleanup();				// filesort did select
    tab->select= 0;
  }
  tab->select_cond=0;
  tab->last_inner= 0;
  tab->first_unmatched= 0;
  tab->type=JT_ALL;				// Read with normal read_record
  tab->read_first_record= join_init_read_record;
  tab->join->examined_rows+=examined_rows;
  if (table->key_read)				// Restore if we used indexes
  {
    table->key_read=0;
    table->file->extra(HA_EXTRA_NO_KEYREAD);
  }
  return(table->sort.found_records == HA_POS_ERROR);
err:
  return(-1);
}

static bool copy_blobs(Field **ptr)
{
  for (; *ptr ; ptr++)
  {
    if ((*ptr)->flags & BLOB_FLAG)
      if (((Field_blob *) (*ptr))->copy())
	return 1;				// Error
  }
  return 0;
}

static void free_blobs(Field **ptr)
{
  for (; *ptr ; ptr++)
  {
    if ((*ptr)->flags & BLOB_FLAG)
      ((Field_blob *) (*ptr))->free();
  }
}


static int
remove_duplicates(JOIN *join, Table *entry,List<Item> &fields, Item *having)
{
  int error;
  uint32_t reclength,offset;
  uint32_t field_count;
  Session *session= join->session;

  entry->reginfo.lock_type=TL_WRITE;

  /* Calculate how many saved fields there is in list */
  field_count=0;
  List_iterator<Item> it(fields);
  Item *item;
  while ((item=it++))
  {
    if (item->get_tmp_table_field() && ! item->const_item())
      field_count++;
  }

  if (!field_count && !(join->select_options & OPTION_FOUND_ROWS) && !having)
  {                    // only const items with no OPTION_FOUND_ROWS
    join->unit->select_limit_cnt= 1;		// Only send first row
    return(0);
  }
  Field **first_field=entry->field+entry->s->fields - field_count;
  offset= (field_count ?
           entry->field[entry->s->fields - field_count]->
           offset(entry->record[0]) : 0);
  reclength= entry->s->reclength-offset;

  free_io_cache(entry);				// Safety
  entry->file->info(HA_STATUS_VARIABLE);
  if (entry->s->db_type() == heap_engine ||
      (!entry->s->blob_fields &&
       ((ALIGN_SIZE(reclength) + HASH_OVERHEAD) * entry->file->stats.records <
	session->variables.sortbuff_size)))
    error= remove_dup_with_hash_index(join->session, entry,
				     field_count, first_field,
				     reclength, having);
  else
    error= remove_dup_with_compare(join->session, entry, first_field, offset,
				  having);

  free_blobs(first_field);
  return(error);
}


static int remove_dup_with_compare(Session *session, Table *table, Field **first_field,
				   uint32_t offset, Item *having)
{
  handler *file=table->file;
  char *org_record,*new_record;
  unsigned char *record;
  int error;
  uint32_t reclength= table->s->reclength-offset;

  org_record=(char*) (record=table->record[0])+offset;
  new_record=(char*) table->record[1]+offset;

  file->ha_rnd_init(1);
  error=file->rnd_next(record);
  for (;;)
  {
    if (session->killed)
    {
      session->send_kill_message();
      error=0;
      goto err;
    }
    if (error)
    {
      if (error == HA_ERR_RECORD_DELETED)
	continue;
      if (error == HA_ERR_END_OF_FILE)
	break;
      goto err;
    }
    if (having && !having->val_int())
    {
      if ((error=file->ha_delete_row(record)))
	goto err;
      error=file->rnd_next(record);
      continue;
    }
    if (copy_blobs(first_field))
    {
      my_message(ER_OUTOFMEMORY, ER(ER_OUTOFMEMORY), MYF(0));
      error=0;
      goto err;
    }
    memcpy(new_record,org_record,reclength);

    /* Read through rest of file and mark duplicated rows deleted */
    bool found=0;
    for (;;)
    {
      if ((error=file->rnd_next(record)))
      {
	if (error == HA_ERR_RECORD_DELETED)
	  continue;
	if (error == HA_ERR_END_OF_FILE)
	  break;
	goto err;
      }
      if (table->compare_record(first_field) == 0)
      {
	if ((error=file->ha_delete_row(record)))
	  goto err;
      }
      else if (!found)
      {
	found=1;
	file->position(record);	// Remember position
      }
    }
    if (!found)
      break;					// End of file
    /* Restart search on next row */
    error=file->restart_rnd_next(record,file->ref);
  }

  file->extra(HA_EXTRA_NO_CACHE);
  return(0);
err:
  file->extra(HA_EXTRA_NO_CACHE);
  if (error)
    file->print_error(error,MYF(0));
  return(1);
}


/**
  Generate a hash index for each row to quickly find duplicate rows.

  @note
    Note that this will not work on tables with blobs!
*/

static int remove_dup_with_hash_index(Session *session, Table *table,
				      uint32_t field_count,
				      Field **first_field,
				      uint32_t key_length,
				      Item *having)
{
  unsigned char *key_buffer, *key_pos, *record=table->record[0];
  int error;
  handler *file= table->file;
  uint32_t extra_length= ALIGN_SIZE(key_length)-key_length;
  uint32_t *field_lengths,*field_length;
  HASH hash;

  if (!my_multi_malloc(MYF(MY_WME),
		       &key_buffer,
		       (uint32_t) ((key_length + extra_length) *
			       (long) file->stats.records),
		       &field_lengths,
		       (uint32_t) (field_count*sizeof(*field_lengths)),
		       NULL))
    return(1);

  {
    Field **ptr;
    uint32_t total_length= 0;
    for (ptr= first_field, field_length=field_lengths ; *ptr ; ptr++)
    {
      uint32_t length= (*ptr)->sort_length();
      (*field_length++)= length;
      total_length+= length;
    }
    assert(total_length <= key_length);
    key_length= total_length;
    extra_length= ALIGN_SIZE(key_length)-key_length;
  }

  if (hash_init(&hash, &my_charset_bin, (uint32_t) file->stats.records, 0,
		key_length, (hash_get_key) 0, 0, 0))
  {
    free((char*) key_buffer);
    return(1);
  }

  file->ha_rnd_init(1);
  key_pos=key_buffer;
  for (;;)
  {
    unsigned char *org_key_pos;
    if (session->killed)
    {
      session->send_kill_message();
      error=0;
      goto err;
    }
    if ((error=file->rnd_next(record)))
    {
      if (error == HA_ERR_RECORD_DELETED)
	continue;
      if (error == HA_ERR_END_OF_FILE)
	break;
      goto err;
    }
    if (having && !having->val_int())
    {
      if ((error=file->ha_delete_row(record)))
	goto err;
      continue;
    }

    /* copy fields to key buffer */
    org_key_pos= key_pos;
    field_length=field_lengths;
    for (Field **ptr= first_field ; *ptr ; ptr++)
    {
      (*ptr)->sort_string(key_pos,*field_length);
      key_pos+= *field_length++;
    }
    /* Check if it exists before */
    if (hash_search(&hash, org_key_pos, key_length))
    {
      /* Duplicated found ; Remove the row */
      if ((error=file->ha_delete_row(record)))
	goto err;
    }
    else
      (void) my_hash_insert(&hash, org_key_pos);
    key_pos+=extra_length;
  }
  free((char*) key_buffer);
  hash_free(&hash);
  file->extra(HA_EXTRA_NO_CACHE);
  (void) file->ha_rnd_end();
  return(0);

err:
  free((char*) key_buffer);
  hash_free(&hash);
  file->extra(HA_EXTRA_NO_CACHE);
  (void) file->ha_rnd_end();
  if (error)
    file->print_error(error,MYF(0));
  return(1);
}


SORT_FIELD *make_unireg_sortorder(order_st *order, uint32_t *length,
                                  SORT_FIELD *sortorder)
{
  uint32_t count;
  SORT_FIELD *sort,*pos;

  count=0;
  for (order_st *tmp = order; tmp; tmp=tmp->next)
    count++;
  if (!sortorder)
    sortorder= (SORT_FIELD*) sql_alloc(sizeof(SORT_FIELD) *
                                       (cmax(count, *length) + 1));
  pos= sort= sortorder;

  if (!pos)
    return 0;

  for (;order;order=order->next,pos++)
  {
    Item *item= order->item[0]->real_item();
    pos->field= 0; pos->item= 0;
    if (item->type() == Item::FIELD_ITEM)
      pos->field= ((Item_field*) item)->field;
    else if (item->type() == Item::SUM_FUNC_ITEM && !item->const_item())
      pos->field= ((Item_sum*) item)->get_tmp_table_field();
    else if (item->type() == Item::COPY_STR_ITEM)
    {						// Blob patch
      pos->item= ((Item_copy_string*) item)->item;
    }
    else
      pos->item= *order->item;
    pos->reverse=! order->asc;
  }
  *length=count;
  return(sort);
}


/*****************************************************************************
  Fill join cache with packed records
  Records are stored in tab->cache.buffer and last record in
  last record is stored with pointers to blobs to support very big
  records
******************************************************************************/

static int
join_init_cache(Session *session,JOIN_TAB *tables,uint32_t table_count)
{
  register unsigned int i;
  unsigned int length, blobs;
  size_t size;
  CACHE_FIELD *copy,**blob_ptr;
  JOIN_CACHE  *cache;
  JOIN_TAB *join_tab;

  cache= &tables[table_count].cache;
  cache->fields=blobs=0;

  join_tab=tables;
  for (i=0 ; i < table_count ; i++,join_tab++)
  {
    if (!join_tab->used_fieldlength)		/* Not calced yet */
      calc_used_field_length(session, join_tab);
    cache->fields+=join_tab->used_fields;
    blobs+=join_tab->used_blobs;

    /* SemiJoinDuplicateElimination: reserve space for rowid */
    if (join_tab->rowid_keep_flags & JOIN_TAB::KEEP_ROWID)
    {
      cache->fields++;
      join_tab->used_fieldlength += join_tab->table->file->ref_length;
    }
  }
  if (!(cache->field=(CACHE_FIELD*)
	sql_alloc(sizeof(CACHE_FIELD)*(cache->fields+table_count*2)+(blobs+1)*

		  sizeof(CACHE_FIELD*))))
  {
    free((unsigned char*) cache->buff);		/* purecov: inspected */
    cache->buff=0;				/* purecov: inspected */
    return(1);				/* purecov: inspected */
  }
  copy=cache->field;
  blob_ptr=cache->blob_ptr=(CACHE_FIELD**)
    (cache->field+cache->fields+table_count*2);

