Merge

include/my_tree.h

@@ -30,7 +30,17 @@ extern "C" {
 
 #define tree_set_pointer(element,ptr) *((uchar **) (element+1))=((uchar*) (ptr))
 
+/*
+  A tree with its flag set to TREE_ONLY_DUPS behaves differently on inserting 
+  an element that is not in the tree:
+  the element is not added at all, but instead tree_insert() returns a special
+  address TREE_ELEMENT_UNIQUE as an indication that the function has not failed
+  due to lack of memory. 
+*/
+
+#define TREE_ELEMENT_UNIQUE ((TREE_ELEMENT *) 1)
 #define TREE_NO_DUPS 1
+#define TREE_ONLY_DUPS 2
 
 typedef enum { left_root_right, right_root_left } TREE_WALK;
 typedef uint32 element_count;

mysql-test/include/index_merge1.inc

@@ -539,7 +539,7 @@ INSERT INTO t1 SELECT * FROM t1;
 INSERT INTO t1 SELECT * FROM t1;
 INSERT INTO t1 SELECT * FROM t1;
 INSERT INTO t1 SELECT * FROM t1;
-SET SESSION sort_buffer_size=1;
+SET SESSION sort_buffer_size=1024*8;
 EXPLAIN 
 SELECT * FROM t1 FORCE INDEX(a,b) WHERE a LIKE 'a%' OR b LIKE 'b%' 
   ORDER BY a,b;

mysql-test/r/index_merge_innodb.result

+set @optimizer_switch_save= @@optimizer_switch;
+set optimizer_switch='index_merge_sort_intersection=off';
 #---------------- Index merge test 2 -------------------------------------------
 SET SESSION STORAGE_ENGINE = InnoDB;
 drop table if exists t1,t2;
@@ -709,3 +711,4 @@ WHERE a BETWEEN 2 AND 7 OR pk=1000000) AS t;
 COUNT(*)
 6145
 DROP TABLE t1;
+set optimizer_switch= @optimizer_switch_save;

mysql-test/r/index_merge_myisam.result

+set @optimizer_switch_save= @@optimizer_switch;
+set optimizer_switch='index_merge_sort_intersection=off';
 #---------------- Index merge test 1 -------------------------------------------
 SET SESSION STORAGE_ENGINE = MyISAM;
 drop table if exists t0, t1, t2, t3, t4;
@@ -569,9 +571,7 @@ INSERT INTO t1 SELECT * FROM t1;
 INSERT INTO t1 SELECT * FROM t1;
 INSERT INTO t1 SELECT * FROM t1;
 INSERT INTO t1 SELECT * FROM t1;
-SET SESSION sort_buffer_size=1;
-Warnings:
-Warning	1292	Truncated incorrect sort_buffer_size value: '1'
+SET SESSION sort_buffer_size=1024*8;
 EXPLAIN 
 SELECT * FROM t1 FORCE INDEX(a,b) WHERE a LIKE 'a%' OR b LIKE 'b%' 
 ORDER BY a,b;
@@ -1494,19 +1494,19 @@ DROP TABLE t1,t2;
 #
 select @@optimizer_switch;
 @@optimizer_switch
-index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on
+index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on,index_merge_sort_intersection=off
 set optimizer_switch='index_merge=off,index_merge_union=off';
 select @@optimizer_switch;
 @@optimizer_switch
-index_merge=off,index_merge_union=off,index_merge_sort_union=on,index_merge_intersection=on
+index_merge=off,index_merge_union=off,index_merge_sort_union=on,index_merge_intersection=on,index_merge_sort_intersection=off
 set optimizer_switch='index_merge_union=on';
 select @@optimizer_switch;
 @@optimizer_switch
-index_merge=off,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on
+index_merge=off,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on,index_merge_sort_intersection=off
 set optimizer_switch='default,index_merge_sort_union=off';
 select @@optimizer_switch;
 @@optimizer_switch
-index_merge=on,index_merge_union=on,index_merge_sort_union=off,index_merge_intersection=on
+index_merge=on,index_merge_union=on,index_merge_sort_union=off,index_merge_intersection=on,index_merge_sort_intersection=off
 set optimizer_switch=4;
 ERROR 42000: Variable 'optimizer_switch' can't be set to the value of '4'
 set optimizer_switch=NULL;
@@ -1533,21 +1533,21 @@ set optimizer_switch=default;
 set optimizer_switch='index_merge=off,index_merge_union=off,default';
 select @@optimizer_switch;
 @@optimizer_switch
-index_merge=off,index_merge_union=off,index_merge_sort_union=on,index_merge_intersection=on
+index_merge=off,index_merge_union=off,index_merge_sort_union=on,index_merge_intersection=on,index_merge_sort_intersection=off
 set optimizer_switch=default;
 select @@global.optimizer_switch;
 @@global.optimizer_switch
-index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on
+index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on,index_merge_sort_intersection=off
 set @@global.optimizer_switch=default;
 select @@global.optimizer_switch;
 @@global.optimizer_switch
-index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on
+index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on,index_merge_sort_intersection=off
 #
 # Check index_merge's @@optimizer_switch flags
 #
 select @@optimizer_switch;
 @@optimizer_switch
-index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on
+index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on,index_merge_sort_intersection=off
 create table t0 (a int);
 insert into t0 values (0),(1),(2),(3),(4),(5),(6),(7),(8),(9);
 create table t1 (a int, b int, c int, filler char(100), 
@@ -1624,7 +1624,7 @@ explain select * from t1 where a=10 and b=10;
 id	select_type	table	type	possible_keys	key	key_len	ref	rows	Extra
 1	SIMPLE	t1	ref	a,b	a	5	const	49	Using where
 No intersect if it is disabled:
-set optimizer_switch='default,index_merge_intersection=off';
+set optimizer_switch='default,index_merge_sort_intersection=off,index_merge_intersection=off';
 explain select * from t1 where a=10 and b=10;
 id	select_type	table	type	possible_keys	key	key_len	ref	rows	Extra
 1	SIMPLE	t1	ref	a,b	a	5	const	49	Using where
@@ -1657,5 +1657,6 @@ id	select_type	table	type	possible_keys	key	key_len	ref	rows	Extra
 set optimizer_switch=default;
 show variables like 'optimizer_switch';
 Variable_name	Value
-optimizer_switch	index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on
+optimizer_switch	index_merge=on,index_merge_union=on,index_merge_sort_union=on,index_merge_intersection=on,index_merge_sort_intersection=off
 drop table t0, t1;
+set optimizer_switch= @optimizer_switch_save;

