Let's start from basics, if it looks boring, feel free skip to the next post :-). I just feel I need to establish some terminology first.
The main building block of SQL is what the standard calls "query specification". It consists of "clauses": SELECT, FROM, WHERE, GROUP BY, HAVING (the last three are optional). They are executed in that order, with one exception: SELECT clause is executed last.
How is this different from a "SELECT statement"? First, UNION. Several query specifications may be combined using "set operators" (we only support UNION and UNION ALL at this point), and that will still be a single SELECT statement. Second, subqueries. Single SELECT statement (or "query") may contain several query specifications, some serving as subqueries (table or scalar). So, "query specification" is an all-encompassing term, meaning that it contains general-purpose stuff that would be good either in a top-level query or in a subquery. Top-level query, aka SELECT statement, also contains some specialized syntax that is not accepted in a subquery. I'll mention ORDER BY clause, COMPUTE clause, and FOR BROWSE as examples.
While I am on the topic of subqueries, table subqueries (aka "derived tables") deserve a special mention. This feature is what makes SQL "fully composable". This means that you can take any SQL query that delivers a relational stream and use it as a source for another query. It's amazing how many "hard to write" queries that don't really fit into a single query specification become possible to express if you use derived tables.
So, here is the test: we can only allow something in a query specification if it produces a regular relational stream, easily composable with other relational operators. So, the three example clauses I mentioned above as "top-level-only" features all have something that is not really relational. ORDER BY introduces ordering, which the relational model does not support. COMPUTE returns multiple result sets. FOR BROWSE adds "hidden columns" to the result set, something that does not completely fit into the relational data model.
This posting is provided "AS IS" with no warranties, and confers no rights.