From c2323dd4d2a65420f77c73f7b55c41ba469a47f8 Mon Sep 17 00:00:00 2001
From: Grazvydas Ignotas <notasas@gmail.com>
Date: Sun, 16 Aug 2015 15:18:42 +0200
Subject: [PATCH 1/7] really fix ARM locking

- was writing 0 to lock variable, so was ineffective
- only exit loop if both lock was 0 and strex was successful
---
 common_arm.h | 20 +++++++-------------
 1 file changed, 7 insertions(+), 13 deletions(-)

diff --git a/common_arm.h b/common_arm.h
index 135191057..2dabd4d7f 100644
--- a/common_arm.h
+++ b/common_arm.h
@@ -59,22 +59,16 @@ static void __inline blas_lock(volatile BLASULONG *address){
     while (*address) {YIELDING;};
 
     __asm__ __volatile__(
-			 "1:								\n\t"
-                         "ldrex r2, [%1]                                                \n\t"
-                         "mov   r2, #0                                                  \n\t"
-                         "strex r3, r2, [%1]                                            \n\t"
-			 "cmp	r3, #0							\n\t"
-			 "bne	1b							\n\t"
-			 "mov	%0 , r3							\n\t"
-                         : "=r"(ret), "=r"(address)
-                         : "1"(address)
-                         : "memory", "r2" , "r3"
-
-
+                         "ldrex r2, [%1]      \n\t"
+                         "strex %0, %2, [%1]  \n\t"
+                         "orr   %0, r2        \n\t"
+                         : "=&r"(ret)
+                         : "r"(address), "r"(1)
+                         : "memory", "r2"
     );
 
   } while (ret);
-
+  MB;
 }
 
 

From d3e2f0a1af73a6e74258294c911e7f4cb72d8ab5 Mon Sep 17 00:00:00 2001
From: Grazvydas Ignotas <notasas@gmail.com>
Date: Sun, 16 Aug 2015 15:37:02 +0200
Subject: [PATCH 2/7] add missing barriers

should fix issue #597
---
 driver/others/blas_server.c | 11 ++++++++++-
 driver/others/memory.c      |  3 +++
 2 files changed, 13 insertions(+), 1 deletion(-)

diff --git a/driver/others/blas_server.c b/driver/others/blas_server.c
index b3b1ce7bd..1fd848c6b 100644
--- a/driver/others/blas_server.c
+++ b/driver/others/blas_server.c
@@ -425,6 +425,10 @@ static int blas_thread_server(void *arg){
       main_status[cpu] = MAIN_FINISH;
 #endif
 
+      // arm: make sure all results are written out _before_
+      // thread is marked as done and other threads use them
+      WMB;
+
       thread_status[cpu].queue = (blas_queue_t * volatile) ((long)thread_status[cpu].queue & 0);  /* Need a trick */
       WMB;
 
@@ -775,7 +779,12 @@ int exec_blas(BLASLONG num, blas_queue_t *queue){
   stop = rpcc();
 #endif
 
-  if ((num > 1) && queue -> next) exec_blas_async_wait(num - 1, queue -> next);
+  if ((num > 1) && queue -> next) {
+    exec_blas_async_wait(num - 1, queue -> next);
+
+    // arm: make sure results from other threads are visible
+    MB;
+  }
 
 #ifdef TIMING_DEBUG
   fprintf(STDERR, "Thread[0] : %16lu %16lu (%8lu cycles)\n",
diff --git a/driver/others/memory.c b/driver/others/memory.c
index a562da377..49c57f911 100644
--- a/driver/others/memory.c
+++ b/driver/others/memory.c
@@ -1153,6 +1153,9 @@ void blas_memory_free(void *free_area){
   printf("  Position : %d\n", position);
 #endif
 
+  // arm: ensure all writes are finished before other thread takes this memory
+  WMB;
+
   memory[position].used = 0;
 
 #ifdef DEBUG

From e12cf1123e8784ce6fe9d2ac14526331fbe2c555 Mon Sep 17 00:00:00 2001
From: Grazvydas Ignotas <notasas@gmail.com>
Date: Sun, 16 Aug 2015 17:27:25 +0200
Subject: [PATCH 3/7] add fallback rpcc implementation

