Why is SysTime.init invalid?

[Jonathan M Davis]

On Tuesday, April 01, 2014 03:54:07 Jonathan M Davis wrote:
> On Tuesday, April 01, 2014 05:35:28 ed wrote:
> > OK, lazy me just read the std.satetime article again. It appears
> > the design is for no invalid values and it is currently a known
> > limitation due to CTFE.
> > 
> > ---
> > d_time_nan	There is no equivalent. SysTime.init, which has a null
> > TimeZone object, would be the closest, but once CTFE advances to
> > the point that you can new up class objects with it,
> > SysTime.init's timezone will be LocalTime, so don't rely on
> > SysTime.init being invalid. std.datetime in general tries to
> > avoid having any invalid states for any of its types. It's
> > intended that creating such values be impossible.
> > ---
> > 
> > I would still like to know if there is a way around this so I can
> > have my struct default init.
> 
> You can certainly have a struct with a SysTime member, but there's no way
> for it to use that SysTime without first assigning a valid value to it - at
> least, not without segfaulting when it tries to use the SysTime in any
> context which it would use its timezone member.
> 
> If you really want to have the SysTime be useable in the init value of your
> struct, then you can make it so that each of its member functions which use
> the SysTime member check it for validity first (or have a separate bool
> which indicates whether the SysTime has been assigned to or not), in which
> case, it could assign something like SysTime(0) to it, which would be the
> valid equivalent of SysTime.init.
> 
> Unfortunately, that does add extra overhead, but there isn't any way around
> it if you want to have the SysTime be used without having the user of the
> struct assign to it first. It's pretty much the same boat that SysTime
> itself is in. I could make it so that it checks the timezone for null every
> time it's used and assign LocalTime to it if it's null, but that adds
> constant overhead. My decision was that it was better to just live with the
> fact that SysTime.init is invalid. It's debatable though, as it's a
> trade-off.

Actually, after thinking about this more after someone opened a bug on the 
fact that SysTime.init.toString() segfaults (since the timezone is null)

https://d.puremagic.com/issues/show_bug.cgi?id=12507

I think that I have a solution. It's now possible to default-intialize the 
timezonet to UTC so that SysTime.init is equivalent to SysTime(0, UTC()). It's 
not ideal, because it's not SysTime(0, LocalTime()), and SysTime normally 
defaults to LocalTime, but it will never be possible to default-initialize it 
to LocalTime thanks to the fact that it has to call tzset when it's 
initialized.

You can look my comments in the bug for more details. I've opened a pull 
request with those changes, so the next release will probably have a valid 
SysTime.init, even if it's not SysTime(0, LocalTime()) like would be ideal.

- Jonathan M Davis