mysql-test/r/order_by.result

@@ -1421,9 +1421,9 @@ DROP TABLE t1;
 #
 create table t1(a int, b tinytext);
 insert into t1 values (1,2),(3,2);
-set session sort_buffer_size= 30000;
+set session sort_buffer_size= 1000;
 Warnings:
-Warning	1292	Truncated incorrect sort_buffer_size value: '30000'
+Warning	1292	Truncated incorrect sort_buffer_size value: '1000'
 set session max_sort_length= 2180;
 select * from t1 order by b;
 ERROR HY001: Out of sort memory; increase server sort buffer size

mysql-test/r/range_vs_index_merge.result

@@ -38,6 +38,7 @@ SELECT COUNT(*) FROM CountryLanguage;
 COUNT(*)
 984
 CREATE INDEX Name ON City(Name);
+set session optimizer_switch='index_merge_sort_intersection=off';
 EXPLAIN
 SELECT * FROM City
 WHERE (Population >= 100000 OR Name LIKE 'P%' OR Population < 100000);
@@ -1366,3 +1367,4 @@ f1	f2	f3	f4
 9	0	2	6
 SET SESSION optimizer_switch=DEFAULT;
 DROP TABLE t1;
+set session optimizer_switch='index_merge_sort_intersection=default';

mysql-test/r/range_vs_index_merge_innodb.result

@@ -39,6 +39,7 @@ SELECT COUNT(*) FROM CountryLanguage;
 COUNT(*)
 984
 CREATE INDEX Name ON City(Name);
+set session optimizer_switch='index_merge_sort_intersection=off';
 EXPLAIN
 SELECT * FROM City
 WHERE (Population >= 100000 OR Name LIKE 'P%' OR Population < 100000);
@@ -1367,4 +1368,5 @@ f1	f2	f3	f4
 9	0	2	6
 SET SESSION optimizer_switch=DEFAULT;
 DROP TABLE t1;
+set session optimizer_switch='index_merge_sort_intersection=default';
 SET SESSION STORAGE_ENGINE=DEFAULT;

mysql-test/r/subselect.result

@@ -3667,8 +3667,6 @@ CREATE TABLE t1 (a int, b int auto_increment, PRIMARY KEY (b));
 CREATE TABLE t2 (x int auto_increment, y int, z int,
 PRIMARY KEY (x), FOREIGN KEY (y) REFERENCES t1 (b));
 SET SESSION sort_buffer_size = 32 * 1024;
-Warnings:
-Warning	1292	Truncated incorrect sort_buffer_size value: '32768'
 SELECT SQL_NO_CACHE COUNT(*)
 FROM (SELECT  a, b, (SELECT x FROM t2 WHERE y=b ORDER BY z DESC LIMIT 1) c
 FROM t1) t;
@@ -4104,8 +4102,6 @@ INSERT INTO `t1` VALUES ('asdf','2007-02-08 01:11:26');
 INSERT INTO `t2` VALUES ('abcdefghijk');
 INSERT INTO `t2` VALUES ('asdf');
 SET session sort_buffer_size=8192;
-Warnings:
-Warning	1292	Truncated incorrect sort_buffer_size value: '8192'
 SELECT (SELECT 1 FROM  t1 WHERE t1.a=t2.a ORDER BY t1.b LIMIT 1) AS d1 FROM t2;
 d1
 1

mysql-test/r/variables.result

@@ -971,6 +971,7 @@ tmpdir	#
 select * from information_schema.session_variables where variable_name like 'tmpdir';
 VARIABLE_NAME	VARIABLE_VALUE
 TMPDIR	#
+set sort_buffer_size=1024*8;
 select @@ssl_ca, @@ssl_capath, @@ssl_cert, @@ssl_cipher, @@ssl_key;
 @@ssl_ca	@@ssl_capath	@@ssl_cert	@@ssl_cipher	@@ssl_key
 #	#	#	#	#

