Merge pull request #2553 from martin-frbg/issue2444

Add a read memory barrier to the traversal of the buffer slot list

common_alpha.h

@@ -43,6 +43,7 @@
 
 #define MB  asm("mb")
 #define WMB asm("wmb")
+#define RMB asm("rmb")
 
 static void __inline blas_lock(unsigned long *address){
 #ifndef __DECC

common_arm.h

@@ -37,11 +37,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define MB
 #define WMB
+#define RMB
 
 #else
 
 #define MB   __asm__ __volatile__ ("dmb  ish" : : : "memory")
 #define WMB  __asm__ __volatile__ ("dmb  ishst" : : : "memory")
+#define RMB  __asm__ __volatile__ ("dmb  ish" : : : "memory")
 
 #endif
 

common_arm64.h

@@ -35,7 +35,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define MB   __asm__ __volatile__ ("dmb  ish" : : : "memory")
 #define WMB  __asm__ __volatile__ ("dmb  ishst" : : : "memory")
-
+#define RMB  __asm__ __volatile__ ("dmb  ishld" : : : "memory")
 
 #define INLINE inline
 

common_ia64.h

@@ -47,6 +47,7 @@
 
 #define MB
 #define WMB
+#define RMB
 
 #ifdef __ECC
 #include <ia64intrin.h>

common_mips.h

@@ -35,6 +35,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define MB  __sync_synchronize()
 #define WMB __sync_synchronize()
+#define RMB __sync_synchronize()
 
 #define INLINE inline
 

common_mips64.h

@@ -73,6 +73,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define MB  __sync_synchronize()
 #define WMB __sync_synchronize()
+#define RMB __sync_synchronize()
 
 #define INLINE inline
 

common_power.h

@@ -71,9 +71,11 @@
 #if defined(POWER8) || defined(POWER9)
 #define MB		__asm__ __volatile__ ("eieio":::"memory")
 #define WMB		__asm__ __volatile__ ("eieio":::"memory")
+#define RMB		__asm__ __volatile__ ("eieio":::"memory")
 #else
 #define MB		__asm__ __volatile__ ("sync")
 #define WMB		__asm__ __volatile__ ("sync")
+#define RMB		__asm__ __volatile__ ("sync")
 #endif
 
 #define INLINE inline

common_sparc.h

@@ -41,6 +41,7 @@
 
 #define MB	__asm__ __volatile__ ("nop")
 #define WMB	__asm__ __volatile__ ("nop")
+#define RMB	__asm__ __volatile__ ("nop")
 
 #ifndef ASSEMBLER
 

common_x86.h

@@ -47,6 +47,7 @@
 
 #define MB
 #define WMB
+#define RMB
 
 #ifdef C_SUN
 #define	__asm__ __asm

common_x86_64.h

@@ -63,13 +63,16 @@
 #ifdef __GNUC__
 #define MB do { __asm__ __volatile__("": : :"memory"); } while (0)
 #define WMB do { __asm__ __volatile__("": : :"memory"); } while (0)
+#define RMB
 #else
 #define MB do {} while (0)
 #define WMB do {} while (0)
+#define RMB
 #endif
 
 static void __inline blas_lock(volatile BLASULONG *address){
 
+	
 #ifndef C_MSVC
   int ret;
 #else

common_zarch.h

@@ -34,9 +34,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define COMMON_ZARCH
 
 #define MB   
-//__asm__ __volatile__ ("dmb  ish" : : : "memory")
 #define WMB  
-//__asm__ __volatile__ ("dmb  ishst" : : : "memory")
+#define RMB
 
 
 #define INLINE inline

driver/others/memory.c

@@ -2741,6 +2741,7 @@ void *blas_memory_alloc(int procpos){
   LOCK_COMMAND(&alloc_lock);
 #endif
   do {
+	  RMB;
 #if defined(USE_OPENMP)	  
     if (!memory[position].used) { 
       blas_lock(&memory[position].lock);