The implementation previously worked for non-short date-times, where
the offset has been remembered since construction. This included the
case of zoned times (and local times more than 2^55 msec away from the
start of 1970) that hit a spring-forward's gap; but excluded local
times that did the same (within 2^55 msec of the epoch).
This precluded an offset check in a spring-forward test, now added.
We can in fact determine the offset whenever we got a valid date and
time (we do so in the course of initializing the object, and when
asked for toMSecsSinceEpoch(), even when invalid), and we should not
use the value of the recorded offset if we didn't get a valid date and
time, so amend to always return 0 if we didn't get valid date and time
and always report the correct offset otherwise.
In the process, amend offsetFromUtc()'s computation to directly
resolve the date-time, rather than doing so via toMSecsSinceEpoch(),
which has to repeat decision-making offsetFromUtc() has already done
by the time it calls it. Also amend toMSecsSinceEpoch() to return 0 if
we didn't have a valid date and time to begin with, so it only
attempts to produce a useful result in the case where construction
attempted to resolve the date-time.
Change-Id: I6574e362275ccc4fbd8de6f0fa875d2e50f3bffe
Reviewed-by: Mårten Nordheim <marten.nordheim@qt.io>
c23d00078c