- check determinism

   - missing syntax:
     TRUNCATE (MySQL and PostgreSQL)
    Probably later:
     SELECT ... *INTO*
     more stuff

   - Have a table of allowed tables to enable more compile-time checks.
     (Any interpolated table will be accepted and checked by the DB
     system instead -- we only want compile-time checks here.)

   - Maybe have a table of columns per table to check at compile time.
     This is more tricky if done with context, i.e., per table, because
     we'd have to fully understand scoping.  A complete list of columns
     is a start, though, particularly for DBs that have been constructed
     to work with NATURAL JOIN, which typically have very unique column
     names in their tables.

   - Maybe have ExprAs and TableAs objects.

   - Maybe have WhenThen objects (two classes: one for expr, one for
     suffix operation, for the two types of operations:
     CASE WHEN and CASE x WHEN (the latter allows both).

   - MAYBBE:
     Allow keywords and constructs to be marked as optional (if the read
     dialect allows them).  Probably we need to list all possible optional
     stuff, because this may also depend on the DB version.  Things
     immediately coming to mind:
         ?IF NOT EXISTS         MySQL only
         ?IF EXISTS             MySQL only
         ?ONLY                  PostgreSQL only

   - for peep-hole optimisions on the generated code, str_append_... should
     not directly construct a string, but return a recursive structure of
     the term.  Each node should be something like:

         { kind => '...', args => [...], ... }

     E.g. the number 5:
         { kind => 'num', args => [ 5 ] }

     Arbitrary perl interpolation in list or scalar context:

         { kind => 'list',   args => [ 'do{perlcode}' ] }
         { kind => 'scalar', args => [ 'do{perlcode}' ] }

     More information could be stored, of course:

         { kind => 'join', args => [ ... ], sep => ',', prefix => 'CONCAT ' }
         { kind => 'map', args => [ ... ], body => '($_)' }

     On this structure, we could run an optimiser that removes maps,
     combines joins, etc.

   - Support translation of IF NOT EXISTS etc. via information_scheme
     tables.  This is tricky with all the different data bases around.

   - Add support for scalar references, which should become bind

         my $uid= ...;
         my $q= sql{ SELECT foo FROM bar WHERE uid = \$uid and name = ? };

     "$q" should stringify as:

         SELECT `foo` FROM `bar` WHERE `uid` = ? and `name` = ?

     There shall be a function $q->bind($name) which produces a list of
     values, parameterised by all the ? found, automatically filling in
     all the \$ references automatically.  So:

           my @a= $q->bind(4);

     Should lead to @a=(99,4);

     Implementing this is a bit tricky, but would be great.  We could
     then prepare statements with references, and very simply execute
     them later:

           my $pq= $dbh->prepare($q);

     LATER: If dbh is set, we could even provide prepare, execute, etc.
     functions ourselves to eliminate the need for the ->bind()
     invocation and for two variables to keep track of:

           my $pq= $q->prepare();

   - LATER: named bind variables:

      my $q= sql{ ... WHERE time_insert BETWEEN ?{time_min} AND ?{time_max} }
      my $pq= $q->prepare();
      $pq->execute(time_min => $time_min, time_max => $time_max);

     Of course, the named bind variable could be used more than

        my $q= sql{ ...
                (time_insert BETWEEN ?{time_min} AND ?{time_max})
            OR  (time_close  BETWEEN ?{time_min} AND ?{time_max})

        my $pq= $q->prepare();
        $pq->execute(time_min => $time_min, time_max => $time_max);

     Problem: mixing of normal and named variables. Solution: normal
     ones first, so we get prototypes like:

        ($$$)        ; three normal bind variables
        ($$$%)       ; three normal plus some named bind variables
        (%)          ; only named bind variables

   - optimise:
     min1, min1default, and max1_if_scalar can usually be elimited by
     looking at the list:

        - min1 and min1default can be dropped if the list is
          guaranteed to be non-empty

        - max1_if_scalar can be dropped if the list guaranteed to have
          at most 1 element.

     It is quite easy to count, actually, since only non-scalar
     'interpol' nodes are dangerous.

   - optimise:

        str_append_str ( str_append_comma str_append_str )*

     -> one str_append_str with the concatenated string.

   - Optimise:
        (map{ SQL::Yapp::column1($_) } SQL::Yapp::ASTERISK)
     is equal to:
     and inside a join(), it can even be optimised to:

     Same for QUESTION and all the other special values.

     Same for literal numbers if inside a join():
        (map{ SQL::Yapp::column1($_) } 5)
     in a join is equal to:

   - encapsulate different syntax (read and write) in different packages
     so that this code has no knowledge about MySQL, PostgreSQL, Oracle,
     etc., but lets the packages handle it.