mysql-test/suite/innodb/r/innodb_mysql.result

@@ -679,8 +679,6 @@ INSERT INTO t1(b,c) SELECT b,c FROM t2;
 UPDATE t2 SET c='2007-01-03';
 INSERT INTO t1(b,c) SELECT b,c FROM t2;
 set @@sort_buffer_size=8192;
-Warnings:
-Warning	1292	Truncated incorrect sort_buffer_size value: '8192'
 SELECT COUNT(*) FROM t1;
 COUNT(*)
 3072

mysql-test/suite/innodb_plugin/r/innodb_mysql.result

@@ -679,8 +679,6 @@ INSERT INTO t1(b,c) SELECT b,c FROM t2;
 UPDATE t2 SET c='2007-01-03';
 INSERT INTO t1(b,c) SELECT b,c FROM t2;
 set @@sort_buffer_size=8192;
-Warnings:
-Warning	1292	Truncated incorrect sort_buffer_size value: '8192'
 SELECT COUNT(*) FROM t1;
 COUNT(*)
 3072

mysql-test/suite/pbxt/r/subselect.result

@@ -3667,8 +3667,6 @@ CREATE TABLE t1 (a int, b int auto_increment, PRIMARY KEY (b));
 CREATE TABLE t2 (x int auto_increment, y int, z int,
 PRIMARY KEY (x), FOREIGN KEY (y) REFERENCES t1 (b));
 SET SESSION sort_buffer_size = 32 * 1024;
-Warnings:
-Warning	1292	Truncated incorrect sort_buffer_size value: '32768'
 SELECT SQL_NO_CACHE COUNT(*) 
 FROM (SELECT  a, b, (SELECT x FROM t2 WHERE y=b ORDER BY z DESC LIMIT 1) c
 FROM t1) t;
@@ -4104,8 +4102,6 @@ INSERT INTO `t1` VALUES ('asdf','2007-02-08 01:11:26');
 INSERT INTO `t2` VALUES ('abcdefghijk');
 INSERT INTO `t2` VALUES ('asdf');
 SET session sort_buffer_size=8192;
-Warnings:
-Warning	1292	Truncated incorrect sort_buffer_size value: '8192'
 SELECT (SELECT 1 FROM  t1 WHERE t1.a=t2.a ORDER BY t1.b LIMIT 1) AS d1 FROM t2;
 d1
 1

mysql-test/t/index_intersect_innodb.test

+--source include/have_innodb.inc
+
+SET SESSION STORAGE_ENGINE='InnoDB';
+
+--source t/index_intersect.test
+
+SET SESSION STORAGE_ENGINE=DEFAULT;

mysql-test/t/index_merge_innodb.test

@@ -18,6 +18,11 @@ let $engine_type= InnoDB;
 # InnoDB does not support Merge tables (affects include/index_merge1.inc)
 let $merge_table_support= 0;
 
+set @optimizer_switch_save= @@optimizer_switch;
+
+set optimizer_switch='index_merge_sort_intersection=off';
+
+
 # The first two tests are disabled because of non deterministic explain output.
 # If include/index_merge1.inc can be enabled for InnoDB and all other
 # storage engines, please remove the subtest for Bug#21277 from
@@ -84,3 +89,5 @@ SELECT COUNT(*) FROM
      WHERE a BETWEEN 2 AND 7 OR pk=1000000) AS t;
 
 DROP TABLE t1;
+
+set optimizer_switch= @optimizer_switch_save;

mysql-test/t/index_merge_myisam.test

@@ -14,6 +14,10 @@ let $engine_type= MyISAM;
 # MyISAM supports Merge tables
 let $merge_table_support= 1;
 
+set @optimizer_switch_save= @@optimizer_switch;
+
+set optimizer_switch='index_merge_sort_intersection=off';
+
 --source include/index_merge1.inc
 --source include/index_merge_ror.inc
 --source include/index_merge2.inc
@@ -164,7 +168,7 @@ set optimizer_switch='default,index_merge=off';
 explain select * from t1 where a=10 and b=10;
 
 --echo No intersect if it is disabled:
-set optimizer_switch='default,index_merge_intersection=off';
+set optimizer_switch='default,index_merge_sort_intersection=off,index_merge_intersection=off';
 explain select * from t1 where a=10 and b=10;
 
 --echo Do intersect when union was disabled
@@ -195,3 +199,5 @@ show variables like 'optimizer_switch';
 
 drop table t0, t1;
 
+set optimizer_switch= @optimizer_switch_save;
+

mysql-test/t/order_by.test

@@ -843,7 +843,7 @@ DROP TABLE t1;
 --echo #
 create table t1(a int, b tinytext);
 insert into t1 values (1,2),(3,2);
-set session sort_buffer_size= 30000;
+set session sort_buffer_size= 1000;
 set session max_sort_length= 2180;
 --error 1038
 select * from t1 order by b;

mysql-test/t/range_vs_index_merge.test

@@ -33,6 +33,8 @@ ANALYZE TABLE City;
 --enable_result_log
 --enable_query_log
 
+set session optimizer_switch='index_merge_sort_intersection=off';
+
 # The following 4 queries are added for code coverage 
 
