ARROW-12011: [C++] Fix crashes and incorrect results when printing extreme date values

[pitrou]

The arrow_vendored::date library represents year numbers as a C short
and may silently wraparound its value
(but also throw an exception if the year has the value -32768).

[github-actions]

https://issues.apache.org/jira/browse/ARROW-12011

[pitrou]

@jorisvandenbossche Would you like to review some C++ code?

[bkietz]

Some nits, otherwise this looks good

[bkietz]

Nit: please extract named constants like

Suggested change

	if (v >= -12687428 && v <= 11248737) {
	if (v >= kMinimumDay && v <= kMaximumDay) {

[bkietz]

The range check could possibly even just be here:

Suggested change

	(*sink_) << arrow_vendored::date::format(fmt, epoch_ + Unit{value});
	constexpr Unit kMin = duration_cast<Unit>(kMinDay);
	constexpr Unit kMax = duration_cast<Unit>(kMaxDay);
	if (Unit{value} >= kMin && Unit{value} <= kMax) {
	(*sink_) << arrow_vendored::date::format(fmt, epoch_ + Unit{value});
	} else {
	WriteOutOfRange(value);
	}

[pitrou]

Hmm... nice idea, but it unfortunately overflows for nanoseconds.

/home/antoine/arrow/dev/cpp/src/arrow/pretty_print.cc: In instantiation of 'void arrow::{anonymous}::ArrayPrinter::FormatDateTime(const char*, int64_t, bool) [with Unit = std::chrono::duration<long int, std::ratio<1, 1000000000> >; int64_t = long int]':
/home/antoine/arrow/dev/cpp/src/arrow/pretty_print.cc:505:71:   required from here
/home/antoine/arrow/dev/cpp/src/arrow/pretty_print.cc:483:59:   in 'constexpr' expansion of 'std::chrono::duration_cast<std::chrono::duration<long int, std::ratio<1, 1000000000> >, int, std::ratio<86400, 1> >(std::chrono::duration<int, std::ratio<86400, 1> >(-12687428))'
/usr/include/c++/9/chrono:200:21:   in 'constexpr' expansion of 'std::chrono::__duration_cast_impl<std::chrono::duration<long int, std::ratio<1, 1000000000> >, std::ratio<86400000000000, 1>, long int, false, true>::__cast<int, std::ratio<86400, 1> >((* & __d))'
/home/antoine/arrow/dev/cpp/src/arrow/pretty_print.cc:483:20: error: overflow in constant expression [-fpermissive]
  483 |     constexpr Unit kMin = std::chrono::duration_cast<Unit>(arrow_vendored::date::days{-12687428});
      |                    ^~~~

[pitrou]

Ok, it works by templating a bit differently.

[jorisvandenbossche]

I didn't yet look at the code, but one note: the strftime compute kernel has the same problem (at least for timestamp type, since it doesn't yet support date32/64 formatting. Checking the limits encoded in this PR for timestamp, using one out of range crashes with strftime).

That can certainly be for another JIRA, but wondering a bit if it might be possible to share some of those checks (the printing should also work if compute is not available I suppose, so cannot directly rely on it)

[pitrou]

@jorisvandenbossche Yes, see https://issues.apache.org/jira/browse/ARROW-13738 and also https://issues.apache.org/jira/browse/ARROW-13736

[pitrou]

I will merge this, feel free to post any comments afterwards, though.