From 6b92204a7ce5faf8dab2301c59aa69a26f6b8a19 Mon Sep 17 00:00:00 2001
From: Grazvydas Ignotas <notasas@gmail.com>
Date: Sun, 16 Aug 2015 18:10:34 +0200
Subject: [PATCH] add fallback blas_lock implementation

to be used on armv5 and new platforms
---
 common.h        | 14 ++++++++++++++
 common_alpha.h  |  1 +
 common_arm.h    |  4 ++++
 common_arm64.h  |  1 +
 common_ia64.h   |  2 ++
 common_mips64.h |  1 +
 common_power.h  |  1 +
 common_sparc.h  |  1 +
 common_x86.h    |  1 +
 common_x86_64.h |  1 +
 10 files changed, 27 insertions(+)

diff --git a/common.h b/common.h
index 5998b56..6073f03 100644
--- a/common.h
+++ b/common.h
@@ -435,9 +435,23 @@ static inline unsigned long long rpcc(void){
 #define RPCC64BIT
 #endif // !RPCC_DEFINED
 
+#if !defined(BLAS_LOCK_DEFINED) && defined(__GNUC__)
+static void __inline blas_lock(volatile BLASULONG *address){
+
+  do {
+    while (*address) {YIELDING;};
+
+  } while (!__sync_bool_compare_and_swap(address, 0, 1));
+}
+#define BLAS_LOCK_DEFINED
+#endif
+
 #ifndef RPCC_DEFINED
 #error "rpcc() implementation is missing for your platform"
 #endif
+#ifndef BLAS_LOCK_DEFINED
+#error "blas_lock() implementation is missing for your platform"
+#endif
 #endif // !ASSEMBLER
 
 #ifdef OS_LINUX
diff --git a/common_alpha.h b/common_alpha.h
index 86f5896..9739c94 100644
--- a/common_alpha.h
+++ b/common_alpha.h
@@ -76,6 +76,7 @@ static void __inline blas_lock(unsigned long *address){
     "30:", address);
 #endif
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline unsigned int rpcc(void){
 
diff --git a/common_arm.h b/common_arm.h
index 74b6378..84691d7 100644
--- a/common_arm.h
+++ b/common_arm.h
@@ -51,6 +51,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifndef ASSEMBLER
 
+#if defined(ARMV6) || defined(ARMV7) || defined(ARMV8)
+
 static void __inline blas_lock(volatile BLASULONG *address){
 
   int register ret;
@@ -71,6 +73,8 @@ static void __inline blas_lock(volatile BLASULONG *address){
   MB;
 }
 
+#define BLAS_LOCK_DEFINED
+#endif
 
 static inline int blas_quickdivide(blasint x, blasint y){
   return x / y;
diff --git a/common_arm64.h b/common_arm64.h
index cc08fa7..c4e588d 100644
--- a/common_arm64.h
+++ b/common_arm64.h
@@ -69,6 +69,7 @@ static void __inline blas_lock(volatile BLASULONG *address){
   } while (ret);
 
 }
+#define BLAS_LOCK_DEFINED
 
 
 static inline int blas_quickdivide(blasint x, blasint y){
diff --git a/common_ia64.h b/common_ia64.h
index d1f2107..72b75fc 100644
--- a/common_ia64.h
+++ b/common_ia64.h
@@ -68,6 +68,7 @@ static __inline void blas_lock(volatile unsigned long *address){
 			  : "ar.ccv", "memory");
   } while (ret);
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline unsigned long rpcc(void) {
   unsigned long clocks;
@@ -100,6 +101,7 @@ static __inline void blas_lock(volatile unsigned long *address){
   while (*address || _InterlockedCompareExchange((volatile int *) address,1,0))
     ;
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline unsigned int rpcc(void) {
   return __getReg(_IA64_REG_AR_ITC);
diff --git a/common_mips64.h b/common_mips64.h
index bc1a52f..f5c0ec7 100644
--- a/common_mips64.h
+++ b/common_mips64.h
@@ -98,6 +98,7 @@ static void INLINE blas_lock(volatile unsigned long *address){
 
   } while (ret);
 }
+#define BLAS_LOCK_DEFINED
 
 static inline unsigned int rpcc(void){
   unsigned long ret;
diff --git a/common_power.h b/common_power.h
index 3b9471a..ab331b0 100644
--- a/common_power.h
+++ b/common_power.h
@@ -87,6 +87,7 @@ static void INLINE blas_lock(volatile unsigned long *address){
 #endif
   } while (ret);
 }
+#define BLAS_LOCK_DEFINED
 
 static inline unsigned long rpcc(void){
   unsigned long ret;
diff --git a/common_sparc.h b/common_sparc.h
index 8a16e3d..f99972d 100644
--- a/common_sparc.h
+++ b/common_sparc.h
@@ -58,6 +58,7 @@ static void __inline blas_lock(volatile unsigned long *address){
 			 : "memory");
   } while (ret);
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline unsigned long rpcc(void){
   unsigned long clocks;
diff --git a/common_x86.h b/common_x86.h
index 9506716..6c90432 100644
--- a/common_x86.h
+++ b/common_x86.h
@@ -65,6 +65,7 @@ static void __inline blas_lock(volatile BLASULONG *address){
   } while (ret);
 
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline unsigned long long rpcc(void){
   unsigned int a, d;
diff --git a/common_x86_64.h b/common_x86_64.h
index 3a02bee..4c783b3 100644
--- a/common_x86_64.h
+++ b/common_x86_64.h
@@ -74,6 +74,7 @@ static void __inline blas_lock(volatile BLASULONG *address){
 
   } while (ret);
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline BLASULONG rpcc(void){
   BLASULONG a, d;
-- 
2.7.4