MWL#90: Subqueries: Inside-out execution for non-semijoin materialized...

MWL#90: Subqueries: Inside-out execution for non-semijoin materialized subqueries that are AND-parts of the WHERE
- Address feedback
- Code cleanup (not finished)

sql/item_cmpfunc.cc

@@ -5733,26 +5733,10 @@ Item_field* Item_equal::get_first(Item_field *field)
       It's a field from an materialized semi-join. We can substitute it only
       for a field from the same semi-join.
     */
-#if 0
-    psergey3:remove:
-    JOIN_TAB *first;
-    JOIN *join= field_tab->join;
-    int tab_idx= field_tab - field_tab->join->join_tab;
 
-    /* Find the first table of this semi-join nest */
-    for (int i= tab_idx; i >= (int)join->const_tables; i--)
-    {
-      if (join->join_tab[i].table->map & emb_nest->sj_inner_tables)
-        first= join->join_tab + i;
-      else
-        // Found first tab that doesn't belong to current SJ.
-        break;
-    }
-#endif    
     /* Find an item to substitute for. */
     while ((item= it++))
     {
-      //if (item->field->table->reginfo.join_tab >= first)
       if (item->field->table->pos_in_table_list->embedding == emb_nest)
       {
         /*

sql/opt_subselect.cc

@@ -45,8 +45,8 @@
   exception that we don't care how many matches a row from outer_tbl has in
   inner_tbl.
 
-  In SQL, that translates into following: a semi-join subquery is an IN 
-  subquery that is an AND-part of the WHERE/ON clause.
+  In SQL terms: a semi-join subquery is an IN subquery that is an AND-part of
+  the WHERE/ON clause.
 
   2. General idea about semi-join execution
   -----------------------------------------
@@ -3957,8 +3957,8 @@ static void remove_subq_pushed_predicates(JOIN *join, Item **where)
 
 bool join_tab_execution_startup(JOIN_TAB *tab)
 {
-  DBUG_ENTER("join_tab_execution_startup");
   Item_in_subselect *in_subs;
+  DBUG_ENTER("join_tab_execution_startup");
   if (tab->table->pos_in_table_list && 
       (in_subs= tab->table->pos_in_table_list->jtbm_subselect))
   {
@@ -3995,13 +3995,9 @@ bool join_tab_execution_startup(JOIN_TAB *tab)
       if ((rc= sub_select(join, join_tab, FALSE/* no EOF */)) < 0 ||
           (rc= sub_select(join, join_tab, TRUE/* now EOF */)) < 0)
       {
-        //psergey3-todo: set sjm->materialized=TRUE here, too??
         join->return_tab= save_return_tab;
         DBUG_RETURN(rc); /* it's NESTED_LOOP_(ERROR|KILLED)*/
       }
-      /*
-        Ok, materialization finished. Initialize the access to the temptable
-      */
       join->return_tab= save_return_tab;
       sjm->materialized= TRUE;
     }

sql/sql_join_cache.cc

@@ -158,42 +158,6 @@ JOIN_TAB *next_linear_tab(JOIN* join, JOIN_TAB* tab, bool include_bush_roots);
 
 void JOIN_CACHE::calc_record_fields()
 {
-  //psergey4-todo: prev_cache, or
-  //  - first non-const table if on top level
-  //  - first table inside SJM nest if within sjm nest
-  // this->join_tab is 'our' join_tab
-
-  // No. the right idea: start from ... and walk to the current join_tab 
-  /// with an iterator, skipping
-  // join nests (can do so for now)
-
-  /*
-    The above sucks, too.
-     The right idea: 
-      - for SJM-inner tables, walk only within the nest
-      - for SJM-outer tables, use all preceding tables, including inner ones.
-      eof
-  */
-
-/*  JOIN_TAB *tab = prev_cache ? prev_cache->join_tab :
-                               join->join_tab+join->const_tables;
-*/
-
-/*  JOIN_TAB *tab;
-  if (prev_cache)
-    tab= prev_cache->join_tab;
-  else
-  {
-    if (tab->bush_root_tab)
-    {
-      ;
-    }
-    else
-    {
-      / * top-level * /
-      tab= join->join_tab+join->const_tables;
-    }
-  }*/
   JOIN_TAB *tab;
   if (prev_cache)
     tab= prev_cache->join_tab;
@@ -201,12 +165,18 @@ void JOIN_CACHE::calc_record_fields()
   {
     if (join_tab->bush_root_tab)
     {
-      // inside SJM-Mat nest: pick first one
+      /* 
+        If the tab we're attached to is inside an SJM-nest, start from the
+        first tab in that SJM nest
+      */
       tab= join_tab->bush_root_tab->bush_children->start;
     }
     else
     {
-      // outside SJM-Mat nest: start from first non-const table
+      /*
+        The tab we're attached to is not inside an SJM-nest. Start from the
+        first non-const table.
+      */
       tab= join->join_tab + join->const_tables;
     }
   }

