Hi Zsolt Parragi, Thanks for the review! Attached is v4. it nice catch with the old bug i have the change draft base on your suggestions The bellowing attachment is current my update approach for your review
I changed the clock-time assignment to addition, added tests for fractional days/weeks in both orders, and fixed the comments. Negation now avoids the newer int128 helpers to ease backpatching. Once we agree on the approach, I think i can prepare and test backpatches for the affected maintained branches, back to PG15 upward ? Could you take another look would you also prefer the fractional day/week fix as a separate patch? Regards, Vào Thứ 3, 17 thg 11, 2026 vào lúc 05:23 Zsolt Parragi < [email protected]> đã viết: > Hello > > Thanks for working on this! > > + itm_in->tm_usec = int64_to_int128(itm.tm_usec); > + int128_add_int64_mul_int64(&itm_in->tm_usec, itm.tm_hour, > USECS_PER_HOUR); > + int128_add_int64_mul_int64(&itm_in->tm_usec, itm.tm_min, > USECS_PER_MINUTE); > > There's also a related preexisting bug: time field overwrites the > fractional part of a day/week unit > > SELECT '01:00:00 1.5 days'::interval AS a, '1.5 days 01:00:00'::interval > AS b; > a | b > ----------------+---------------- > 1 day 01:00:00 | 1 day 13:00:00 > > SELECT '01:00 0.5 weeks'::interval, '0.5 weeks 01:00'::interval; > 3 days 01:00:00 | 3 days 13:00:00 > > int128_add_int64(&itm_in->tm_usec, itm.tm_usec) would also fix this > > > There are also some stale comments now: > > * dtype and itm_in are output parameters. > ... > * itm_in remains undefined for infinite interval values for which dtype > alone > * suffices. > > > +/* Private wide arithmetic for the general interval parser. */ > +static INT128 > +NegateIntervalValue(INT128 value) > +{ > + INT128 result = make_int128(~PG_INT128_HI_INT64(value), > + > ~PG_INT128_LO_UINT64(value)); > + > + int128_add_int64(&result, 1); > + return result; > +} > + > > The bug exists on 15+, and this function uses PG19 functions - that is > fine, but this part will need a different solution on older branches. > > > +/* > + * Keep general interval input wider than its stored representation until > AGO > + * has been applied. There are at most MAXDATEFIELDS int64 input values, > and > + * even the largest unit scale fits comfortably in this accumulator. > + */ > > I am not sure how good this comment is outside this patch, as it > mostly seems to explain implementation details, not the type. >
v4-0001-Fix-interval-input-at-minimum-signed-field-values.patch
Description: Binary data