   -  Klarstellen, warum `...`-Syntax benutzt wird.  Evtl. auch [...] erlauben
      mit der entsprechenden Dialekt-Einstellung.

    - In void context, prepare, execute and possibly fetch all rows and store
      them somewhere.  Maybe don't fetch, but provide access to the handle
      somehow.  Maybe have a special callback hook for running
      commands in void mode.
   - ARRAY[...]

   - C<//> will be normalised to C<COALESCE>?

   - Normalisation:

           COALESCE()               MySQL, PostgreSQL
       vs. NVL()                    Oracle             <-- unfold

           SELECT ... <nothing>     MySQL, PostgreSQL
       vs. SELECT ... FROM dual     Oracle

           CASE WHEN ... END        Std
       vs. DECODE()                 Oracle

    With read_dialect containing 'perl',
    allow: &&, ||, ^^, ! as boolean operators.  This requires
    substitution based in read_dialect in set_expr_functor.

   - Handle INSERT...SET without set2value if $write_dialect == 'mysql'.

   - Handle LIMIT without large number if $write_dialect == 'postgresql'.

   - Statements:
   - Expressions
     - special Expr syntax:
           CASE ... (WHEN ... THEN ...)* [ ELSE ... ] END)

   - functions with special syntax, e.g. CAST(...)
         CAST      ( <expr> AS <type> )
         TREAT     ( <expr> AS <type> )
         POSITION  ( <expr> IN <expr> [ USING <char length units> ] )
         SUBSTRING ( <expr> FROM <expr> [ FOR <expr> ] [ USING <char length units> ] )
         { CHAR_LENGTH | CHARACTER_LENGTH } ( <expr> [ USING <char length units> ] )
         CONVERT   ( <expr> USING <transcoding name> )
         OVERLAY   ( <expr> PLACING <expr> FROM <expr> [ FOR <expr> ] [ USING <char length units> ] )
         EXTRACT   ( <expr> FROM <expr> )
         UNNEST    ( <expr> ) [ WITH ORDINALITY ]
         TRANSLATE ( <expr> USING <transliteration name> )
   - XOR

   - Interpolation:
     Allow (arrays of) array references after INSERT ... VALUES:

       my $a= ['a', 'b'];


       my @a= (['a','b'],[1,2]);


     For this, maybe allow reference syntax:

       my @a= ('a', 'b');


     Here, it is not needed, however, because you can write:

       INSERT INTO t VALUES (@a)

     This even works now.

   - subqueries: ANY, SOME, ALL

   - implement context: sql{...} as expression will expand as an SQL::Yapp::Expr

   - Allow Stmt interpolation in subqueries:

         my $stmt= sql{ SELECT ... };

         ... WHERE a = ANY (Stmt $stmt) ...

     This requires us to check for a SELECT Stmt, not only a Stmt, which
     could be wrong.  Maybe we leave that problem to the SQL server?
     In that case, it would be easy, just add another function
     parse_select_stmt() that's like parse_stmt() but only accepts SELECT
     and interpol.

     Think about interpolating ORDER BY / GROUP BY stuff.  It will
     happen in applications where the user can set the order.
     At least we need to interpolate the direction DESC/ASC.
     Better yet: a sequence of Order:

         my $dir=   $desc ? sqlDir{DESC} : sqlDir{ASC};
         my @order= sqlOrder{ a Dir $dir, b ASC }
         my $q=     sql{ SELECT * FROM B ORDER BY @order };

     It would be interesting whether it's possible to implement

         my $q=     sql{ SELECT * FROM B ORDER BY @order DESC };

         --> ... a ASC, b DESC

     Store direction in bool variable and then use:

         $order= sqlOrder{ $order DESC }
             if $direction eq 'DESC';

     This works already.

   - query types:

   - think about handling table and column names.

     - The goal is to modify table names on the fly, e.g. to prefix
       each table name with a user-defined string.  This means we
       need to finely distinguish different types of names.

     - How to handle unqualified and qualified column specifications
       in interpolations?  I.e.:

           SELECT foo.$col_name1 FROM ...
           SELECT $col_name2 FROM ...

       $col_name2 may contain a table name, while $col_name1 may not.
       Maybe we must require that $col_name1 is a simple Perl string,
       so that we can invoke:

           quote_identifier('foo', $col_name1)

       while $col_name2 may be either a Perl string, or a Column
       object which is already in proper syntax.  In any case, for
       qualified column names, we *must* invoke quote_identifier
       with both arguments.