  length=0;
  for (i=0 ; i < table_count ; i++)
  {
    uint32_t null_fields=0, used_fields;
    Field **f_ptr,*field;
    for (f_ptr=tables[i].table->field,used_fields=tables[i].used_fields ;
	 used_fields ;
	 f_ptr++)
    {
      field= *f_ptr;
      if (field->isReadSet())
      {
	used_fields--;
	length+=field->fill_cache_field(copy);
	if (copy->blob_field)
	  (*blob_ptr++)=copy;
	if (field->maybe_null())
	  null_fields++;
        copy->get_rowid= NULL;
	copy++;
      }
    }
    /* Copy null bits from table */
    if (null_fields && tables[i].table->getNullFields())
    {						/* must copy null bits */
      copy->str= tables[i].table->null_flags;
      copy->length= tables[i].table->s->null_bytes;
      copy->strip=0;
      copy->blob_field=0;
      copy->get_rowid= NULL;
      length+=copy->length;
      copy++;
      cache->fields++;
    }
    /* If outer join table, copy null_row flag */
    if (tables[i].table->maybe_null)
    {
      copy->str= (unsigned char*) &tables[i].table->null_row;
      copy->length=sizeof(tables[i].table->null_row);
      copy->strip=0;
      copy->blob_field=0;
      copy->get_rowid= NULL;
      length+=copy->length;
      copy++;
      cache->fields++;
    }
    /* SemiJoinDuplicateElimination: Allocate space for rowid if needed */
    if (tables[i].rowid_keep_flags & JOIN_TAB::KEEP_ROWID)
    {
      copy->str= tables[i].table->file->ref;
      copy->length= tables[i].table->file->ref_length;
      copy->strip=0;
      copy->blob_field=0;
      copy->get_rowid= NULL;
      if (tables[i].rowid_keep_flags & JOIN_TAB::CALL_POSITION)
      {
        /* We will need to call h->position(): */
        copy->get_rowid= tables[i].table;
        /* And those after us won't have to: */
        tables[i].rowid_keep_flags &=  ~((int)JOIN_TAB::CALL_POSITION);
      }
      copy++;
    }
  }

  cache->length=length+blobs*sizeof(char*);
  cache->blobs=blobs;
  *blob_ptr= NULL;					/* End sequentel */
  size= max((size_t)session->variables.join_buff_size,
            (size_t)cache->length);
  if (!(cache->buff=(unsigned char*) malloc(size)))
    return 1;				/* Don't use cache */ /* purecov: inspected */
  cache->end=cache->buff+size;
  reset_cache_write(cache);
  return 0;
}


static uint32_t used_blob_length(CACHE_FIELD **ptr)
{
  uint32_t length,blob_length;
  for (length=0 ; *ptr ; ptr++)
  {
    (*ptr)->blob_length=blob_length=(*ptr)->blob_field->get_length();
    length+=blob_length;
    (*ptr)->blob_field->get_ptr(&(*ptr)->str);
  }
  return length;
}


static bool
store_record_in_cache(JOIN_CACHE *cache)
{
  uint32_t length;
  unsigned char *pos;
  CACHE_FIELD *copy,*end_field;
  bool last_record;

  pos=cache->pos;
  end_field=cache->field+cache->fields;

  length=cache->length;
  if (cache->blobs)
    length+= used_blob_length(cache->blob_ptr);
  if ((last_record= (length + cache->length > (size_t) (cache->end - pos))))
    cache->ptr_record=cache->records;
  /*
    There is room in cache. Put record there
  */
  cache->records++;
  for (copy=cache->field ; copy < end_field; copy++)
  {
    if (copy->blob_field)
    {
      if (last_record)
      {
	copy->blob_field->get_image(pos, copy->length+sizeof(char*),
				    copy->blob_field->charset());
	pos+=copy->length+sizeof(char*);
      }
      else
      {
	copy->blob_field->get_image(pos, copy->length, // blob length
				    copy->blob_field->charset());
	memcpy(pos+copy->length,copy->str,copy->blob_length);  // Blob data
	pos+=copy->length+copy->blob_length;
      }
    }
    else
    {
      // SemiJoinDuplicateElimination: Get the rowid into table->ref:
      if (copy->get_rowid)
        copy->get_rowid->file->position(copy->get_rowid->record[0]);

      if (copy->strip)
      {
	unsigned char *str,*end;
	for (str=copy->str,end= str+copy->length;
	     end > str && end[-1] == ' ' ;
	     end--) ;
	length=(uint32_t) (end-str);
	memcpy(pos+2, str, length);
        int2store(pos, length);
	pos+= length+2;
      }
      else
      {
	memcpy(pos,copy->str,copy->length);
	pos+=copy->length;
      }
    }
  }
  cache->pos=pos;
  return last_record || (size_t) (cache->end - pos) < cache->length;
}


static void
reset_cache_read(JOIN_CACHE *cache)
{
  cache->record_nr=0;
  cache->pos=cache->buff;
}


static void reset_cache_write(JOIN_CACHE *cache)
{
  reset_cache_read(cache);
  cache->records= 0;
  cache->ptr_record= UINT32_MAX;
}


static void
read_cached_record(JOIN_TAB *tab)
{
  unsigned char *pos;
  uint32_t length;
  bool last_record;
  CACHE_FIELD *copy,*end_field;

  last_record=tab->cache.record_nr++ == tab->cache.ptr_record;
  pos=tab->cache.pos;
  for (copy=tab->cache.field,end_field=copy+tab->cache.fields ;
       copy < end_field;
       copy++)
  {
    if (copy->blob_field)
    {
      if (last_record)
      {
	copy->blob_field->set_image(pos, copy->length+sizeof(char*),
				    copy->blob_field->charset());
	pos+=copy->length+sizeof(char*);
      }
      else
      {
	copy->blob_field->set_ptr(pos, pos+copy->length);
	pos+=copy->length+copy->blob_field->get_length();
      }
    }
    else
    {
      if (copy->strip)
      {
        length= uint2korr(pos);
	memcpy(copy->str, pos+2, length);
	memset(copy->str+length, ' ', copy->length-length);
	pos+= 2 + length;
      }
      else
      {
	memcpy(copy->str,pos,copy->length);
	pos+=copy->length;
      }
    }
  }
  tab->cache.pos=pos;
  return;
}


/*
  eq_ref: Create the lookup key and check if it is the same as saved key

  SYNOPSIS
    cmp_buffer_with_ref()
      tab  Join tab of the accessed table

  DESCRIPTION
    Used by eq_ref access method: create the index lookup key and check if
    we've used this key at previous lookup (If yes, we don't need to repeat
    the lookup - the record has been already fetched)

  RETURN
    true   No cached record for the key, or failed to create the key (due to
           out-of-domain error)
    false  The created key is the same as the previous one (and the record
           is already in table->record)
*/

static bool
cmp_buffer_with_ref(JOIN_TAB *tab)
{
  bool no_prev_key;
  if (!tab->ref.disable_cache)
  {
    if (!(no_prev_key= tab->ref.key_err))
    {
      /* Previous access found a row. Copy its key */
      memcpy(tab->ref.key_buff2, tab->ref.key_buff, tab->ref.key_length);
    }
  }
  else
    no_prev_key= true;
  if ((tab->ref.key_err= cp_buffer_from_ref(tab->join->session, &tab->ref)) ||
      no_prev_key)
    return 1;
  return memcmp(tab->ref.key_buff2, tab->ref.key_buff, tab->ref.key_length)
    != 0;
}


bool
cp_buffer_from_ref(Session *session, TABLE_REF *ref)
{
  enum enum_check_fields save_count_cuted_fields= session->count_cuted_fields;
  session->count_cuted_fields= CHECK_FIELD_IGNORE;
  bool result= 0;

  for (store_key **copy=ref->key_copy ; *copy ; copy++)
  {
    if ((*copy)->copy() & 1)
    {
      result= 1;
      break;
    }
  }
  session->count_cuted_fields= save_count_cuted_fields;
  return result;
}


/*****************************************************************************
  Group and order functions
*****************************************************************************/

/**
  Resolve an order_st BY or GROUP BY column reference.

  Given a column reference (represented by 'order') from a GROUP BY or order_st
  BY clause, find the actual column it represents. If the column being
  resolved is from the GROUP BY clause, the procedure searches the SELECT
  list 'fields' and the columns in the FROM list 'tables'. If 'order' is from
  the order_st BY clause, only the SELECT list is being searched.

  If 'order' is resolved to an Item, then order->item is set to the found
  Item. If there is no item for the found column (that is, it was resolved
  into a table field), order->item is 'fixed' and is added to all_fields and
  ref_pointer_array.

  ref_pointer_array and all_fields are updated.

  @param[in] session		     Pointer to current thread structure
  @param[in,out] ref_pointer_array  All select, group and order by fields
  @param[in] tables                 List of tables to search in (usually
    FROM clause)
  @param[in] order                  Column reference to be resolved
  @param[in] fields                 List of fields to search in (usually
    SELECT list)
  @param[in,out] all_fields         All select, group and order by fields
  @param[in] is_group_field         True if order is a GROUP field, false if
    order_st by field

  @retval
    false if OK
  @retval
    true  if error occurred
*/

static bool
find_order_in_list(Session *session, Item **ref_pointer_array, TableList *tables,
                   order_st *order, List<Item> &fields, List<Item> &all_fields,
                   bool is_group_field)
{
  Item *order_item= *order->item; /* The item from the GROUP/order_st caluse. */
  Item::Type order_item_type;
  Item **select_item; /* The corresponding item from the SELECT clause. */
  Field *from_field;  /* The corresponding field from the FROM clause. */
  uint32_t counter;
  enum_resolution_type resolution;

  /*
    Local SP variables may be int but are expressions, not positions.
    (And they can't be used before fix_fields is called for them).
  */
  if (order_item->type() == Item::INT_ITEM && order_item->basic_const_item())
  {						/* Order by position */
    uint32_t count= (uint32_t) order_item->val_int();
    if (!count || count > fields.elements)
    {
      my_error(ER_BAD_FIELD_ERROR, MYF(0),
               order_item->full_name(), session->where);
      return true;
    }
    order->item= ref_pointer_array + count - 1;
    order->in_field_list= 1;
    order->counter= count;
    order->counter_used= 1;
    return false;
  }
  /* Lookup the current GROUP/order_st field in the SELECT clause. */
  select_item= find_item_in_list(order_item, fields, &counter,
                                 REPORT_EXCEPT_NOT_FOUND, &resolution);
  if (!select_item)
    return true; /* The item is not unique, or some other error occured. */


  /* Check whether the resolved field is not ambiguos. */
  if (select_item != not_found_item)
  {
    Item *view_ref= NULL;
    /*
      If we have found field not by its alias in select list but by its
      original field name, we should additionaly check if we have conflict
      for this name (in case if we would perform lookup in all tables).
    */
    if (resolution == RESOLVED_BEHIND_ALIAS && !order_item->fixed &&
        order_item->fix_fields(session, order->item))
      return true;

    /* Lookup the current GROUP field in the FROM clause. */
    order_item_type= order_item->type();
    from_field= (Field*) not_found_field;
    if ((is_group_field && order_item_type == Item::FIELD_ITEM) ||
        order_item_type == Item::REF_ITEM)
    {
      from_field= find_field_in_tables(session, (Item_ident*) order_item, tables,
                                       NULL, &view_ref, IGNORE_ERRORS, true,
                                       false);
      if (!from_field)
        from_field= (Field*) not_found_field;
    }

    if (from_field == not_found_field ||
        (from_field != view_ref_found ?
         /* it is field of base table => check that fields are same */
         ((*select_item)->type() == Item::FIELD_ITEM &&
          ((Item_field*) (*select_item))->field->eq(from_field)) :
         /*
           in is field of view table => check that references on translation
           table are same
         */
         ((*select_item)->type() == Item::REF_ITEM &&
          view_ref->type() == Item::REF_ITEM &&
          ((Item_ref *) (*select_item))->ref ==
          ((Item_ref *) view_ref)->ref)))
    {
      /*
        If there is no such field in the FROM clause, or it is the same field
        as the one found in the SELECT clause, then use the Item created for
        the SELECT field. As a result if there was a derived field that
        'shadowed' a table field with the same name, the table field will be
        chosen over the derived field.
      */
      order->item= ref_pointer_array + counter;
      order->in_field_list=1;
      return false;
    }
    else
    {
      /*
        There is a field with the same name in the FROM clause. This
        is the field that will be chosen. In this case we issue a
        warning so the user knows that the field from the FROM clause
        overshadows the column reference from the SELECT list.
      */
      push_warning_printf(session, DRIZZLE_ERROR::WARN_LEVEL_WARN, ER_NON_UNIQ_ERROR,
                          ER(ER_NON_UNIQ_ERROR),
                          ((Item_ident*) order_item)->field_name,
                          current_session->where);
    }
  }

  order->in_field_list=0;
  /*
    The call to order_item->fix_fields() means that here we resolve
    'order_item' to a column from a table in the list 'tables', or to
    a column in some outer query. Exactly because of the second case
    we come to this point even if (select_item == not_found_item),
    inspite of that fix_fields() calls find_item_in_list() one more
    time.