 #the exptected # of rows differ on 32-bit and 64-bit platforms for innodb 
@@ -1007,3 +1009,6 @@ SELECT * FROM t1 FORCE KEY (PRIMARY,f3,f4)
 SET SESSION optimizer_switch=DEFAULT;
 
 DROP TABLE t1;
+
+#the following command must be the last one in the file
+set session optimizer_switch='index_merge_sort_intersection=default';

mysql-test/t/variables.test

@@ -736,6 +736,7 @@ select * from information_schema.session_variables where variable_name like 'tmp
 # Bug #19606: make ssl settings available via SHOW VARIABLES and @@variables
 # 
 # Don't actually output, since it depends on the system
+set sort_buffer_size=1024*8;
 --replace_column 1 # 2 # 3 # 4 # 5 #
 select @@ssl_ca, @@ssl_capath, @@ssl_cert, @@ssl_cipher, @@ssl_key;
 --replace_column 2 #

mysys/tree.c

@@ -221,6 +221,8 @@ TREE_ELEMENT *tree_insert(TREE *tree, void *key, uint key_size,
   }
   if (element == &tree->null_element)
   {
+    if (tree->flag & TREE_ONLY_DUPS)
+      return((TREE_ELEMENT *) 1);
     uint alloc_size=sizeof(TREE_ELEMENT)+key_size+tree->size_of_element;
     tree->allocated+=alloc_size;
 

sql/filesort.cc

@@ -50,10 +50,6 @@ static int write_keys(SORTPARAM *param,uchar * *sort_keys,
 		      uint count, IO_CACHE *buffer_file, IO_CACHE *tempfile);
 static void make_sortkey(SORTPARAM *param,uchar *to, uchar *ref_pos);
 static void register_used_fields(SORTPARAM *param);
-static int merge_index(SORTPARAM *param,uchar *sort_buffer,
-		       BUFFPEK *buffpek,
-		       uint maxbuffer,IO_CACHE *tempfile,
-		       IO_CACHE *outfile);
 static bool save_index(SORTPARAM *param,uchar **sort_keys, uint count, 
                        FILESORT_INFO *table_sort);
 static uint suffix_length(ulong string_length);
@@ -145,8 +141,6 @@ ha_rows filesort(THD *thd, TABLE *table, SORT_FIELD *sortorder, uint s_length,
   /* filesort cannot handle zero-length records. */
   DBUG_ASSERT(param.sort_length);
   param.ref_length= table->file->ref_length;
-  param.addon_field= 0;
-  param.addon_length= 0;
   if (!(table->file->ha_table_flags() & HA_FAST_KEY_READ) &&
       !table->fulltext_searched && !sort_positions)
   {
@@ -1200,8 +1194,11 @@ int merge_buffers(SORTPARAM *param, IO_CACHE *from_file,
   QUEUE queue;
   qsort2_cmp cmp;
   void *first_cmp_arg;
-  volatile THD::killed_state *killed= &current_thd->killed;
+  element_count dupl_count;
+  uchar *src;
   THD::killed_state not_killable;
+  uchar *unique_buff= param->unique_buff;
+  volatile THD::killed_state *killed= &current_thd->killed;
   DBUG_ENTER("merge_buffers");
 
   status_var_increment(current_thd->status_var.filesort_merge_passes);
@@ -1216,7 +1213,13 @@ int merge_buffers(SORTPARAM *param, IO_CACHE *from_file,
   rec_length= param->rec_length;
   res_length= param->res_length;
   sort_length= param->sort_length;
-  offset= rec_length-res_length;
+  uint dupl_count_ofs= rec_length-sizeof(element_count);
+  uint min_dupl_count= param->min_dupl_count;
+  bool check_dupl_count= flag && min_dupl_count;
+  offset= (rec_length-
+           (flag && min_dupl_count ? sizeof(dupl_count) : 0)-res_length);
+  uint wr_len= flag ? res_length : rec_length;
+  uint wr_offset= flag ? offset : 0;
   maxcount= (ulong) (param->keys/((uint) (Tb-Fb) +1));
   to_start_filepos= my_b_tell(to_file);
   strpos= sort_buffer;
@@ -1225,7 +1228,7 @@ int merge_buffers(SORTPARAM *param, IO_CACHE *from_file,
   /* The following will fire if there is not enough space in sort_buffer */
   DBUG_ASSERT(maxcount!=0);
   
-  if (param->unique_buff)
+  if (unique_buff)
   {
     cmp= param->compare;
     first_cmp_arg= (void *) &param->cmp_context;
@@ -1250,28 +1253,29 @@ int merge_buffers(SORTPARAM *param, IO_CACHE *from_file,
     queue_insert(&queue, (uchar*) buffpek);
   }
 
