Chapter 9: Statements Summary and Error Checking

The following is a brief discussion of the major statement types and the semantic rules that CQL enforces for each of the statements. A detailed discussion of SQL statements (the bulk of these) is beyond the scope of this document and you should refer to the SQLite documentation for most details. However, in many cases CQL does provide additional enforcement and it is helpful to describe the basic checking that happens for each fragment of CQL. A much more authoritative list of the things CQL checks for can be inferred from the error documentation. "Tricky" errors have examples and suggested remediation.

The Primary SQL Statements

These are, roughly, the statements that involve the database.

The SELECT Statement

Top level statement list processing for select. This is easily the hardest statement to process. Each clause has its own set of complex rules and the result of previous clauses constrains the next in a complex fashion. Among the things that are verified:

• the mentioned tables exist and have the mentioned columns
• the columns are type compatible in their context
• any variables in the expressions are compatible
• aggregate functions are used only in places where aggregation makes sense
• column and table names are unambiguous, especially when self-joins are involved
• compound selects (e.g. with UNION) are type-consistent in all the fragments
• the projection of a select has unique column labels if they are used

The SELECT * Statement

SELECT * is special in that it creates its own struct type by assembling all the columns of all the tables in the select's join result. CQL rewrites these column names into a new SELECT with the specific columns explicitly listed. While this makes the program slightly bigger it means that logically deleted columns are never present in results because SELECT * won't select them and attempting to use a logically deleted column results in an error.

The CREATE TABLE Statement

Unlike the other parts of DDL we actually deeply care about the tables. We have to grab all the columns and column types out of it and create the appropriate structure type for the table. Along the way we validate a bunch of stuff like:

• verify unique table name
• no duplicate column names
• recursive correctness of constraints (see constraints discussion below)

The UNIQUE KEY Clause

Similar to other constraints, we don't actually do anything with this other than offer some validation. Again, we use the usual helpers for name lookup within the context of the table that contains the constraint.

The FOREIGN KEY Clause

Similar to other constraints, we don't actually do anything with this other than offer some validation. Again, we use the usual helpers for name lookup within the context of the table with the foreign key. Note that the foreign key has to be validated against two tables to fully validate it.

The PRIMARY KEY Clause

Similar to other constraints, we don't actually do anything with this other than offer some validation. Again, we use the usual helpers for name lookup within the context of the table with the primary key.

The CHECK Clause

Similar to other constraints, we don't actually do anything with this other than offer some validation. The CHECK clause is validated after the entire table has been processed so that even if it appears early in the table, the clause can use any columns defined later in the table.

The CREATE INDEX Statement

CQL doesn't really do anything with indices but we do validate that they make sense (so we lookup all the names of all the columns and so forth.)

The CREATE VIEW Statement

Create view analysis is very simple because the select analysis does the heavy lifting. All we have to do is validate that the view is unique, then validate the select statement.

Additionally, views must not be allowed to have any NULL type columns; all nulls must be converted to some type with a CAST. e.g. create view foo as select NULL n is not valid. NULL is not a real storage type.

The CREATE TRIGGER Statement

The create trigger statement is quite a beast, and validations include:

• The trigger name must be unique
• For insert the "new.*" table is available in expressions/statement
• For delete the "old.*" table is available in expressions/statements
• For update both are available
  • If optional columns present in the update, they must be unique/valid
• The when expression must evaluate to a numeric
• The statement list must be error free with the usual rules plus new/old
• The raise function may be used inside a trigger (NYI)
• The table name must be a table (not a view) UNLESS the trigger type is INSTEAD OF
• Select statements inside the statement block do not count as returns for the procedure and that includes the create trigger

The DROP TABLE Statement

This is the basic checking for the drop table statement:

• the table must exist in some version
• it has to be a table and not a view

The DROP VIEW Statement

This is the basic checking for the drop view statement:

• the view must exist in some version
• it has to be a view and not a table

The DROP INDEX Statement

This is the basic checking for the drop index statement:

• the index must exist in some version
• it could be deleted now, that's ok, but the name has to be valid

The DROP TRIGGER Statement

This is the basic checking for the drop trigger statement

• the trigger must exist in some version
• it could be deleted now, that's ok, but the name has to be valid