    We check order_item->fixed because Item_func_group_concat can put
    arguments for which fix_fields already was called.
  */
  if (!order_item->fixed &&
      (order_item->fix_fields(session, order->item) ||
       (order_item= *order->item)->check_cols(1) ||
       session->is_fatal_error))
    return true; /* Wrong field. */

  uint32_t el= all_fields.elements;
  all_fields.push_front(order_item); /* Add new field to field list. */
  ref_pointer_array[el]= order_item;
  order->item= ref_pointer_array + el;
  return false;
}


/**
  Change order to point at item in select list.

  If item isn't a number and doesn't exits in the select list, add it the
  the field list.
*/

int setup_order(Session *session, Item **ref_pointer_array, TableList *tables,
		List<Item> &fields, List<Item> &all_fields, order_st *order)
{
  session->where="order clause";
  for (; order; order=order->next)
  {
    if (find_order_in_list(session, ref_pointer_array, tables, order, fields,
			   all_fields, false))
      return 1;
  }
  return 0;
}


/**
  Intitialize the GROUP BY list.

  @param session			Thread handler
  @param ref_pointer_array	We store references to all fields that was
                               not in 'fields' here.
  @param fields		All fields in the select part. Any item in
                               'order' that is part of these list is replaced
                               by a pointer to this fields.
  @param all_fields		Total list of all unique fields used by the
                               select. All items in 'order' that was not part
                               of fields will be added first to this list.
  @param order			The fields we should do GROUP BY on.
  @param hidden_group_fields	Pointer to flag that is set to 1 if we added
                               any fields to all_fields.

  @todo
    change ER_WRONG_FIELD_WITH_GROUP to more detailed
    ER_NON_GROUPING_FIELD_USED

  @retval
    0  ok
  @retval
    1  error (probably out of memory)
*/

int
setup_group(Session *session, Item **ref_pointer_array, TableList *tables,
	    List<Item> &fields, List<Item> &all_fields, order_st *order,
	    bool *hidden_group_fields)
{
  *hidden_group_fields=0;
  order_st *ord;

  if (!order)
    return 0;				/* Everything is ok */

  uint32_t org_fields=all_fields.elements;

  session->where="group statement";
  for (ord= order; ord; ord= ord->next)
  {
    if (find_order_in_list(session, ref_pointer_array, tables, ord, fields,
			   all_fields, true))
      return 1;
    (*ord->item)->marker= UNDEF_POS;		/* Mark found */
    if ((*ord->item)->with_sum_func)
    {
      my_error(ER_WRONG_GROUP_FIELD, MYF(0), (*ord->item)->full_name());
      return 1;
    }
  }
  /* MODE_ONLY_FULL_GROUP_BY */
  {
    /*
      Don't allow one to use fields that is not used in GROUP BY
      For each select a list of field references that aren't under an
      aggregate function is created. Each field in this list keeps the
      position of the select list expression which it belongs to.

      First we check an expression from the select list against the GROUP BY
      list. If it's found there then it's ok. It's also ok if this expression
      is a constant or an aggregate function. Otherwise we scan the list
      of non-aggregated fields and if we'll find at least one field reference
      that belongs to this expression and doesn't occur in the GROUP BY list
      we throw an error. If there are no fields in the created list for a
      select list expression this means that all fields in it are used under
      aggregate functions.
    */
    Item *item;
    Item_field *field;
    int cur_pos_in_select_list= 0;
    List_iterator<Item> li(fields);
    List_iterator<Item_field> naf_it(session->lex->current_select->non_agg_fields);

    field= naf_it++;
    while (field && (item=li++))
    {
      if (item->type() != Item::SUM_FUNC_ITEM && item->marker >= 0 &&
          !item->const_item() &&
          !(item->real_item()->type() == Item::FIELD_ITEM &&
            item->used_tables() & OUTER_REF_TABLE_BIT))
      {
        while (field)
        {
          /* Skip fields from previous expressions. */
          if (field->marker < cur_pos_in_select_list)
            goto next_field;
          /* Found a field from the next expression. */
          if (field->marker > cur_pos_in_select_list)
            break;
          /*
            Check whether the field occur in the GROUP BY list.
            Throw the error later if the field isn't found.
          */
          for (ord= order; ord; ord= ord->next)
            if ((*ord->item)->eq((Item*)field, 0))
              goto next_field;
          /*
            TODO: change ER_WRONG_FIELD_WITH_GROUP to more detailed
            ER_NON_GROUPING_FIELD_USED
          */
          my_error(ER_WRONG_FIELD_WITH_GROUP, MYF(0), field->full_name());
          return 1;
next_field:
          field= naf_it++;
        }
      }
      cur_pos_in_select_list++;
    }
  }
  if (org_fields != all_fields.elements)
    *hidden_group_fields=1;			// group fields is not used
  return 0;
}

/**
  Create a group by that consist of all non const fields.

  Try to use the fields in the order given by 'order' to allow one to
  optimize away 'order by'.
*/

static order_st *
create_distinct_group(Session *session, Item **ref_pointer_array,
                      order_st *order_list, List<Item> &fields,
                      List<Item> &, bool *all_order_by_fields_used)
{
  List_iterator<Item> li(fields);
  Item *item;
  order_st *order,*group,**prev;

  *all_order_by_fields_used= 1;
  while ((item=li++))
    item->marker=0;			/* Marker that field is not used */

  prev= &group;  group=0;
  for (order=order_list ; order; order=order->next)
  {
    if (order->in_field_list)
    {
      order_st *ord=(order_st*) session->memdup((char*) order,sizeof(order_st));
      if (!ord)
	return 0;
      *prev=ord;
      prev= &ord->next;
      (*ord->item)->marker=1;
    }
    else
      *all_order_by_fields_used= 0;
  }

  li.rewind();
  while ((item=li++))
  {
    if (!item->const_item() && !item->with_sum_func && !item->marker)
    {
      /*
        Don't put duplicate columns from the SELECT list into the
        GROUP BY list.
      */
      order_st *ord_iter;
      for (ord_iter= group; ord_iter; ord_iter= ord_iter->next)
        if ((*ord_iter->item)->eq(item, 1))
          goto next_item;

      order_st *ord=(order_st*) session->calloc(sizeof(order_st));
      if (!ord)
	return 0;

      /*
        We have here only field_list (not all_field_list), so we can use
        simple indexing of ref_pointer_array (order in the array and in the
        list are same)
      */
      ord->item= ref_pointer_array;
      ord->asc=1;
      *prev=ord;
      prev= &ord->next;
    }
next_item:
    ref_pointer_array++;
  }
  *prev=0;
  return group;
}


/**
  Update join with count of the different type of fields.
*/

void
count_field_types(Select_Lex *select_lex, Tmp_Table_Param *param,
                  List<Item> &fields, bool reset_with_sum_func)
{
  List_iterator<Item> li(fields);
  Item *field;

  param->field_count=param->sum_func_count=param->func_count=
    param->hidden_field_count=0;
  param->quick_group=1;
  while ((field=li++))
  {
    Item::Type real_type= field->real_item()->type();
    if (real_type == Item::FIELD_ITEM)
      param->field_count++;
    else if (real_type == Item::SUM_FUNC_ITEM)
    {
      if (! field->const_item())
      {
	Item_sum *sum_item=(Item_sum*) field->real_item();
        if (!sum_item->depended_from() ||
            sum_item->depended_from() == select_lex)
        {
          if (!sum_item->quick_group)
            param->quick_group=0;			// UDF SUM function
          param->sum_func_count++;

          for (uint32_t i=0 ; i < sum_item->arg_count ; i++)
          {
            if (sum_item->args[0]->real_item()->type() == Item::FIELD_ITEM)
              param->field_count++;
            else
              param->func_count++;
          }
        }
        param->func_count++;
      }
    }
    else
    {
      param->func_count++;
      if (reset_with_sum_func)
	field->with_sum_func=0;
    }
  }
}


/**
  Return 1 if second is a subpart of first argument.

  If first parts has different direction, change it to second part
  (group is sorted like order)
*/

static bool
test_if_subpart(order_st *a,order_st *b)
{
  for (; a && b; a=a->next,b=b->next)
  {
    if ((*a->item)->eq(*b->item,1))
      a->asc=b->asc;
    else
      return 0;
  }
  return test(!b);
}

/**
  Return table number if there is only one table in sort order
  and group and order is compatible, else return 0.
*/

static Table *
get_sort_by_table(order_st *a,order_st *b,TableList *tables)
{
  table_map map= (table_map) 0;

  if (!a)
    a=b;					// Only one need to be given
  else if (!b)
    b=a;

  for (; a && b; a=a->next,b=b->next)
  {
    if (!(*a->item)->eq(*b->item,1))
      return(0);
    map|=a->item[0]->used_tables();
  }
  if (!map || (map & (RAND_TABLE_BIT | OUTER_REF_TABLE_BIT)))
    return(0);

  for (; !(map & tables->table->map); tables= tables->next_leaf) {};
  if (map != tables->table->map)
    return(0);				// More than one table
  return(tables->table);
}


/**
  calc how big buffer we need for comparing group entries.
*/

static void
calc_group_buffer(JOIN *join,order_st *group)
{
  uint32_t key_length=0, parts=0, null_parts=0;

  if (group)
    join->group= 1;
  for (; group ; group=group->next)
  {
    Item *group_item= *group->item;
    Field *field= group_item->get_tmp_table_field();
    if (field)
    {
      enum_field_types type;
      if ((type= field->type()) == DRIZZLE_TYPE_BLOB)
	key_length+=MAX_BLOB_WIDTH;		// Can't be used as a key
      else if (type == DRIZZLE_TYPE_VARCHAR)
        key_length+= field->field_length + HA_KEY_BLOB_LENGTH;
      else
	key_length+= field->pack_length();
    }
    else
    {
      switch (group_item->result_type()) {
      case REAL_RESULT:
        key_length+= sizeof(double);
        break;
      case INT_RESULT:
        key_length+= sizeof(int64_t);
        break;
      case DECIMAL_RESULT:
        key_length+= my_decimal_get_binary_size(group_item->max_length -
                                                (group_item->decimals ? 1 : 0),
                                                group_item->decimals);
        break;
      case STRING_RESULT:
      {
        enum enum_field_types type= group_item->field_type();
        /*
          As items represented as DATE/TIME fields in the group buffer
          have STRING_RESULT result type, we increase the length
          by 8 as maximum pack length of such fields.
        */
        if (type == DRIZZLE_TYPE_DATE ||
            type == DRIZZLE_TYPE_DATETIME ||
            type == DRIZZLE_TYPE_TIMESTAMP)
        {
          key_length+= 8;
        }
        else
        {
          /*
            Group strings are taken as varstrings and require an length field.
            A field is not yet created by create_tmp_field()
            and the sizes should match up.
          */
          key_length+= group_item->max_length + HA_KEY_BLOB_LENGTH;
        }
        break;
      }
      default:
        /* This case should never be choosen */
        assert(0);
        my_error(ER_OUT_OF_RESOURCES, MYF(ME_FATALERROR));
      }
    }
    parts++;
    if (group_item->maybe_null)
      null_parts++;
  }
  join->tmp_table_param.group_length=key_length+null_parts;
  join->tmp_table_param.group_parts=parts;
  join->tmp_table_param.group_null_parts=null_parts;
}


/**
  allocate group fields or take prepared (cached).

