[base] Implement CPU time on Windows.

src/base/platform/time.cc

Index: src/base/platform/time.cc
diff --git a/src/base/platform/time.cc b/src/base/platform/time.cc
index b2355a33bde275695d13fe49ceb619ddf12911d5..80f2eefa924c1a21da38524bb1bfd2ef036056c8 100644
--- a/src/base/platform/time.cc
+++ b/src/base/platform/time.cc
@@ -67,6 +67,23 @@ inline int64_t ClockNow(clockid_t clk_id) {
   return 0;
 #endif
 }
+#elif V8_OS_WIN
+inline bool IsBuggyAthlon(const v8::base::CPU& cpu) {

[fmeawad, 2016/05/24 22:56:59]

nit:Rename to IsQPCReliable

[lpy, 2016/05/24 23:29:16]

Done.

+  // On Athlon X2 CPUs (e.g. model 15) QueryPerformanceCounter is unreliable.
+  return strcmp(cpu.vendor(), "AuthenticAMD") == 0 && cpu.family() == 15;
+}
+
+// Returns the current value of the performance counter.
+inline uint64_t QPCNowRaw() {
+  LARGE_INTEGER perf_counter_now = {};
+  // According to the MSDN documentation for QueryPerformanceCounter(), this
+  // will never fail on systems that run XP or later.
+  // https://msdn.microsoft.com/library/windows/desktop/ms644904.aspx
+  BOOL result = ::QueryPerformanceCounter(&perf_counter_now);
+  DCHECK(result);
+  USE(result);
+  return perf_counter_now.QuadPart;
+}
 #endif  // V8_OS_MACOSX
 
 
@@ -456,15 +473,12 @@ class HighResolutionTickClock final : public TickClock {
   virtual ~HighResolutionTickClock() {}
 
   int64_t Now() override {
-    LARGE_INTEGER now;
-    BOOL result = QueryPerformanceCounter(&now);
-    DCHECK(result);
-    USE(result);
+    uint64_t now = QPCNowRaw();
 
     // Intentionally calculate microseconds in a round about manner to avoid
     // overflow and precision issues. Think twice before simplifying!
-    int64_t whole_seconds = now.QuadPart / ticks_per_second_;
-    int64_t leftover_ticks = now.QuadPart % ticks_per_second_;
+    int64_t whole_seconds = now / ticks_per_second_;
+    int64_t leftover_ticks = now % ticks_per_second_;
     int64_t ticks = (whole_seconds * Time::kMicrosecondsPerSecond) +
         ((leftover_ticks * Time::kMicrosecondsPerSecond) / ticks_per_second_);
 
@@ -529,10 +543,10 @@ struct CreateHighResTickClockTrait {
       return tick_clock.Pointer();
     }
 
-    // On Athlon X2 CPUs (e.g. model 15) the QueryPerformanceCounter
-    // is unreliable, fallback to the low-resolution tick clock.
+    // Fallback to the low-resolution tick clock when on buggy Cpu
+    // where QueryPerformanceCounter is unreliable.

[fmeawad, 2016/05/24 22:56:59]

Change comment to: If QPC not reliable, fallback to low-resolution tick clock.

[lpy, 2016/05/24 23:29:16]

Done.

     CPU cpu;

[fmeawad, 2016/05/24 22:56:59]

CPU cpu should move inside IsBuggyAthlon since it is not used anywhere else?

[lpy, 2016/05/24 23:29:16]

Done.

-    if (strcmp(cpu.vendor(), "AuthenticAMD") == 0 && cpu.family() == 15) {
+    if (IsBuggyAthlon(cpu)) {
       return tick_clock.Pointer();
     }
 
@@ -621,11 +635,106 @@ ThreadTicks ThreadTicks::Now() {
 #elif(defined(_POSIX_THREAD_CPUTIME) && (_POSIX_THREAD_CPUTIME >= 0)) || \
   defined(V8_OS_ANDROID)
   return ThreadTicks(ClockNow(CLOCK_THREAD_CPUTIME_ID));
+#elif V8_OS_WIN
+  return ThreadTicks::GetForThread(::GetCurrentThread());
 #else
   UNREACHABLE();
   return ThreadTicks();
 #endif
 }
 
