Documentation Source Text

    foreach {s_title s_tag s_kw} $section {}
    if {$s_kw == ""} {
      set s_kw $s_title
    }
    lappend lx [list $s_kw $s_title 0]
  }
  hd_list_of_links {} 340 $lx






</tcl>
</ul></td></tr></table>

<p>^The routines [sqlite3_prepare_v2()], [sqlite3_prepare()],
[sqlite3_prepare16()], [sqlite3_prepare16_v2()], 
[sqlite3_exec()], and [sqlite3_get_table()] accept
an SQL statement list (sql-stmt-list) which is a semicolon-separated
................................................................................

<tcl>BubbleDiagram sql-stmt-list</tcl>

<p>Each SQL statement in the statement list is an instance of the
following:</p>

<tcl>BubbleDiagram sql-stmt</tcl>

<tcl>
proc Operator {name} {
  return "<font color=\"#2c2cf0\"><big>$name</big></font>"
}
proc Example {text} {
  hd_puts "<blockquote><pre>$text</pre></blockquote>"
}

proc Section {name label keywords} {
  global DOC
  hd_close_main
  hd_open_main lang_$label.html
  hd_header "SQLite Query Language: $name" $DOC/pages/lang.in
  eval hd_keywords $keywords
  if {[lsearch $keywords $name] == -1 && [lsearch $keywords *$name] == -1} {
    eval hd_keywords { $name }
  }
  hd_puts {<div class=nosearch>}
  hd_puts {<h1 align="center">SQL As Understood By SQLite</h1>}
  hd_puts {<p><a href="lang.html">[Top]</a></p>}
  hd_puts "<h2>$name</h2>"
  hd_puts {</div>}
}
##############################################################################
Section {CREATE TRIGGER} createtrigger {{CREATE TRIGGER}}

RecursiveBubbleDiagram create-trigger-stmt
</tcl>

<p>^The CREATE TRIGGER statement is used to add triggers to the 
database schema. ^Triggers are database operations 
that are automatically performed when a specified database event
occurs.  </p>

<p>^Each trigger must specify that it will fire for one of
the following operations: [DELETE], [INSERT], [UPDATE].
The trigger fires once for each row that is deleted, inserted,
or updated.  If the "UPDATE OF <yyterm>column-name</yyterm>"
syntax is used, then the trigger will only fire if
<yyterm>column-name</yyterm> appears on the left-hand side of
one of the terms in the SET clause of the [UPDATE] statement.</p>

<p>Due to an historical oversight, columns named in the "UPDATE OF"
clause do not actually have to exist in the table being updated.
Unrecognized column names are silently ignored.
It would be more helpful if SQLite would fail the CREATE TRIGGER
statement if any of the names in the "UPDATE OF" clause are not
columns in the table.  However, as this problem was discovered
many years after SQLite was widely deployed, we have resisted
fixing the problem for fear of breaking legacy applications.</p>

<p>^At this time SQLite supports only FOR EACH ROW triggers, not FOR EACH
STATEMENT triggers. ^Hence explicitly specifying FOR EACH ROW is optional.
^FOR EACH ROW implies that the SQL statements specified in the trigger
may be executed (depending on the WHEN clause) for each database row being
inserted, updated or deleted by the statement causing the trigger to fire.</p>

<p>^(Both the WHEN clause and the trigger actions may access elements of 
the row being inserted, deleted or updated using references of the form 
"NEW.<i>column-name</i>" and "OLD.<i>column-name</i>", where
<i>column-name</i> is the name of a column from the table that the trigger
is associated with.)^ ^(OLD and NEW references may only be used in triggers on
events for which they are relevant, as follows:</p>

<table border=0 cellpadding=10>
<tr>
<td valign="top" align="right" width=120><i>INSERT</i></td>
<td valign="top">NEW references are valid</td>
</tr>
<tr>
<td valign="top" align="right" width=120><i>UPDATE</i></td>
<td valign="top">NEW and OLD references are valid</td>
</tr>
<tr>
<td valign="top" align="right" width=120><i>DELETE</i></td>
<td valign="top">OLD references are valid</td>
</tr>
</table>
</p>)^

<p>^If a WHEN clause is supplied, the SQL statements specified
are only executed if the WHEN clause is true.
^If no WHEN clause is supplied, the SQL statements
are executed every time the trigger fires.</p>

<p>^The BEFORE or AFTER keyword determines when the trigger actions
will be executed relative to the insertion, modification or removal of the
associated row.</p>

<p>^An [ON CONFLICT] clause may be specified as part of an [UPDATE] or [INSERT]
action within the body of the trigger.
^However if an [ON CONFLICT] clause is specified as part of 
the statement causing the trigger to fire, then conflict handling
policy of the outer statement is used instead.</p>

<p>^Triggers are automatically [DROP TRIGGER | dropped]
when the table that they are 
associated with (the <i>table-name</i> table) is 
[DROP TABLE | dropped].  ^However if the trigger actions reference
other tables, the trigger is not dropped or modified if those other
tables are [DROP TABLE | dropped] or [ALTER TABLE | modified].</p>

<p>^Triggers are removed using the [DROP TRIGGER] statement.</p>

<h3>Syntax Restrictions On UPDATE, DELETE, and INSERT Statements Within
    Triggers</h3>

<p>^The [UPDATE], [DELETE], and [INSERT]
statements within triggers do not support
the full syntax for [UPDATE], [DELETE], and [INSERT] statements.  The following
restrictions apply:</p>

<ul>
<li><p>
  ^(The name of the table to be modified in an [UPDATE], [DELETE], or [INSERT]
  statement must be an unqualified table name.  In other words, one must
  use just "<i>tablename</i>" not "<i>database</i><b>.</b><i>tablename</i>"
  when specifying the table.)^ </p></li>

<li><p>
 ^For non-TEMP triggers,
  the table to be modified or queried must exist in the
  same database as the table or view to which the trigger is attached.
  ^(TEMP triggers are not subject to the same-database rule.  A TEMP
  trigger is allowed to query or modify any table in any [ATTACH]-ed database.)^
  </p></li>

<li><p>
  ^The "INSERT INTO <i>table</i> DEFAULT VALUES" form of the [INSERT] statement
  is not supported.
  </p></li>

<li><p>
  ^The INDEXED BY and NOT INDEXED clauses are not supported for [UPDATE] and
  [DELETE] statements.
  </p></li>

<li><p>
  ^(The ORDER BY and LIMIT clauses on [UPDATE] and [DELETE] statements are not
  supported.  ORDER BY and LIMIT are not normally supported for [UPDATE] or
  [DELETE] in any context but can be enabled for top-level statements
  using the [SQLITE_ENABLE_UPDATE_DELETE_LIMIT] compile-time option.  However,
  that compile-time option only applies to top-level [UPDATE] and [DELETE]
  statements, not [UPDATE] and [DELETE] statements within triggers.)^
  </p></li>

<li><p>
  ^[common-table-expression|Common table expression] are not supported for
  statements inside of triggers.
  </p></li>
</ul>

<tcl>hd_fragment instead_of_trigger {INSTEAD OF} {INSTEAD OF trigger}</tcl>
<h3>INSTEAD OF triggers</h3>

<p>^Triggers may be created on [views], as well as ordinary tables, by
specifying INSTEAD OF in the CREATE TRIGGER statement. 
^If one or more ON INSERT, ON DELETE
or ON UPDATE triggers are defined on a view, then it is not an
error to execute an INSERT, DELETE or UPDATE statement on the view, 
respectively.  ^Instead,
executing an INSERT, DELETE or UPDATE on the view causes the associated
triggers to fire. ^The real tables underlying the view are not modified
(except possibly explicitly, by a trigger program).</p>

<p>^Note that the [sqlite3_changes()] and [sqlite3_total_changes()] interfaces
do not count INSTEAD OF trigger firings, but the
[count_changes pragma] does count INSTEAD OF trigger firing.</p>

<h3>Some Example Triggers</h3>

<p>^(Assuming that customer records are stored in the "customers" table, and
that order records are stored in the "orders" table, the following
UPDATE trigger
ensures that all associated orders are redirected when a customer changes
his or her address:</p>

<tcl>Example {
CREATE TRIGGER update_customer_address UPDATE OF address ON customers 
  BEGIN
    UPDATE orders SET address = new.address WHERE customer_name = old.name;
  END;
}</tcl>

<p>With this trigger installed, executing the statement:</p>

<tcl>Example {
UPDATE customers SET address = '1 Main St.' WHERE name = 'Jack Jones';
}</tcl>

<p>causes the following to be automatically executed:</p>

<tcl>Example {
UPDATE orders SET address = '1 Main St.' WHERE customer_name = 'Jack Jones';
}</tcl>)^

<p>For an example of an INSTEAD OF trigger, consider the following schema:

<tcl>Example {
CREATE TABLE customer(
  cust_id INTEGER PRIMARY KEY,
  cust_name TEXT,
  cust_addr TEXT
);
CREATE VIEW customer_address AS
   SELECT cust_id, cust_addr FROM customer;
CREATE TRIGGER cust_addr_chng
INSTEAD OF UPDATE OF cust_addr ON customer_address
BEGIN
  UPDATE customer SET cust_addr=NEW.cust_addr
   WHERE cust_id=NEW.cust_id;
END;
}</tcl>

<p>With the schema above, a statement of the form:</p>

<tcl>Example {
UPDATE customer_address SET cust_addr=$new_address WHERE cust_id=$cust_id;
}</tcl>

<p>Causes the customer.cust_addr field to be updated for a specific
customer entry that has customer.cust_id equal to the $cust_id parameter.
Note how the values assigned to the view are made available as field
in the special "NEW" table within the trigger body.</p>

<tcl>hd_fragment undef_before {undefined BEFORE trigger behavior}</tcl>
<h3>Cautions On The Use Of BEFORE triggers</h3>

<p>If a BEFORE UPDATE or BEFORE DELETE trigger modifies or deletes a row
that was to have been updated or deleted, then the result of the subsequent
update or delete operation is undefined.  Furthermore, if a BEFORE trigger
modifies or deletes a row, then it is undefined whether or not AFTER triggers
that would have otherwise run on those rows will in fact run.
</p>

<p>The value of NEW.rowid is undefined in a BEFORE INSERT trigger in which
the rowid is not explicitly set to an integer.</p>

<p>Because of the behaviors described above, programmers are encouraged to
prefer AFTER triggers over BEFORE triggers.</p>

<tcl>hd_fragment raise {RAISE function}</tcl>
<h3>The RAISE() function</h3>

<p>^(A special SQL function RAISE() may be used within a trigger-program,)^
with the following syntax</p> 

<tcl>BubbleDiagram raise-function</tcl>

<p>^(When one of RAISE(ROLLBACK,...), RAISE(ABORT,...) or RAISE(FAIL,...)
is called during trigger-program
execution, the specified [ON CONFLICT] processing is performed
the current query terminates.)^
An error code of [SQLITE_CONSTRAINT] is returned to the application,
along with the specified error message.</p>

<p>^When RAISE(IGNORE) is called, the remainder of the current trigger program,
the statement that caused the trigger program to execute and any subsequent
trigger programs that would have been executed are abandoned. ^No database
changes are rolled back.  ^If the statement that caused the trigger program
to execute is itself part of a trigger program, then that trigger program
resumes execution at the beginning of the next step.
</p>

<tcl>hd_fragment temptrig {TEMP triggers on non-TEMP tables}</tcl>
<h3>TEMP Triggers on Non-TEMP Tables</h3>

<p>^(A trigger normally exists in the same database as the table named
after the "ON" keyword in the CREATE TRIGGER statement.  Except, it is
possible to create a TEMP TRIGGER on a table in another database.)^  
Such a trigger will only fire when changes
are made to the target table by the application that defined the trigger.
Other applications that modify the database will not be able to see the
TEMP trigger and hence cannot run the trigger.</p>

<p>When defining a TEMP trigger on a non-TEMP table, it is important to
specify the database holding the non-TEMP table.  For example,
in the following statement, it is important to say "main.tab1" instead
of just "tab1":</p>

<blockquote><pre>
CREATE TEMP TRIGGER ex1 AFTER INSERT ON <b>main.</b>tab1 BEGIN ...
</pre></blockquote>

<p>Failure to specify the schema name on the target table could result
in the TEMP trigger being reattached to a table with the same name in
another database whenever any schema change occurs.</p>

<tcl>
###############################################################################
Section {CREATE VIEW} {createview} {{CREATE VIEW} view VIEW *views}

RecursiveBubbleDiagram create-view-stmt
</tcl>

<p>^The CREATE VIEW command assigns a name to a pre-packaged 
[SELECT] statement. 
^Once the view is created, it can be used in the FROM clause
of another [SELECT] in place of a table name.
</p>

<p>^If the "TEMP" or "TEMPORARY" keyword occurs in between "CREATE"
and "VIEW" then the view that is created is only visible to the
[database connection] that created it and is automatically deleted when
the database connection is closed.</p>

<p> ^If a <yyterm>schema-name</yyterm> is specified, then the view 
is created in the specified database.
^It is an error to specify both a <yyterm>schema-name</yyterm>
and the TEMP keyword on a VIEW, unless the <yyterm>schema-name</yyterm> 
is "temp".
^If no schema name is specified, and the TEMP keyword is not present,
the VIEW is created in the main database.</p>

<p>^You cannot [DELETE], [INSERT], or [UPDATE] a view.  ^Views are read-only 
in SQLite.  ^However, in many cases you can use an
[INSTEAD OF trigger] on the view to accomplish 
the same thing.  ^Views are removed 
with the [DROP VIEW] command.</p>

<p>^If a <yyterm>column-name</yyterm> list follows 
the <yyterm>view-name</yyterm>, then that list determines
the names of the columns for the view.  ^If the <yyterm>column-name</yyterm>
list is omitted, then the names of the columns in the view are derived
from the names of the result-set columns in the [select-stmt].
The use of <yyterm>column-name</yyterm> list is recommended.  Or, if
<yyterm>column-name</yyterm> list is omitted, then the result
columns in the [SELECT] statement that defines the view should have
well-defined names using the 
"[result-column|AS column-alias]" syntax.
SQLite allows you to create views that depend on automatically 
generated column names, but you should avoid using them since the 
rules used to generate column names are not a defined part of the
interface and might change in future releases of SQLite.

<p>The <yyterm>column-name</yyterm> list syntax was added in
SQLite versions 3.9.0 ([dateof:3.9.0]).

<tcl>
##############################################################################
Section {CREATE VIRTUAL TABLE} {createvtab} {{CREATE VIRTUAL TABLE}}

RecursiveBubbleDiagram create-virtual-table-stmt
</tcl>

<p>A [virtual table] is an interface to an external storage or computation
engine that appears to be a table but does not actually store information
in the database file.</p>

<p>In general, you can do anything with a [virtual table] that can be done
with an ordinary table, except that ^you cannot create indices or triggers on a
virtual table.  ^Some virtual table implementations might impose additional
restrictions.  For example, many virtual tables are read-only.</p>

<p>The <yyterm>module-name</yyterm> is the name of an object that implements
the virtual table.  ^The <yyterm>module-name</yyterm> must be registered with
the SQLite database connection using
[sqlite3_create_module()] or [sqlite3_create_module_v2()]
prior to issuing the CREATE VIRTUAL TABLE statement.
^The module takes zero or more comma-separated arguments.
^The arguments can be just about any text as long as it has balanced
parentheses.  The argument syntax is sufficiently general that the
arguments can be made to appear as [column definitions] in a traditional
[CREATE TABLE] statement.  
^SQLite passes the module arguments directly
to the [xCreate] and [xConnect] methods of the module implementation
without any interpretation.  It is the responsibility
of the module implementation to parse and interpret its own arguments.</p>

<p>^A virtual table is destroyed using the ordinary
[DROP TABLE] statement.  There is no
DROP VIRTUAL TABLE statement.</p>

<tcl>
##############################################################################
Section DELETE delete {DELETE *DELETEs}

RecursiveBubbleDiagram delete-stmt
</tcl>

<p>The DELETE command removes records from the table identified by the
   [qualified-table-name].

<p>^If the WHERE clause is not present, all records in the table are deleted.
   ^If a WHERE clause is supplied, then only those rows for which the
   WHERE clause [boolean expression] is true are deleted.
   ^Rows for which the expression is false or NULL are retained.

<h3>Restrictions on DELETE Statements Within CREATE TRIGGER</h3>

<p>The following restrictions apply to DELETE statements that occur within the
   body of a [CREATE TRIGGER] statement:

<ul>
  <li><p>^The <yyterm>table-name</yyterm> specified as part of a 
    DELETE statement within
    a trigger body must be unqualified.  ^(In other words, the
    <i>schema-name</i><b>.</b> prefix on the table name is not allowed 
    within triggers.)^ ^If the table to which the trigger is attached is
    not in the temp database, then DELETE statements within the trigger
    body must operate on tables within the same database as it. ^If the table
    to which the trigger is attached is in the TEMP database, then the
    unqualified name of the table being deleted is resolved in the same way as
    it is for a top-level statement (by searching first the TEMP database, then
    the main database, then any other databases in the order they were
    attached).
    
  <li><p>^The INDEXED BY and NOT INDEXED clauses are not allowed on DELETE
    statements within triggers.</p>

  <li><p>^The LIMIT and ORDER BY clauses (described below) are unsupported for
    DELETE statements within triggers.</p>
</ul>

<h3>Optional LIMIT and ORDER BY clauses</h3>

<p>^(If SQLite is compiled with the [SQLITE_ENABLE_UPDATE_DELETE_LIMIT]
compile-time option, then the syntax of the DELETE statement is
extended by the addition of optional ORDER BY and LIMIT clauses:)^</p>

<tcl>BubbleDiagram delete-stmt-limited</tcl>

<p>^If a DELETE statement has a LIMIT clause, the maximum number of rows that
will be deleted is found by evaluating the accompanying expression and casting
it to an integer value. ^If the result of the evaluating the LIMIT clause
cannot be losslessly converted to an integer value, it is an error. ^A 
negative LIMIT value is interpreted as "no limit". ^(If the DELETE statement 
also has an OFFSET clause, then it is similarly evaluated and cast to an
integer value. Again, it is an error if the value cannot be losslessly
converted to an integer.)^ ^If there is no OFFSET clause, or the calculated
integer value is negative, the effective OFFSET value is zero.

<p>^(If the DELETE statement has an ORDER BY clause, then all rows that would 
be deleted in the absence of the LIMIT clause are sorted according to the 
ORDER BY. The first <i>M</i> rows, where <i>M</i> is the value found by
evaluating the OFFSET clause expression, are skipped, and the following 
<i>N</i>, where <i>N</i> is the value of the LIMIT expression, are deleted.)^
^If there are less than <i>N</i> rows remaining after taking the OFFSET clause
into account, or if the LIMIT clause evaluated to a negative value, then all
remaining rows are deleted.

<p>^If the DELETE statement has no ORDER BY clause, then all rows that
would be deleted in the absence of the LIMIT clause are assembled in an
arbitrary order before applying the LIMIT and OFFSET clauses to determine 
the subset that are actually deleted.

<p>^(The ORDER BY clause on a DELETE statement is used only to determine which
rows fall within the LIMIT. The order in which rows are deleted is arbitrary
and is not influenced by the ORDER BY clause.)^

<tcl>hd_fragment truncateopt {truncate optimization}</tcl>
<h3>The Truncate Optimization</h3>

<p>^When the WHERE is omitted from a DELETE statement and the table
being deleted has no triggers,
SQLite uses an optimization to erase the entire table content
without having to visit each row of the table individually.
This "truncate" optimization makes the delete run much faster.
Prior to SQLite [version 3.6.5] ([dateof:3.6.5]), the truncate optimization
also meant that the [sqlite3_changes()] and
[sqlite3_total_changes()] interfaces
and the [count_changes pragma]
will not actually return the number of deleted rows.  
That problem has been fixed as of [version 3.6.5] ([dateof:3.6.5]).

<p>^The truncate optimization can be permanently disabled for all queries
by recompiling
SQLite with the [SQLITE_OMIT_TRUNCATE_OPTIMIZATION] compile-time switch.</p>

<p>The truncate optimization can also be disabled at runtime using
the [sqlite3_set_authorizer()] interface.  ^If an authorizer callback
returns [SQLITE_IGNORE] for an [SQLITE_DELETE] action code, then
the DELETE operation will proceed but the truncate optimization will
be bypassed and rows will be deleted one by one.</p>

<tcl>
###############################################################################
Section {DETACH DATABASE} detach *DETACH

RecursiveBubbleDiagram detach-stmt
</tcl>

<p>^This statement detaches an additional database connection previously 
attached using the [ATTACH] statement.  
^When not in [shared cache mode], 
it is possible to have the same database file attached multiple times using 
different names, and detaching one connection to a file will leave the 
others intact.</p>
^In [shared cache mode], attempting to attach the same database file more
than once results in an error.


<tcl>
##############################################################################
Section {DROP INDEX} dropindex {{DROP INDEX}}

RecursiveBubbleDiagram drop-index-stmt
</tcl>

<p>^The DROP INDEX statement removes an index added
with the [CREATE INDEX] statement.  The index is completely removed from
the disk.  The only way to recover the index is to reenter the
appropriate [CREATE INDEX] command.</p>

<tcl>
##############################################################################
Section {DROP TABLE} droptable {{DROP TABLE}}

RecursiveBubbleDiagram drop-table-stmt
</tcl>

<p>^The DROP TABLE statement removes a table added with the
[CREATE TABLE] statement.  The name specified is the
table name.  ^The dropped table is completely removed from the database 
schema and the disk file.  The table can not be recovered.  
^All indices and triggers
associated with the table are also deleted.</p>

<p>^The optional IF EXISTS clause suppresses the error that would normally
result if the table does not exist.</p>

<p>^If [foreign key constraints] are enabled, a DROP TABLE command performs an
implicit [DELETE | DELETE FROM] command before removing the
table from the database schema. ^Any triggers attached to the table are
dropped from the database schema before the implicit DELETE FROM
is executed, so this cannot cause any triggers to fire. By contrast, ^an
implicit DELETE FROM does cause any configured
[foreign key actions] to take place. 
^If the implicit DELETE FROM executed
as part of a DROP TABLE command violates any immediate foreign key constraints,
an error is returned and the table is not dropped. ^If 
the implicit DELETE FROM causes any 
deferred foreign key constraints to be violated, and the violations still
exist when the transaction is committed, an error is returned at the time
of commit.

<tcl>
##############################################################################
Section {DROP TRIGGER} droptrigger {{DROP TRIGGER}}

RecursiveBubbleDiagram drop-trigger-stmt
</tcl>

<p>^The DROP TRIGGER statement removes a trigger created by the 
[CREATE TRIGGER] statement. ^Once removed, the trigger definition is no
longer present in the sqlite_master (or sqlite_temp_master) table and is
not fired by any subsequent INSERT, UPDATE or DELETE statements.

<p>^Note that triggers are automatically dropped when the associated table is
dropped.

<tcl>
##############################################################################
Section {DROP VIEW} dropview {{DROP VIEW}}

RecursiveBubbleDiagram drop-view-stmt
</tcl>

<p>^The DROP VIEW statement removes a view created by the [CREATE VIEW] 
  statement. ^The view definition is removed from the database schema, but 
  no actual data in the underlying base tables is modified.

<p>^(The view to drop is identified by the view-name and optional 
  schema-name specified as part of the DROP VIEW statement. This 
  reference is resolved using the standard procedure for [object resolution].)^

<p>
  ^If the specified view cannot be found and the IF EXISTS clause is not 
  present, it is an error. ^If the specified view cannot be found and an IF
  EXISTS clause is present in the DROP VIEW statement, then the statement
  is a no-op.


<tcl>
##############################################################################
Section {Database Object Name Resolution} {naming} {{object resolution}}
</tcl>

<p>
  In SQLite, a database object (a table, index, trigger or view) is identified
  by the name of the object and the name of the database that it resides in. 
  Database objects may reside in the main database, the temp database, or in
  an [ATTACH|attached database].

<p>
  The syntax of the [DROP TABLE], [DROP INDEX], [DROP VIEW], [DROP TRIGGER],
  [REINDEX], [ALTER TABLE] and many other commands all permit the user to
  specify a database object either by its name alone, or by a combination of
  its name and the name of its database. ^(If no database is specified as part
  of the object reference, then SQLite searches the main, temp and all attached
  databases for an object with a matching name. The temp database is searched
  first, followed by the main database, followed all attached databases in the
  order that they were attached. The reference resolves to the first match
  found.)^ For example:

<pre>^(
      /* Add a table named 't1' to the temp, main and an attached database */
      ATTACH 'file.db' AS aux;
      CREATE TABLE t1(x, y);
      CREATE TEMP TABLE t1(x, y);
      CREATE TABLE aux.t1(x, y);

      DROP TABLE t1;         /* Drop table in temp database */
      DROP TABLE t1;         /* Drop table in main database */
      DROP TABLE t1;         /* Drop table in aux database */
)^</pre>

<p>
  ^If a schema name is specified as part of an object reference, it must be
  either "main", or "temp" or the schema-name of an attached database.
  ^Like other SQL identifiers, schema names are case-insensitive.
  ^If a schema name is specified, then only that one schema is searched for
  the named object.

<p>
  Most object references may only resolve to a specific type of object (for
  example a reference that is part of a DROP TABLE statement may only resolve
  to a table object, not an index, trigger or view). However in some contexts 
  (e.g. [REINDEX]) an object reference may be resolve to more than one type
  of object. ^When searching database schemas for a named object, objects of
  types that cannot be used in the context of the reference are always 
  ignored.

<tcl>
##############################################################################
Section EXPLAIN explain EXPLAIN

BubbleDiagram sql-stmt
</tcl>

<p>^An SQL statement can be preceded by the keyword "EXPLAIN" or
by the phrase "EXPLAIN QUERY PLAN".  ^Either modification causes the
SQL statement to behave as a query and to return information about
how the SQL statement would have operated if the EXPLAIN keyword or
phrase had been omitted.</p>

<p>The output from EXPLAIN and EXPLAIN QUERY PLAN is intended for
interactive analysis and troubleshooting only.  The details of the 
output format are subject to change from one release of SQLite to the next.
Applications should not use EXPLAIN or EXPLAIN QUERY PLAN since
their exact behavior is variable and only partially documented.</p>

<p>^When the EXPLAIN keyword appears by itself it causes the statement
to behave as a query that returns the sequence of 
[virtual machine instructions] it would have used to execute the command had
the EXPLAIN keyword not been present. ^When the EXPLAIN QUERY PLAN phrase
appears, the statement returns high-level information regarding the query
plan that would have been used.

<p>The EXPLAIN QUERY PLAN command is described in 
[explain query plan|more detail here].

<h3>EXPLAIN operates at run-time, not at prepare-time</h3>

<p>The EXPLAIN and EXPLAIN QUERY PLAN prefixes affect the behavior of
running a [prepared statement] using [sqlite3_step()].  The process of
generating a new prepared statement using [sqlite3_prepare()] or similar
is (mostly) unaffected by EXPLAIN.  (The exception to the previous sentence
is that some special opcodes used by EXPLAIN QUERY PLAN are omitted when
building an EXPLAIN QUERY PLAN prepared statement, as a performance
optimization.)

<p>This means that actions that occur during sqlite3_prepare() are
unaffected by EXPLAIN.

<ul>
<li><p>
Some [PRAGMA] statements do their work during sqlite3_prepare() rather
than during sqlite3_step().  Those PRAGMA statements are unaffected
by EXPLAIN.  They operate the same with or without the EXPLAIN prefix.
The set of PRAGMA statements that are unaffected by EXPLAIN can vary
from one release to the next.  Some PRAGMA statements operate during
sqlite3_prepare() depending on their arguments.  For consistent
results, avoid using EXPLAIN on PRAGMA statements.

<li><p>
The [authorizer callback] is invoked regardless of the presence of
EXPLAIN or EXPLAIN QUERY PLAN.
</ul>
<tcl>
##############################################################################
Section expression expr {*expression {expression syntax}}

RecursiveBubbleDiagram expr
</tcl>

<p>This section is different from the others.  Most other sections of
this document talks about a particular SQL command.  This section does
not talk about a standalone command but about "expressions" which are 
subcomponents of most other commands.</p>

<tcl>hd_fragment binaryops {binary operators}</tcl>
<h3>Operators</h3>
<p>^(SQLite understands the following binary operators, in order from
highest to lowest precedence:</p>

<blockquote><pre>
<font color="#2c2cf0"><big>||
*    /    %
+    -
&lt;&lt;   &gt;&gt;   &amp;    |
&lt;    &lt;=   &gt;    &gt;=
=    ==   !=   &lt;&gt;   </big>IS   IS NOT   IN   LIKE   GLOB   MATCH   REGEXP
AND   
OR</font>
</pre></blockquote>)^

<p>^(Supported unary prefix operators are these:</p>

<blockquote><pre>
<font color="#2c2cf0"><big>-    +    ~    NOT</big></font>
</pre></blockquote>)^

<tcl>hd_fragment collateop {COLLATE operator}</tcl>
<p>^The COLLATE operator is a unary postfix
operator that assigns a [collating sequence] to an expression.
^The COLLATE operator has a higher precedence (binds more tightly) than any
binary operator and any unary prefix operator except "~".
(COLLATE and "~" are associative so their binding order does not matter.)
^The collating sequence set by the COLLATE operator overrides the
collating sequence determined by the COLLATE clause in a table
[column definition].
See the [collating sequence| detailed discussion on collating sequences]
in the [datatype | Datatype In SQLite3] document for additional information.
</p>

<tcl>hd_puts "
<p>^The unary operator [Operator +] is a no-op.  ^It can be applied
to strings, numbers, blobs or NULL and it always returns a result
with the same value as the operand.</p>"</tcl>

<p>Note that there are two variations of the equals and not equals
operators.  ^Equals can be either

<tcl>
hd_puts "[Operator =] or [Operator ==].
^The non-equals operator can be either
[Operator !=] or [Operator {&lt;&gt;}].
^The [Operator ||] operator is \"concatenate\" - it joins together
the two strings of its operands.
^The operator [Operator %] outputs the integer value of its left 
operand modulo its right operand.  In other words, the
operator [Operator %] works in SQLite the same as it does
in ANSI-C.</p>

<p>^The result of any binary operator is either a numeric value or 
NULL, except for the [Operator ||] concatenation operator which always 
evaluates to either NULL or a text value.</p>"

hd_fragment {isisnot} {IS operator} {IS NOT operator}

hd_puts "<p>^The [Operator IS] and [Operator {IS NOT}] operators work
like [Operator =] and [Operator !=] except when one or both of the
operands are NULL. ^In this case, if both operands are NULL, then the
IS operator evaluates to 1 (true) and the IS NOT operator evaluates
to 0 (false). ^If one operand is NULL and the other is not, then the
IS operator evaluates to 0 (false) and the IS NOT operator is 1 (true).
^It is not possible for an IS or IS NOT expression to evaluate to NULL.
^Operators [Operator IS] and [Operator {IS NOT}] have the same 
precedence as [Operator =]."

<tcl>hd_fragment litvalue {literal value}</tcl>
<h3>Literal Values (Constants)</h3>
<p>
A literal value represents a constant.
^Literal values may be integers, floating point numbers, strings,
BLOBs, or NULLs.
<p>
The syntax for integer and floating point literals (collectively
"numeric literals") is shown by the following diagram:</p>

<tcl>BubbleDiagram numeric-literal</tcl>

<p>
^(If a numeric literal has a decimal point or an exponentiation
clause or if its magnitude is less than -9223372036854775808 or
greater than 9223372036854775807, then it is a floating point literal.
Otherwise is it is an  integer literal.)^
^The "E" character that begins the exponentiation
clause of a floating point literal can be either upper or lower case.
^(The "." character is always used 
as the decimal point even if the locale setting specifies "," for
this role - the use of "," for the decimal point would result in
syntactic ambiguity.)^

<tcl>hd_fragment hexint {hexadecimal integer literals} {hexadecimal integers}</tcl>
<p>^Hexadecimal integer literals follow the C-language notation of
"0x" or "0X" followed by hexadecimal digits.
^For example, 0x1234 means the same as 4660
and 0x8000000000000000 means the same as -9223372036854775808.
 ^(Hexadecimal integer literals are interpreted as 64-bit
two's-complement integers and are thus limited
to sixteen significant digits of precision.)^
Support for hexadecimal integers was added to SQLite
version 3.8.6 ([dateof:3.8.6]).
^For backwards compatibility, the "0x" hexadecimal integer
notation is only understood by the SQL language parser, not by the
type conversions routines.
^(String variables that
contain text formatted like hexadecimal integers are not
interpreted as hexadecimal integers when coercing the string value
into an integer due to a [CAST expression] or for a [column affinity]
transformation or prior to performing a numeric operation or for
any other run-time conversions.)^  ^(When coercing a
string value in the format of a hexadecimal integer into an integer
value, the conversion process stops when the 'x' character is seen
so the resulting integer value is always zero.)^
^(SQLite only understands the hexadecimal integer notation when it
appears in the SQL statement text, not when it appears as
part of the content of the database.)^

<p> ^A string constant is formed by enclosing the
string in single quotes (').  ^A single quote within the string can
be encoded by putting two single quotes in a row - as in Pascal.
C-style escapes using the backslash character are not supported because
they are not standard SQL.

<p> ^BLOB literals are string literals containing hexadecimal data and
preceded by a single "x" or "X" character.  ^(Example: X'53514C697465')^

<p>
^A literal value can also be the token "NULL".
</p>

<tcl>hd_fragment varparam parameter parameters {bound parameter} {bound parameters}</tcl>
<h3>Parameters</h3>
<p>
A "variable" or "parameter" token
specifies a placeholder in the expression for a 
value that is filled in at runtime using the
[sqlite3_bind_blob() | sqlite3_bind()] family of C/C++ interfaces.
Parameters can take several forms:
</p>

<blockquote>
<table>
<tr>
<td align="right" valign="top"><b>?</b><i>NNN</i></td><td width="20"></td>
<td>^(A question mark followed by a number <i>NNN</i> holds a spot for the
NNN-th parameter.  NNN must be between 1 and [SQLITE_MAX_VARIABLE_NUMBER].)^
</td>
</tr>
<tr>
<td align="right" valign="top"><b>?</b></td><td width="20"></td>
<td>^A question mark that is not followed by a number creates a parameter
with a number one greater than the largest parameter number already assigned.
^If this means the parameter number is greater than
[SQLITE_MAX_VARIABLE_NUMBER], it is an error.
This parameter format is provided for compatibility with other database
engines.  But because it is easy to miscount the question marks, the
use of this parameter format is discouraged.  Programmers are encouraged
to use one of the symbolic formats below or the ?NNN format above instead.
</td>
</tr>
<tr>
<td align="right" valign="top"><b>:</b><i>AAAA</i></td><td width="20"></td>
<td>^A colon followed by an identifier name holds a spot for a 
[sqlite3_bind_parameter_name|named parameter] with the name :AAAA.  
^(Named parameters are also numbered. The number assigned is one greater than
the largest parameter number already assigned.)^ ^If this means the parameter
would be assigned a number greater than [SQLITE_MAX_VARIABLE_NUMBER], it is
an error. To avoid confusion, it is best to avoid mixing named and numbered
parameters.</td>
</tr>
<tr>
<td align="right" valign="top"><b>@</b><i>AAAA</i></td><td width="20"></td>
<td>^An "at" sign works exactly like a colon, except that the name of
the parameter created is @AAAA.</td>
</tr>
<tr>
<td align="right" valign="top"><b>$</b><i>AAAA</i></td><td width="20"></td>
<td>^A dollar-sign followed by an identifier name also holds a spot for a named
parameter with the name $AAAA.  ^(The identifier name in this case can include
one or more occurrences of "::" and a suffix enclosed in "(...)" containing
any text at all.)^  This syntax is the form of a variable name in the
[http://www.tcl-lang.org/ | Tcl programming language].  The presence
of this syntax results from the fact that SQLite is really a 
[Tcl extension] that has escaped into the wild.</td>
</tr>
</table>
</blockquote>

<p>^Parameters that are not assigned values using
[sqlite3_bind_blob() | sqlite3_bind()] are treated
as NULL.  The [sqlite3_bind_parameter_index()] interface can be used
to translate a symbolic parameter name into its equivalent numeric index.</p>

<p>^The maximum parameter number is set at compile-time by
the [SQLITE_MAX_VARIABLE_NUMBER] macro.  ^(An individual [database connection]
D can reduce its maximum parameter number below the compile-time maximum
using the [sqlite3_limit](D, [SQLITE_LIMIT_VARIABLE_NUMBER],...) interface.)^
</p>

<tcl>hd_fragment like LIKE ESCAPE</tcl>
<h3>The LIKE, GLOB, REGEXP, and MATCH operators</h3>
<p>^The LIKE operator does a pattern matching comparison. ^The operand
to the right of the LIKE operator contains the pattern and the left hand
operand contains the string to match against the pattern.

<tcl>hd_puts "^A percent symbol (\"%\") in the LIKE pattern matches any
sequence of zero or more characters in the string.  ^An underscore
(\"_\") in the LIKE pattern matches any single character in the
string.  ^(Any other character matches itself or its lower/upper case
equivalent (i.e. case-insensitive matching).)^

<u>Important Note:</u> ^SQLite only
understands upper/lower case for ASCII characters by default.  ^The
LIKE operator is case sensitive by default for unicode characters that are
beyond the ASCII range.  For example,
^the expression <b>'a'&nbsp;LIKE&nbsp;'A'</b>
is TRUE but <b>'&aelig;'&nbsp;LIKE&nbsp;'&AElig;'</b> is FALSE."</tcl>
The ICU extension to SQLite includes an enhanced version of the
LIKE operator that does case folding across all unicode characters.</p>

<p>^If the optional ESCAPE clause is present, then the expression
following the ESCAPE keyword must evaluate to a string consisting of
a single character. ^This character may be used in the LIKE pattern
to include literal percent or underscore characters. ^The escape
character followed by a percent symbol (%), underscore (_), or a second
instance of the escape character itself matches a
literal percent symbol, underscore, or a single escape character,
respectively.

<p>^The infix LIKE operator is implemented by calling the
application-defined SQL functions [like(<i>Y</i>,<i>X</i>)] or
[like(<i>Y</i>,<i>X</i>,<i>Z</i>)]</a>.</p>

<p>^The LIKE operator can be made case sensitive using the
[case_sensitive_like pragma].</p>

<tcl>hd_fragment glob GLOB</tcl>
<p>^The GLOB operator is similar to LIKE but uses the Unix
file globbing syntax for its wildcards.  ^Also, GLOB is case
sensitive, unlike LIKE.  ^Both GLOB and LIKE may be preceded by
the NOT keyword to invert the sense of the test.  ^The infix GLOB 
operator is implemented by calling the function
[glob(<i>Y</i>,<i>X</i>)] and can be modified by overriding
that function.</p>

<tcl>hd_fragment regexp REGEXP</tcl>
<p>^The REGEXP operator is a special syntax for the regexp()
user function.  ^No regexp() user function is defined by default
and so use of the REGEXP operator will normally result in an
error message.  ^If an [application-defined SQL function] named "regexp"
is added at run-time, then the "<i>X</i> REGEXP <i>Y</i>" operator will
be implemented as a call to "regexp(<i>Y</i>,<i>X</i>)".</p>

<tcl>hd_fragment match MATCH</tcl>
<p>^The MATCH operator is a special syntax for the match()
application-defined function.  ^The default match() function implementation
raises an exception and is not really useful for anything.
^But extensions can override the match() function with more
helpful logic.</p>

<tcl>hd_fragment between BETWEEN</tcl>
<h3>The BETWEEN operator</h3>
<p>^(The BETWEEN operator is logically equivalent to a pair of comparisons.
"<i>x</i> <b>BETWEEN</b> <i>y</i> <b>AND</b> <i>z</i>" is 
equivalent to 
"<i>x</i><b>&gt;=</b><i>y</i> <b>AND</b> <i>x</i><b>&lt;=</b><i>z</i>" except
that with BETWEEN, the <i>x</i> expression is only evaluated once.)^
^The precedence of the BETWEEN operator is the same as the precedence
as operators <b>==</b> and <b>!=</b> and <b>LIKE</b> and groups left to right.

<tcl>hd_fragment case {CASE expression}</tcl>
<h3>The CASE expression</h3>
<p>A CASE expression serves a role similar to IF-THEN-ELSE in other
programming languages.  

<p>The optional expression that occurs in between the CASE keyword and the
first WHEN keyword is called the "base" expression. ^There are two basic forms
of the CASE expression: those with a base expression and those without.

<p>^In a CASE without a base expression, each WHEN expression is evaluated
and the result treated as a boolean, starting with the leftmost and continuing
to the right. ^The result of the CASE expression is the evaluation of the THEN
expression that corresponds to the first WHEN expression that evaluates to
true. ^Or, if none of the WHEN expressions evaluate to true, the result of
evaluating the ELSE expression, if any. ^If there is no ELSE expression and
none of the WHEN expressions are true, then the overall result is NULL.

<p>^A NULL result is considered untrue when evaluating WHEN terms.

<p>^In a CASE with a base expression, the base expression is evaluated just
once and the result is compared against the evaluation of each WHEN 
expression from left to right. ^The result of the CASE expression is the 
evaluation of the THEN expression that corresponds to the first WHEN
expression for which the comparison is true. ^Or, if none of the WHEN
expressions evaluate to a value equal to the base expression, the result
of evaluating the ELSE expression, if any. ^If there is no ELSE expression and
none of the WHEN expressions produce a result equal to the base expression,
the overall result is NULL.

<p>^When comparing a base expression against a WHEN expression, the same
collating sequence, affinity, and NULL-handling rules apply as if the
base expression and WHEN expression are respectively the left- and
right-hand operands of an <big><b>=</b></big> operator.</p> ^If the base 
expression is NULL then the result of the CASE is always the result 
of evaluating the ELSE expression if it exists, or NULL if it does not.

<p>^Both forms of the CASE expression use lazy, or short-circuit, 
evaluation.