-  if (param->unique_buff)
+  if (unique_buff)
   {
     /* 
        Called by Unique::get()
-       Copy the first argument to param->unique_buff for unique removal.
+       Copy the first argument to unique_buff for unique removal.
        Store it also in 'to_file'.
-
-       This is safe as we know that there is always more than one element
-       in each block to merge (This is guaranteed by the Unique:: algorithm
     */
     buffpek= (BUFFPEK*) queue_top(&queue);
-    memcpy(param->unique_buff, buffpek->key, rec_length);
-    if (my_b_write(to_file, (uchar*) buffpek->key, rec_length))
-    {
-      error=1; goto err;                        /* purecov: inspected */
-    }
+    memcpy(unique_buff, buffpek->key, rec_length);
+    if (min_dupl_count)
+      memcpy(&dupl_count, unique_buff+dupl_count_ofs, 
+             sizeof(dupl_count));
     buffpek->key+= rec_length;
-    buffpek->mem_count--;
-    if (!--max_rows)
+    if (! --buffpek->mem_count)
     {
-      error= 0;                                       /* purecov: inspected */
-      goto end;                                       /* purecov: inspected */
+      if (!(error= (int) read_to_buffer(from_file, buffpek,
+                                        rec_length)))
+      {
+        VOID(queue_remove(&queue,0));
+        reuse_freed_buff(&queue, buffpek, rec_length);
+      }
+      else if (error == -1)
+        goto err;                        /* purecov: inspected */ 
     }
     queue_replaced(&queue);                        // Top element has been used
   }
@@ -1287,27 +1291,50 @@ int merge_buffers(SORTPARAM *param, IO_CACHE *from_file,
     for (;;)
     {
       buffpek= (BUFFPEK*) queue_top(&queue);
+      src= buffpek->key;
       if (cmp)                                        // Remove duplicates
       {
-        if (!(*cmp)(first_cmp_arg, &(param->unique_buff),
+        if (!(*cmp)(first_cmp_arg, &unique_buff,
                     (uchar**) &buffpek->key))
+	{
+          if (min_dupl_count)
+	  {
+            element_count cnt;
+            memcpy(&cnt, (uchar *) buffpek->key+dupl_count_ofs, sizeof(cnt));
+            dupl_count+= cnt;
+          }
           goto skip_duplicate;
-            memcpy(param->unique_buff, (uchar*) buffpek->key, rec_length);
         }
-      if (flag == 0)
+        if (min_dupl_count)
 	{
-        if (my_b_write(to_file,(uchar*) buffpek->key, rec_length))
-        {
-          error=1; goto err;                        /* purecov: inspected */
+          memcpy(unique_buff+dupl_count_ofs, &dupl_count,
+                 sizeof(dupl_count));
         }
+	src= unique_buff;
       }
-      else
+        
+      /* 
+        Do not write into the output file if this is the final merge called
+        for a Unique object used for intersection and dupl_count is less
+        than min_dupl_count.
+        If the Unique object is used to intersect N sets of unique elements
+        then for any element:
+        dupl_count >= N <=> the element is occurred in each of these N sets.
+      */          
+      if (!check_dupl_count || dupl_count >= min_dupl_count)
       {
-        if (my_b_write(to_file, (uchar*) buffpek->key+offset, res_length))
+        if (my_b_write(to_file, src+wr_offset, wr_len))
         {
           error=1; goto err;                        /* purecov: inspected */
         }
       }
+      if (cmp)
+      {   
+        memcpy(unique_buff, (uchar*) buffpek->key, rec_length);
+        if (min_dupl_count)
+          memcpy(&dupl_count, unique_buff+dupl_count_ofs, 
+                 sizeof(dupl_count));
+      }
       if (!--max_rows)
       {
         error= 0;                               /* purecov: inspected */
@@ -1318,7 +1345,7 @@ int merge_buffers(SORTPARAM *param, IO_CACHE *from_file,
       buffpek->key+= rec_length;
       if (! --buffpek->mem_count)
       {
-        if (!(error= (int) read_to_buffer(from_file,buffpek,
+        if (!(error= (int) read_to_buffer(from_file, buffpek,
                                           rec_length)))
         {
           VOID(queue_remove(&queue,0));
@@ -1341,11 +1368,35 @@ int merge_buffers(SORTPARAM *param, IO_CACHE *from_file,
   */
   if (cmp)
   {
-    if (!(*cmp)(first_cmp_arg, &(param->unique_buff), (uchar**) &buffpek->key))
+    if (!(*cmp)(first_cmp_arg, &unique_buff, (uchar**) &buffpek->key))
     {
-      buffpek->key+= rec_length;         // Remove duplicate
+      if (min_dupl_count)
+      {
+        element_count cnt;
+        memcpy(&cnt, (uchar *) buffpek->key+dupl_count_ofs, sizeof(cnt));
+        dupl_count+= cnt;
+      }
+      buffpek->key+= rec_length;         
       --buffpek->mem_count;
     }
+
+    if (min_dupl_count)
+      memcpy(unique_buff+dupl_count_ofs, &dupl_count,
+             sizeof(dupl_count));
+
+    if (!check_dupl_count || dupl_count >= min_dupl_count)
+    {
+      src= unique_buff;
+      if (my_b_write(to_file, src+wr_offset, wr_len))
+      {
+        error=1; goto err;                        /* purecov: inspected */
+      }
+      if (!--max_rows)
+      {
+        error= 0;                               
+        goto end;                             
+      }
+    }   
   }
 
