Backport from 10.0:

MDEV-6450 - MariaDB crash on Power8 when built with advance tool
            chain

InnoDB mutex_exit() function calls __sync_test_and_set() to release
the lock. According to manual this function is supposed to create
"acquire" memory barrier whereas in fact we need "release" memory
barrier at mutex_exit().

The problem isn't repeatable with gcc because it creates
"acquire-release" memory barrier for __sync_test_and_set().
ATC creates just "acquire" barrier.

Fixed by creating proper barrier at mutex_exit() by using
__sync_lock_release() instead of __sync_test_and_set().

storage/innobase/include/os0sync.h

@@ -310,6 +310,9 @@ Returns the old value of *ptr, atomically sets *ptr to new_val */
 # define os_atomic_test_and_set_byte(ptr, new_val) \
 	__sync_lock_test_and_set(ptr, (byte) new_val)
 
+# define os_atomic_lock_release_byte(ptr) \
+	__sync_lock_release(ptr)
+
 #elif defined(HAVE_IB_SOLARIS_ATOMICS)
 
 #define HAVE_ATOMIC_BUILTINS
@@ -363,6 +366,9 @@ Returns the old value of *ptr, atomically sets *ptr to new_val */
 # define os_atomic_test_and_set_byte(ptr, new_val) \
 	atomic_swap_uchar(ptr, new_val)
 
+# define os_atomic_lock_release_byte(ptr) \
+	(void) atomic_swap_uchar(ptr, 0)
+
 #elif defined(HAVE_WINDOWS_ATOMICS)
 
 #define HAVE_ATOMIC_BUILTINS
@@ -456,6 +462,9 @@ clobbered */
 # define IB_MEMORY_BARRIER_STARTUP_MSG \
 	"_mm_lfence() and _mm_sfence() are used for memory barrier"
 
+# define os_atomic_lock_release_byte(ptr) \
+	(void) InterlockedExchange(ptr, 0)
+
 #else
 # define os_rmb do { } while(0)
 # define os_wmb do { } while(0)

storage/innobase/include/sync0sync.ic

@@ -109,10 +109,7 @@ mutex_reset_lock_word(
 	mutex_t*	mutex)	/*!< in: mutex */
 {
 #if defined(HAVE_ATOMIC_BUILTINS)
-	/* In theory __sync_lock_release should be used to release the lock.
-	Unfortunately, it does not work properly alone. The workaround is
-	that more conservative __sync_lock_test_and_set is used instead. */
-	os_atomic_test_and_set_byte(&mutex->lock_word, 0);
+	os_atomic_lock_release_byte(&mutex->lock_word);
 #else
 	mutex->lock_word = 0;
 

storage/xtradb/include/os0sync.h

@@ -317,6 +317,9 @@ Returns the old value of *ptr, atomically sets *ptr to new_val */
 # define os_atomic_test_and_set_byte(ptr, new_val) \
 	__sync_lock_test_and_set(ptr, (byte) new_val)
 
+# define os_atomic_lock_release_byte(ptr) \
+	__sync_lock_release(ptr)
+
 #elif defined(HAVE_IB_SOLARIS_ATOMICS)
 
 # define HAVE_ATOMIC_BUILTINS
@@ -374,6 +377,9 @@ Returns the old value of *ptr, atomically sets *ptr to new_val */
 # define os_atomic_test_and_set_byte(ptr, new_val) \
 	atomic_swap_uchar(ptr, new_val)
 
+# define os_atomic_lock_release_byte(ptr) \
+	(void) atomic_swap_uchar(ptr, 0)
+
 #elif defined(HAVE_WINDOWS_ATOMICS)
 
 # define HAVE_ATOMIC_BUILTINS
@@ -476,6 +482,9 @@ clobbered */
 # define IB_MEMORY_BARRIER_STARTUP_MSG \
 	"_mm_lfence() and _mm_sfence() are used for memory barrier"
 
+# define os_atomic_lock_release_byte(ptr) \
+	(void) InterlockedExchange(ptr, 0)
+
 #else
 # define os_rmb do { } while(0)
 # define os_wmb do { } while(0)

storage/xtradb/include/sync0sync.ic

@@ -109,10 +109,7 @@ mutex_reset_lock_word(
 	mutex_t*	mutex)	/*!< in: mutex */
 {
 #if defined(HAVE_ATOMIC_BUILTINS)
-	/* In theory __sync_lock_release should be used to release the lock.
-	Unfortunately, it does not work properly alone. The workaround is
-	that more conservative __sync_lock_test_and_set is used instead. */
-	os_atomic_test_and_set_byte(&mutex->lock_word, 0);
+	os_atomic_lock_release_byte(&mutex->lock_word);
 #else
 	mutex->lock_word = 0;