Thanks! We have a helper function for verifying the Windows version:

can you please restart "LGTM analysis: C/C++ — Analysis failed (could not build the merge commit)" step.

When you force push it should restart

Yes, I had to come up with some random change to force push something new :)

Still investigating, see https://stackoverflow.com/questions/44020619/queryperformancecounter-on-multi-core-processor-under-windows-10-behaves-erratic

How does QueryPerformanceCounter() behave on e.g. a laptop in low-power mode? I'd expect the CPU-frequency to drop and Curl_now() to behave strange.

How about timespec_get (&tspec_val, TIME_UTC) as a base? This function has been in the WindowsKit for some time. I assume it is much more monotonic.

AFAIK QueryPerformanceCounter is guaranteed to provide consistent results for quite some time. It does not use true CPU frequency. It used to have some of these issues long time ago, so in this change I set it to use QPC only starting from Vista (even though QPC is available starting from win95 probably)

Still investigating, see https://stackoverflow.com/questions/44020619/queryperformancecounter-on-multi-core-processor-under-windows-10-behaves-erratic

Strange, they mention that TSC is used for QPC timer, however, I haven't seen that. AFAIK QPC always uses something else, as I've never seen that QPF returns some huge values closely related to CPU frequency; instead I've always seen values around 1000 times smaller than CPU frequency.

Also, keep in mind originally I simply matched implementation of std::chrono::steady_clock, that would be ridiculous if std::steady_clock was more like erratic_clock. Airplanes could fall, bank machines could throw out money with such bugs :)
I was aware about those QPC issues, but always thought it was resolved now and when I wanted to make this change I actually stepped through steady_clock and it simply uses QPC timer with no attempts to "align" or "check" returned clock values.

How does QueryPerformanceCounter() behave on e.g. a laptop in low-power mode? I'd expect the CPU-frequency to drop and Curl_now() to behave strange.

How about timespec_get (&tspec_val, TIME_UTC) as a base? This function has been in the WindowsKit for some time. I assume it is much more monotonic.

No idea. You'd need to step into these functions to see what they use internally (what API). There aren't many choices. GetSystemTimeAsFileTime is very likely candidate here, but it returns calendar time (e.g. user can change system time and your clock will jump, same applies for auto time adjustments done by the system).

@MarcelRaad @gvanem @jay I'll appreciate your input and guidance on where to go with this PR. Users will certainly appreciate higher resolution timers.

@gvanem

How does QueryPerformanceCounter() behave on e.g. a laptop in low-power mode? I'd expect the CPU-frequency to drop and Curl_now() to behave strange.

The result of QueryPerformanceFrequency remains constant until the next system boot.

@MarcelRaad Okay, then this look all good to me.

AFAICT from Acquiring high-resolution time stamps Windows Vista+ will automatically select a source for the QPC that is not subject to synchronization issues across cores. Specifically it says that "the performance counter results are consistent across all processors in multi-core and multi-processor systems, even when measured on different threads or processes" with the two noted exceptions, neither of which affect us here. If there is a bug in the user's TSC or some other source Windows Vista+ should detect and avoid it, and if it's not then QPC is not working as documented. Though that stackoverflow report sounds credible I don't think it's significant enough to block this.

Thanks