Modify handling of SQLITE_SCHEMA in fts2 code. An SQLITE_SCHEMA error may cause SQLite to reload the internal schema, deleting and recreating v-table objects. So the sqlite3_vtab structure can be deleted out from under a v-table implementation. (CVS 4183) (check-in: [f9020cff] user: shess branch: trunk, size: 100359)

When creating fts tables in an attached database, the backing tables are created in database 'main'. This change propagates the appropriate database name to the routines which build sql statements.

Note that I propagate the database name and table name separately. I briefly considered just making the table name be "db.table", but it didn't fit so well in the model used to store the table name and other information, and having the db name passed separately seemed a bit more transparent. (CVS 3631) (check-in: [283385d2] user: shess branch: trunk, size: 99576)

Calling UPDATE against an fts table in a UTF-16 database inserts corrupted data into the database. The UTF-8 data is being inserted directly. This appears to happen because sqlite3_ value_text() destructively coerces a value to UTF-8, and it's never converted back when updating the table. This works around the problem by rearranging things so that the update happens before the coercion. (CVS 3596) (check-in: [4f2ab4b6] user: shess branch: trunk, size: 99070)

The virtual table interface allows for a cursor to field multiple xFilter() calls. For instance, if a join is done with a virtual table, there could be a call for each row which potentially matches. Unfortunately, fulltextFilter() assumes that it has a fresh cursor, and overwrites a prepared statement and a malloc'ed pointer, resulting in unfinalized statements and a memory leak.

This change hacks the code to manually clean up offending items in fulltextFilter(), emphasis on "hacks", since it's a fragile fix insofar as future additions to fulltext_cursor could continue to have the problem. (CVS 3521) (check-in: [18142fdb] user: shess branch: trunk, size: 99090)

We handle an UPDATE to a row by performing an UPDATE on the content table and by building new position lists for each term which appears in either the old or new versions of the row. We write these position lists all at once; this is presumably more efficient than a delete followed by an insert (which would first write empty position lists, then new position lists). (CVS 3434) (check-in: [757fa224] user: adamd branch: trunk, size: 96723)

For now, each posting list stores position/offset information for multiple columns. We may implement separate posting lists for separate columns at some future point. (CVS 3408) (check-in: [366a70b0] user: adamd branch: trunk, size: 65307)

An earlier change had additionally changed appropriate sqlite3_bind_text() calls to sqlite3_bind_blob(). I've found that this changes what's actually stored in the database, so backed those changes out. Also (and this is weird), I found that I could no longer do straight-forward = queries against %_term.term at a command-line. (CVS 3379) (check-in: [5844db1a] user: shess branch: trunk, size: 45799)