MWL#68 Subquery optimization: Efficient NOT IN execution with NULLs

This patch implements correct NULL semantics for materialized subquery execution.
The implementation has the following properties and main limitations:
- It passes all query result tests, but fails a number of EXPLAIN tests because of
  changed plans.
- The EXPLAIN output for partial matching is not decided yet.
- It works only when all necessary indexes fit into main memory. Notice that these
  are not the general B-tree/Hash indexes, but instead much more compact ones,
  therefore this limitation may not be a problem in many practical cases.
- It doesn't contain specialized tests.
- In several places the implementation uses methods that are modified copies of
  other similar methods. These cases need to be refactored to avoid code duplication.
- Add a test if the predicate is top-level just before deciding on partial matching.
  If it is top-level, use a more efficient exec method (index lookup).
- Add sorting of indexes according to their selectivity. The code is almost there.
- Needs more comments, and to sync existing ones with the implementation.

sql/item_cmpfunc.h:
  Expose the Arg_comparator of a comparison predicate. This makes it possible to
  directly get the comparison result {-1,0,1}, which is not possible through the
  val_XXX() methods which "fold" such results into a boolean.
sql/item_subselect.cc:
  The core of the implementation of MWL#68.
sql/item_subselect.h:
  The core of the implementation of MWL#68.
sql/opt_subselect.cc:
  Removed the limitation for materialized subquery execution that it is applicable only
  for top-level predicates.
sql/sql_class.cc:
  New class select_materialize_with_stats that collects data statistics about
  the data being inserted into the target table.
sql/sql_class.h:
  New class select_materialize_with_stats that collects data statistics about
  the data being inserted into the target table.
sql/sql_select.cc:
  - more complete initialization of the TABLE object of a temp table.
  - call setup_subquery_materialization at one more exit point.

sql/item_cmpfunc.h

@@ -350,6 +350,7 @@ public:
   CHARSET_INFO *compare_collation() { return cmp.cmp_collation.collation; }
   uint decimal_precision() const { return 1; }
   void top_level_item() { abort_on_null= TRUE; }
+  Arg_comparator *get_comparator() { return &cmp; }
 
   friend class  Arg_comparator;
 };

sql/opt_subselect.cc

@@ -187,11 +187,7 @@ int check_and_do_in_subquery_rewrites(JOIN *join)
              does not call setup_subquery_materialization(). We could make 
              SELECT ... FROM DUAL call that function but that doesn't seem
              to be the case that is worth handling.
-        4. Subquery predicate is a top-level predicate
-           (this implies it is not negated)
-           TODO: this is a limitation that should be lifted once we
-           implement correct NULL semantics (WL#3830)
-        5. Subquery is non-correlated
+        4. Subquery is non-correlated
            TODO:
            This is an overly restrictive condition. It can be extended to:
            (Subquery is non-correlated ||
@@ -199,7 +195,7 @@ int check_and_do_in_subquery_rewrites(JOIN *join)
             (Subquery is correlated to the immediate outer query &&
              Subquery !contains {GROUP BY, ORDER BY [LIMIT],
              aggregate functions}) && subquery predicate is not under "NOT IN"))
-        6. No execution method was already chosen (by a prepared statement).
+        5. 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.
@@ -218,9 +214,8 @@ int check_and_do_in_subquery_rewrites(JOIN *join)
           subquery_types_allow_materialization(in_subs))
       {
         // psergey-todo: duplicated_subselect_card_check: where it's done?
-        if (in_subs->is_top_level_item() &&                             // 4
-            !in_subs->is_correlated &&                                  // 5
-            in_subs->exec_method == Item_in_subselect::NOT_TRANSFORMED) // 6
+        if (!in_subs->is_correlated &&                                  // 4
+            in_subs->exec_method == Item_in_subselect::NOT_TRANSFORMED) // 5
           in_subs->exec_method= Item_in_subselect::MATERIALIZATION;
       }
 

sql/sql_class.cc

@@ -42,6 +42,7 @@
 
 #include "sp_rcontext.h"
 #include "sp_cache.h"
+#include "sql_select.h" /* declares create_tmp_table() */
 
 /*
   The following is used to initialise Table_ident with a internal
@@ -2877,6 +2878,71 @@ bool select_dumpvar::send_eof()
   return 0;
 }
 
+
+bool
+select_materialize_with_stats::
+create_result_table(THD *thd_arg, List<Item> *column_types,
+                    bool is_union_distinct, ulonglong options,
+                    const char *table_alias, bool bit_fields_as_long)
+{
+  DBUG_ASSERT(table == 0);
+  tmp_table_param.field_count= column_types->elements;
+  tmp_table_param.bit_fields_as_long= bit_fields_as_long;
+
+  if (! (table= create_tmp_table(thd_arg, &tmp_table_param, *column_types,
+                                 (ORDER*) 0, is_union_distinct, 1,
+                                 options, HA_POS_ERROR, (char*) table_alias)))
+    return TRUE;
+
+  col_stat= (Column_statistics*) table->in_use->alloc(table->s->fields *
+                                                      sizeof(Column_statistics));
+  if (!stat)
+    return TRUE;
+
+  cleanup();
+
+  table->file->extra(HA_EXTRA_WRITE_CACHE);
+  table->file->extra(HA_EXTRA_IGNORE_DUP_KEY);
+  return FALSE;
+}
+
+
+/**
+  Override select_union::send_data to analyze each row for NULLs and to
+  update null_statistics before sending data to the client.
+
+  @return TRUE if fatal error when sending data to the client
+  @return FALSE on success
+*/
+
+bool select_materialize_with_stats::send_data(List<Item> &items)
+{
+  List_iterator_fast<Item> item_it(items);
+  Item *cur_item;
+  Column_statistics *cur_col_stat= col_stat;
+  uint nulls_in_row= 0;
+
+  ++count_rows;
+
+  while ((cur_item= item_it++))
+  {
+    if (cur_item->is_null())
+    {
+      ++cur_col_stat->null_count;
+      cur_col_stat->max_null_row= count_rows;
+      if (!cur_col_stat->min_null_row)
+        cur_col_stat->min_null_row= count_rows;
+      ++nulls_in_row;
+    }
+    ++cur_col_stat;
+  }
+  if (nulls_in_row > max_nulls_in_row)
+    max_nulls_in_row= nulls_in_row;
+
+  return select_union::send_data(items);
+}
+
+
 /****************************************************************************
   TMP_TABLE_PARAM
 ****************************************************************************/