The RAISE Statement

CQL validates that RAISE is being used in the context of a trigger and that it has the correct arguments

The ALTER TABLE ADD COLUMN Statement

To validate alter table add column we check the following:

• the table must exist and not be a view (in any version)
• the column definition of the new column must be self-consistent
• no auto-increment columns may be added
• added columns must be either nullable or have a default value

Note: Alter statements are typically used in the context of migration, so it's possible the table that is mentioned is condemned in a future version. We still have to run the intervening upgrade steps so basically DDL gets to ignore the current deadness of the table as in context it might be "not dead yet". This will be more obvious in the context of the schema maintenance features. (q.v.)

The DELETE Statement

The delete analyzer sets up a scope for the table being deleted and then validates the WHERE clause, if present, against that scope. Additionally, we verify that the table actually was defined and is not a view.

The UPDATE Statement

The update analyzer sets up the scope for the table(s) being updated. If there are optional clauses (e.g. LIMIT), they are evaluated just like in a select statement with those same helper methods. Expression fragments are evaluated similarly as in a select statement.

The INSERT Statement

We check that the table exists and then we walk the columns and the value list to make sure they are valid for the table. Also, we cannot insert into a view.

Details:

• The column list specifies the columns we will provide; they must exist and be unique.
• The columns specified must suffice to insert a row (all not nulls and not default present.)
• The insert list specifies the values that are to be inserted.
• The type of each value must match the type of the column.
• Auto-increment columns may be specified as NULL.
• If there are too many or too few columns, that is considered an error.
• If no columns are specified, that is the same as if all columns had been specified, in table order.

The THROW Statement

Throw can literally go anywhere, so it's always ok.

The BEGIN TRANSACTION Statement

Begin transaction can go anywhere, so it's always ok.

The sqlite documentation can be helpful here (CQL syntax is a subset). See: https://www.sqlite.org/lang_transaction.html

The COMMIT TRANSACTION Statement

Commit transaction can go anywhere, so it's always ok.

The sqlite documentation can be helpful here (CQL syntax is a subset). See: https://www.sqlite.org/lang_transaction.html

The ROLLBACK TRANSACTION Statement

Rollback transaction can go anywhere but if you're using the format where you rollback to a particular save point, then we must have seen that name in a savepoint statement previously. Otherwise, it's an error.

The sqlite documentation can be helpful here again (CQL syntax is a subset). See: https://www.sqlite.org/lang_transaction.html

The SAVEPOINT Statement

The savepoint statement can go anywhere but we do record this savepoint name as having been seen, so that we can verify it in rollback. So this is sort of a weak declaration of the savepoint name.

The sqlite documentation can be helpful here (CQL syntax is a subset). https://www.sqlite.org/lang_savepoint.html

The RELEASE SAVEPOINT Statement

Release savepoint can go anywhere but we must have seen that name in a previous savepoint statement, otherwise it's an error.

The sqlite documentation can be helpful here (CQL syntax is a subset). https://www.sqlite.org/lang_savepoint.html

The PROCEDURE SAVEPOINT Statement

A common pattern is to have a savepoint associated with a particular procedure. The savepoint's scope is the same as the procedure's scope. More precisely

create procedure foo()
begin
  proc savepoint
  begin
   -- your code
  end;
end;

becomes:

create procedure foo()
begin
  savepoint @proc;  -- @proc is always the name of the current procedure
  begin try
    -- your code
    release savepoint @proc;
  end try;
  begin catch
    rollback transaction to savepoint @proc;
    release savepoint @proc;
    throw;
  end catch;
end;

This form is not quite syntactic sugar because there are some interesting rules:

• the proc savepoint form must be used at the top level of the procedure, hence no leave or continue may escape it
• within begin/end the return form may not be used; you must use rollback return or commit return (see below)
• throw may be used to return an error as usual
• proc savepoint may be used again, at the top level, in the same procedure, if there are, for instance, several sequential stages
• a procedure using proc savepoint could call another such procedure, or a procedure that manipulates savepoints in some other way

The ROLLBACK RETURN Statement

This form may be used only inside of a proc savepoint block. It indicates that the savepoint should be rolled back and then the procedure should return. It is exactly equivalent to:

  rollback transaction to savepoint @proc;
  release savepoint @proc;
  return; -- wouldn't actually be allowed inside of proc savepoint; see note below

Note: to avoid errors, the loose return above is not actually allowed inside of proc savepoint -- you must use rollback return or commit return.