sql/sql_select.cc

@@ -1011,10 +1011,10 @@ JOIN::optimize()
   {
     List_iterator<JOIN_TAB_RANGE> it(join_tab_ranges);
     JOIN_TAB_RANGE *jt_range;
-    bool first= TRUE;
+    uint first_tab_offs= const_tables;
     while ((jt_range= it++))
     {
-      for (JOIN_TAB *tab= jt_range->start + (first ? const_tables : 0); 
+      for (JOIN_TAB *tab= jt_range->start + first_tab_offs;
            tab < jt_range->end; tab++)
       {
         if (*tab->on_expr_ref)
@@ -1025,7 +1025,7 @@ JOIN::optimize()
           (*tab->on_expr_ref)->update_used_tables();
         }
       }
-      first= FALSE;
+      first_tab_offs= 0;
     }
   }
 
@@ -1298,9 +1298,11 @@ JOIN::optimize()
   */
   if (need_tmp || select_distinct || group_list || order)
   {
-    for (uint i = const_tables; i < tables; i++)
+    for (uint i= 0; i < tables; i++)
+    {
+      if (!(table[i]->map & const_table_map))
         table[i]->prepare_for_position();
-
+    }
   }
 
   DBUG_EXECUTE("info",TEST_join(this););
@@ -5868,10 +5870,9 @@ JOIN_TAB *first_linear_tab(JOIN *join, bool after_const_tables)
 {
   JOIN_TAB *first= join->join_tab;
   if (after_const_tables)
-    first += join->const_tables;
+    first+= join->const_tables;
   if (first < join->join_tab + join->top_jtrange_tables)
     return first;
-  else
   return NULL;
 }
 
@@ -5888,24 +5889,24 @@ JOIN_TAB *first_linear_tab(JOIN *join, bool after_const_tables)
     to.)
 */
 
-JOIN_TAB *next_linear_tab(JOIN* join, JOIN_TAB* tab, bool include_bush_roots) //psergey2: added
+JOIN_TAB *next_linear_tab(JOIN* join, JOIN_TAB* tab, bool include_bush_roots)
 {
   if (include_bush_roots && tab->bush_children)
     return tab->bush_children->start;
 
+  DBUG_ASSERT(!tab->last_leaf_in_bush || tab->bush_root_tab);
   if (tab->last_leaf_in_bush)
     tab= tab->bush_root_tab;
 
   if (tab->bush_root_tab)
     return ++tab;
 
-  if (++tab == join->join_tab + join->top_jtrange_tables /*join->join_tab_ranges.head()->end*/)
+  if (++tab == join->join_tab + join->top_jtrange_tables)
     return NULL;
 
   if (!include_bush_roots && tab->bush_children)
-  {
     tab= tab->bush_children->start;
-  }
+
   return tab;
 }
 
@@ -5929,12 +5930,12 @@ JOIN_TAB *next_linear_tab(JOIN* join, JOIN_TAB* tab, bool include_bush_roots) //
 
 */
 
-JOIN_TAB *next_depth_first_tab(JOIN* join, JOIN_TAB* tab) //psergey2: added
+JOIN_TAB *next_depth_first_tab(JOIN* join, JOIN_TAB* tab)
 {
   bool start= FALSE;
   if (tab == NULL)
   {
-    /* This means we're starting. */
+    /* This means we're starting the enumeration */
     if (join->const_tables == join->top_jtrange_tables)
       return NULL;
 
@@ -5945,7 +5946,11 @@ JOIN_TAB *next_depth_first_tab(JOIN* join, JOIN_TAB* tab) //psergey2: added
   if (tab->last_leaf_in_bush)
     return tab->bush_root_tab;
   
-  if ((start? tab: ++tab) == join->join_tab_ranges.head()->end)
+  /* Move to next tab in the array we're traversing*/
+  if (!start)
+    tab++;
+
+  if (tab == join->join_tab_ranges.head()->end)
     return NULL; /* End */
 
   if (tab->bush_children)
@@ -5955,7 +5960,7 @@ JOIN_TAB *next_depth_first_tab(JOIN* join, JOIN_TAB* tab) //psergey2: added
 }
 
 
-static Item *null_ptr= NULL;
+static Item * const null_ptr= NULL;
 