sql/sql_class.h

@@ -2740,17 +2740,18 @@ public:
 
 class select_union :public select_result_interceptor
 {
+protected:
   TMP_TABLE_PARAM tmp_table_param;
 public:
   TABLE *table;
 
-  select_union() :table(0) {}
+  select_union() :table(0) { tmp_table_param.init(); }
   int prepare(List<Item> &list, SELECT_LEX_UNIT *u);
   bool send_data(List<Item> &items);
   bool send_eof();
   bool flush();
 
-  bool create_result_table(THD *thd, List<Item> *column_types,
+  virtual bool create_result_table(THD *thd, List<Item> *column_types,
                                    bool is_distinct, ulonglong options,
                                    const char *alias, bool bit_fields_as_long);
 };
@@ -2776,6 +2777,74 @@ public:
   bool send_data(List<Item> &items);
 };
 
+
+/*
+  This class specializes select_union to collect statistics about the
+  data stored in the temp table. Currently the class collects statistcs
+  about NULLs.
+*/
+
+class select_materialize_with_stats : public select_union
+{
+protected:
+  class Column_statistics
+  {
+  public:
+    /* Count of NULLs per column. */
+    ha_rows null_count;
+    /* The row number that contains the first NULL in a column. */
+    ha_rows min_null_row;
+    /* The row number that contains the last NULL in a column. */
+    ha_rows max_null_row;
+  };
+
+  /* Array of statistics data per column. */
+  Column_statistics* col_stat;
+
+  /*
+    The number of columns in the biggest sub-row that consists of only
+    NULL values.
+  */
+  ha_rows max_nulls_in_row;
+  /*
+    Count of rows writtent to the temp table. This is redundant as it is
+    already stored in handler::stats.records, however that one is relatively
+    expensive to compute (given we need that for evry row).
+  */
+  ha_rows count_rows;
+
+public:
+  select_materialize_with_stats() {}
+  virtual bool create_result_table(THD *thd, List<Item> *column_types,
+                                   bool is_distinct, ulonglong options,
+                                   const char *alias, bool bit_fields_as_long);
+  bool init_result_table(ulonglong select_options);
+  bool send_data(List<Item> &items);
+  void cleanup()
+  {
+    memset(col_stat, 0, table->s->fields * sizeof(Column_statistics));
+    max_nulls_in_row= 0;
+    count_rows= 0;
+  }
+  ha_rows get_null_count_of_col(uint idx)
+  {
+    DBUG_ASSERT(idx < table->s->fields);
+    return col_stat[idx].null_count;
+  }
+  ha_rows get_max_null_of_col(uint idx)
+  {
+    DBUG_ASSERT(idx < table->s->fields);
+    return col_stat[idx].max_null_row;
+  }
+  ha_rows get_min_null_of_col(uint idx)
+  {
+    DBUG_ASSERT(idx < table->s->fields);
+    return col_stat[idx].min_null_row;
+  }
+  ha_rows get_max_nulls_in_row() { return max_nulls_in_row; }
+};
+
+
 /* used in independent ALL/ANY optimisation */
 class select_max_min_finder_subselect :public select_subselect
 {

sql/sql_select.cc

@@ -874,6 +874,9 @@ JOIN::optimize()
   {
     DBUG_PRINT("info",("No tables"));
     error= 0;
+    /* Create all structures needed for materialized subquery execution. */
+    if (setup_subquery_materialization())
+      DBUG_RETURN(1);
     DBUG_RETURN(0);
   }
   error= -1;					// Error is sent to client
@@ -11233,7 +11236,7 @@ create_tmp_table(THD *thd,TMP_TABLE_PARAM *param,List<Item> &fields,
     param->group_buff=group_buff;
     share->keys=1;
     share->uniques= test(using_unique_constraint);
-    table->key_info=keyinfo;
+    table->key_info= table->s->key_info= keyinfo;
     keyinfo->key_part=key_part_info;
     keyinfo->flags=HA_NOSAME;
     keyinfo->usable_key_parts=keyinfo->key_parts= param->group_parts;
@@ -11319,7 +11322,7 @@ create_tmp_table(THD *thd,TMP_TABLE_PARAM *param,List<Item> &fields,
                      keyinfo->key_parts * sizeof(KEY_PART_INFO))))
       goto err;
     bzero((void*) key_part_info, keyinfo->key_parts * sizeof(KEY_PART_INFO));
-    table->key_info=keyinfo;
+    table->key_info= table->s->key_info= keyinfo;
     keyinfo->key_part=key_part_info;
     keyinfo->flags=HA_NOSAME | HA_NULL_ARE_EQUAL;
     keyinfo->key_length= 0;  // Will compute the sum of the parts below.