Use thread_local compiler specifics

Much faster then pthread_getspecific() but still a
speed regression against the original code.

Following are the nps on a bench:

Position
454165
454838
455433

tls
441046
442767
442767

ms (Win)
450521
447510
451105

ms (pthread)
422115
422115
424276

Signed-off-by: Marco Costalba <mcostalba@gmail.com>

src/evaluate.cpp

@@ -371,7 +371,7 @@ Value do_evaluate(const Position& pos, Value& margin) {
   margins[WHITE] = margins[BLACK] = VALUE_ZERO;
 
   // Probe the material hash table
-  ei.mi = Threads.this_thread()->materialTable.probe(pos);
+  ei.mi = this_thread->materialTable.probe(pos);
   score += ei.mi->material_value();
 
   // If we have a specialized evaluation function for the current material
@@ -383,7 +383,7 @@ Value do_evaluate(const Position& pos, Value& margin) {
   }
 
   // Probe the pawn hash table
-  ei.pi = Threads.this_thread()->pawnTable.probe(pos);
+  ei.pi = this_thread->pawnTable.probe(pos);
   score += ei.pi->pawns_value();
 
   // Initialize attack and king safety bitboards

src/platform.h

@@ -54,7 +54,6 @@ inline uint64_t time_to_msec(const sys_time_t& t) { return t.tv_sec * 1000LL + t
 typedef pthread_mutex_t Lock;
 typedef pthread_cond_t WaitCondition;
 typedef pthread_t NativeHandle;
-typedef pthread_key_t ThreadLocalStorageKey;
 typedef void*(*pt_start_fn)(void*);
 
 #  define lock_init(x) pthread_mutex_init(&(x), NULL)
@@ -68,10 +67,6 @@ typedef void*(*pt_start_fn)(void*);
 #  define cond_timedwait(x,y,z) pthread_cond_timedwait(&(x),&(y),z)
 #  define thread_create(x,f,t) !pthread_create(&(x),NULL,(pt_start_fn)f,t)
 #  define thread_join(x) pthread_join(x, NULL)
-#  define tls_init(k) pthread_key_create(&k,NULL)
-#  define tls_get(k) pthread_getspecific(k)
-#  define tls_set(k,x) pthread_setspecific(k,x)
-#  define tls_destroy(k) pthread_key_delete(k)
 
 #else // Windows and MinGW
 
@@ -96,7 +91,6 @@ inline uint64_t time_to_msec(const sys_time_t& t) { return t.time * 1000LL + t.m
 typedef CRITICAL_SECTION Lock;
 typedef HANDLE WaitCondition;
 typedef HANDLE NativeHandle;
-typedef DWORD ThreadLocalStorageKey;
 
 #  define lock_init(x) InitializeCriticalSection(&(x))
 #  define lock_grab(x) EnterCriticalSection(&(x))
@@ -109,10 +103,6 @@ typedef DWORD ThreadLocalStorageKey;
 #  define cond_timedwait(x,y,z) { lock_release(y); WaitForSingleObject(x,z); lock_grab(y); }
 #  define thread_create(x,f,t) (x = CreateThread(NULL,0,(LPTHREAD_START_ROUTINE)f,t,0,NULL), x != NULL)
 #  define thread_join(x) { WaitForSingleObject(x, INFINITE); CloseHandle(x); }
-#  define tls_init(k) do { k = TlsAlloc(); } while(0)
-#  define tls_get(k) TlsGetValue(k)
-#  define tls_set(k,x) TlsSetValue(k,x)
-#  define tls_destroy(k) TlsFree(k)
 
 #endif
 

src/position.cpp

@@ -895,8 +895,8 @@ void Position::do_move(Move m, StateInfo& newSt, const CheckInfo& ci, bool moveI
   }
 
   // Prefetch pawn and material hash tables
-  prefetch((char*)Threads.this_thread()->pawnTable.entries[st->pawnKey]);
+  prefetch((char*)this_thread->pawnTable.entries[st->pawnKey]);
-  prefetch((char*)Threads.this_thread()->materialTable.entries[st->materialKey]);
+  prefetch((char*)this_thread->materialTable.entries[st->materialKey]);
 