  @param main_join   join of current select
  @param curr_join   current join (join of current select or temporary copy
                     of it)

  @retval
    0   ok
  @retval
    1   failed
*/

static bool
make_group_fields(JOIN *main_join, JOIN *curr_join)
{
  if (main_join->group_fields_cache.elements)
  {
    curr_join->group_fields= main_join->group_fields_cache;
    curr_join->sort_and_group= 1;
  }
  else
  {
    if (alloc_group_fields(curr_join, curr_join->group_list))
      return (1);
    main_join->group_fields_cache= curr_join->group_fields;
  }
  return (0);
}


/**
  Get a list of buffers for saveing last group.

  Groups are saved in reverse order for easyer check loop.
*/

static bool
alloc_group_fields(JOIN *join,order_st *group)
{
  if (group)
  {
    for (; group ; group=group->next)
    {
      Cached_item *tmp=new_Cached_item(join->session, *group->item, false);
      if (!tmp || join->group_fields.push_front(tmp))
	return true;
    }
  }
  join->sort_and_group=1;			/* Mark for do_select */
  return false;
}


/*
  Test if a single-row cache of items changed, and update the cache.

  @details Test if a list of items that typically represents a result
  row has changed. If the value of some item changed, update the cached
  value for this item.

  @param list list of <item, cached_value> pairs stored as Cached_item.

  @return -1 if no item changed
  @return index of the first item that changed
*/

int test_if_item_cache_changed(List<Cached_item> &list)
{
  List_iterator<Cached_item> li(list);
  int idx= -1,i;
  Cached_item *buff;

  for (i=(int) list.elements-1 ; (buff=li++) ; i--)
  {
    if (buff->cmp())
      idx=i;
  }
  return(idx);
}


/**
  Setup copy_fields to save fields at start of new group.

  Setup copy_fields to save fields at start of new group

  Only FIELD_ITEM:s and FUNC_ITEM:s needs to be saved between groups.
  Change old item_field to use a new field with points at saved fieldvalue
  This function is only called before use of send_fields.

  @param session                   Session pointer
  @param param                 temporary table parameters
  @param ref_pointer_array     array of pointers to top elements of filed list
  @param res_selected_fields   new list of items of select item list
  @param res_all_fields        new list of all items
  @param elements              number of elements in select item list
  @param all_fields            all fields list

  @todo
    In most cases this result will be sent to the user.
    This should be changed to use copy_int or copy_real depending
    on how the value is to be used: In some cases this may be an
    argument in a group function, like: IF(ISNULL(col),0,COUNT(*))

  @retval
    0     ok
  @retval
    !=0   error
*/

bool
setup_copy_fields(Session *session, Tmp_Table_Param *param,
		  Item **ref_pointer_array,
		  List<Item> &res_selected_fields, List<Item> &res_all_fields,
		  uint32_t elements, List<Item> &all_fields)
{
  Item *pos;
  List_iterator_fast<Item> li(all_fields);
  Copy_field *copy= NULL;
  res_selected_fields.empty();
  res_all_fields.empty();
  List_iterator_fast<Item> itr(res_all_fields);
  List<Item> extra_funcs;
  uint32_t i, border= all_fields.elements - elements;

  if (param->field_count &&
      !(copy=param->copy_field= new Copy_field[param->field_count]))
    goto err2;

  param->copy_funcs.empty();
  for (i= 0; (pos= li++); i++)
  {
    Field *field;
    unsigned char *tmp;
    Item *real_pos= pos->real_item();
    if (real_pos->type() == Item::FIELD_ITEM)
    {
      Item_field *item;
      if (!(item= new Item_field(session, ((Item_field*) real_pos))))
	goto err;
      if (pos->type() == Item::REF_ITEM)
      {
        /* preserve the names of the ref when dereferncing */
        Item_ref *ref= (Item_ref *) pos;
        item->db_name= ref->db_name;
        item->table_name= ref->table_name;
        item->name= ref->name;
      }
      pos= item;
      if (item->field->flags & BLOB_FLAG)
      {
	if (!(pos= new Item_copy_string(pos)))
	  goto err;
       /*
         Item_copy_string::copy for function can call
         Item_copy_string::val_int for blob via Item_ref.
         But if Item_copy_string::copy for blob isn't called before,
         it's value will be wrong
         so let's insert Item_copy_string for blobs in the beginning of
         copy_funcs
         (to see full test case look at having.test, BUG #4358)
       */
	if (param->copy_funcs.push_front(pos))
	  goto err;
      }
      else
      {
	/*
	   set up save buffer and change result_field to point at
	   saved value
	*/
	field= item->field;
	item->result_field=field->new_field(session->mem_root,field->table, 1);
        /*
          We need to allocate one extra byte for null handling and
          another extra byte to not get warnings from purify in
          Field_varstring::val_int
        */
	if (!(tmp= (unsigned char*) sql_alloc(field->pack_length()+2)))
	  goto err;
        if (copy)
        {
          copy->set(tmp, item->result_field);
          item->result_field->move_field(copy->to_ptr,copy->to_null_ptr,1);
#ifdef HAVE_purify
          copy->to_ptr[copy->from_length]= 0;
#endif
          copy++;
        }
      }
    }
    else if ((real_pos->type() == Item::FUNC_ITEM ||
	      real_pos->type() == Item::SUBSELECT_ITEM ||
	      real_pos->type() == Item::CACHE_ITEM ||
	      real_pos->type() == Item::COND_ITEM) &&
	     !real_pos->with_sum_func)
    {						// Save for send fields
      pos= real_pos;
      /* TODO:
	 In most cases this result will be sent to the user.
	 This should be changed to use copy_int or copy_real depending
	 on how the value is to be used: In some cases this may be an
	 argument in a group function, like: IF(ISNULL(col),0,COUNT(*))
      */
      if (!(pos=new Item_copy_string(pos)))
	goto err;
      if (i < border)                           // HAVING, order_st and GROUP BY
      {
        if (extra_funcs.push_back(pos))
          goto err;
      }
      else if (param->copy_funcs.push_back(pos))
	goto err;
    }
    res_all_fields.push_back(pos);
    ref_pointer_array[((i < border)? all_fields.elements-i-1 : i-border)]=
      pos;
  }
  param->copy_field_end= copy;

  for (i= 0; i < border; i++)
    itr++;
  itr.sublist(res_selected_fields, elements);
  /*
    Put elements from HAVING, order_st BY and GROUP BY last to ensure that any
    reference used in these will resolve to a item that is already calculated
  */
  param->copy_funcs.concat(&extra_funcs);

  return(0);

 err:
  if (copy)
    delete [] param->copy_field;			// This is never 0
  param->copy_field=0;
err2:
  return(true);
}


/**
  Make a copy of all simple SELECT'ed items.

  This is done at the start of a new group so that we can retrieve
  these later when the group changes.
*/

void
copy_fields(Tmp_Table_Param *param)
{
  Copy_field *ptr=param->copy_field;
  Copy_field *end=param->copy_field_end;

  for (; ptr != end; ptr++)
    (*ptr->do_copy)(ptr);

  List_iterator_fast<Item> it(param->copy_funcs);
  Item_copy_string *item;
  while ((item = (Item_copy_string*) it++))
    item->copy();
}


/**
  Make an array of pointers to sum_functions to speed up
  sum_func calculation.

  @retval
    0	ok
  @retval
    1	Error
*/

bool JOIN::alloc_func_list()
{
  uint32_t func_count, group_parts;

  func_count= tmp_table_param.sum_func_count;
  /*
    If we are using rollup, we need a copy of the summary functions for
    each level
  */
  if (rollup.state != ROLLUP::STATE_NONE)
    func_count*= (send_group_parts+1);

  group_parts= send_group_parts;
  /*
    If distinct, reserve memory for possible
    disctinct->group_by optimization
  */
  if (select_distinct)
  {
    group_parts+= fields_list.elements;
    /*
      If the order_st clause is specified then it's possible that
      it also will be optimized, so reserve space for it too
    */
    if (order)
    {
      order_st *ord;
      for (ord= order; ord; ord= ord->next)
        group_parts++;
    }
  }

  /* This must use calloc() as rollup_make_fields depends on this */
  sum_funcs= (Item_sum**) session->calloc(sizeof(Item_sum**) * (func_count+1) +
				      sizeof(Item_sum***) * (group_parts+1));
  sum_funcs_end= (Item_sum***) (sum_funcs+func_count+1);
  return(sum_funcs == 0);
}


/**
  Initialize 'sum_funcs' array with all Item_sum objects.

  @param field_list        All items
  @param send_fields       Items in select list
  @param before_group_by   Set to 1 if this is called before GROUP BY handling
  @param recompute         Set to true if sum_funcs must be recomputed

  @retval
    0  ok
  @retval
    1  error
*/

bool JOIN::make_sum_func_list(List<Item> &field_list, List<Item> &send_fields,
			      bool before_group_by, bool recompute)
{
  List_iterator_fast<Item> it(field_list);
  Item_sum **func;
  Item *item;

  if (*sum_funcs && !recompute)
    return(false); /* We have already initialized sum_funcs. */

  func= sum_funcs;
  while ((item=it++))
  {
    if (item->type() == Item::SUM_FUNC_ITEM && !item->const_item() &&
        (!((Item_sum*) item)->depended_from() ||
         ((Item_sum *)item)->depended_from() == select_lex))
      *func++= (Item_sum*) item;
  }
  if (before_group_by && rollup.state == ROLLUP::STATE_INITED)
  {
    rollup.state= ROLLUP::STATE_READY;
    if (rollup_make_fields(field_list, send_fields, &func))
      return(true);			// Should never happen
  }
  else if (rollup.state == ROLLUP::STATE_NONE)
  {
    for (uint32_t i=0 ; i <= send_group_parts ;i++)
      sum_funcs_end[i]= func;
  }
  else if (rollup.state == ROLLUP::STATE_READY)
    return(false);                         // Don't put end marker
  *func=0;					// End marker
  return(false);
}


/**
  Change all funcs and sum_funcs to fields in tmp table, and create
  new list of all items.