 /*
   Set up join struct according to the picked join order in
@@ -6004,10 +6009,6 @@ get_best_combination(JOIN *join)
 
   fix_semijoin_strategies_for_picked_join_order(join);
    
-  /* 
-    psergey2-todo:  Here: switch to nested structure when copying.
-  */
-
   JOIN_TAB_RANGE *root_range= new JOIN_TAB_RANGE;
   root_range->start= join->join_tab;
   /* root_range->end will be set later */
@@ -6041,7 +6042,7 @@ get_best_combination(JOIN *join)
       j->ref.key_parts=0;
       j->loosescan_match_tab= NULL;  //non-nulls will be set later
       j->use_join_cache= FALSE;
-      j->on_expr_ref= &null_ptr;
+      j->on_expr_ref= (Item**) &null_ptr;
       j->cache= NULL;
 
       /*
@@ -6363,9 +6364,6 @@ JOIN::make_simple_join(JOIN *parent, TABLE *tmp_table)
     DBUG_RETURN(TRUE);                        /* purecov: inspected */
 
   join_tab= parent->join_tab_reexec;
-  //psergey2: hopefully this is ok:
- // join_tab_ranges.head()->start= join_tab;
- // join_tab_ranges.head()->end= join_tab + 1;
   top_jtrange_tables= 1;
 
   table= &parent->table_reexec[0]; parent->table_reexec[0]= tmp_table;
@@ -7701,21 +7699,8 @@ make_join_readinfo(JOIN *join, ulonglong options, uint no_jbuf_after)
     tab->sorted= sorted;
     sorted= 0;                                  // only first must be sorted
 
-
-    //if (sj_is_materialize_strategy(join->best_positions[i].sj_strategy))
-    if (tab->bush_children) // SJM
+    if (tab->bush_children)
     {
-      /* This is a start of semi-join nest */
-      //first_sjm_table= i;
-      //last_sjm_table= i + join->best_positions[i].n_sj_tables;
-      /*
-      psergey2: dont:
-      if (i == join->const_tables)
-        join->first_select= sub_select_sjm;
-      else
-       tab[-1].next_select= sub_select_sjm;
-      */
-
       if (setup_sj_materialization(tab))
         return TRUE;
       table= tab->table;
@@ -12717,9 +12702,9 @@ do_select(JOIN *join,List<Item> *fields,TABLE *table,Procedure *procedure)
   else
   {
     DBUG_ASSERT(join->tables);
-    error= join->first_select(join,join_tab,0);
+    error= sub_select(join,join_tab,0);
     if (error == NESTED_LOOP_OK || error == NESTED_LOOP_NO_MORE_ROWS)
-      error= join->first_select(join,join_tab,1);
+      error= sub_select(join,join_tab,1);
     if (error == NESTED_LOOP_QUERY_LIMIT)
       error= NESTED_LOOP_OK;                    /* select_limit used */
   }

sql/sql_select.h

@@ -190,8 +190,15 @@ typedef struct st_join_table {
     psergey2:  for join tabs that are inside a bush: root of this bush.
   */
   st_join_table *bush_root_tab;
+
+  /* TRUE <=> This join_tab is inside a join bush and is the last leaf tab here */
   bool          last_leaf_in_bush;
   
+  /*
+    ptr  - this is a bush, and ptr points to description of child join_tab
+           range
+    NULL - this join tab has no bush children
+  */
   JOIN_TAB_RANGE *bush_children;
   
   /* Special content for EXPLAIN 'Extra' column or NULL if none */
@@ -500,13 +507,13 @@ protected:
     context can be accessed.  
   */   
   JOIN *join;  
-#if 0
   /*
-    Cardinality of the range of join tables whose fields can be put into the
-    cache. (A table from the range not necessarily contributes to the cache.)
+    JOIN_TAB of the first table that can have it's fields in the join cache. 
+    That is, tables in the [start_tab, tab) range can have their fields in the
+    join cache. 
+    If a join tab in the range represents an SJM-nest, then all tables from the
+    nest can have their fields in the join cache, too.
   */
-  uint tables;
-#endif
   JOIN_TAB *start_tab;
 
   /* 
@@ -1505,7 +1512,6 @@ public:
 
   /* We also maintain a stack of join optimization states in * join->positions[] */
 /******* Join optimization state members end *******/
-  Next_select_func first_select;
   /*
     The cost of best complete join plan found so far during optimization,
     after optimization phase - cost of picked join order (not taking into
@@ -1691,7 +1697,6 @@ public:
     rollup.state= ROLLUP::STATE_NONE;
 
     no_const_tables= FALSE;
-    first_select= sub_select;
   }
 
   int prepare(Item ***rref_pointer_array, TABLE_LIST *tables, uint wind_num,