Regression. Result of cross join of two procedures + unioned with common datasource (e.g. rdb$database) is wrong since 5.0.3.1675

[pavel-zotov]

Consider following script:

set bail on;
shell del r:\temp\tmp4test.fdb 2>nul;
create database 'localhost:r:\temp\tmp4test.fdb' user 'sysdba' password 'masterkey';
set echo on;

recreate table test(x int);
insert into test select row_number()over() from rdb$types rows 5;
commit;

set term ^;
create or alter procedure sp_issue_row_1 (a_x int default 1) returns (x int) as
begin
    x = a_x + 1;
    suspend;
end
^ 
set term ;^
commit;

set explain on;
with
c as (
    select first 1 p2.x
    from sp_issue_row_1 p1
    cross join sp_issue_row_1 (p1.x) p2
    union all
    select 4 from rdb$database
)
select * from c
join test n on n.x = c.x
;

Its output on 5.0.3.1674:

       X            X
======== ============
       3            3
       4            4

(and this is expected)

But since 5.0.3.1675 result became wrong:

       X            X
======== ============
       3            1
       3            2
       3            3
       3            4
       3            5
       4            1
       4            2
       4            3
       4            4
       4            5

(same on 6.x)

Key notes:

1. p1 must be CROSS joined with p2
2. input argument (p1.x) must be specified for procedure that is in the right part of cross join (alias p2)