BUG #17795: Erroneous parsing of floating-poing components in DecodeISO8601Interval() - Mailing list pgsql-bugs

From PG Bug reporting form
Subject BUG #17795: Erroneous parsing of floating-poing components in DecodeISO8601Interval()
Date
Msg-id 17795-748d6db3ed95d313@postgresql.org
Whole thread Raw
Responses Re: BUG #17795: Erroneous parsing of floating-poing components in DecodeISO8601Interval()  (Tom Lane <tgl@sss.pgh.pa.us>)
List pgsql-bugs
The following bug has been logged on the website:

Bug reference:      17795
Logged by:          Alexander Lakhin
Email address:      exclusion@gmail.com
PostgreSQL version: 15.2
Operating system:   Ubuntu 22.04
Description:

I've encountered a pair of relatively new anomalies when using ISO-8601
intervals.
Firstly, a float value passed as a component of the interval
can produce an overflow (I performed the following exercises with clang 14
and gcc 11.3).
Let's begin with an integer component:
select interval 'P178956970Y';
178956970 years -- OK (178956970 * 12 == 2 ^ 31 - 8)
select interval 'P178956971Y';
interval out of range -- OK (178956971 * 12 == 2 ^ 31 + 4)

Compare with a float value:
select interval 'P.178956970e9Y';
178956970 years -- OK
And:
select interval 'P.178956970e9Y6M';
178956970 years 6 mons -- OK
But:
select interval 'P.178956971e9Y';
-178956970 years -8 mons -- not OK, previously: interval out of range

Though:
select interval 'P.178956970e9Y8M';
interval field value out of range -- OK
But:
select interval 'P.178956971e9Y8M';
-178956970 years  -- not OK, previously: interval out of range

As I can see, this is explained by the following code:
int         extra_months = (int) rint(frac * scale * MONTHS_PER_YEAR);
return !pg_add_s32_overflow(itm_in->tm_mon, extra_months,
&itm_in->tm_mon);

Here, when calculating extra_months with an out-of-range float,
we get the sentinel value: -2147483648 == -0x80000000.
And then pg_add_s32_overflow(0, -0x80000000, ...) (internally
__builtin_add_overflow() for me)
happily returns "no overflow".
For the case 'P.178956970e9Y8M' an overflow detected because extra_months
is
near the upper limit, but is not the sentinel value.
(BTW, clang' ubsan doesn't like conversion
"(int)out_of_int_range_number".)

And for the sake of completeness, the upper limit of the double type:
select interval 'P.17976931348623158e309Y';
-178956969 years -8 mons -- not OK, previously: interval field value out of
range
select interval 'P.17976931348623159e309Y';
invalid input syntax for type interval -- OK (out of the double range)

Here "previously" is the behavior observed on e39f99046~1.

The second anomaly is with parsing float values with an integer part:
select interval 'P.0e100Y';
00:00:00 -- OK
But:
select interval 'P1.0e100Y';
1 year -- previously: invalid input syntax for type interval
select interval 'P1.0e-100Y';
1 year -- previously: invalid input syntax for type interval

IIUC, the integer part is just added to the result of parsing a rest with
the scientific notation.

Similar anomaly can be seen with the D part:
select interval 'P.1e9D';
2400000000:00:00 -- OK, previously: 100000000 days
But
select interval 'P.1e10D';
-2562047788:00:54.775807 -- not OK, previously: 1000000000 days

select interval 'P1.1e9D';
1 day 2400000000:00:00 -- not OK, previously: 1100000000 days

Probably all these anomalies can be eliminated by disabling
the scientific notation (though it more or less worked previously),
just as it is not supported for ordinary (not ISO-8601) intervals.


pgsql-bugs by date:

Previous
From: Alexander Bluce
Date:
Subject: Re: BUG #17782: ERROR: variable not found in subplan target lists
Next
From: Kyotaro Horiguchi
Date:
Subject: Re: BUG #17789: process_pgfdw_appname() fails for autovacuum workers