<p>^(The only difference between the following two CASE expressions is that 
the <i>x</i> expression is evaluated exactly once in the first example but 
might be evaluated multiple times in the second:

<ul><pre>
<li>CASE x WHEN w1 THEN r1 WHEN w2 THEN r2 ELSE r3 END
<li>CASE WHEN x=w1 THEN r1 WHEN x=w2 THEN r2 ELSE r3 END
</pre></ul>)^


<tcl>hd_fragment in_op {IN operator} {NOT IN operator}</tcl>
<h3>The IN and NOT IN operators</h3>
<p>^The IN and NOT IN operators take an expression on the
left and a list of values or a subquery on the right.
^When the right operand of an IN or NOT IN operator is a subquery, the
subquery must have the same number of columns as there are columns in
the [row value] of the left operand.  ^The subquery on the
right of an IN or NOT IN operator must be a scalar subquery if the left
expression is not a [row value] expression.
^If the right operand of an IN or NOT IN operator is a list of values,
each of those values must be scalars and the left expression must also
be a scalar.
^The right-hand side of an IN or NOT IN operator can be a
table <i>name</i> or [table-valued function] <i>name</i> in which 
case the right-hand side is understood to be subquery of
the form "(SELECT * FROM <i>name</i>)".
^When the right operand is an empty set, the result of IN is false and the
result of NOT IN is true, regardless of the left operand and even if the
left operand is NULL.
<p>^(The result of an IN or NOT IN operator is determined by the following
matrix:

<center>
<table border=1>
<tr>
<th>Left operand <br>is NULL
<th>Right operand <br>contains NULL
<th>Right operand <br>is an empty set
<th>Left operand found <br>within right operand
<th>Result of <br>IN operator
<th>Result of <br>NOT IN operator
<tr>
<td align="center">no
<td align="center">no
<td align="center">no
<td align="center">no
<td align="center">false
<td align="center">true
<tr>
<td align="center">does not matter
<td align="center">no
<td align="center">yes
<td align="center">no
<td align="center">false
<td align="center">true
<tr>
<td align="center">no
<td align="center">does not matter
<td align="center">no
<td align="center">yes
<td align="center">true
<td align="center">false
<tr>
<td align="center">no
<td align="center">yes
<td align="center">no
<td align="center">no
<td align="center">NULL
<td align="center">NULL
<tr>
<td align="center">yes
<td align="center">does not matter
<td align="center">no
<td align="center">does not matter
<td align="center">NULL
<td align="center">NULL
</table>
</center>)^

<p>^Note that SQLite allows the parenthesized list of scalar values on
the right-hand side of an IN or NOT IN operator to be an empty list but
most other SQL database database engines and the SQL92 standard require
the list to contain at least one element.</p>

<h3>Table Column Names</h3>

<p>^A column name can be any of the names defined in the [CREATE TABLE]
statement or one of the following special identifiers: "<b>ROWID</b>",
"<b>OID</b>", or "<b>_ROWID_</b>".
^The three special identifiers describe the
unique integer key (the [rowid]) associated with every 
row of every table and so are not available on [WITHOUT ROWID] tables.
^The special identifiers only refer to the row key if the [CREATE TABLE]
statement does not define a real column with the same name.
^The rowid can be used anywhere a regular
column can be used.</p>

<tcl>hd_fragment exists_op {EXISTS operator} {NOT EXISTS operator}</tcl>
<h3>The EXISTS operator</h3>

<p>^The EXISTS operator always evaluates to one of the integer values 0 
and 1. ^If executing the SELECT statement specified as the right-hand 
operand of the EXISTS operator would return one or more rows, then the
EXISTS operator evaluates to 1. ^If executing the SELECT would return
no rows at all, then the EXISTS operator evaluates to 0. 

<p>^The number of columns in each row returned by the SELECT statement
(if any) and the specific values returned have no effect on the results
of the EXISTS operator. ^In particular, rows containing NULL values are
not handled any differently from rows without NULL values.

<tcl>hd_fragment subq {Subqueries}</tcl>
<h3>Subquery Expressions</h3>

<p>^A [SELECT] statement enclosed in parentheses is a subquery.
^All types of SELECT statement, including
aggregate and compound SELECT queries (queries with keywords like
UNION or EXCEPT) are allowed as scalar subqueries.
^The value of a subquery expression is the first row of the result
from the enclosed [SELECT] statement.
^The value of a subquery expression is NULL if the enclosed
[SELECT] statement returns no rows.

<p>A subquery that returns a single column is a scalar subquery and can
be used most anywhere.
^A subquery that returns two or more columns is a [row value]
subquery and can only be used as the operand of a comparison operator.

<tcl>hd_fragment cosub {correlated subqueries}</tcl>
<h3>Correlated Subqueries</h3>

<p>^A [SELECT] statement used as either a scalar subquery or as the 
right-hand operand of an IN, NOT IN or EXISTS expression may contain 
references to columns in the outer query. Such a subquery is known as
a correlated subquery. ^A correlated subquery is reevaluated each time
its result is required. ^An uncorrelated subquery is evaluated only once
and the result reused as necessary.

<tcl>hd_fragment castexpr {CAST expression} {CAST} {cast} {CAST operator}</tcl>
<h3>CAST expressions</h3>

<p>A CAST expression of the form "CAST(<i>expr</i> AS <i>type-name</i>)"
is used to convert the value of <i>expr</i> to 
a different [storage class] specified by <yyterm>type-name</yyterm>.
^A CAST conversion is similar to the conversion that takes
place when a [column affinity] is applied to a value except that with
the CAST operator the conversion always takes place even if the conversion
lossy and irreversible, whereas column affinity only changes the data type
of a value if the change is lossless and reversible.

<p>^If the value of <i>expr</i> is NULL, then the result of the CAST
expression is also NULL. ^Otherwise, the storage class of the result
is determined by applying the [rules for determining column affinity] to
the <yyterm>type-name</yyterm>.

<table border=1>
<tr>
  <th> Affinity of <yyterm><nobr>type-name</nobr></yyterm>
  <th> Conversion Processing
<tr>
  <td> NONE 
  <td> ^Casting a value to a <yyterm>type-name</yyterm> with no affinity 
  causes the value to
  be converted into a BLOB.  ^Casting to a BLOB consists of first casting
  the value to TEXT in the [encoding] of the database connection, then
  interpreting the resulting byte sequence as a BLOB instead of as TEXT.

<tr>
  <td> TEXT
  <td> ^To cast a BLOB value to TEXT, the sequence of bytes that make up the
  BLOB is interpreted as text encoded using the database encoding.
  <p>
   ^Casting an INTEGER or REAL value into TEXT renders the value as if via 
    [sqlite3_snprintf()] except that the resulting TEXT uses the [encoding] of
    the database connection.

<tr>
  <td> REAL
  <td> ^When casting a BLOB value to a REAL, the value is first converted to
        TEXT.
       <p>^When casting a TEXT value to REAL, the longest possible prefix of
        the value that can be interpreted as a real number is extracted from
        the TEXT value and the remainder ignored. ^Any leading spaces in the
        TEXT value are ignored when converging from TEXT to REAL. ^(If there is
        no prefix that can be interpreted as a real number, the result of the
        conversion is 0.0.)^

<tr>
  <td> INTEGER
  <td> ^When casting a BLOB value to INTEGER, the value is first converted to
        TEXT.
       <p>^When casting a TEXT value to INTEGER, the longest possible prefix of
        the value that can be interpreted as an integer number is extracted from
        the TEXT value and the remainder ignored. ^Any leading spaces in the
        TEXT value when converting from TEXT to INTEGER are ignored. ^If there
        is no prefix that can be interpreted as an integer number, the result
        of the conversion is 0.  ^If the prefix integer is greater than
        +9223372036854775807 then the result of the cast is exactly
        +9223372036854775807.  ^Similarly, if the prefix integer is
        less than -9223372036854775808 then the result of the cast is
        exactly -9223372036854775808.

        <p>^When casting to INTEGER, if the text looks like a floating point
        value with an exponent, the exponent will be ignored because it is
        no part of the integer prefix.  ^For example,
        "(CAST '123e+5' AS INTEGER)" results in 123, not in 12300000.

        <p> ^(The CAST operator understands decimal
        integers only &mdash; conversion of [hexadecimal integers] stops 
        at the "x" in the "0x" prefix of the hexadecimal integer string 
        and thus result of the CAST is always zero.)^

      <p>^A cast of a REAL value into an INTEGER results in the integer
      between the REAL value and zero that is closest to the REAL value.
      ^If a REAL is greater than the greatest possible signed
      integer (+9223372036854775807) then the result is the greatest possible
      signed integer and if the REAL is less than the least possible signed
      integer (-9223372036854775808) then the result is the least possible
      signed integer.

      <p>Prior to SQLite version 3.8.2 ([dateof:3.8.2]),
      casting a REAL value greater than
      +9223372036854775807.0 into an integer resulted in the most negative
      integer, -9223372036854775808.  This behavior was meant to emulate the
      behavior of x86/x64 hardware when doing the equivalent cast.

<tr>
  <td> NUMERIC
  <td> ^Casting a TEXT or BLOB value into NUMERIC yields either an INTEGER or
   a REAL result.
   ^If the input text looks like an integer (there is no decimal point nor
   exponent) and the value is small enough to fit in a 64-bit signed integer,
   then the result will be INTEGER.
   ^Input text that looks like floating point (there is a decimal point and/or
   an exponent) and the text describes a value that 
   can be losslessly converted back and forth between IEEE 754 64-bit float and a
   51-bit signed integer, then the result is INTEGER.
   (In the previous sentence, a 51-bit integer is specified since that is one
   bit less than the length of the mantissa of an IEEE 754 64-bit float and
   thus provides a 1-bit of margin for the text-to-float conversion operation.)
   ^Any text input that describes a value outside the range of a 64-bit
   signed integer yields a REAL result.
   <p> ^Casting a REAL or INTEGER value to NUMERIC is a no-op, even if a real
   value could be losslessly converted to an integer.

</tr>

</table>

<p>^Note that the result from casting any non-BLOB value into a 
BLOB and the result from casting any BLOB value into a non-BLOB value
may be different depending on whether the database [encoding] is UTF-8,
UTF-16be, or UTF-16le.


<tcl>hd_fragment booleanexpr {boolean expression}</tcl>
<h3>Boolean Expressions</h3>

<p>The SQL language features several contexts where an expression is 
evaluated and the result converted to a boolean (true or false) value. These
contexts are:

  <ul>
    <li> the WHERE clause of a SELECT, UPDATE or DELETE statement,
    <li> the ON or USING clause of a join in a SELECT statement,
    <li> the HAVING clause of a SELECT statement,
    <li> the WHEN clause of an SQL trigger, and
    <li> the WHEN clause or clauses of some CASE expressions.
  </ul>

<p>^(To convert the results of an SQL expression to a boolean value, SQLite
first casts the result to a NUMERIC value in the same way as a 
[CAST expression]. A numeric zero value (integer value 0 or real 
value 0.0) is considered to be false.  A NULL value is still NULL.
All other values are considered true.)^

<p>^(For example, the values NULL, 0.0, 0, 'english' and '0' are all considered
to be false.)^ ^(Values 1, 1.0, 0.1, -0.1 and '1english' are considered to 
be true.)^

<p>Beginning with SQLite 3.23.0 ([dateof:3.23.0]), SQLite recognizes the
identifiers "TRUE" and "FALSE" as boolean literals, if and only if those
identifiers are not already used for some other meaning.  If there already
exists columns or tables or other objects named TRUE or FALSE, then for
the sake of backwards compatibility, the TRUE and FALSE identifiers refer
to those other objects, not to the boolean values.

<p>The boolean identifiers TRUE and FALSE are usually just aliases for
the integer values 1 and 0, respectively.  However, if TRUE or FALSE
occur on the right-hand side of an IS operator, then they form new
unary postfix operators "IS TRUE" and "IS FALSE" which test the boolean
value of the operand on the left.

<h3>Functions</h3>
<p>SQLite supports many [corefunc|simple], [aggfunc|aggregate],
and [window functions|window]
SQL functions.  For presentation purposes, simple functions are further
subdivided into [corefunc | core functions], [datefunc|date-time functions],
and [json1|JSON functions].
Applications can add new functions, written in C/C++, using the
[sqlite3_create_function()] interface.
</p>

<p>
The main expression bubble diagram above shows a single syntax for
all function invocations.  But this is merely to simplify the expression
bubble diagram.  In reality, each type of function has a slightly different
syntax, shown below.  The function invocation syntax shown in the main
expression bubble diagram is the union of the three syntaxes shown here:

<tcl>
BubbleDiagram simple-function-invocation
BubbleDiagram aggregate-function-invocation
BubbleDiagram window-function-invocation
</tcl>

<p>
The OVER clause is required for [window functions] and is prohibited
otherwise.  The DISTINCT keyword is only allowed in [aggfunc|aggregate functions].
The FILTER clause may not appear on a [corefunc|simple function].

<p>^It is possible to have an aggregate function with the same name as a
simple function, as long as the number of arguments for the two forms of the
function are different.  ^For example, the [agg_max|max()] function with a
single argument is an aggregate and the [max()] function with two or more
arguments is a simple function.  Aggregate functions can usually also
be used as window functions.

<tcl>
##############################################################################
Section {Core Functions} corefunc {*corefunc}
unset -nocomplain corefuncset
proc funcdef {syntax keywords desc} {
  global corefuncset
  regexp {^[a-z_]+} $syntax basesyntax
  set corefuncset($basesyntax) [list $syntax $keywords $desc]
}
funcdef {abs(X)} {} {
  ^The abs(X) function returns the absolute value of the numeric
  argument X.  ^Abs(X) returns NULL if X is NULL. 
  ^(Abs(X) returns 0.0 if X is a string or blob
  that cannot be converted to a numeric value.)^  ^If X is the 
  integer -9223372036854775808 then abs(X) throws an integer overflow
  error since there is no equivalent positive 64-bit two complement value.
}

funcdef {changes()} {} {
  ^The changes() function returns the number of database rows that were changed
  or inserted or deleted by the most recently completed INSERT, DELETE,
  or UPDATE statement, exclusive of statements in lower-level triggers.
  ^The changes() SQL function is a wrapper around the [sqlite3_changes()]
  C/C++ function and hence follows the same rules for counting changes.
}

funcdef {char(X1,X2,...,XN)} {} {
  ^(The char(X1,X2,...,XN) function returns a string composed of characters having the
   unicode code point values of integers X1 through XN, respectively.)^
}

funcdef {coalesce(X,Y,...)} {} {
  ^The coalesce() function returns a copy of its first non-NULL argument, or
  NULL if all arguments are NULL.  ^Coalesce() must have at least 
  2 arguments.
}

funcdef {glob(X,Y)} {} {
  ^The glob(X,Y) function is equivalent to the
  expression "<b>Y GLOB X</b>".
  Note that the X and Y arguments are reversed in the glob() function
  relative to the infix [GLOB] operator.  Y is the string and X is the
  pattern.  So, for example, the following expressions are equivalent:
  <blockquote><pre>
     name GLOB '*helium*'
     glob('*helium*',name)
  </pre></blockquote>
  <p>^If the [sqlite3_create_function()] interface is used to
  override the glob(X,Y) function with an alternative implementation then
  the [GLOB] operator will invoke the alternative implementation.
}

funcdef {ifnull(X,Y)} {} {
  ^The ifnull() function returns a copy of its first non-NULL argument, or
  NULL if both arguments are NULL.  ^Ifnull() must have exactly 2 arguments.
  ^The ifnull() function is equivalent to [coalesce()] with two arguments.
}

funcdef {instr(X,Y)} {} {
  ^The instr(X,Y) function finds the first occurrence of string Y within 
  string X and returns the number of prior characters plus 1, or 0 if
  Y is nowhere found within X.
  ^Or, if X and Y are both BLOBs, then instr(X,Y) returns one
  more than the number bytes prior to the first occurrence of Y, or 0 if
  Y does not occur anywhere within X.
  ^If both arguments X and Y to instr(X,Y) are non-NULL and are not BLOBs
  then both are interpreted as strings.
  ^If either X or Y are NULL in instr(X,Y) then the result is NULL.
}

funcdef {hex(X)} {} {
  ^The hex() function interprets its argument as a BLOB and returns
  a string which is the upper-case hexadecimal rendering of the content of
  that blob.
}

funcdef {last_insert_rowid()} {} {
  ^The last_insert_rowid() function returns the [ROWID]
  of the last row insert from the database connection which invoked the
  function.
  ^The last_insert_rowid() SQL function is a wrapper around the
  [sqlite3_last_insert_rowid()] C/C++ interface function.
}

funcdef {length(X)} {} {
  ^For a string value X, the length(X) function returns the number of 
  characters (not bytes) in X prior to the first NUL character.
  Since SQLite strings do not normally contain NUL characters, the length(X)
  function will usually return the total number of characters in the string X.
  ^For a blob value X, length(X) returns the number of bytes in the blob.
  ^If X is NULL then length(X) is NULL.
  ^If X is numeric then length(X) returns the length of a string
  representation of X.
}

funcdef {like(X,Y) like(X,Y,Z)} {} {
  ^The like() function is used to implement the
  "<b>Y LIKE X &#91;ESCAPE Z&#93;</b>" expression. 
  ^If the optional ESCAPE clause is present, then the
  like() function is invoked with three arguments.  ^Otherwise, it is
  invoked with two arguments only. Note that the X and Y parameters are
  reversed in the like() function relative to the infix [LIKE] operator.
  X is the pattern and Y is the string to match against that pattern.
  Hence, the following expressions are equivalent:
  <blockquote><pre>
     name LIKE '%neon%'
     like('%neon%',name)
  </pre></blockquote>
  ^The [sqlite3_create_function()] interface can be used to override the
  like() function and thereby change the operation of the
  [LIKE] operator.  When overriding the like() function, it may be important
  to override both the two and three argument versions of the like() 
  function. Otherwise, different code may be called to implement the
  [LIKE] operator depending on whether or not an ESCAPE clause was 
  specified.
}


funcdef {likelihood(X,Y)} {} {
  ^The likelihood(X,Y) function returns argument X unchanged.
  ^(The value Y in likelihood(X,Y) must be a floating point constant
  between 0.0 and 1.0, inclusive.)^
  ^The likelihood(X) function is a no-op that the code generator
  optimizes away so that it consumes no CPU cycles during run-time
  (that is, during calls to [sqlite3_step()]).
  ^The purpose of the likelihood(X,Y) function is to provide a hint
  to the query planner that the argument X is a boolean that is
  true with a probability of approximately Y.
  ^(The [unlikely(X)] function is short-hand for likelihood(X,0.0625).)^
  ^(The [likely(X)] function is short-hand for likelihood(X,0.9375).)^
}

funcdef {likely(X)} {} {
  ^The likely(X) function returns the argument X unchanged.
  ^The likely(X) function is a no-op that the code generator
  optimizes away so that it consumes no CPU cycles at
  run-time (that is, during calls to [sqlite3_step()]).
  ^The purpose of the likely(X) function is to provide a hint
  to the query planner that the argument X is a boolean value
  that is usually true. ^(The likely(X) function is equivalent
  to [likelihood](X,0.9375).)^ See also: [unlikely(X)].
}

funcdef {load_extension(X) load_extension(X,Y)} {} {
  ^The load_extension(X,Y) function loads [SQLite extensions] out of the shared
  library file named X using the entry point Y.  ^The result of load_extension()
  is always a NULL.  ^If Y is omitted then the default entry point name is used.
  ^The load_extension() function raises an exception if the extension fails to
  load or initialize correctly.

  <p>^The load_extension() function will fail if the extension attempts to 
  modify or delete an SQL function or collating sequence.  ^The
  extension can add new functions or collating sequences, but cannot
  modify or delete existing functions or collating sequences because
  those functions and/or collating sequences might be used elsewhere
  in the currently running SQL statement.  To load an extension that
  changes or deletes functions or collating sequences, use the
  [sqlite3_load_extension()] C-language API.</p>

  <p>For security reasons, extension loaded is turned off by default and must
  be enabled by a prior call to [sqlite3_enable_load_extension()].</p>
}

funcdef {lower(X)} {} {
  ^The lower(X) function returns a copy of string X with all ASCII characters
  converted to lower case.  ^The default built-in lower() function works
  for ASCII characters only.  To do case conversions on non-ASCII
  characters, load the ICU extension.
}

funcdef {ltrim(X) ltrim(X,Y)} {} {
  ^The ltrim(X,Y) function returns a string formed by removing any and all
  characters that appear in Y from the left side of X.
  ^If the Y argument is omitted, ltrim(X) removes spaces from the left side
  of X.
}

funcdef {max(X,Y,...)} {*maxCoreFunc *max {max() SQL function}} {
  ^The multi-argument max() function returns the argument with the 
  maximum value, or return NULL if any argument is NULL. 
  ^The multi-argument max() function searches its arguments from left to right
  for an argument that defines a collating function and uses that collating
  function for all string comparisons.  ^If none of the arguments to max()
  define a collating function, then the BINARY collating function is used.
  ^(Note that <b>max()</b> is a simple function when
  it has 2 or more arguments but operates as an
  [maxAggFunc | aggregate function] if given only a single argument.)^
}

funcdef {min(X,Y,...)} {*minCoreFunc *min {min() SQL function}} {
  ^The multi-argument min() function returns the argument with the
  minimum value.
  ^The multi-argument min() function searches its arguments from left to right
  for an argument that defines a collating function and uses that collating
  function for all string comparisons.  ^If none of the arguments to min()
  define a collating function, then the BINARY collating function is used.
  ^(Note that <b>min()</b> is a simple function when
  it has 2 or more arguments but operates as an 
  [minAggFunc | aggregate function] if given
  only a single argument.)^
}

funcdef {nullif(X,Y)} {} {
  ^The nullif(X,Y) function returns its first argument if the arguments are
  different and NULL if the arguments are the same.  ^The nullif(X,Y) function
  searches its arguments from left to right for an argument that defines a
  collating function and uses that collating function for all string
  comparisons.  ^If neither argument to nullif() defines a collating function
  then the BINARY is used.
}

funcdef {printf(FORMAT,...)} {} {
  ^(The printf(FORMAT,...) SQL function works like the [sqlite3_mprintf()] C-language
  function and the printf() function from the standard C library.)^
  The first argument is a format string that specifies how to construct the output
  string using values taken from subsequent arguments.  ^If the FORMAT argument is
  missing or NULL then the result is NULL.  ^The %n format is silently ignored and
  does not consume an argument.  ^The %p format is an alias for %X.  ^The %z format
  is interchangeable with %s.  ^(If there are too few arguments in the argument list,
  missing arguments are assumed to have a NULL value, which is translated into
  0 or 0.0 for numeric formats or an empty string for %s.)^  See the
  [built-in printf()] documentation for additional information.
}
  

funcdef {quote(X)} {} {
  ^The quote(X) function returns the text of an SQL literal which
  is the value of its argument suitable for inclusion into an SQL statement.
  ^Strings are surrounded by single-quotes with escapes on interior quotes
  as needed.  ^BLOBs are encoded as hexadecimal literals.
  ^Strings with embedded NUL characters cannot be represented as string
  literals in SQL and hence the returned string literal is truncated prior
  to the first NUL.
}

funcdef {random()} {} {
  ^The random() function returns a pseudo-random integer
  between -9223372036854775808 and +9223372036854775807.
}

funcdef {randomblob(N)} {} {
  ^The randomblob(N) function return an N-byte blob containing pseudo-random
  bytes. ^If N is less than 1 then a 1-byte random blob is returned.

  <p>Hint:  applications can generate globally unique identifiers
  using this function together with [hex()] and/or
  [lower()] like this:</p>

  <blockquote>
  hex(randomblob(16))<br></br>
  lower(hex(randomblob(16)))
  </blockquote>
}

funcdef {replace(X,Y,Z)} {} {
  ^The replace(X,Y,Z) function returns a string formed by substituting
  string Z for every occurrence of string Y in string X.  ^The [BINARY]
  collating sequence is used for comparisons.  ^If Y is an empty
  string then return X unchanged.  ^If Z is not initially
  a string, it is cast to a UTF-8 string prior to processing.
}

funcdef {round(X) round(X,Y)} {} {
  ^The round(X,Y) function returns a floating-point
  value X rounded to Y digits to the right of the decimal point.
  ^If the Y argument is omitted, it is assumed to be 0.
}

funcdef {rtrim(X) rtrim(X,Y)} {} {
  ^The rtrim(X,Y) function returns a string formed by removing any and all
  characters that appear in Y from the right side of X.
  ^If the Y argument is omitted, rtrim(X) removes spaces from the right
  side of X.
}

funcdef {soundex(X)} {} {
  ^The soundex(X) function returns a string that is the soundex encoding 
  of the string X.
  ^The string "?000" is returned if the argument is NULL or contains
  no ASCII alphabetic characters.
  ^(This function is omitted from SQLite by default.
  It is only available if the [SQLITE_SOUNDEX] compile-time option
  is used when SQLite is built.)^
}

funcdef {sqlite_compileoption_get(N)} {} {
  ^The sqlite_compileoption_get() SQL function is a wrapper around the
  [sqlite3_compileoption_get()] C/C++ function.
  ^This routine returns the N-th compile-time option used to build SQLite
  or NULL if N is out of range.  See also the [compile_options pragma].
}

funcdef {sqlite_compileoption_used(X)} {} {
  ^The sqlite_compileoption_used() SQL function is a wrapper around the
  [sqlite3_compileoption_used()] C/C++ function.
  ^When the argument X to sqlite_compileoption_used(X) is a string which
  is the name of a compile-time option, this routine returns true (1) or
  false (0) depending on whether or not that option was used during the
  build.
}

funcdef {sqlite_offset(X)} {} {
  ^The sqlite_offset(X) function returns the byte offset in the database
  file for the beginning of the record from which value would be read.
  If X is not a column in an ordinary table, then sqlite_offset(X) returns
  NULL.  The value returned by sqlite_offset(X) might reference either the
  original table or an index, depending on the query.  If the value X would
  normally be extracted from an index, the sqlite_offset(X) returns the
  offset to the corresponding index record.  If the value X would be
  extracted from the original table, then sqlite_offset(X) returns the offset
  to the table record.

  <p>The sqlite_offset(X) SQL function is only available if SQLite is built
  using the [-DSQLITE_ENABLE_OFFSET_SQL_FUNC] compile-time option.
}

funcdef {sqlite_source_id()} {} {
  ^The sqlite_source_id() function returns a string that identifies the
  specific version of the source code that was used to build the SQLite
  library.  ^The string returned by sqlite_source_id() is
  the date and time that the source code was checked in followed by
  the SHA1 hash for that check-in.  ^This function is
  an SQL wrapper around the [sqlite3_sourceid()] C interface.
}

funcdef {sqlite_version()} {} {
  ^The sqlite_version() function returns the version string for the SQLite
  library that is running.  ^This function is an SQL
  wrapper around the [sqlite3_libversion()] C-interface.
}

funcdef {substr(X,Y,Z) substr(X,Y)} {} {
  ^The substr(X,Y,Z) function returns a substring of input string X that begins
  with the Y-th character and which is Z characters long.
  ^If Z is omitted then substr(X,Y) returns all characters through the end
  of the string X beginning with the Y-th.
  ^The left-most character of X is number 1.  ^If Y is negative
  then the first character of the substring is found by counting from the
  right rather than the left.  ^If Z is negative then
  the abs(Z) characters preceding the Y-th character are returned.
  ^If X is a string then characters indices refer to actual UTF-8 
  characters.  ^If X is a BLOB then the indices refer to bytes.
}

funcdef {total_changes()} {} {
  ^The total_changes() function returns the number of row changes
  caused by INSERT, UPDATE or DELETE
  statements since the current database connection was opened.
  ^This function is a wrapper around the [sqlite3_total_changes()]
  C/C++ interface.
}

funcdef {trim(X) trim(X,Y)} {} {
  ^The trim(X,Y) function returns a string formed by removing any and all
  characters that appear in Y from both ends of X.
  ^If the Y argument is omitted, trim(X) removes spaces from both ends of X.
}

funcdef {typeof(X)} {} {
  ^The typeof(X) function returns a string that indicates the [datatype] of
  the expression X: "null", "integer", "real", "text", or "blob".
}

funcdef {unlikely(X)} {} {
  ^The unlikely(X) function returns the argument X unchanged.
  ^The unlikely(X) function is a no-op that the code generator
  optimizes away so that it consumes no CPU cycles at
  run-time (that is, during calls to [sqlite3_step()]).
  ^The purpose of the unlikely(X) function is to provide a hint
  to the query planner that the argument X is a boolean value
  that is usually not true. ^(The unlikely(X) function is equivalent
  to [likelihood](X, 0.0625).)^
}

funcdef {unicode(X)} {} {
  ^The unicode(X) function returns the numeric unicode code point corresponding to
  the first character of the string X.  If the argument to unicode(X) is not a string
  then the result is undefined.
}

funcdef {upper(X)} {} {
  ^The upper(X) function returns a copy of input string X in which all 
  lower-case ASCII characters are converted to their upper-case equivalent.
}

funcdef {zeroblob(N)} {} {
  ^The zeroblob(N) function returns a BLOB consisting of N bytes of 0x00.
  SQLite manages these zeroblobs very efficiently.  Zeroblobs can be used to
  reserve space for a BLOB that is later written using 
  [sqlite3_blob_open() | incremental BLOB I/O].
  ^This SQL function is implemented using the [sqlite3_result_zeroblob()]
  routine from the C/C++ interface.
}
</tcl>

<p>The core functions shown below are available by default. 
[datefunc | Date &amp; Time functions],
[aggfunc | aggregate functions],
[builtin window functions|window functions], and
[json1 | JSON functions] are documented separately.  An
application may define additional
functions written in C and added to the database engine using
the [sqlite3_create_function()] API.</p>

<tcl>
set lx {}
foreach basesyntax [array names corefuncset] {
  foreach {syntax keywords desc} $corefuncset($basesyntax) break
  regexp {^[a-z_]+} $basesyntax fragment
  foreach coresyntax $syntax {
    lappend lx [list $fragment $coresyntax 0]
  }
}
hd_list_of_links {} 225 [lsort -index 1 $lx]

hd_putsnl "<hr class='xhr'>"
hd_putsnl "<dl>"
foreach basesyntax [lsort [array names corefuncset]] {
  foreach {syntax keywords desc} $corefuncset($basesyntax) break
  regsub -all {\s+} [string trim $syntax] {<br />} syntax
  regsub -all {\(([^*)]+)\)} $syntax {(<i>\1</i>)} syntax
  regsub -all {,} $syntax {</i>,<i>} syntax
  regsub -all {<i>\.\.\.</i>} $syntax {...} syntax
  regexp {^[a-z]+} $basesyntax fragment
  if {[llength $keywords]==0} {
    regexp {[a-z_]+} $syntax name
    hd_fragment $name *$name "${name}() SQL function"
  } else {
    set fragname [lindex $keywords 0]
    regsub -all {[^a-z]} $fragname {} fragname
    hd_fragment $fragname
    eval hd_keywords [string map {\n { }} $keywords]
  }
  hd_puts "<dt><p><b>$syntax</b></dt>"
  hd_resolve "<dd><p>$desc</dd>\n"
}
hd_putsnl "</dl>"

##############################################################################
Section {Date And Time Functions} datefunc {*datefunc {date and time functions}}
hd_keywords {date() SQL function} {time() SQL function}
hd_keywords {datetime() SQL function} {julianday() SQL function}
hd_keywords {strftime() SQL function}
</tcl>

<p>
SQLite supports five date and time functions as follows:
</p>

<p>
<ol>
<li> ^(<b>date(</b><i>timestring, modifier, modifier, ...</i><b>)</b>)^ </li>
<li> ^(<b>time(</b><i>timestring, modifier, modifier, ...</i><b>)</b>)^ </li>
<li> ^(<b>datetime(</b><i>timestring, modifier, modifier, ...</i><b>)</b>)^ </li>
<li> ^(<b>julianday(</b><i>timestring, modifier, modifier, ...</i><b>)</b>)^ </li>
<li> ^(<b>strftime(</b><i>format, timestring, modifier, modifier, ...</i><b>)</b>)^ </li>
</ol>

<p>
^All five date and time functions take a time string as an argument. 
^The time string is followed by zero or more modifiers. 
^The strftime() function also takes a format string as its first argument.
</p>

<p>
The date and time functions use a subset of
[http://en.wikipedia.org/wiki/ISO_8601 | IS0-8601] date and time
formats.
^The date() function returns the date in this format: YYYY-MM-DD. 
^The time() function returns the time as HH:MM:SS. 
^The datetime() function returns "YYYY-MM-DD HH:MM:SS". 
^(The julianday() function returns the 
[http://en.wikipedia.org/wiki/Julian_day | Julian day] - the
number of days since noon in Greenwich on November 24, 4714 B.C. 
([http://en.wikipedia.org/wiki/Proleptic_Gregorian_calendar | Proleptic Gregorian calendar]).)^
^The strftime() routine returns the date formatted according to 
the format string specified as the first argument.
^The format string supports the most common substitutions found in the 
[http://opengroup.org/onlinepubs/007908799/xsh/strftime.html | strftime() function]
from the standard C library plus two new substitutions, %f and %J.
^(The following is a complete list of valid strftime() substitutions:
</p>

<blockquote>
<table border="0" cellpadding="0" cellspacing="0">
<tr><td><td width="10"><td></tr>

<tr><td> %d <td><td> day of month: 00
<tr><td> %f <td><td> fractional seconds: SS.SSS
<tr><td> %H <td><td> hour: 00-24 
<tr><td> %j <td><td> day of year: 001-366
<tr><td> %J <td><td> Julian day number
<tr><td> %m <td><td> month: 01-12
<tr><td> %M <td><td> minute: 00-59
<tr><td> %s <td><td> seconds since 1970-01-01
<tr><td> %S <td><td> seconds: 00-59
<tr><td> %w <td><td> day of week 0-6 with Sunday==0
<tr><td> %W <td><td> week of year: 00-53
<tr><td> %Y <td><td> year: 0000-9999
<tr><td> %% <td><td> %
</table>
</blockquote>)^

<p>
^(Notice that all other date and time functions can be expressed
in terms of strftime():
</p>

<blockquote>
<table border="0" cellpadding="0" cellspacing="0">
<tr><td><b>Function</b><td width="30"><td><b>Equivalent strftime()</b>
<tr><td>   date(...)      <td><td>  strftime('%Y-%m-%d', ...)
<tr><td>   time(...)      <td><td>  strftime('%H:%M:%S', ...)
<tr><td>   datetime(...)  <td><td>  strftime('%Y-%m-%d %H:%M:%S', ...)
<tr><td>   julianday(...) <td><td>  strftime('%J', ...)
</table>
</blockquote>)^

<p>
The only reasons for providing functions other than strftime() is
for convenience and for efficiency.
</p>

<h3>Time Strings</h3>

<p>^(A time string can be in any of the following formats:</p>

<ol>
<li> <i>YYYY-MM-DD</i>
<li> <i>YYYY-MM-DD HH:MM</i>
<li> <i>YYYY-MM-DD HH:MM:SS</i>
<li> <i>YYYY-MM-DD HH:MM:SS.SSS</i>
<li> <i>YYYY-MM-DD</i><b>T</b><i>HH:MM</i>
<li> <i>YYYY-MM-DD</i><b>T</b><i>HH:MM:SS</i>
<li> <i>YYYY-MM-DD</i><b>T</b><i>HH:MM:SS.SSS</i>
<li> <i>HH:MM</i>
<li> <i>HH:MM:SS</i>
<li> <i>HH:MM:SS.SSS</i>
<li> <b>now</b>
<li> <i>DDDDDDDDDD</i>
</ol>)^

<p>
^In formats 5 through 7, the "T" is a literal character separating 
the date and the time, as required by 
[http://www.w3c.org/TR/NOTE-datetime | ISO-8601]. 
^Formats 8 through 10 that specify only a time assume a date of 
2000-01-01. Format 11, the string 'now', is converted into the 
current date and time as obtained from the xCurrentTime method
of the [sqlite3_vfs] object in use.
^The 'now' argument to date and time functions always returns exactly the
same value for multiple invocations within the same [sqlite3_step()] call.
^[http://en.wikipedia.org/wiki/Coordinated_Universal_Time | Universal Coordinated Time (UTC)] is used. 
^Format 12 is the 
[http://en.wikipedia.org/wiki/Julian_day | Julian day number]
expressed as a floating point value.
</p>

<p>
Formats 2 through 10 may be optionally followed by a timezone indicator of the form
"<i>&#91;+-&#93;HH:MM</i>" or just "<i>Z</i>".  The date and time functions use UTC or "zulu"
time internally, and so the "Z" suffix is a no-op.  Any non-zero "HH:MM" suffix is
subtracted from the indicated date and time in order to compute zulu time.
For example, all of the following time strings are equivalent:
</p>

<blockquote>
2013-10-07 08:23:19.120<br>
2013-10-07T08:23:19.120Z<br>
2013-10-07 04:23:19.120-04:00<br>
2456572.84952685
</blockquote>

<p>
^(In formats 4, 7, and 10, the fractional seconds value SS.SSS can have
one or more digits following the decimal point.  Exactly three digits are
shown in the examples because only the first three digits are significant
to the result, but the input string can have fewer or more than three digits
and the date/time functions will still operate correctly.)^
Similarly, format 12 is shown with 10 significant digits, but the date/time
functions will really accept as many or as few digits as are necessary to
represent the Julian day number.
</p>

<h3>Modifiers</h3>

<p>^The time string can be followed by zero or more modifiers that 
alter date and/or time. ^Each modifier
is a transformation that is applied to the time value to its left.
^Modifiers are applied from left to right; order is important.
^(The available modifiers are as follows.</p>

<ol>
<li> NNN days
<li> NNN hours
<li> NNN minutes
<li> NNN.NNNN seconds
<li> NNN months
<li> NNN years
<li> start of month
<li> start of year
<li> start of day
<li> weekday N
<li> unixepoch
<li> localtime
<li> utc 
</ol>)^

<p>^The first six modifiers (1 through 6) 
simply add the specified amount of time to the date and time 
specified by the preceding timestring and modifiers.
^The 's' character at the end of the modifier names is optional.
^Note that "&plusmn;NNN months" works by rendering the original date into
the YYYY-MM-DD format, adding the &plusmn;NNN to the MM month value, then
normalizing the result.  ^Thus, for example, the data 2001-03-31 modified
by '+1 month' initially yields 2001-04-31, but April only has 30 days
so the date is normalized to 2001-05-01.  ^A similar effect occurs when
the original date is February 29 of a leapyear and the modifier is
&plusmn;N years where N is not a multiple of four.</p>

<p>^The "start of" modifiers (7 through 9) shift the date backwards 
to the beginning of the current month, year or day.</p>

<p>^(The "weekday" modifier advances the date forward, if necessary,
to the next date where the weekday number is N. Sunday is 0, Monday is 1,
and so forth.)^
^If the date is already on the desired weekday, the "weekday" modifier
leaves the date unchanged.  </p>

<p>^The "unixepoch" modifier (11) only works if it immediately follows 
a timestring in the DDDDDDDDDD format. 
^This modifier causes the DDDDDDDDDD to be interpreted not 
as a Julian day number as it normally would be, but as
[http://en.wikipedia.org/wiki/Unix_time | Unix Time] - the 
number of seconds since 1970.  If the "unixepoch" modifier does not
follow a timestring of the form DDDDDDDDDD which expresses the number
of seconds since 1970 or if other modifiers
separate the "unixepoch" modifier from prior DDDDDDDDDD then the
behavior is undefined.
For SQLite versions before 3.16.0 ([dateof:3.16.0]), 
the "unixepoch" modifier only works for
dates between 0000-01-01 00:00:00 and 5352-11-01 10:52:47 (unix times
of -62167219200 through 106751991167).</p>

<tcl>hd_fragment localtime {localtime modifier} {'utc' modifier}</tcl>
<p>^The "localtime" modifier (12) assumes the time string to its left is in
Universal Coordinated Time (UTC) and adjusts the time
string so that it displays localtime.  If "localtime"
follows a time that is not UTC, then the behavior is undefined.
^(The "utc" modifier is the opposite of "localtime".  
"utc" assumes that the string
to its left is in the local timezone and adjusts that string to be in UTC.)^
If the prior string is not in localtime, then the result of "utc" is
undefined.</p>

<h3>Examples</h3>

^(<p>Compute the current date.<p>

<blockquote>SELECT date('now');</blockquote>)^

^(<p>Compute the last day of the current month.</p>

<blockquote>SELECT date('now','start of month','+1 month','-1 day');
</blockquote>)^

^(<p>Compute the date and time given a unix timestamp 1092941466.</p>

<blockquote>
    SELECT datetime(1092941466, 'unixepoch');
</blockquote>)^

^(<p>Compute the date and time given a unix timestamp 1092941466, and 
compensate for your local timezone.</p>

<blockquote>
  SELECT datetime(1092941466, 'unixepoch', 'localtime');
</blockquote>)^

^(<p>Compute the current unix timestamp.</p>

<blockquote>
  SELECT strftime('%s','now');
</blockquote>)^

^(<p>Compute the number of days since the signing of the US Declaration
of Independence.</p>

<blockquote>
  SELECT julianday('now') - julianday('1776-07-04');
</blockquote>)^

^(<p>Compute the number of seconds since a particular moment in 2004:</p>

<blockquote>
  SELECT strftime('%s','now') - strftime('%s','2004-01-01 02:34:56');
</blockquote>)^

^(<p>
Compute the date of the first Tuesday in October
for the current year.
</p>

<blockquote>
  SELECT date('now','start of year','+9 months','weekday 2');
</blockquote>)^

^(<p>Compute the time since the unix epoch in seconds 
(like strftime('%s','now') except includes fractional part):</p>

<blockquote>
  SELECT (julianday('now') - 2440587.5)*86400.0;
</blockquote>)^

<h3>Caveats And Bugs</h3>

<p>The computation of local time depends heavily on the whim 
of politicians and is thus difficult to get correct for 
all locales. ^In this implementation, the standard C library 
function localtime_r() is used to assist in the calculation of 
local time.  ^(The 
localtime_r() C function normally only works for years
between 1970 and 2037. For dates outside this range, SQLite 
attempts to map the year into an equivalent year within 
this range, do the calculation, then map the year back.)^</p>


<p>^(These functions only work for dates between 0000-01-01 00:00:00
and 9999-12-31 23:59:59 (julian day numbers 1721059.5 through 5373484.5).)^
For dates outside that range, the results of these
functions are undefined.</p>

<p>Non-Vista Windows platforms only support one set of DST rules. 
Vista only supports two. Therefore, on these platforms, 
historical DST calculations will be incorrect. 
For example, in the US, in 2007 the DST rules changed. 
Non-Vista Windows platforms apply the new 2007 DST rules 
to all previous years as well. Vista does somewhat better
getting results correct back to 1986, when the rules were also changed.</p>

<p>All internal computations assume the 
[http://en.wikipedia.org/wiki/Gregorian_calendar | Gregorian calendar]
system.  It is also assumed that every
day is exactly 86400 seconds in duration.</p>

<tcl>
##############################################################################
Section {Aggregate Functions} aggfunc {*aggfunc}
unset -nocomplain corefuncset

funcdef {avg(X)} {*avg {avg() aggregate function}} {
  ^The avg() function
  returns the average value of all non-NULL <i>X</i> within a
  group.  ^String and BLOB values that do not look like numbers are
  interpreted as 0.
  ^The result of avg() is always a floating point value as long as
  at there is at least one non-NULL input even if all
  inputs are integers.  ^The result of avg() is NULL if and only if
  there are no non-NULL inputs.  
}

funcdef {count(X) count(*)} {*count {count() aggregate function}} {
  ^The count(X) function returns
  a count of the number of times
  that <i>X</i> is not NULL in a group.  ^The count(*) function
  (with no arguments) returns the total number of rows in the group.
}

funcdef {group_concat(X) group_concat(X,Y)} {
  *group_concat {group_concat() aggregate function}
} {
  ^The group_concat() function returns
  a string which is the concatenation of
  all non-NULL values of <i>X</i>.  ^If parameter <i>Y</i> is present then
  it is used as the separator
  between instances of <i>X</i>.  ^A comma (",") is used as the separator
  if <i>Y</i> is omitted.  The order of the concatenated elements is
  arbitrary.
}

funcdef {max(X)} {*maxAggFunc *agg_max {max() aggregate function}} {
  ^The max() aggregate function
  returns the maximum value of all values in the group.
  ^The maximum value is the value that would be returned last in an
  ORDER BY on the same column.  ^Aggregate max() returns NULL 
  if and only if there are no non-NULL values in the group.
}

funcdef {min(X)} {*minAggFunc *agg_min {min() aggregate function}} {
  ^The min() aggregate function
  returns the minimum non-NULL value of all values in the group.
  ^The minimum value is the first non-NULL value that would appear
  in an ORDER BY of the column.
  ^Aggregate min() returns NULL if and only if there are no non-NULL
  values in the group.
}

funcdef {sum(X) total(X)} {
  *sumFunc *sum *total
  {sum() aggregate function}
  {total() aggregate function}
} {
  ^The sum() and total() aggregate functions
  return sum of all non-NULL values in the group.
  ^If there are no non-NULL input rows then sum() returns
  NULL but total() returns 0.0.
  NULL is not normally a helpful result for the sum of no rows
  but the SQL standard requires it and most other
  SQL database engines implement sum() that way so SQLite does it in the
  same way in order to be compatible.   The non-standard total() function
  is provided as a convenient way to work around this design problem
  in the SQL language.</p>

  <p>^The result of total() is always a floating point value.
  ^The result of sum() is an integer value if all non-NULL inputs are integers.
  ^If any input to sum() is neither an integer or a NULL
  then sum() returns a floating point value
  which might be an approximation to the true sum.</p>

  <p>^Sum() will throw an "integer overflow" exception if all inputs
  are integers or NULL
  and an integer overflow occurs at any point during the computation.
  ^Total() never throws an integer overflow.
}
</tcl>


<tcl>
RecursiveBubbleDiagram aggregate-function-invocation
</tcl>

<p>
The aggregate functions shown below are available by default.  Additional
aggregate functions written in C may be added using the 
[sqlite3_create_function()]</a>
API.</p>

<p>
^In any aggregate function that takes a single argument, that argument
can be preceded by the keyword DISTINCT.  ^In such cases, duplicate
elements are filtered before being passed into the aggregate function.
^For example, the function "count(distinct X)" will return the number
of distinct values of column X instead of the total number of non-null
values in column X.
</p>

<tcl>hd_fragment aggfilter {FILTER clause on aggregate functions}</tcl>
<p>
If a FILTER clause is provided, then only rows for which the <i>expr</i> is
true are included in the aggregate.
</p>

<tcl>
set lx {}
foreach basesyntax [array names corefuncset] {
  foreach {syntax keywords desc} $corefuncset($basesyntax) break
  regexp {^[a-z_]+} $basesyntax fragment
  foreach coresyntax $syntax {
    lappend lx [list $fragment $coresyntax 0]
  }
}
hd_list_of_links {} 225 [lsort -index 1 $lx]

hd_putsnl "<hr class='xhr'>"
hd_putsnl "<dl>"
foreach basesyntax [lsort [array names corefuncset]] {
  foreach {syntax keywords desc} $corefuncset($basesyntax) break
  regsub -all {\s+} [string trim $syntax] {<br />} syntax
  regsub -all {\(([^*)]+)\)} $syntax {(<i>\1</i>)} syntax
  regsub -all {,} $syntax {</i>,<i>} syntax
  regsub -all {<i>\.\.\.</i>} $syntax {...} syntax
  regexp {^[a-z]+} $basesyntax fragment
  if {[llength $keywords]==0} {
    regexp {[a-z_]+} $syntax name
    hd_fragment $name *$name "${name}() SQL function"
  } else {
    set fragname [lindex $keywords 0]
    regsub -all {[^a-z]} $fragname {} fragname
    hd_fragment $fragname
    eval hd_keywords [string map {\n { }} $keywords]
  }
  hd_puts "<dt><p><b>$syntax</b></dt>"
  hd_resolve "<dd><p>$desc</dd>\n"
}
hd_putsnl "</dl>"

##############################################################################
Section INSERT insert {INSERT *INSERTs}

RecursiveBubbleDiagram insert-stmt
</tcl>

<p>The INSERT statement comes in three basic forms.  
<ol>
<li><p><b>INSERT INTO </b><i>table</i><b> VALUES(...);</b>
<p>^The first form (with the "VALUES" keyword) creates one or more
new rows in
an existing table. ^If the <yyterm>column-name</yyterm> list after
<yyterm>table-name</yyterm> is omitted then the number
of values inserted into each row
must be the same as the number of columns in the table. ^In this case
the result of evaluating the left-most expression from each term of
the VALUES list is inserted into the left-most column of each new row,
and so forth for each subsequent expression. ^If a <yyterm>column-name</yyterm>
list is specified, then the number of values in each term of the
VALUE list must match the number of
specified columns. ^Each of the named columns of the new row is populated
with the results of evaluating the corresponding VALUES expression. ^Table
columns that do not appear in the column list are populated with the 
[default column value] (specified as part of the [CREATE TABLE] statement), or
with NULL if no [default value] is specified.