- use on arm, arm64 and any new platform
- use faster integer math instead of double
- use similar scale as rdtsc so that timeouts work
---
 common.h        | 28 ++++++++++++++++++++++++++++
 common_alpha.h  |  1 +
 common_arm.h    | 10 ----------
 common_arm64.h  | 10 ----------
 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, 36 insertions(+), 20 deletions(-)

diff --git a/common.h b/common.h
index 320adadcb..5998b5608 100644
--- a/common.h
+++ b/common.h
@@ -410,7 +410,35 @@ typedef char env_var_t[MAX_PATH];
 typedef char* env_var_t;
 #define readenv(p, n) ((p)=getenv(n))
 #endif
+
+#if !defined(RPCC_DEFINED) && !defined(OS_WINDOWS)
+#ifdef _POSIX_MONOTONIC_CLOCK
+#if defined(__GNUC_PREREQ) && __GLIBC_PREREQ(2, 17) // don't require -lrt
+#define USE_MONOTONIC
+#elif defined(OS_ANDROID)
+#define USE_MONOTONIC
 #endif
+#endif
+/* use similar scale as x86 rdtsc for timeouts to work correctly */
+static inline unsigned long long rpcc(void){
+#ifdef USE_MONOTONIC
+  struct timespec ts;
+  clock_gettime(CLOCK_MONOTONIC, &ts);
+  return (unsigned long long)ts.tv_sec * 1000000000ull + ts.tv_nsec;
+#else
+  struct timeval tv;
+  gettimeofday(&tv,NULL);
+  return (unsigned long long)tv.tv_sec * 1000000000ull + tv.tv_usec * 1000;
+#endif
+}
+#define RPCC_DEFINED
+#define RPCC64BIT
+#endif // !RPCC_DEFINED
+
+#ifndef RPCC_DEFINED
+#error "rpcc() implementation is missing for your platform"
+#endif
+#endif // !ASSEMBLER
 
 #ifdef OS_LINUX
 #include "common_linux.h"
diff --git a/common_alpha.h b/common_alpha.h
index 845fb316a..86f58966a 100644
--- a/common_alpha.h
+++ b/common_alpha.h
@@ -89,6 +89,7 @@ static __inline unsigned int rpcc(void){
 
   return r0;
 }
+#define RPCC_DEFINED
 
 
 #define HALT 	ldq	$0, 0($0)
diff --git a/common_arm.h b/common_arm.h
index 2dabd4d7f..7e0c02306 100644
--- a/common_arm.h
+++ b/common_arm.h
@@ -72,16 +72,6 @@ static void __inline blas_lock(volatile BLASULONG *address){
 }
 
 
-static inline unsigned long long rpcc(void){
-  unsigned long long ret=0;
-  double v;
-  struct timeval tv;
-  gettimeofday(&tv,NULL);
-  v=(double) tv.tv_sec + (double) tv.tv_usec * 1e-6;
-  ret = (unsigned long long) ( v * 1000.0d );
-  return ret;
-}
-
 static inline int blas_quickdivide(blasint x, blasint y){
   return x / y;
 }
diff --git a/common_arm64.h b/common_arm64.h
index aa310c5f2..cc08fa75b 100644
--- a/common_arm64.h
+++ b/common_arm64.h
@@ -71,16 +71,6 @@ static void __inline blas_lock(volatile BLASULONG *address){
 }
 
 
-static inline unsigned long long rpcc(void){
-  unsigned long long ret=0;
-  double v;
-  struct timeval tv;
-  gettimeofday(&tv,NULL);
-  v=(double) tv.tv_sec + (double) tv.tv_usec * 1e-6;
-  ret = (unsigned long long) ( v * 1000.0d );
-  return ret;
-}
-
 static inline int blas_quickdivide(blasint x, blasint y){
   return x / y;
 }
diff --git a/common_ia64.h b/common_ia64.h
index 8e92b5992..d1f210749 100644
--- a/common_ia64.h
+++ b/common_ia64.h
@@ -75,6 +75,7 @@ static __inline unsigned long rpcc(void) {
   __asm__ __volatile__ ("mov %0=ar.itc" : "=r"(clocks));
   return clocks;
 }
+#define RPCC_DEFINED
 
