Re: Regression in statement locations

On Tue, May 20, 2025 at 08:38:47AM +0900, Michael Paquier wrote:
> With the semicolon in place, stmt_len gets set for the last query of
> the string.  Still digging more..

And got it.  The problem is that we are failing to update the
statement location in a couple of cases with subqueries, and that we
should handle (p_stmt_len == 0) as of using the remaining bytes in
the string when a location is available, but the code was too
aggressive in thinking that the length = 0 case should be always
discarded.

Once I have fixed that, I've been a bit puzzled by the difference in
output in the tests of pg_overexplain, but I think that the new output
is actually the correct one: the queries whose plan outputs have
changed are passed as arguments of the explain_filter() function,
hence the location of the inner queries point at the start location of
the inner query instead of the start of the top-level query.  Note
that if you add a semicolon at the end of these three queries in the
pg_overexplain tests, we finish with an end location reported.

I have also played with 499edb0 reverted and noted that the results of
pg_overexplain were inconsistent when the module has been originally
introduced, with two queries choking a bit.
--
Michael