New lock implementation for mac

webrtc/base/criticalsection_unittest.cc

Index: webrtc/base/criticalsection_unittest.cc
diff --git a/webrtc/base/criticalsection_unittest.cc b/webrtc/base/criticalsection_unittest.cc
index d6990c0023d8b02a33a97abdfce4da99ded00bc4..a0e103386280dbfd8cef4e5d8c2bd0fe1beb0f60 100644
--- a/webrtc/base/criticalsection_unittest.cc
+++ b/webrtc/base/criticalsection_unittest.cc
@@ -11,9 +11,12 @@
 #include <set>
 #include <vector>
 
+#include "webrtc/base/arraysize.h"
+#include "webrtc/base/checks.h"
 #include "webrtc/base/criticalsection.h"
 #include "webrtc/base/event.h"
 #include "webrtc/base/gunit.h"
+#include "webrtc/base/platform_thread.h"
 #include "webrtc/base/scoped_ptr.h"
 #include "webrtc/base/scopedptrcollection.h"
 #include "webrtc/base/thread.h"
@@ -320,4 +323,109 @@ TEST(CriticalSectionTest, IsLocked) {
 }
 #endif
 
+class PerfTestData {
+ public:
+  PerfTestData(int expected_count, Event* event)
+      : cache_line_barrier_1_(), cache_line_barrier_2_(),
+        expected_count_(expected_count), event_(event) {
+    cache_line_barrier_1_[0]++;  // Avoid 'is not used'.
+    cache_line_barrier_2_[0]++;  // Avoid 'is not used'.
+  }
+  ~PerfTestData() {}
+
+  void AddToCounter(int add) {
+    rtc::CritScope cs(&lock_);
+    my_counter_ += add;
+    if (my_counter_ == expected_count_)
+      event_->Set();
+  }
+
+  int64_t total() const {
+    // Assume that only one thread is running now.
+    return my_counter_;
+  }
+
+ private:
+  uint8_t cache_line_barrier_1_[64];
+  CriticalSection lock_;
+  uint8_t cache_line_barrier_2_[64];
+  int64_t my_counter_ = 0;
+  const int expected_count_;
+  Event* const event_;
+};
+
+class PerfTestThread {
+ public:
+  PerfTestThread() : thread_(&ThreadFunc, this, "CsPerf") {}
+
+  void Start(PerfTestData* data, int repeats, int id) {
+    RTC_DCHECK(!thread_.IsRunning());
+    RTC_DCHECK(!data_);
+    data_ = data;
+    repeats_ = repeats;
+    my_id_ = id;
+    thread_.Start();
+  }
+
+  void Stop() {
+    RTC_DCHECK(thread_.IsRunning());
+    RTC_DCHECK(data_);
+    thread_.Stop();
+    repeats_ = 0;
+    data_ = nullptr;
+    my_id_ = 0;
+  }
+
+ private:
+  static bool ThreadFunc(void* param) {
+    PerfTestThread* me = static_cast<PerfTestThread*>(param);
+    for (int i = 0; i < me->repeats_; ++i)
+      me->data_->AddToCounter(me->my_id_);
+    return false;
+  }
+
+  PlatformThread thread_;
+  PerfTestData* data_ = nullptr;
+  int repeats_ = 0;
+  int my_id_ = 0;
+};
+
+// Comparison of output of this test as tested on a MacBook Pro Retina, 15-inch,
+// Mid 2014, 2,8 GHz Intel Core i7, 16 GB 1600 MHz DDR3,
+// running OS X El Capitan, 10.11.2.
+//
+// Native mutex implementation:
+// Approximate CPU usage:
+//   System: ~16%
+//   User mode: ~1.3%
+//   Idle: ~82%
+// Unit test output:
+// [       OK ] CriticalSectionTest.Performance (234545 ms)
+//
+// Special partially spin lock based implementation:
+// Approximate CPU usage:
+//   System: ~75%
+//   User mode: ~16%
+//   Idle: ~8%
+// Unit test output:
+// [       OK ] CriticalSectionTest.Performance (2107 ms)
+//
+// The test is disabled by default to avoid unecessarily loading the bots.
+TEST(CriticalSectionTest, DISABLED_Performance) {
+  PerfTestThread threads[8];
+  Event event(false, false);
+
+  static const int kThreadRepeats = 10000000;
+  static const int kExpectedCount = kThreadRepeats * arraysize(threads);
+  PerfTestData test_data(kExpectedCount, &event);
+
+  for (auto& t : threads)
+    t.Start(&test_data, kThreadRepeats, 1);
+
+  event.Wait(Event::kForever);
+
+  for (auto& t : threads)
+    t.Stop();
+}
+
 }  // namespace rtc