ext/sqlite3: fix wrong pointer types passed to the free list comparator.

[devnexen]

SQLite3Stmt::close() and SQLite3Result::finalize() passed php_sqlite3_stmt pointer types instead of sqlite3_stmt pointers to zend_llist_del_element, causing the comparator to never match and both methods to silently become no-ops.

Regression introduced in 5eae6d1 ("Don't store the object zval directly").

[ndossche]

Are you sure?

https://github.com/devnexen/php-src/blob/18d287e5b50301dd36171e4f3fb0760b23a95021/ext/sqlite3/sqlite3.c#L2265
shows that ->stmt is dereferenced there, and the initialization shows:
https://github.com/devnexen/php-src/blob/18d287e5b50301dd36171e4f3fb0760b23a95021/ext/sqlite3/sqlite3.c#L2417
i.e. elements of type php_sqlite3_stmt * should be added.

[devnexen]

The pb resides in the second argument not the first

static int php_sqlite3_compare_stmt_free(php_sqlite3_stmt **stmt_obj_ptr, sqlite3_stmt *statement )

[devnexen]

by the way here an unchanged line which does not need fix.

[ndossche]

Okay but then the patch here is incomplete as well.
The signature of php_sqlite3_compare_stmt_free is wrong then, as well as its contents. Probably other calls to the linked list code should also be adapted.
In fact, looking at devnexen@bc74cff this code might've always been wrong?

[ndossche]

I believe this is right, but would like a second opinion too.

[ndossche]

I also believe this fixes a real long-standing bug that should be backported at a later point after this one is merged in 8.5.

[Girgias]

This looks sensible, but maybe @kamil-tekiela or @SakiTakamachi would be better reviewer as they are more familiar with DB extensions.

[kamil-tekiela]

If it fixes a bug, can we get a unit test?

[kamil-tekiela]

The test passes with no issues without this PR

[devnexen]

You're right, the test passes either way. This is a C-level type correctness fix — the comparator was receiving mismatched pointer types (php_sqlite3_stmt ** vs
sqlite3_stmt *), which is undefined behavior even if it happens to work on common platforms. There's no way to reliably trigger the UB from PHP userland, so the test
just exercises the code path but can't demonstrate the actual fix.