[libcxx] std::chrono::steady_clock is not working as expected

[antonic901]

While working on my Kodi port I faced lot of issues with:

• frame time (CTimeUtils::GetFrameTime)
• stopwatch (CStopwatch)
• events (CEvent)

I inspected this and everything leads to use of std::chrono::steady_clock::now(). This is used by std::condition_variable_any::wait_for (used in CEvent), it's used inside CStopwatch to calculate elapsed time and it's used to update frame time inside CTimeUtils. When app is started it behaves normally but after some time it starts behaving unpredictably giving negative and zero values which leads to crashes in case of std::condition_variable_any and brokes complete logic in Kodi because returned value of this functions should be always greater then previous one (somerhing like GetTickCount).

Astarivi helped me to debug this and we came to conslusion that problem is inside kernel functin KeQueryPerformanceFrequency which is invoked when std::chrono::steady_clock::now() is called. This is actually a know problem for Win 2K and older systems. In short, KeQueryPerformanceFrequency on Xbox can fail and LLVM assumes this will always succeed. See here. There are also more details on MS docs

Temporary solution for this is to replace steady_clock with system_clock but this should be fixed anyway. Or at least until this is sorted steady_clock should be disabled so it doesnt lead to confusion.

[antonic901]

This is minimal example to reproduce this:

#include <hal/video.h>
#include <hal/debug.h>

#include <assert.h>
#include <chrono>
#include <windows.h>
#include <xboxkrnl/xboxkrnl.h>

int main()
{
  XVideoSetMode(640, 480, 32, REFRESH_DEFAULT);
  auto start = std::chrono::steady_clock::now();
  while (true)
  {
    Sleep(2000);
    auto now = std::chrono::steady_clock::now();
    auto duration = std::chrono::duration_cast<std::chrono::milliseconds>(now - start);
    if (duration.count() < 0)
    {
      debugPrint("Wrong!\n");
      assert(0);
    }
    else
    {
      debugPrint("Elapsed %lld\n", duration.count());
    }
  }

  return 0;
}

[thrimbor]

KeQueryPerformanceFrequency cannot fail on the Xbox, it returns a hardcoded value.

It's not being used by libc++ though, that instead depends on QueryPerformanceFrequency which returns a fixed value computed at startup and thus also can't randomly fail.

The real issue is the integer overflow in counter.QuadPart * nano::den / freq.QuadPart. The counter increases by 733333333 per second on a stock system, nano::den is 1000000000. It just takes a few seconds for that multiplication to overflow the signed 64-bit integer.

I don't have a clever solution for this rn, the easiest thing would be to modify libc++ and introduce a division by some sufficiently large constant - that would reduce the precision, but also work around the overflow when it gets folded into the nano::den constant.

/edit:
I wonder whether this affects Windows as well, LARGE_INTEGER afaik is still only 64 bits on x86_64.

[LukeUsher]

llvm/llvm-project@02f1d28 seems to be related to this issue;

[antonic901]

Thanks guys, this commit which Luke provided fixed the issue. However problem with std::condition_variable is still present. I will see if I can make simple example to reproduce that and then I will open new issue related to that. Thanks again! When can I except fix to be on main?