The COMMIT RETURN Statement

This form may be used only inside of a proc savepoint block. It indicates that the savepoint should be released and then the procedure should return. It is exactly equivalent to:

  release savepoint @proc;
  return; -- wouldn't actually be allowed inside of proc savepoint; see note below

Of course this isn't exactly a commit, in that there might be an outer savepoint or outer transaction that might still be rolled back, but it is commited at its level of nesting, if you will. Or, equivalently, you can think of it as merging the savepoint into the transaction in flight.

Note: to avoid errors, the loose return above is not actually allowed inside of proc savepoint and you must use rollback return or commit return.

The CREATE VIRTUAL TABLE Statement

The SQLite CREATE VIRTUAL TABLE form (https://sqlite.org/lang_createvtab.html) is problematic from CQL because:

• it is not parseable, because the module arguments can be literally anything (or nothing), even a letter to your grandma
• the arguments do not necessarily say anything about the table's schema at all

So the CQL form departs from the standard syntax to this form:

create virtual table virt_table using my_module [(module arguments)]  as (
  id integer not null,
  name text
);

The part after the AS is used by CQL as a table declaration for the virtual table. The grammar for that is exactly the same as a normal CREATE TABLE statement. However, that part is not transmitted to SQLite; when the table is created, SQLite sees only the part it cares about, which is the part before the AS.

In order to have strict parsing rules, the module arguments follow one of these forms:

1. no arguments at all
2. a list of identifiers, constants, and parenthesized sublists, just like in the @attribute form
3. the words arguments following

Case 1 Example

create virtual table virt_table using my_module as (
  id integer not null,
  name text
);

becomes (to SQLite)

CREATE VIRTUAL TABLE virt_table USING my_module;

Note: empty arguments USING my_module() are not allowed in the SQLite docs but do seem to work in SQLite. We take the position that no args should be formatted with no parentheses, at least for now.

Case 2 Example

create virtual table virt_table using my_module(foo, 'goo', (1.5, (bar, baz))) as (
  id integer not null,
  name text
);

CREATE VIRTUAL TABLE virt_table USING my_module(foo, "goo", (1.5, (bar, baz)));

This form allows for very flexible arguments but not totally arbitrary arguments, so it can still be parsed and validated.

Case 3 Example

This case recognizes the popular choice that the arguments are often the actual schema declaration for the table in question. So:

create virtual table virt_table using my_module(arguments following) as (
  id integer not null,
  name text
);

becomes

CREATE VIRTUAL TABLE virt_table USING my_module(
  id INTEGER NOT NULL,
  name TEXT
);

The normalized text (keywords capitalized, whitespace normalized) of the table declaration in the as clause is used as the arguments.

Other details

Virtual tables go into their own section in the JSON and they include the module and moduleArgs entries; they are additionally marked isVirtual in case you want to use the same processing code for virtual tables as normal tables. The JSON format is otherwise the same, although some things can't happen in virtual tables (e.g. there is no TEMP option so "isTemp" must be false in the JSON.)

For purposes of schema processing, virtual tables are on the @recreate plan, just like indices, triggers, etc. This is the only option since the alter table form is not allowed on a virtual table.

Semantic validation enforces "no alter statements on virtual tables" as well as other things like no indices, and no triggers, since SQLite does not support any of those things.

CQL supports the notion of eponymous virtual tables. If you intend to register the virtual table's module in this fashion, you can use create virtual table @eponymous ... to declare this to CQL. The only effect this has is to ensure that CQL will not try to drop this table during schema maintenance as dropping such a table is an invalid operation. In all other ways, the fact that the table is eponymous makes no difference.

Finally, because virtual tables are on the @recreate plan, you may not have foreign keys that reference virtual tables. Such keys seem like a bad idea in any case.

The Primary Procedure Statements

These are the statements which form the language of procedures, and do not involve the database.