  @param session                   Session pointer
  @param ref_pointer_array     array of pointers to top elements of filed list
  @param res_selected_fields   new list of items of select item list
  @param res_all_fields        new list of all items
  @param elements              number of elements in select item list
  @param all_fields            all fields list

  @retval
    0     ok
  @retval
    !=0   error
*/

static bool
change_to_use_tmp_fields(Session *session, Item **ref_pointer_array,
			 List<Item> &res_selected_fields,
			 List<Item> &res_all_fields,
			 uint32_t elements, List<Item> &all_fields)
{
  List_iterator_fast<Item> it(all_fields);
  Item *item_field,*item;

  res_selected_fields.empty();
  res_all_fields.empty();

  uint32_t i, border= all_fields.elements - elements;
  for (i= 0; (item= it++); i++)
  {
    Field *field;

    if ((item->with_sum_func && item->type() != Item::SUM_FUNC_ITEM) ||
        (item->type() == Item::FUNC_ITEM &&
         ((Item_func*)item)->functype() == Item_func::SUSERVAR_FUNC))
      item_field= item;
    else
    {
      if (item->type() == Item::FIELD_ITEM)
      {
	item_field= item->get_tmp_table_item(session);
      }
      else if ((field= item->get_tmp_table_field()))
      {
	if (item->type() == Item::SUM_FUNC_ITEM && field->table->group)
	  item_field= ((Item_sum*) item)->result_item(field);
	else
	  item_field= (Item*) new Item_field(field);
	if (!item_field)
	  return(true);                    // Fatal error

        if (item->real_item()->type() != Item::FIELD_ITEM)
          field->orig_table= 0;
	item_field->name= item->name;
        if (item->type() == Item::REF_ITEM)
        {
          Item_field *ifield= (Item_field *) item_field;
          Item_ref *iref= (Item_ref *) item;
          ifield->table_name= iref->table_name;
          ifield->db_name= iref->db_name;
        }
      }
      else
	item_field= item;
    }
    res_all_fields.push_back(item_field);
    ref_pointer_array[((i < border)? all_fields.elements-i-1 : i-border)]=
      item_field;
  }

  List_iterator_fast<Item> itr(res_all_fields);
  for (i= 0; i < border; i++)
    itr++;
  itr.sublist(res_selected_fields, elements);
  return(false);
}


/**
  Change all sum_func refs to fields to point at fields in tmp table.
  Change all funcs to be fields in tmp table.

  @param session                   Session pointer
  @param ref_pointer_array     array of pointers to top elements of filed list
  @param res_selected_fields   new list of items of select item list
  @param res_all_fields        new list of all items
  @param elements              number of elements in select item list
  @param all_fields            all fields list

  @retval
    0	ok
  @retval
    1	error
*/

static bool
change_refs_to_tmp_fields(Session *session, Item **ref_pointer_array,
			  List<Item> &res_selected_fields,
			  List<Item> &res_all_fields, uint32_t elements,
			  List<Item> &all_fields)
{
  List_iterator_fast<Item> it(all_fields);
  Item *item, *new_item;
  res_selected_fields.empty();
  res_all_fields.empty();

  uint32_t i, border= all_fields.elements - elements;
  for (i= 0; (item= it++); i++)
  {
    res_all_fields.push_back(new_item= item->get_tmp_table_item(session));
    ref_pointer_array[((i < border)? all_fields.elements-i-1 : i-border)]=
      new_item;
  }

  List_iterator_fast<Item> itr(res_all_fields);
  for (i= 0; i < border; i++)
    itr++;
  itr.sublist(res_selected_fields, elements);

  return session->is_fatal_error;
}



/******************************************************************************
  Code for calculating functions
******************************************************************************/


/**
  Call ::setup for all sum functions.

  @param session           thread handler
  @param func_ptr      sum function list

  @retval
    false  ok
  @retval
    true   error
*/

static bool setup_sum_funcs(Session *session, Item_sum **func_ptr)
{
  Item_sum *func;
  while ((func= *(func_ptr++)))
  {
    if (func->setup(session))
      return(true);
  }
  return(false);
}


static void
init_tmptable_sum_functions(Item_sum **func_ptr)
{
  Item_sum *func;
  while ((func= *(func_ptr++)))
    func->reset_field();
}


/** Update record 0 in tmp_table from record 1. */

static void
update_tmptable_sum_func(Item_sum **func_ptr, Table *)
{
  Item_sum *func;
  while ((func= *(func_ptr++)))
    func->update_field();
}


/** Copy result of sum functions to record in tmp_table. */

static void
copy_sum_funcs(Item_sum **func_ptr, Item_sum **end_ptr)
{
  for (; func_ptr != end_ptr ; func_ptr++)
    (void) (*func_ptr)->save_in_result_field(1);
  return;
}


static bool
init_sum_functions(Item_sum **func_ptr, Item_sum **end_ptr)
{
  for (; func_ptr != end_ptr ;func_ptr++)
  {
    if ((*func_ptr)->reset())
      return 1;
  }
  /* If rollup, calculate the upper sum levels */
  for ( ; *func_ptr ; func_ptr++)
  {
    if ((*func_ptr)->add())
      return 1;
  }
  return 0;
}


static bool
update_sum_func(Item_sum **func_ptr)
{
  Item_sum *func;
  for (; (func= (Item_sum*) *func_ptr) ; func_ptr++)
    if (func->add())
      return 1;
  return 0;
}

/** Copy result of functions to record in tmp_table. */

void
copy_funcs(Item **func_ptr)
{
  Item *func;
  for (; (func = *func_ptr) ; func_ptr++)
    func->save_in_result_field(1);
}


/**
  Create a condition for a const reference and add this to the
  currenct select for the table.
*/

static bool add_ref_to_table_cond(Session *session, JOIN_TAB *join_tab)
{
  if (!join_tab->ref.key_parts)
    return(false);

  Item_cond_and *cond=new Item_cond_and();
  Table *table=join_tab->table;
  int error;
  if (!cond)
    return(true);

  for (uint32_t i=0 ; i < join_tab->ref.key_parts ; i++)
  {
    Field *field=table->field[table->key_info[join_tab->ref.key].key_part[i].
			      fieldnr-1];
    Item *value=join_tab->ref.items[i];
    cond->add(new Item_func_equal(new Item_field(field), value));
  }
  if (session->is_fatal_error)
    return(true);

  if (!cond->fixed)
    cond->fix_fields(session, (Item**)&cond);
  if (join_tab->select)
  {
    error=(int) cond->add(join_tab->select->cond);
    join_tab->select_cond=join_tab->select->cond=cond;
  }
  else if ((join_tab->select= make_select(join_tab->table, 0, 0, cond, 0,
                                          &error)))
    join_tab->select_cond=cond;

  return(error ? true : false);
}


/**
  Free joins of subselect of this select.

  @param session      Session pointer
  @param select   pointer to Select_Lex which subselects joins we will free
*/

void free_underlaid_joins(Session *, Select_Lex *select)
{
  for (Select_Lex_Unit *unit= select->first_inner_unit();
       unit;
       unit= unit->next_unit())
    unit->cleanup();
}

/****************************************************************************
  ROLLUP handling
****************************************************************************/

/**
  Replace occurences of group by fields in an expression by ref items.

  The function replaces occurrences of group by fields in expr
  by ref objects for these fields unless they are under aggregate
  functions.
  The function also corrects value of the the maybe_null attribute
  for the items of all subexpressions containing group by fields.

  @b EXAMPLES
    @code
      SELECT a+1 FROM t1 GROUP BY a WITH ROLLUP
      SELECT SUM(a)+a FROM t1 GROUP BY a WITH ROLLUP
  @endcode

  @b IMPLEMENTATION

    The function recursively traverses the tree of the expr expression,
    looks for occurrences of the group by fields that are not under
    aggregate functions and replaces them for the corresponding ref items.

  @note
    This substitution is needed GROUP BY queries with ROLLUP if
    SELECT list contains expressions over group by attributes.

  @param session                  reference to the context
  @param expr                 expression to make replacement
  @param group_list           list of references to group by items
  @param changed        out:  returns 1 if item contains a replaced field item

  @todo
    - TODO: Some functions are not null-preserving. For those functions
    updating of the maybe_null attribute is an overkill.

  @retval
    0	if ok
  @retval
    1   on error
*/

static bool change_group_ref(Session *session, Item_func *expr, order_st *group_list,
                             bool *changed)
{
  if (expr->arg_count)
  {
    Name_resolution_context *context= &session->lex->current_select->context;
    Item **arg,**arg_end;
    bool arg_changed= false;
    for (arg= expr->arguments(),
         arg_end= expr->arguments()+expr->arg_count;
         arg != arg_end; arg++)
    {
      Item *item= *arg;
      if (item->type() == Item::FIELD_ITEM || item->type() == Item::REF_ITEM)
      {
        order_st *group_tmp;
        for (group_tmp= group_list; group_tmp; group_tmp= group_tmp->next)
        {
          if (item->eq(*group_tmp->item,0))
          {
            Item *new_item;
            if (!(new_item= new Item_ref(context, group_tmp->item, 0,
                                        item->name)))
              return 1;                                 // fatal_error is set
            session->change_item_tree(arg, new_item);
            arg_changed= true;
          }
        }
      }
      else if (item->type() == Item::FUNC_ITEM)
      {
        if (change_group_ref(session, (Item_func *) item, group_list, &arg_changed))
          return 1;
      }
    }
    if (arg_changed)
    {
      expr->maybe_null= 1;
      *changed= true;
    }
  }
  return 0;
}


/** Allocate memory needed for other rollup functions. */

bool JOIN::rollup_init()
{
  uint32_t i,j;
  Item **ref_array;

  tmp_table_param.quick_group= 0;	// Can't create groups in tmp table
  rollup.state= ROLLUP::STATE_INITED;

  /*
    Create pointers to the different sum function groups
    These are updated by rollup_make_fields()
  */
  tmp_table_param.group_parts= send_group_parts;

  if (!(rollup.null_items= (Item_null_result**) session->alloc((sizeof(Item*) +
                                                sizeof(Item**) +
                                                sizeof(List<Item>) +
				                ref_pointer_array_size)
				                * send_group_parts )))
    return 1;

  rollup.fields= (List<Item>*) (rollup.null_items + send_group_parts);
  rollup.ref_pointer_arrays= (Item***) (rollup.fields + send_group_parts);
  ref_array= (Item**) (rollup.ref_pointer_arrays+send_group_parts);

  /*
    Prepare space for field list for the different levels
    These will be filled up in rollup_make_fields()
  */
  for (i= 0 ; i < send_group_parts ; i++)
  {
    rollup.null_items[i]= new (session->mem_root) Item_null_result();
    List<Item> *rollup_fields= &rollup.fields[i];
    rollup_fields->empty();
    rollup.ref_pointer_arrays[i]= ref_array;
    ref_array+= all_fields.elements;
  }
  for (i= 0 ; i < send_group_parts; i++)
  {
    for (j=0 ; j < fields_list.elements ; j++)
      rollup.fields[i].push_back(rollup.null_items[i]);
  }
  List_iterator<Item> it(all_fields);
  Item *item;
  while ((item= it++))
  {
    order_st *group_tmp;
    bool found_in_group= 0;

    for (group_tmp= group_list; group_tmp; group_tmp= group_tmp->next)
    {
      if (*group_tmp->item == item)
      {
        item->maybe_null= 1;
        found_in_group= 1;
        if (item->const_item())
        {
          /*
            For ROLLUP queries each constant item referenced in GROUP BY list
            is wrapped up into an Item_func object yielding the same value
            as the constant item. The objects of the wrapper class are never
            considered as constant items and besides they inherit all
            properties of the Item_result_field class.
            This wrapping allows us to ensure writing constant items
            into temporary tables whenever the result of the ROLLUP
            operation has to be written into a temporary table, e.g. when
            ROLLUP is used together with DISTINCT in the SELECT list.
            Usually when creating temporary tables for a intermidiate
            result we do not include fields for constant expressions.
	  */
          Item* new_item= new Item_func_rollup_const(item);
          if (!new_item)
            return 1;
          new_item->fix_fields(session, (Item **) 0);
          session->change_item_tree(it.ref(), new_item);
          for (order_st *tmp= group_tmp; tmp; tmp= tmp->next)
          {
            if (*tmp->item == item)
              session->change_item_tree(tmp->item, new_item);
          }
        }
      }
    }
    if (item->type() == Item::FUNC_ITEM && !found_in_group)
    {
      bool changed= false;
      if (change_group_ref(session, (Item_func *) item, group_list, &changed))
        return 1;
      /*
        We have to prevent creation of a field in a temporary table for
        an expression that contains GROUP BY attributes.
        Marking the expression item as 'with_sum_func' will ensure this.
      */
      if (changed)
        item->with_sum_func= 1;
    }
  }
  return 0;
}


/**
  Fill up rollup structures with pointers to fields to use.