   // Update incremental scores
   st->psqScore += psq_delta(piece, from, to);

src/search.cpp

@@ -332,7 +332,7 @@ finalize:
   // but if we are pondering or in infinite search, we shouldn't print the best
   // move before we are told to do so.
   if (!Signals.stop && (Limits.ponder || Limits.infinite))
-      Threads.this_thread()->wait_for_stop_or_ponderhit();
+      this_thread->wait_for_stop_or_ponderhit();
 
   // Best move could be MOVE_NONE when searching on a stalemate position
   cout << "bestmove " << move_to_uci(RootMoves[0].pv[0], Chess960)
@@ -543,7 +543,7 @@ namespace {
     bool isPvMove, inCheck, singularExtensionNode, givesCheck;
     bool captureOrPromotion, dangerous, doFullDepthSearch;
     int moveCount = 0, playedMoveCount = 0;
-    Thread* thisThread = Threads.this_thread();
+    Thread* thisThread = this_thread;
     SplitPoint* sp = NULL;
 
     refinedValue = bestValue = value = -VALUE_INFINITE;

src/thread.cpp

@@ -28,6 +28,7 @@
 using namespace Search;
 
 ThreadsManager Threads; // Global object
+THREAD_LOCAL Thread* this_thread; // Thread local variable
 
 namespace { extern "C" {
 
@@ -36,7 +37,7 @@ namespace { extern "C" {
 
  long start_routine(Thread* th) {
 
-   Threads.set_this_thread(th); // Save pointer into thread local storage
+   this_thread = th; // Save pointer into thread local storage
    (th->*(th->start_fn))();
    return 0;
  }
@@ -205,12 +206,11 @@ bool Thread::is_available_to(Thread* master) const {
 
 void ThreadsManager::init() {
 
-  tls_init(tlsKey);
   cond_init(sleepCond);
   lock_init(splitLock);
   timer = new Thread(&Thread::timer_loop);
   threads.push_back(new Thread(&Thread::main_loop));
-  set_this_thread(main_thread()); // Use main thread's resources
+  this_thread = main_thread(); // Use main thread's resources
   read_uci_options();
 }
 
@@ -225,7 +225,6 @@ ThreadsManager::~ThreadsManager() {
   delete timer;
   lock_destroy(splitLock);
   cond_destroy(sleepCond);
-  tls_destroy(tlsKey);
 }
 
 
@@ -314,7 +313,7 @@ Value ThreadsManager::split(Position& pos, Stack* ss, Value alpha, Value beta,
   assert(beta <= VALUE_INFINITE);
   assert(depth > DEPTH_ZERO);
 
-  Thread* master = this_thread();
+  Thread* master = this_thread;
 
   if (master->splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD)
       return bestValue;

src/thread.h

@@ -120,8 +120,6 @@ public:
   int min_split_depth() const { return minimumSplitDepth; }
   int size() const { return (int)threads.size(); }
   Thread* main_thread() const { return threads[0]; }
-  Thread* this_thread() const { return (Thread*)tls_get(tlsKey); }
-  void set_this_thread(Thread* th) const { tls_set(tlsKey, th); }
 
   void wake_up() const;
   void sleep() const;
@@ -140,7 +138,6 @@ private:
 
   std::vector<Thread*> threads;
   Thread* timer;
-  ThreadLocalStorageKey tlsKey;
   Lock splitLock;
   WaitCondition sleepCond;
   Depth minimumSplitDepth;
@@ -149,5 +146,6 @@ private:
 };
 
 extern ThreadsManager Threads;
+extern THREAD_LOCAL Thread* this_thread;
 
 #endif // !defined(THREAD_H_INCLUDED)

src/types.h

@@ -64,6 +64,12 @@
 #  define FORCE_INLINE  inline
 #endif
 
+#if defined(__GNUC__)
+#  define THREAD_LOCAL  __thread
+#else
+#  define THREAD_LOCAL  __declspec(thread)
+#endif
+
 #if defined(USE_POPCNT)
 const bool HasPopCnt = true;
 #else