<li><p><b>INSERT INTO </b><i>table</i><b> SELECT ...;</b>
<p>The second form of the INSERT statement contains a [SELECT] statement
instead of a VALUES clause. ^A new entry is inserted into the table for each
row of data returned by executing the SELECT statement. ^If a column-list is
specified, the number of columns in the result of the SELECT must be the same
as the number of items in the column-list. ^Otherwise, if no column-list is
specified, the number of columns in the result of the SELECT must be the same
as the number of columns in the table. ^Any SELECT statement, including
[compound SELECTs] and SELECT statements with [ORDER BY] and/or [LIMIT] clauses, 
may be used in an INSERT statement of this form.

<p>To avoid a parsing ambiguity, the SELECT statement should always
contain a WHERE clause, even if that clause is simply "WHERE true",
if the [upsert-clause] is present.  Without the WHERE clause, the
parser does not know if the token "ON" is part of a join constraint
on the SELECT, or the beginning of the [upsert-clause].

<li><p><b>INSERT INTO </b><i>table</i><b> DEFAULT VALUES;</b>
<p>The third form of an INSERT statement is with DEFAULT VALUES.
^(The INSERT ... DEFAULT VALUES statement inserts a single new row into the
named table.)^ ^Each column of the new row is populated with its 
[default value], or with a NULL if no default value is specified 
as part of the column definition in the [CREATE TABLE] statement.
The [upsert-clause] is not supported after DEFAULT VALUES.

</ol>

<p>
^The initial "INSERT" keyword can be replaced by
"REPLACE" or "INSERT OR <i>action</i>" to specify an alternative
constraint [ON CONFLICT|conflict resolution algorithm] to use during 
that one INSERT command.
For compatibility with MySQL, ^the parser allows the use of the
single keyword <a href="lang_replace.html">REPLACE</a> as an 
alias for "INSERT OR REPLACE".

<p>^(The optional "<i>schema-name</i><b>.</b>" prefix on the 
<yyterm>table-name</yyterm>
is supported for top-level INSERT statements only.)^  ^The table name must be
unqualified for INSERT statements that occur within [CREATE TRIGGER] statements.
^Similarly, the "DEFAULT VALUES" form of the INSERT statement is supported for
top-level INSERT statements only and not for INSERT statements within
triggers.
<p>

<p>The optional "AS <yyterm>alias</yyterm>" phrase provides an alternative
name for the table into which content is being inserted.  The alias name
can be used within WHERE and SET clauses of the [UPSERT].  If there is no
[upsert-clause], then the <yyterm>alias</yyterm> is pointless, but also
harmless.

<p>See the separate [UPSERT] documentation for the additional trailing
syntax that can cause an INSERT to behave as an UPDATE if the INSERT would
otherwise violate a uniqueness constraint.  The [upsert clause] is not
allowed on an "INSERT ... DEFAULT VALUES".


<tcl>
##############################################################################
Section upsert UPSERT {UPSERT {upsert clause}}

RecursiveBubbleDiagram upsert-clause
</tcl>

<p>UPSERT is a special syntax addition to [INSERT] that causes the
INSERT to behave as an [UPDATE] or a no-op if the INSERT would violate
a uniqueness constraint.
UPSERT is not standard SQL.  UPSERT in SQLite follows the
syntax established by PostgreSQL.
UPSERT syntax was added to SQLite with version 3.24.0 ([dateof:3.24.0]).

<p>An UPSERT is an ordinary [INSERT] statement that is followed by
the special ON CONFLICT clause shown above.

<p>The syntax that occurs in between the "ON CONFLICT" and "DO" keywords
is called the "conflict target".  The conflict target specifies a specific
uniqueness constraint that will trigger the upsert.  The conflict target
is required for DO UPDATE upserts, but is optional for DO NOTHING.  When
the conflict target is omitted, the upsert behavior is triggered by a
violation of any uniqueness constraint on the table of the INSERT.

<p>If the insert operation would cause the uniqueness constraint identified
by the conflict-target clause to fail, then the insert is omitted and
either the DO NOTHING or DO UPDATE operation is performed instead.
In the case of a multi-row insert, this decision is made separately
for each row of the insert.

<p>The special UPSERT processing happens only for uniqueness constraint on
the table that is receiving the INSERT. A "uniqueness constraint"
is an explicit UNIQUE or PRIMARY KEY constraint within
the CREATE TABLE statement, or a [unique index].
UPSERT does not intervene for failed NOT NULL or foreign key constraints
or for constraints that are implemented using triggers.

<p>Column names in the expressions of a DO UPDATE refer to the original
unchanged value of the column, before the attempted INSERT.  To use the
value that would have been inserted had the constraint not failed,
add the special "excluded." table qualifier to the column name.

<p>Some examples will help illustrate the difference:

<blockquote><pre>
CREATE TABLE vocabulary(word TEXT PRIMARY KEY, count INT DEFAULT 1);
INSERT INTO vocabulary(word) VALUES('jovial')
  ON CONFLICT(word) DO UPDATE SET count=count+1;
</pre></blockquote>

<p>The upsert above inserts the new vocabulary word "jovial" if that
word is not already in the dictionary, or if it is already in the
dictionary, it increments the counter.  The "count+1" expression
could also be written as "vocabulary.count".  PostgreSQL requires the
second form, but SQLite accepts either.

<blockquote><pre>
CREATE TABLE phonebook(name TEXT PRIMARY KEY, phonenumber TEXT);
INSERT INTO phonebook(name,phonenumber) VALUES('Alice','704-555-1212')
  ON CONFLICT(name) DO UPDATE SET phonenumber=excluded.phonenumber;
</pre></blockquote>

<p>In the second example, the expression in the DO UPDATE clause is
of the form "excluded.phonenumber".  The "excluded." prefix causes the
"phonenumber" to refer to the value for phonenumber that would have been
inserted had there been no conflict.  Hence, the effect of the upsert
is to insert a phonenumber of Alice if none exists, or to overwrite
any prior phonenumber for Alice with the new one.

<p>Note that the DO UPDATE clause acts only on the single row
that experienced the constraint error during INSERT.  It is not
necessary to include a WHERE clause that restrictions the action
to that one row.  The only use for the WHERE clause at
the end of the DO UPDATE is to optionally change the DO UPDATE
into a no-op depending on the original and/or new values.
For example:

<blockquote><pre>
CREATE TABLE phonebook2(
  name TEXT PRIMARY KEY,
  phonenumber TEXT,
  validDate DATE
);
INSERT INTO phonebook2(name,phonenumber,validDate)
  VALUES('Alice','704-555-1212','2018-05-08')
  ON CONFLICT(name) DO UPDATE SET
    phonenumber=excluded.phonenumber,
    validDate=excluded.validDate
  WHERE excluded.validDate>phonebook2.validDate;
</pre></blockquote>

<p>In this last example, the phonebook2 entry is only
updated if the validDate for the newly inserted value is
newer than the entry already in the table.  If the table already
contains an entry with the same name and a current validDate,
then the WHERE clause causes the DO UPDATE to become a no-op.

<tcl>hd_fragment parseambig {UPSERT parsing ambiguity}</tcl>
<h3>Parsing Ambiguity</h3>

<p>When the [INSERT] statement to which the UPSERT is attached
takes its values from a [SELECT] statement, there is a potential
parsing ambiguity.  The parser might not be able to tell if the
"ON" keyword is introducing the UPSERT or if it is the ON clause
of a join.  To work around this, the SELECT statement should always
include a WHERE clause, even if that WHERE clause is just
"WHERE true".

<p>Ambiguous use of ON:

<blockquote><pre>
INSERT INTO t1 SELECT * FROM t2
ON CONFLICT(x) DO UPDATE SET y=excluded.y;
</pre></blockquote>

<p>Ambiguity resolved using a WHERE clause:

<blockquote><pre>
INSERT INTO t1 SELECT * FROM t2 <font color="blue">WHERE true</font>
ON CONFLICT(x) DO UPDATE SET y=excluded.y;
</pre></blockquote>

<h3>Limitations</h3>

<p>UPSERT does not currently work for [virtual tables].


<tcl>
##############################################################################
Section {ON CONFLICT clause} conflict {{conflict clause} {ON CONFLICT}}

RecursiveBubbleDiagram conflict-clause
</tcl>

<p>The ON CONFLICT clause is a
non-standard extension specific to SQLite 
that can appear in many other SQL commands.
It is given its own section in this document because it is not
part of standard SQL and therefore might not be familiar.</p>

<p>The ON CONFLICT clause described here has been a part of
SQLite since before version 3.0.0 ([dateof:3.0.0]).  The phrase
"ON&nbsp;CONFLICT" is also part of [UPSERT], which is an extension
to [INSERT] added in version 3.24.0 ([dateof:3.24.0]).  Do not
confuse these two separate uses of the "ON&nbsp;CONFLICT" phrase.

<p>^The syntax for the ON CONFLICT clause is as shown above for
the CREATE TABLE command.  ^For the INSERT and
UPDATE commands, the keywords "ON CONFLICT" are replaced by "OR" so that
the syntax reads more naturally.  For example, instead of
"INSERT ON CONFLICT IGNORE" we have "INSERT OR IGNORE".
The keywords change but the meaning of the clause is the same
either way.</p>

<p>The ON CONFLICT clause applies to [UNIQUE], [NOT NULL],
[CHECK], and [PRIMARY KEY] constraints.
The ON CONFLICT algorithm does not
apply to [FOREIGN KEY constraints].
There are five conflict resolution algorithm choices:
ROLLBACK, ABORT, FAIL, IGNORE, and REPLACE.
^The default conflict resolution algorithm is ABORT.  This
is what they mean:</p>

<dl>
<dt><b>ROLLBACK</b></dt>
<dd><p> ^When an applicable constraint violation occurs, the ROLLBACK
resolution algorithm aborts the current SQL statement with
an SQLITE_CONSTRAINT error and rolls back the current transaction.
^If no transaction is
active (other than the implied transaction that is created on every
command) then the ROLLBACK resolution algorithm works the same as the
ABORT algorithm.</p></dd>

<dt><b>ABORT</b></dt>
<dd><p> ^When an applicable constraint violation occurs, the ABORT
resolution algorithm aborts the current SQL statement
with an SQLITE_CONSTRAINT error and backs out any changes
made by the current SQL statement; but changes caused
by prior SQL statements within the same transaction are preserved and the
transaction remains active.
This is the default behavior and the behavior specified by the SQL
standard.</p></dd>

<dt><b>FAIL</b></dt>
<dd><p> ^When an applicable constraint violation occurs, the FAIL
resolution algorithm aborts the current SQL statement with an
SQLITE_CONSTRAINT error.  ^But the FAIL resolution does not
back out prior changes of the SQL statement that failed nor does
it end the transaction.
^For example, if an UPDATE
statement encountered a constraint violation on the 100th row that
it attempts to update, then the first 99 row changes are preserved
but changes to rows 100 and beyond never occur.</p>

<p>The FAIL behavior only works for uniqueness, NOT NULL, and CHECK 
constraints.  A [foreign key constraint] violation causes an ABORT.
</p></dd>

<dt><b>IGNORE</b></dt>
<dd><p> ^When an applicable constraint violation occurs, 
the IGNORE resolution algorithm skips the one row that contains
the constraint violation and continues processing subsequent rows
of the SQL statement as if nothing went wrong.
Other rows before and after the row that
contained the constraint violation are inserted or updated
normally. ^No error is returned for uniqueness, NOT NULL, and
UNIQUE constraint errors when the IGNORE conflict resolution
algorithm is used.  ^However, the IGNORE conflict resolution
algorithm works like ABORT for [foreign key constraint] errors.
</p>
</dd>

<dt><b>REPLACE</b></dt>
<dd><p> ^When a [UNIQUE] or [PRIMARY KEY] constraint violation occurs, 
the REPLACE algorithm
deletes pre-existing rows that are causing the constraint violation
prior to inserting or updating the current row and the command continues 
executing normally.
^If a [NOT NULL] constraint violation occurs, the REPLACE conflict
resolution replaces the NULL value with
the default value for that column, or if the column has no default
value, then the ABORT algorithm is used.
^If a [CHECK constraint] or [foreign key constraint] violation occurs, 
the REPLACE conflict resolution algorithm works like ABORT.</p>

<p>^When the REPLACE conflict resolution strategy deletes rows in order to
satisfy a constraint, [CREATE TRIGGER | delete triggers] fire if and only if
[recursive_triggers pragma | recursive triggers] are enabled.</p>

<p>^The [sqlite3_update_hook | update hook] is not invoked for rows that
are deleted by the REPLACE conflict resolution strategy.  ^Nor does
REPLACE increment the [sqlite3_changes | change counter].
The exceptional behaviors defined in this paragraph might change 
in a future release.</p>
</dl>

<p>^The algorithm specified in the OR clause of an INSERT or UPDATE
overrides any algorithm specified in a CREATE TABLE.
^If no algorithm is specified anywhere, the ABORT algorithm is used.</p>

<tcl>
##############################################################################
Section REINDEX reindex REINDEX

RecursiveBubbleDiagram reindex-stmt
</tcl>

<p>^The REINDEX command is used to delete and recreate indices from scratch.
This is useful when the definition of a collation sequence has changed.
</p>

<p>^If the REINDEX keyword is not followed by a collation-sequence or database 
object identifier, then all indices in all attached databases are rebuilt.

<p>^If the REINDEX keyword is followed by a collation-sequence name, then
all indices in all attached databases that use the named collation sequences
are recreated. 

<p>^Or, if the argument attached to the REINDEX identifies a specific 
database table, then all indices attached to the database table are rebuilt. 
^If it identifies a specific database index, then just that index is recreated.

<p>^For a command of the form "REINDEX <i>name</i>", a match
against <yyterm>collation-name</yyterm> takes precedence over a match
against <yyterm>index-name</yyterm> or <yyterm>table-name</yyterm>.
This ambiguity in the syntax may be avoided by always specifying a
<yyterm>schema-name</yyterm> when reindexing a specific table or index.

<tcl>
###############################################################################
Section REPLACE replace REPLACE

</tcl>

<p>^The REPLACE command is an alias for the "[ON CONFLICT | INSERT OR REPLACE]"
variant of the [INSERT] command.  
This alias is provided for compatibility other SQL database engines.  See the 
[INSERT] command documentation for additional information.</p>  

<tcl>
###############################################################################
Section {WITH clause} with {{common table expressions} WITH}

RecursiveBubbleDiagram with-clause
</tcl>

<p>Common Table Expressions or CTEs act like temporary [views] that exist
only for the duration of a single SQL statement.  There are two kinds of
common table expressions: "ordinary" and "recursive". Ordinary 
common table expressions are helpful for making
queries easier to understand by factoring
subqueries out of the main SQL statement.
Recursive common table expressions
provide the ability to do hierarchical or
recursive queries of trees and graphs, a capability
that is not otherwise available in the SQL language.

<p>All common table expressions (ordinary and recursive) are 
created by prepending a WITH clause in front of a [SELECT], [INSERT], [DELETE],
or [UPDATE] statement.  A single WITH clause can specify one or more
common table expressions, some of which are ordinary and some of which
are recursive.

<tcl>hd_fragment ordinarycte {ordinary common table expressions}</tcl>
<h3>Ordinary Common Table Expressions</h3>

<p>An ordinary common table expression works as if it were a [view] that
exists for the duration of a single statement.  Ordinary common table
expressions are useful for factoring out subqueries and making the overall
SQL statement easier to read and understand.

<p>A WITH clause can contain ordinary common table expressions even if
it includes the RECURSIVE keyword.  The use of RECURSIVE does not force
common table expressions to be recursive.

<tcl>
hd_fragment recursivecte {recursive common table expressions} \
{recursive common table expression} {recursive query}
</tcl>
<h3>Recursive Common Table Expressions</h3>

<p>A recursive common table expression can be used to write a query that
walks a tree or graph.  A recursive common table expression has the same
basic syntax as an ordinary common table expression, but with the following
additional features:

<ol>
<li> The "[select-stmt]"
     must be a [compound select] where the right-most [compound-operator] is
     either UNION or UNION ALL.
<li> The table named on the left-hand side of the AS keyword must appear
     exactly once in the FROM clause of the right-most SELECT statement
     of the compound select, and nowhere else.
<li> The right-most SELECT of the compound select must not make use of
     [Aggregate Functions|aggregate] or [window functions].
</ol>

<p>To put it another way, a recursive common table expression must
look like the following:

<tcl>RecursiveBubbleDiagram recursive-cte</tcl>

<p>Call the table named by the [cte-table-name] in a recursive
common table expression the "recursive table".
In the [recursive-cte] bubble diagram above, the recursive
table must appear exactly once in the FROM clause of the 
<yyterm>recursive-select</yyterm>
and must not appear anywhere else in either the
<yyterm>initial-select</yyterm> or the
<yyterm>recursive-select</yyterm>, including subqueries.
The <yyterm>initial-select</yyterm> may be
a [compound select], but it may not include an ORDER BY, LIMIT, or OFFSET.
The recursive-select must be a simple select, not a compound.  The
recursive-select is allowed to include an ORDER BY, LIMIT, and/or OFFSET.

<p>The basic algorithm for computing the content of the recursive table
is as follows:

<ol>
<li> Run the <yyterm>initial-select</yyterm> and add the results to a queue.
<li> While the queue is not empty:
<ol type="a">
<li> Extract a single row from the queue.
<li> Insert that single row into the recursive table
<li> Pretend that the single row just extracted is the only
     row in the recursive table and run the recursive-select,
     adding all results to the queue.
</ol>
</ol>

<p>The basic procedure above may modified by the following additional rules:

<ul>
<li><p>
  If a UNION operator connects the <yyterm>initial-select</yyterm> with the
  <yyterm>recursive-select</yyterm>, then only add rows to the queue if 
  no identical row has
  been previously added to the queue.  Repeated rows are discarded before being
  added to the queue even if the repeated rows have already been extracted
  from the queue by the recursion step.  If the operator is UNION ALL,
  then all rows generated by both the <yyterm>initial-select</yyterm> and the
  <yyterm>recursive-select</yyterm> are always added to the queue even if
  they are repeats.
  When determining if a row is repeated, NULL values compare
  equal to one another and not equal to any other value.
<li><p>
  The LIMIT clause, if present, determines the maximum number of rows that
  will ever be added to the recursive table in step 2b.
  Once the limit is reached, the recursion stops.
  A limit of zero means that no rows are ever added to the
  recursive table, and a negative limit means an unlimited number of rows
  may be added to the recursive table.
<li><p>
  The OFFSET clause, if it is present and has a positive value N, prevents the
  first N rows from being added to the recursive table.
  The first N rows are still processed
  by the <yyterm>recursive-select</yyterm> &mdash; they
  just are not added to the recursive table.  Rows are not counted toward
  fulfilling the LIMIT until all OFFSET rows have been skipped.
<li><p>
  If an ORDER BY clause is present, it determines the order in which rows
  are extracted from the queue in step 2a.  If there is no ORDER BY clause,
  then the order in which rows are extracted is undefined.  (In the current
  implementation, the queue becomes a FIFO if the ORDER BY clause is omitted,
  but applications should not depend on that fact since it might change.)
</ul>

<tcl>hd_fragment rcex1</tcl>
<h4>Recursive Query Examples</h4>

<p>The following query returns all integers between 1 and 1000000:

<blockquote><pre>
WITH RECURSIVE
  cnt(x) AS (VALUES(1) UNION ALL SELECT x+1 FROM cnt WHERE x<1000000)
SELECT x FROM cnt;
</pre></blockquote>

<p>Consider how this query works.  The initial-select
runs first and returns a single row
with a single column "1".  This one row is added to the queue.  In
step 2a, that one row is extracted from the queue and added to "cnt".
Then the recursive-select is run in accordance with step 2c generating
a single new row with value "2" to add to the queue.  The queue still
has one row, so step 2 repeats.  The "2" row is extracted and added to the
recursive table by steps 2a and 2b.  Then the row containing 2 is used 
as if it were the complete content of the recursive table and the 
recursive-select is run again, resulting in a row with value "3" being added
to the queue.  This repeats 999999 times until finally at step 2a the
only value on the queue is a row containing 1000000.  That row is
extracted and added to the recursive table.  But this time, the
WHERE clause causes the recursive-select to return no rows, so the
queue remains empty and the recursion stops.

<p><b>Optimization note:</b>
In the discussion above, statements like "insert the row into
the recursive table" should be understood conceptually, not literally.
It sounds as if SQLite is accumulating a huge table
containing one million rows, then going back and scanning that table
from top to bottom to generate the result.  What really happens
is that the query optimizer sees that values in the
"cnt" recursive table are only used once.  So as each row is added to
the recursive table, that row is immediately returned as a result of the main
SELECT statement and then discarded.  SQLite does <em>not</em> accumulate
a temporary table containing a million rows.  Very little memory is
needed to run the above example.  However, if the example had used
UNION instead of UNION ALL, then SQLite would have had to keep around
all previously generated content in order to check for duplicates.
For this reason, programmers should strive to use UNION ALL instead
of UNION when feasible.

<p>Here is a variation on the previous example:

<blockquote><pre>
WITH RECURSIVE
  cnt(x) AS (
     SELECT 1
     UNION ALL
     SELECT x+1 FROM cnt
      LIMIT 1000000
  )
SELECT x FROM cnt;
</pre></blockquote>

<p>There are two differences in this variation.  The initial-select is
"SELECT 1" instead of "VALUES(1)".  But those are just different
syntaxes for saying exactly the same thing.  The other change is that the
recursion is stopped by a LIMIT rather than a WHERE clause.  The use of
LIMIT means that when the one-millionth row is added to the "cnt" table
(and returned by the main SELECT, thanks to the query optimizer)
then the recursion stops immediately regardless of how many rows might be
left in the queue.  On more complex queries, it can sometimes be
difficult to ensure that the WHERE clause will eventually cause the
queue to drain and the recursion to terminate.  But the LIMIT clause will
always stop the recursion.  So it is good practice to always include a
LIMIT clause as a safety if an upper bound on the size of the recursion 
is known.

<tcl>hd_fragment rcex2</tcl>
<h4>Hierarchical Query Examples</h4>

<p>Consider a table that describes the members of an organization as
well as the chain-of-command within that organization:

<blockquote><pre>
CREATE TABLE org(
  name TEXT PRIMARY KEY,
  boss TEXT REFERENCES org,
  height INT,
  -- other content omitted
);
</pre></blockquote>

<p>Every member in the organization has a name, and most members have
a single boss.  (The head of the whole organization has a NULL
"boss" field.) The rows of the "org" table form a tree.

<p>Here is a query that computes the average height over everyone
in Alice's organization, including Alice:

<blockquote><pre>
WITH RECURSIVE
  works_for_alice(n) AS (
    VALUES('Alice')
    UNION
    SELECT name FROM org, works_for_alice
     WHERE org.boss=works_for_alice.n
  )
SELECT avg(height) FROM org
 WHERE org.name IN works_for_alice;
</pre></blockquote>

<p>The next example uses two 
common table expressions in a single WITH clause.  
The following table records a family tree:

<blockquote><pre>
CREATE TABLE family(
  name TEXT PRIMARY KEY,
  mom TEXT REFERENCES family,
  dad TEXT REFERENCES family,
  born DATETIME,
  died DATETIME, -- NULL if still alive
  -- other content
);
</pre></blockquote>

<p>The "family" table is similar to the earlier "org" table except that 
now there are two parents to each member.
We want to know all living ancestors of Alice, from oldest to youngest.
An ordinary common table expression, "parent_of", is defined first.  That
ordinary CTE is a view that can be used to find all parents of any
individual.  That ordinary CTE is then used in the "ancestor_of_alice"
recursive CTE.  The recursive CTE is then used in the final query:

<blockquote><pre>
WITH RECURSIVE
  parent_of(name, parent) AS
    (SELECT name, mom FROM family UNION SELECT name, dad FROM family),
  ancestor_of_alice(name) AS
    (SELECT parent FROM parent_of WHERE name='Alice'
     UNION ALL
     SELECT parent FROM parent_of JOIN ancestor_of_alice USING(name))
SELECT family.name FROM ancestor_of_alice, family
 WHERE ancestor_of_alice.name=family.name
   AND died IS NULL
 ORDER BY born;
</pre></blockquote>

<tcl>hd_fragment rcex3</tcl>
<h4>Queries Against A Graph</h4>

<p>A version control system (VCS) will typically store the evolving
versions of a project as a directed acyclic graph (DAG).  Call each
version of the project a "checkin".  A single
checkin can have zero or more parents.  Most checkins (except the
first) have a single parent, but in the case of a merge, a checkin
might have two or three or more parents.  A schema to keep track of
checkins and the order in which they occur might look something like
this:

<blockquote><pre>
CREATE TABLE checkin(
  id INTEGER PRIMARY KEY,
  mtime INTEGER -- timestamp when this checkin occurred
);
CREATE TABLE derivedfrom(
  xfrom INTEGER NOT NULL REFERENCES checkin, -- parent checkin
  xto INTEGER NOT NULL REFERENCES checkin,   -- derived checkin
  PRIMARY KEY(xfrom,xto)
);
CREATE INDEX derivedfrom_back ON derivedfrom(xto,xfrom);
</pre></blockquote>

<p>This graph is acyclic.  And we assume that the mtime of every
child checkin is no less than the mtime of all its parents.  But
unlike the earlier examples, this graph might have multiple paths of
differing lengths between any two checkins.

<p>We want to know the twenty most recent ancestors in time (out of
the thousands and thousands of ancestors in the whole DAG) for
checkin "@BASELINE".  (A query similar to this is used
by the <a href="http://www.fossil-scm.org/">Fossil</a> VCS to
show the N most recent ancestors of a check.  For example:
<a href="http://www.sqlite.org/src/timeline?p=trunk&n=30">http://www.sqlite.org/src/timeline?p=trunk&n=30</a>.)

<blockquote><pre>
WITH RECURSIVE
  ancestor(id,mtime) AS (
    SELECT id, mtime FROM checkin WHERE id=@BASELINE
    UNION
    SELECT derivedfrom.xfrom, checkin.mtime
      FROM ancestor, derivedfrom, checkin
     WHERE ancestor.id=derivedfrom.xto
       AND checkin.id=derivedfrom.xfrom
     ORDER BY checkin.mtime DESC
     LIMIT 20
  )
SELECT * FROM checkin JOIN ancestor USING(id);
</pre></blockquote>

<p>
The "ORDER BY checkin.mtime DESC" term in the recursive-select makes
the query run much faster by preventing it from following
branches that merge checkins
from long ago.  The ORDER BY forces the recursive-select to focus
on the most recent checkins, the ones we want.  Without the ORDER BY
on the recursive-select, one would be forced to compute the complete set of
thousands of ancestors, sort them all by mtime, then take the top twenty.
The ORDER BY essentially sets up a priority queue that
forces the recursive query to look at the most recent ancestors first,
allowing the use of a LIMIT clause to restrict the scope of the
query to just the checkins of interest.

<tcl>hd_fragment withorderby</tcl>
<h4>Controlling Depth-First Versus Breadth-First Search Of a Tree
Using ORDER BY</h4>

<p>An ORDER BY clause on the recursive-select can be used to control
whether the search of a tree is depth-first or breadth-first.  To
illustrate, we will use a variation on the "org" table from an example
above, without the "height" column, and with some real data inserted:

<blockquote><pre>
CREATE TABLE org(
  name TEXT PRIMARY KEY,
  boss TEXT REFERENCES org
) WITHOUT ROWID;
INSERT INTO org VALUES('Alice',NULL);
INSERT INTO org VALUES('Bob','Alice');
INSERT INTO org VALUES('Cindy','Alice');
INSERT INTO org VALUES('Dave','Bob');
INSERT INTO org VALUES('Emma','Bob');
INSERT INTO org VALUES('Fred','Cindy');
INSERT INTO org VALUES('Gail','Cindy');
</pre></blockquote>

<p>Here is a query to show the tree structure in a breadth-first pattern:

<blockquote><pre>
WITH RECURSIVE
  under_alice(name,level) AS (
    VALUES('Alice',0)
    UNION ALL
    SELECT org.name, under_alice.level+1
      FROM org JOIN under_alice ON org.boss=under_alice.name
     ORDER BY 2
  )
SELECT substr('..........',1,level*3) || name FROM under_alice;
</pre></blockquote>

<p>The "ORDER BY 2" (which means the same as "ORDER BY under_alice.level+1")
causes higher levels in the organization chart (with smaller "level" values)
to be processed first, resulting in a breadth-first search.  The output is:

<blockquote><pre>
Alice
...Bob
...Cindy
......Dave
......Emma
......Fred
......Gail
</pre></blockquote>

<p>But if we change the ORDER BY clause to add the "DESC" modifier, that will
cause lower levels in the organization (with larger "level" values) to be
processed first by the recursive-select, resulting in a depth-first search:

<blockquote><pre>
WITH RECURSIVE
  under_alice(name,level) AS (
    VALUES('Alice',0)
    UNION ALL
    SELECT org.name, under_alice.level+1
      FROM org JOIN under_alice ON org.boss=under_alice.name
     ORDER BY 2 <b>DESC</b>
  )
SELECT substr('..........',1,level*3) || name FROM under_alice;
</pre></blockquote>

<p>The output of this revised query is:

<blockquote><pre>
Alice
...Bob
......Dave
......Emma
...Cindy
......Fred
......Gail
</pre></blockquote>

<p>When the ORDER BY clause is omitted from the recursive-select, the
queue behaves as a FIFO, which results in a breadth-first search.


<tcl>hd_fragment mandelbrot {compute the Mandelbrot set}</tcl>
<h4>Outlandish Recursive Query Examples</h4>

<p>The following query computes an approximation of the Mandelbrot Set
and outputs the result as ASCII-art:

<blockquote><pre>
WITH RECURSIVE
  xaxis(x) AS (VALUES(-2.0) UNION ALL SELECT x+0.05 FROM xaxis WHERE x&lt;1.2),
  yaxis(y) AS (VALUES(-1.0) UNION ALL SELECT y+0.1 FROM yaxis WHERE y&lt;1.0),
  m(iter, cx, cy, x, y) AS (
    SELECT 0, x, y, 0.0, 0.0 FROM xaxis, yaxis
    UNION ALL
    SELECT iter+1, cx, cy, x*x-y*y + cx, 2.0*x*y + cy FROM m 
     WHERE (x*x + y*y) &lt; 4.0 AND iter&lt;28
  ),
  m2(iter, cx, cy) AS (
    SELECT max(iter), cx, cy FROM m GROUP BY cx, cy
  ),
  a(t) AS (
    SELECT group_concat( substr(' .+*#', 1+min(iter/7,4), 1), '') 
    FROM m2 GROUP BY cy
  )
SELECT group_concat(rtrim(t),x'0a') FROM a;
</pre></blockquote>

<p>In this query, the "xaxis" and "yaxis" CTEs define the grid of points for
which the Mandelbrot Set will be approximated.  Each row in the
"m(iter,cx,cy,x,y)" CTE means that after "iter" iterations, the Mandelbrot
iteration starting at cx,cy has reached point x,y.  The number of iterations
in this example is limited to 28 (which severely limits the resolution of
the computation, but is sufficient for low-resolution ASCII-art output).
The "m2(iter,cx,cy)" CTE holds the maximum number of iterations reached when
starting at point cx,cy.
Finally, each row in the "a(t)" CTE holds a string 
which is a single line of the output ASCII-art.
The SELECT statement at the end just queries the "a" CTE to
retrieve all lines of ASCII-art, one by one.

<p>Running the query above in an SQLite [command-line shell] results
in the following output:

<blockquote><pre>
                                    ....#
                                   ..#*..
                                 ..+####+.
                            .......+####....   +
                           ..##+*##########+.++++
                          .+.##################+.
              .............+###################+.+
              ..++..#.....*#####################+.
             ...+#######++#######################.
          ....+*################################.
 #############################################...
          ....+*################################.
             ...+#######++#######################.
              ..++..#.....*#####################+.
              .............+###################+.+
                          .+.##################+.
                           ..##+*##########+.++++
                            .......+####....   +
                                 ..+####+.
                                   ..#*..
                                    ....#
                                    +.
</pre></blockquote>

<tcl>hd_fragment sudoku {solve Sudoku puzzles} </tcl>
<p>This next query solves a Sudoku puzzle.  The state of the puzzle is
defined by an 81-character string formed by reading entries from the
puzzle box row by row from left to right and then from top to bottom.
Blank squares in the puzzle are denoted by a "." character.  
Thus the input string:

<blockquote>
53..7....6..195....98....6.8...6...34..8.3..17...2...6.6....28....419..5....8..79
</blockquote>

<p>Corresponds to a puzzle like this:

<blockquote>
<table border="1" cellpadding="5">
<tr><td>5<td>3<td> <td> <td>7<td> <td> <td> <td>
<tr><td>6<td> <td> <td>1<td>9<td>5<td> <td> <td>
<tr><td> <td>9<td>8<td> <td> <td> <td> <td>6<td>
<tr><td>8<td> <td> <td> <td>6<td> <td> <td> <td>3
<tr><td>4<td> <td> <td>8<td> <td>3<td> <td> <td>1
<tr><td>7<td> <td> <td> <td>2<td> <td> <td> <td>6
<tr><td> <td>6<td> <td> <td> <td> <td>2<td>8<td>
<tr><td> <td> <td> <td>4<td>1<td>9<td> <td> <td>5
<tr><td> <td> <td> <td> <td>8<td> <td> <td>7<td>9
</table>
</blockquote>

<p>This is the query that solves the puzzle:

<blockquote><pre>
WITH RECURSIVE
  input(sud) AS (
    VALUES('53..7....6..195....98....6.8...6...34..8.3..17...2...6.6....28....419..5....8..79')
  ),
  digits(z, lp) AS (
    VALUES('1', 1)
    UNION ALL SELECT
    CAST(lp+1 AS TEXT), lp+1 FROM digits WHERE lp&lt;9
  ),
  x(s, ind) AS (
    SELECT sud, instr(sud, '.') FROM input
    UNION ALL
    SELECT
      substr(s, 1, ind-1) || z || substr(s, ind+1),
      instr( substr(s, 1, ind-1) || z || substr(s, ind+1), '.' )
     FROM x, digits AS z
    WHERE ind>0
      AND NOT EXISTS (
            SELECT 1
              FROM digits AS lp
             WHERE z.z = substr(s, ((ind-1)/9)*9 + lp, 1)
                OR z.z = substr(s, ((ind-1)%9) + (lp-1)*9 + 1, 1)
                OR z.z = substr(s, (((ind-1)/3) % 3) * 3
                        + ((ind-1)/27) * 27 + lp
                        + ((lp-1) / 3) * 6, 1)
         )
  )
SELECT s FROM x WHERE ind=0;
</pre></blockquote>

<p>The "input" CTE defines the input puzzle.
The "digits" CTE defines a table that holds all digits between 1 and 9.
The work of solving the puzzle is undertaken by the "x" CTE.
An entry in x(s,ind) means that the 81-character string "s" is a valid
sudoku puzzle (it has no conflicts) and that the first unknown character
is at position "ind", or ind==0 if all character positions are filled in.
The goal, then, is to compute entries for "x" with an "ind" of 0.

<p>The solver works by adding new entries to the "x" recursive table.
Given prior entries, the recursive-select tries to fill in a single new
position with all values between 1 and 9 that actually work in that
position.  The complicated "NOT EXISTS" subquery is the magic that
figures out whether or not each candidate "s" string is a valid
sudoku puzzle or not.

<p>The final answer is found by looking for a string with ind==0.
If the original sudoku problem did not have a unique solution, then
the query will return all possible solutions.  If the original problem
was unsolvable, then no rows will be returned.  In this case, the unique
answer is:

<blockquote>
534678912672195348198342567859761423426853791713924856961537284287419635345286179
</blockquote>

<p>The solution was computed in less than 300 milliseconds on a modern
workstation.

<h3>Limitations And Caveats</h3>

<ul>
<li><p>
The WITH clause cannot be used within a [CREATE TRIGGER].
<li><p>
The WITH clause must appear at the beginning of a top-level [SELECT] statement
or at the beginning of a subquery.  The WITH clause cannot be prepended to
the second or subsequent SELECT statement of a [compound select].
<li><p>
The SQL:1999 spec requires that the RECURSIVE keyword follow WITH in any
WITH clause that includes a recursive common table expression.  However, for
compatibility with SqlServer and Oracle, SQLite does not enforce this rule.
</ul>

<tcl>
###############################################################################
Section SELECT select {SELECT query}

RecursiveBubbleDiagram select-stmt
</tcl>


<p>The SELECT statement is used to query the database.  The
result of a SELECT is zero or more rows of data where each row
has a fixed number of columns.  A SELECT statement does not make
any changes to the database.

<p>The "[select-stmt]" syntax diagram above attempts to show as much of the
SELECT statement syntax as possible in a single diagram, because some readers
find that helpful.  The following "[factored-select-stmt]" is an alternative
syntax diagrams that expresses the same syntax but tries to break the syntax 
down into smaller chunks.

<tcl>
RecursiveBubbleDiagram --initially-hidden factored-select-stmt select-core
</tcl>

<p>Note that there are paths through the syntax diagrams that
are not allowed in practice.  Some examples:
<ul>
<li>A [VALUES] clause can be the first element in a [compound SELECT]
    that uses a [WITH] clause, but a [simple SELECT] that consists of
    just a [VALUES] clause cannot be preceded by a [WITH] clause.
<li>The [WITH] clause must occur on the first SELECT of a [compound SELECT].
    It cannot follow a [compound-operator].
</ul>
These and other similar syntax restrictions are described in the text.

<p>The SELECT statement is the most complicated command in the SQL language.
To make the description easier to follow, some of the passages below describe
the way the data returned by a SELECT statement is determined as a series of
steps. It is important to keep in mind that this is purely illustrative -
in practice neither SQLite nor any other SQL engine is required to follow 
this or any other specific process.

<tcl>hd_fragment simpleselect {simple SELECT}</tcl>
<h3>Simple Select Processing</h3>

<p>The core of a SELECT statement is a "simple SELECT" shown by the
[select-core] and [simple-select-stmt] syntax diagrams below.  
In practice, most SELECT statements are simple SELECT statements.