  Creates copies of item_sum items for each sum level.

  @param fields_arg		List of all fields (hidden and real ones)
  @param sel_fields		Pointer to selected fields
  @param func			Store here a pointer to all fields

  @retval
    0	if ok;
    In this case func is pointing to next not used element.
  @retval
    1    on error
*/

bool JOIN::rollup_make_fields(List<Item> &fields_arg, List<Item> &sel_fields,
			      Item_sum ***func)
{
  List_iterator_fast<Item> it(fields_arg);
  Item *first_field= sel_fields.head();
  uint32_t level;

  /*
    Create field lists for the different levels

    The idea here is to have a separate field list for each rollup level to
    avoid all runtime checks of which columns should be NULL.

    The list is stored in reverse order to get sum function in such an order
    in func that it makes it easy to reset them with init_sum_functions()

    Assuming:  SELECT a, b, c SUM(b) FROM t1 GROUP BY a,b WITH ROLLUP

    rollup.fields[0] will contain list where a,b,c is NULL
    rollup.fields[1] will contain list where b,c is NULL
    ...
    rollup.ref_pointer_array[#] points to fields for rollup.fields[#]
    ...
    sum_funcs_end[0] points to all sum functions
    sum_funcs_end[1] points to all sum functions, except grand totals
    ...
  */

  for (level=0 ; level < send_group_parts ; level++)
  {
    uint32_t i;
    uint32_t pos= send_group_parts - level -1;
    bool real_fields= 0;
    Item *item;
    List_iterator<Item> new_it(rollup.fields[pos]);
    Item **ref_array_start= rollup.ref_pointer_arrays[pos];
    order_st *start_group;

    /* Point to first hidden field */
    Item **ref_array= ref_array_start + fields_arg.elements-1;

    /* Remember where the sum functions ends for the previous level */
    sum_funcs_end[pos+1]= *func;

    /* Find the start of the group for this level */
    for (i= 0, start_group= group_list ;
	 i++ < pos ;
	 start_group= start_group->next)
      ;

    it.rewind();
    while ((item= it++))
    {
      if (item == first_field)
      {
	real_fields= 1;				// End of hidden fields
	ref_array= ref_array_start;
      }

      if (item->type() == Item::SUM_FUNC_ITEM && !item->const_item() &&
          (!((Item_sum*) item)->depended_from() ||
           ((Item_sum *)item)->depended_from() == select_lex))

      {
	/*
	  This is a top level summary function that must be replaced with
	  a sum function that is reset for this level.

	  NOTE: This code creates an object which is not that nice in a
	  sub select.  Fortunately it's not common to have rollup in
	  sub selects.
	*/
	item= item->copy_or_same(session);
	((Item_sum*) item)->make_unique();
	*(*func)= (Item_sum*) item;
	(*func)++;
      }
      else
      {
	/* Check if this is something that is part of this group by */
	order_st *group_tmp;
	for (group_tmp= start_group, i= pos ;
             group_tmp ; group_tmp= group_tmp->next, i++)
	{
          if (*group_tmp->item == item)
	  {
	    /*
	      This is an element that is used by the GROUP BY and should be
	      set to NULL in this level
	    */
            Item_null_result *null_item= new (session->mem_root) Item_null_result();
            if (!null_item)
              return 1;
	    item->maybe_null= 1;		// Value will be null sometimes
            null_item->result_field= item->get_tmp_table_field();
            item= null_item;
	    break;
	  }
	}
      }
      *ref_array= item;
      if (real_fields)
      {
	(void) new_it++;			// Point to next item
	new_it.replace(item);			// Replace previous
	ref_array++;
      }
      else
	ref_array--;
    }
  }
  sum_funcs_end[0]= *func;			// Point to last function
  return 0;
}

/**
  Send all rollup levels higher than the current one to the client.

  @b SAMPLE
    @code
      SELECT a, b, c SUM(b) FROM t1 GROUP BY a,b WITH ROLLUP
  @endcode

  @param idx		Level we are on:
                        - 0 = Total sum level
                        - 1 = First group changed  (a)
                        - 2 = Second group changed (a,b)

  @retval
    0   ok
  @retval
    1   If send_data_failed()
*/

int JOIN::rollup_send_data(uint32_t idx)
{
  uint32_t i;
  for (i= send_group_parts ; i-- > idx ; )
  {
    /* Get reference pointers to sum functions in place */
    memcpy(ref_pointer_array, rollup.ref_pointer_arrays[i],
	   ref_pointer_array_size);
    if ((!having || having->val_int()))
    {
      if (send_records < unit->select_limit_cnt && do_send_rows &&
	  result->send_data(rollup.fields[i]))
	return 1;
      send_records++;
    }
  }
  /* Restore ref_pointer_array */
  set_items_ref_array(current_ref_pointer_array);
  return 0;
}

/**
  Write all rollup levels higher than the current one to a temp table.

  @b SAMPLE
    @code
      SELECT a, b, SUM(c) FROM t1 GROUP BY a,b WITH ROLLUP
  @endcode

  @param idx                 Level we are on:
                               - 0 = Total sum level
                               - 1 = First group changed  (a)
                               - 2 = Second group changed (a,b)
  @param table               reference to temp table

  @retval
    0   ok
  @retval
    1   if write_data_failed()
*/

int JOIN::rollup_write_data(uint32_t idx, Table *table_arg)
{
  uint32_t i;
  for (i= send_group_parts ; i-- > idx ; )
  {
    /* Get reference pointers to sum functions in place */
    memcpy(ref_pointer_array, rollup.ref_pointer_arrays[i],
	   ref_pointer_array_size);
    if ((!having || having->val_int()))
    {
      int write_error;
      Item *item;
      List_iterator_fast<Item> it(rollup.fields[i]);
      while ((item= it++))
      {
        if (item->type() == Item::NULL_ITEM && item->is_result_field())
          item->save_in_result_field(1);
      }
      copy_sum_funcs(sum_funcs_end[i+1], sum_funcs_end[i]);
      if ((write_error= table_arg->file->ha_write_row(table_arg->record[0])))
      {
	if (create_myisam_from_heap(session, table_arg,
                                    tmp_table_param.start_recinfo,
                                    &tmp_table_param.recinfo,
                                    write_error, 0))
	  return 1;
      }
    }
  }
  /* Restore ref_pointer_array */
  set_items_ref_array(current_ref_pointer_array);
  return 0;
}

/**
  clear results if there are not rows found for group
  (end_send_group/end_write_group)
*/

void JOIN::clear()
{
  clear_tables(this);
  copy_fields(&tmp_table_param);

  if (sum_funcs)
  {
    Item_sum *func, **func_ptr= sum_funcs;
    while ((func= *(func_ptr++)))
      func->clear();
  }
}

/**
  EXPLAIN handling.

  Send a description about what how the select will be done to stdout.
*/

void select_describe(JOIN *join, bool need_tmp_table, bool need_order,
                     bool distinct,const char *message)
{
  List<Item> field_list;
  List<Item> item_list;
  Session *session=join->session;
  select_result *result=join->result;
  Item *item_null= new Item_null();
  const CHARSET_INFO * const cs= system_charset_info;
  int quick_type;
  /* Don't log this into the slow query log */
  session->server_status&= ~(SERVER_QUERY_NO_INDEX_USED | SERVER_QUERY_NO_GOOD_INDEX_USED);
  join->unit->offset_limit_cnt= 0;

  /*
    NOTE: the number/types of items pushed into item_list must be in sync with
    EXPLAIN column types as they're "defined" in Session::send_explain_fields()
  */
  if (message)
  {
    item_list.push_back(new Item_int((int32_t)
				     join->select_lex->select_number));
    item_list.push_back(new Item_string(join->select_lex->type,
					strlen(join->select_lex->type), cs));
    for (uint32_t i=0 ; i < 7; i++)
      item_list.push_back(item_null);
    if (join->session->lex->describe & DESCRIBE_EXTENDED)
      item_list.push_back(item_null);

    item_list.push_back(new Item_string(message,strlen(message),cs));
    if (result->send_data(item_list))
      join->error= 1;
  }
  else if (join->select_lex == join->unit->fake_select_lex)
  {
    /*
      here we assume that the query will return at least two rows, so we
      show "filesort" in EXPLAIN. Of course, sometimes we'll be wrong
      and no filesort will be actually done, but executing all selects in
      the UNION to provide precise EXPLAIN information will hardly be
      appreciated :)
    */
    char table_name_buffer[NAME_LEN];
    item_list.empty();
    /* id */
    item_list.push_back(new Item_null);
    /* select_type */
    item_list.push_back(new Item_string(join->select_lex->type,
					strlen(join->select_lex->type),
					cs));
    /* table */
    {
      Select_Lex *sl= join->unit->first_select();
      uint32_t len= 6, lastop= 0;
      memcpy(table_name_buffer, STRING_WITH_LEN("<union"));
      for (; sl && len + lastop + 5 < NAME_LEN; sl= sl->next_select())
      {
        len+= lastop;
        lastop= snprintf(table_name_buffer + len, NAME_LEN - len,
                         "%u,", sl->select_number);
      }
      if (sl || len + lastop >= NAME_LEN)
      {
        memcpy(table_name_buffer + len, STRING_WITH_LEN("...>") + 1);
        len+= 4;
      }
      else
      {
        len+= lastop;
        table_name_buffer[len - 1]= '>';  // change ',' to '>'
      }
      item_list.push_back(new Item_string(table_name_buffer, len, cs));
    }
    /* type */
    item_list.push_back(new Item_string(join_type_str[JT_ALL],
					  strlen(join_type_str[JT_ALL]),
					  cs));
    /* possible_keys */
    item_list.push_back(item_null);
    /* key*/
    item_list.push_back(item_null);
    /* key_len */
    item_list.push_back(item_null);
    /* ref */
    item_list.push_back(item_null);
    /* in_rows */
    if (join->session->lex->describe & DESCRIBE_EXTENDED)
      item_list.push_back(item_null);
    /* rows */
    item_list.push_back(item_null);
    /* extra */
    if (join->unit->global_parameters->order_list.first)
      item_list.push_back(new Item_string("Using filesort",
					  14, cs));
    else
      item_list.push_back(new Item_string("", 0, cs));

    if (result->send_data(item_list))
      join->error= 1;
  }
  else
  {
    table_map used_tables=0;
    for (uint32_t i=0 ; i < join->tables ; i++)
    {
      JOIN_TAB *tab=join->join_tab+i;
      Table *table=tab->table;
      TableList *table_list= tab->table->pos_in_table_list;
      char buff[512];
      char buff1[512], buff2[512], buff3[512];
      char keylen_str_buf[64];
      String extra(buff, sizeof(buff),cs);
      char table_name_buffer[NAME_LEN];
      String tmp1(buff1,sizeof(buff1),cs);
      String tmp2(buff2,sizeof(buff2),cs);
      String tmp3(buff3,sizeof(buff3),cs);
      extra.length(0);
      tmp1.length(0);
      tmp2.length(0);
      tmp3.length(0);

