Thread: BUG #4231: Interval silent seconds 32-bit wraparound when given as text

The following bug has been logged online:

Bug reference:      4231
Logged by:          Geoff Tolley
Email address:      geoff@polimetrix.com
PostgreSQL version: 8.3.1
Operating system:   CentOS 4.6 x86_64 (custom-compiled Linux 2.6.24.2 kernel
unpatched from kernel.org)
Description:        Interval silent seconds 32-bit wraparound when given as
text
Details:

When you make an interval by giving it a text representation of a number of
seconds, it appears to have a silent 32-bit wraparound (the example here is
from some raw SPSS data from a colleague; apparently it uses October 14th
1584 as its base so the seconds output tends to be >> 2^31):


select '1584-10-14 00:00'::timestamp + '13424373496 seconds'::interval;
      ?column?
---------------------
 1601-11-17 21:13:28
(1 row)

select 13424373496::bigint % (2 ^ 32)::bigint;
 ?column?
-----------
 539471608
(1 row)

select '1584-10-14 00:00'::timestamp + '539471608 seconds'::interval;
      ?column?
---------------------
 1601-11-17 21:13:28
(1 row)


However, the error is actually raised if it's not a string:

select '1584-10-14 00:00'::timestamp + 13424373496::interval;
ERROR:  cannot cast type bigint to interval
LINE 1: ...elect '1584-10-14 00:00'::timestamp + 13424373496::interval;
                                                              ^


HTH,
Geoff

"Geoff Tolley" <geoff@polimetrix.com> writes:
> When you make an interval by giving it a text representation of a number of
> seconds, it appears to have a silent 32-bit wraparound (the example here is
> from some raw SPSS data from a colleague; apparently it uses October 14th
> 1584 as its base so the seconds output tends to be >> 2^31):

Hmm ... it works properly on a 32-bit machine ...

There are a boatload of occurrences of this pattern in datetime.c:

                    integer_variable = strtol(field[i], &cp, 10);
                    if (errno == ERANGE)
                        return DTERR_FIELD_OVERFLOW;

which is okay if long and int are the same size, but not so much
otherwise.  Looks like we need to fix that.

            regards, tom lane