<tcl>
RecursiveBubbleDiagram simple-select-stmt select-core
</tcl>

<p>Generating the results of a simple SELECT
statement is presented as a four step process in the description below:

<ol>
  <li> <p>[FROM clause] processing: The input data for the simple SELECT is
       determined. The input data is either implicitly a single row with 0
       columns (if there is no FROM clause) or is determined by the FROM
       clause.
  <li> <p>[WHERE clause] processing: The input data is filtered using the WHERE
       clause expression.  
  <li> <p>[GROUP BY|GROUP BY, HAVING and result-column expression] processing: 
       The set of result rows is computed by aggregating the data according to
       any GROUP BY clause and calculating the result-set expressions for the
       rows of the filtered input dataset.  
  <li> <p>[DISTINCT|DISTINCT/ALL keyword] processing: If the query is a "SELECT
       DISTINCT" query, duplicate rows are removed from the set of result rows.
</ol>

<p>There are two types of simple SELECT statement - aggregate and 
non-aggregate queries. ^A simple SELECT statement is an aggregate query if
it contains either a GROUP BY clause or one or more aggregate functions
in the result-set. ^Otherwise, if a simple SELECT contains no aggregate
functions or a GROUP BY clause, it is a non-aggregate query.

<p><b>1. Determination of input data (FROM clause processing).</b>
<tcl>hd_fragment fromclause</tcl>
<tcl>hd_keywords {FROM clause}</tcl>

<p>The input data used by a simple SELECT query is a set of <i>N</i> rows 
each <i>M</i> columns wide.

<p>^(If the FROM clause is omitted from a simple SELECT statement, then the 
input data is implicitly a single row zero columns wide)^ (i.e. <i>N</i>=1 and
<i>M</i>=0).

<p>If a FROM clause is specified, the data on which a simple SELECT query
operates comes from the one or more tables or subqueries (SELECT statements
in parentheses) specified following the FROM keyword. ^A subquery specified
in the <yyterm>table-or-subquery</yyterm> following the FROM clause in a 
simple SELECT statement is
handled as if it was a table containing the data returned by executing the
subquery statement. ^Each column of the subquery has the
[collation|collation sequence] and [affinity] of the corresponding expression
in the subquery statement.

<p>^If there is only a single table or subquery in the FROM
clause, then the input data used by the SELECT statement is the contents of the
named table. ^If there is more than one table or subquery in FROM clause
then the contents of all tables and/or subqueries
are joined into a single dataset for the simple SELECT statement to operate on.
Exactly how the data is combined depends on the specific [join-operator] and
[join-constraint] used to connect the tables or subqueries together.

<p>All joins in SQLite are based on the cartesian product of the left and
right-hand datasets. ^The columns of the cartesian product dataset are, in 
order, all the columns of the left-hand dataset followed by all the columns
of the right-hand dataset. ^There is a row in the cartesian product dataset
formed by combining each unique combination of a row from the left-hand 
and right-hand datasets. ^(In other words, if the left-hand dataset consists of
<i>N<sub><small>left</small></sub></i> rows of 
<i>M<sub><small>left</small></sub></i> columns, and the right-hand dataset of
<i>N<sub><small>right</small></sub></i> rows of
<i>M<sub><small>right</small></sub></i> columns, then the cartesian product is a
dataset of 
<i>N<sub><small>left</small></sub>&times;N<sub><small>right</small></sub></i>
rows, each containing 
<i>M<sub><small>left</small></sub>+M<sub><small>right</small></sub></i> columns.)^

<p>^If the join-operator is "CROSS JOIN", "INNER JOIN", "JOIN" or a comma
(",") and there is no ON or USING clause, then the result of the join is
simply the cartesian product of the left and right-hand datasets. 
If join-operator does have ON or USING clauses, those are handled according to
the following bullet points:

<ul>
  <li> <p>^(If there is an ON clause then the ON expression is
       evaluated for each row of the cartesian product as a 
       [boolean expression]. Only rows for which the expression evaluates to 
       true are included from the dataset.)^

  <li> <p>^If there is a USING clause
       then each of the column names specified must exist in the datasets to 
       both the left and right of the join-operator. ^(For each pair of named
       columns, the expression "lhs.X = rhs.X" is evaluated for each row of
       the cartesian product as a [boolean expression]. Only rows for which
       all such expressions evaluates to true are included from the
       result set.)^ ^When comparing values as a result of a USING clause, the
       normal rules for handling affinities, collation sequences and NULL
       values in comparisons apply. ^The column from the dataset on the
       left-hand side of the join-operator is considered to be on the left-hand
       side of the comparison operator (=) for the purposes of collation 
       sequence and affinity precedence.

       <p>^For each pair of columns identified by a USING clause, the column
       from the right-hand dataset is omitted from the joined dataset. ^This 
       is the only difference between a USING clause and its equivalent ON
       constraint.

  <li> <p>^(If the NATURAL keyword is in the join-operator then an
       implicit USING clause is added to the join-constraints. The implicit
       USING clause contains each of the column names that appear in both
       the left and right-hand input datasets.)^ ^If the left and right-hand
       input datasets feature no common column names, then the NATURAL keyword
       has no effect on the results of the join. ^A USING or ON clause may
       not be added to a join that specifies the NATURAL keyword.

  <li> <p>^(If the join-operator is a "LEFT JOIN" or "LEFT OUTER JOIN", then
       after
       the ON or USING filtering clauses have been applied, an extra row is 
       added to the output for each row in the original left-hand input 
       dataset that corresponds to no rows at all in the composite
       dataset (if any).)^ ^The added rows contain NULL values in the columns
       that would normally contain values copied from the right-hand input
       dataset.  
</ul>

<p>^(When more than two tables are joined together as part of a FROM clause,
the join operations are processed in order from left to right. In other 
words, the FROM clause (A join-op-1 B join-op-2 C) is computed as 
((A join-op-1 B) join-op-2 C).)^

<tcl>hd_fragment crossjoin {treats the CROSS JOIN operator specially}</tcl>
<p><b>Side note: Special handling of CROSS JOIN.</b>
^There is no difference between the "INNER JOIN", "JOIN" and "," join
operators. They are completely interchangeable in SQLite.
^(The "CROSS JOIN" join operator produces the same result as the 
"INNER JOIN", "JOIN" and "," operators)^, but is 
[CROSS JOIN|handled differently by the query optimizer]
in that it prevents the query optimizer from reordering
the tables in the join.  An application programmer can use the CROSS JOIN 
operator to directly influence the algorithm that is chosen to implement
the SELECT statement.  Avoid using CROSS JOIN except in specific situations 
where manual control of the query optimizer is desired.  Avoid using
CROSS JOIN early in the development of an application as doing so is
a <a href="http://c2.com/cgi/wiki?PrematureOptimization">premature
optimization</a>.  The special handling of CROSS JOIN is an SQLite-specific
feature and is not a part of standard SQL.
       

<tcl>hd_fragment whereclause</tcl>
<tcl>hd_keywords {WHERE clause}</tcl>
<p><b>2. WHERE clause filtering.</b>

<p>^(If a WHERE clause is specified, the WHERE expression is evaluated for 
each row in the input data as a [boolean expression]. Only rows for which the
WHERE clause expression evaluates to true are included from the dataset before
continuing.)^  Rows are excluded from the result if the WHERE clause
evaluates to either false or NULL.

<p>For a JOIN or INNER JOIN or CROSS JOIN, there is no difference between 
a constraint expression in the WHERE clause and one in the ON clause.  However,
for a LEFT JOIN or LEFT OUTER JOIN, the difference is very important.  
In a LEFT JOIN,
the extra NULL row for the right-hand table is added after ON clause processing
but before WHERE clause processing.  A constraint of the form "left.x=right.y"
in an ON clause will therefore allow through the added all-NULL rows of the
right table.  But if that same constraint is in the WHERE clause a NULL in
"right.y" will prevent the expression "left.x=right.y" from being true, and
thus exclude that row from the output.

<p><b>3. Generation of the set of result rows.</b>
<tcl>hd_fragment resultset</tcl>
<tcl>hd_keywords {result-set expressions} {GROUP BY}</tcl>

<p>Once the input data from the FROM clause has been filtered by the
WHERE clause expression (if any), the set of result rows for the simple 
SELECT are calculated. Exactly how this is done depends on whether the simple 
SELECT is an aggregate or non-aggregate query, and whether or not a GROUP
BY clause was specified.

<p> The list of expressions between the SELECT and FROM keywords is known as
the result expression list.  ^If a result expression is the special expression
"*" then all columns in the input data are substituted for that one expression.
^(If the expression is the alias of a table or subquery in the FROM clause
followed by ".*" then all columns from the named table or subquery are
substituted for the single expression.)^ ^(It is an error to use a "*" or
"alias.*" expression in any context other than a result expression list.)^
^(It is also an error to use a "*" or "alias.*" expression in a simple SELECT
query that does not have a FROM clause.)^

<p> ^(The number of columns in the rows returned by a simple SELECT statement
is equal to the number of expressions in the result expression list after
substitution of * and alias.* expressions.)^ Each result row is calculated by
evaluating the expressions in the result expression list with respect to a
single row of input data or, for aggregate queries, with respect to a group
of rows.

<ul>
  <li><p>^If the SELECT statement is <b>a non-aggregate query</b>, then 
    each expression in the result expression list is evaluated for each row in
    the dataset filtered by the WHERE clause.

  <li><p>^If the SELECT statement is <b>an aggregate query without a GROUP
    BY</b> clause, then each aggregate expression in the result-set is 
    evaluated once across the entire dataset. ^Each non-aggregate expression
    in the result-set is evaluated once for an arbitrarily selected row of
    the dataset. ^The same arbitrarily selected row is used for each
    non-aggregate expression. ^Or, if the dataset contains zero rows, then 
    each non-aggregate expression is evaluated against a row consisting
    entirely of NULL values.

   <p>^The single row of result-set data created by evaluating the aggregate
    and non-aggregate expressions in the result-set forms the result of an
    aggregate query without a GROUP BY clause. ^An aggregate query without a
    GROUP BY clause always returns exactly one row of data, even if there are
    zero rows of input data.

  <li><p>^(If the SELECT statement is <b>an aggregate query with a GROUP
    BY</b> clause, then each of the expressions specified as part of the
    GROUP BY clause is evaluated for each row of the dataset. Each row
    is then assigned to a "group" based on the results; rows for which
    the results of evaluating the GROUP BY expressions are the same get
    assigned to the same group.)^ ^For the purposes of grouping rows, NULL 
    values are considered equal. ^The usual rules for [collation|selecting a
    collation sequence] with which to compare text values apply when evaluating
    expressions in a GROUP BY clause.  ^The expressions in the GROUP BY clause
    do <em>not</em> have to be expressions that appear in the result. ^The
    expressions in a GROUP BY clause may not be aggregate expressions.

    <p>^(If a HAVING clause is specified, it is evaluated once for each group 
    of rows as a [boolean expression]. If the result of evaluating the
    HAVING clause is false, the group is discarded.)^ ^If the HAVING clause is
    an aggregate expression, it is evaluated across all rows in the group. ^If
    a HAVING clause is a non-aggregate expression, it is evaluated with respect
    to an arbitrarily selected row from the group.  ^The HAVING expression may
    refer to values, even aggregate functions, that are not in the result.</p>

    <p>^Each expression in the result-set is then evaluated once for each
    group of rows. ^If the expression is an aggregate expression, it is 
    evaluated across all rows in the group. ^Otherwise, it is evaluated against
    a single arbitrarily chosen row from within the group. ^If there is more
    than one non-aggregate expression in the result-set, then all such
    expressions are evaluated for the same row.

    <p>^Each group of input dataset rows contributes a single row to the 
    set of result rows. ^Subject to filtering associated with the DISTINCT
    keyword, the number of rows returned by an aggregate query with a GROUP
    BY clause is the same as the number of groups of rows produced by applying
    the GROUP BY and HAVING clauses to the filtered input dataset.
</ul>

<tcl>hd_fragment bareagg {bare aggregate terms}</tcl>
<p><b>Side note: Bare columns in an aggregate queries.</b>
The usual case is that all column names in an aggregate query are either
arguments to [aggfunc|aggregate functions] or else appear in the GROUP BY clause.
A result column which contains a column name that is not within an
aggregate function and that does not appear in the GROUP BY clause (if
one exists) is called a "bare" column.
Example:
<blockquote><pre>
SELECT a, b, sum(c) FROM tab1 GROUP BY a;
</pre></blockquote>
<p>In the query above, the "a" column is part of the GROUP BY clause
and so each row of the output contains one of the distinct values for
"a".  The "c" column is contained within the [sum()] aggregate function
and so that output column is the sum of all "c" values in rows that
have the same value for "a".  But what is the result of the bare 
column "b"? The answer is that the "b" result will be the
value for "b" in one of the input rows that form the aggregate.
The problem is that you usually do not know which input row is used
to compute "b", and so in many cases the value for "b" is undefined.
</p>
<p>
Special processing occurs when the aggregate function is 
either [maxAggFunc|min()] or [minAggFunc|max()].  Example:
<blockquote><pre>
SELECT a, b, max(c) FROM tab1 GROUP BY a;
</pre></blockquote>
When the [maxAggFunc|min()] or [minAggFunc|max()] aggregate functions 
are used in an aggregate query, all bare columns in the result set 
take values from the input
row which also contains the minimum or maximum.  So in the query
above, the value of the "b" column in the output will be the value of
the "b" column in the input row that has the largest "c" value.
There is still an ambiguity if two or more of the input rows have the
same minimum or maximum value or if the query contains more than one
min() and/or max() aggregate function. Only the built-in [minAggFunc|min()] 
and [maxAggFunc|max()] functions work this way.
</p>

<p><b>4. Removal of duplicate rows (DISTINCT processing).</b>
<tcl>hd_fragment distinct</tcl>
<tcl>hd_keywords {DISTINCT}</tcl>

<p>^One of the ALL or DISTINCT keywords may follow the SELECT keyword in a 
simple SELECT statement. ^If the simple SELECT is a SELECT ALL, then the
entire set of result rows are returned by the SELECT. ^If neither ALL or
DISTINCT are present, then the behavior is as if ALL were specified. 
^If the simple SELECT is a SELECT DISTINCT, then duplicate rows are removed
from the set of result rows before it is returned. ^For the purposes of
detecting duplicate rows, two NULL values are considered to be equal. ^The
[how collation is determined|usual rules] apply for selecting a collation
sequence to compare text values.

<tcl>
hd_fragment compound {compound select} {compound query} {compound SELECT} {compound SELECTs}
</tcl>
<h3>Compound Select Statements</h3>

<p>Two or more [simple SELECT] statements may be connected together to form
a compound SELECT using the UNION, UNION ALL, INTERSECT or EXCEPT operator,
as shown by the following diagram:

<tcl>
RecursiveBubbleDiagram compound-select-stmt
</tcl>

<p>^In a compound SELECT, all the constituent SELECTs must return the same 
number of result columns. ^As the components of a compound SELECT must
be simple SELECT statements, they may not contain [ORDER BY] or [LIMIT] clauses.
^[ORDER BY] and [LIMIT] clauses may only occur at the end of the entire compound
SELECT, and then only if the final element of the compound is not a [VALUES] clause.

<p>^A compound SELECT created using UNION ALL operator returns all the rows
from the SELECT to the left of the UNION ALL operator, and all the rows
from the SELECT to the right of it. ^The UNION operator works the same way as
UNION ALL, except that duplicate rows are removed from the final result set.
^The INTERSECT operator returns the intersection of the results of the left and
right SELECTs.  ^The EXCEPT operator returns the subset of rows returned by the
left SELECT that are not also returned by the right-hand SELECT. ^Duplicate
rows are removed from the results of INTERSECT and EXCEPT operators before the
result set is returned.

<p>^For the purposes of determining duplicate rows for the results of compound
SELECT operators, NULL values are considered equal to other NULL values and
distinct from all non-NULL values. ^The collation sequence used to compare 
two text values is determined as if the columns of the left and right-hand
SELECT statements were the left and right-hand operands of the equals (=)
operator, except that greater precedence is not assigned to a collation 
sequence specified with the postfix COLLATE operator. ^No affinity
transformations are applied to any values when comparing rows as part of a
compound SELECT. 

<p>^(When three or more simple SELECTs are connected into a compound SELECT,
they group from left to right. In other words, if "A", "B" and "C" are all
simple SELECT statements, (A op B op C) is processed as ((A op B) op C).)^

</p>

<tcl>hd_fragment orderby {order by} {ORDER BY}</tcl>
<h3>The ORDER BY clause</h3>

<p>If a SELECT statement that returns more than one row does not have an
ORDER BY clause, the order in which the rows are returned is undefined.
Or, if a SELECT statement does have an ORDER BY clause, then the list of
expressions attached to the ORDER BY determine the order in which rows
are returned to the user.

<p>
^In a [compound SELECT] statement, only the last or right-most [simple SELECT]
may have an ORDER BY clause. ^That ORDER BY clause will apply across all elements of
the compound. ^If the right-most element of a [compound SELECT] is a [VALUES] clause,
then no ORDER BY clause is allowed on that statement.


<p>^Rows are first sorted based on the results of
evaluating the left-most expression in the ORDER BY list, then ties are broken
by evaluating the second left-most expression and so on. The order in which
two rows for which all ORDER BY expressions evaluate to equal values are
returned is undefined. ^Each ORDER BY expression may be optionally followed
by one of the keywords ASC (smaller values are returned first) or DESC (larger
values are returned first). ^If neither ASC or DESC are specified, rows
are sorted in ascending (smaller values first) order by default.

<tcl>hd_fragment nullslast {NULLS FIRST} {NULLS LAST}</tcl>
<p>SQLite considers NULL values to be smaller than any other values
for sorting purposes.  Hence, NULLs naturally appear at the beginning
of an ASC order-by and at the end of a DESC order-by.  This can be changed
using the "ASC NULLS LAST" or "DESC NULLS FIRST" syntax.

<p>Each ORDER BY expression is processed as follows:</p>

<ol>
<li><p>^If the ORDER BY expression is a constant integer K then the
expression is considered an alias for the K-th column of the result set
(columns are numbered from left to right starting with 1).

<li><p>^If the ORDER BY expression is an identifier that corresponds to
the alias of one of the output columns, then the expression is considered
an alias for that column.

<li><p>^Otherwise, if the ORDER BY expression is any other expression, it 
is evaluated and the returned value used to order the output rows. ^If
the SELECT statement is a simple SELECT, then an ORDER BY may contain any
arbitrary expressions. ^However, if the SELECT is a compound SELECT, then
ORDER BY expressions that are not aliases to output columns must be exactly
the same as an expression used as an output column.
</ol>

<p>^For the purposes of sorting rows, values are compared in the same way
as for [comparison expressions]. The collation sequence used to compare
two text values is determined as follows:

<ol>
  <li><p>^If the ORDER BY expression is assigned a collation sequence using
  the postfix [COLLATE operator], then the specified collation sequence is
  used.
  <li><p>^Otherwise, if the ORDER BY expression is an alias to an expression
  that has been assigned a collation sequence using the postfix 
  [COLLATE operator], then the collation sequence assigned to the aliased
  expression is used.
  <li><p>^Otherwise, if the ORDER BY expression is a column or an alias of
  an expression that is a column, then the default collation sequence for
  the column is used. 
  <li><p>^Otherwise, the [BINARY] collation sequence is used.
</ol>

<p>In a [compound SELECT] statement, all ORDER BY expressions are handled
as aliases for one of the result columns of the compound.
^(If an ORDER BY expression is not an integer alias, then SQLite searches
the left-most SELECT in the compound for a result column that matches either
the second or third rules above. If a match is found, the search stops and
the expression is handled as an alias for the result column that it has been
matched against. Otherwise, the next SELECT to the right is tried, and so on.)^
^If no matching expression can be found in the result columns of any
constituent SELECT, it is an error. ^Each term of the ORDER BY clause is
processed separately and may be matched against result columns from different
SELECT statements in the compound.</p>

<tcl>hd_fragment limitoffset {LIMIT} {OFFSET}</tcl>
<h3>The LIMIT clause</h3>

<p>^The LIMIT clause is used to place an upper bound on the number of rows
returned by the entire SELECT statement.

<p>^In a [compound SELECT], only the
last or right-most [simple SELECT] may contain a LIMIT clause.  
^In a [compound SELECT], 
the LIMIT clause applies to the entire compound, not just the final SELECT.
^If the right-most [simple SELECT] is a [VALUES clause] then no LIMIT clause
is allowed.

<p>^Any scalar expression may be used in the 
LIMIT clause, so long as it evaluates to an integer or a value that can be
losslessly converted to an integer. ^If the expression evaluates to a NULL 
value or any other value that cannot be losslessly converted to an integer, an
error is returned. ^If the LIMIT expression evaluates to a negative value,
then there is no upper bound on the number of rows returned. ^Otherwise, the
SELECT returns the first N rows of its result set only, where N is the value
that the LIMIT expression evaluates to. ^Or, if the SELECT statement would
return less than N rows without a LIMIT clause, then the entire result set is
returned. 

<p>^The expression attached to the optional OFFSET clause that may follow a
LIMIT clause must also evaluate to an integer, or a value that can be
losslessly converted to an integer. ^If an expression has an OFFSET clause,
then the first M rows are omitted from the result set returned by the SELECT
statement and the next N rows are returned, where M and N are the values that
the OFFSET and LIMIT clauses evaluate to, respectively. ^Or, if the SELECT
would return less than M+N rows if it did not have a LIMIT clause, then the
first M rows are skipped and the remaining rows (if any) are returned. ^If the
OFFSET clause evaluates to a negative value, the results are the same as if it
had evaluated to zero.

<p>^Instead of a separate OFFSET clause, the LIMIT clause may specify two
scalar expressions separated by a comma. ^In this case, the first expression
is used as the OFFSET expression and the second as the LIMIT expression.
This is counter-intuitive, as when using the OFFSET clause the second of
the two expressions is the OFFSET and the first the LIMIT. 
This reversal of the offset and limit is intentional
- it maximizes compatibility with other SQL database systems.
However, to avoid confusion, programmers are strongly encouraged to use
the form of the LIMIT clause that uses the "OFFSET" keyword and avoid
using a LIMIT clause with a comma-separated offset.

<tcl>hd_fragment values {VALUES clause} VALUES</tcl>
<h3>The VALUES clause</h3>

<p>^(The phrase "VALUES(<i>expr-list</i>)" means the same thing
as "SELECT <i>expr-list</i>".)^  ^(The phrase
"VALUES(<i>expr-list-1</i>),...,(<i>expr-list-N</i>)" means the same
thing as "SELECT <i>expr-list-1</i> UNION ALL ... UNION ALL
SELECT <i>expr-list-N</i>".)^  Both forms are the same, except that
the number of SELECT statements in a compound is limited by
[SQLITE_LIMIT_COMPOUND_SELECT] whereas the number of rows in a
VALUES clause has no arbitrary limit.

<p>There are some restrictions on the use of a VALUES clause that are
not shown on the syntax diagrams:

<ul>
<li><p>
^A VALUES clause cannot be followed by [ORDER BY].

<li><p>
^A VALUES clause cannot be followed by [LIMIT].
</ul>


<h3>The WITH Clause</h3>

<p>^SELECT statements may be optionally preceded by a single
[WITH clause] that defines one or more [common table expressions]
for use within the SELECT statement.

<tcl>hd_fragment tabfunc1 {table-valued functions in the FROM clause}</tcl>
<h3>Table-valued Functions In The FROM Clause</h3>

<p>A [virtual table] that contains [hidden columns] can be used like
a [table-valued function] in the FROM clause.  The arguments to the
table-valued function become constraints on the HIDDEN columns of
the virtual table.  Additional information can be found in the
[table-valued functions|virtual table documentation].

<tcl>
##############################################################################
Section UPDATE update {UPDATE *UPDATEs}

RecursiveBubbleDiagram update-stmt
</tcl>

<p>^An UPDATE statement is used to modify a subset of the values stored in 
zero or more rows of the database table identified by the 
[qualified-table-name] specified as part of the UPDATE statement.

<p>^If the UPDATE statement does not have a WHERE clause, all rows in the
table are modified by the UPDATE. ^Otherwise, the UPDATE affects only those
rows for which the WHERE clause
[boolean expression|boolean expression is true]. ^It is not an error if the
WHERE clause does not evaluate to true for any row in the table - this just
means that the UPDATE statement affects zero rows.

<p>The modifications made to each row affected by an UPDATE statement are
determined by the list of assignments following the SET keyword. Each
assignment specifies a <yyterm>column-name</yyterm> to the left of the 
equals sign and a scalar expression to the right. 
^For each affected row, the named columns
are set to the values found by evaluating the corresponding scalar 
expressions. ^If a single column-name appears more than once in the list of
assignment expressions, all but the rightmost occurrence is ignored. ^Columns
that do not appear in the list of assignments are left unmodified. ^The scalar
expressions may refer to columns of the row being updated. ^In this case all
scalar expressions are evaluated before any assignments are made.

<p>Beginning in SQLite [version 3.15.0] ([dateof:3.15.0]), an assignment in
the SET clause can be a 
[column-name-list|parenthesized list of column names] on the left and a
[row value] of the same size on the right.


<p>^The optional "OR <i>action</i>" conflict clause that follows the
UPDATE keyword allows the user to nominate a specific
constraint conflict resolution algorithm to use during this one UPDATE command.
Refer to the section entitled [ON CONFLICT] for additional information.

<h3>Restrictions on UPDATE Statements Within CREATE TRIGGER</h3>

<p>The following additional syntax restrictions apply to UPDATE statements that
occur within the body of a [CREATE TRIGGER] statement.  

<ul>
  <li><p>^The <yyterm>table-name</yyterm> specified as part of an UPDATE 
      statement within
      a trigger body must be unqualified. ^(In other words, the
      <i>schema-name</i><b>.</b> prefix on the table name of the UPDATE is
      not allowed within triggers.)^ ^Unless the table to which the trigger
      is attached is in the TEMP database, the table being updated by the
      trigger program must reside in the same database as it. ^If the table
      to which the trigger is attached is in the TEMP database, then the
      unqualified name of the table being updated is resolved in the same way
      as it is for a top-level statement (by searching first the TEMP database,
      then the main database, then any other databases in the order they were
      attached).

  <li><p>^The INDEXED BY and NOT INDEXED clauses are not allowed on UPDATE
      statements within triggers.</p>

  <li><p>^The LIMIT and ORDER BY clauses for UPDATE are unsupported within
      triggers, regardless of the compilation options used to build SQLite.
</ul>

<h3>Optional LIMIT and ORDER BY Clauses</h3>

<p>^(If SQLite is built with the [SQLITE_ENABLE_UPDATE_DELETE_LIMIT]
compile-time option then the syntax of the UPDATE statement is extended
with optional ORDER BY and LIMIT clauses)^ as follows:</p>

<tcl>BubbleDiagram update-stmt-limited</tcl>

<p>^If an UPDATE statement has a LIMIT clause, the maximum number of rows that
will be updated is found by evaluating the accompanying expression and casting
it to an integer value. ^A negative value is interpreted as "no limit".

<p>^If the LIMIT expression evaluates to non-negative value <i>N</i> and the
UPDATE statement has an ORDER BY clause, then all rows that would be updated in
the absence of the LIMIT clause are sorted according to the ORDER BY and the
first <i>N</i> updated. ^(If the UPDATE statement also has an OFFSET clause,
then it is similarly evaluated and cast to an integer value. If the OFFSET
expression evaluates to a non-negative value <i>M</i>, then the first <i>M</i>
rows are skipped and the following <i>N</i> rows updated instead.)^

<p>^If the UPDATE statement has no ORDER BY clause, then all rows that
would be updated in the absence of the LIMIT clause are assembled in an
arbitrary order before applying the LIMIT and OFFSET clauses to determine 
which are actually updated.

<p>^(The ORDER BY clause on an UPDATE statement is used only to determine which
rows fall within the LIMIT. The order in which rows are modified is arbitrary
and is not influenced by the ORDER BY clause.)^

<tcl>
##############################################################################
Section VACUUM vacuum VACUUM

RecursiveBubbleDiagram vacuum-stmt
</tcl>

<p>
  The VACUUM command rebuilds the database file, repacking it into a minimal
  amount of disk space. There are several 
  reasons an application might do this:

<ul>
  <li> <p> ^Unless SQLite is running in "auto_vacuum=FULL" mode, when a large
     amount of data is deleted from the database file it leaves behind empty
     space, or "free" database pages. This means the database file might
     be larger than strictly necessary. ^Running VACUUM to rebuild the 
     database reclaims this space and reduces the size of the database file.

  <li> <p> ^Frequent inserts, updates, and deletes can cause the database file
     to become fragmented - where data for a single table or index is scattered 
     around the database file. ^Running VACUUM ensures that each table and
     index is largely stored contiguously within the database file. ^In some
     cases, VACUUM may also reduce the number of partially filled pages in
     the database, reducing the size of the database file further.

  <li> <p> When content is deleted from an SQLite database, the content is not
     usually erased but rather the space used to hold the content is marked as
     being available for reuse.  This can allow deleted content to be recovered
     by a hacker or by forensic analysis.  Running VACUUM will clean the database
     of all traces of deleted content, thus preventing an adversary from recovering
     deleted content.  Using VACUUM in this way is an alternative to setting
     [PRAGMA secure_delete=ON]. 

  <li> <p> ^Normally, the database [page_size] and whether or not the database
     supports [auto_vacuum] must be configured before the database file is
     actually created. ^However, when not in [write-ahead log] mode, the 
     [page_size] and/or [auto_vacuum] properties of an existing database may be
     changed by using the [page_size pragma|page_size]  and/or 
     [auto_vacuum|pragma auto_vacuum] pragmas and then immediately VACUUMing
     the database. ^When in [write-ahead log] mode, only the [auto_vacuum]
     support property can be changed using VACUUM.
</ul>

<p>^By default, VACUUM only works only on the main database.
^[ATTACH|Attached databases] can be vacuumed by appending the appropriate
<yyterm>schema-name</yyterm> to the VACUUM statement.

<p><b>Compatibility Warning:</b> The ability to vacuum attached databases was
added in [version 3.15.0] ([dateof:3.15.0]).  Prior to that, a 
<yyterm>schema-name</yyterm> added to the
VACUUM statement would be silently ignored and the "main" schema would be
vacuumed.</p>

<tcl>hd_fragment vacuuminto {VACUUM INTO}</tcl>
<h3>VACUUM with an INTO clause</h3>

<p>^If the INTO clause is included, then the original database file is
unchanged and a new database is created in the filename given by the
argument to the INTO clause.  The new database will contain the same
logical content as the original database, fully vacuumed.

<p>
The VACUUM command with an INTO clause is an alternative to the
[backup API] for generating backup copies of a live database.
The advantage of using VACUUM INTO is that the resulting backup
database is minimal in size and hence the amount of filesystem
I/O may be reduced.  Also, all deleted content is purged from the
backup, leaving behind no forensic traces.  On the other hand,
the [backup API] uses fewer CPU cycles and can be executed
incrementally.

<p>
^The filename in the INTO clause can be an arbitrary SQL expression
that evaluates to a string.
^The file named by the INTO clause must not previously exist, or
else it must be an empty file, or the VACUUM INTO command will
fail with an error.

<p>
^The argument to INTO can be a [URI filename] if URI filenames
are enabled.
URL filenames are enabled if any of the following are true:
<ul>
<li> The SQLite library was compiled with [-DSQLITE_USE_URI=1].
<li> The [sqlite3_config]([SQLITE_CONFIG_URI],1) interfaces was
     invoked at start-time.
<li> The [database connection] that is running the VACUUM INTO
     statement was originally opened using the
     [SQLITE_OPEN_URI] flag.
</ul>

<p>
The VACUUM INTO command is transactional in the sense that
the generated output database is a consistent snapshot of the
original database.  However, if the VACUUM INTO command is
interrupted by an unplanned shutdown or power lose, then
the generated output database might be incomplete and corrupt.
Also, SQLite does not invoke fsync() or FlushFileBuffers()
on the generated database to ensure that it has reached
non-volatile storage before completing.


<tcl>hd_fragment howvacuumworks {how vacuum works}</tcl>
<h3>How VACUUM works</h3>

<p>^The VACUUM command works by copying the contents of the database into
a temporary database file and then overwriting the original with the 
contents of the temporary file. ^When overwriting the original, a rollback
journal or [write-ahead log] WAL file is used just as it would be for any
other database transaction. ^This means that when VACUUMing a database, 
as much as twice the size of the original database file is required in free
disk space.

<p>^The VACUUM INTO command works the same way except that it uses the file
named on the INTO clause in place of the temporary database and omits the
step of copying the vacuumed database back over top of the original database.

<p>^The VACUUM command may change the [ROWID | ROWIDs] of entries in any
tables that do not have an explicit [INTEGER PRIMARY KEY].
</p>

<p>^A VACUUM will fail if there is an open transaction on the database
connection that is attempting to run the VACUUM.  ^Unfinalized SQL
statements typically hold a read transaction open, so the VACUUM
might fail if there are unfinalized SQL statements on the same connection.
^VACUUM (but not VACUUM INTO) is a write operation and so if another
database connection is holding a lock that prevents writes, then
the VACUUM will fail.

<p>^(An alternative to using the VACUUM command to
reclaim space after data has been deleted is auto-vacuum mode, enabled using
the [auto_vacuum] pragma.)^ ^When [auto_vacuum] is enabled for a database
free pages may be reclaimed after deleting data, causing the file to shrink,
without rebuilding the entire database using VACUUM.  ^However, using
[auto_vacuum] can lead to extra database file fragmentation.  ^And [auto_vacuum]
does not compact partially filled pages of the database as VACUUM does.

</p>


<tcl>
##############################################################################
Section {INDEXED BY} indexedby {{INDEXED BY} {NOT INDEXED}}

</tcl>
<p>^The INDEXED BY phrase forces the [SQLite query planner] to use a
particular named index on a [DELETE], [SELECT], or [UPDATE] statement.
The INDEXED BY phrase is an SQLite extension and
is not portable to other SQL database engines.</p>

<tcl>
RecursiveBubbleDiagram qualified-table-name
</tcl>

<p>^The "INDEXED BY <yyterm>index-name</yyterm>" phrase specifies 
that the named index
must be used in order to look up values on the preceding table.
^If <yyterm>index-name</yyterm> does not exist or cannot be used 
for the query, then the preparation of the SQL statement fails.
^(The "NOT INDEXED" clause specifies that no index shall be used when
accessing the preceding table, including implied indices create by
UNIQUE and PRIMARY KEY constraints.  However, the [rowid]
can still be used to look up entries even when "NOT INDEXED" is specified.)^</p>

<p>Some SQL database engines provide non-standard "hint" mechanisms which
can be used to give the query optimizer clues about what indices it should
use for a particular statement.  The INDEX BY clause of SQLite is 
<em>not</em> a hinting mechanism and it should not be used as such.
^The INDEXED BY clause does not give the optimizer hints about which index
to use; it gives the optimizer a requirement of which index to use.
^If the query optimizer is unable to use the index specified by the
INDEX BY clause, then the query will fail with an error.</p>

<p>The INDEXED BY clause is <em>not</em> intended for use in tuning
the performance of a query.  The intent of the INDEXED BY clause is
to raise a run-time error if a schema change, such as dropping or
creating an index, causes the query plan for a time-sensitive query
to change.  The INDEXED BY clause is designed to help detect
undesirable query plan changes during regression testing.
Application 
developers are admonished to omit all use of INDEXED BY during
application design, implementation, testing, and tuning.  If
INDEXED BY is to be used at all, it should be inserted at the very
end of the development process when "locking down" a design.</p>

<h3>See Also:</h3>

<ol>
<li><p>The [query planner checklist] describes steps that application
developers should following to help resolve query planner problems.
Notice the that the use of INDEXED BY is a last resort, to be used only
when all other measures fail.</p>

<li><p>[upluscontrol|The unary "+" operator]
can be used to disqualify terms in the WHERE clause from use by indices.
Careful use of unary + can sometimes help prevent the query planner from
choosing a poor index without restricting it to using one specific index.
Careful placement of unary + operators is a better method for controlling 
which indices are used by a query.</p>

<li><p>The [sqlite3_stmt_status()] C/C++ interface together with the
[SQLITE_STMTSTATUS_FULLSCAN_STEP] and [SQLITE_STMTSTATUS_SORT] verbs
can be used to detect at run-time when an SQL statement is not
making effective use of indices.  Many applications may prefer to
use the [sqlite3_stmt_status()] interface to detect index misuse
rather than the INDEXED BY phrase described here.</p>
</ol>

<tcl>
set keyword_list [lsort {
   ABORT
   ACTION
   ADD
   AFTER
   ALL
   ALTER
   ALWAYS
   ANALYZE
   AND
   AS
   ASC
   ATTACH
   AUTOINCREMENT
   BEFORE
   BEGIN
   BETWEEN
   BY
   CASCADE
   CASE
   CAST
   CHECK
   COLLATE
   COLUMN
   COMMIT
   CONFLICT
   CONSTRAINT
   CREATE
   CROSS
   CURRENT
   CURRENT_DATE
   CURRENT_TIME
   CURRENT_TIMESTAMP
   DATABASE
   DEFAULT
   DEFERRED
   DEFERRABLE
   DELETE
   DESC
   DETACH
   DISTINCT
   DO
   DROP
   END
   EACH
   ELSE
   ESCAPE
   EXCEPT
   EXCLUSIVE
   EXCLUDE
   EXISTS
   EXPLAIN
   FAIL
   FILTER
   FIRST
   FOLLOWING
   FOR
   FOREIGN
   FROM
   FULL
   GENERATED
   GLOB
   GROUP
   GROUPS
   HAVING
   IF
   IGNORE
   IMMEDIATE
   IN
   INDEX
   INDEXED
   INITIALLY
   INNER
   INSERT
   INSTEAD
   INTERSECT
   INTO
   IS
   ISNULL
   JOIN
   KEY
   LAST
   LEFT
   LIKE
   LIMIT
   MATCH
   NATURAL
   NO
   NOT
   NOTHING
   NOTNULL
   NULL
   NULLS
   OF
   OFFSET
   ON
   OR
   ORDER
   OTHERS
   OUTER
   OVER
   PARTITION
   PLAN
   PRAGMA
   PRECEDING
   PRIMARY
   QUERY
   RAISE
   RANGE
   RECURSIVE
   REFERENCES
   REGEXP
   REINDEX
   RELEASE
   RENAME
   REPLACE
   RESTRICT
   RIGHT
   ROLLBACK
   ROW
   ROWS
   SAVEPOINT
   SELECT
   SET
   TABLE
   TEMP
   TEMPORARY
   THEN
   TIES
   TO
   TRANSACTION
   TRIGGER
   UNBOUNDED
   UNION
   UNIQUE
   UPDATE
   USING
   VACUUM
   VALUES
   VIEW
   VIRTUAL
   WHEN
   WHERE
   WINDOW
   WITH
   WITHOUT
}]

hd_puts {<DIV class="pdf_section">}
Section {SQLite Keywords} keywords {{*SQL keyword} {SQL keywords}}
hd_puts {</DIV>}
</tcl>

<p>The SQL standard specifies a large number of keywords which may not
be used as the names of tables, indices, columns, databases, user-defined
functions, collations, virtual table modules, or any other named object.
The list of keywords is so long that few people can remember them all.
For most SQL code, your safest bet is to never use any English language
word as the name of a user-defined object.</p>

<p>If you want to use a keyword as a name, you need to quote it.  There
are four ways of quoting keywords in SQLite:</p>

<p>
<blockquote>
<table>
<tr>	<td valign="top"><b>'keyword'</b></td><td width="20"></td>
	<td>^A keyword in single quotes is a string literal.</td></tr>

<tr>	<td valign="top"><b>"keyword"</b></td><td></td>
	<td>^A keyword in double-quotes is an identifier.</td></tr>

<tr>	<td valign="top"><b>&#91;keyword&#93;</b></td><td></td>
	<td>^A keyword enclosed in square brackets is 
        an identifier.  This is not standard SQL.  This quoting mechanism
        is used by MS Access and SQL Server and is included in SQLite for
        compatibility.</td></tr>

<tr>	<td valign="top"><b>&#96;keyword&#96;</b></td><td></td>
	<td>^A keyword enclosed in grave accents (ASCII code 96) is 
        an identifier.  This is not standard SQL.  This quoting mechanism
        is used by MySQL and is included in SQLite for
        compatibility.</td></tr>
</table>
</blockquote>
</p>

<p>For resilience when confronted with historical SQL statements, SQLite
will sometimes bend the quoting rules above:</p>

<ul>
<li><p>^If a keyword in single
quotes (ex: <b>'key'</b> or <b>'glob'</b>) is used in a context where
an identifier is allowed but where a string literal is not allowed, then
the token is understood to be an identifier instead of a string literal.
</p></li>

<li><p>^If a keyword in double
quotes (ex: <b>"key"</b> or <b>"glob"</b>) is used in a context where
it cannot be resolved to an identifier but where a string literal
is allowed, then the token is understood to be a string literal instead
of an identifier.</p></li>
</ul>

<p>Programmers are cautioned not to use the two exceptions described in
the previous bullets.  We emphasize that they exist only so that old
and ill-formed SQL statements will run correctly.  Future versions of
SQLite might raise errors instead of accepting the malformed
statements covered by the exceptions above.</p>

<p>
SQLite adds new keywords from time to time when it takes on new features.
So to prevent your code from being broken by future enhancements, you should
normally quote any identifier that is an English language word, even if
you do not have to.
</p>

<p>
The list below shows all possible keywords used by any build of
SQLite regardless of [compile-time options].  
Most reasonable configurations use most or all of these keywords,
but some keywords may be omitted when SQL language features are
disabled.
Applications can use the
[sqlite3_keyword_count()], [sqlite3_keyword_name()], and
[sqlite3_keyword_check()] interfaces to determine the keywords
recognized by SQLite at run-time.
^(Regardless of the compile-time configuration, any identifier that is not on
the following <tcl>hd_puts [llength $keyword_list]</tcl> element
list is not a keyword to the SQL parser in SQLite:
</p>

<ol>
<tcl>
set lx {}
foreach word [lsort $keyword_list] {
  hd_putsnl "<li>$word</li>"
}
hd_putsnl "</ol>)^"
</tcl>

    foreach {s_title s_tag s_kw} $section {}
    if {$s_kw == ""} {
      set s_kw $s_title
    }
    lappend lx [list $s_kw $s_title 0]
  }
  hd_list_of_links {} 340 $lx
  proc Operator {name} {
    return "<font color=\"#2c2cf0\"><big>$name</big></font>"
  }
  proc Example {text} {
    hd_puts "<blockquote><pre>$text</pre></blockquote>"
  }
</tcl>
</ul></td></tr></table>

<p>^The routines [sqlite3_prepare_v2()], [sqlite3_prepare()],
[sqlite3_prepare16()], [sqlite3_prepare16_v2()], 
[sqlite3_exec()], and [sqlite3_get_table()] accept
an SQL statement list (sql-stmt-list) which is a semicolon-separated
................................................................................