 
 static __inline unsigned long stmxcsr(void){
@@ -103,6 +104,7 @@ static __inline void blas_lock(volatile unsigned long *address){
 static __inline unsigned int rpcc(void) {
   return __getReg(_IA64_REG_AR_ITC);
 }
+#define RPCC_DEFINED
 
 static __inline unsigned int stmxcsr(void) {
   return __getReg(_IA64_REG_AR_FPSR);
diff --git a/common_mips64.h b/common_mips64.h
index 7cd86b375..bc1a52fb4 100644
--- a/common_mips64.h
+++ b/common_mips64.h
@@ -118,6 +118,7 @@ static inline unsigned int rpcc(void){
 #endif
   return ret;
 }
+#define RPCC_DEFINED
 
 #if defined(LOONGSON3A) || defined(LOONGSON3B)
 #ifndef NO_AFFINITY
diff --git a/common_power.h b/common_power.h
index e9b5cb630..3b9471a17 100644
--- a/common_power.h
+++ b/common_power.h
@@ -103,6 +103,7 @@ static inline unsigned long rpcc(void){
 #endif
 
 }
+#define RPCC_DEFINED
 
 #ifdef __64BIT__
 #define RPCC64BIT
diff --git a/common_sparc.h b/common_sparc.h
index 87ef75276..8a16e3d3a 100644
--- a/common_sparc.h
+++ b/common_sparc.h
@@ -66,6 +66,7 @@ static __inline unsigned long rpcc(void){
 
   return clocks;
 };
+#define RPCC_DEFINED
 
 #ifdef __64BIT__
 #define RPCC64BIT
diff --git a/common_x86.h b/common_x86.h
index 99a723fd7..9506716ce 100644
--- a/common_x86.h
+++ b/common_x86.h
@@ -73,6 +73,7 @@ static __inline unsigned long long rpcc(void){
 
   return ((unsigned long long)a + ((unsigned long long)d << 32));
 };
+#define RPCC_DEFINED
 
 static __inline unsigned long getstackaddr(void){
   unsigned long addr;
diff --git a/common_x86_64.h b/common_x86_64.h
index efb902416..3a02beefb 100644
--- a/common_x86_64.h
+++ b/common_x86_64.h
@@ -82,6 +82,7 @@ static __inline BLASULONG rpcc(void){
 
   return ((BLASULONG)a + ((BLASULONG)d << 32));
 }
+#define RPCC_DEFINED
 
 #define RPCC64BIT
 

From f2ac1a5cee9eebfaad33194e362fa2c05e2b05d9 Mon Sep 17 00:00:00 2001
From: Grazvydas Ignotas <notasas@gmail.com>
Date: Sun, 16 Aug 2015 18:08:45 +0200
Subject: [PATCH 4/7] set ARMV7 for Cortex-A9 and Cortex-A15

otherwise some macros like YIELDING are not defined correctly
---
 common_arm.h | 4 ++++
 cpuid_arm.c  | 2 ++
 2 files changed, 6 insertions(+)

diff --git a/common_arm.h b/common_arm.h
index 7e0c02306..74b6378dd 100644
--- a/common_arm.h
+++ b/common_arm.h
@@ -124,4 +124,8 @@ REALNAME:
 #define MAP_ANONYMOUS MAP_ANON
 #endif
 
+#if !defined(ARMV5) && !defined(ARMV6) && !defined(ARMV7) && !defined(ARMV8)
+#error "you must define ARMV5, ARMV6, ARMV7 or ARMV8"
+#endif
+
 #endif
diff --git a/cpuid_arm.c b/cpuid_arm.c
index 51ba72d70..6485003f3 100644
--- a/cpuid_arm.c
+++ b/cpuid_arm.c
@@ -192,6 +192,7 @@ void get_cpuconfig(void)
 	{
 	       case CPU_CORTEXA9:
     			printf("#define CORTEXA9\n");
+    			printf("#define ARMV7\n");
     			printf("#define HAVE_VFP\n");
     			printf("#define HAVE_VFPV3\n");
 			if ( get_feature("neon"))	printf("#define HAVE_NEON\n");
@@ -207,6 +208,7 @@ void get_cpuconfig(void)
 
 	       case CPU_CORTEXA15:
     			printf("#define CORTEXA15\n");
+    			printf("#define ARMV7\n");
     			printf("#define HAVE_VFP\n");
     			printf("#define HAVE_VFPV3\n");
 			if ( get_feature("neon"))	printf("#define HAVE_NEON\n");

