Обсуждение: [PATCH] replace float8 with int in date2isoweek() and date2isoyear()
Hi PG hackers,
I found suspicious use of float8 in date2isoweek() and date2isoyear(). In both
cases float8 is only used for storing the value, while the entire calculation
on the right happens in integers:
float8 result = (dayn - (day4 - day0)) / 7 + 1;
At the end date2isoweek() returns `result' converted back to int:
return (int) result;
float8 here is confusing and a bit slow.
I run a benchmark with optimization -O3, and found that using int is 3% faster:
BM_Float/1000/3/8_mean 5.12 ns 5.12 ns 100
BM_Int/1000/3/8_mean 4.95 ns 4.95 ns 100
BM_Long/1000/3/8_mean 4.96 ns 4.96 ns 100
Without optimization (-O0), speedup even better around 30%:
BM_Float/1000/3/8_mean 19.6 ns 19.6 ns 100
BM_Int/1000/3/8_mean 14.8 ns 14.8 ns 100
BM_Long/1000/3/8_mean 16.7 ns 16.7 ns 100
Additionally, as a paranoia measure I run a comparison test with the following
ranges: year={-100000,100000}, month={-100,+100}, day={-100,+100}.
As expected float and int did not produce any difference.
Attached patch replaces `float8 result' with `int result'.
I think there is no need in adding an extra test case here, because
date2isoweek and date2isoyear are covered by three regression tests:
* date
SELECT f1 as "date",
...
date_part('isoyear', f1) AS isoyear,
date_part('week', f1) AS week,
date_part('dow', f1) AS dow,
date_part('isodow', f1) AS isodow,
...
FROM date_tbl;
* timestamp
SELECT date_trunc( 'week', timestamp '2004-02-29 15:44:17.71393' ) AS week_trunc;
* timestamptz
SELECT date_trunc( 'week', timestamp with time zone '2004-02-29 15:44:17.71393' ) AS week_trunc;
---
Sergey
Вложения
=?utf-8?q?=D0=A4=D1=83=D0=BA=D0=B0=D0=BD=D1=87=D0=B8=D0=BA_=D0=A1=D0=B5=D1=80=D0=B3=D0=B5=D0=B9?=
<s.fukanchik@postgrespro.ru>writes:
> Hi PG hackers,
> I found suspicious use of float8 in date2isoweek() and date2isoyear(). In both
> cases float8 is only used for storing the value, while the entire calculation
> on the right happens in integers:
> float8 result = (dayn - (day4 - day0)) / 7 + 1;
> At the end date2isoweek() returns `result' converted back to int:
> return (int) result;
> float8 here is confusing and a bit slow.
I looked into our git history to try to find out why it's like this.
The answer seems to be that commit dffd8cac3 created date2isoweek()
by splitting out pre-existing code that had been in timestamp_part().
In that context the code had been using a float8 "result" variable
that was shared with other switch cases, and that variable's type
was just blindly copied into date2isoweek(). Then 1c757c49f again
copied-and-pasted while creating date2isoyear().
I agree with getting rid of the unnecessary usage of float8 here,
but there's another aspect that's bugging me: "result" is a totally
misleading variable name in date2isoyear(), because it's *not*
the function's result. I'm inclined to rename it to "week", and
then to keep these functions looking as parallel as possible,
I'd probably do the same in date2isoweek().
> I think there is no need in adding an extra test case here, because
> date2isoweek and date2isoyear are covered by three regression tests:
Agreed, the code coverage report shows these are covered.
regards, tom lane
Sure, I'm attaching v2 of the patch with "result" renamed to "week". -- Sergey On 7/12/25 18:15, Tom Lane wrote: > =?utf-8?q?=D0=A4=D1=83=D0=BA=D0=B0=D0=BD=D1=87=D0=B8=D0=BA_=D0=A1=D0=B5=D1=80=D0=B3=D0=B5=D0=B9?= <s.fukanchik@postgrespro.ru>writes: >> Hi PG hackers, >> I found suspicious use of float8 in date2isoweek() and date2isoyear(). In both >> cases float8 is only used for storing the value, while the entire calculation >> on the right happens in integers: >> float8 result = (dayn - (day4 - day0)) / 7 + 1; >> At the end date2isoweek() returns `result' converted back to int: >> return (int) result; >> float8 here is confusing and a bit slow. > I looked into our git history to try to find out why it's like this. > The answer seems to be that commit dffd8cac3 created date2isoweek() > by splitting out pre-existing code that had been in timestamp_part(). > In that context the code had been using a float8 "result" variable > that was shared with other switch cases, and that variable's type > was just blindly copied into date2isoweek(). Then 1c757c49f again > copied-and-pasted while creating date2isoyear(). > > I agree with getting rid of the unnecessary usage of float8 here, > but there's another aspect that's bugging me: "result" is a totally > misleading variable name in date2isoyear(), because it's *not* > the function's result. I'm inclined to rename it to "week", and > then to keep these functions looking as parallel as possible, > I'd probably do the same in date2isoweek(). > >> I think there is no need in adding an extra test case here, because >> date2isoweek and date2isoyear are covered by three regression tests: > Agreed, the code coverage report shows these are covered. > > regards, tom lane
Вложения
Sergey Fukanchik <s.fukanchik@postgrespro.ru> writes:
> I'm attaching v2 of the patch with "result" renamed to "week".
I pushed it already, actually.
regards, tom lane
Thank you! -- Sergey On 7/13/25 18:39, Tom Lane wrote: > Sergey Fukanchik <s.fukanchik@postgrespro.ru> writes: >> I'm attaching v2 of the patch with "result" renamed to "week". > I pushed it already, actually. > > regards, tom lane