<tcl>BubbleDiagram sql-stmt-list</tcl>

<p>Each SQL statement in the statement list is an instance of the
following:</p>

<tcl>BubbleDiagram sql-stmt</tcl>

<title>Built-in Aggregate Functions</title>
<tcl>
hd_keywords *aggfunc {Aggregate Functions}

unset -nocomplain corefuncset
proc funcdef {syntax keywords desc} {
  global corefuncset
  regexp {^[a-z_]+} $syntax basesyntax
  set corefuncset($basesyntax) [list $syntax $keywords $desc]
}

funcdef {avg(X)} {*avg {avg() aggregate function}} {
  ^The avg() function
  returns the average value of all non-NULL <i>X</i> within a
  group.  ^String and BLOB values that do not look like numbers are
  interpreted as 0.
  ^The result of avg() is always a floating point value as long as
  at there is at least one non-NULL input even if all
  inputs are integers.  ^The result of avg() is NULL if and only if
  there are no non-NULL inputs.  
}

funcdef {count(X) count(*)} {*count {count() aggregate function}} {
  ^The count(X) function returns
  a count of the number of times
  that <i>X</i> is not NULL in a group.  ^The count(*) function
  (with no arguments) returns the total number of rows in the group.
}

funcdef {group_concat(X) group_concat(X,Y)} {
  *group_concat {group_concat() aggregate function}
} {
  ^The group_concat() function returns
  a string which is the concatenation of
  all non-NULL values of <i>X</i>.  ^If parameter <i>Y</i> is present then
  it is used as the separator
  between instances of <i>X</i>.  ^A comma (",") is used as the separator
  if <i>Y</i> is omitted.  The order of the concatenated elements is
  arbitrary.
}

funcdef {max(X)} {*maxAggFunc *agg_max {max() aggregate function}} {
  ^The max() aggregate function
  returns the maximum value of all values in the group.
  ^The maximum value is the value that would be returned last in an
  ORDER BY on the same column.  ^Aggregate max() returns NULL 
  if and only if there are no non-NULL values in the group.
}

funcdef {min(X)} {*minAggFunc *agg_min {min() aggregate function}} {
  ^The min() aggregate function
  returns the minimum non-NULL value of all values in the group.
  ^The minimum value is the first non-NULL value that would appear
  in an ORDER BY of the column.
  ^Aggregate min() returns NULL if and only if there are no non-NULL
  values in the group.
}

funcdef {sum(X) total(X)} {
  *sumFunc *sum *total
  {sum() aggregate function}
  {total() aggregate function}
} {
  ^The sum() and total() aggregate functions
  return sum of all non-NULL values in the group.
  ^If there are no non-NULL input rows then sum() returns
  NULL but total() returns 0.0.
  NULL is not normally a helpful result for the sum of no rows
  but the SQL standard requires it and most other
  SQL database engines implement sum() that way so SQLite does it in the
  same way in order to be compatible.   The non-standard total() function
  is provided as a convenient way to work around this design problem
  in the SQL language.</p>

  <p>^The result of total() is always a floating point value.
  ^The result of sum() is an integer value if all non-NULL inputs are integers.
  ^If any input to sum() is neither an integer or a NULL
  then sum() returns a floating point value
  which might be an approximation to the true sum.</p>

  <p>^Sum() will throw an "integer overflow" exception if all inputs
  are integers or NULL
  and an integer overflow occurs at any point during the computation.
  ^Total() never throws an integer overflow.
}
</tcl>

<fancy_format>
<tcl>
RecursiveBubbleDiagram aggregate-function-invocation
</tcl>

<p>
The aggregate functions shown below are available by default.  Additional
aggregate functions written in C may be added using the 
[sqlite3_create_function()]</a>
API.</p>

<p>
^In any aggregate function that takes a single argument, that argument
can be preceded by the keyword DISTINCT.  ^In such cases, duplicate
elements are filtered before being passed into the aggregate function.
^For example, the function "count(distinct X)" will return the number
of distinct values of column X instead of the total number of non-null
values in column X.
</p>

<tcl>hd_fragment aggfilter {FILTER clause on aggregate functions}</tcl>
<p>
If a FILTER clause is provided, then only rows for which the <i>expr</i> is
true are included in the aggregate.
</p>

<tcl>
set lx {}
foreach basesyntax [array names corefuncset] {
  foreach {syntax keywords desc} $corefuncset($basesyntax) break
  regexp {^[a-z_]+} $basesyntax fragment
  foreach coresyntax $syntax {
    lappend lx [list $fragment $coresyntax 0]
  }
}
hd_list_of_links {} 225 [lsort -index 1 $lx]

hd_putsnl "<hr class='xhr'>"
hd_putsnl "<dl>"
foreach basesyntax [lsort [array names corefuncset]] {
  foreach {syntax keywords desc} $corefuncset($basesyntax) break
  regsub -all {\s+} [string trim $syntax] {<br />} syntax
  regsub -all {\(([^*)]+)\)} $syntax {(<i>\1</i>)} syntax
  regsub -all {,} $syntax {</i>,<i>} syntax
  regsub -all {<i>\.\.\.</i>} $syntax {...} syntax
  regexp {^[a-z]+} $basesyntax fragment
  if {[llength $keywords]==0} {
    regexp {[a-z_]+} $syntax name
    hd_fragment $name *$name "${name}() SQL function"
  } else {
    set fragname [lindex $keywords 0]
    regsub -all {[^a-z]} $fragname {} fragname
    hd_fragment $fragname
    eval hd_keywords [string map {\n { }} $keywords]
  }
  hd_puts "<dt><p><b>$syntax</b></dt>"
  hd_resolve "<dd><p>$desc</dd>\n"
}
hd_putsnl "</dl>"

<title>The ON CONFLICT Clause</title>
<tcl>
hd_keywords *conflict {conflict clause} {ON CONFLICT} {ON CONFLICT clause}
</tcl>

<fancy_format>
<tcl>
RecursiveBubbleDiagram conflict-clause
</tcl>

<p>The ON CONFLICT clause is a
non-standard extension specific to SQLite 
that can appear in many other SQL commands.
It is given its own section in this document because it is not
part of standard SQL and therefore might not be familiar.</p>

<p>The ON CONFLICT clause described here has been a part of
SQLite since before version 3.0.0 ([dateof:3.0.0]).  The phrase
"ON&nbsp;CONFLICT" is also part of [UPSERT], which is an extension
to [INSERT] added in version 3.24.0 ([dateof:3.24.0]).  Do not
confuse these two separate uses of the "ON&nbsp;CONFLICT" phrase.

<p>^The syntax for the ON CONFLICT clause is as shown above for
the CREATE TABLE command.  ^For the INSERT and
UPDATE commands, the keywords "ON CONFLICT" are replaced by "OR" so that
the syntax reads more naturally.  For example, instead of
"INSERT ON CONFLICT IGNORE" we have "INSERT OR IGNORE".
The keywords change but the meaning of the clause is the same
either way.</p>

<p>The ON CONFLICT clause applies to [UNIQUE], [NOT NULL],
[CHECK], and [PRIMARY KEY] constraints.
The ON CONFLICT algorithm does not
apply to [FOREIGN KEY constraints].
There are five conflict resolution algorithm choices:
ROLLBACK, ABORT, FAIL, IGNORE, and REPLACE.
^The default conflict resolution algorithm is ABORT.  This
is what they mean:</p>

<dl>
<dt><b>ROLLBACK</b></dt>
<dd><p> ^When an applicable constraint violation occurs, the ROLLBACK
resolution algorithm aborts the current SQL statement with
an SQLITE_CONSTRAINT error and rolls back the current transaction.
^If no transaction is
active (other than the implied transaction that is created on every
command) then the ROLLBACK resolution algorithm works the same as the
ABORT algorithm.</p></dd>

<dt><b>ABORT</b></dt>
<dd><p> ^When an applicable constraint violation occurs, the ABORT
resolution algorithm aborts the current SQL statement
with an SQLITE_CONSTRAINT error and backs out any changes
made by the current SQL statement; but changes caused
by prior SQL statements within the same transaction are preserved and the
transaction remains active.
This is the default behavior and the behavior specified by the SQL
standard.</p></dd>

<dt><b>FAIL</b></dt>
<dd><p> ^When an applicable constraint violation occurs, the FAIL
resolution algorithm aborts the current SQL statement with an
SQLITE_CONSTRAINT error.  ^But the FAIL resolution does not
back out prior changes of the SQL statement that failed nor does
it end the transaction.
^For example, if an UPDATE
statement encountered a constraint violation on the 100th row that
it attempts to update, then the first 99 row changes are preserved
but changes to rows 100 and beyond never occur.</p>

<p>The FAIL behavior only works for uniqueness, NOT NULL, and CHECK 
constraints.  A [foreign key constraint] violation causes an ABORT.
</p></dd>

<dt><b>IGNORE</b></dt>
<dd><p> ^When an applicable constraint violation occurs, 
the IGNORE resolution algorithm skips the one row that contains
the constraint violation and continues processing subsequent rows
of the SQL statement as if nothing went wrong.
Other rows before and after the row that
contained the constraint violation are inserted or updated
normally. ^No error is returned for uniqueness, NOT NULL, and
UNIQUE constraint errors when the IGNORE conflict resolution
algorithm is used.  ^However, the IGNORE conflict resolution
algorithm works like ABORT for [foreign key constraint] errors.
</p>
</dd>

<dt><b>REPLACE</b></dt>
<dd><p> ^When a [UNIQUE] or [PRIMARY KEY] constraint violation occurs, 
the REPLACE algorithm
deletes pre-existing rows that are causing the constraint violation
prior to inserting or updating the current row and the command continues 
executing normally.
^If a [NOT NULL] constraint violation occurs, the REPLACE conflict
resolution replaces the NULL value with
the default value for that column, or if the column has no default
value, then the ABORT algorithm is used.
^If a [CHECK constraint] or [foreign key constraint] violation occurs, 
the REPLACE conflict resolution algorithm works like ABORT.</p>

<p>^When the REPLACE conflict resolution strategy deletes rows in order to
satisfy a constraint, [CREATE TRIGGER | delete triggers] fire if and only if
[recursive_triggers pragma | recursive triggers] are enabled.</p>

<p>^The [sqlite3_update_hook | update hook] is not invoked for rows that
are deleted by the REPLACE conflict resolution strategy.  ^Nor does
REPLACE increment the [sqlite3_changes | change counter].
The exceptional behaviors defined in this paragraph might change 
in a future release.</p>
</dl>

<p>^The algorithm specified in the OR clause of an INSERT or UPDATE
overrides any algorithm specified in a CREATE TABLE.
^If no algorithm is specified anywhere, the ABORT algorithm is used.</p>

<title>Built-In Scalar SQL Functions</title>
<tcl>
hd_keywords *corefunc {Core Functions}

unset -nocomplain corefuncset
proc funcdef {syntax keywords desc} {
  global corefuncset
  regexp {^[a-z_]+} $syntax basesyntax
  set corefuncset($basesyntax) [list $syntax $keywords $desc]
}
funcdef {abs(X)} {} {
  ^The abs(X) function returns the absolute value of the numeric
  argument X.  ^Abs(X) returns NULL if X is NULL. 
  ^(Abs(X) returns 0.0 if X is a string or blob
  that cannot be converted to a numeric value.)^  ^If X is the 
  integer -9223372036854775808 then abs(X) throws an integer overflow
  error since there is no equivalent positive 64-bit two complement value.
}

funcdef {changes()} {} {
  ^The changes() function returns the number of database rows that were changed
  or inserted or deleted by the most recently completed INSERT, DELETE,
  or UPDATE statement, exclusive of statements in lower-level triggers.
  ^The changes() SQL function is a wrapper around the [sqlite3_changes()]
  C/C++ function and hence follows the same rules for counting changes.
}

funcdef {char(X1,X2,...,XN)} {} {
  ^(The char(X1,X2,...,XN) function returns a string composed of characters having the
   unicode code point values of integers X1 through XN, respectively.)^
}

funcdef {coalesce(X,Y,...)} {} {
  ^The coalesce() function returns a copy of its first non-NULL argument, or
  NULL if all arguments are NULL.  ^Coalesce() must have at least 
  2 arguments.
}

funcdef {glob(X,Y)} {} {
  ^The glob(X,Y) function is equivalent to the
  expression "<b>Y GLOB X</b>".
  Note that the X and Y arguments are reversed in the glob() function
  relative to the infix [GLOB] operator.  Y is the string and X is the
  pattern.  So, for example, the following expressions are equivalent:
  <blockquote><pre>
     name GLOB '*helium*'
     glob('*helium*',name)
  </pre></blockquote>
  <p>^If the [sqlite3_create_function()] interface is used to
  override the glob(X,Y) function with an alternative implementation then
  the [GLOB] operator will invoke the alternative implementation.
}

funcdef {ifnull(X,Y)} {} {
  ^The ifnull() function returns a copy of its first non-NULL argument, or
  NULL if both arguments are NULL.  ^Ifnull() must have exactly 2 arguments.
  ^The ifnull() function is equivalent to [coalesce()] with two arguments.
}

funcdef {instr(X,Y)} {} {
  ^The instr(X,Y) function finds the first occurrence of string Y within 
  string X and returns the number of prior characters plus 1, or 0 if
  Y is nowhere found within X.
  ^Or, if X and Y are both BLOBs, then instr(X,Y) returns one
  more than the number bytes prior to the first occurrence of Y, or 0 if
  Y does not occur anywhere within X.
  ^If both arguments X and Y to instr(X,Y) are non-NULL and are not BLOBs
  then both are interpreted as strings.
  ^If either X or Y are NULL in instr(X,Y) then the result is NULL.
}

funcdef {hex(X)} {} {
  ^The hex() function interprets its argument as a BLOB and returns
  a string which is the upper-case hexadecimal rendering of the content of
  that blob.
}

funcdef {last_insert_rowid()} {} {
  ^The last_insert_rowid() function returns the [ROWID]
  of the last row insert from the database connection which invoked the
  function.
  ^The last_insert_rowid() SQL function is a wrapper around the
  [sqlite3_last_insert_rowid()] C/C++ interface function.
}

funcdef {length(X)} {} {
  ^For a string value X, the length(X) function returns the number of 
  characters (not bytes) in X prior to the first NUL character.
  Since SQLite strings do not normally contain NUL characters, the length(X)
  function will usually return the total number of characters in the string X.
  ^For a blob value X, length(X) returns the number of bytes in the blob.
  ^If X is NULL then length(X) is NULL.
  ^If X is numeric then length(X) returns the length of a string
  representation of X.
}

funcdef {like(X,Y) like(X,Y,Z)} {} {
  ^The like() function is used to implement the
  "<b>Y LIKE X &#91;ESCAPE Z&#93;</b>" expression. 
  ^If the optional ESCAPE clause is present, then the
  like() function is invoked with three arguments.  ^Otherwise, it is
  invoked with two arguments only. Note that the X and Y parameters are
  reversed in the like() function relative to the infix [LIKE] operator.
  X is the pattern and Y is the string to match against that pattern.
  Hence, the following expressions are equivalent:
  <blockquote><pre>
     name LIKE '%neon%'
     like('%neon%',name)
  </pre></blockquote>
  ^The [sqlite3_create_function()] interface can be used to override the
  like() function and thereby change the operation of the
  [LIKE] operator.  When overriding the like() function, it may be important
  to override both the two and three argument versions of the like() 
  function. Otherwise, different code may be called to implement the
  [LIKE] operator depending on whether or not an ESCAPE clause was 
  specified.
}


funcdef {likelihood(X,Y)} {} {
  ^The likelihood(X,Y) function returns argument X unchanged.
  ^(The value Y in likelihood(X,Y) must be a floating point constant
  between 0.0 and 1.0, inclusive.)^
  ^The likelihood(X) function is a no-op that the code generator
  optimizes away so that it consumes no CPU cycles during run-time
  (that is, during calls to [sqlite3_step()]).
  ^The purpose of the likelihood(X,Y) function is to provide a hint
  to the query planner that the argument X is a boolean that is
  true with a probability of approximately Y.
  ^(The [unlikely(X)] function is short-hand for likelihood(X,0.0625).)^
  ^(The [likely(X)] function is short-hand for likelihood(X,0.9375).)^
}

funcdef {likely(X)} {} {
  ^The likely(X) function returns the argument X unchanged.
  ^The likely(X) function is a no-op that the code generator
  optimizes away so that it consumes no CPU cycles at
  run-time (that is, during calls to [sqlite3_step()]).
  ^The purpose of the likely(X) function is to provide a hint
  to the query planner that the argument X is a boolean value
  that is usually true. ^(The likely(X) function is equivalent
  to [likelihood](X,0.9375).)^ See also: [unlikely(X)].
}

funcdef {load_extension(X) load_extension(X,Y)} {} {
  ^The load_extension(X,Y) function loads [SQLite extensions] out of the shared
  library file named X using the entry point Y.  ^The result of load_extension()
  is always a NULL.  ^If Y is omitted then the default entry point name is used.
  ^The load_extension() function raises an exception if the extension fails to
  load or initialize correctly.

  <p>^The load_extension() function will fail if the extension attempts to 
  modify or delete an SQL function or collating sequence.  ^The
  extension can add new functions or collating sequences, but cannot
  modify or delete existing functions or collating sequences because
  those functions and/or collating sequences might be used elsewhere
  in the currently running SQL statement.  To load an extension that
  changes or deletes functions or collating sequences, use the
  [sqlite3_load_extension()] C-language API.</p>

  <p>For security reasons, extension loaded is turned off by default and must
  be enabled by a prior call to [sqlite3_enable_load_extension()].</p>
}

funcdef {lower(X)} {} {
  ^The lower(X) function returns a copy of string X with all ASCII characters
  converted to lower case.  ^The default built-in lower() function works
  for ASCII characters only.  To do case conversions on non-ASCII
  characters, load the ICU extension.
}

funcdef {ltrim(X) ltrim(X,Y)} {} {
  ^The ltrim(X,Y) function returns a string formed by removing any and all
  characters that appear in Y from the left side of X.
  ^If the Y argument is omitted, ltrim(X) removes spaces from the left side
  of X.
}

funcdef {max(X,Y,...)} {*maxCoreFunc *max {max() SQL function}} {
  ^The multi-argument max() function returns the argument with the 
  maximum value, or return NULL if any argument is NULL. 
  ^The multi-argument max() function searches its arguments from left to right
  for an argument that defines a collating function and uses that collating
  function for all string comparisons.  ^If none of the arguments to max()
  define a collating function, then the BINARY collating function is used.
  ^(Note that <b>max()</b> is a simple function when
  it has 2 or more arguments but operates as an
  [maxAggFunc | aggregate function] if given only a single argument.)^
}

funcdef {min(X,Y,...)} {*minCoreFunc *min {min() SQL function}} {
  ^The multi-argument min() function returns the argument with the
  minimum value.
  ^The multi-argument min() function searches its arguments from left to right
  for an argument that defines a collating function and uses that collating
  function for all string comparisons.  ^If none of the arguments to min()
  define a collating function, then the BINARY collating function is used.
  ^(Note that <b>min()</b> is a simple function when
  it has 2 or more arguments but operates as an 
  [minAggFunc | aggregate function] if given
  only a single argument.)^
}

funcdef {nullif(X,Y)} {} {
  ^The nullif(X,Y) function returns its first argument if the arguments are
  different and NULL if the arguments are the same.  ^The nullif(X,Y) function
  searches its arguments from left to right for an argument that defines a
  collating function and uses that collating function for all string
  comparisons.  ^If neither argument to nullif() defines a collating function
  then the BINARY is used.
}

funcdef {printf(FORMAT,...)} {} {
  ^(The printf(FORMAT,...) SQL function works like the [sqlite3_mprintf()] C-language
  function and the printf() function from the standard C library.)^
  The first argument is a format string that specifies how to construct the output
  string using values taken from subsequent arguments.  ^If the FORMAT argument is
  missing or NULL then the result is NULL.  ^The %n format is silently ignored and
  does not consume an argument.  ^The %p format is an alias for %X.  ^The %z format
  is interchangeable with %s.  ^(If there are too few arguments in the argument list,
  missing arguments are assumed to have a NULL value, which is translated into
  0 or 0.0 for numeric formats or an empty string for %s.)^  See the
  [built-in printf()] documentation for additional information.
}
  

funcdef {quote(X)} {} {
  ^The quote(X) function returns the text of an SQL literal which
  is the value of its argument suitable for inclusion into an SQL statement.
  ^Strings are surrounded by single-quotes with escapes on interior quotes
  as needed.  ^BLOBs are encoded as hexadecimal literals.
  ^Strings with embedded NUL characters cannot be represented as string
  literals in SQL and hence the returned string literal is truncated prior
  to the first NUL.
}

funcdef {random()} {} {
  ^The random() function returns a pseudo-random integer
  between -9223372036854775808 and +9223372036854775807.
}

funcdef {randomblob(N)} {} {
  ^The randomblob(N) function return an N-byte blob containing pseudo-random
  bytes. ^If N is less than 1 then a 1-byte random blob is returned.

  <p>Hint:  applications can generate globally unique identifiers
  using this function together with [hex()] and/or
  [lower()] like this:</p>

  <blockquote>
  hex(randomblob(16))<br></br>
  lower(hex(randomblob(16)))
  </blockquote>
}

funcdef {replace(X,Y,Z)} {} {
  ^The replace(X,Y,Z) function returns a string formed by substituting
  string Z for every occurrence of string Y in string X.  ^The [BINARY]
  collating sequence is used for comparisons.  ^If Y is an empty
  string then return X unchanged.  ^If Z is not initially
  a string, it is cast to a UTF-8 string prior to processing.
}

funcdef {round(X) round(X,Y)} {} {
  ^The round(X,Y) function returns a floating-point
  value X rounded to Y digits to the right of the decimal point.
  ^If the Y argument is omitted, it is assumed to be 0.
}

funcdef {rtrim(X) rtrim(X,Y)} {} {
  ^The rtrim(X,Y) function returns a string formed by removing any and all
  characters that appear in Y from the right side of X.
  ^If the Y argument is omitted, rtrim(X) removes spaces from the right
  side of X.
}

funcdef {soundex(X)} {} {
  ^The soundex(X) function returns a string that is the soundex encoding 
  of the string X.
  ^The string "?000" is returned if the argument is NULL or contains
  no ASCII alphabetic characters.
  ^(This function is omitted from SQLite by default.
  It is only available if the [SQLITE_SOUNDEX] compile-time option
  is used when SQLite is built.)^
}

funcdef {sqlite_compileoption_get(N)} {} {
  ^The sqlite_compileoption_get() SQL function is a wrapper around the
  [sqlite3_compileoption_get()] C/C++ function.
  ^This routine returns the N-th compile-time option used to build SQLite
  or NULL if N is out of range.  See also the [compile_options pragma].
}

funcdef {sqlite_compileoption_used(X)} {} {
  ^The sqlite_compileoption_used() SQL function is a wrapper around the
  [sqlite3_compileoption_used()] C/C++ function.
  ^When the argument X to sqlite_compileoption_used(X) is a string which
  is the name of a compile-time option, this routine returns true (1) or
  false (0) depending on whether or not that option was used during the
  build.
}

funcdef {sqlite_offset(X)} {} {
  ^The sqlite_offset(X) function returns the byte offset in the database
  file for the beginning of the record from which value would be read.
  If X is not a column in an ordinary table, then sqlite_offset(X) returns
  NULL.  The value returned by sqlite_offset(X) might reference either the
  original table or an index, depending on the query.  If the value X would
  normally be extracted from an index, the sqlite_offset(X) returns the
  offset to the corresponding index record.  If the value X would be
  extracted from the original table, then sqlite_offset(X) returns the offset
  to the table record.

  <p>The sqlite_offset(X) SQL function is only available if SQLite is built
  using the [-DSQLITE_ENABLE_OFFSET_SQL_FUNC] compile-time option.
}

funcdef {sqlite_source_id()} {} {
  ^The sqlite_source_id() function returns a string that identifies the
  specific version of the source code that was used to build the SQLite
  library.  ^The string returned by sqlite_source_id() is
  the date and time that the source code was checked in followed by
  the SHA1 hash for that check-in.  ^This function is
  an SQL wrapper around the [sqlite3_sourceid()] C interface.
}

funcdef {sqlite_version()} {} {
  ^The sqlite_version() function returns the version string for the SQLite
  library that is running.  ^This function is an SQL
  wrapper around the [sqlite3_libversion()] C-interface.
}

funcdef {substr(X,Y,Z) substr(X,Y)} {} {
  ^The substr(X,Y,Z) function returns a substring of input string X that begins
  with the Y-th character and which is Z characters long.
  ^If Z is omitted then substr(X,Y) returns all characters through the end
  of the string X beginning with the Y-th.
  ^The left-most character of X is number 1.  ^If Y is negative
  then the first character of the substring is found by counting from the
  right rather than the left.  ^If Z is negative then
  the abs(Z) characters preceding the Y-th character are returned.
  ^If X is a string then characters indices refer to actual UTF-8 
  characters.  ^If X is a BLOB then the indices refer to bytes.
}

funcdef {total_changes()} {} {
  ^The total_changes() function returns the number of row changes
  caused by INSERT, UPDATE or DELETE
  statements since the current database connection was opened.
  ^This function is a wrapper around the [sqlite3_total_changes()]
  C/C++ interface.
}

funcdef {trim(X) trim(X,Y)} {} {
  ^The trim(X,Y) function returns a string formed by removing any and all
  characters that appear in Y from both ends of X.
  ^If the Y argument is omitted, trim(X) removes spaces from both ends of X.
}

funcdef {typeof(X)} {} {
  ^The typeof(X) function returns a string that indicates the [datatype] of
  the expression X: "null", "integer", "real", "text", or "blob".
}

funcdef {unlikely(X)} {} {
  ^The unlikely(X) function returns the argument X unchanged.
  ^The unlikely(X) function is a no-op that the code generator
  optimizes away so that it consumes no CPU cycles at
  run-time (that is, during calls to [sqlite3_step()]).
  ^The purpose of the unlikely(X) function is to provide a hint
  to the query planner that the argument X is a boolean value
  that is usually not true. ^(The unlikely(X) function is equivalent
  to [likelihood](X, 0.0625).)^
}

funcdef {unicode(X)} {} {
  ^The unicode(X) function returns the numeric unicode code point corresponding to
  the first character of the string X.  If the argument to unicode(X) is not a string
  then the result is undefined.
}

funcdef {upper(X)} {} {
  ^The upper(X) function returns a copy of input string X in which all 
  lower-case ASCII characters are converted to their upper-case equivalent.
}

funcdef {zeroblob(N)} {} {
  ^The zeroblob(N) function returns a BLOB consisting of N bytes of 0x00.
  SQLite manages these zeroblobs very efficiently.  Zeroblobs can be used to
  reserve space for a BLOB that is later written using 
  [sqlite3_blob_open() | incremental BLOB I/O].
  ^This SQL function is implemented using the [sqlite3_result_zeroblob()]
  routine from the C/C++ interface.
}
</tcl>

<fancy_format>

<p>The core functions shown below are available by default. 
[datefunc | Date &amp; Time functions],
[aggfunc | aggregate functions],
[builtin window functions|window functions], and
[json1 | JSON functions] are documented separately.  An
application may define additional
functions written in C and added to the database engine using
the [sqlite3_create_function()] API.</p>

<tcl>
set lx {}
foreach basesyntax [array names corefuncset] {
  foreach {syntax keywords desc} $corefuncset($basesyntax) break
  regexp {^[a-z_]+} $basesyntax fragment
  foreach coresyntax $syntax {
    lappend lx [list $fragment $coresyntax 0]
  }
}
hd_putsnl "<p>"
hd_list_of_links {} 300 [lsort -index 1 $lx]

hd_putsnl "<hr class='xhr'>"
hd_putsnl "<dl>"
foreach basesyntax [lsort [array names corefuncset]] {
  foreach {syntax keywords desc} $corefuncset($basesyntax) break
  regsub -all {\s+} [string trim $syntax] {<br />} syntax
  regsub -all {\(([^*)]+)\)} $syntax {(<i>\1</i>)} syntax
  regsub -all {,} $syntax {</i>,<i>} syntax
  regsub -all {<i>\.\.\.</i>} $syntax {...} syntax
  regexp {^[a-z]+} $basesyntax fragment
  if {[llength $keywords]==0} {
    regexp {[a-z_]+} $syntax name
    hd_fragment $name *$name "${name}() SQL function"
  } else {
    set fragname [lindex $keywords 0]
    regsub -all {[^a-z]} $fragname {} fragname
    hd_fragment $fragname
    eval hd_keywords [string map {\n { }} $keywords]
  }
  hd_puts "<dt><p><b>$syntax</b></dt>"
  hd_resolve "<dd><p>$desc</dd>\n"
}
hd_putsnl "</dl>"

<title>CREATE TRIGGER</title>
<tcl>
hd_keywords *createtrigger {CREATE TRIGGER}
</tcl>

<table_of_contents>
<h1>Syntax</h1>

<tcl>
RecursiveBubbleDiagram create-trigger-stmt
</tcl>

<h1>Description</h1>
<p>^The CREATE TRIGGER statement is used to add triggers to the 
database schema. ^Triggers are database operations 
that are automatically performed when a specified database event
occurs.  </p>

<p>^Each trigger must specify that it will fire for one of
the following operations: [DELETE], [INSERT], [UPDATE].
The trigger fires once for each row that is deleted, inserted,
or updated.  If the "UPDATE OF <yyterm>column-name</yyterm>"
syntax is used, then the trigger will only fire if
<yyterm>column-name</yyterm> appears on the left-hand side of
one of the terms in the SET clause of the [UPDATE] statement.</p>

<p>Due to an historical oversight, columns named in the "UPDATE OF"
clause do not actually have to exist in the table being updated.
Unrecognized column names are silently ignored.
It would be more helpful if SQLite would fail the CREATE TRIGGER
statement if any of the names in the "UPDATE OF" clause are not
columns in the table.  However, as this problem was discovered
many years after SQLite was widely deployed, we have resisted
fixing the problem for fear of breaking legacy applications.</p>

<p>^At this time SQLite supports only FOR EACH ROW triggers, not FOR EACH
STATEMENT triggers. ^Hence explicitly specifying FOR EACH ROW is optional.
^FOR EACH ROW implies that the SQL statements specified in the trigger
may be executed (depending on the WHEN clause) for each database row being
inserted, updated or deleted by the statement causing the trigger to fire.</p>

<p>^(Both the WHEN clause and the trigger actions may access elements of 
the row being inserted, deleted or updated using references of the form 
"NEW.<i>column-name</i>" and "OLD.<i>column-name</i>", where
<i>column-name</i> is the name of a column from the table that the trigger
is associated with.)^ ^(OLD and NEW references may only be used in triggers on
events for which they are relevant, as follows:</p>

<table border=0 cellpadding=10>
<tr>
<td valign="top" align="right" width=120><i>INSERT</i></td>
<td valign="top">NEW references are valid</td>
</tr>
<tr>
<td valign="top" align="right" width=120><i>UPDATE</i></td>
<td valign="top">NEW and OLD references are valid</td>
</tr>
<tr>
<td valign="top" align="right" width=120><i>DELETE</i></td>
<td valign="top">OLD references are valid</td>
</tr>
</table>
</p>)^

<p>^If a WHEN clause is supplied, the SQL statements specified
are only executed if the WHEN clause is true.
^If no WHEN clause is supplied, the SQL statements
are executed every time the trigger fires.</p>

<p>^The BEFORE or AFTER keyword determines when the trigger actions
will be executed relative to the insertion, modification or removal of the
associated row.</p>

<p>^An [ON CONFLICT] clause may be specified as part of an [UPDATE] or [INSERT]
action within the body of the trigger.
^However if an [ON CONFLICT] clause is specified as part of 
the statement causing the trigger to fire, then conflict handling
policy of the outer statement is used instead.</p>

<p>^Triggers are automatically [DROP TRIGGER | dropped]
when the table that they are 
associated with (the <i>table-name</i> table) is 
[DROP TABLE | dropped].  ^However if the trigger actions reference
other tables, the trigger is not dropped or modified if those other
tables are [DROP TABLE | dropped] or [ALTER TABLE | modified].</p>

<p>^Triggers are removed using the [DROP TRIGGER] statement.</p>

<h2>Syntax Restrictions On UPDATE, DELETE, and INSERT Statements Within
    Triggers</h2>

<p>^The [UPDATE], [DELETE], and [INSERT]
statements within triggers do not support
the full syntax for [UPDATE], [DELETE], and [INSERT] statements.  The following
restrictions apply:</p>

<ul>
<li><p>
  ^(The name of the table to be modified in an [UPDATE], [DELETE], or [INSERT]
  statement must be an unqualified table name.  In other words, one must
  use just "<i>tablename</i>" not "<i>database</i><b>.</b><i>tablename</i>"
  when specifying the table.)^ </p></li>

<li><p>
 ^For non-TEMP triggers,
  the table to be modified or queried must exist in the
  same database as the table or view to which the trigger is attached.
  ^(TEMP triggers are not subject to the same-database rule.  A TEMP
  trigger is allowed to query or modify any table in any [ATTACH]-ed database.)^
  </p></li>

<li><p>
  ^The "INSERT INTO <i>table</i> DEFAULT VALUES" form of the [INSERT] statement
  is not supported.
  </p></li>

<li><p>
  ^The INDEXED BY and NOT INDEXED clauses are not supported for [UPDATE] and
  [DELETE] statements.
  </p></li>

<li><p>
  ^(The ORDER BY and LIMIT clauses on [UPDATE] and [DELETE] statements are not
  supported.  ORDER BY and LIMIT are not normally supported for [UPDATE] or
  [DELETE] in any context but can be enabled for top-level statements
  using the [SQLITE_ENABLE_UPDATE_DELETE_LIMIT] compile-time option.  However,
  that compile-time option only applies to top-level [UPDATE] and [DELETE]
  statements, not [UPDATE] and [DELETE] statements within triggers.)^
  </p></li>

<li><p>
  ^[common-table-expression|Common table expression] are not supported for
  statements inside of triggers.
  </p></li>
</ul>

<tcl>hd_fragment instead_of_trigger {INSTEAD OF} {INSTEAD OF trigger}</tcl>
<h1>INSTEAD OF triggers</h1>

<p>^Triggers may be created on [views], as well as ordinary tables, by
specifying INSTEAD OF in the CREATE TRIGGER statement. 
^If one or more ON INSERT, ON DELETE
or ON UPDATE triggers are defined on a view, then it is not an
error to execute an INSERT, DELETE or UPDATE statement on the view, 
respectively.  ^Instead,
executing an INSERT, DELETE or UPDATE on the view causes the associated
triggers to fire. ^The real tables underlying the view are not modified
(except possibly explicitly, by a trigger program).</p>

<p>^Note that the [sqlite3_changes()] and [sqlite3_total_changes()] interfaces
do not count INSTEAD OF trigger firings, but the
[count_changes pragma] does count INSTEAD OF trigger firing.</p>

<h1>Some Example Triggers</h1>

<p>^(Assuming that customer records are stored in the "customers" table, and
that order records are stored in the "orders" table, the following
UPDATE trigger
ensures that all associated orders are redirected when a customer changes
his or her address:</p>

<tcl>Example {
CREATE TRIGGER update_customer_address UPDATE OF address ON customers 
  BEGIN
    UPDATE orders SET address = new.address WHERE customer_name = old.name;
  END;
}</tcl>

<p>With this trigger installed, executing the statement:</p>

<tcl>Example {
UPDATE customers SET address = '1 Main St.' WHERE name = 'Jack Jones';
}</tcl>

<p>causes the following to be automatically executed:</p>

<tcl>Example {
UPDATE orders SET address = '1 Main St.' WHERE customer_name = 'Jack Jones';
}</tcl>)^

<p>For an example of an INSTEAD OF trigger, consider the following schema:

<tcl>Example {
CREATE TABLE customer(
  cust_id INTEGER PRIMARY KEY,
  cust_name TEXT,
  cust_addr TEXT
);
CREATE VIEW customer_address AS
   SELECT cust_id, cust_addr FROM customer;
CREATE TRIGGER cust_addr_chng
INSTEAD OF UPDATE OF cust_addr ON customer_address
BEGIN
  UPDATE customer SET cust_addr=NEW.cust_addr
   WHERE cust_id=NEW.cust_id;
END;
}</tcl>

<p>With the schema above, a statement of the form:</p>

<tcl>Example {
UPDATE customer_address SET cust_addr=$new_address WHERE cust_id=$cust_id;
}</tcl>

<p>Causes the customer.cust_addr field to be updated for a specific
customer entry that has customer.cust_id equal to the $cust_id parameter.
Note how the values assigned to the view are made available as field
in the special "NEW" table within the trigger body.</p>

<tcl>hd_fragment undef_before {undefined BEFORE trigger behavior}</tcl>
<h1>Cautions On The Use Of BEFORE triggers</h1>

<p>If a BEFORE UPDATE or BEFORE DELETE trigger modifies or deletes a row
that was to have been updated or deleted, then the result of the subsequent
update or delete operation is undefined.  Furthermore, if a BEFORE trigger
modifies or deletes a row, then it is undefined whether or not AFTER triggers
that would have otherwise run on those rows will in fact run.
</p>

<p>The value of NEW.rowid is undefined in a BEFORE INSERT trigger in which
the rowid is not explicitly set to an integer.</p>

<p>Because of the behaviors described above, programmers are encouraged to
prefer AFTER triggers over BEFORE triggers.</p>

<tcl>hd_fragment raise {RAISE function}</tcl>
<h1>The RAISE() function</h1>

<p>^(A special SQL function RAISE() may be used within a trigger-program,)^
with the following syntax</p> 

<tcl>BubbleDiagram raise-function</tcl>

<p>^(When one of RAISE(ROLLBACK,...), RAISE(ABORT,...) or RAISE(FAIL,...)
is called during trigger-program
execution, the specified [ON CONFLICT] processing is performed
the current query terminates.)^
An error code of [SQLITE_CONSTRAINT] is returned to the application,
along with the specified error message.</p>

<p>^When RAISE(IGNORE) is called, the remainder of the current trigger program,
the statement that caused the trigger program to execute and any subsequent
trigger programs that would have been executed are abandoned. ^No database
changes are rolled back.  ^If the statement that caused the trigger program
to execute is itself part of a trigger program, then that trigger program
resumes execution at the beginning of the next step.
</p>

<tcl>hd_fragment temptrig {TEMP triggers on non-TEMP tables}</tcl>
<h1>TEMP Triggers on Non-TEMP Tables</h1>

<p>^(A trigger normally exists in the same database as the table named
after the "ON" keyword in the CREATE TRIGGER statement.  Except, it is
possible to create a TEMP TRIGGER on a table in another database.)^  
Such a trigger will only fire when changes
are made to the target table by the application that defined the trigger.
Other applications that modify the database will not be able to see the
TEMP trigger and hence cannot run the trigger.</p>

<p>When defining a TEMP trigger on a non-TEMP table, it is important to
specify the database holding the non-TEMP table.  For example,
in the following statement, it is important to say "main.tab1" instead
of just "tab1":</p>

<blockquote><pre>
CREATE TEMP TRIGGER ex1 AFTER INSERT ON <b>main.</b>tab1 BEGIN ...
</pre></blockquote>

<p>Failure to specify the schema name on the target table could result
in the TEMP trigger being reattached to a table with the same name in
another database whenever any schema change occurs.</p>

<title>CREATE VIEW</title>
<tcl>
hd_keywords *createview {CREATE VIEW} view VIEW *views
</tcl>

<fancy_format>
<h1>Syntax</h1>

<tcl>
RecursiveBubbleDiagram create-view-stmt
</tcl>

<h1>Description</h1>

<p>^The CREATE VIEW command assigns a name to a pre-packaged 
[SELECT] statement. 
^Once the view is created, it can be used in the FROM clause
of another [SELECT] in place of a table name.
</p>

<p>^If the "TEMP" or "TEMPORARY" keyword occurs in between "CREATE"
and "VIEW" then the view that is created is only visible to the
[database connection] that created it and is automatically deleted when
the database connection is closed.</p>

<p> ^If a <yyterm>schema-name</yyterm> is specified, then the view 
is created in the specified database.
^It is an error to specify both a <yyterm>schema-name</yyterm>
and the TEMP keyword on a VIEW, unless the <yyterm>schema-name</yyterm> 
is "temp".
^If no schema name is specified, and the TEMP keyword is not present,
the VIEW is created in the main database.</p>

<p>^You cannot [DELETE], [INSERT], or [UPDATE] a view.  ^Views are read-only 
in SQLite.  ^However, in many cases you can use an
[INSTEAD OF trigger] on the view to accomplish 
the same thing.  ^Views are removed 
with the [DROP VIEW] command.</p>

<p>^If a <yyterm>column-name</yyterm> list follows 
the <yyterm>view-name</yyterm>, then that list determines
the names of the columns for the view.  ^If the <yyterm>column-name</yyterm>
list is omitted, then the names of the columns in the view are derived
from the names of the result-set columns in the [select-stmt].
The use of <yyterm>column-name</yyterm> list is recommended.  Or, if
<yyterm>column-name</yyterm> list is omitted, then the result
columns in the [SELECT] statement that defines the view should have
well-defined names using the 
"[result-column|AS column-alias]" syntax.
SQLite allows you to create views that depend on automatically 
generated column names, but you should avoid using them since the 
rules used to generate column names are not a defined part of the
interface and might change in future releases of SQLite.

<p>The <yyterm>column-name</yyterm> list syntax was added in
SQLite versions 3.9.0 ([dateof:3.9.0]).

