time: Time.Sub regression

[dsnet]

Consider the following:

func main() {
	t := time.Date(2311, 11, 26, 02, 16, 47, 63535996, time.UTC)
	u := time.Date(2019, 8, 16, 02, 29, 30, 268436582, time.UTC)

	fmt.Println(int(time.Second) + t.Nanosecond() - u.Nanosecond())
	fmt.Println(t.Sub(u))
}

On go1.12, this prints:

795099414
2562047h47m16.795099414s

On go1.13beta1, this prints:

795099414
2562047h47m16.854775807s

Notice how the nanoseconds are off? On 1.12, it is correctly .795099414, while on 1.13, it is incorrectly .854775807. This regression causes the calculation of timeouts for network connections to go wonky in RPC implementations.

The culprit for this regression is https://golang.org/cl/131196

\cc @pongad @ianlancetaylor @cybrcodr @paranoiacblack

[gopherbot]

Change https://golang.org/cl/190497 mentions this issue: Revert "time: optimize Sub"

[gopherbot]

Change https://golang.org/cl/190524 mentions this issue: time: add TestSub test case to avoid future regressions

[gopherbot]

Change https://golang.org/cl/190524 mentions this issue: time: update TestSub to avoid future regressions

[pongad]

Wanted to give a little update on this. (Apologies if this should be written elsewhere.)

There is a fix for this that passes all tests by checking t.Before(u) against the sign of the computed duration. At least on my laptop, the performance benefit of this is rather small at ~20% in an already-fast function. I'm not sure if this is actually worth doing. Unless someone can think of a better algorithm for this?