      quick_type= -1;
      item_list.empty();
      /* id */
      item_list.push_back(new Item_uint((uint32_t)
				       join->select_lex->select_number));
      /* select_type */
      item_list.push_back(new Item_string(join->select_lex->type,
					  strlen(join->select_lex->type),
					  cs));
      if (tab->type == JT_ALL && tab->select && tab->select->quick)
      {
        quick_type= tab->select->quick->get_type();
        if ((quick_type == QUICK_SELECT_I::QS_TYPE_INDEX_MERGE) ||
            (quick_type == QUICK_SELECT_I::QS_TYPE_ROR_INTERSECT) ||
            (quick_type == QUICK_SELECT_I::QS_TYPE_ROR_UNION))
          tab->type = JT_INDEX_MERGE;
        else
	  tab->type = JT_RANGE;
      }
      /* table */
      if (table->derived_select_number)
      {
	/* Derived table name generation */
	int len= snprintf(table_name_buffer, sizeof(table_name_buffer)-1,
		          "<derived%u>",
                          table->derived_select_number);
	item_list.push_back(new Item_string(table_name_buffer, len, cs));
      }
      else
      {
        TableList *real_table= table->pos_in_table_list;
	item_list.push_back(new Item_string(real_table->alias,
					    strlen(real_table->alias),
					    cs));
      }
      /* "type" column */
      item_list.push_back(new Item_string(join_type_str[tab->type],
					  strlen(join_type_str[tab->type]),
					  cs));
      /* Build "possible_keys" value and add it to item_list */
      if (tab->keys.any())
      {
        uint32_t j;
        for (j=0 ; j < table->s->keys ; j++)
        {
          if (tab->keys.test(j))
          {
            if (tmp1.length())
              tmp1.append(',');
            tmp1.append(table->key_info[j].name,
			strlen(table->key_info[j].name),
			system_charset_info);
          }
        }
      }
      if (tmp1.length())
	item_list.push_back(new Item_string(tmp1.ptr(),tmp1.length(),cs));
      else
	item_list.push_back(item_null);

      /* Build "key", "key_len", and "ref" values and add them to item_list */
      if (tab->ref.key_parts)
      {
	KEY *key_info=table->key_info+ tab->ref.key;
        register uint32_t length;
	item_list.push_back(new Item_string(key_info->name,
					    strlen(key_info->name),
					    system_charset_info));
        length= int64_t2str(tab->ref.key_length, keylen_str_buf, 10) -
                keylen_str_buf;
        item_list.push_back(new Item_string(keylen_str_buf, length,
                                            system_charset_info));
	for (store_key **ref=tab->ref.key_copy ; *ref ; ref++)
	{
	  if (tmp2.length())
	    tmp2.append(',');
	  tmp2.append((*ref)->name(), strlen((*ref)->name()),
		      system_charset_info);
	}
	item_list.push_back(new Item_string(tmp2.ptr(),tmp2.length(),cs));
      }
      else if (tab->type == JT_NEXT)
      {
	KEY *key_info=table->key_info+ tab->index;
        register uint32_t length;
	item_list.push_back(new Item_string(key_info->name,
					    strlen(key_info->name),cs));
        length= int64_t2str(key_info->key_length, keylen_str_buf, 10) -
                keylen_str_buf;
        item_list.push_back(new Item_string(keylen_str_buf,
                                            length,
                                            system_charset_info));
	item_list.push_back(item_null);
      }
      else if (tab->select && tab->select->quick)
      {
        tab->select->quick->add_keys_and_lengths(&tmp2, &tmp3);
	item_list.push_back(new Item_string(tmp2.ptr(),tmp2.length(),cs));
	item_list.push_back(new Item_string(tmp3.ptr(),tmp3.length(),cs));
	item_list.push_back(item_null);
      }
      else
      {
        if (table_list->schema_table && table_list->schema_table->i_s_requested_object & OPTIMIZE_I_S_TABLE)
        {
          const char *tmp_buff;
          int f_idx;
          if (table_list->has_db_lookup_value)
          {
            f_idx= table_list->schema_table->idx_field1;
            tmp_buff= table_list->schema_table->fields_info[f_idx].field_name;
            tmp2.append(tmp_buff, strlen(tmp_buff), cs);
          }
          if (table_list->has_table_lookup_value)
          {
            if (table_list->has_db_lookup_value)
              tmp2.append(',');
            f_idx= table_list->schema_table->idx_field2;
            tmp_buff= table_list->schema_table->fields_info[f_idx].field_name;
            tmp2.append(tmp_buff, strlen(tmp_buff), cs);
          }
          if (tmp2.length())
            item_list.push_back(new Item_string(tmp2.ptr(),tmp2.length(),cs));
          else
            item_list.push_back(item_null);
        }
        else
          item_list.push_back(item_null);
	item_list.push_back(item_null);
	item_list.push_back(item_null);
      }

      /* Add "rows" field to item_list. */
      if (table_list->schema_table)
      {
        /* in_rows */
        if (join->session->lex->describe & DESCRIBE_EXTENDED)
          item_list.push_back(item_null);
        /* rows */
        item_list.push_back(item_null);
      }
      else
      {
        double examined_rows;
        if (tab->select && tab->select->quick)
          examined_rows= rows2double(tab->select->quick->records);
        else if (tab->type == JT_NEXT || tab->type == JT_ALL)
          examined_rows= rows2double(tab->limit ? tab->limit :
                                     tab->table->file->records());
        else
          examined_rows= join->best_positions[i].records_read;

        item_list.push_back(new Item_int((int64_t) (uint64_t) examined_rows,
                                         MY_INT64_NUM_DECIMAL_DIGITS));

        /* Add "filtered" field to item_list. */
        if (join->session->lex->describe & DESCRIBE_EXTENDED)
        {
          float f= 0.0;
          if (examined_rows)
            f= (float) (100.0 * join->best_positions[i].records_read /
                        examined_rows);
          item_list.push_back(new Item_float(f, 2));
        }
      }

      /* Build "Extra" field and add it to item_list. */
      bool key_read=table->key_read;
      if ((tab->type == JT_NEXT || tab->type == JT_CONST) &&
          table->covering_keys.test(tab->index))
	key_read=1;
      if (quick_type == QUICK_SELECT_I::QS_TYPE_ROR_INTERSECT &&
          !((QUICK_ROR_INTERSECT_SELECT*)tab->select->quick)->need_to_fetch_row)
        key_read=1;

      if (tab->info)
	item_list.push_back(new Item_string(tab->info,strlen(tab->info),cs));
      else if (tab->packed_info & TAB_INFO_HAVE_VALUE)
      {
        if (tab->packed_info & TAB_INFO_USING_INDEX)
          extra.append(STRING_WITH_LEN("; Using index"));
        if (tab->packed_info & TAB_INFO_USING_WHERE)
          extra.append(STRING_WITH_LEN("; Using where"));
        if (tab->packed_info & TAB_INFO_FULL_SCAN_ON_NULL)
          extra.append(STRING_WITH_LEN("; Full scan on NULL key"));
        /* Skip initial "; "*/
        const char *str= extra.ptr();
        uint32_t len= extra.length();
        if (len)
        {
          str += 2;
          len -= 2;
        }
	item_list.push_back(new Item_string(str, len, cs));
      }
      else
      {
        uint32_t keyno= MAX_KEY;
        if (tab->ref.key_parts)
          keyno= tab->ref.key;
        else if (tab->select && tab->select->quick)
          keyno = tab->select->quick->index;

        if (keyno != MAX_KEY && keyno == table->file->pushed_idx_cond_keyno &&
            table->file->pushed_idx_cond)
          extra.append(STRING_WITH_LEN("; Using index condition"));

        if (quick_type == QUICK_SELECT_I::QS_TYPE_ROR_UNION ||
            quick_type == QUICK_SELECT_I::QS_TYPE_ROR_INTERSECT ||
            quick_type == QUICK_SELECT_I::QS_TYPE_INDEX_MERGE)
        {
          extra.append(STRING_WITH_LEN("; Using "));
          tab->select->quick->add_info_string(&extra);
        }
          if (tab->select)
	{
	  if (tab->use_quick == 2)
	  {
            /*
             * To print out the bitset in tab->keys, we go through
             * it 32 bits at a time. We need to do this to ensure
             * that the to_ulong() method will not throw an
             * out_of_range exception at runtime which would happen
             * if the bitset we were working with was larger than 64
             * bits on a 64-bit platform (for example).
             */
            stringstream s, w;
            string str;
            w << tab->keys;
            w >> str;
            for (uint32_t pos= 0; pos < tab->keys.size(); pos+= 32)
            {
              bitset<32> tmp(str, pos, 32);
              if (tmp.any())
                s << uppercase << hex << tmp.to_ulong();
            }
            extra.append(STRING_WITH_LEN("; Range checked for each "
                                         "record (index map: 0x"));
            extra.append(s.str().c_str());
            extra.append(')');
	  }
	  else if (tab->select->cond)
          {
            const COND *pushed_cond= tab->table->file->pushed_cond;

            if (session->variables.engine_condition_pushdown && pushed_cond)
            {
              extra.append(STRING_WITH_LEN("; Using where with pushed "
                                           "condition"));
              if (session->lex->describe & DESCRIBE_EXTENDED)
              {
                extra.append(STRING_WITH_LEN(": "));
                ((COND *)pushed_cond)->print(&extra, QT_ORDINARY);
              }
            }
            else
              extra.append(STRING_WITH_LEN("; Using where"));
          }
        }
        if (key_read)
        {
          if (quick_type == QUICK_SELECT_I::QS_TYPE_GROUP_MIN_MAX)
            extra.append(STRING_WITH_LEN("; Using index for group-by"));
          else
            extra.append(STRING_WITH_LEN("; Using index"));
        }
        if (table->reginfo.not_exists_optimize)
          extra.append(STRING_WITH_LEN("; Not exists"));

        if (quick_type == QUICK_SELECT_I::QS_TYPE_RANGE &&
            !(((QUICK_RANGE_SELECT*)(tab->select->quick))->mrr_flags &
             HA_MRR_USE_DEFAULT_IMPL))
        {
	  extra.append(STRING_WITH_LEN("; Using MRR"));
        }

        if (table_list->schema_table &&
            table_list->schema_table->i_s_requested_object & OPTIMIZE_I_S_TABLE)
        {
          if (!table_list->table_open_method)
            extra.append(STRING_WITH_LEN("; Skip_open_table"));
          else if (table_list->table_open_method == OPEN_FRM_ONLY)
            extra.append(STRING_WITH_LEN("; Open_frm_only"));
          else
            extra.append(STRING_WITH_LEN("; Open_full_table"));
          if (table_list->has_db_lookup_value &&
              table_list->has_table_lookup_value)
            extra.append(STRING_WITH_LEN("; Scanned 0 databases"));
          else if (table_list->has_db_lookup_value ||
                   table_list->has_table_lookup_value)
            extra.append(STRING_WITH_LEN("; Scanned 1 database"));
          else
            extra.append(STRING_WITH_LEN("; Scanned all databases"));
        }
        if (need_tmp_table)
        {
          need_tmp_table=0;
          extra.append(STRING_WITH_LEN("; Using temporary"));
        }
        if (need_order)
        {
          need_order=0;
          extra.append(STRING_WITH_LEN("; Using filesort"));
        }
        if (distinct & test_all_bits(used_tables,session->used_tables))
          extra.append(STRING_WITH_LEN("; Distinct"));

        if (tab->insideout_match_tab)
        {
          extra.append(STRING_WITH_LEN("; LooseScan"));
        }

        if (tab->flush_weedout_table)
          extra.append(STRING_WITH_LEN("; Start temporary"));
        else if (tab->check_weed_out_table)
          extra.append(STRING_WITH_LEN("; End temporary"));
        else if (tab->do_firstmatch)
        {
          extra.append(STRING_WITH_LEN("; FirstMatch("));
          Table *prev_table=tab->do_firstmatch->table;
          if (prev_table->derived_select_number)
          {
            char namebuf[NAME_LEN];
            /* Derived table name generation */
            int len= snprintf(namebuf, sizeof(namebuf)-1,
                              "<derived%u>",
                              prev_table->derived_select_number);
            extra.append(namebuf, len);
          }
          else
            extra.append(prev_table->pos_in_table_list->alias);
          extra.append(STRING_WITH_LEN(")"));
        }

        for (uint32_t part= 0; part < tab->ref.key_parts; part++)
        {
          if (tab->ref.cond_guards[part])
          {
            extra.append(STRING_WITH_LEN("; Full scan on NULL key"));
            break;
          }
        }

        if (i > 0 && tab[-1].next_select == sub_select_cache)
          extra.append(STRING_WITH_LEN("; Using join buffer"));