+
+#if V8_OS_WIN
+ThreadTicks ThreadTicks::GetForThread(const HANDLE& thread_handle) {
+  DCHECK(IsSupported());
+
+  // Get the number of TSC ticks used by the current thread.
+  ULONG64 thread_cycle_time = 0;
+  ::QueryThreadCycleTime(thread_handle, &thread_cycle_time);
+
+  // Get the frequency of the TSC.
+  double tsc_ticks_per_second = TSCTicksPerSecond();
+  if (tsc_ticks_per_second == 0)
+    return ThreadTicks();
+
+  // Return the CPU time of the current thread.
+  double thread_time_seconds = thread_cycle_time / tsc_ticks_per_second;
+  return ThreadTicks(
+      static_cast<int64_t>(thread_time_seconds * Time::kMicrosecondsPerSecond));
+}
+
+// static
+bool ThreadTicks::IsSupportedWin() {
+  static bool is_supported = base::CPU().has_non_stop_time_stamp_counter() &&
+                             !IsBuggyAthlon(base::CPU());
+  return is_supported;
+}
+
+// static
+void ThreadTicks::WaitUntilInitializedWin() {
+  while (TSCTicksPerSecond() == 0)
+    ::Sleep(10);
+}
+
+double ThreadTicks::TSCTicksPerSecond() {
+  DCHECK(IsSupported());
+
+  // The value returned by QueryPerformanceFrequency() cannot be used as the TSC
+  // frequency, because there is no guarantee that the TSC frequency is equal to
+  // the performance counter frequency.
+
+  // The TSC frequency is cached in a static variable because it takes some time
+  // to compute it.
+  static double tsc_ticks_per_second = 0;
+  if (tsc_ticks_per_second != 0)
+    return tsc_ticks_per_second;
+
+  // Increase the thread priority to reduces the chances of having a context
+  // switch during a reading of the TSC and the performance counter.
+  int previous_priority = ::GetThreadPriority(::GetCurrentThread());
+  ::SetThreadPriority(::GetCurrentThread(), THREAD_PRIORITY_HIGHEST);
+
+  // The first time that this function is called, make an initial reading of the
+  // TSC and the performance counter.
+  static const uint64_t tsc_initial = __rdtsc();
+  static const uint64_t perf_counter_initial = QPCNowRaw();
+
+  // Make a another reading of the TSC and the performance counter every time
+  // that this function is called.
+  uint64_t tsc_now = __rdtsc();
+  uint64_t perf_counter_now = QPCNowRaw();
+
+  // Reset the thread priority.
+  ::SetThreadPriority(::GetCurrentThread(), previous_priority);
+
+  // Make sure that at least 50 ms elapsed between the 2 readings. The first
+  // time that this function is called, we don't expect this to be the case.
+  // Note: The longer the elapsed time between the 2 readings is, the more
+  //   accurate the computed TSC frequency will be. The 50 ms value was
+  //   chosen because local benchmarks show that it allows us to get a
+  //   stddev of less than 1 tick/us between multiple runs.
+  // Note: According to the MSDN documentation for QueryPerformanceFrequency(),
+  //   this will never fail on systems that run XP or later.
+  //   https://msdn.microsoft.com/library/windows/desktop/ms644905.aspx
+  LARGE_INTEGER perf_counter_frequency = {};
+  ::QueryPerformanceFrequency(&perf_counter_frequency);
+  DCHECK_GE(perf_counter_now, perf_counter_initial);
+  uint64_t perf_counter_ticks = perf_counter_now - perf_counter_initial;
+  double elapsed_time_seconds =
+      perf_counter_ticks / static_cast<double>(perf_counter_frequency.QuadPart);
+
+  const double kMinimumEvaluationPeriodSeconds = 0.05;
+  if (elapsed_time_seconds < kMinimumEvaluationPeriodSeconds)
+    return 0;
+
+  // Compute the frequency of the TSC.
+  DCHECK_GE(tsc_now, tsc_initial);
+  uint64_t tsc_ticks = tsc_now - tsc_initial;
+  tsc_ticks_per_second = tsc_ticks / elapsed_time_seconds;
+
+  return tsc_ticks_per_second;
+}
+#endif  // V8_OS_WIN
+
 }  // namespace base
 }  // namespace v8