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 5/7] 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 5998b5608..6073f037f 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 86f58966a..9739c941d 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 74b6378dd..84691d766 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 cc08fa75b..c4e588d1f 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 d1f210749..72b75fc4e 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 bc1a52fb4..f5c0ec7cf 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 3b9471a17..ab331b04a 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 8a16e3d3a..f99972db9 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 9506716ce..6c90432a2 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 3a02beefb..4c783b315 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;

From d38a1ddc7a4ef8c10017ae5b81a447e322721b94 Mon Sep 17 00:00:00 2001
From: Grazvydas Ignotas <notasas@gmail.com>
Date: Sun, 16 Aug 2015 18:13:30 +0200
Subject: [PATCH 6/7] use real armv5 support

there is no more requirement for ARMv6 instructions,
and VFP on ARMv5 is uncommon
---
 Makefile.arm | 4 ++--
 common_arm.h | 5 ++++-
 getarch.c    | 3 +--
 3 files changed, 7 insertions(+), 5 deletions(-)

diff --git a/Makefile.arm b/Makefile.arm
index 2f7b33730..272220ca9 100644
--- a/Makefile.arm
+++ b/Makefile.arm
@@ -26,8 +26,8 @@ endif
 
 
 ifeq ($(CORE), ARMV5)
-CCOMMON_OPT += -marm -mfpu=vfp -mfloat-abi=hard  -march=armv6
-FCOMMON_OPT += -marm -mfpu=vfp -mfloat-abi=hard  -march=armv6
+CCOMMON_OPT += -marm -march=armv5
+FCOMMON_OPT += -marm -march=armv5
 endif
 
 
diff --git a/common_arm.h b/common_arm.h
index 84691d766..6bf836835 100644
--- a/common_arm.h
+++ b/common_arm.h
@@ -80,7 +80,10 @@ static inline int blas_quickdivide(blasint x, blasint y){
   return x / y;
 }
 
-#if defined(DOUBLE)
+#if !defined(HAVE_VFP)
+/* no FPU, soft float */
+#define GET_IMAGE(res)
+#elif defined(DOUBLE)
 #define GET_IMAGE(res)  __asm__ __volatile__("vstr.f64 d1, %0" : "=m"(res) : : "memory")
 #else
 #define GET_IMAGE(res)  __asm__ __volatile__("vstr.f32 s1, %0" : "=m"(res) : : "memory")
diff --git a/getarch.c b/getarch.c
index d56a37a7a..89e736a31 100644
--- a/getarch.c
+++ b/getarch.c
@@ -798,8 +798,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ARCHCONFIG   "-DARMV5 " \
        "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=32 " \
        "-DL2_SIZE=512488 -DL2_LINESIZE=32 " \
-       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=4 " \
-       "-DHAVE_VFP"
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=4 "
 #define LIBNAME   "armv5"
 #define CORENAME  "ARMV5"
 #else

From 3efeaed0d867c9d54701e9351de44e747cd21578 Mon Sep 17 00:00:00 2001
From: Grazvydas Ignotas <notasas@gmail.com>
Date: Sun, 16 Aug 2015 20:11:13 +0200
Subject: [PATCH 7/7] correct a minor mistake

---
 common.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/common.h b/common.h
index 6073f037f..a607c888b 100644
--- a/common.h
+++ b/common.h
@@ -413,7 +413,7 @@ typedef char* env_var_t;
 
 #if !defined(RPCC_DEFINED) && !defined(OS_WINDOWS)
 #ifdef _POSIX_MONOTONIC_CLOCK
-#if defined(__GNUC_PREREQ) && __GLIBC_PREREQ(2, 17) // don't require -lrt
+#if defined(__GLIBC_PREREQ) && __GLIBC_PREREQ(2, 17) // don't require -lrt
 #define USE_MONOTONIC
 #elif defined(OS_ANDROID)
 #define USE_MONOTONIC