        /* Skip initial "; "*/
        const char *str= extra.ptr();
        uint32_t len= extra.length();
        if (len)
        {
          str += 2;
          len -= 2;
        }
        item_list.push_back(new Item_string(str, len, cs));
      }
      // For next iteration
      used_tables|=table->map;
      if (result->send_data(item_list))
	join->error= 1;
    }
  }
  for (Select_Lex_Unit *unit= join->select_lex->first_inner_unit();
       unit;
       unit= unit->next_unit())
  {
    if (mysql_explain_union(session, unit, result))
      return;
  }
  return;
}


bool mysql_explain_union(Session *session, Select_Lex_Unit *unit, select_result *result)
{
  bool res= false;
  Select_Lex *first= unit->first_select();

  for (Select_Lex *sl= first;
       sl;
       sl= sl->next_select())
  {
    // drop UNCACHEABLE_EXPLAIN, because it is for internal usage only
    uint8_t uncacheable= (sl->uncacheable & ~UNCACHEABLE_EXPLAIN);
    sl->type= (((&session->lex->select_lex)==sl)?
	       (sl->first_inner_unit() || sl->next_select() ?
		"PRIMARY" : "SIMPLE"):
	       ((sl == first)?
		((sl->linkage == DERIVED_TABLE_TYPE) ?
		 "DERIVED":
		 ((uncacheable & UNCACHEABLE_DEPENDENT) ?
		  "DEPENDENT SUBQUERY":
		  (uncacheable?"UNCACHEABLE SUBQUERY":
		   "SUBQUERY"))):
		((uncacheable & UNCACHEABLE_DEPENDENT) ?
		 "DEPENDENT UNION":
		 uncacheable?"UNCACHEABLE UNION":
		 "UNION")));
    sl->options|= SELECT_DESCRIBE;
  }
  if (unit->is_union())
  {
    unit->fake_select_lex->select_number= UINT_MAX; // jost for initialization
    unit->fake_select_lex->type= "UNION RESULT";
    unit->fake_select_lex->options|= SELECT_DESCRIBE;
    if (!(res= unit->prepare(session, result, SELECT_NO_UNLOCK | SELECT_DESCRIBE)))
      res= unit->exec();
    res|= unit->cleanup();
  }
  else
  {
    session->lex->current_select= first;
    unit->set_limit(unit->global_parameters);
    res= mysql_select(session, &first->ref_pointer_array,
			(TableList*) first->table_list.first,
			first->with_wild, first->item_list,
			first->where,
			first->order_list.elements +
			first->group_list.elements,
			(order_st*) first->order_list.first,
			(order_st*) first->group_list.first,
			first->having,
			first->options | session->options | SELECT_DESCRIBE,
			result, unit, first);
  }
  return(res || session->is_error());
}


static void print_table_array(Session *session, String *str, TableList **table,
                              TableList **end)
{
  (*table)->print(session, str, QT_ORDINARY);

  for (TableList **tbl= table + 1; tbl < end; tbl++)
  {
    TableList *curr= *tbl;
    if (curr->outer_join)
    {
      /* MySQL converts right to left joins */
      str->append(STRING_WITH_LEN(" left join "));
    }
    else if (curr->straight)
      str->append(STRING_WITH_LEN(" straight_join "));
    else if (curr->sj_inner_tables)
      str->append(STRING_WITH_LEN(" semi join "));
    else
      str->append(STRING_WITH_LEN(" join "));
    curr->print(session, str, QT_ORDINARY);
    if (curr->on_expr)
    {
      str->append(STRING_WITH_LEN(" on("));
      curr->on_expr->print(str, QT_ORDINARY);
      str->append(')');
    }
  }
}


/**
  Print joins from the FROM clause.
  @param session     thread handler
  @param str     string where table should be printed
  @param tables  list of tables in join
  @query_type    type of the query is being generated
*/

static void print_join(Session *session, String *str,
                       List<TableList> *tables, enum_query_type)
{
  /* List is reversed => we should reverse it before using */
  List_iterator_fast<TableList> ti(*tables);
  TableList **table= (TableList **)session->alloc(sizeof(TableList*) *
                                                tables->elements);
  if (table == 0)
    return;  // out of memory

  for (TableList **t= table + (tables->elements - 1); t >= table; t--)
    *t= ti++;

  /*
    If the first table is a semi-join nest, swap it with something that is
    not a semi-join nest.
  */
  if ((*table)->sj_inner_tables)
  {
    TableList **end= table + tables->elements;
    for (TableList **t2= table; t2!=end; t2++)
    {
      if (!(*t2)->sj_inner_tables)
      {
        TableList *tmp= *t2;
        *t2= *table;
        *table= tmp;
        break;
      }
    }
  }
  assert(tables->elements >= 1);
  print_table_array(session, str, table, table + tables->elements);
}

/**
  Print table as it should be in join list.

  @param str   string where table should be printed
*/
void TableList::print(Session *session, String *str, enum_query_type query_type)
{
  if (nested_join)
  {
    str->append('(');
    print_join(session, str, &nested_join->join_list, query_type);
    str->append(')');
  }
  else
  {
    const char *cmp_name;                         // Name to compare with alias
    if (derived)
    {
      // A derived table
      str->append('(');
      derived->print(str, query_type);
      str->append(')');
      cmp_name= "";                               // Force printing of alias
    }
    else
    {
      // A normal table
      {
        str->append_identifier(db, db_length);
        str->append('.');
      }
      if (schema_table)
      {
        str->append_identifier(schema_table_name, strlen(schema_table_name));
        cmp_name= schema_table_name;
      }
      else
      {
        str->append_identifier(table_name, table_name_length);
        cmp_name= table_name;
      }
    }
    if (my_strcasecmp(table_alias_charset, cmp_name, alias))
    {

      if (alias && alias[0])
      {
        str->append(' ');

        string t_alias(alias);
        if (lower_case_table_names== 1)
          transform(t_alias.begin(), t_alias.end(),
                    t_alias.begin(), ::tolower);

        str->append_identifier(t_alias.c_str(), t_alias.length());
      }

    }

    if (index_hints)
    {
      List_iterator<Index_hint> it(*index_hints);
      Index_hint *hint;

      while ((hint= it++))
      {
        str->append (STRING_WITH_LEN(" "));
        hint->print (session, str);
      }
    }
  }
}


void Select_Lex::print(Session *session, String *str, enum_query_type query_type)
{
  /* QQ: session may not be set for sub queries, but this should be fixed */
  if (!session)
    session= current_session;

  str->append(STRING_WITH_LEN("select "));

  /* First add options */
  if (options & SELECT_STRAIGHT_JOIN)
    str->append(STRING_WITH_LEN("straight_join "));
  if (options & SELECT_DISTINCT)
    str->append(STRING_WITH_LEN("distinct "));
  if (options & SELECT_SMALL_RESULT)
    str->append(STRING_WITH_LEN("sql_small_result "));
  if (options & SELECT_BIG_RESULT)
    str->append(STRING_WITH_LEN("sql_big_result "));
  if (options & OPTION_BUFFER_RESULT)
    str->append(STRING_WITH_LEN("sql_buffer_result "));
  if (options & OPTION_FOUND_ROWS)
    str->append(STRING_WITH_LEN("sql_calc_found_rows "));

  //Item List
  bool first= 1;
  List_iterator_fast<Item> it(item_list);
  Item *item;
  while ((item= it++))
  {
    if (first)
      first= 0;
    else
      str->append(',');
    item->print_item_w_name(str, query_type);
  }

  /*
    from clause
    TODO: support USING/FORCE/IGNORE index
  */
  if (table_list.elements)
  {
    str->append(STRING_WITH_LEN(" from "));
    /* go through join tree */
    print_join(session, str, &top_join_list, query_type);
  }
  else if (where)
  {
    /*
      "SELECT 1 FROM DUAL WHERE 2" should not be printed as
      "SELECT 1 WHERE 2": the 1st syntax is valid, but the 2nd is not.
    */
    str->append(STRING_WITH_LEN(" from DUAL "));
  }

  // Where
  Item *cur_where= where;
  if (join)
    cur_where= join->conds;
  if (cur_where || cond_value != Item::COND_UNDEF)
  {
    str->append(STRING_WITH_LEN(" where "));
    if (cur_where)
      cur_where->print(str, query_type);
    else
      str->append(cond_value != Item::COND_FALSE ? "1" : "0");
  }

  // group by & olap
  if (group_list.elements)
  {
    str->append(STRING_WITH_LEN(" group by "));
    print_order(str, (order_st *) group_list.first, query_type);
    switch (olap)
    {
      case CUBE_TYPE:
	str->append(STRING_WITH_LEN(" with cube"));
	break;
      case ROLLUP_TYPE:
	str->append(STRING_WITH_LEN(" with rollup"));
	break;
      default:
	;  //satisfy compiler
    }
  }

  // having
  Item *cur_having= having;
  if (join)
    cur_having= join->having;

  if (cur_having || having_value != Item::COND_UNDEF)
  {
    str->append(STRING_WITH_LEN(" having "));
    if (cur_having)
      cur_having->print(str, query_type);
    else
      str->append(having_value != Item::COND_FALSE ? "1" : "0");
  }

  if (order_list.elements)
  {
    str->append(STRING_WITH_LEN(" order by "));
    print_order(str, (order_st *) order_list.first, query_type);
  }

  // limit
  print_limit(session, str, query_type);

  // PROCEDURE unsupported here
}


/**
  change select_result object of JOIN.

  @param res		new select_result object

  @retval
    false   OK
  @retval
    true    error
*/

bool JOIN::change_result(select_result *res)
{
  result= res;
  if (result->prepare(fields_list, select_lex->master_unit()))
  {
    return(true);
  }
  return(false);
}

/**
  @} (end of group Query_Optimizer)
*/