stephan [Tue, 12 Jul 2022 09:40:27 +0000 (09:40 +0000)]
Renamed SQLITE_SHELL_WASM_MODE to SQLITE_SHELL_WASM_WEB_MODE and no longer automatically enable it if __EMSCRIPTEN__ is defined, in order to facilitate using Emscripten to build the shell for CLI-based WASM runtimes (which cannot make use of the web-specific user input changes). The fiddle build now explicitly passes the new flag on at compile-time.
Back out the pager performance enhancement at [a1c090e08139f99d3], because it
turns out we should never allow a zero key into the pcache interface according
to the design specs, even if that page is immediately released without ever
being used.
Add the --config option to wapptest.tcl. The argument is a glob pattern. All
configurations must match the glob pattern in order to run. The default value
is "*".
In the sqlite_stmt extension, store the result of strlen() in a 64-bit
integer to avoid a compiler warning, even though we know that the length
will always fit comfortably in 32 bits.
Ensure that the Parse.nErr flag is set following an SQLITE_TOOBIG error on
a nested parse. Fix for the problem identified by
[forum/forumpost/d5a82ba9eedee30c | forum post d5a82ba9eedee30c].
Also, remove unnecessary clearing of the Parse.zErrMsg field
following a nested parse.
Omit an unnecessary structure initialization in whereLoopAddAll(), replacing
it with an assert() to show that the initialization has already occurred, for
a small performance increase and size reduction.
In wherePathSolver(), defer initializing variables until they are actually
needed, in case they are not needed. This gives a small performance increase.
dan [Wed, 6 Jul 2022 15:44:57 +0000 (15:44 +0000)]
Avoid ignoring the last line of a csv file if the final field is empty and
there is no trailing CFLS. Also have the csv extension treat the last line of a
file in the same way as any other line if it is short fields.
Enhance the REGEXP extension so that the end-of-input indicate ("$") is
allowed to occur on one branch of an OR ("|").
[forum:/forumpost/0107d5d40dd273e2|Forum post 0107d5d40dd273e2], second
issue.
Fix the initial-prefix optimization for the REGEXP extension such that it
works even if the prefix contains characters that require a 3-byte UTF8
encoding. This should fix the problem reported by
[forum:/forumpost/96692f8ba5|forum post 96692f8ba5].
Performance improvement in resolveP2Values(). Save over 1 million CPU
cycles by omitting the loop termination condition and exiting when the
OP_Init (always the first opcode in any VDBE program) is encountered.
drh [Sat, 25 Jun 2022 19:43:44 +0000 (19:43 +0000)]
Allow flattening of a subquery that is the right operand of a LEFT JOIN
in an aggregate query as long as there is no GROUP BY clause. (The GROUP BY
clause will interfere with the operation of the TK_IF_NULL_ROW expression
nodes.)
drh [Sat, 25 Jun 2022 02:39:29 +0000 (02:39 +0000)]
Change a harmless assert() into a testcase(). The actually test case code
is in TH3. Fix for the problem described at
[forum:/info/ed29e196d5c4f3d5|forum post ed29e196d5c4f3d5].
drh [Wed, 22 Jun 2022 14:43:52 +0000 (14:43 +0000)]
Allow more line with on the debugging output for SrcItem elements in the
parse tree. This affects debugging builds only and is a no-op for production
builds.
drh [Tue, 21 Jun 2022 13:41:24 +0000 (13:41 +0000)]
Allow a HAVING clause on any aggregate query, even if there is no GROUP BY
clause. This brings SQLite into closer agreement with PostgreSQL and fixes
the concern raised by
[forum:/forumpost/1a7fea4651|forum post 1a7fea4651].
drh [Mon, 20 Jun 2022 18:26:14 +0000 (18:26 +0000)]
Do not allow FROM-clause terms on the left side of a RIGHT or FULL JOIN to
be reordered. [forum:/forumpost/6650cd40b5634f35|forum post 6650cd40b5634f35].
This is probably more strict that necessary to get correct behavior,
but for the first release that supports RIGHT/FULL JOIN it is perhaps better
to be correct than fast. A less strict constraint might be to prohibit
FROM-clause terms that originate on the left side of a RIGHT JOIN from
crossing from the right side to the left side of a LEFT JOIN. Revisit this
later.
drh [Mon, 20 Jun 2022 17:04:44 +0000 (17:04 +0000)]
Do not allow an ON clause to references tables to its right if there is a
RIGHT or LEFT join anywhere in the query. Other RDBMSes prohibit this always,
but SQLite must allow ON clauses to reference tables to their right for legacy
compatibility, unless there is a RIGHT or LEFT join someplace in the query,
in which case there is no legacy to support.
drh [Mon, 20 Jun 2022 12:42:28 +0000 (12:42 +0000)]
The fix at [cab9b4cccd13bf0a] was incomplete, as demonstrated by
[forum:/forumpost/57bdf2217d|forum post 57bdf2217d]. This check-in
should complete the fix.
drh [Sun, 19 Jun 2022 16:55:07 +0000 (16:55 +0000)]
Follow-up to check-in [0057bbb508e7662b] - ensure that the database page
has been initialized prior to continuing with the optimization. If the page
is not initialized, that indicates that the database is corrupt.
dbsqlfuzz 09ee46becd5e6d1b2a55c9f8ad767335a90aadb0.
drh [Sat, 18 Jun 2022 20:20:30 +0000 (20:20 +0000)]
Enable query invariant checking in fuzzcheck by default. There is no way
to turn it off. Update the invariant checking logic to be consistant with
dbsqlfuzz.
drh [Fri, 17 Jun 2022 21:31:30 +0000 (21:31 +0000)]
Fix the OP_Concat operator such that when concatenating a BLOB with an
odd number of bytes on a database that is UTF16, the size of the resulting
string is reduced to a multiple of two.
drh [Fri, 17 Jun 2022 15:11:31 +0000 (15:11 +0000)]
Fix the virtual table detection mechanism to avoid false-positives that were
blocking all failures. Then fix a few of the additional problems that are
revealed by that fix. More fixes are needed.
drh [Fri, 17 Jun 2022 12:25:33 +0000 (12:25 +0000)]
Fix the new --query-invariants option on fuzzcheck so that it does not
use an unprotected sqlite3_value object as an argument to sqlite3_value_int64().
drh [Wed, 15 Jun 2022 14:57:04 +0000 (14:57 +0000)]
When running an incremental vacuum, detect growth in the size of the database
file (which can only occur if the file is corrupt) and fail with SQLITE_CORRUPT.
drh [Wed, 15 Jun 2022 12:32:27 +0000 (12:32 +0000)]
Dbsqlfuzz discovered a case where a bytecode branch is in fact taken, so change
the designator from VdbeCoverageNeverTaken() to VdbeCoverage(). Test case
in TH3.