The CREATE PROCEDURE Statement

Semantic analysis of stored procedures is fairly easy at the core:

• check for duplicate names
• validate the parameters are well formed
• set the current proc in flight (this not allowed to nest)
• recurse on the statement list and prop errors
• record the name of the procedure for callers In addition, while processing the statement:
• we determine if it uses the database; this will change the emitted signature of the proc to include a sqlite3 *db input argument and it will return a sqlite error code (e.g. SQLITE_OK)
• select statements that are loose in the proc represent the "return" of that select; this changes the signature to include a sqlite3_stmt **pstmt parameter corresponding to the returned statement

The IF Statement

The top level if node links the initial condition with a possible series of else_if nodes and then the else node. Each condition is checked for validity. The conditions must be valid expressions that can each be converted to a boolean.

The SET Statement

The set statement is for variable assignment. We just validate that the target exists and is compatible with the source. Cursor variables cannot be set with simple assignment and CQL generates errors if you attempt to do so.

The LET Statement

Let combines a DECLARE and a SET. The variable is declared to be the exact type of the right hand side. All the validations for DECLARE and SET are applicable, but there is no chance that the variable will not be compatible with the expression. The expression could still be erroneous in the first place. The variable could be a duplicate.

The SWITCH Statement

The SWITCH form requires a number of conditions to successfully map down to a C switch statement. These are:

• the switch-expression must be a not-null integral type (integer not null or long integer not null)
  • the WHEN expressions must be losslessly promotable to the type of the switch-expression
• the values in the WHEN clauses must be unique
• If ALL VALUES is present then:
  • the switch-expression must be of an enum type
  • the WHEN values must cover every value of the enum except those beginning with '_'
  • there can be no extra WHEN values not in the enum
  • there can be no ELSE clause

The DECLARE PROCEDURE Statement

There are three forms of this declaration:

• a regular procedure with no DML
  • e.g. declare proc X(id integer);
• a regular procedure that uses DML (it will need a db parameter and returns a result code)
  • e.g. declare proc X(id integer) using transaction;
• a procedure that returns a result set, and you provide the result columns
  • e.g. declare proc X(id integer) : (A bool not null, B text); The main validations here are that there are no duplicate parameter names, or return value columns.

The DECLARE FUNCTION Statement

Function declarations are similar to procedures; there must be a return type (use proc if there is none). The DECLARE SELECT FUNCTION form indicates a function visible to SQLite; other functions are usable in the call statement.

The DECLARE Variable Statement

This declares a new local or global variable that is not a cursor. The type is computed with the same helper that is used for analyzing column definitions. Once we have the type we walk the list of variable names, check them for duplicates and such (see above) and assign their type. The canonical name of the variable is defined here. If it is later used with a different casing the output will always be as declared. e.g. declare Foo integer; set foo = 1; is legal but the output will always contain the variable written as Foo.

The DECLARE Cursor Statement

There are two forms of the declare cursor, both of which allow CQL to infer the exact type of the cursor.

• declare foo cursor for select etc.
  • the type of the cursor is the net struct type of the select list
• declare foo cursor for call proc();
  • proc must be statement that produces a result set via select (see above)
  • the type of the cursor is the struct of the select returned by the proc
  • note if there is more than one loose select in the proc they must match exactly
• cursor names have the same rules regarding duplicates as other variables With this in mind, both cases simply recurse on either the select or the call and then pull out the structure type of that thing and use it for the cursor's shape. If the call is not semantically valid according to the rules for calls or the select is not semantically valid, then of course this declaration will generate errors.

The DECLARE Value Cursor Statement

This statement declares a cursor that will be based on the return type of a procedure. When using this form the cursor is also fetched, hence the name. The fetch result of the stored proc will be used for the value. At this point, we use its type only.

• the call must be semantically valid
• the procedure must return an OUT parameter (not a result set)
• the cursor name must be unique

The WHILE Statement

While semantic analysis is super simple.

• the condition must be numeric
• the statement list must be error-free
• loop_depth is increased allowing the use of interior leave/continue

The LOOP Statement

Loop analysis is just as simple as "while" -- because the loop_stmt literally has an embedded fetch, you simply use the fetch helper to validate that the fetch is good and then visit the statement list. Loop depth is increased as it is with while.

