Common: Add Threading::SleepUntil()

common/Darwin/DarwinMisc.cpp

@@ -19,11 +19,13 @@
 #include <cstdlib>
 #include <sys/types.h>
 #include <sys/sysctl.h>
+#include <time.h>
 #include <mach/mach_time.h>
 #include <IOKit/pwr_mgt/IOPMLib.h>
 
 #include "common/Pcsx2Types.h"
 #include "common/General.h"
+#include "common/Threading.h"
 #include "common/WindowInfo.h"
 
 // Darwin (OSX) is a bit different from Linux when requesting properties of
@@ -47,13 +49,13 @@ u64 GetPhysicalMemory()
 }
 
 static u64 tickfreq;
+static mach_timebase_info_data_t s_timebase_info;
 
 void InitCPUTicks()
 {
-	mach_timebase_info_data_t info;
-	if (mach_timebase_info(&info) != KERN_SUCCESS)
+	if (mach_timebase_info(&s_timebase_info) != KERN_SUCCESS)
 		abort();
-	tickfreq = (u64)1e9 * (u64)info.denom / (u64)info.numer;
+	tickfreq = (u64)1e9 * (u64)s_timebase_info.denom / (u64)s_timebase_info.numer;
 }
 
 // returns the performance-counter frequency: ticks per second (Hz)
@@ -110,4 +112,27 @@ bool WindowInfo::InhibitScreensaver(const WindowInfo& wi, bool inhibit)
 
 	return true;
 }
+
+void Threading::Sleep(int ms)
+{
+	usleep(1000 * ms);
+}
+
+void Threading::SleepUntil(u64 ticks)
+{
+	// This is definitely sub-optimal, but apparently clock_nanosleep() doesn't exist.
+	const s64 diff = static_cast<s64>(ticks - GetCPUTicks());
+	if (diff <= 0)
+		return;
+
+	const u64 nanos = (static_cast<u64>(diff) * static_cast<u64>(s_timebase_info.denom)) / static_cast<u64>(s_timebase_info.numer);
+	if (nanos == 0)
+		return;
+
+	struct timespec ts;
+	ts.tv_sec = nanos / 1000000000ULL;
+	ts.tv_nsec = nanos % 1000000000ULL;
+	nanosleep(&ts, nullptr);
+}
+
 #endif

common/Darwin/DarwinThreads.cpp

@@ -30,11 +30,6 @@
 // the LOCK prefix.  The prefix works on single core CPUs fine (but is slow), but not
 // having the LOCK prefix is very bad indeed.
 
-__forceinline void Threading::Sleep(int ms)
-{
-	usleep(1000 * ms);
-}
-
 __forceinline void Threading::Timeslice()
 {
 	sched_yield();

common/Linux/LnxMisc.cpp

@@ -28,6 +28,7 @@
 #include "common/Pcsx2Types.h"
 #include "common/General.h"
 #include "common/StringUtil.h"
+#include "common/Threading.h"
 #include "common/WindowInfo.h"
 
 // Returns 0 on failure (not supported by the operating system).
@@ -149,4 +150,17 @@ bool Common::PlaySoundAsync(const char* path)
 #endif
 }
 
+void Threading::Sleep(int ms)
+{
+	usleep(1000 * ms);
+}
+
+void Threading::SleepUntil(u64 ticks)
+{
+	struct timespec ts;
+	ts.tv_sec = static_cast<time_t>(ticks / 1000000000ULL);
+	ts.tv_nsec = static_cast<long>(ticks % 1000000000ULL);
+	clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME, &ts, nullptr);
+}
+
 #endif

common/Linux/LnxThreads.cpp

@@ -54,11 +54,6 @@
 // the LOCK prefix.  The prefix works on single core CPUs fine (but is slow), but not
 // having the LOCK prefix is very bad indeed.
 
-__forceinline void Threading::Sleep(int ms)
-{
-	usleep(1000 * ms);
-}
-
 __forceinline void Threading::Timeslice()
 {
 	sched_yield();

common/Threading.h

@@ -55,6 +55,9 @@ namespace Threading
 	// sleeps the current thread for the given number of milliseconds.
 	extern void Sleep(int ms);
 
+	// sleeps the current thread until the specified time point, or later.
+	extern void SleepUntil(u64 ticks);
+
 	// --------------------------------------------------------------------------------------
 	//  ThreadHandle
 	// --------------------------------------------------------------------------------------

common/Windows/WinMisc.cpp

@@ -19,6 +19,7 @@
 #include "common/RedtapeWindows.h"
 #include "common/Exceptions.h"
 #include "common/StringUtil.h"
+#include "common/Threading.h"
 #include "common/General.h"
 #include "common/WindowInfo.h"
 
@@ -30,6 +31,23 @@
 
 alignas(16) static LARGE_INTEGER lfreq;
 
+// This gets leaked... oh well.
+static thread_local HANDLE s_sleep_timer;
+static thread_local bool s_sleep_timer_created = false;
+
+static HANDLE GetSleepTimer()
+{
+	if (s_sleep_timer_created)
+		return s_sleep_timer;
+
+	s_sleep_timer_created = true;
+	s_sleep_timer = CreateWaitableTimerEx(nullptr, nullptr, CREATE_WAITABLE_TIMER_HIGH_RESOLUTION, TIMER_ALL_ACCESS);
+	if (!s_sleep_timer)
+		s_sleep_timer = CreateWaitableTimer(nullptr, TRUE, nullptr);
+
+	return s_sleep_timer;
+}
+
 void InitCPUTicks()
 {
 	QueryPerformanceFrequency(&lfreq);
@@ -91,4 +109,34 @@ bool Common::PlaySoundAsync(const char* path)
 	return PlaySoundW(wpath.c_str(), NULL, SND_ASYNC | SND_NODEFAULT);
 }
 
+void Threading::Sleep(int ms)
+{
+	::Sleep(ms);
+}
+
+void Threading::SleepUntil(u64 ticks)
+{
+	// This is definitely sub-optimal, but there's no way to sleep until a QPC timestamp on Win32.
+	const s64 diff = static_cast<s64>(ticks - GetCPUTicks());
+	if (diff <= 0)
+		return;
+
+	const HANDLE hTimer = GetSleepTimer();
+	if (!hTimer)
+		return;
+
+	const u64 one_hundred_nanos_diff = (static_cast<u64>(diff) * 10000000ULL) / GetTickFrequency();
+	if (one_hundred_nanos_diff == 0)
+		return;
+
+	LARGE_INTEGER fti;
+	fti.QuadPart = -static_cast<s64>(one_hundred_nanos_diff);
+
+	if (SetWaitableTimer(hTimer, &fti, 0, nullptr, nullptr, FALSE))
+	{
+		WaitForSingleObject(hTimer, INFINITE);
+		return;
+	}
+}
+
 #endif

common/Windows/WinThreads.cpp

@@ -23,11 +23,6 @@
 #include <process.h>
 #include <timeapi.h>
 
-__fi void Threading::Sleep(int ms)
-{
-	::Sleep(ms);
-}
-
 __fi void Threading::Timeslice()
 {
 	::Sleep(0);