   do
@@ -1358,7 +1409,7 @@ int merge_buffers(SORTPARAM *param, IO_CACHE *from_file,
     max_rows-= buffpek->mem_count;
     if (flag == 0)
     {
-      if (my_b_write(to_file,(uchar*) buffpek->key,
+      if (my_b_write(to_file, (uchar*) buffpek->key,
                      (rec_length*buffpek->mem_count)))
       {
         error= 1; goto err;                        /* purecov: inspected */
@@ -1367,19 +1418,25 @@ int merge_buffers(SORTPARAM *param, IO_CACHE *from_file,
     else
     {
       register uchar *end;
-      strpos= buffpek->key+offset;
-      for (end= strpos+buffpek->mem_count*rec_length ;
-           strpos != end ;
-           strpos+= rec_length)
+      src= buffpek->key+offset;
+      for (end= src+buffpek->mem_count*rec_length ;
+           src != end ;
+           src+= rec_length)
       {
-        if (my_b_write(to_file, strpos, res_length))
+        if (check_dupl_count)
+        {
+          memcpy((uchar *) &dupl_count, src+dupl_count_ofs, sizeof(dupl_count)); 
+          if (dupl_count < min_dupl_count)
+	    continue;
+        }
+        if (my_b_write(to_file, src, wr_len))
         {
           error=1; goto err;                        
         }
       }
     }
   }
-  while ((error=(int) read_to_buffer(from_file,buffpek, rec_length))
+  while ((error=(int) read_to_buffer(from_file, buffpek, rec_length))
          != -1 && error != 0);
 
 end:
@@ -1393,7 +1450,7 @@ int merge_buffers(SORTPARAM *param, IO_CACHE *from_file,
 
 	/* Do a merge to output-file (save only positions) */
 
-static int merge_index(SORTPARAM *param, uchar *sort_buffer,
+int merge_index(SORTPARAM *param, uchar *sort_buffer,
 		       BUFFPEK *buffpek, uint maxbuffer,
 		       IO_CACHE *tempfile, IO_CACHE *outfile)
 {
@@ -1699,3 +1756,4 @@ void change_double_for_sort(double nr,uchar *to)
     }
   }
 }
+

sql/mysql_priv.h

@@ -340,6 +340,9 @@ class Default_object_creation_ctx : public Object_creation_ctx
 */
 #define TIME_FOR_COMPARE_ROWID  (TIME_FOR_COMPARE*100)
 
+/* cost1 is better that cost2 only if cost1 + COST_EPS < cost2 */
+#define COST_EPS  0.001
+
 /*
   For sequential disk seeks the cost formula is:
     DISK_SEEK_BASE_COST + DISK_SEEK_PROP_COST * #blocks_to_skip  
@@ -542,12 +545,13 @@ class Default_object_creation_ctx : public Object_creation_ctx
 #define OPTIMIZER_SWITCH_INDEX_MERGE_UNION 2
 #define OPTIMIZER_SWITCH_INDEX_MERGE_SORT_UNION 4
 #define OPTIMIZER_SWITCH_INDEX_MERGE_INTERSECT 8
+#define OPTIMIZER_SWITCH_INDEX_MERGE_SORT_INTERSECT 16
 
 #ifdef DBUG_OFF
-#  define OPTIMIZER_SWITCH_LAST 16
-#else
-#  define OPTIMIZER_SWITCH_TABLE_ELIMINATION 16
 #  define OPTIMIZER_SWITCH_LAST 32
+#else
+#  define OPTIMIZER_SWITCH_TABLE_ELIMINATION 32
+#  define OPTIMIZER_SWITCH_LAST 64
 #endif
 
 #ifdef DBUG_OFF 
@@ -2233,6 +2237,8 @@ ha_rows filesort(THD *thd, TABLE *form,struct st_sort_field *sortorder,
 		 ha_rows max_rows, bool sort_positions,
                  ha_rows *examined_rows);
 void filesort_free_buffers(TABLE *table, bool full);
+double get_merge_many_buffs_cost(uint *buffer, uint last_n_elems,
+                                 int elem_size);
 void change_double_for_sort(double nr,uchar *to);
 double my_double_round(double value, longlong dec, bool dec_unsigned,
                        bool truncate);

sql/mysqld.cc

@@ -338,7 +338,7 @@ TYPELIB sql_mode_typelib= { array_elements(sql_mode_names)-1,"",
 static const char *optimizer_switch_names[]=
 {
   "index_merge","index_merge_union","index_merge_sort_union",
-  "index_merge_intersection",
+  "index_merge_intersection","index_merge_sort_intersection",
 #ifndef DBUG_OFF
   "table_elimination",
 #endif
@@ -352,6 +352,7 @@ static const unsigned int optimizer_switch_names_len[]=
   sizeof("index_merge_union") - 1,
   sizeof("index_merge_sort_union") - 1,
   sizeof("index_merge_intersection") - 1,
+  sizeof("index_merge_sort_intersection") - 1,
 #ifndef DBUG_OFF
   sizeof("table_elimination") - 1,
 #endif
@@ -431,7 +432,8 @@ static const char *sql_mode_str= "OFF";
 /* Text representation for OPTIMIZER_SWITCH_DEFAULT */
 static const char *optimizer_switch_str="index_merge=on,index_merge_union=on,"
                                         "index_merge_sort_union=on,"
-                                        "index_merge_intersection=on"
+                                        "index_merge_intersection=on,"
+                                        "index_merge_sort_intersection=off"
 #ifndef DBUG_OFF                                        
                                         ",table_elimination=on";
 #else
@@ -7297,7 +7299,8 @@ thread is in the relay logs.",
    0, GET_ULONG, OPT_ARG, MAX_TABLES+1, 0, MAX_TABLES+2, 0, 1, 0},
   {"optimizer_switch", OPT_OPTIMIZER_SWITCH,
    "optimizer_switch=option=val[,option=val...], where option={index_merge, "
-   "index_merge_union, index_merge_sort_union, index_merge_intersection"
+   "index_merge_union, index_merge_sort_union, index_merge_intersection, "
+   "index_merge_sort_intersection"
 #ifndef DBUG_OFF
    ", table_elimination"
 #endif 