The CALL Statement

There are three ways that a call can happen:

• signatures of procedures that we know in full:
  • call foo();
  • declare cursor for call foo();
• some external call to some outside function we don't know
  • e.g. call printf('hello, world\n');

The cursor form can be used if and only if the procedure has a loose select or a call to a procedure with a loose select. In that case, the procedure will have a structure type, rather than just "ok" (the normal signature for a proc). If the user is attempting to do the second case, cursor_name will be set and the appropriate verification happens here.

Note: Recursively calling fetch cursor is not really doable in general because at the point in the call we might not yet know that the method does in fact return a select. You could make it work if you put the select before the recursive call.

Semantic rules:

• for all cases each argument must be error-free (no internal type conflicts)
• for known procs
  • the call has to have the correct number of arguments
  • if the formal is an out parameter the argument must be a variable
    • the type of the variable must be an exact type match for the formal
  • non-out parameters must be type-compatible, but exact match is not required

The DECLARE OUT CALL Statement

This form is syntactic sugar and corresponds to declaring any OUT parameters of the CALL portion that are not already declared as the exact type of the OUT parameter. This is intended to save you from declaring a lot of variables just so that you can use them as OUT arguments.

Since any variables that already exist are not re-declared, there are no additional semantic rules beyond the normal call except that it is an error to use this form if no OUT variables needed to be declared.

The FETCH Statement

The fetch statement has two forms:

• fetch C into var1, var2, var3 etc.
• fetch C; The second form is called the auto_cursor. In the first form the variables of the cursor must be assignment compatible with declared structure type of the cursor and the count must be correct. In the second form, the codegen will implicitly create local variables that are exactly the correct type, but we'll cover that later. Since no semantic error is possible in that case, we simply record that this is an auto_cursor and then later we will allow the use of C.field during analysis. Of course "C" must be a valid cursor.

The CONTINUE Statement

We just need to ensure that continue is inside a loop or while.

The LEAVE Statement

We only need to ensure that leave is inside a loop, while or switch.

The TRY/CATCH Statements

No analysis needed here other than that the two statement lists are ok.

The CLOSE CURSOR Statement

For close [cursor], we just validate that the name is in fact a cursor and it is not a boxed cursor. Boxed cursor lifetime is managed by the box object so manually closing it is not allowed. Instead, the usual reference-counting semantics apply; the boxed cursor variable typically falls out of scope and is released, or is perhaps set to NULL to release its reference early.

The OUT CURSOR Statement

For out [cursor], we first validate that the name is a cursor then we set the output type of the procedure we're in accordingly.

The "Meta" Statements

The program's control/ the overall meaning of the program / or may give the compiler specific directives as to how the program should be compiled.

The @ECHO Statement

Echo is valid in any top level contexts.

The @PREVIOUS SCHEMA Statement

Begins the region where previous schema will be compared against what has been declared before this directive for alterations that could not be upgraded.

The @SCHEMA_UPGRADE_SCRIPT Statement

When upgrading the DDL, it's necessary to emit create table statements for the original version of the schema. These create statements may conflict with the current version of the schema. This attribute tells CQL to 1) ignore DDL in stored procedures for declaration purposes; only DDL outside of a proc counts 2) do not make any columns "hidden" thereby allowing all annotations to be present so they can be used to validate other aspects of the migration script.

The @SCHEMA_UPGRADE_VERSION Statement

For sql stored procedures that are supposed to update previous schema versions you can use this attribute to put CQL into that mindset. This will make the columns hidden for the version in question rather than the current version. This is important because older schema migration procedures might still refer to old columns. Those columns truly exist at that schema version.

