boost :: locale :: date_time的陷阱

Question

Could you provide the pitfalls you meet in using boost::locale::date_time in boost 1.55 or tell me what's wrong I have made in the following examples? Do I misunderstand boost::locale::date_time API?

day_of_week Example 1: Shift A Date to A Sunday

#include <iostream>
#include <boost/locale.hpp>

int main() {
  using namespace boost::locale;
  generator gen;
  std::locale locale = gen("en_US.UTF-8");
  std::locale::global(locale);
  std::cout.imbue(locale);

  date_time_period_set s;
  s.add(period::year(2013));
  s.add(period::month(2));
  s.add(period::day(5));
  s.add(period::hour(9));
  s.add(period::minute(0));
  s.add(period::second(0));

  // 2013-03-05 is a Tuesday, let's shift it to a Sunday
  s.add(period::day_of_week(1));

  date_time now(s);// why it's not 2013-03-02 or 2013-03-10?
  std::cout << now << std::endl;
}

Output: Mar 30, 2014, 9:00:00 AM , why it's not 2013-03-02, 9:00:00 AM or 2013-03-10, 9:00:00 AM ? It's even not a Sunday.

day_of_week Example 2: Last Saturday of March 2013

#include <iostream>
#include <boost/locale.hpp>

int main() {
  using namespace boost::locale;
  generator gen;
  std::locale locale = gen("en_US.UTF-8");
  std::locale::global(locale);
  std::cout.imbue(locale);

  date_time_period_set s;
  s.add(period::year(2013));
  s.add(period::month(2));
  s.add(period::hour(9));
  s.add(period::minute(0));
  s.add(period::second(0));


  // swapping the following 2 lines rusults in a wrong date
  s.add(period::day_of_week(7));
  s.add(period::day_of_week_in_month(-1));


  date_time now(s);// last Saturday of March 2013
  std::cout << now << std::endl;
}

It outputs: Mar 30, 2013, 9:00:00 AM

But if you change 2 lines marked in the source code, you might see: Apr 5, 2014, 9:00:00 AM

Where and How to Use date_time_period boost::locale::period::first_day_of_week(int v) ?

See date_time_period boost::locale::period::first_day_of_week(int v) .

Answer 1

This is not an answer to this question. I am not well-versed in boost::locale::date_time . However I am responding anyway with a possible alternative. This alternative is not as well packaged as boost::date_time , nor near as complete, especially with respect to locales. However it might be a viable alternative for you, because your needs look simple, and this alternative is very simple.

I am speaking about:

chrono-Compatible Low-Level Date Algorithms

This is nothing but a paper outlining 10 very low-level building blocks for date (not time) computations, such as those in your question. The above paper does not present a library you can download. Instead it just lists and explains the 10 algorithms. The code for each algorithm is so short, it is quite practical to copy/paste into your own code, and modify for your specific needs. The code is in the public domain, so no worries on copyright issues.

Using 5 of the 10 algorithm for this paper, combined with the following bit of boiler-plate for outputting a date (I've used std::tuple to hold the date, use whatever you like), I've come up with very simple solutions to your simple problems:

// here is my copy of the 10 algorithms
#include "../date_performance/date_algorithms"
#include <string>
#include <iostream>

// I'm being lazy in defining my date class
using date = std::tuple<int, unsigned, unsigned>;

// I'm being lazy in defining I/O for my date class
std::string
to_string(date const& d)
{
    int year;
    unsigned month;
    unsigned day;
    std::tie(year, month, day) = d;
    std::string r = std::to_string(year) + '-';
    if (month < 10)
        r += '0';
    r += std::to_string(month) + '-';
    if (day < 10)
        r += '0';
    r += std::to_string(day);
    return r;
}

day_of_week Example 1: Shift A Date to A Sunday

Here is how I would construct a date such as Mar. 5, 2013:

date d(2013, 3, 5);

In order to find the Sunday on or prior to this date, I first need to find out what day of the week this date is on. This is a two-step process using these low-level building blocks:

1. Convert this date to a serial date: A count of days since some epoch.
2. Convert the serial date to a day of the week.

