[2.0>master] [MERGE chakra-core#2777 @obastemur] xplat: Improve GetCurrent*Thread performance

obastemur · obastemur · commit d0b5b90d307e · 2017-04-05T03:18:08.000-07:00
Merge pull request chakra-core#2777 from thread_wait_fix This PR targets node-chakracore server load testing scenarios. Perf hungry JS application triggers GetCurrent*Thread more than 10K times per second. Use front thread_local caching if possible.
diff --git a/lib/Runtime/PlatformAgnostic/Platform/Common/HiResTimer.cpp b/lib/Runtime/PlatformAgnostic/Platform/Common/HiResTimer.cpp
@@ -77,7 +77,7 @@ namespace DateTime
             // in case the system time wasn't updated backwards, and cache is still beyond...
             if (currentTime >= data.cacheSysTime && currentTime < data.cacheSysTime + INTERVAL_FOR_TICK_BACKUP)
             {
-                return data.cacheSysTime + INTERVAL_FOR_TICK_BACKUP - 1; // wait for real time
+                return data.cacheSysTime + INTERVAL_FOR_TICK_BACKUP; // wait for real time
             }
 
             data.cacheSysTime = currentTime;
diff --git a/pal/src/include/pal/thread.hpp b/pal/src/include/pal/thread.hpp
@@ -737,20 +737,9 @@ namespace CorUnix
     extern "C" CPalThread *CreateCurrentThreadData();
 #endif // FEATURE_PAL_SXS
 
-    inline CPalThread *GetCurrentPalThread()
-    {
-        return reinterpret_cast<CPalThread*>(pthread_getspecific(thObjKey));
-    }
+    CPalThread *GetCurrentPalThread(bool force = false);
 
-    inline CPalThread *InternalGetCurrentThread()
-    {
-        CPalThread *pThread = GetCurrentPalThread();
-#if defined(FEATURE_PAL_SXS)
-        if (pThread == nullptr)
-            pThread = CreateCurrentThreadData();
-#endif // FEATURE_PAL_SXS
-        return pThread;
-    }
+    CPalThread *InternalGetCurrentThread();
 
 /***
 
diff --git a/pal/src/thread/pal_thread.cpp b/pal/src/thread/pal_thread.cpp
@@ -2911,3 +2911,47 @@ bool IsAddressOnStack(ULONG_PTR address)
 
     return false;
 }
+
+#ifndef __IOS__
+// why _Thread_local ? Because it is faster(.)
+// why not replace PAL to use _Thread_local instead of front caching? _Thread_local is not cross platform
+
+// Why ULONG_PTR? keeping type for localThread `simple` may affect implementation hence the perf. positively.
+THREAD_LOCAL ULONG_PTR localThread = 0;
+CPalThread *CorUnix::GetCurrentPalThread(bool force)
+{
+    ULONG_PTR pThread = localThread;
+    if (pThread == 0)
+    {
+        pThread = (ULONG_PTR) reinterpret_cast<CPalThread*>(pthread_getspecific(thObjKey));
+#ifdef FEATURE_PAL_SXS
+        if (pThread == 0 && force)
+        {
+            pThread = (ULONG_PTR) CreateCurrentThreadData();
+        }
+#endif
+        localThread = pThread;
+    }
+
+    return (CPalThread*)pThread;
+}
+#else // !__IOS__
+CPalThread *CorUnix::GetCurrentPalThread(bool force)
+{
+    CPalThread *pThread = reinterpret_cast<CPalThread*>(pthread_getspecific(thObjKey));
+
+#ifdef FEATURE_PAL_SXS
+    if (pThread == nullptr && force)
+    {
+        pThread = CreateCurrentThreadData();
+    }
+#endif
+
+    return pThread;
+}
+#endif
+
+CPalThread *CorUnix::InternalGetCurrentThread()
+{
+    return GetCurrentPalThread(true);
+}

Original file line number	Diff line number	Diff line change
`@@ -77,7 +77,7 @@ namespace DateTime`
`77`	`77`	`// in case the system time wasn't updated backwards, and cache is still beyond...`
`78`	`78`	`if (currentTime >= data.cacheSysTime && currentTime < data.cacheSysTime + INTERVAL_FOR_TICK_BACKUP)`
`79`	`79`	`{`
`80`		`- return data.cacheSysTime + INTERVAL_FOR_TICK_BACKUP - 1; // wait for real time`
	`80`	`+ return data.cacheSysTime + INTERVAL_FOR_TICK_BACKUP; // wait for real time`
`81`	`81`	`}`
`82`	`82`
`83`	`83`	`data.cacheSysTime = currentTime;`