The @ENFORCE_STRICT Statement

Switch to strict mode for the indicated item. The choices and their meanings are:

• "FOREIGN KEY ON DELETE" indicates there must be some ON DELETE action in every FK
• "FOREIGN KEY ON UPDATE" indicates there must be some ON UPDATE action in every FK
• "INSERT SELECT" indicates that insert with SELECT for values may not include top level joins (avoiding a SQLite bug)
• "IS TRUE" indicates that IS TRUE IS FALSE IS NOT TRUE IS NOT FALSE may not be used (*)
• "JOIN" indicates only ANSI style joins may be used, and "from A,B" is rejected
• "PROCEDURE" indicates no calls to undeclared procedures (like loose printf calls)
• "SELECT IF NOTHING" indicates (select ...) expressions must include an IF NOTHING clause if they have a FROM part
• "TABLE FUNCTIONS" indicates table valued functions cannot be used on left/right joins (avoiding a SQLite bug)
• "TRANSACTION" indicates no transactions may be started, committed, or aborted
• "UPSERT" indicates no upsert statement may be used (*)
• "WINDOW FUNCTION" indicates no window functions may be used (*)
• "WITHOUT ROWID" indicates WITHOUT ROWID may not be used

The items marked with * are present so that features can be disabled to target downlevel versions of SQLite that may not have those features.

See the grammar details for exact syntax.

The @ENFORCE_NORMAL Statement

Turn off strict enforcement for the indicated item.

The @ENFORCE_PUSH Statement

Push the current strict settings onto the enforcement stack. This does not change the current settings.

The @ENFORCE_POP Statement

Pop the previous current strict settings from the enforcement stack.

The @ENFORCE_RESET Statement

Turns off all the strict modes. Best used immediately after @ENFORCE_PUSH.

The @DECLARE_SCHEMA_REGION Statement

A schema region is a partitioning of the schema such that it only uses objects in the same partition or one of its declared dependencies. One schema region may be upgraded independently from any others (assuming they happen such that dependents are done first.) Here we validate:

• the region name is unique
• the dependencies (if any) are unique and exist
• the directive is not inside a procedure

The @BEGIN_SCHEMA_REGION Statement

Entering a schema region makes all the objects that follow part of that region. It also means that all the contained objects must refer to only pieces of schema that are in the same region or a dependent region. Here we validate that region we are entering is in fact a valid region and that there isn't already a schema region.

The @END_SCHEMA_REGION Statement

Leaving a schema region puts you back in the default region. Here we check that we are in a schema region.

The @EMIT_ENUMS Statement

Declared enumarations can be voluminous and it is undesirable for every emitted .h file to contain every enumeration. To avoid this problem you can emit enumeration values of your choice using @emit_enums x, y, z which places the named enumerations into the .h file associated with the current translation unit. If no enumerations are listed, all enums are emitted.

Note: generated enum definitions are protected by #ifndef X ... #endif so multiple definitions are harmless and hence you can afford to use @emit_enums for the same enum in several translations units, if desired.

Note: Enumeration values also appear in the JSON output in their own section.

The @EMIT_CONSTANTS Statement

This statement is entirely analogous to the the @EMIT_ENUMS except that the parameters are one or more constant groups. In fact constants are put into groups precisely so that they can be emitted in logical bundles (and to encourage keeping related constants together). Placing @EMIT_CONSTANTS causes the C version of the named groups to go into the current .h file.

Note: Global constants also appear in the JSON output in their own section.

Important Program Fragments

These items appear in a variety of places and are worthy of discussion. They are generally handled uniformly.

Argument Lists

In each case we walk the entire list and do the type inference on each argument. Note that this happens in the context of a function call, and depending on what the function is, there may be additional rules for compatibility of the arguments with the function. The generic code doesn't do those checks, there is per-function code that handles that sort of thing.

At this stage the compiler computes the type of each argument and makes sure that, independently, they are not bogus.

Procedures that return a Result Set

If a procedure is returning a select statement then we need to attach a result type to the procedure's semantic info. We have to do some extra validation at this point, especially if the procedure already has some other select that might be returned. The compiler ensures that all the possible select results are are 100% compatible.

General Name Lookups

Every name is checked in a series of locations. If the name is known to be a table, view, cursor, or some other specific type of object then only those name are considered. If the name is more general a wider search is used.

Among the places that are considered:

• columns in the current join if any (this must not conflict with #2)
• local or global variables
• fields in an open cursor
• fields in enumerations and global constants

Data Types with a Discriminator

Discriminators can appear on any type, int, real, object, etc.

Where there is a discriminator the compiler checks that (e.g.) object<Foo> only combines with object<Foo> or object. real<meters> only combines with real<meters> or real. In this way its not possible to accidentally add meters to kilograms or to store an int<task_id> where an int<person_id> is required.

