What is the C++ time library? Got a link?

What is the C++ time library? Got a link?

I'm not sure whether it's public. So, no.

time.Duration does do saturated arithmetic. For time.Time that really seems like overkill. There's no reasonable way you could end up with times that large anyway.

There's no reasonable way you could end up with times that large anyway.

The time.Time data structure itself is surely capable of representing usefully-distant values: int64 seconds gives a range of ~292 billion years in both directions, which is longer than virtually all estimates for the current age of the universe (albeit shorter than some estimates of the remaining duration of the Stelliferous Era). So having limits on the range supported by the implementation seems fine.

The trouble is, at the moment those limits are ad-hoc: the point at which methods stop working is a complicated function of both the method (AddDate, Add, String, and Unix all break down at different points) and CPU architecture (specifically, the size of int), and many of the individual methods break down long before the limits inherent to the representation.

Some points of interest are shown here: https://play.golang.org/p/-WgwEfN1Vs

In practice, it appears that the portable range is limited by the functions involving dates, which break down outside the range [-2147483648-01-01 00:00:00, 2147483647-12-31 23:59:59.999999999] on 32-bit platforms. That gives a minimum time.Time value that lies well short of most scientific estimates for the age of the universe, one measure of "reasonable" times. (For example, you could imagine attempting to use time.Time to sort a list of events interesting to humans, starting with the Big Bang.)

For the use-case I have in mind, I explicitly need two distinct unreasonable times: one far enough in the past to represent "infinite past", and one far enough in the future to represent "infinite future". It doesn't really matter what the specific values are, but they need to be well-behaved w.r.t. all "reasonable" values: for example, infiniteFuture.AddDate(<any reasonable offset>).After(<any reasonable date>) and infiniteFuture.Add(<any reasonable duration>).After(<any reasonable date>) must always be true.

It would be nice to also have infiniteFuture.Add(<any reasonable duration>).Equal(infiniteFuture) , but with enough space between the "reasonable" and "infinite" values I can probably work around that. I could consider values of extreme magnitude at either end of the range to be "infinite", but the overflow issues in the date functions in particular make me worry about "infinite" values overflowing into the "reasonable" range.

Fixed the link: Stelliferous Era

If you want to make a proposal for the proper min/max and do the implementation, I guess it is OK with me. It does match Duration.

What is the C++ time library? Got a link?

I'm interested in interoperating with two libraries in particular:

1. absl::Time, which supports “precision of at least one nanosecond, and range +/-100 billion years”. It implements saturating integer arithmetic.

2. std::chrono, which represents counts of ticks parameterized on both the representation and the tick period. It implements saturating arithmetic if the underlying representation saturates (e.g., is a floating-point type).

Ah, fun: there may have been an undetected overflow in Sub and/or Unix too:
https://play.golang.org/p/Tdo-5VgCJto reported -1s in Go 1.12, even though the dates are supposed to be 64 bits apart!

If the overflow was in Sub, it may have been fixed in CL 131196 (#17858), although I haven't verified either way.

(CC @pongad @martisch @rasky )

Looks like there is definitely an overflow in Unix, and it is not fixed in go1.13beta1.

~/tmp$ go1.13beta1 run main.go
true
292277026304-08-26 10:42:50.145224192 -0500 EST - 292277026304-08-26 10:42:51.145224192 -0500 EST = -1s

If we do anything more, I think we should clamp the year to |y|<100e9, similar to absl (same as absl? unclear).
There is definitely an overflow in Unix if you pass in a large positive value,
since Unix adds ~62e9 seconds to get the internal (Jan 1 year 1-based) representation.