sql/opt_range.h

@@ -274,12 +274,13 @@ class QUICK_SELECT_I
 
   enum {
     QS_TYPE_RANGE = 0,
-    QS_TYPE_INDEX_MERGE = 1,
-    QS_TYPE_RANGE_DESC = 2,
-    QS_TYPE_FULLTEXT   = 3,
-    QS_TYPE_ROR_INTERSECT = 4,
-    QS_TYPE_ROR_UNION = 5,
-    QS_TYPE_GROUP_MIN_MAX = 6
+    QS_TYPE_INDEX_INTERSECT = 1,
+    QS_TYPE_INDEX_MERGE = 2,
+    QS_TYPE_RANGE_DESC = 3,
+    QS_TYPE_FULLTEXT   = 4,
+    QS_TYPE_ROR_INTERSECT = 5,
+    QS_TYPE_ROR_UNION = 6,
+    QS_TYPE_GROUP_MIN_MAX = 7
   };
 
   /* Get type of this quick select - one of the QS_TYPE_* values */
@@ -306,6 +307,10 @@ class QUICK_SELECT_I
   */
   virtual void save_last_pos(){};
   
+  void add_key_and_length(String *key_names,
+                          String *used_lengths,
+                          bool *first);
+
   /*
     Append comma-separated list of keys this quick select uses to key_names;
     append comma-separated list of corresponding used lengths to used_lengths.
@@ -314,13 +319,16 @@ class QUICK_SELECT_I
   virtual void add_keys_and_lengths(String *key_names,
                                     String *used_lengths)=0;
 
+  void add_key_name(String *str, bool *first);
+
   /*
     Append text representation of quick select structure (what and how is
     merged) to str. The result is added to "Extra" field in EXPLAIN output.
     This function is implemented only by quick selects that merge other quick
     selects output and/or can produce output suitable for merging.
   */
-  virtual void add_info_string(String *str) {};
+  virtual void add_info_string(String *str) {}
+
   /*
     Return 1 if any index used by this quick select
     uses field which is marked in passed bitmap.
@@ -393,8 +401,19 @@ class QUICK_RANGE_SELECT : public QUICK_SELECT_I
   friend QUICK_RANGE_SELECT *get_quick_select(PARAM*,uint idx,
                                               SEL_ARG *key_tree,
                                               MEM_ROOT *alloc);
+  friend 
+  int read_keys_and_merge_scans(THD *thd, TABLE *head,
+                                List<QUICK_RANGE_SELECT> quick_selects,
+                                QUICK_RANGE_SELECT *pk_quick_select,
+                                READ_RECORD *read_record,
+                                bool intersection,
+                                key_map *filtered_scans,
+                                Unique **unique_ptr);
+
   friend class QUICK_SELECT_DESC;
+  friend class QUICK_INDEX_SORT_SELECT;
   friend class QUICK_INDEX_MERGE_SELECT;
+  friend class QUICK_INDEX_INTERSECT_SELECT;
   friend class QUICK_ROR_INTERSECT_SELECT;
   friend class QUICK_GROUP_MIN_MAX_SELECT;
 
@@ -448,37 +467,43 @@ class QUICK_RANGE_SELECT_GEOM: public QUICK_RANGE_SELECT
 
 
 /*
-  QUICK_INDEX_MERGE_SELECT - index_merge access method quick select.
+  QUICK_INDEX_SORT_SELECT is the base class for the common functionality of:
+  - QUICK_INDEX_MERGE_SELECT, access based on multi-index merge/union 
+  - QUICK_INDEX_INTERSECT_SELECT, access based on  multi-index intersection 
+    
 
-    QUICK_INDEX_MERGE_SELECT uses
+    QUICK_INDEX_SORT_SELECT uses
      * QUICK_RANGE_SELECTs to get rows
-     * Unique class to remove duplicate rows
+     * Unique class
+       - to remove duplicate rows for QUICK_INDEX_MERGE_SELECT
+       - to intersect rows for QUICK_INDEX_INTERSECT_SELECT
 
   INDEX MERGE OPTIMIZER
-    Current implementation doesn't detect all cases where index_merge could
+    Current implementation doesn't detect all cases where index merge could
     be used, in particular:
 
      * index_merge+'using index' is not supported
 
      * If WHERE part contains complex nested AND and OR conditions, some ways
-       to retrieve rows using index_merge will not be considered. The choice
+       to retrieve rows using index merge will not be considered. The choice
        of read plan may depend on the order of conjuncts/disjuncts in WHERE
        part of the query, see comments near imerge_list_or_list and
        SEL_IMERGE::or_sel_tree_with_checks functions for details.
 