The CASE Expression

There are two parts to this: the "when" expression and the "then" expression. We compute the aggregate type of the "when" expressions as we go, promoting it up to a larger type if needed (e.g. if one "when" is an int and the other is a real, then the result is a real). Likewise, nullability is computed as the aggregate. Note that if nothing matches, the result is null, so we always get a nullable resultm unless there is an "else" expression. If we started with case expression, then each "when" expression must be comparable to the case expression. If we started with case when xx then yy; then each case expression must be numeric (typically boolean).

The BETWEEN EXPRESSIONS

Between requires type compatibility between all three of its arguments. Nullability follows the usual rules: if any might be null then the result type might be null. In any case, the result's core type is BOOL.

The CAST Expression

For cast expressions we use the provided semantic type; the only trick is that we preserve the extra properties of the input argument. e.g. CAST does not remove NOT NULL.

The COALESCE Function

Coalesce requires type compatibility between all of its arguments. The result is a not null type if we find a not null item in the list. There should be nothing after that item. Note that ifnull and coalesce are really the same thing except ifnull must have exactly two arguments.

The IN AND NOT IN Expressions

The in predicate is like many of the other multi-argument operators. All the items must be type compatible. Note that in this case the nullablity of the items does not matter, only the nullability of the item being tested. Note that null in (null) is null, not true.

Aggregate Functions

Aggregate functions can only be used in certain places. For instance they may not appear in a WHERE clause.

User Defined Functions

User defined function - this is an external function. There are a few things to check:

• If this is declared without the select keyword then
  • we can't use these in SQL, so this has to be a loose expression
• If this is declared with the select keyword then
  • we can ONLY use these in SQL, not in a loose expression
• args have to be compatible with formals

Calling a procedure as a function

There are a few things to check:

• we can't use these in SQL, so this has to be a loose expression
• args have to be compatible with formals, except
• the last formal must be an OUT arg and it must be a scalar type
• that out arg will be treated as the return value of the "function"
• in code-gen we will create a temporary for it; semantic analysis doesn't care

Root Expressions

A top level expression defines the context for that evaluation. Different expressions can have constraints. e.g. aggregate functions may not appear in the WHERE clause of a statement. There are cases where expression nesting can happen. This nesting changes the evaluation context accordingly, e.g. you can put a nested select in a where clause and that nested select could legally have aggregates. Root expressions keep a stack of nested contexts to facilitate the changes.

Table Factors

A table factor is one of three things:

• a table name (a string) select * from X
• a select subquery (select X,Y from..) as T2
• a list of table references select * from (X, Y, Z) Here we dispatch to the appropriate helper for each case.

Joining with the USING Clause

When specifying joins, one of the alternatives is to give the shared columns in the join e.g. select * from X inner join Y using (a,b). This method validates that all the columns are present on both sides of the join, that they are unique, and they are comparable. The return code tells us if any columns had SENSITIVE data. See Special Note on JOIN...USING below

JOIN WITH THE ON Clause

The most explicit join condition is a full expression in an ON clause this is like select a,b from X inner join Y on X.id = Y.id; The on expression should be something that can be used as a bool, so any numeric will do. The return code tells us if the ON condition used SENSITIVE data.

TABLE VALUED FUNCTIONS

Table valued functions can appear anywhere a table is allowed. The validation rules are:

• must be a valid function

• must return a struct type (i.e. a table-valued-function)

• must have valid arg expressions

• arg expressions must match formal parameters The name of the resulting table is the name of the function

• but it can be aliased later with "AS"

  Special Note on the select * and select T.* forms

  The select * construct is very popular in many codebases but it can be unsafe to use in production code because, if the schema changes, the code might get columns it does not expect. Note the extra columns could have appeared anywhere in the result set because the * applies to the entire result of the FROM clause, joins and all, so extra columns are not necessarily at the end and column ordinals are not preserved. CQL mitigates this situation somewhat with some useful constraints/features:

• in a select *, and indeed in any query, the column names of the select must be unique, this is because:

  • they could form the field names of an automatically generated cursor (see the section on cursors)
  • they could form the field names in a CQL result set (see section on result sets)
  • it's weird/confusing to not have unique names generally

• when issuing a select * or a select T.* CQL will automatically expand the * into the actual logical columns that exist in the schema at the time the code was compiled

  • this is important because if a column had been logically deleted from a table it would be unexpected in the result set even though it is still present in the database and would throw everything off
  • likewise if the schema were to change without updating the code, the code will still get the columns it was compiled with, not new columns

Expanding the * at compile time means Sqlite cannot see anything that might tempt it to include different columns in the result.