<tcl>

<title>CREATE VIRTUAL TABLE</title>
<tcl>
hd_keywords *createvtab {CREATE VIRTUAL TABLE}
</tcl>

<fancy_format>

<tcl>
RecursiveBubbleDiagram create-virtual-table-stmt
</tcl>

<p>A [virtual table] is an interface to an external storage or computation
engine that appears to be a table but does not actually store information
in the database file.</p>

<p>In general, you can do anything with a [virtual table] that can be done
with an ordinary table, except that ^you cannot create indices or triggers on a
virtual table.  ^Some virtual table implementations might impose additional
restrictions.  For example, many virtual tables are read-only.</p>

<p>The <yyterm>module-name</yyterm> is the name of an object that implements
the virtual table.  ^The <yyterm>module-name</yyterm> must be registered with
the SQLite database connection using
[sqlite3_create_module()] or [sqlite3_create_module_v2()]
prior to issuing the CREATE VIRTUAL TABLE statement.
^The module takes zero or more comma-separated arguments.
^The arguments can be just about any text as long as it has balanced
parentheses.  The argument syntax is sufficiently general that the
arguments can be made to appear as [column definitions] in a traditional
[CREATE TABLE] statement.  
^SQLite passes the module arguments directly
to the [xCreate] and [xConnect] methods of the module implementation
without any interpretation.  It is the responsibility
of the module implementation to parse and interpret its own arguments.</p>

<p>^A virtual table is destroyed using the ordinary
[DROP TABLE] statement.  There is no
DROP VIRTUAL TABLE statement.</p>

<title>Date And Time Functions</title>
<tcl>
hd_keywords *datefunc {date and time functions} {Date And Time Functions}
hd_keywords {date() SQL function} {time() SQL function}
hd_keywords {datetime() SQL function} {julianday() SQL function}
hd_keywords {strftime() SQL function}
</tcl>

<table_of_contents>
<h1>Overview</h1>

<p>
SQLite supports five date and time functions as follows:
</p>

<p>
<ol>
<li> ^(<b>date(</b><i>timestring, modifier, modifier, ...</i><b>)</b>)^ </li>
<li> ^(<b>time(</b><i>timestring, modifier, modifier, ...</i><b>)</b>)^ </li>
<li> ^(<b>datetime(</b><i>timestring, modifier, modifier, ...</i><b>)</b>)^ </li>
<li> ^(<b>julianday(</b><i>timestring, modifier, modifier, ...</i><b>)</b>)^ </li>
<li> ^(<b>strftime(</b><i>format, timestring, modifier, modifier, ...</i><b>)</b>)^ </li>
</ol>

<p>
^All five date and time functions take a time string as an argument. 
^The time string is followed by zero or more modifiers. 
^The strftime() function also takes a format string as its first argument.
</p>

<p>
The date and time functions use a subset of
[http://en.wikipedia.org/wiki/ISO_8601 | IS0-8601] date and time
formats.
^The date() function returns the date in this format: YYYY-MM-DD. 
^The time() function returns the time as HH:MM:SS. 
^The datetime() function returns "YYYY-MM-DD HH:MM:SS". 
^(The julianday() function returns the 
[http://en.wikipedia.org/wiki/Julian_day | Julian day] - the
number of days since noon in Greenwich on November 24, 4714 B.C. 
([http://en.wikipedia.org/wiki/Proleptic_Gregorian_calendar | Proleptic Gregorian calendar]).)^
^The strftime() routine returns the date formatted according to 
the format string specified as the first argument.
^The format string supports the most common substitutions found in the 
[http://opengroup.org/onlinepubs/007908799/xsh/strftime.html | strftime() function]
from the standard C library plus two new substitutions, %f and %J.
^(The following is a complete list of valid strftime() substitutions:
</p>

<blockquote>
<table border="0" cellpadding="0" cellspacing="0">
<tr><td><td width="10"><td></tr>

<tr><td> %d <td><td> day of month: 00
<tr><td> %f <td><td> fractional seconds: SS.SSS
<tr><td> %H <td><td> hour: 00-24 
<tr><td> %j <td><td> day of year: 001-366
<tr><td> %J <td><td> Julian day number
<tr><td> %m <td><td> month: 01-12
<tr><td> %M <td><td> minute: 00-59
<tr><td> %s <td><td> seconds since 1970-01-01
<tr><td> %S <td><td> seconds: 00-59
<tr><td> %w <td><td> day of week 0-6 with Sunday==0
<tr><td> %W <td><td> week of year: 00-53
<tr><td> %Y <td><td> year: 0000-9999
<tr><td> %% <td><td> %
</table>
</blockquote>)^

<p>
^(Notice that all other date and time functions can be expressed
in terms of strftime():
</p>

<blockquote>
<table border="0" cellpadding="0" cellspacing="0">
<tr><td><b>Function</b><td width="30"><td><b>Equivalent strftime()</b>
<tr><td>   date(...)      <td><td>  strftime('%Y-%m-%d', ...)
<tr><td>   time(...)      <td><td>  strftime('%H:%M:%S', ...)
<tr><td>   datetime(...)  <td><td>  strftime('%Y-%m-%d %H:%M:%S', ...)
<tr><td>   julianday(...) <td><td>  strftime('%J', ...)
</table>
</blockquote>)^

<p>
The only reasons for providing functions other than strftime() is
for convenience and for efficiency.
</p>

<h1>Time Strings</h1>

<p>^(A time string can be in any of the following formats:</p>

<ol>
<li> <i>YYYY-MM-DD</i>
<li> <i>YYYY-MM-DD HH:MM</i>
<li> <i>YYYY-MM-DD HH:MM:SS</i>
<li> <i>YYYY-MM-DD HH:MM:SS.SSS</i>
<li> <i>YYYY-MM-DD</i><b>T</b><i>HH:MM</i>
<li> <i>YYYY-MM-DD</i><b>T</b><i>HH:MM:SS</i>
<li> <i>YYYY-MM-DD</i><b>T</b><i>HH:MM:SS.SSS</i>
<li> <i>HH:MM</i>
<li> <i>HH:MM:SS</i>
<li> <i>HH:MM:SS.SSS</i>
<li> <b>now</b>
<li> <i>DDDDDDDDDD</i>
</ol>)^

<p>
^In formats 5 through 7, the "T" is a literal character separating 
the date and the time, as required by 
[http://www.w3c.org/TR/NOTE-datetime | ISO-8601]. 
^Formats 8 through 10 that specify only a time assume a date of 
2000-01-01. Format 11, the string 'now', is converted into the 
current date and time as obtained from the xCurrentTime method
of the [sqlite3_vfs] object in use.
^The 'now' argument to date and time functions always returns exactly the
same value for multiple invocations within the same [sqlite3_step()] call.
^[http://en.wikipedia.org/wiki/Coordinated_Universal_Time | Universal Coordinated Time (UTC)] is used. 
^Format 12 is the 
[http://en.wikipedia.org/wiki/Julian_day | Julian day number]
expressed as a floating point value.
</p>

<p>
Formats 2 through 10 may be optionally followed by a timezone indicator of the form
"<i>&#91;+-&#93;HH:MM</i>" or just "<i>Z</i>".  The date and time functions use UTC or "zulu"
time internally, and so the "Z" suffix is a no-op.  Any non-zero "HH:MM" suffix is
subtracted from the indicated date and time in order to compute zulu time.
For example, all of the following time strings are equivalent:
</p>

<blockquote>
2013-10-07 08:23:19.120<br>
2013-10-07T08:23:19.120Z<br>
2013-10-07 04:23:19.120-04:00<br>
2456572.84952685
</blockquote>

<p>
^(In formats 4, 7, and 10, the fractional seconds value SS.SSS can have
one or more digits following the decimal point.  Exactly three digits are
shown in the examples because only the first three digits are significant
to the result, but the input string can have fewer or more than three digits
and the date/time functions will still operate correctly.)^
Similarly, format 12 is shown with 10 significant digits, but the date/time
functions will really accept as many or as few digits as are necessary to
represent the Julian day number.
</p>

<h1>Modifiers</h1>

<p>^The time string can be followed by zero or more modifiers that 
alter date and/or time. ^Each modifier
is a transformation that is applied to the time value to its left.
^Modifiers are applied from left to right; order is important.
^(The available modifiers are as follows.</p>

<ol>
<li> NNN days
<li> NNN hours
<li> NNN minutes
<li> NNN.NNNN seconds
<li> NNN months
<li> NNN years
<li> start of month
<li> start of year
<li> start of day
<li> weekday N
<li> unixepoch
<li> localtime
<li> utc 
</ol>)^

<p>^The first six modifiers (1 through 6) 
simply add the specified amount of time to the date and time 
specified by the preceding timestring and modifiers.
^The 's' character at the end of the modifier names is optional.
^Note that "&plusmn;NNN months" works by rendering the original date into
the YYYY-MM-DD format, adding the &plusmn;NNN to the MM month value, then
normalizing the result.  ^Thus, for example, the data 2001-03-31 modified
by '+1 month' initially yields 2001-04-31, but April only has 30 days
so the date is normalized to 2001-05-01.  ^A similar effect occurs when
the original date is February 29 of a leapyear and the modifier is
&plusmn;N years where N is not a multiple of four.</p>

<p>^The "start of" modifiers (7 through 9) shift the date backwards 
to the beginning of the current month, year or day.</p>

<p>^(The "weekday" modifier advances the date forward, if necessary,
to the next date where the weekday number is N. Sunday is 0, Monday is 1,
and so forth.)^
^If the date is already on the desired weekday, the "weekday" modifier
leaves the date unchanged.  </p>

<p>^The "unixepoch" modifier (11) only works if it immediately follows 
a timestring in the DDDDDDDDDD format. 
^This modifier causes the DDDDDDDDDD to be interpreted not 
as a Julian day number as it normally would be, but as
[http://en.wikipedia.org/wiki/Unix_time | Unix Time] - the 
number of seconds since 1970.  If the "unixepoch" modifier does not
follow a timestring of the form DDDDDDDDDD which expresses the number
of seconds since 1970 or if other modifiers
separate the "unixepoch" modifier from prior DDDDDDDDDD then the
behavior is undefined.
For SQLite versions before 3.16.0 ([dateof:3.16.0]), 
the "unixepoch" modifier only works for
dates between 0000-01-01 00:00:00 and 5352-11-01 10:52:47 (unix times
of -62167219200 through 106751991167).</p>

<tcl>hd_fragment localtime {localtime modifier} {'utc' modifier}</tcl>
<p>^The "localtime" modifier (12) assumes the time string to its left is in
Universal Coordinated Time (UTC) and adjusts the time
string so that it displays localtime.  If "localtime"
follows a time that is not UTC, then the behavior is undefined.
^(The "utc" modifier is the opposite of "localtime".  
"utc" assumes that the string
to its left is in the local timezone and adjusts that string to be in UTC.)^
If the prior string is not in localtime, then the result of "utc" is
undefined.</p>

<h1>Examples</h1>

^(<p>Compute the current date.<p>

<blockquote>SELECT date('now');</blockquote>)^

^(<p>Compute the last day of the current month.</p>

<blockquote>SELECT date('now','start of month','+1 month','-1 day');
</blockquote>)^

^(<p>Compute the date and time given a unix timestamp 1092941466.</p>

<blockquote>
    SELECT datetime(1092941466, 'unixepoch');
</blockquote>)^

^(<p>Compute the date and time given a unix timestamp 1092941466, and 
compensate for your local timezone.</p>

<blockquote>
  SELECT datetime(1092941466, 'unixepoch', 'localtime');
</blockquote>)^

^(<p>Compute the current unix timestamp.</p>

<blockquote>
  SELECT strftime('%s','now');
</blockquote>)^

^(<p>Compute the number of days since the signing of the US Declaration
of Independence.</p>

<blockquote>
  SELECT julianday('now') - julianday('1776-07-04');
</blockquote>)^

^(<p>Compute the number of seconds since a particular moment in 2004:</p>

<blockquote>
  SELECT strftime('%s','now') - strftime('%s','2004-01-01 02:34:56');
</blockquote>)^

^(<p>
Compute the date of the first Tuesday in October
for the current year.
</p>

<blockquote>
  SELECT date('now','start of year','+9 months','weekday 2');
</blockquote>)^

^(<p>Compute the time since the unix epoch in seconds 
(like strftime('%s','now') except includes fractional part):</p>

<blockquote>
  SELECT (julianday('now') - 2440587.5)*86400.0;
</blockquote>)^

<h1>Caveats And Bugs</h1>

<p>The computation of local time depends heavily on the whim 
of politicians and is thus difficult to get correct for 
all locales. ^In this implementation, the standard C library 
function localtime_r() is used to assist in the calculation of 
local time.  ^(The 
localtime_r() C function normally only works for years
between 1970 and 2037. For dates outside this range, SQLite 
attempts to map the year into an equivalent year within 
this range, do the calculation, then map the year back.)^</p>


<p>^(These functions only work for dates between 0000-01-01 00:00:00
and 9999-12-31 23:59:59 (julian day numbers 1721059.5 through 5373484.5).)^
For dates outside that range, the results of these
functions are undefined.</p>

<p>Non-Vista Windows platforms only support one set of DST rules. 
Vista only supports two. Therefore, on these platforms, 
historical DST calculations will be incorrect. 
For example, in the US, in 2007 the DST rules changed. 
Non-Vista Windows platforms apply the new 2007 DST rules 
to all previous years as well. Vista does somewhat better
getting results correct back to 1986, when the rules were also changed.</p>

<p>All internal computations assume the 
[http://en.wikipedia.org/wiki/Gregorian_calendar | Gregorian calendar]
system.  It is also assumed that every
day is exactly 86400 seconds in duration.</p>

<title>DELETE</title>
<tcl>
hd_keywords *delete DELETE DELETEs
</tcl>

<table_of_contents>
<h1>Overview</h1>

<tcl>
RecursiveBubbleDiagram delete-stmt
</tcl>

<p>The DELETE command removes records from the table identified by the
   [qualified-table-name].

<p>^If the WHERE clause is not present, all records in the table are deleted.
   ^If a WHERE clause is supplied, then only those rows for which the
   WHERE clause [boolean expression] is true are deleted.
   ^Rows for which the expression is false or NULL are retained.

<h1>Restrictions on DELETE Statements Within CREATE TRIGGER</h1>

<p>The following restrictions apply to DELETE statements that occur within the
   body of a [CREATE TRIGGER] statement:

<ul>
  <li><p>^The <yyterm>table-name</yyterm> specified as part of a 
    DELETE statement within
    a trigger body must be unqualified.  ^(In other words, the
    <i>schema-name</i><b>.</b> prefix on the table name is not allowed 
    within triggers.)^ ^If the table to which the trigger is attached is
    not in the temp database, then DELETE statements within the trigger
    body must operate on tables within the same database as it. ^If the table
    to which the trigger is attached is in the TEMP database, then the
    unqualified name of the table being deleted is resolved in the same way as
    it is for a top-level statement (by searching first the TEMP database, then
    the main database, then any other databases in the order they were
    attached).
    
  <li><p>^The INDEXED BY and NOT INDEXED clauses are not allowed on DELETE
    statements within triggers.</p>

  <li><p>^The LIMIT and ORDER BY clauses (described below) are unsupported for
    DELETE statements within triggers.</p>
</ul>

<h1>Optional LIMIT and ORDER BY clauses</h1>

<p>^(If SQLite is compiled with the [SQLITE_ENABLE_UPDATE_DELETE_LIMIT]
compile-time option, then the syntax of the DELETE statement is
extended by the addition of optional ORDER BY and LIMIT clauses:)^</p>

<tcl>BubbleDiagram delete-stmt-limited</tcl>

<p>^If a DELETE statement has a LIMIT clause, the maximum number of rows that
will be deleted is found by evaluating the accompanying expression and casting
it to an integer value. ^If the result of the evaluating the LIMIT clause
cannot be losslessly converted to an integer value, it is an error. ^A 
negative LIMIT value is interpreted as "no limit". ^(If the DELETE statement 
also has an OFFSET clause, then it is similarly evaluated and cast to an
integer value. Again, it is an error if the value cannot be losslessly
converted to an integer.)^ ^If there is no OFFSET clause, or the calculated
integer value is negative, the effective OFFSET value is zero.

<p>^(If the DELETE statement has an ORDER BY clause, then all rows that would 
be deleted in the absence of the LIMIT clause are sorted according to the 
ORDER BY. The first <i>M</i> rows, where <i>M</i> is the value found by
evaluating the OFFSET clause expression, are skipped, and the following 
<i>N</i>, where <i>N</i> is the value of the LIMIT expression, are deleted.)^
^If there are less than <i>N</i> rows remaining after taking the OFFSET clause
into account, or if the LIMIT clause evaluated to a negative value, then all
remaining rows are deleted.

<p>^If the DELETE statement has no ORDER BY clause, then all rows that
would be deleted in the absence of the LIMIT clause are assembled in an
arbitrary order before applying the LIMIT and OFFSET clauses to determine 
the subset that are actually deleted.

<p>^(The ORDER BY clause on a DELETE statement is used only to determine which
rows fall within the LIMIT. The order in which rows are deleted is arbitrary
and is not influenced by the ORDER BY clause.)^

<tcl>hd_fragment truncateopt {truncate optimization}</tcl>
<h1>The Truncate Optimization</h1>

<p>^When the WHERE is omitted from a DELETE statement and the table
being deleted has no triggers,
SQLite uses an optimization to erase the entire table content
without having to visit each row of the table individually.
This "truncate" optimization makes the delete run much faster.
Prior to SQLite [version 3.6.5] ([dateof:3.6.5]), the truncate optimization
also meant that the [sqlite3_changes()] and
[sqlite3_total_changes()] interfaces
and the [count_changes pragma]
will not actually return the number of deleted rows.  
That problem has been fixed as of [version 3.6.5] ([dateof:3.6.5]).

<p>^The truncate optimization can be permanently disabled for all queries
by recompiling
SQLite with the [SQLITE_OMIT_TRUNCATE_OPTIMIZATION] compile-time switch.</p>

<p>The truncate optimization can also be disabled at runtime using
the [sqlite3_set_authorizer()] interface.  ^If an authorizer callback
returns [SQLITE_IGNORE] for an [SQLITE_DELETE] action code, then
the DELETE operation will proceed but the truncate optimization will
be bypassed and rows will be deleted one by one.</p>

<title>DETACH</title>
<tcl>
hd_keywords *detach *DETACH {DETACH DATABASE}
</tcl>

<fancy_format>
<tcl>
RecursiveBubbleDiagram detach-stmt
</tcl>

<p>^This statement detaches an additional database connection previously 
attached using the [ATTACH] statement.  
^When not in [shared cache mode], 
it is possible to have the same database file attached multiple times using 
different names, and detaching one connection to a file will leave the 
others intact.</p>

<p>
^In [shared cache mode], attempting to attach the same database file more
than once results in an error.

<title>DROP INDEX</title>
<tcl>
hd_keywords *dropindex {DROP INDEX}
</tcl>

<fancy_format>
<tcl>
RecursiveBubbleDiagram drop-index-stmt
</tcl>

<p>^The DROP INDEX statement removes an index added
with the [CREATE INDEX] statement.  The index is completely removed from
the disk.  The only way to recover the index is to reenter the
appropriate [CREATE INDEX] command.</p>

<tcl>

<title>DROP TABLE</title>
<tcl>
hd_keywords *droptable {DROP TABLE}
</tcl>

<fancy_format>
<tcl>
RecursiveBubbleDiagram drop-table-stmt
</tcl>

<p>^The DROP TABLE statement removes a table added with the
[CREATE TABLE] statement.  The name specified is the
table name.  ^The dropped table is completely removed from the database 
schema and the disk file.  The table can not be recovered.  
^All indices and triggers
associated with the table are also deleted.</p>

<p>^The optional IF EXISTS clause suppresses the error that would normally
result if the table does not exist.</p>

<p>^If [foreign key constraints] are enabled, a DROP TABLE command performs an
implicit [DELETE | DELETE FROM] command before removing the
table from the database schema. ^Any triggers attached to the table are
dropped from the database schema before the implicit DELETE FROM
is executed, so this cannot cause any triggers to fire. By contrast, ^an
implicit DELETE FROM does cause any configured
[foreign key actions] to take place. 
^If the implicit DELETE FROM executed
as part of a DROP TABLE command violates any immediate foreign key constraints,
an error is returned and the table is not dropped. ^If 
the implicit DELETE FROM causes any 
deferred foreign key constraints to be violated, and the violations still
exist when the transaction is committed, an error is returned at the time
of commit.

<title>DROP TRIGGER</title>
<tcl>
hd_keywords *droptrigger {DROP TRIGGER}
</tcl>

<tcl>
RecursiveBubbleDiagram drop-trigger-stmt
</tcl>

<p>^The DROP TRIGGER statement removes a trigger created by the 
[CREATE TRIGGER] statement. ^Once removed, the trigger definition is no
longer present in the sqlite_master (or sqlite_temp_master) table and is
not fired by any subsequent INSERT, UPDATE or DELETE statements.

<p>^Note that triggers are automatically dropped when the associated table is
dropped.

<title>DROP VIEW</title>
<tcl>
hd_keywords *dropview {DROP VIEW}
</tcl>

<fancy_format>
<tcl>
RecursiveBubbleDiagram drop-view-stmt
</tcl>

<p>^The DROP VIEW statement removes a view created by the [CREATE VIEW] 
  statement. ^The view definition is removed from the database schema, but 
  no actual data in the underlying base tables is modified.

<p>^(The view to drop is identified by the view-name and optional 
  schema-name specified as part of the DROP VIEW statement. This 
  reference is resolved using the standard procedure for [object resolution].)^

<p>
  ^If the specified view cannot be found and the IF EXISTS clause is not 
  present, it is an error. ^If the specified view cannot be found and an IF
  EXISTS clause is present in the DROP VIEW statement, then the statement
  is a no-op.


<tcl>

<title>EXPLAIN</title>
<tcl>
hd_keywords *explain EXPLAIN
</tcl>

<fancy_format>
<h1>Syntax</h1>
<tcl>
BubbleDiagram sql-stmt
</tcl>

<h1>Description</h1>

<p>^An SQL statement can be preceded by the keyword "EXPLAIN" or
by the phrase "EXPLAIN QUERY PLAN".  ^Either modification causes the
SQL statement to behave as a query and to return information about
how the SQL statement would have operated if the EXPLAIN keyword or
phrase had been omitted.</p>

<p>The output from EXPLAIN and EXPLAIN QUERY PLAN is intended for
interactive analysis and troubleshooting only.  The details of the 
output format are subject to change from one release of SQLite to the next.
Applications should not use EXPLAIN or EXPLAIN QUERY PLAN since
their exact behavior is variable and only partially documented.</p>

<p>^When the EXPLAIN keyword appears by itself it causes the statement
to behave as a query that returns the sequence of 
[virtual machine instructions] it would have used to execute the command had
the EXPLAIN keyword not been present. ^When the EXPLAIN QUERY PLAN phrase
appears, the statement returns high-level information regarding the query
plan that would have been used.

<p>The EXPLAIN QUERY PLAN command is described in 
[explain query plan|more detail here].

<h2>EXPLAIN operates at run-time, not at prepare-time</h2>

<p>The EXPLAIN and EXPLAIN QUERY PLAN prefixes affect the behavior of
running a [prepared statement] using [sqlite3_step()].  The process of
generating a new prepared statement using [sqlite3_prepare()] or similar
is (mostly) unaffected by EXPLAIN.  (The exception to the previous sentence
is that some special opcodes used by EXPLAIN QUERY PLAN are omitted when
building an EXPLAIN QUERY PLAN prepared statement, as a performance
optimization.)

<p>This means that actions that occur during sqlite3_prepare() are
unaffected by EXPLAIN.

<ul>
<li><p>
Some [PRAGMA] statements do their work during sqlite3_prepare() rather
than during sqlite3_step().  Those PRAGMA statements are unaffected
by EXPLAIN.  They operate the same with or without the EXPLAIN prefix.
The set of PRAGMA statements that are unaffected by EXPLAIN can vary
from one release to the next.  Some PRAGMA statements operate during
sqlite3_prepare() depending on their arguments.  For consistent
results, avoid using EXPLAIN on PRAGMA statements.

<li><p>
The [authorizer callback] is invoked regardless of the presence of
EXPLAIN or EXPLAIN QUERY PLAN.
</ul>

<title>SQL Langauge Expressions</title>
<tcl>
hd_keywords *langexpr *expression {expression syntax}
</tcl>

<table_of_contents>
<h1>Syntax</h1>

<tcl>
RecursiveBubbleDiagram expr
</tcl>

<tcl>hd_fragment binaryops {binary operators}</tcl>
<h1>Operators</h1>
<p>^(SQLite understands the following binary operators, in order from
highest to lowest precedence:</p>

<blockquote><pre>
<font color="#2c2cf0"><big>||
*    /    %
+    -
&lt;&lt;   &gt;&gt;   &amp;    |
&lt;    &lt;=   &gt;    &gt;=
=    ==   !=   &lt;&gt;   </big>IS   IS NOT   IN   LIKE   GLOB   MATCH   REGEXP
AND   
OR</font>
</pre></blockquote>)^

<p>^(Supported unary prefix operators are these:</p>

<blockquote><pre>
<font color="#2c2cf0"><big>-    +    ~    NOT</big></font>
</pre></blockquote>)^

<tcl>hd_fragment collateop {COLLATE operator}</tcl>
<p>^The COLLATE operator is a unary postfix
operator that assigns a [collating sequence] to an expression.
^The COLLATE operator has a higher precedence (binds more tightly) than any
binary operator and any unary prefix operator except "~".
(COLLATE and "~" are associative so their binding order does not matter.)
^The collating sequence set by the COLLATE operator overrides the
collating sequence determined by the COLLATE clause in a table
[column definition].
See the [collating sequence| detailed discussion on collating sequences]
in the [datatype | Datatype In SQLite3] document for additional information.
</p>

<tcl>hd_puts "
<p>^The unary operator [Operator +] is a no-op.  ^It can be applied
to strings, numbers, blobs or NULL and it always returns a result
with the same value as the operand.</p>"</tcl>

<p>Note that there are two variations of the equals and not equals
operators.  ^Equals can be either

<tcl>
hd_puts "[Operator =] or [Operator ==].
^The non-equals operator can be either
[Operator !=] or [Operator {&lt;&gt;}].
^The [Operator ||] operator is \"concatenate\" - it joins together
the two strings of its operands.
^The operator [Operator %] outputs the integer value of its left 
operand modulo its right operand.  In other words, the
operator [Operator %] works in SQLite the same as it does
in ANSI-C.</p>

<p>^The result of any binary operator is either a numeric value or 
NULL, except for the [Operator ||] concatenation operator which always 
evaluates to either NULL or a text value.</p>"

hd_fragment {isisnot} {IS operator} {IS NOT operator}

hd_puts "<p>^The [Operator IS] and [Operator {IS NOT}] operators work
like [Operator =] and [Operator !=] except when one or both of the
operands are NULL. ^In this case, if both operands are NULL, then the
IS operator evaluates to 1 (true) and the IS NOT operator evaluates
to 0 (false). ^If one operand is NULL and the other is not, then the
IS operator evaluates to 0 (false) and the IS NOT operator is 1 (true).
^It is not possible for an IS or IS NOT expression to evaluate to NULL.
^Operators [Operator IS] and [Operator {IS NOT}] have the same 
precedence as [Operator =]."

<tcl>hd_fragment litvalue {literal value}</tcl>
<h1>Literal Values (Constants)</h1>
<p>
A literal value represents a constant.
^Literal values may be integers, floating point numbers, strings,
BLOBs, or NULLs.
<p>
The syntax for integer and floating point literals (collectively
"numeric literals") is shown by the following diagram:</p>

<tcl>BubbleDiagram numeric-literal</tcl>

<p>
^(If a numeric literal has a decimal point or an exponentiation
clause or if its magnitude is less than -9223372036854775808 or
greater than 9223372036854775807, then it is a floating point literal.
Otherwise is it is an  integer literal.)^
^The "E" character that begins the exponentiation
clause of a floating point literal can be either upper or lower case.
^(The "." character is always used 
as the decimal point even if the locale setting specifies "," for
this role - the use of "," for the decimal point would result in
syntactic ambiguity.)^

<tcl>hd_fragment hexint {hexadecimal integer literals} {hexadecimal integers}</tcl>
<p>^Hexadecimal integer literals follow the C-language notation of
"0x" or "0X" followed by hexadecimal digits.
^For example, 0x1234 means the same as 4660
and 0x8000000000000000 means the same as -9223372036854775808.
 ^(Hexadecimal integer literals are interpreted as 64-bit
two's-complement integers and are thus limited
to sixteen significant digits of precision.)^
Support for hexadecimal integers was added to SQLite
version 3.8.6 ([dateof:3.8.6]).
^For backwards compatibility, the "0x" hexadecimal integer
notation is only understood by the SQL language parser, not by the
type conversions routines.
^(String variables that
contain text formatted like hexadecimal integers are not
interpreted as hexadecimal integers when coercing the string value
into an integer due to a [CAST expression] or for a [column affinity]
transformation or prior to performing a numeric operation or for
any other run-time conversions.)^  ^(When coercing a
string value in the format of a hexadecimal integer into an integer
value, the conversion process stops when the 'x' character is seen
so the resulting integer value is always zero.)^
^(SQLite only understands the hexadecimal integer notation when it
appears in the SQL statement text, not when it appears as
part of the content of the database.)^

<p> ^A string constant is formed by enclosing the
string in single quotes (').  ^A single quote within the string can
be encoded by putting two single quotes in a row - as in Pascal.
C-style escapes using the backslash character are not supported because
they are not standard SQL.

<p> ^BLOB literals are string literals containing hexadecimal data and
preceded by a single "x" or "X" character.  ^(Example: X'53514C697465')^

<p>
^A literal value can also be the token "NULL".
</p>

<tcl>hd_fragment varparam parameter parameters {bound parameter} {bound parameters}</tcl>
<h1>Parameters</h1>
<p>
A "variable" or "parameter" token
specifies a placeholder in the expression for a 
value that is filled in at runtime using the
[sqlite3_bind_blob() | sqlite3_bind()] family of C/C++ interfaces.
Parameters can take several forms:
</p>

<blockquote>
<table>
<tr>
<td align="right" valign="top"><b>?</b><i>NNN</i></td><td width="20"></td>
<td>^(A question mark followed by a number <i>NNN</i> holds a spot for the
NNN-th parameter.  NNN must be between 1 and [SQLITE_MAX_VARIABLE_NUMBER].)^
</td>
</tr>
<tr>
<td align="right" valign="top"><b>?</b></td><td width="20"></td>
<td>^A question mark that is not followed by a number creates a parameter
with a number one greater than the largest parameter number already assigned.
^If this means the parameter number is greater than
[SQLITE_MAX_VARIABLE_NUMBER], it is an error.
This parameter format is provided for compatibility with other database
engines.  But because it is easy to miscount the question marks, the
use of this parameter format is discouraged.  Programmers are encouraged
to use one of the symbolic formats below or the ?NNN format above instead.
</td>
</tr>
<tr>
<td align="right" valign="top"><b>:</b><i>AAAA</i></td><td width="20"></td>
<td>^A colon followed by an identifier name holds a spot for a 
[sqlite3_bind_parameter_name|named parameter] with the name :AAAA.  
^(Named parameters are also numbered. The number assigned is one greater than
the largest parameter number already assigned.)^ ^If this means the parameter
would be assigned a number greater than [SQLITE_MAX_VARIABLE_NUMBER], it is
an error. To avoid confusion, it is best to avoid mixing named and numbered
parameters.</td>
</tr>
<tr>
<td align="right" valign="top"><b>@</b><i>AAAA</i></td><td width="20"></td>
<td>^An "at" sign works exactly like a colon, except that the name of
the parameter created is @AAAA.</td>
</tr>
<tr>
<td align="right" valign="top"><b>$</b><i>AAAA</i></td><td width="20"></td>
<td>^A dollar-sign followed by an identifier name also holds a spot for a named
parameter with the name $AAAA.  ^(The identifier name in this case can include
one or more occurrences of "::" and a suffix enclosed in "(...)" containing
any text at all.)^  This syntax is the form of a variable name in the
[http://www.tcl-lang.org/ | Tcl programming language].  The presence
of this syntax results from the fact that SQLite is really a 
[Tcl extension] that has escaped into the wild.</td>
</tr>
</table>
</blockquote>

<p>^Parameters that are not assigned values using
[sqlite3_bind_blob() | sqlite3_bind()] are treated
as NULL.  The [sqlite3_bind_parameter_index()] interface can be used
to translate a symbolic parameter name into its equivalent numeric index.</p>

<p>^The maximum parameter number is set at compile-time by
the [SQLITE_MAX_VARIABLE_NUMBER] macro.  ^(An individual [database connection]
D can reduce its maximum parameter number below the compile-time maximum
using the [sqlite3_limit](D, [SQLITE_LIMIT_VARIABLE_NUMBER],...) interface.)^
</p>

<tcl>hd_fragment like LIKE ESCAPE</tcl>
<h1>The LIKE, GLOB, REGEXP, and MATCH operators</h1>
<p>^The LIKE operator does a pattern matching comparison. ^The operand
to the right of the LIKE operator contains the pattern and the left hand
operand contains the string to match against the pattern.

<tcl>hd_puts "^A percent symbol (\"%\") in the LIKE pattern matches any
sequence of zero or more characters in the string.  ^An underscore
(\"_\") in the LIKE pattern matches any single character in the
string.  ^(Any other character matches itself or its lower/upper case
equivalent (i.e. case-insensitive matching).)^

<u>Important Note:</u> ^SQLite only
understands upper/lower case for ASCII characters by default.  ^The
LIKE operator is case sensitive by default for unicode characters that are
beyond the ASCII range.  For example,
^the expression <b>'a'&nbsp;LIKE&nbsp;'A'</b>
is TRUE but <b>'&aelig;'&nbsp;LIKE&nbsp;'&AElig;'</b> is FALSE."</tcl>
The ICU extension to SQLite includes an enhanced version of the
LIKE operator that does case folding across all unicode characters.</p>

<p>^If the optional ESCAPE clause is present, then the expression
following the ESCAPE keyword must evaluate to a string consisting of
a single character. ^This character may be used in the LIKE pattern
to include literal percent or underscore characters. ^The escape
character followed by a percent symbol (%), underscore (_), or a second
instance of the escape character itself matches a
literal percent symbol, underscore, or a single escape character,
respectively.

<p>^The infix LIKE operator is implemented by calling the
application-defined SQL functions [like(<i>Y</i>,<i>X</i>)] or
[like(<i>Y</i>,<i>X</i>,<i>Z</i>)]</a>.</p>

<p>^The LIKE operator can be made case sensitive using the
[case_sensitive_like pragma].</p>

<tcl>hd_fragment glob GLOB</tcl>
<p>^The GLOB operator is similar to LIKE but uses the Unix
file globbing syntax for its wildcards.  ^Also, GLOB is case
sensitive, unlike LIKE.  ^Both GLOB and LIKE may be preceded by
the NOT keyword to invert the sense of the test.  ^The infix GLOB 
operator is implemented by calling the function
[glob(<i>Y</i>,<i>X</i>)] and can be modified by overriding
that function.</p>

<tcl>hd_fragment regexp REGEXP</tcl>
<p>^The REGEXP operator is a special syntax for the regexp()
user function.  ^No regexp() user function is defined by default
and so use of the REGEXP operator will normally result in an
error message.  ^If an [application-defined SQL function] named "regexp"
is added at run-time, then the "<i>X</i> REGEXP <i>Y</i>" operator will
be implemented as a call to "regexp(<i>Y</i>,<i>X</i>)".</p>

<tcl>hd_fragment match MATCH</tcl>
<p>^The MATCH operator is a special syntax for the match()
application-defined function.  ^The default match() function implementation
raises an exception and is not really useful for anything.
^But extensions can override the match() function with more
helpful logic.</p>

<tcl>hd_fragment between BETWEEN</tcl>
<h1>The BETWEEN operator</h1>
<p>^(The BETWEEN operator is logically equivalent to a pair of comparisons.
"<i>x</i> <b>BETWEEN</b> <i>y</i> <b>AND</b> <i>z</i>" is 
equivalent to 
"<i>x</i><b>&gt;=</b><i>y</i> <b>AND</b> <i>x</i><b>&lt;=</b><i>z</i>" except
that with BETWEEN, the <i>x</i> expression is only evaluated once.)^
^The precedence of the BETWEEN operator is the same as the precedence
as operators <b>==</b> and <b>!=</b> and <b>LIKE</b> and groups left to right.

<tcl>hd_fragment case {CASE expression}</tcl>
<h1>The CASE expression</h1>
<p>A CASE expression serves a role similar to IF-THEN-ELSE in other
programming languages.  

<p>The optional expression that occurs in between the CASE keyword and the
first WHEN keyword is called the "base" expression. ^There are two basic forms
of the CASE expression: those with a base expression and those without.

<p>^In a CASE without a base expression, each WHEN expression is evaluated
and the result treated as a boolean, starting with the leftmost and continuing
to the right. ^The result of the CASE expression is the evaluation of the THEN
expression that corresponds to the first WHEN expression that evaluates to
true. ^Or, if none of the WHEN expressions evaluate to true, the result of
evaluating the ELSE expression, if any. ^If there is no ELSE expression and
none of the WHEN expressions are true, then the overall result is NULL.

<p>^A NULL result is considered untrue when evaluating WHEN terms.

<p>^In a CASE with a base expression, the base expression is evaluated just
once and the result is compared against the evaluation of each WHEN 
expression from left to right. ^The result of the CASE expression is the 
evaluation of the THEN expression that corresponds to the first WHEN
expression for which the comparison is true. ^Or, if none of the WHEN
expressions evaluate to a value equal to the base expression, the result
of evaluating the ELSE expression, if any. ^If there is no ELSE expression and
none of the WHEN expressions produce a result equal to the base expression,
the overall result is NULL.

<p>^When comparing a base expression against a WHEN expression, the same
collating sequence, affinity, and NULL-handling rules apply as if the
base expression and WHEN expression are respectively the left- and
right-hand operands of an <big><b>=</b></big> operator.</p> 

<p>^If the base 
expression is NULL then the result of the CASE is always the result 
of evaluating the ELSE expression if it exists, or NULL if it does not.

<p>^Both forms of the CASE expression use lazy, or short-circuit, 
evaluation.

<p>^(The only difference between the following two CASE expressions is that 
the <i>x</i> expression is evaluated exactly once in the first example but 
might be evaluated multiple times in the second:

<ul><pre>
<li>CASE x WHEN w1 THEN r1 WHEN w2 THEN r2 ELSE r3 END
<li>CASE WHEN x=w1 THEN r1 WHEN x=w2 THEN r2 ELSE r3 END
</pre></ul>)^


<tcl>hd_fragment in_op {IN operator} {NOT IN operator}</tcl>
<h1>The IN and NOT IN operators</h1>
<p>^The IN and NOT IN operators take an expression on the
left and a list of values or a subquery on the right.
^When the right operand of an IN or NOT IN operator is a subquery, the
subquery must have the same number of columns as there are columns in
the [row value] of the left operand.  ^The subquery on the
right of an IN or NOT IN operator must be a scalar subquery if the left
expression is not a [row value] expression.
^If the right operand of an IN or NOT IN operator is a list of values,
each of those values must be scalars and the left expression must also
be a scalar.
^The right-hand side of an IN or NOT IN operator can be a
table <i>name</i> or [table-valued function] <i>name</i> in which 
case the right-hand side is understood to be subquery of
the form "(SELECT * FROM <i>name</i>)".
^When the right operand is an empty set, the result of IN is false and the
result of NOT IN is true, regardless of the left operand and even if the
left operand is NULL.
<p>^(The result of an IN or NOT IN operator is determined by the following
matrix:

<blockquote>
<table border=1>
<tr>
<th>Left operand <br>is NULL
<th>Right operand <br>contains NULL
<th>Right operand <br>is an empty set
<th>Left operand found <br>within right operand
<th>Result of <br>IN operator
<th>Result of <br>NOT IN operator
<tr>
<td align="center">no
<td align="center">no
<td align="center">no
<td align="center">no
<td align="center">false
<td align="center">true
<tr>
<td align="center">does not matter
<td align="center">no
<td align="center">yes
<td align="center">no
<td align="center">false
<td align="center">true
<tr>
<td align="center">no
<td align="center">does not matter
<td align="center">no
<td align="center">yes
<td align="center">true
<td align="center">false
<tr>
<td align="center">no
<td align="center">yes
<td align="center">no
<td align="center">no
<td align="center">NULL
<td align="center">NULL
<tr>
<td align="center">yes
<td align="center">does not matter
<td align="center">no
<td align="center">does not matter
<td align="center">NULL
<td align="center">NULL
</table>
</blockquote>)^

<p>^Note that SQLite allows the parenthesized list of scalar values on
the right-hand side of an IN or NOT IN operator to be an empty list but
most other SQL database database engines and the SQL92 standard require
the list to contain at least one element.</p>

<h1>Table Column Names</h1>

<p>^A column name can be any of the names defined in the [CREATE TABLE]
statement or one of the following special identifiers: "<b>ROWID</b>",
"<b>OID</b>", or "<b>_ROWID_</b>".
^The three special identifiers describe the
unique integer key (the [rowid]) associated with every 
row of every table and so are not available on [WITHOUT ROWID] tables.
^The special identifiers only refer to the row key if the [CREATE TABLE]
statement does not define a real column with the same name.
^The rowid can be used anywhere a regular
column can be used.</p>

<tcl>hd_fragment exists_op {EXISTS operator} {NOT EXISTS operator}</tcl>
<h1>The EXISTS operator</h1>

<p>^The EXISTS operator always evaluates to one of the integer values 0 
and 1. ^If executing the SELECT statement specified as the right-hand 
operand of the EXISTS operator would return one or more rows, then the
EXISTS operator evaluates to 1. ^If executing the SELECT would return
no rows at all, then the EXISTS operator evaluates to 0. 

<p>^The number of columns in each row returned by the SELECT statement
(if any) and the specific values returned have no effect on the results
of the EXISTS operator. ^In particular, rows containing NULL values are
not handled any differently from rows without NULL values.

<tcl>hd_fragment subq {Subqueries}</tcl>
<h1>Subquery Expressions</h1>

<p>^A [SELECT] statement enclosed in parentheses is a subquery.
^All types of SELECT statement, including
aggregate and compound SELECT queries (queries with keywords like
UNION or EXCEPT) are allowed as scalar subqueries.
^The value of a subquery expression is the first row of the result
from the enclosed [SELECT] statement.
^The value of a subquery expression is NULL if the enclosed
[SELECT] statement returns no rows.

<p>A subquery that returns a single column is a scalar subquery and can
be used most anywhere.
^A subquery that returns two or more columns is a [row value]
subquery and can only be used as the operand of a comparison operator.

<tcl>hd_fragment cosub {correlated subqueries}</tcl>
<h1>Correlated Subqueries</h1>

<p>^A [SELECT] statement used as either a scalar subquery or as the 
right-hand operand of an IN, NOT IN or EXISTS expression may contain 
references to columns in the outer query. Such a subquery is known as
a correlated subquery. ^A correlated subquery is reevaluated each time
its result is required. ^An uncorrelated subquery is evaluated only once
and the result reused as necessary.

<tcl>hd_fragment castexpr {CAST expression} {CAST} {cast} {CAST operator}</tcl>
<h1>CAST expressions</h1>

<p>A CAST expression of the form "CAST(<i>expr</i> AS <i>type-name</i>)"
is used to convert the value of <i>expr</i> to 
a different [storage class] specified by <yyterm>type-name</yyterm>.
^A CAST conversion is similar to the conversion that takes
place when a [column affinity] is applied to a value except that with
the CAST operator the conversion always takes place even if the conversion
lossy and irreversible, whereas column affinity only changes the data type
of a value if the change is lossless and reversible.

<p>^If the value of <i>expr</i> is NULL, then the result of the CAST
expression is also NULL. ^Otherwise, the storage class of the result
is determined by applying the [rules for determining column affinity] to
the <yyterm>type-name</yyterm>.

<blockquote>
<table border=1>
<tr>
  <th> Affinity of <yyterm><nobr>type-name</nobr></yyterm>
  <th> Conversion Processing
<tr>
  <td> NONE 
  <td> ^Casting a value to a <yyterm>type-name</yyterm> with no affinity 
  causes the value to
  be converted into a BLOB.  ^Casting to a BLOB consists of first casting
  the value to TEXT in the [encoding] of the database connection, then
  interpreting the resulting byte sequence as a BLOB instead of as TEXT.