-     * There is no "index_merge_ref" method (but index_merge on non-first
+     * There is no "index_merge_ref" method (but index merge on non-first
        table in join is possible with 'range checked for each record').
 
-    See comments around SEL_IMERGE class and test_quick_select for more
-    details.
 
   ROW RETRIEVAL ALGORITHM
 
-    index_merge uses Unique class for duplicates removal.  index_merge takes
-    advantage of Clustered Primary Key (CPK) if the table has one.
-    The index_merge algorithm consists of two phases:
+    index merge/intersection uses Unique class for duplicates removal. 
+    index merge/intersection takes advantage of Clustered Primary Key (CPK)
+    if the table has one.
+    The index merge/intersection algorithm consists of two phases:
+
+    Phase 1 
+    (implemented by a QUICK_INDEX_MERGE_SELECT::read_keys_and_merge call):
 
-    Phase 1 (implemented in QUICK_INDEX_MERGE_SELECT::prepare_unique):
     prepare()
     {
       activate 'index only';
@@ -492,32 +517,31 @@ class QUICK_RANGE_SELECT_GEOM: public QUICK_RANGE_SELECT
       deactivate 'index only';
     }
 
-    Phase 2 (implemented as sequence of QUICK_INDEX_MERGE_SELECT::get_next
-    calls):
+    Phase 2 
+    (implemented as sequence of QUICK_INDEX_MERGE_SELECT::get_next calls):
 
     fetch()
     {
-      retrieve all rows from row pointers stored in Unique;
+      retrieve all rows from row pointers stored in Unique
+      (merging/intersecting them);
       free Unique;
+      if (! intersection) 
         retrieve all rows for CPK scan;
     }
 */
 
-class QUICK_INDEX_MERGE_SELECT : public QUICK_SELECT_I
+class QUICK_INDEX_SORT_SELECT : public QUICK_SELECT_I
 {
+protected:
   Unique *unique;
 public:
-  QUICK_INDEX_MERGE_SELECT(THD *thd, TABLE *table);
-  ~QUICK_INDEX_MERGE_SELECT();
+  QUICK_INDEX_SORT_SELECT(THD *thd, TABLE *table);
+  ~QUICK_INDEX_SORT_SELECT();
 
   int  init();
   int  reset(void);
-  int  get_next();
   bool reverse_sorted() { return false; }
   bool unique_key_range() { return false; }
-  int get_type() { return QS_TYPE_INDEX_MERGE; }
-  void add_keys_and_lengths(String *key_names, String *used_lengths);
-  void add_info_string(String *str);
   bool is_keys_used(const MY_BITMAP *fields);
 #ifndef DBUG_OFF
   void dbug_dump(int indent, bool verbose);
@@ -525,24 +549,57 @@ class QUICK_INDEX_MERGE_SELECT : public QUICK_SELECT_I
 
   bool push_quick_back(QUICK_RANGE_SELECT *quick_sel_range);
 
-  /* range quick selects this index_merge read consists of */
+  /* range quick selects this index merge/intersect consists of */
   List<QUICK_RANGE_SELECT> quick_selects;
 
   /* quick select that uses clustered primary key (NULL if none) */
   QUICK_RANGE_SELECT* pk_quick_select;
 
-  /* true if this select is currently doing a clustered PK scan */
-  bool  doing_pk_scan;
-
   MEM_ROOT alloc;
   THD *thd;
-  int read_keys_and_merge();
+  virtual int read_keys_and_merge()= 0;
 
   /* used to get rows collected in Unique */
   READ_RECORD read_record;
 };
 
 
+
+class QUICK_INDEX_MERGE_SELECT : public QUICK_INDEX_SORT_SELECT
+{
+private:
+  /* true if this select is currently doing a clustered PK scan */
+  bool  doing_pk_scan;
+protected:
+  int read_keys_and_merge();
+
+public:
+  QUICK_INDEX_MERGE_SELECT(THD *thd, TABLE *table)
+    :QUICK_INDEX_SORT_SELECT(thd, table) {}
+
+  int get_next();
+  int get_type() { return QS_TYPE_INDEX_MERGE; }
+  void add_keys_and_lengths(String *key_names, String *used_lengths);
+  void add_info_string(String *str);
+};
+
+class QUICK_INDEX_INTERSECT_SELECT : public QUICK_INDEX_SORT_SELECT
+{
+protected:
+  int read_keys_and_merge();
+
+public:
+  QUICK_INDEX_INTERSECT_SELECT(THD *thd, TABLE *table)
+    :QUICK_INDEX_SORT_SELECT(thd, table) {}
+
+  key_map filtered_scans;
+  int get_next();
+  int get_type() { return QS_TYPE_INDEX_INTERSECT; }
+  void add_keys_and_lengths(String *key_names, String *used_lengths);
+  void add_info_string(String *str);
+};
+
+
 /*
   Rowid-Ordered Retrieval (ROR) index intersection quick select.
   This quick select produces intersection of row sequences returned

sql/table.h

@@ -683,7 +683,7 @@ struct st_table {
     needed by the query without reading the row.
   */
   key_map covering_keys;
-  key_map quick_keys, merge_keys;
+  key_map quick_keys, merge_keys,intersect_keys;
   /*
     A set of keys that can be used in the query that references this
     table.