Finally fix prefetch on Linux

It was due to a missing -msse compiler option !

Without this option the CPU silently discards
prefetcht2 instructions during execution.

Also added a (gcc documented) hack to prevent Intel
compiler to optimize away the prefetches.

Special thanks to Heinz for testing and suggesting
improvments. And for Jim for testing icc on Windows.

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

diff --git a/src/Makefile b/src/Makefile
index 7ca0495e3f4c76495503b96db16020e45b669179..ff3405f7883c08e2cc379ca76d6bf6c312629bd4 100644 (file)
--- a/src/Makefile
+++ b/src/Makefile
@@ -26,8 +26,8 @@ EXE = stockfish
 ### Compiler speed switches for both GCC and ICC. These settings are generally
 ### fast on a broad range of systems, but may be changed experimentally
 ### ==========================================================================
 ### Compiler speed switches for both GCC and ICC. These settings are generally
 ### fast on a broad range of systems, but may be changed experimentally
 ### ==========================================================================

-GCCFLAGS = -O3
-ICCFLAGS = -fast
+GCCFLAGS = -O3 -msse
+ICCFLAGS = -fast -msse

 
 
 ### ==========================================================================
 
 
 ### ==========================================================================


@@ -169,6 +169,6 @@ $(EXE): $(OBJS)
 
 ### Dependencies. Do not change
 .depend:
 
 ### Dependencies. Do not change
 .depend:

-       $(CXX) -MM $(OBJS:.o=.cpp) > $@
+       $(CXX) -msse -MM $(OBJS:.o=.cpp) > $@

 
 include .depend
 
 include .depend

diff --git a/src/tt.cpp b/src/tt.cpp
index 5ea6a808acf86b2ad256f8715bf5c59772e015b1..e140a0ba10ec5fa2183b506b68750a4dcbc39289 100644 (file)
--- a/src/tt.cpp
+++ b/src/tt.cpp
@@ -25,14 +25,11 @@
 #include <cassert>
 #include <cmath>
 #include <cstring>
 #include <cassert>
 #include <cmath>
 #include <cstring>

+#include <xmmintrin.h>

 
 #include "movegen.h"
 #include "tt.h"
 
 
 #include "movegen.h"
 #include "tt.h"
 

-#if defined(_MSC_VER)
-#include <xmmintrin.h>
-#endif
-

 // The main transposition table
 TranspositionTable TT;
 
 // The main transposition table
 TranspositionTable TT;
 


@@ -175,16 +172,15 @@ TTEntry* TranspositionTable::retrieve(const Key posKey) const {
 
 void TranspositionTable::prefetch(const Key posKey) const {
 
 
 void TranspositionTable::prefetch(const Key posKey) const {
 

-#if defined(_MSC_VER)
-   char* addr = (char*)first_entry(posKey);
-  _mm_prefetch(addr, _MM_HINT_T0);
-  _mm_prefetch(addr+64, _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));
+#if defined(__INTEL_COMPILER) || defined(__ICL)
+   // This hack prevents prefetches to be optimized away by the
+   // Intel compiler. Both MSVC and gcc seems not affected.
+   __asm__ ("");

 #endif
 #endif

+
+   char const* addr = (char*)first_entry(posKey);
+  _mm_prefetch(addr, _MM_HINT_T2);
+  _mm_prefetch(addr+64, _MM_HINT_T2); // 64 bytes ahead

 }
 
 
 }