Avoid Intel compiler optimizes away prefetching

[stockfish] / src / tt.cpp

diff --git a/src/tt.cpp b/src/tt.cpp
index e9b6c1758c7eb5c7ead2bae915fccd2e7bc243ab..0396b287cc5734b9517569e1124f91ddf0fcde73 100644 (file)
--- a/src/tt.cpp
+++ b/src/tt.cpp
@@ -29,10 +29,15 @@
 #include "movegen.h"
 #include "tt.h"
 
+#if defined(_MSC_VER)
+#include <xmmintrin.h>
+#endif
 
-/// This is the number of TTEntry slots for each position
-static const int ClusterSize = 5;
+// This is the number of TTEntry slots for each position
+static const int ClusterSize = 4;
 
+// The main transposition table
+TranspositionTable TT;
 
 ////
 //// Functions
@@ -92,6 +97,16 @@ void TranspositionTable::clear() {
 }
 
 
+/// TranspositionTable::first_entry returns a pointer to the first
+/// entry of a cluster given a position. The low 32 bits of the key
+/// are used to get the index in the table.
+
+inline TTEntry* TranspositionTable::first_entry(const Key posKey) const {
+
+  return entries + ((uint32_t(posKey) & (size - 1)) * ClusterSize);
+}
+
+
 /// TranspositionTable::store writes a new entry containing a position,
 /// a value, a value type, a search depth, and a best move to the
 /// transposition table. Transposition table is organized in clusters of
@@ -144,7 +159,7 @@ void TranspositionTable::store(const Key posKey, Value v, ValueType t, Depth d,
 TTEntry* TranspositionTable::retrieve(const Key posKey) const {
 
   uint32_t posKey32 = posKey >> 32;
-  TTEntry *tte = first_entry(posKey);
+  TTEntry* tte = first_entry(posKey);
 
   for (int i = 0; i < ClusterSize; i++, tte++)
       if (tte->key() == posKey32)
@@ -154,15 +169,26 @@ TTEntry* TranspositionTable::retrieve(const Key posKey) const {
 }
 
 
-/// TranspositionTable::first_entry returns a pointer to the first
-/// entry of a cluster given a position. The low 32 bits of the key
-/// are used to get the index in the table.
+/// TranspositionTable::prefetch looks up the current position in the
+/// transposition table and load it in L1/L2 cache. This is a non
+/// blocking function and do not stalls the CPU waiting for data
+/// to be loaded from RAM, that can be very slow. When we will
+/// subsequently call retrieve() the TT data will be already
+/// quickly accessible in L1/l2 CPU cache.
 
-inline TTEntry* TranspositionTable::first_entry(const Key posKey) const {
+void TranspositionTable::prefetch(const Key posKey) const {
 
-  return entries + ((uint32_t(posKey) & (size - 1)) * ClusterSize);
+#if defined(_MSC_VER)
+  _mm_prefetch((char*)first_entry(posKey), _MM_HINT_T0);
+#else
+  // We need to force an asm volatile here because gcc builtin
+  // is optimized away by Intel compiler.
+  char* addr = (char*)first_entry(posKey);
+  asm volatile("prefetcht0 %0" :: "m" (addr));
+#endif
 }
 
+
 /// TranspositionTable::new_search() is called at the beginning of every new
 /// search. It increments the "generation" variable, which is used to
 /// distinguish transposition table entries from previous searches from