With this done we just have to look at the places a select * might appear so we can see if it is safe (or at least reasonably safe) to use * and, by extension of the same argument, T.*.

In an EXISTS or NOT EXISTS clause like `where not exists (select from x)`*

• this is perfectly safe; the particular columns do not matter; select * is not even expanded in this case.

In a statement that produces a result set like `select from table_or_view`*

• binding to a CQL result set is done by column name and we know those names are unique
• we won't include any columns that are logically deleted, so if you try to use a deleted column you'll get a compile time error

In a cursor statement like declare C cursor for select * from table_or_view there are two cases here:

Automatic Fetch fetch C;

• in this case you don't specify the column names yourself;2 they are inferred
• you are therefore binding to the columns by name, so new columns in the cursor would be unused (until you choose to start using them)
• if you try to access a deleted column you get a compile-time error

Manual Fetch: fetch C into a, b, c;

• In this case the number and type of the columns must match exactly with the specified variables
• If new columns are added, deleted, or changed, the above code will not compile

So considering the cases above we can conclude that auto expanding the * into the exact columns present in the compile-time schema version ensures that any incompatible changes result in compile time errors. Adding columns to tables does not cause problems even if the code is not recompiled. This makes the * construct much safer, if not perfect, but no semantic would be safe from arbitrary schema changes without recompilation. At the very least here we can expect a meaningful runtime error rather than silently fetching the wrong columns.

Special Note on the JOIN...USING form

CQL varies slightly from SQLite in terms of the expected results for joins if the USING syntax is employed. This is not the most common syntax (typically an ON clause is used) but Sqlite has special rules for this kind of join.

Let's take a quick look. First some sample data:

create table A( id integer, a text, b text);
create table B( id integer, c text, d text);

insert into A values(1, 'a1', 'b1');
insert into B values(1, 'c1', 'd1');
insert into A values(2, 'a2', 'b2');
insert into B values(2, 'c2', 'd2');

Now let's look at the normal join; this is our reference:

select * from A T1 inner join B T2 on T1.id = T2.id;

result:

1|a1|b1|1|c1|d1
2|a2|b2|2|c2|d2

As expected, you get all the columns of A, and all the columns of B. The 'id' column appears twice.

However, with the USING syntax:

select * T1 inner join B T2 using (id);

result:

1|a1|b1|c1|d1
2|a2|b2|c2|d2

The id column is now appearing exactly once. However, the situation is not so simple as that. It seems that what hapened was that the * expansion has not included two copies of the id. The following cases show that both copies of id are still logically in the join.

select T1.*, 'xxx', T2.* from A T1 inner join B T2 using (id);

result:

1|a1|b1|xxx|1|c1|d1
2|a2|b2|xxx|2|c2|d2

The T2.id column is part of the join, it just wasn't part of the *

In fact, looking further:

select T1.id, T1.a, T1.b, 'xxx', T2.id, T2.c, T2.d from A T1 inner join B T2 using (id);

result:

1|a1|b1|xxx|1|c1|d1
2|a2|b2|xxx|2|c2|d2

There is no doubt, T2.id is a valid column and can be used in expressions freely. That means the column cannot be removed from the type calculus.

Now in CQL, the * and T.* forms are automatically expanded; SQLite doesn't see the *. This is done so that if any columns have been logically deleted they can be elided from the result set. Given that this happens, the * operator will expand to ALL the columns. Just the same as if you did T1.* and T2.*.

As a result, in CQL, there is no difference between the USING form of a join and the ON form of a join.

In fact, only the select * form could possibly be different, so in most cases this ends up being moot anyway. Typically, you don't need to use * in the presence of joins because of name duplication and ambiguity of the column names of the result set. CQL's automatic expansion means you have a much better idea exactly what columns you will get - those that were present in the schema you declared.