This is done with:

int s = days_from_civil(std::get<0>(d), std::get<1>(d), std::get<2>(d));
unsigned wd = weekday_from_days(s);

Or if you prefer:

int s = days_from_civil(2013, 3, 5);
unsigned wd = weekday_from_days(s);

Next it is good practice to name the constants for the days of the week. This code follows the C and C++ conventions, for example:

constexpr unsigned sun = 0;

Now the day preceding this day that is a Sunday is the serial date, subtracted by the number of days this weekday is past Sunday (which might be 0 if this is a Sunday, or as great as 6 if this weekday is a Saturday):

date d1 = civil_from_days(s - weekday_difference(wd, sun));

When I print this out:

std::cout << "The Sunday before was " << to_string(d1) << '\n';

I get:

The Sunday before was 2013-03-03

This is very little code to address your first question. If instead you wanted the Sunday after 2013-03-05, that is just as easy:

date d2 = civil_from_days(s + weekday_difference(sun, wd));
std::cout << "The Sunday after was " << to_string(d2) << '\n';

which outputs:

The Sunday after was 2013-03-10

You could easily use these low-level algorithms to write your own higher-level algorithm to compute what you want.

day_of_week Example 2: Last Saturday of March 2013

The first thing you need is find the last day of the month for this year:

unsigned last = last_day_of_month(2013, 3);

In this example we all know that this is the same as:

unsigned last = 31;

But if the year and month were run time values, last_day_of_month comes in handy. Now we need the weekday of this last day of the month. As before:

unsigned wd_last = weekday_from_days(days_from_civil(2013, 3, last));

And now we just need to subtract from last the number of days this last weekday is past Saturday. Assuming we had set up an informative constant ( sat == 6 ), this looks like:

unsigned day = last - weekday_difference(wd_last, sat);

day is the day of the month that is the last Saturday of March in 2013. We could print that out with:

std::cout << "The last Saturday of March 2013 was " << to_string(date(2013, 3, day)) << '\n';

And get:

The last Saturday of March 2013 was 2013-03-30

And again, it would be easy for you to wrap all this up in a neat function. Indeed, the paper even has such an example function in it.

If these low-level date algorithms are helpful to you, or any other readers, then great, otherwise never mind. :-)

Answer 2

It's rather an old question, but I leave my findings here for the record.

Looking up the source, the only place in date_time class that actually uses first_day_of_week is date_time::get() function. You cannot not use it for building date_time , or other fancy things. The only place you are allowed to use it is when you do something like mydt.get(period::first_day_of_week) , which returns a int in the range [1,7] . It is essentially the same as calendar::first_day_of_week() .

Note that int v in boost::locale::period::first_day_of_week(int v) will be effectivly ignored. int v sets amount of date_time_period , but no API will respect amount value of first_day_of_week . It is just auto-generated API, I guess.

It is helpful to look into ICU documentation rather than Boost.locale documentation, because Boost.locale is really a thin layer on top of ICU. You will find documentation of ICU calendar class that is used inside Boost.locale here .

Your example 1 leads to what ICU documentation refers to "Inconsistent information". You specified day and day-in-week that is inconsistent, and ICU will not try to correct day using provided day-in-week. It just consider it inconsistent, thus showing strange behavior you mentioned. You will need to specify week-in-month instead of day to avoid this inconsistency.

Your example 2 is much more interesting. You got it right when you specified the day-of-week first, and got it wrong when you set day-of-week-in-month first. I guess the logic is like this:

When you first set day-of-week-in-month to -1, ICU has insufficient information. So as in documentation, it tries to fill remaining field(day-of-week) to default, ie 1. The calendar is set to Mar 31, 2013 now, which is last sunday in 2013 March. Then you set day-of-week to 7, so it travels to saturday of THAT WEEK, which is apr...6? I don't know why exactly you got Apr 5, 2014, 9:00:00 AM , but I believe the logic process is similar to this. Always set day-of-week first to avoid this weird behavior.

I think ICU calendars are unintuitive a bit. Java also used DateTime API very similar to ICU calendar until SE 7, and Java programmers considered the standard DateTime library a nightmare. (On release of SE 8, they replaced the ICU-like API with better one .)

You can consider switching to Howard Hinnant's Date library , which has far better API.