# HG changeset patch
# User fyang
# Date 1573715257 -28800
#      Thu Nov 14 15:07:37 2019 +0800
# Node ID aa0f4a1592004a4f2b786204190b699ae9bfe19b
# Parent  b987ea528c21997ebb11b48079c672786f8f9b0c
8233912: aarch64: minor improvements of atomic operations
Reviewed-by: aph

diff --git a/src/hotspot/os_cpu/linux_aarch64/atomic_linux_aarch64.hpp b/src/hotspot/os_cpu/linux_aarch64/atomic_linux_aarch64.hpp
--- a/src/hotspot/os_cpu/linux_aarch64/atomic_linux_aarch64.hpp
+++ b/src/hotspot/os_cpu/linux_aarch64/atomic_linux_aarch64.hpp
@@ -29,6 +29,8 @@
 #include "vm_version_aarch64.hpp"
 
 // Implementation of class atomic
+// Note that memory_order_conservative requires a full barrier after atomic stores.
+// See https://patchwork.kernel.org/patch/3575821/
 
 #define FULL_MEM_BARRIER  __sync_synchronize()
 #define READ_MEM_BARRIER  __atomic_thread_fence(__ATOMIC_ACQUIRE);
@@ -52,7 +54,7 @@
                                                      T volatile* dest,
                                                      atomic_memory_order order) const {
   STATIC_ASSERT(byte_size == sizeof(T));
-  T res = __sync_lock_test_and_set(dest, exchange_value);
+  T res = __atomic_exchange_n(dest, exchange_value, __ATOMIC_RELEASE);
   FULL_MEM_BARRIER;
   return res;
 }
@@ -70,7 +72,12 @@
                               __ATOMIC_RELAXED, __ATOMIC_RELAXED);
     return value;
   } else {
-    return __sync_val_compare_and_swap(dest, compare_value, exchange_value);
+    T value = compare_value;
+    FULL_MEM_BARRIER;
+    __atomic_compare_exchange(dest, &value, &exchange_value, /*weak*/false,
+                              __ATOMIC_RELAXED, __ATOMIC_RELAXED);
+    FULL_MEM_BARRIER;
+    return value;
   }
 }