<tr>
  <td> TEXT
  <td> ^To cast a BLOB value to TEXT, the sequence of bytes that make up the
  BLOB is interpreted as text encoded using the database encoding.
  <p>
   ^Casting an INTEGER or REAL value into TEXT renders the value as if via 
    [sqlite3_snprintf()] except that the resulting TEXT uses the [encoding] of
    the database connection.

<tr>
  <td> REAL
  <td> ^When casting a BLOB value to a REAL, the value is first converted to
        TEXT.
       <p>^When casting a TEXT value to REAL, the longest possible prefix of
        the value that can be interpreted as a real number is extracted from
        the TEXT value and the remainder ignored. ^Any leading spaces in the
        TEXT value are ignored when converging from TEXT to REAL. ^(If there is
        no prefix that can be interpreted as a real number, the result of the
        conversion is 0.0.)^

<tr>
  <td> INTEGER
  <td> ^When casting a BLOB value to INTEGER, the value is first converted to
        TEXT.
       <p>^When casting a TEXT value to INTEGER, the longest possible prefix of
        the value that can be interpreted as an integer number is extracted from
        the TEXT value and the remainder ignored. ^Any leading spaces in the
        TEXT value when converting from TEXT to INTEGER are ignored. ^If there
        is no prefix that can be interpreted as an integer number, the result
        of the conversion is 0.  ^If the prefix integer is greater than
        +9223372036854775807 then the result of the cast is exactly
        +9223372036854775807.  ^Similarly, if the prefix integer is
        less than -9223372036854775808 then the result of the cast is
        exactly -9223372036854775808.

        <p>^When casting to INTEGER, if the text looks like a floating point
        value with an exponent, the exponent will be ignored because it is
        no part of the integer prefix.  ^For example,
        "(CAST '123e+5' AS INTEGER)" results in 123, not in 12300000.

        <p> ^(The CAST operator understands decimal
        integers only &mdash; conversion of [hexadecimal integers] stops 
        at the "x" in the "0x" prefix of the hexadecimal integer string 
        and thus result of the CAST is always zero.)^

      <p>^A cast of a REAL value into an INTEGER results in the integer
      between the REAL value and zero that is closest to the REAL value.
      ^If a REAL is greater than the greatest possible signed
      integer (+9223372036854775807) then the result is the greatest possible
      signed integer and if the REAL is less than the least possible signed
      integer (-9223372036854775808) then the result is the least possible
      signed integer.

      <p>Prior to SQLite version 3.8.2 ([dateof:3.8.2]),
      casting a REAL value greater than
      +9223372036854775807.0 into an integer resulted in the most negative
      integer, -9223372036854775808.  This behavior was meant to emulate the
      behavior of x86/x64 hardware when doing the equivalent cast.

<tr>
  <td> NUMERIC
  <td> ^Casting a TEXT or BLOB value into NUMERIC yields either an INTEGER or
   a REAL result.
   ^If the input text looks like an integer (there is no decimal point nor
   exponent) and the value is small enough to fit in a 64-bit signed integer,
   then the result will be INTEGER.
   ^Input text that looks like floating point (there is a decimal point and/or
   an exponent) and the text describes a value that 
   can be losslessly converted back and forth between IEEE 754 64-bit float and a
   51-bit signed integer, then the result is INTEGER.
   (In the previous sentence, a 51-bit integer is specified since that is one
   bit less than the length of the mantissa of an IEEE 754 64-bit float and
   thus provides a 1-bit of margin for the text-to-float conversion operation.)
   ^Any text input that describes a value outside the range of a 64-bit
   signed integer yields a REAL result.
   <p> ^Casting a REAL or INTEGER value to NUMERIC is a no-op, even if a real
   value could be losslessly converted to an integer.
</tr>
</table>
</blockquote>

<p>^Note that the result from casting any non-BLOB value into a 
BLOB and the result from casting any BLOB value into a non-BLOB value
may be different depending on whether the database [encoding] is UTF-8,
UTF-16be, or UTF-16le.


<tcl>hd_fragment booleanexpr {boolean expression}</tcl>
<h1>Boolean Expressions</h1>

<p>The SQL language features several contexts where an expression is 
evaluated and the result converted to a boolean (true or false) value. These
contexts are:

  <ul>
    <li> the WHERE clause of a SELECT, UPDATE or DELETE statement,
    <li> the ON or USING clause of a join in a SELECT statement,
    <li> the HAVING clause of a SELECT statement,
    <li> the WHEN clause of an SQL trigger, and
    <li> the WHEN clause or clauses of some CASE expressions.
  </ul>

<p>^(To convert the results of an SQL expression to a boolean value, SQLite
first casts the result to a NUMERIC value in the same way as a 
[CAST expression]. A numeric zero value (integer value 0 or real 
value 0.0) is considered to be false.  A NULL value is still NULL.
All other values are considered true.)^

<p>^(For example, the values NULL, 0.0, 0, 'english' and '0' are all considered
to be false.)^ ^(Values 1, 1.0, 0.1, -0.1 and '1english' are considered to 
be true.)^

<p>Beginning with SQLite 3.23.0 ([dateof:3.23.0]), SQLite recognizes the
identifiers "TRUE" and "FALSE" as boolean literals, if and only if those
identifiers are not already used for some other meaning.  If there already
exists columns or tables or other objects named TRUE or FALSE, then for
the sake of backwards compatibility, the TRUE and FALSE identifiers refer
to those other objects, not to the boolean values.

<p>The boolean identifiers TRUE and FALSE are usually just aliases for
the integer values 1 and 0, respectively.  However, if TRUE or FALSE
occur on the right-hand side of an IS operator, then they form new
unary postfix operators "IS TRUE" and "IS FALSE" which test the boolean
value of the operand on the left.

<h1>Functions</h1>
<p>SQLite supports many [corefunc|simple], [aggfunc|aggregate],
and [window functions|window]
SQL functions.  For presentation purposes, simple functions are further
subdivided into [corefunc | core functions], [datefunc|date-time functions],
and [json1|JSON functions].
Applications can add new functions, written in C/C++, using the
[sqlite3_create_function()] interface.
</p>

<p>
The main expression bubble diagram above shows a single syntax for
all function invocations.  But this is merely to simplify the expression
bubble diagram.  In reality, each type of function has a slightly different
syntax, shown below.  The function invocation syntax shown in the main
expression bubble diagram is the union of the three syntaxes shown here:

<tcl>
BubbleDiagram simple-function-invocation
BubbleDiagram aggregate-function-invocation
BubbleDiagram window-function-invocation
</tcl>

<p>
The OVER clause is required for [window functions] and is prohibited
otherwise.  The DISTINCT keyword is only allowed in [aggfunc|aggregate functions].
The FILTER clause may not appear on a [corefunc|simple function].

<p>^It is possible to have an aggregate function with the same name as a
simple function, as long as the number of arguments for the two forms of the
function are different.  ^For example, the [agg_max|max()] function with a
single argument is an aggregate and the [max()] function with two or more
arguments is a simple function.  Aggregate functions can usually also
be used as window functions.

<title>The INDEXED BY Clause</title>
<tcl>
hd_keywords *langindexedby {INDEXED BY} {NOT INDEXED}
</tcl>

<fancy_format>
<h1>How INDEXED BY Works</h1>

<p>^The INDEXED BY phrase forces the [SQLite query planner] to use a
particular named index on a [DELETE], [SELECT], or [UPDATE] statement.
The INDEXED BY phrase is an SQLite extension and
is not portable to other SQL database engines.</p>

<tcl>
RecursiveBubbleDiagram qualified-table-name
</tcl>

<p>^The "INDEXED BY <yyterm>index-name</yyterm>" phrase specifies 
that the named index
must be used in order to look up values on the preceding table.
^If <yyterm>index-name</yyterm> does not exist or cannot be used 
for the query, then the preparation of the SQL statement fails.
^(The "NOT INDEXED" clause specifies that no index shall be used when
accessing the preceding table, including implied indices create by
UNIQUE and PRIMARY KEY constraints.  However, the [rowid]
can still be used to look up entries even when "NOT INDEXED" is specified.)^</p>

<p>Some SQL database engines provide non-standard "hint" mechanisms which
can be used to give the query optimizer clues about what indices it should
use for a particular statement.  The INDEX BY clause of SQLite is 
<em>not</em> a hinting mechanism and it should not be used as such.
^The INDEXED BY clause does not give the optimizer hints about which index
to use; it gives the optimizer a requirement of which index to use.
^If the query optimizer is unable to use the index specified by the
INDEX BY clause, then the query will fail with an error.</p>

<p>The INDEXED BY clause is <em>not</em> intended for use in tuning
the performance of a query.  The intent of the INDEXED BY clause is
to raise a run-time error if a schema change, such as dropping or
creating an index, causes the query plan for a time-sensitive query
to change.  The INDEXED BY clause is designed to help detect
undesirable query plan changes during regression testing.
Application 
developers are admonished to omit all use of INDEXED BY during
application design, implementation, testing, and tuning.  If
INDEXED BY is to be used at all, it should be inserted at the very
end of the development process when "locking down" a design.</p>

<h1>See Also</h1>

<ol>
<li><p>The [query planner checklist] describes steps that application
developers should following to help resolve query planner problems.
Notice the that the use of INDEXED BY is a last resort, to be used only
when all other measures fail.</p>

<li><p>[upluscontrol|The unary "+" operator]
can be used to disqualify terms in the WHERE clause from use by indices.
Careful use of unary + can sometimes help prevent the query planner from
choosing a poor index without restricting it to using one specific index.
Careful placement of unary + operators is a better method for controlling 
which indices are used by a query.</p>

<li><p>The [sqlite3_stmt_status()] C/C++ interface together with the
[SQLITE_STMTSTATUS_FULLSCAN_STEP] and [SQLITE_STMTSTATUS_SORT] verbs
can be used to detect at run-time when an SQL statement is not
making effective use of indices.  Many applications may prefer to
use the [sqlite3_stmt_status()] interface to detect index misuse
rather than the INDEXED BY phrase described here.</p>
</ol>

<title>INSERT</title>
<tcl>
hd_keywords *insert INSERT *INSERTs
</tcl>

<fancy_format>

<tcl>
RecursiveBubbleDiagram insert-stmt
</tcl>

<p>The INSERT statement comes in three basic forms.  
<ol>
<li><p><b>INSERT INTO </b><i>table</i><b> VALUES(...);</b>
<p>^The first form (with the "VALUES" keyword) creates one or more
new rows in
an existing table. ^If the <yyterm>column-name</yyterm> list after
<yyterm>table-name</yyterm> is omitted then the number
of values inserted into each row
must be the same as the number of columns in the table. ^In this case
the result of evaluating the left-most expression from each term of
the VALUES list is inserted into the left-most column of each new row,
and so forth for each subsequent expression. ^If a <yyterm>column-name</yyterm>
list is specified, then the number of values in each term of the
VALUE list must match the number of
specified columns. ^Each of the named columns of the new row is populated
with the results of evaluating the corresponding VALUES expression. ^Table
columns that do not appear in the column list are populated with the 
[default column value] (specified as part of the [CREATE TABLE] statement), or
with NULL if no [default value] is specified.

<li><p><b>INSERT INTO </b><i>table</i><b> SELECT ...;</b>
<p>The second form of the INSERT statement contains a [SELECT] statement
instead of a VALUES clause. ^A new entry is inserted into the table for each
row of data returned by executing the SELECT statement. ^If a column-list is
specified, the number of columns in the result of the SELECT must be the same
as the number of items in the column-list. ^Otherwise, if no column-list is
specified, the number of columns in the result of the SELECT must be the same
as the number of columns in the table. ^Any SELECT statement, including
[compound SELECTs] and SELECT statements with [ORDER BY] and/or [LIMIT] clauses, 
may be used in an INSERT statement of this form.

<p>To avoid a parsing ambiguity, the SELECT statement should always
contain a WHERE clause, even if that clause is simply "WHERE true",
if the [upsert-clause] is present.  Without the WHERE clause, the
parser does not know if the token "ON" is part of a join constraint
on the SELECT, or the beginning of the [upsert-clause].

<li><p><b>INSERT INTO </b><i>table</i><b> DEFAULT VALUES;</b>
<p>The third form of an INSERT statement is with DEFAULT VALUES.
^(The INSERT ... DEFAULT VALUES statement inserts a single new row into the
named table.)^ ^Each column of the new row is populated with its 
[default value], or with a NULL if no default value is specified 
as part of the column definition in the [CREATE TABLE] statement.
The [upsert-clause] is not supported after DEFAULT VALUES.

</ol>

<p>
^The initial "INSERT" keyword can be replaced by
"REPLACE" or "INSERT OR <i>action</i>" to specify an alternative
constraint [ON CONFLICT|conflict resolution algorithm] to use during 
that one INSERT command.
For compatibility with MySQL, ^the parser allows the use of the
single keyword <a href="lang_replace.html">REPLACE</a> as an 
alias for "INSERT OR REPLACE".

<p>^(The optional "<i>schema-name</i><b>.</b>" prefix on the 
<yyterm>table-name</yyterm>
is supported for top-level INSERT statements only.)^  ^The table name must be
unqualified for INSERT statements that occur within [CREATE TRIGGER] statements.
^Similarly, the "DEFAULT VALUES" form of the INSERT statement is supported for
top-level INSERT statements only and not for INSERT statements within
triggers.
<p>

<p>The optional "AS <yyterm>alias</yyterm>" phrase provides an alternative
name for the table into which content is being inserted.  The alias name
can be used within WHERE and SET clauses of the [UPSERT].  If there is no
[upsert-clause], then the <yyterm>alias</yyterm> is pointless, but also
harmless.

<p>See the separate [UPSERT] documentation for the additional trailing
syntax that can cause an INSERT to behave as an UPDATE if the INSERT would
otherwise violate a uniqueness constraint.  The [upsert clause] is not
allowed on an "INSERT ... DEFAULT VALUES".

<title>SQLite Keywords</title>
<tcl>
hd_keywords *langkeywords {SQLite Keywords} keywords \
            {*SQL keyword} {SQL keywords}
set keyword_list [lsort {
   ABORT
   ACTION
   ADD
   AFTER
   ALL
   ALTER
   ALWAYS
   ANALYZE
   AND
   AS
   ASC
   ATTACH
   AUTOINCREMENT
   BEFORE
   BEGIN
   BETWEEN
   BY
   CASCADE
   CASE
   CAST
   CHECK
   COLLATE
   COLUMN
   COMMIT
   CONFLICT
   CONSTRAINT
   CREATE
   CROSS
   CURRENT
   CURRENT_DATE
   CURRENT_TIME
   CURRENT_TIMESTAMP
   DATABASE
   DEFAULT
   DEFERRED
   DEFERRABLE
   DELETE
   DESC
   DETACH
   DISTINCT
   DO
   DROP
   END
   EACH
   ELSE
   ESCAPE
   EXCEPT
   EXCLUSIVE
   EXCLUDE
   EXISTS
   EXPLAIN
   FAIL
   FILTER
   FIRST
   FOLLOWING
   FOR
   FOREIGN
   FROM
   FULL
   GENERATED
   GLOB
   GROUP
   GROUPS
   HAVING
   IF
   IGNORE
   IMMEDIATE
   IN
   INDEX
   INDEXED
   INITIALLY
   INNER
   INSERT
   INSTEAD
   INTERSECT
   INTO
   IS
   ISNULL
   JOIN
   KEY
   LAST
   LEFT
   LIKE
   LIMIT
   MATCH
   NATURAL
   NO
   NOT
   NOTHING
   NOTNULL
   NULL
   NULLS
   OF
   OFFSET
   ON
   OR
   ORDER
   OTHERS
   OUTER
   OVER
   PARTITION
   PLAN
   PRAGMA
   PRECEDING
   PRIMARY
   QUERY
   RAISE
   RANGE
   RECURSIVE
   REFERENCES
   REGEXP
   REINDEX
   RELEASE
   RENAME
   REPLACE
   RESTRICT
   RIGHT
   ROLLBACK
   ROW
   ROWS
   SAVEPOINT
   SELECT
   SET
   TABLE
   TEMP
   TEMPORARY
   THEN
   TIES
   TO
   TRANSACTION
   TRIGGER
   UNBOUNDED
   UNION
   UNIQUE
   UPDATE
   USING
   VACUUM
   VALUES
   VIEW
   VIRTUAL
   WHEN
   WHERE
   WINDOW
   WITH
   WITHOUT
}]
</tcl>

<fancy_format>

<p>The SQL standard specifies a large number of keywords which may not
be used as the names of tables, indices, columns, databases, user-defined
functions, collations, virtual table modules, or any other named object.
The list of keywords is so long that few people can remember them all.
For most SQL code, your safest bet is to never use any English language
word as the name of a user-defined object.</p>

<p>If you want to use a keyword as a name, you need to quote it.  There
are four ways of quoting keywords in SQLite:</p>

<p>
<blockquote>
<table>
<tr>	<td valign="top"><b>'keyword'</b></td><td width="20"></td>
	<td>^A keyword in single quotes is a string literal.</td></tr>

<tr>	<td valign="top"><b>"keyword"</b></td><td></td>
	<td>^A keyword in double-quotes is an identifier.</td></tr>

<tr>	<td valign="top"><b>&#91;keyword&#93;</b></td><td></td>
	<td>^A keyword enclosed in square brackets is 
        an identifier.  This is not standard SQL.  This quoting mechanism
        is used by MS Access and SQL Server and is included in SQLite for
        compatibility.</td></tr>

<tr>	<td valign="top"><b>&#96;keyword&#96;</b></td><td></td>
	<td>^A keyword enclosed in grave accents (ASCII code 96) is 
        an identifier.  This is not standard SQL.  This quoting mechanism
        is used by MySQL and is included in SQLite for
        compatibility.</td></tr>
</table>
</blockquote>
</p>

<p>For resilience when confronted with historical SQL statements, SQLite
will sometimes bend the quoting rules above:</p>

<ul>
<li><p>^If a keyword in single
quotes (ex: <b>'key'</b> or <b>'glob'</b>) is used in a context where
an identifier is allowed but where a string literal is not allowed, then
the token is understood to be an identifier instead of a string literal.
</p></li>

<li><p>^If a keyword in double
quotes (ex: <b>"key"</b> or <b>"glob"</b>) is used in a context where
it cannot be resolved to an identifier but where a string literal
is allowed, then the token is understood to be a string literal instead
of an identifier.</p></li>
</ul>

<p>Programmers are cautioned not to use the two exceptions described in
the previous bullets.  We emphasize that they exist only so that old
and ill-formed SQL statements will run correctly.  Future versions of
SQLite might raise errors instead of accepting the malformed
statements covered by the exceptions above.</p>

<p>
SQLite adds new keywords from time to time when it takes on new features.
So to prevent your code from being broken by future enhancements, you should
normally quote any identifier that is an English language word, even if
you do not have to.
</p>

<p>
The list below shows all possible keywords used by any build of
SQLite regardless of [compile-time options].  
Most reasonable configurations use most or all of these keywords,
but some keywords may be omitted when SQL language features are
disabled.
Applications can use the
[sqlite3_keyword_count()], [sqlite3_keyword_name()], and
[sqlite3_keyword_check()] interfaces to determine the keywords
recognized by SQLite at run-time.
^(Regardless of the compile-time configuration, any identifier that is not on
the following <tcl>hd_puts [llength $keyword_list]</tcl> element
list is not a keyword to the SQL parser in SQLite:
</p>

<div class="columns" style="columns:11em auto;">
<ul style="padding-top:0;">
<tcl>
set lx {}
foreach word [lsort $keyword_list] {
  hd_putsnl "<li>$word</li>"
}
hd_putsnl "</ul></div>)^"
</tcl>

<title>Database Object Name Resolution</title>
<tcl>
hd_keywords *naming {object resolution}
</tcl>

<p>
  In SQLite, a database object (a table, index, trigger or view) is identified
  by the name of the object and the name of the database that it resides in. 
  Database objects may reside in the main database, the temp database, or in
  an [ATTACH|attached database].

<p>
  The syntax of the [DROP TABLE], [DROP INDEX], [DROP VIEW], [DROP TRIGGER],
  [REINDEX], [ALTER TABLE] and many other commands all permit the user to
  specify a database object either by its name alone, or by a combination of
  its name and the name of its database. ^(If no database is specified as part
  of the object reference, then SQLite searches the main, temp and all attached
  databases for an object with a matching name. The temp database is searched
  first, followed by the main database, followed all attached databases in the
  order that they were attached. The reference resolves to the first match
  found.)^ For example:

<pre>^(
      /* Add a table named 't1' to the temp, main and an attached database */
      ATTACH 'file.db' AS aux;
      CREATE TABLE t1(x, y);
      CREATE TEMP TABLE t1(x, y);
      CREATE TABLE aux.t1(x, y);

      DROP TABLE t1;         /* Drop table in temp database */
      DROP TABLE t1;         /* Drop table in main database */
      DROP TABLE t1;         /* Drop table in aux database */
)^</pre>

<p>
  ^If a schema name is specified as part of an object reference, it must be
  either "main", or "temp" or the schema-name of an attached database.
  ^Like other SQL identifiers, schema names are case-insensitive.
  ^If a schema name is specified, then only that one schema is searched for
  the named object.

<p>
  Most object references may only resolve to a specific type of object (for
  example a reference that is part of a DROP TABLE statement may only resolve
  to a table object, not an index, trigger or view). However in some contexts 
  (e.g. [REINDEX]) an object reference may be resolve to more than one type
  of object. ^When searching database schemas for a named object, objects of
  types that cannot be used in the context of the reference are always 
  ignored.

<title>REINDEX</title>
<tcl>
hd_keywords *reindex REINDEX
</tcl>

<fancy_format>
<tcl>
RecursiveBubbleDiagram reindex-stmt
</tcl>

<p>^The REINDEX command is used to delete and recreate indices from scratch.
This is useful when the definition of a collation sequence has changed.
</p>

<p>^If the REINDEX keyword is not followed by a collation-sequence or database 
object identifier, then all indices in all attached databases are rebuilt.

<p>^If the REINDEX keyword is followed by a collation-sequence name, then
all indices in all attached databases that use the named collation sequences
are recreated. 

<p>^Or, if the argument attached to the REINDEX identifies a specific 
database table, then all indices attached to the database table are rebuilt. 
^If it identifies a specific database index, then just that index is recreated.

<p>^For a command of the form "REINDEX <i>name</i>", a match
against <yyterm>collation-name</yyterm> takes precedence over a match
against <yyterm>index-name</yyterm> or <yyterm>table-name</yyterm>.
This ambiguity in the syntax may be avoided by always specifying a
<yyterm>schema-name</yyterm> when reindexing a specific table or index.

<title>REPLACE</title>
<tcl>
hd_keywords *langreplace REPLACE
</tcl>

<fancy_format>
<p>^The REPLACE command is an alias for the "[ON CONFLICT | INSERT OR REPLACE]"
variant of the [INSERT] command.  
This alias is provided for compatibility other SQL database engines.  See the 
[INSERT] command documentation for additional information.</p>

<title>SELECT</title>
<tcl>
hd_keywords select {SELECT query} SELECT
</tcl>
<table_of_contents>

<h1>Overview</h1>
<tcl>
RecursiveBubbleDiagram select-stmt
</tcl>


<p>The SELECT statement is used to query the database.  The
result of a SELECT is zero or more rows of data where each row
has a fixed number of columns.  A SELECT statement does not make
any changes to the database.

<p>The "[select-stmt]" syntax diagram above attempts to show as much of the
SELECT statement syntax as possible in a single diagram, because some readers
find that helpful.  The following "[factored-select-stmt]" is an alternative
syntax diagrams that expresses the same syntax but tries to break the syntax 
down into smaller chunks.

<tcl>
RecursiveBubbleDiagram --initially-hidden factored-select-stmt select-core
</tcl>

<p>Note that there are paths through the syntax diagrams that
are not allowed in practice.  Some examples:
<ul>
<li>A [VALUES] clause can be the first element in a [compound SELECT]
    that uses a [WITH] clause, but a [simple SELECT] that consists of
    just a [VALUES] clause cannot be preceded by a [WITH] clause.
<li>The [WITH] clause must occur on the first SELECT of a [compound SELECT].
    It cannot follow a [compound-operator].
</ul>

<p>These and other similar syntax restrictions are described in the text.

<p>The SELECT statement is the most complicated command in the SQL language.
To make the description easier to follow, some of the passages below describe
the way the data returned by a SELECT statement is determined as a series of
steps. It is important to keep in mind that this is purely illustrative -
in practice neither SQLite nor any other SQL engine is required to follow 
this or any other specific process.

<tcl>hd_fragment simpleselect {simple SELECT}</tcl>
<h1>Simple Select Processing</h1>

<p>The core of a SELECT statement is a "simple SELECT" shown by the
[select-core] and [simple-select-stmt] syntax diagrams below.  
In practice, most SELECT statements are simple SELECT statements.

<tcl>
RecursiveBubbleDiagram simple-select-stmt select-core
</tcl>

<p>Generating the results of a simple SELECT
statement is presented as a four step process in the description below:

<ol>
  <li> <p>[FROM clause] processing: The input data for the simple SELECT is
       determined. The input data is either implicitly a single row with 0
       columns (if there is no FROM clause) or is determined by the FROM
       clause.
  <li> <p>[WHERE clause] processing: The input data is filtered using the WHERE
       clause expression.  
  <li> <p>[GROUP BY|GROUP BY, HAVING and result-column expression] processing: 
       The set of result rows is computed by aggregating the data according to
       any GROUP BY clause and calculating the result-set expressions for the
       rows of the filtered input dataset.  
  <li> <p>[DISTINCT|DISTINCT/ALL keyword] processing: If the query is a "SELECT
       DISTINCT" query, duplicate rows are removed from the set of result rows.
</ol>

<p>There are two types of simple SELECT statement - aggregate and 
non-aggregate queries. ^A simple SELECT statement is an aggregate query if
it contains either a GROUP BY clause or one or more aggregate functions
in the result-set. ^Otherwise, if a simple SELECT contains no aggregate
functions or a GROUP BY clause, it is a non-aggregate query.

<p><b>1. Determination of input data (FROM clause processing).</b>
<tcl>hd_fragment fromclause</tcl>
<tcl>hd_keywords {FROM clause}</tcl>

<p>The input data used by a simple SELECT query is a set of <i>N</i> rows 
each <i>M</i> columns wide.

<p>^(If the FROM clause is omitted from a simple SELECT statement, then the 
input data is implicitly a single row zero columns wide)^ (i.e. <i>N</i>=1 and
<i>M</i>=0).

<p>If a FROM clause is specified, the data on which a simple SELECT query
operates comes from the one or more tables or subqueries (SELECT statements
in parentheses) specified following the FROM keyword. ^A subquery specified
in the <yyterm>table-or-subquery</yyterm> following the FROM clause in a 
simple SELECT statement is
handled as if it was a table containing the data returned by executing the
subquery statement. ^Each column of the subquery has the
[collation|collation sequence] and [affinity] of the corresponding expression
in the subquery statement.

<p>^If there is only a single table or subquery in the FROM
clause, then the input data used by the SELECT statement is the contents of the
named table. ^If there is more than one table or subquery in FROM clause
then the contents of all tables and/or subqueries
are joined into a single dataset for the simple SELECT statement to operate on.
Exactly how the data is combined depends on the specific [join-operator] and
[join-constraint] used to connect the tables or subqueries together.

<p>All joins in SQLite are based on the cartesian product of the left and
right-hand datasets. ^The columns of the cartesian product dataset are, in 
order, all the columns of the left-hand dataset followed by all the columns
of the right-hand dataset. ^There is a row in the cartesian product dataset
formed by combining each unique combination of a row from the left-hand 
and right-hand datasets. ^(In other words, if the left-hand dataset consists of
<i>N<sub><small>left</small></sub></i> rows of 
<i>M<sub><small>left</small></sub></i> columns, and the right-hand dataset of
<i>N<sub><small>right</small></sub></i> rows of
<i>M<sub><small>right</small></sub></i> columns, then the cartesian product is a
dataset of 
<i>N<sub><small>left</small></sub>&times;N<sub><small>right</small></sub></i>
rows, each containing 
<i>M<sub><small>left</small></sub>+M<sub><small>right</small></sub></i> columns.)^

<p>^If the join-operator is "CROSS JOIN", "INNER JOIN", "JOIN" or a comma
(",") and there is no ON or USING clause, then the result of the join is
simply the cartesian product of the left and right-hand datasets. 
If join-operator does have ON or USING clauses, those are handled according to
the following bullet points:

<ul>
  <li> <p>^(If there is an ON clause then the ON expression is
       evaluated for each row of the cartesian product as a 
       [boolean expression]. Only rows for which the expression evaluates to 
       true are included from the dataset.)^

  <li> <p>^If there is a USING clause
       then each of the column names specified must exist in the datasets to 
       both the left and right of the join-operator. ^(For each pair of named
       columns, the expression "lhs.X = rhs.X" is evaluated for each row of
       the cartesian product as a [boolean expression]. Only rows for which
       all such expressions evaluates to true are included from the
       result set.)^ ^When comparing values as a result of a USING clause, the
       normal rules for handling affinities, collation sequences and NULL
       values in comparisons apply. ^The column from the dataset on the
       left-hand side of the join-operator is considered to be on the left-hand
       side of the comparison operator (=) for the purposes of collation 
       sequence and affinity precedence.

       <p>^For each pair of columns identified by a USING clause, the column
       from the right-hand dataset is omitted from the joined dataset. ^This 
       is the only difference between a USING clause and its equivalent ON
       constraint.

  <li> <p>^(If the NATURAL keyword is in the join-operator then an
       implicit USING clause is added to the join-constraints. The implicit
       USING clause contains each of the column names that appear in both
       the left and right-hand input datasets.)^ ^If the left and right-hand
       input datasets feature no common column names, then the NATURAL keyword
       has no effect on the results of the join. ^A USING or ON clause may
       not be added to a join that specifies the NATURAL keyword.

  <li> <p>^(If the join-operator is a "LEFT JOIN" or "LEFT OUTER JOIN", then
       after
       the ON or USING filtering clauses have been applied, an extra row is 
       added to the output for each row in the original left-hand input 
       dataset that corresponds to no rows at all in the composite
       dataset (if any).)^ ^The added rows contain NULL values in the columns
       that would normally contain values copied from the right-hand input
       dataset.  
</ul>

<p>^(When more than two tables are joined together as part of a FROM clause,
the join operations are processed in order from left to right. In other 
words, the FROM clause (A join-op-1 B join-op-2 C) is computed as 
((A join-op-1 B) join-op-2 C).)^

<tcl>hd_fragment crossjoin {treats the CROSS JOIN operator specially}</tcl>
<p><b>Side note: Special handling of CROSS JOIN.</b>
^There is no difference between the "INNER JOIN", "JOIN" and "," join
operators. They are completely interchangeable in SQLite.
^(The "CROSS JOIN" join operator produces the same result as the 
"INNER JOIN", "JOIN" and "," operators)^, but is 
[CROSS JOIN|handled differently by the query optimizer]
in that it prevents the query optimizer from reordering
the tables in the join.  An application programmer can use the CROSS JOIN 
operator to directly influence the algorithm that is chosen to implement
the SELECT statement.  Avoid using CROSS JOIN except in specific situations 
where manual control of the query optimizer is desired.  Avoid using
CROSS JOIN early in the development of an application as doing so is
a <a href="http://c2.com/cgi/wiki?PrematureOptimization">premature
optimization</a>.  The special handling of CROSS JOIN is an SQLite-specific
feature and is not a part of standard SQL.
       

<tcl>hd_fragment whereclause</tcl>
<tcl>hd_keywords {WHERE clause}</tcl>
<p><b>2. WHERE clause filtering.</b>

<p>^(If a WHERE clause is specified, the WHERE expression is evaluated for 
each row in the input data as a [boolean expression]. Only rows for which the
WHERE clause expression evaluates to true are included from the dataset before
continuing.)^  Rows are excluded from the result if the WHERE clause
evaluates to either false or NULL.

<p>For a JOIN or INNER JOIN or CROSS JOIN, there is no difference between 
a constraint expression in the WHERE clause and one in the ON clause.  However,
for a LEFT JOIN or LEFT OUTER JOIN, the difference is very important.  
In a LEFT JOIN,
the extra NULL row for the right-hand table is added after ON clause processing
but before WHERE clause processing.  A constraint of the form "left.x=right.y"
in an ON clause will therefore allow through the added all-NULL rows of the
right table.  But if that same constraint is in the WHERE clause a NULL in
"right.y" will prevent the expression "left.x=right.y" from being true, and
thus exclude that row from the output.

<p><b>3. Generation of the set of result rows.</b>
<tcl>hd_fragment resultset</tcl>
<tcl>hd_keywords {result-set expressions} {GROUP BY}</tcl>

<p>Once the input data from the FROM clause has been filtered by the
WHERE clause expression (if any), the set of result rows for the simple 
SELECT are calculated. Exactly how this is done depends on whether the simple 
SELECT is an aggregate or non-aggregate query, and whether or not a GROUP
BY clause was specified.

<p> The list of expressions between the SELECT and FROM keywords is known as
the result expression list.  ^If a result expression is the special expression
"*" then all columns in the input data are substituted for that one expression.
^(If the expression is the alias of a table or subquery in the FROM clause
followed by ".*" then all columns from the named table or subquery are
substituted for the single expression.)^ ^(It is an error to use a "*" or
"alias.*" expression in any context other than a result expression list.)^
^(It is also an error to use a "*" or "alias.*" expression in a simple SELECT
query that does not have a FROM clause.)^

<p> ^(The number of columns in the rows returned by a simple SELECT statement
is equal to the number of expressions in the result expression list after
substitution of * and alias.* expressions.)^ Each result row is calculated by
evaluating the expressions in the result expression list with respect to a
single row of input data or, for aggregate queries, with respect to a group
of rows.

<ul>
  <li><p>^If the SELECT statement is <b>a non-aggregate query</b>, then 
    each expression in the result expression list is evaluated for each row in
    the dataset filtered by the WHERE clause.

  <li><p>^If the SELECT statement is <b>an aggregate query without a GROUP
    BY</b> clause, then each aggregate expression in the result-set is 
    evaluated once across the entire dataset. ^Each non-aggregate expression
    in the result-set is evaluated once for an arbitrarily selected row of
    the dataset. ^The same arbitrarily selected row is used for each
    non-aggregate expression. ^Or, if the dataset contains zero rows, then 
    each non-aggregate expression is evaluated against a row consisting
    entirely of NULL values.

   <p>^The single row of result-set data created by evaluating the aggregate
    and non-aggregate expressions in the result-set forms the result of an
    aggregate query without a GROUP BY clause. ^An aggregate query without a
    GROUP BY clause always returns exactly one row of data, even if there are
    zero rows of input data.

  <li><p>^(If the SELECT statement is <b>an aggregate query with a GROUP
    BY</b> clause, then each of the expressions specified as part of the
    GROUP BY clause is evaluated for each row of the dataset. Each row
    is then assigned to a "group" based on the results; rows for which
    the results of evaluating the GROUP BY expressions are the same get
    assigned to the same group.)^ ^For the purposes of grouping rows, NULL 
    values are considered equal. ^The usual rules for [collation|selecting a
    collation sequence] with which to compare text values apply when evaluating
    expressions in a GROUP BY clause.  ^The expressions in the GROUP BY clause
    do <em>not</em> have to be expressions that appear in the result. ^The
    expressions in a GROUP BY clause may not be aggregate expressions.

    <p>^(If a HAVING clause is specified, it is evaluated once for each group 
    of rows as a [boolean expression]. If the result of evaluating the
    HAVING clause is false, the group is discarded.)^ ^If the HAVING clause is
    an aggregate expression, it is evaluated across all rows in the group. ^If
    a HAVING clause is a non-aggregate expression, it is evaluated with respect
    to an arbitrarily selected row from the group.  ^The HAVING expression may
    refer to values, even aggregate functions, that are not in the result.</p>

    <p>^Each expression in the result-set is then evaluated once for each
    group of rows. ^If the expression is an aggregate expression, it is 
    evaluated across all rows in the group. ^Otherwise, it is evaluated against
    a single arbitrarily chosen row from within the group. ^If there is more
    than one non-aggregate expression in the result-set, then all such
    expressions are evaluated for the same row.

    <p>^Each group of input dataset rows contributes a single row to the 
    set of result rows. ^Subject to filtering associated with the DISTINCT
    keyword, the number of rows returned by an aggregate query with a GROUP
    BY clause is the same as the number of groups of rows produced by applying
    the GROUP BY and HAVING clauses to the filtered input dataset.
</ul>

<tcl>hd_fragment bareagg {bare aggregate terms}</tcl>
<p><b>Side note: Bare columns in an aggregate queries.</b>
The usual case is that all column names in an aggregate query are either
arguments to [aggfunc|aggregate functions] or else appear in the GROUP BY clause.
A result column which contains a column name that is not within an
aggregate function and that does not appear in the GROUP BY clause (if
one exists) is called a "bare" column.
Example:
<blockquote><pre>
SELECT a, b, sum(c) FROM tab1 GROUP BY a;
</pre></blockquote>
<p>In the query above, the "a" column is part of the GROUP BY clause
and so each row of the output contains one of the distinct values for
"a".  The "c" column is contained within the [sum()] aggregate function
and so that output column is the sum of all "c" values in rows that
have the same value for "a".  But what is the result of the bare 
column "b"? The answer is that the "b" result will be the
value for "b" in one of the input rows that form the aggregate.
The problem is that you usually do not know which input row is used
to compute "b", and so in many cases the value for "b" is undefined.
</p>
<p>
Special processing occurs when the aggregate function is 
either [maxAggFunc|min()] or [minAggFunc|max()].  Example:
<blockquote><pre>
SELECT a, b, max(c) FROM tab1 GROUP BY a;
</pre></blockquote>

<p>When the [maxAggFunc|min()] or [minAggFunc|max()] aggregate functions 
are used in an aggregate query, all bare columns in the result set 
take values from the input
row which also contains the minimum or maximum.  So in the query
above, the value of the "b" column in the output will be the value of
the "b" column in the input row that has the largest "c" value.
There is still an ambiguity if two or more of the input rows have the
same minimum or maximum value or if the query contains more than one
min() and/or max() aggregate function. Only the built-in [minAggFunc|min()] 
and [maxAggFunc|max()] functions work this way.
</p>

<p><b>4. Removal of duplicate rows (DISTINCT processing).</b>
<tcl>hd_fragment distinct</tcl>
<tcl>hd_keywords {DISTINCT}</tcl>

<p>^One of the ALL or DISTINCT keywords may follow the SELECT keyword in a 
simple SELECT statement. ^If the simple SELECT is a SELECT ALL, then the
entire set of result rows are returned by the SELECT. ^If neither ALL or
DISTINCT are present, then the behavior is as if ALL were specified. 
^If the simple SELECT is a SELECT DISTINCT, then duplicate rows are removed
from the set of result rows before it is returned. ^For the purposes of
detecting duplicate rows, two NULL values are considered to be equal. ^The
[how collation is determined|usual rules] apply for selecting a collation
sequence to compare text values.

<tcl>
hd_fragment compound {compound select} {compound query} {compound SELECT} {compound SELECTs}
</tcl>
<h1>Compound Select Statements</h1>

<p>Two or more [simple SELECT] statements may be connected together to form
a compound SELECT using the UNION, UNION ALL, INTERSECT or EXCEPT operator,
as shown by the following diagram:

<tcl>
RecursiveBubbleDiagram compound-select-stmt
</tcl>

<p>^In a compound SELECT, all the constituent SELECTs must return the same 
number of result columns. ^As the components of a compound SELECT must
be simple SELECT statements, they may not contain [ORDER BY] or [LIMIT] clauses.
^[ORDER BY] and [LIMIT] clauses may only occur at the end of the entire compound
SELECT, and then only if the final element of the compound is not a [VALUES] clause.

<p>^A compound SELECT created using UNION ALL operator returns all the rows
from the SELECT to the left of the UNION ALL operator, and all the rows
from the SELECT to the right of it. ^The UNION operator works the same way as
UNION ALL, except that duplicate rows are removed from the final result set.
^The INTERSECT operator returns the intersection of the results of the left and
right SELECTs.  ^The EXCEPT operator returns the subset of rows returned by the
left SELECT that are not also returned by the right-hand SELECT. ^Duplicate
rows are removed from the results of INTERSECT and EXCEPT operators before the
result set is returned.

<p>^For the purposes of determining duplicate rows for the results of compound
SELECT operators, NULL values are considered equal to other NULL values and
distinct from all non-NULL values. ^The collation sequence used to compare 
two text values is determined as if the columns of the left and right-hand
SELECT statements were the left and right-hand operands of the equals (=)
operator, except that greater precedence is not assigned to a collation 
sequence specified with the postfix COLLATE operator. ^No affinity
transformations are applied to any values when comparing rows as part of a
compound SELECT. 

<p>^(When three or more simple SELECTs are connected into a compound SELECT,
they group from left to right. In other words, if "A", "B" and "C" are all
simple SELECT statements, (A op B op C) is processed as ((A op B) op C).)^

</p>

<tcl>hd_fragment orderby {order by} {ORDER BY}</tcl>
<h1>The ORDER BY clause</h1>

<p>If a SELECT statement that returns more than one row does not have an
ORDER BY clause, the order in which the rows are returned is undefined.
Or, if a SELECT statement does have an ORDER BY clause, then the list of
expressions attached to the ORDER BY determine the order in which rows
are returned to the user.

<p>
^In a [compound SELECT] statement, only the last or right-most [simple SELECT]
may have an ORDER BY clause. ^That ORDER BY clause will apply across all elements of
the compound. ^If the right-most element of a [compound SELECT] is a [VALUES] clause,
then no ORDER BY clause is allowed on that statement.


<p>^Rows are first sorted based on the results of
evaluating the left-most expression in the ORDER BY list, then ties are broken
by evaluating the second left-most expression and so on. The order in which
two rows for which all ORDER BY expressions evaluate to equal values are
returned is undefined. ^Each ORDER BY expression may be optionally followed
by one of the keywords ASC (smaller values are returned first) or DESC (larger
values are returned first). ^If neither ASC or DESC are specified, rows
are sorted in ascending (smaller values first) order by default.

<tcl>hd_fragment nullslast {NULLS FIRST} {NULLS LAST}</tcl>
<p>SQLite considers NULL values to be smaller than any other values
for sorting purposes.  Hence, NULLs naturally appear at the beginning
of an ASC order-by and at the end of a DESC order-by.  This can be changed
using the "ASC NULLS LAST" or "DESC NULLS FIRST" syntax.

<p>Each ORDER BY expression is processed as follows:</p>

<ol>
<li><p>^If the ORDER BY expression is a constant integer K then the
expression is considered an alias for the K-th column of the result set
(columns are numbered from left to right starting with 1).

<li><p>^If the ORDER BY expression is an identifier that corresponds to
the alias of one of the output columns, then the expression is considered
an alias for that column.

