Instant::now can go backward

[goffrie]

As in #51648 (comment).

I've seen "specified instant was later than self" panics in the wild, in particular on Windows (32-bit x86). Googling reveals stuff like https://stackoverflow.com/questions/44020619/queryperformancecounter-on-multi-core-processor-under-windows-10-behaves-erratic which suggests that QueryPerformanceCounter can't really be relied on... frustratingly.

Perhaps libstd should work around this and enforce that Instant::now() never go backward (with an atomic global, or something?)

[goffrie]

Ok, I'm not sure that SO link was the best reference, but e.g. https://chromium.googlesource.com/chromium/src/base/+/master/time/time_win.cc#14 also claims that QPC is buggy.

[the8472]

The linked bug appears to be on a virtual machine, Some hypervisors may emulate TSC instructions with coarse granularity and possibly not core-synchronized. Operating systems try to detect the accuracy of the TSCs based on CPU flags (invariant, nonstop tsc). If the hypervisor is not setting those correctly that's basically out-of-spec behavior.

[froydnj]

We've seen things like this in Firefox and have deployed large hammers to fix. Further investigation suggests that the problems may be limited to certain AMD CPUs.

[alexcrichton]

I've opened #56988 to apply such a hammer to libstd as well