REGEXP_MATCHES() strange behavior with '^' and '$' pattern

Hi,

While playing with regular expression I found some strange behavior of
regexp_matches() function.

Consider following sql query and its output:

postgres=# select regexp_matches('1' || chr(10) || '2' || chr(10) || '3' ||
chr(10) || '4', '^', 'mg');
regexp_matches
----------------
{""}
{""}
{""}
{""}
{""}
{""}
{""}
(7 rows)

It suppose to return me 4 rows and not 7. Similar behavior found with
pattern '$'.

It seems that these start and end anchor characters are not matching
correctly. Or rather they are matching twice.

To get a root cause of it, I put elog(INFO,..) into the
setup_regexp_matches() function where we copy matches into the struct and
found following values.

postgres=# select regexp_matches('1' || chr(10) || '2' || chr(10) || '3' ||
chr(10) || '4', '^', 'mg');
INFO: start_search: 0 rm_so: 0 rm_eo: 0
INFO: updated start_search: 1
INFO: start_search: 1 rm_so: 2 rm_eo: 2
INFO: updated start_search: 2
INFO: start_search: 2 rm_so: 2 rm_eo: 2
INFO: updated start_search: 3
INFO: start_search: 3 rm_so: 4 rm_eo: 4
INFO: updated start_search: 4
INFO: start_search: 4 rm_so: 4 rm_eo: 4
INFO: updated start_search: 5
INFO: start_search: 5 rm_so: 6 rm_eo: 6
INFO: updated start_search: 6
INFO: start_search: 6 rm_so: 6 rm_eo: 6
INFO: updated start_search: 7

Certainly, after second pass, updated start_search should be 3 as last
matched pattern was at 2 and of zero length since so = eo.

I have modified that logic to look similar as that of replace_text_regexp()
function. As regexp_replace works well.

Attached patch with test-case. Please have a look and let me know if I
assumed something wrong.

Thanks

--
Jeevan B Chalke