<li><p>^Otherwise, if the ORDER BY expression is any other expression, it 
is evaluated and the returned value used to order the output rows. ^If
the SELECT statement is a simple SELECT, then an ORDER BY may contain any
arbitrary expressions. ^However, if the SELECT is a compound SELECT, then
ORDER BY expressions that are not aliases to output columns must be exactly
the same as an expression used as an output column.
</ol>

<p>^For the purposes of sorting rows, values are compared in the same way
as for [comparison expressions]. The collation sequence used to compare
two text values is determined as follows:

<ol>
  <li><p>^If the ORDER BY expression is assigned a collation sequence using
  the postfix [COLLATE operator], then the specified collation sequence is
  used.
  <li><p>^Otherwise, if the ORDER BY expression is an alias to an expression
  that has been assigned a collation sequence using the postfix 
  [COLLATE operator], then the collation sequence assigned to the aliased
  expression is used.
  <li><p>^Otherwise, if the ORDER BY expression is a column or an alias of
  an expression that is a column, then the default collation sequence for
  the column is used. 
  <li><p>^Otherwise, the [BINARY] collation sequence is used.
</ol>

<p>In a [compound SELECT] statement, all ORDER BY expressions are handled
as aliases for one of the result columns of the compound.
^(If an ORDER BY expression is not an integer alias, then SQLite searches
the left-most SELECT in the compound for a result column that matches either
the second or third rules above. If a match is found, the search stops and
the expression is handled as an alias for the result column that it has been
matched against. Otherwise, the next SELECT to the right is tried, and so on.)^
^If no matching expression can be found in the result columns of any
constituent SELECT, it is an error. ^Each term of the ORDER BY clause is
processed separately and may be matched against result columns from different
SELECT statements in the compound.</p>

<tcl>hd_fragment limitoffset {LIMIT} {OFFSET}</tcl>
<h1>The LIMIT clause</h1>

<p>^The LIMIT clause is used to place an upper bound on the number of rows
returned by the entire SELECT statement.

<p>^In a [compound SELECT], only the
last or right-most [simple SELECT] may contain a LIMIT clause.  
^In a [compound SELECT], 
the LIMIT clause applies to the entire compound, not just the final SELECT.
^If the right-most [simple SELECT] is a [VALUES clause] then no LIMIT clause
is allowed.

<p>^Any scalar expression may be used in the 
LIMIT clause, so long as it evaluates to an integer or a value that can be
losslessly converted to an integer. ^If the expression evaluates to a NULL 
value or any other value that cannot be losslessly converted to an integer, an
error is returned. ^If the LIMIT expression evaluates to a negative value,
then there is no upper bound on the number of rows returned. ^Otherwise, the
SELECT returns the first N rows of its result set only, where N is the value
that the LIMIT expression evaluates to. ^Or, if the SELECT statement would
return less than N rows without a LIMIT clause, then the entire result set is
returned. 

<p>^The expression attached to the optional OFFSET clause that may follow a
LIMIT clause must also evaluate to an integer, or a value that can be
losslessly converted to an integer. ^If an expression has an OFFSET clause,
then the first M rows are omitted from the result set returned by the SELECT
statement and the next N rows are returned, where M and N are the values that
the OFFSET and LIMIT clauses evaluate to, respectively. ^Or, if the SELECT
would return less than M+N rows if it did not have a LIMIT clause, then the
first M rows are skipped and the remaining rows (if any) are returned. ^If the
OFFSET clause evaluates to a negative value, the results are the same as if it
had evaluated to zero.

<p>^Instead of a separate OFFSET clause, the LIMIT clause may specify two
scalar expressions separated by a comma. ^In this case, the first expression
is used as the OFFSET expression and the second as the LIMIT expression.
This is counter-intuitive, as when using the OFFSET clause the second of
the two expressions is the OFFSET and the first the LIMIT. 
This reversal of the offset and limit is intentional
- it maximizes compatibility with other SQL database systems.
However, to avoid confusion, programmers are strongly encouraged to use
the form of the LIMIT clause that uses the "OFFSET" keyword and avoid
using a LIMIT clause with a comma-separated offset.

<tcl>hd_fragment values {VALUES clause} VALUES</tcl>
<h1>The VALUES clause</h1>

<p>^(The phrase "VALUES(<i>expr-list</i>)" means the same thing
as "SELECT <i>expr-list</i>".)^  ^(The phrase
"VALUES(<i>expr-list-1</i>),...,(<i>expr-list-N</i>)" means the same
thing as "SELECT <i>expr-list-1</i> UNION ALL ... UNION ALL
SELECT <i>expr-list-N</i>".)^  Both forms are the same, except that
the number of SELECT statements in a compound is limited by
[SQLITE_LIMIT_COMPOUND_SELECT] whereas the number of rows in a
VALUES clause has no arbitrary limit.

<p>There are some restrictions on the use of a VALUES clause that are
not shown on the syntax diagrams:

<ul>
<li><p>
^A VALUES clause cannot be followed by [ORDER BY].

<li><p>
^A VALUES clause cannot be followed by [LIMIT].
</ul>


<h1>The WITH Clause</h1>

<p>^SELECT statements may be optionally preceded by a single
[WITH clause] that defines one or more [common table expressions]
for use within the SELECT statement.

<tcl>hd_fragment tabfunc1 {table-valued functions in the FROM clause}</tcl>
<h1>Table-valued Functions In The FROM Clause</h1>

<p>A [virtual table] that contains [hidden columns] can be used like
a [table-valued function] in the FROM clause.  The arguments to the
table-valued function become constraints on the HIDDEN columns of
the virtual table.  Additional information can be found in the
[table-valued functions|virtual table documentation].

<title>UPDATE</title>
<tcl>
hd_keywords *langupdate UPDATE *UPDATEs
</tcl>

<table_of_contents>
<h1>Overview</h1>

<tcl>
RecursiveBubbleDiagram update-stmt
</tcl>

<p>^An UPDATE statement is used to modify a subset of the values stored in 
zero or more rows of the database table identified by the 
[qualified-table-name] specified as part of the UPDATE statement.

<h1>Details</h1>

<p>^If the UPDATE statement does not have a WHERE clause, all rows in the
table are modified by the UPDATE. ^Otherwise, the UPDATE affects only those
rows for which the WHERE clause
[boolean expression|boolean expression is true]. ^It is not an error if the
WHERE clause does not evaluate to true for any row in the table - this just
means that the UPDATE statement affects zero rows.

<p>The modifications made to each row affected by an UPDATE statement are
determined by the list of assignments following the SET keyword. Each
assignment specifies a <yyterm>column-name</yyterm> to the left of the 
equals sign and a scalar expression to the right. 
^For each affected row, the named columns
are set to the values found by evaluating the corresponding scalar 
expressions. ^If a single column-name appears more than once in the list of
assignment expressions, all but the rightmost occurrence is ignored. ^Columns
that do not appear in the list of assignments are left unmodified. ^The scalar
expressions may refer to columns of the row being updated. ^In this case all
scalar expressions are evaluated before any assignments are made.

<p>Beginning in SQLite [version 3.15.0] ([dateof:3.15.0]), an assignment in
the SET clause can be a 
[column-name-list|parenthesized list of column names] on the left and a
[row value] of the same size on the right.


<p>^The optional "OR <i>action</i>" conflict clause that follows the
UPDATE keyword allows the user to nominate a specific
constraint conflict resolution algorithm to use during this one UPDATE command.
Refer to the section entitled [ON CONFLICT] for additional information.

<h2>Restrictions on UPDATE Statements Within CREATE TRIGGER</h2>

<p>The following additional syntax restrictions apply to UPDATE statements that
occur within the body of a [CREATE TRIGGER] statement.  

<ul>
  <li><p>^The <yyterm>table-name</yyterm> specified as part of an UPDATE 
      statement within
      a trigger body must be unqualified. ^(In other words, the
      <i>schema-name</i><b>.</b> prefix on the table name of the UPDATE is
      not allowed within triggers.)^ ^Unless the table to which the trigger
      is attached is in the TEMP database, the table being updated by the
      trigger program must reside in the same database as it. ^If the table
      to which the trigger is attached is in the TEMP database, then the
      unqualified name of the table being updated is resolved in the same way
      as it is for a top-level statement (by searching first the TEMP database,
      then the main database, then any other databases in the order they were
      attached).

  <li><p>^The INDEXED BY and NOT INDEXED clauses are not allowed on UPDATE
      statements within triggers.</p>

  <li><p>^The LIMIT and ORDER BY clauses for UPDATE are unsupported within
      triggers, regardless of the compilation options used to build SQLite.
</ul>

<h2>Optional LIMIT and ORDER BY Clauses</h2>

<p>^(If SQLite is built with the [SQLITE_ENABLE_UPDATE_DELETE_LIMIT]
compile-time option then the syntax of the UPDATE statement is extended
with optional ORDER BY and LIMIT clauses)^ as follows:</p>

<tcl>BubbleDiagram update-stmt-limited</tcl>

<p>^If an UPDATE statement has a LIMIT clause, the maximum number of rows that
will be updated is found by evaluating the accompanying expression and casting
it to an integer value. ^A negative value is interpreted as "no limit".

<p>^If the LIMIT expression evaluates to non-negative value <i>N</i> and the
UPDATE statement has an ORDER BY clause, then all rows that would be updated in
the absence of the LIMIT clause are sorted according to the ORDER BY and the
first <i>N</i> updated. ^(If the UPDATE statement also has an OFFSET clause,
then it is similarly evaluated and cast to an integer value. If the OFFSET
expression evaluates to a non-negative value <i>M</i>, then the first <i>M</i>
rows are skipped and the following <i>N</i> rows updated instead.)^

<p>^If the UPDATE statement has no ORDER BY clause, then all rows that
would be updated in the absence of the LIMIT clause are assembled in an
arbitrary order before applying the LIMIT and OFFSET clauses to determine 
which are actually updated.

<p>^(The ORDER BY clause on an UPDATE statement is used only to determine which
rows fall within the LIMIT. The order in which rows are modified is arbitrary
and is not influenced by the ORDER BY clause.)^

<title>UPSERT</title>
<tcl>
hd_keywords *upsert UPSERT {upsert clause}
</tcl>

<fancy_format>
<h1>Syntax</h1>
<tcl>
RecursiveBubbleDiagram upsert-clause
</tcl>

<h1>Description</h1>

<p>UPSERT is a special syntax addition to [INSERT] that causes the
INSERT to behave as an [UPDATE] or a no-op if the INSERT would violate
a uniqueness constraint.
UPSERT is not standard SQL.  UPSERT in SQLite follows the
syntax established by PostgreSQL.
UPSERT syntax was added to SQLite with version 3.24.0 ([dateof:3.24.0]).

<p>An UPSERT is an ordinary [INSERT] statement that is followed by
the special ON CONFLICT clause shown above.

<p>The syntax that occurs in between the "ON CONFLICT" and "DO" keywords
is called the "conflict target".  The conflict target specifies a specific
uniqueness constraint that will trigger the upsert.  The conflict target
is required for DO UPDATE upserts, but is optional for DO NOTHING.  When
the conflict target is omitted, the upsert behavior is triggered by a
violation of any uniqueness constraint on the table of the INSERT.

<p>If the insert operation would cause the uniqueness constraint identified
by the conflict-target clause to fail, then the insert is omitted and
either the DO NOTHING or DO UPDATE operation is performed instead.
In the case of a multi-row insert, this decision is made separately
for each row of the insert.

<p>The special UPSERT processing happens only for uniqueness constraint on
the table that is receiving the INSERT. A "uniqueness constraint"
is an explicit UNIQUE or PRIMARY KEY constraint within
the CREATE TABLE statement, or a [unique index].
UPSERT does not intervene for failed NOT NULL or foreign key constraints
or for constraints that are implemented using triggers.

<p>Column names in the expressions of a DO UPDATE refer to the original
unchanged value of the column, before the attempted INSERT.  To use the
value that would have been inserted had the constraint not failed,
add the special "excluded." table qualifier to the column name.

<p>Some examples will help illustrate the difference:

<blockquote><pre>
CREATE TABLE vocabulary(word TEXT PRIMARY KEY, count INT DEFAULT 1);
INSERT INTO vocabulary(word) VALUES('jovial')
  ON CONFLICT(word) DO UPDATE SET count=count+1;
</pre></blockquote>

<p>The upsert above inserts the new vocabulary word "jovial" if that
word is not already in the dictionary, or if it is already in the
dictionary, it increments the counter.  The "count+1" expression
could also be written as "vocabulary.count".  PostgreSQL requires the
second form, but SQLite accepts either.

<blockquote><pre>
CREATE TABLE phonebook(name TEXT PRIMARY KEY, phonenumber TEXT);
INSERT INTO phonebook(name,phonenumber) VALUES('Alice','704-555-1212')
  ON CONFLICT(name) DO UPDATE SET phonenumber=excluded.phonenumber;
</pre></blockquote>

<p>In the second example, the expression in the DO UPDATE clause is
of the form "excluded.phonenumber".  The "excluded." prefix causes the
"phonenumber" to refer to the value for phonenumber that would have been
inserted had there been no conflict.  Hence, the effect of the upsert
is to insert a phonenumber of Alice if none exists, or to overwrite
any prior phonenumber for Alice with the new one.

<p>Note that the DO UPDATE clause acts only on the single row
that experienced the constraint error during INSERT.  It is not
necessary to include a WHERE clause that restrictions the action
to that one row.  The only use for the WHERE clause at
the end of the DO UPDATE is to optionally change the DO UPDATE
into a no-op depending on the original and/or new values.
For example:

<blockquote><pre>
CREATE TABLE phonebook2(
  name TEXT PRIMARY KEY,
  phonenumber TEXT,
  validDate DATE
);
INSERT INTO phonebook2(name,phonenumber,validDate)
  VALUES('Alice','704-555-1212','2018-05-08')
  ON CONFLICT(name) DO UPDATE SET
    phonenumber=excluded.phonenumber,
    validDate=excluded.validDate
  WHERE excluded.validDate>phonebook2.validDate;
</pre></blockquote>

<p>In this last example, the phonebook2 entry is only
updated if the validDate for the newly inserted value is
newer than the entry already in the table.  If the table already
contains an entry with the same name and a current validDate,
then the WHERE clause causes the DO UPDATE to become a no-op.

<tcl>hd_fragment parseambig {UPSERT parsing ambiguity}</tcl>

<h2>Parsing Ambiguity</h2>

<p>When the [INSERT] statement to which the UPSERT is attached
takes its values from a [SELECT] statement, there is a potential
parsing ambiguity.  The parser might not be able to tell if the
"ON" keyword is introducing the UPSERT or if it is the ON clause
of a join.  To work around this, the SELECT statement should always
include a WHERE clause, even if that WHERE clause is just
"WHERE true".

<p>Ambiguous use of ON:

<blockquote><pre>
INSERT INTO t1 SELECT * FROM t2
ON CONFLICT(x) DO UPDATE SET y=excluded.y;
</pre></blockquote>

<p>Ambiguity resolved using a WHERE clause:

<blockquote><pre>
INSERT INTO t1 SELECT * FROM t2 <font color="blue">WHERE true</font>
ON CONFLICT(x) DO UPDATE SET y=excluded.y;
</pre></blockquote>

<h1>Limitations</h1>

<p>UPSERT does not currently work for [virtual tables].

<title>VACUUM</title>
<tcl>
hd_keywords *langvacuum vacuum VACUUM
</tcl>
<fancy_format>

<h1>Syntax</h1>
<tcl>
RecursiveBubbleDiagram vacuum-stmt
</tcl>

<h1>Description</h1>
<p>
  The VACUUM command rebuilds the database file, repacking it into a minimal
  amount of disk space. There are several 
  reasons an application might do this:

<ul>
  <li> <p> ^Unless SQLite is running in "auto_vacuum=FULL" mode, when a large
     amount of data is deleted from the database file it leaves behind empty
     space, or "free" database pages. This means the database file might
     be larger than strictly necessary. ^Running VACUUM to rebuild the 
     database reclaims this space and reduces the size of the database file.

  <li> <p> ^Frequent inserts, updates, and deletes can cause the database file
     to become fragmented - where data for a single table or index is scattered 
     around the database file. ^Running VACUUM ensures that each table and
     index is largely stored contiguously within the database file. ^In some
     cases, VACUUM may also reduce the number of partially filled pages in
     the database, reducing the size of the database file further.

  <li> <p> When content is deleted from an SQLite database, the content is not
     usually erased but rather the space used to hold the content is marked as
     being available for reuse.  This can allow deleted content to be recovered
     by a hacker or by forensic analysis.  Running VACUUM will clean the database
     of all traces of deleted content, thus preventing an adversary from recovering
     deleted content.  Using VACUUM in this way is an alternative to setting
     [PRAGMA secure_delete=ON]. 

  <li> <p> ^Normally, the database [page_size] and whether or not the database
     supports [auto_vacuum] must be configured before the database file is
     actually created. ^However, when not in [write-ahead log] mode, the 
     [page_size] and/or [auto_vacuum] properties of an existing database may be
     changed by using the [page_size pragma|page_size]  and/or 
     [auto_vacuum|pragma auto_vacuum] pragmas and then immediately VACUUMing
     the database. ^When in [write-ahead log] mode, only the [auto_vacuum]
     support property can be changed using VACUUM.
</ul>

<p>^By default, VACUUM only works only on the main database.
^[ATTACH|Attached databases] can be vacuumed by appending the appropriate
<yyterm>schema-name</yyterm> to the VACUUM statement.

<p><b>Compatibility Warning:</b> The ability to vacuum attached databases was
added in [version 3.15.0] ([dateof:3.15.0]).  Prior to that, a 
<yyterm>schema-name</yyterm> added to the
VACUUM statement would be silently ignored and the "main" schema would be
vacuumed.</p>

<tcl>hd_fragment vacuuminto {VACUUM INTO}</tcl>
<h2>VACUUM with an INTO clause</h2>

<p>^If the INTO clause is included, then the original database file is
unchanged and a new database is created in the filename given by the
argument to the INTO clause.  The new database will contain the same
logical content as the original database, fully vacuumed.

<p>
The VACUUM command with an INTO clause is an alternative to the
[backup API] for generating backup copies of a live database.
The advantage of using VACUUM INTO is that the resulting backup
database is minimal in size and hence the amount of filesystem
I/O may be reduced.  Also, all deleted content is purged from the
backup, leaving behind no forensic traces.  On the other hand,
the [backup API] uses fewer CPU cycles and can be executed
incrementally.

<p>
^The filename in the INTO clause can be an arbitrary SQL expression
that evaluates to a string.
^The file named by the INTO clause must not previously exist, or
else it must be an empty file, or the VACUUM INTO command will
fail with an error.

<p>
^The argument to INTO can be a [URI filename] if URI filenames
are enabled.
URL filenames are enabled if any of the following are true:
<ul>
<li> The SQLite library was compiled with [-DSQLITE_USE_URI=1].
<li> The [sqlite3_config]([SQLITE_CONFIG_URI],1) interfaces was
     invoked at start-time.
<li> The [database connection] that is running the VACUUM INTO
     statement was originally opened using the
     [SQLITE_OPEN_URI] flag.
</ul>

<p>
The VACUUM INTO command is transactional in the sense that
the generated output database is a consistent snapshot of the
original database.  However, if the VACUUM INTO command is
interrupted by an unplanned shutdown or power lose, then
the generated output database might be incomplete and corrupt.
Also, SQLite does not invoke fsync() or FlushFileBuffers()
on the generated database to ensure that it has reached
non-volatile storage before completing.


<tcl>hd_fragment howvacuumworks {how vacuum works}</tcl>
<h1>How VACUUM works</h1>

<p>^The VACUUM command works by copying the contents of the database into
a temporary database file and then overwriting the original with the 
contents of the temporary file. ^When overwriting the original, a rollback
journal or [write-ahead log] WAL file is used just as it would be for any
other database transaction. ^This means that when VACUUMing a database, 
as much as twice the size of the original database file is required in free
disk space.

<p>^The VACUUM INTO command works the same way except that it uses the file
named on the INTO clause in place of the temporary database and omits the
step of copying the vacuumed database back over top of the original database.

<p>^The VACUUM command may change the [ROWID | ROWIDs] of entries in any
tables that do not have an explicit [INTEGER PRIMARY KEY].
</p>

<p>^A VACUUM will fail if there is an open transaction on the database
connection that is attempting to run the VACUUM.  ^Unfinalized SQL
statements typically hold a read transaction open, so the VACUUM
might fail if there are unfinalized SQL statements on the same connection.
^VACUUM (but not VACUUM INTO) is a write operation and so if another
database connection is holding a lock that prevents writes, then
the VACUUM will fail.

<p>^(An alternative to using the VACUUM command to
reclaim space after data has been deleted is auto-vacuum mode, enabled using
the [auto_vacuum] pragma.)^ ^When [auto_vacuum] is enabled for a database
free pages may be reclaimed after deleting data, causing the file to shrink,
without rebuilding the entire database using VACUUM.  ^However, using
[auto_vacuum] can lead to extra database file fragmentation.  ^And [auto_vacuum]
does not compact partially filled pages of the database as VACUUM does.

</p>

<title>The WITH Clause</title>
<tcl>
hd_keywords *with {common table expressions} WITH {WITH clause}
</tcl>
<table_of_contents>

<h1>Overview</h1>
<tcl>
RecursiveBubbleDiagram with-clause
</tcl>

<p>Common Table Expressions or CTEs act like temporary [views] that exist
only for the duration of a single SQL statement.  There are two kinds of
common table expressions: "ordinary" and "recursive". Ordinary 
common table expressions are helpful for making
queries easier to understand by factoring
subqueries out of the main SQL statement.
Recursive common table expressions
provide the ability to do hierarchical or
recursive queries of trees and graphs, a capability
that is not otherwise available in the SQL language.

<p>All common table expressions (ordinary and recursive) are 
created by prepending a WITH clause in front of a [SELECT], [INSERT], [DELETE],
or [UPDATE] statement.  A single WITH clause can specify one or more
common table expressions, some of which are ordinary and some of which
are recursive.

<tcl>hd_fragment ordinarycte {ordinary common table expressions}</tcl>
<h1>Ordinary Common Table Expressions</h1>

<p>An ordinary common table expression works as if it were a [view] that
exists for the duration of a single statement.  Ordinary common table
expressions are useful for factoring out subqueries and making the overall
SQL statement easier to read and understand.

<p>A WITH clause can contain ordinary common table expressions even if
it includes the RECURSIVE keyword.  The use of RECURSIVE does not force
common table expressions to be recursive.

<tcl>
hd_fragment recursivecte {recursive common table expressions} \
{recursive common table expression} {recursive query}
</tcl>
<h1>Recursive Common Table Expressions</h1>

<p>A recursive common table expression can be used to write a query that
walks a tree or graph.  A recursive common table expression has the same
basic syntax as an ordinary common table expression, but with the following
additional features:

<ol>
<li> The "[select-stmt]"
     must be a [compound select] where the right-most [compound-operator] is
     either UNION or UNION ALL.
<li> The table named on the left-hand side of the AS keyword must appear
     exactly once in the FROM clause of the right-most SELECT statement
     of the compound select, and nowhere else.
<li> The right-most SELECT of the compound select must not make use of
     [Aggregate Functions|aggregate] or [window functions].
</ol>

<p>To put it another way, a recursive common table expression must
look like the following:

<tcl>RecursiveBubbleDiagram recursive-cte</tcl>

<p>Call the table named by the [cte-table-name] in a recursive
common table expression the "recursive table".
In the [recursive-cte] bubble diagram above, the recursive
table must appear exactly once in the FROM clause of the 
<yyterm>recursive-select</yyterm>
and must not appear anywhere else in either the
<yyterm>initial-select</yyterm> or the
<yyterm>recursive-select</yyterm>, including subqueries.
The <yyterm>initial-select</yyterm> may be
a [compound select], but it may not include an ORDER BY, LIMIT, or OFFSET.
The recursive-select must be a simple select, not a compound.  The
recursive-select is allowed to include an ORDER BY, LIMIT, and/or OFFSET.

<p>The basic algorithm for computing the content of the recursive table
is as follows:

<ol>
<li> Run the <yyterm>initial-select</yyterm> and add the results to a queue.
<li> While the queue is not empty:
<ol type="a">
<li> Extract a single row from the queue.
<li> Insert that single row into the recursive table
<li> Pretend that the single row just extracted is the only
     row in the recursive table and run the recursive-select,
     adding all results to the queue.
</ol>
</ol>

<p>The basic procedure above may modified by the following additional rules:

<ul>
<li><p>
  If a UNION operator connects the <yyterm>initial-select</yyterm> with the
  <yyterm>recursive-select</yyterm>, then only add rows to the queue if 
  no identical row has
  been previously added to the queue.  Repeated rows are discarded before being
  added to the queue even if the repeated rows have already been extracted
  from the queue by the recursion step.  If the operator is UNION ALL,
  then all rows generated by both the <yyterm>initial-select</yyterm> and the
  <yyterm>recursive-select</yyterm> are always added to the queue even if
  they are repeats.
  When determining if a row is repeated, NULL values compare
  equal to one another and not equal to any other value.
<li><p>
  The LIMIT clause, if present, determines the maximum number of rows that
  will ever be added to the recursive table in step 2b.
  Once the limit is reached, the recursion stops.
  A limit of zero means that no rows are ever added to the
  recursive table, and a negative limit means an unlimited number of rows
  may be added to the recursive table.
<li><p>
  The OFFSET clause, if it is present and has a positive value N, prevents the
  first N rows from being added to the recursive table.
  The first N rows are still processed
  by the <yyterm>recursive-select</yyterm> &mdash; they
  just are not added to the recursive table.  Rows are not counted toward
  fulfilling the LIMIT until all OFFSET rows have been skipped.
<li><p>
  If an ORDER BY clause is present, it determines the order in which rows
  are extracted from the queue in step 2a.  If there is no ORDER BY clause,
  then the order in which rows are extracted is undefined.  (In the current
  implementation, the queue becomes a FIFO if the ORDER BY clause is omitted,
  but applications should not depend on that fact since it might change.)
</ul>

<tcl>hd_fragment rcex1</tcl>
<h2>Recursive Query Examples</h2>

<p>The following query returns all integers between 1 and 1000000:

<blockquote><pre>
WITH RECURSIVE
  cnt(x) AS (VALUES(1) UNION ALL SELECT x+1 FROM cnt WHERE x<1000000)
SELECT x FROM cnt;
</pre></blockquote>

<p>Consider how this query works.  The initial-select
runs first and returns a single row
with a single column "1".  This one row is added to the queue.  In
step 2a, that one row is extracted from the queue and added to "cnt".
Then the recursive-select is run in accordance with step 2c generating
a single new row with value "2" to add to the queue.  The queue still
has one row, so step 2 repeats.  The "2" row is extracted and added to the
recursive table by steps 2a and 2b.  Then the row containing 2 is used 
as if it were the complete content of the recursive table and the 
recursive-select is run again, resulting in a row with value "3" being added
to the queue.  This repeats 999999 times until finally at step 2a the
only value on the queue is a row containing 1000000.  That row is
extracted and added to the recursive table.  But this time, the
WHERE clause causes the recursive-select to return no rows, so the
queue remains empty and the recursion stops.

<p><b>Optimization note:</b>
In the discussion above, statements like "insert the row into
the recursive table" should be understood conceptually, not literally.
It sounds as if SQLite is accumulating a huge table
containing one million rows, then going back and scanning that table
from top to bottom to generate the result.  What really happens
is that the query optimizer sees that values in the
"cnt" recursive table are only used once.  So as each row is added to
the recursive table, that row is immediately returned as a result of the main
SELECT statement and then discarded.  SQLite does <em>not</em> accumulate
a temporary table containing a million rows.  Very little memory is
needed to run the above example.  However, if the example had used
UNION instead of UNION ALL, then SQLite would have had to keep around
all previously generated content in order to check for duplicates.
For this reason, programmers should strive to use UNION ALL instead
of UNION when feasible.

<p>Here is a variation on the previous example:

<blockquote><pre>
WITH RECURSIVE
  cnt(x) AS (
     SELECT 1
     UNION ALL
     SELECT x+1 FROM cnt
      LIMIT 1000000
  )
SELECT x FROM cnt;
</pre></blockquote>

<p>There are two differences in this variation.  The initial-select is
"SELECT 1" instead of "VALUES(1)".  But those are just different
syntaxes for saying exactly the same thing.  The other change is that the
recursion is stopped by a LIMIT rather than a WHERE clause.  The use of
LIMIT means that when the one-millionth row is added to the "cnt" table
(and returned by the main SELECT, thanks to the query optimizer)
then the recursion stops immediately regardless of how many rows might be
left in the queue.  On more complex queries, it can sometimes be
difficult to ensure that the WHERE clause will eventually cause the
queue to drain and the recursion to terminate.  But the LIMIT clause will
always stop the recursion.  So it is good practice to always include a
LIMIT clause as a safety if an upper bound on the size of the recursion 
is known.

<tcl>hd_fragment rcex2</tcl>
<h2>Hierarchical Query Examples</h2>

<p>Consider a table that describes the members of an organization as
well as the chain-of-command within that organization:

<blockquote><pre>
CREATE TABLE org(
  name TEXT PRIMARY KEY,
  boss TEXT REFERENCES org,
  height INT,
  -- other content omitted
);
</pre></blockquote>

<p>Every member in the organization has a name, and most members have
a single boss.  (The head of the whole organization has a NULL
"boss" field.) The rows of the "org" table form a tree.

<p>Here is a query that computes the average height over everyone
in Alice's organization, including Alice:

<blockquote><pre>
WITH RECURSIVE
  works_for_alice(n) AS (
    VALUES('Alice')
    UNION
    SELECT name FROM org, works_for_alice
     WHERE org.boss=works_for_alice.n
  )
SELECT avg(height) FROM org
 WHERE org.name IN works_for_alice;
</pre></blockquote>

<p>The next example uses two 
common table expressions in a single WITH clause.  
The following table records a family tree:

<blockquote><pre>
CREATE TABLE family(
  name TEXT PRIMARY KEY,
  mom TEXT REFERENCES family,
  dad TEXT REFERENCES family,
  born DATETIME,
  died DATETIME, -- NULL if still alive
  -- other content
);
</pre></blockquote>

<p>The "family" table is similar to the earlier "org" table except that 
now there are two parents to each member.
We want to know all living ancestors of Alice, from oldest to youngest.
An ordinary common table expression, "parent_of", is defined first.  That
ordinary CTE is a view that can be used to find all parents of any
individual.  That ordinary CTE is then used in the "ancestor_of_alice"
recursive CTE.  The recursive CTE is then used in the final query:

<blockquote><pre>
WITH RECURSIVE
  parent_of(name, parent) AS
    (SELECT name, mom FROM family UNION SELECT name, dad FROM family),
  ancestor_of_alice(name) AS
    (SELECT parent FROM parent_of WHERE name='Alice'
     UNION ALL
     SELECT parent FROM parent_of JOIN ancestor_of_alice USING(name))
SELECT family.name FROM ancestor_of_alice, family
 WHERE ancestor_of_alice.name=family.name
   AND died IS NULL
 ORDER BY born;
</pre></blockquote>

<tcl>hd_fragment rcex3</tcl>
<h2>Queries Against A Graph</h2>

<p>A version control system (VCS) will typically store the evolving
versions of a project as a directed acyclic graph (DAG).  Call each
version of the project a "checkin".  A single
checkin can have zero or more parents.  Most checkins (except the
first) have a single parent, but in the case of a merge, a checkin
might have two or three or more parents.  A schema to keep track of
checkins and the order in which they occur might look something like
this:

<blockquote><pre>
CREATE TABLE checkin(
  id INTEGER PRIMARY KEY,
  mtime INTEGER -- timestamp when this checkin occurred
);
CREATE TABLE derivedfrom(
  xfrom INTEGER NOT NULL REFERENCES checkin, -- parent checkin
  xto INTEGER NOT NULL REFERENCES checkin,   -- derived checkin
  PRIMARY KEY(xfrom,xto)
);
CREATE INDEX derivedfrom_back ON derivedfrom(xto,xfrom);
</pre></blockquote>

<p>This graph is acyclic.  And we assume that the mtime of every
child checkin is no less than the mtime of all its parents.  But
unlike the earlier examples, this graph might have multiple paths of
differing lengths between any two checkins.

<p>We want to know the twenty most recent ancestors in time (out of
the thousands and thousands of ancestors in the whole DAG) for
checkin "@BASELINE".  (A query similar to this is used
by the <a href="http://www.fossil-scm.org/">Fossil</a> VCS to
show the N most recent ancestors of a check.  For example:
<a href="http://www.sqlite.org/src/timeline?p=trunk&n=30">http://www.sqlite.org/src/timeline?p=trunk&n=30</a>.)

<blockquote><pre>
WITH RECURSIVE
  ancestor(id,mtime) AS (
    SELECT id, mtime FROM checkin WHERE id=@BASELINE
    UNION
    SELECT derivedfrom.xfrom, checkin.mtime
      FROM ancestor, derivedfrom, checkin
     WHERE ancestor.id=derivedfrom.xto
       AND checkin.id=derivedfrom.xfrom
     ORDER BY checkin.mtime DESC
     LIMIT 20
  )
SELECT * FROM checkin JOIN ancestor USING(id);
</pre></blockquote>

<p>
The "ORDER BY checkin.mtime DESC" term in the recursive-select makes
the query run much faster by preventing it from following
branches that merge checkins
from long ago.  The ORDER BY forces the recursive-select to focus
on the most recent checkins, the ones we want.  Without the ORDER BY
on the recursive-select, one would be forced to compute the complete set of
thousands of ancestors, sort them all by mtime, then take the top twenty.
The ORDER BY essentially sets up a priority queue that
forces the recursive query to look at the most recent ancestors first,
allowing the use of a LIMIT clause to restrict the scope of the
query to just the checkins of interest.

<tcl>hd_fragment withorderby</tcl>
<h2>Controlling Depth-First Versus Breadth-First Search Of a Tree
Using ORDER BY</h2>

<p>An ORDER BY clause on the recursive-select can be used to control
whether the search of a tree is depth-first or breadth-first.  To
illustrate, we will use a variation on the "org" table from an example
above, without the "height" column, and with some real data inserted:

<blockquote><pre>
CREATE TABLE org(
  name TEXT PRIMARY KEY,
  boss TEXT REFERENCES org
) WITHOUT ROWID;
INSERT INTO org VALUES('Alice',NULL);
INSERT INTO org VALUES('Bob','Alice');
INSERT INTO org VALUES('Cindy','Alice');
INSERT INTO org VALUES('Dave','Bob');
INSERT INTO org VALUES('Emma','Bob');
INSERT INTO org VALUES('Fred','Cindy');
INSERT INTO org VALUES('Gail','Cindy');
</pre></blockquote>

<p>Here is a query to show the tree structure in a breadth-first pattern:

<blockquote><pre>
WITH RECURSIVE
  under_alice(name,level) AS (
    VALUES('Alice',0)
    UNION ALL
    SELECT org.name, under_alice.level+1
      FROM org JOIN under_alice ON org.boss=under_alice.name
     ORDER BY 2
  )
SELECT substr('..........',1,level*3) || name FROM under_alice;
</pre></blockquote>

<p>The "ORDER BY 2" (which means the same as "ORDER BY under_alice.level+1")
causes higher levels in the organization chart (with smaller "level" values)
to be processed first, resulting in a breadth-first search.  The output is:

<blockquote><pre>
Alice
...Bob
...Cindy
......Dave
......Emma
......Fred
......Gail
</pre></blockquote>

<p>But if we change the ORDER BY clause to add the "DESC" modifier, that will
cause lower levels in the organization (with larger "level" values) to be
processed first by the recursive-select, resulting in a depth-first search:

<blockquote><pre>
WITH RECURSIVE
  under_alice(name,level) AS (
    VALUES('Alice',0)
    UNION ALL
    SELECT org.name, under_alice.level+1
      FROM org JOIN under_alice ON org.boss=under_alice.name
     ORDER BY 2 <b>DESC</b>
  )
SELECT substr('..........',1,level*3) || name FROM under_alice;
</pre></blockquote>

<p>The output of this revised query is:

<blockquote><pre>
Alice
...Bob
......Dave
......Emma
...Cindy
......Fred
......Gail
</pre></blockquote>

<p>When the ORDER BY clause is omitted from the recursive-select, the
queue behaves as a FIFO, which results in a breadth-first search.


<tcl>hd_fragment mandelbrot {compute the Mandelbrot set}</tcl>
<h2>Outlandish Recursive Query Examples</h2>

<p>The following query computes an approximation of the Mandelbrot Set
and outputs the result as ASCII-art:

<blockquote><pre>
WITH RECURSIVE
  xaxis(x) AS (VALUES(-2.0) UNION ALL SELECT x+0.05 FROM xaxis WHERE x&lt;1.2),
  yaxis(y) AS (VALUES(-1.0) UNION ALL SELECT y+0.1 FROM yaxis WHERE y&lt;1.0),
  m(iter, cx, cy, x, y) AS (
    SELECT 0, x, y, 0.0, 0.0 FROM xaxis, yaxis
    UNION ALL
    SELECT iter+1, cx, cy, x*x-y*y + cx, 2.0*x*y + cy FROM m 
     WHERE (x*x + y*y) &lt; 4.0 AND iter&lt;28
  ),
  m2(iter, cx, cy) AS (
    SELECT max(iter), cx, cy FROM m GROUP BY cx, cy
  ),
  a(t) AS (
    SELECT group_concat( substr(' .+*#', 1+min(iter/7,4), 1), '') 
    FROM m2 GROUP BY cy
  )
SELECT group_concat(rtrim(t),x'0a') FROM a;
</pre></blockquote>

<p>In this query, the "xaxis" and "yaxis" CTEs define the grid of points for
which the Mandelbrot Set will be approximated.  Each row in the
"m(iter,cx,cy,x,y)" CTE means that after "iter" iterations, the Mandelbrot
iteration starting at cx,cy has reached point x,y.  The number of iterations
in this example is limited to 28 (which severely limits the resolution of
the computation, but is sufficient for low-resolution ASCII-art output).
The "m2(iter,cx,cy)" CTE holds the maximum number of iterations reached when
starting at point cx,cy.
Finally, each row in the "a(t)" CTE holds a string 
which is a single line of the output ASCII-art.
The SELECT statement at the end just queries the "a" CTE to
retrieve all lines of ASCII-art, one by one.

<p>Running the query above in an SQLite [command-line shell] results
in the following output:

<blockquote><pre>
                                    ....#
                                   ..#*..
                                 ..+####+.
                            .......+####....   +
                           ..##+*##########+.++++
                          .+.##################+.
              .............+###################+.+
              ..++..#.....*#####################+.
             ...+#######++#######################.
          ....+*################################.
 #############################################...
          ....+*################################.
             ...+#######++#######################.
              ..++..#.....*#####################+.
              .............+###################+.+
                          .+.##################+.
                           ..##+*##########+.++++
                            .......+####....   +
                                 ..+####+.
                                   ..#*..
                                    ....#
                                    +.
</pre></blockquote>

<tcl>hd_fragment sudoku {solve Sudoku puzzles} </tcl>
<p>This next query solves a Sudoku puzzle.  The state of the puzzle is
defined by an 81-character string formed by reading entries from the
puzzle box row by row from left to right and then from top to bottom.
Blank squares in the puzzle are denoted by a "." character.  
Thus the input string:

<blockquote>
53..7....6..195....98....6.8...6...34..8.3..17...2...6.6....28....419..5....8..79
</blockquote>

<p>Corresponds to a puzzle like this:

<blockquote>
<table border="1" cellpadding="5">
<tr><td>5<td>3<td> <td> <td>7<td> <td> <td> <td>
<tr><td>6<td> <td> <td>1<td>9<td>5<td> <td> <td>
<tr><td> <td>9<td>8<td> <td> <td> <td> <td>6<td>
<tr><td>8<td> <td> <td> <td>6<td> <td> <td> <td>3
<tr><td>4<td> <td> <td>8<td> <td>3<td> <td> <td>1
<tr><td>7<td> <td> <td> <td>2<td> <td> <td> <td>6
<tr><td> <td>6<td> <td> <td> <td> <td>2<td>8<td>
<tr><td> <td> <td> <td>4<td>1<td>9<td> <td> <td>5
<tr><td> <td> <td> <td> <td>8<td> <td> <td>7<td>9
</table>
</blockquote>

<p>This is the query that solves the puzzle:

<blockquote><pre>
WITH RECURSIVE
  input(sud) AS (
    VALUES('53..7....6..195....98....6.8...6...34..8.3..17...2...6.6....28....419..5....8..79')
  ),
  digits(z, lp) AS (
    VALUES('1', 1)
    UNION ALL SELECT
    CAST(lp+1 AS TEXT), lp+1 FROM digits WHERE lp&lt;9
  ),
  x(s, ind) AS (
    SELECT sud, instr(sud, '.') FROM input
    UNION ALL
    SELECT
      substr(s, 1, ind-1) || z || substr(s, ind+1),
      instr( substr(s, 1, ind-1) || z || substr(s, ind+1), '.' )
     FROM x, digits AS z
    WHERE ind>0
      AND NOT EXISTS (
            SELECT 1
              FROM digits AS lp
             WHERE z.z = substr(s, ((ind-1)/9)*9 + lp, 1)
                OR z.z = substr(s, ((ind-1)%9) + (lp-1)*9 + 1, 1)
                OR z.z = substr(s, (((ind-1)/3) % 3) * 3
                        + ((ind-1)/27) * 27 + lp
                        + ((lp-1) / 3) * 6, 1)
         )
  )
SELECT s FROM x WHERE ind=0;
</pre></blockquote>

<p>The "input" CTE defines the input puzzle.
The "digits" CTE defines a table that holds all digits between 1 and 9.
The work of solving the puzzle is undertaken by the "x" CTE.
An entry in x(s,ind) means that the 81-character string "s" is a valid
sudoku puzzle (it has no conflicts) and that the first unknown character
is at position "ind", or ind==0 if all character positions are filled in.
The goal, then, is to compute entries for "x" with an "ind" of 0.

<p>The solver works by adding new entries to the "x" recursive table.
Given prior entries, the recursive-select tries to fill in a single new
position with all values between 1 and 9 that actually work in that
position.  The complicated "NOT EXISTS" subquery is the magic that
figures out whether or not each candidate "s" string is a valid
sudoku puzzle or not.

<p>The final answer is found by looking for a string with ind==0.
If the original sudoku problem did not have a unique solution, then
the query will return all possible solutions.  If the original problem
was unsolvable, then no rows will be returned.  In this case, the unique
answer is:

<blockquote>
534678912672195348198342567859761423426853791713924856961537284287419635345286179
</blockquote>

<p>The solution was computed in less than 300 milliseconds on a modern
workstation.

<h1>Limitations And Caveats</h1>

<ul>
<li><p>
The WITH clause cannot be used within a [CREATE TRIGGER].
<li><p>
The WITH clause must appear at the beginning of a top-level [SELECT] statement
or at the beginning of a subquery.  The WITH clause cannot be prepended to
the second or subsequent SELECT statement of a [compound select].
<li><p>
The SQL:1999 spec requires that the RECURSIVE keyword follow WITH in any
WITH clause that includes a recursive common table expression.  However, for
compatibility with SqlServer and Oracle, SQLite does not enforce this rule.
</ul>