Better atomic support for gcc and clang

include/ucore/atomic.h

@@ -3,59 +3,131 @@
 
 /** @file
  * Atomic functions (macros). The ptr argument must be a pointer to 
- * an integer type.
+ * an UC_ATOMIC(integer type)
  * For gcc, see http://gcc.gnu.org/onlinedocs/gcc-4.8.1/gcc/_005f_005fsync-Builtins.html
  */
+#ifdef DOXYGEN
 
-#ifdef __GNUC__
+/** 
+ * Define an atomic type.
+ * e.g UC_ATOMIC(int) foo;
+ */
+#define UC_ATOMIC(type)
 
 /**
  * Atomically subtract val from *ptr
  * @return *ptr before the subtraction
  */
-#define UC_ATOMIC_SUB(ptr, val) __sync_fetch_and_add((ptr), -(val))
+#define UC_ATOMIC_SUB(ptr, val) 
 
 /**
  * Atomically add val to *ptr
  * @return *ptr before the addition
  */
-#define UC_ATOMIC_ADD(ptr, val) __sync_fetch_and_add((ptr), (val))
-
+#define UC_ATOMIC_ADD(ptr, val)
 
 /**
  * Atomically decrement *ptr
  * @return *ptr before the decrement
  */
-#define UC_ATOMIC_DEC(ptr) UC_ATOMIC_SUB((ptr), 1)
+#define UC_ATOMIC_DEC(ptr) 
 
 /**
  * Atomically increment *ptr
  * @return *ptr before the increment
  */
-#define UC_ATOMIC_INC(ptr) UC_ATOMIC_ADD((ptr), 1)
+#define UC_ATOMIC_INC(ptr) 
 
 /**
  * Atomically get the value of *ptr
  * @return value of *ptr
  */
-#define UC_ATOMIC_GET(ptr) UC_ATOMIC_ADD((ptr), 0)
+#define UC_ATOMIC_GET(ptr) 
 
 /**
  * Atomic Compare-And-Swap. If *ptr is oldval, write newwal to *ptr. 
  * @return *ptr that was before the operation.
  */
-#define UC_ATOMIC_CAS(ptr, oldval, newval) __sync_val_compare_and_swap((ptr), (oldval), (newval))
+#define UC_ATOMIC_CAS(ptr, oldval, newval) 
 
 /** Issue a full (hardware) memory barrier.
  * preventing the cpu  to move load/stores across
  * the barrier. (This implies a compiler barrier as well)
  * */
-#define UC_MEM_BARRIER() __sync_synchronize()
+#define UC_MEM_BARRIER() 
 
 /** Issue a full compiler/software only memory barrier 
  * preventing compiler optimization to move load/stores across
  * the barrier.
  * */
+#define UC_COMPILER_BARRIER() 
+
+/* clang with atomic extensions */
+#elif __clang__
+
+/* clang defines 0 to be memory_order_relaxed and 5 to be memory_order_seq_cst */
+#define UC_ORDER_RELAXED 0
+#define UC_ORDER_CST 5
+
+#define UC_ATOMIC(type) _Atomic(type)
+
+#define UC_ATOMIC_SUB(ptr, val) __c11_atomic_fetch_sub((ptr), (val), UC_ORDER_CST)
+
+#define UC_ATOMIC_ADD(ptr, val) __c11_atomic_fetch_add((ptr), (val), UC_ORDER_CST)
+
+#define UC_ATOMIC_DEC(ptr) UC_ATOMIC_SUB((ptr), 1)
+
+#define UC_ATOMIC_INC(ptr) UC_ATOMIC_ADD((ptr), 1)
+
+#define UC_ATOMIC_GET(ptr) __c11_atomic_load((ptr), UC_ORDER_CST)
+
+#define UC_ATOMIC_CAS(ptr, oldval, newval) __atomic_compare_exchange((ptr), (oldval), (newval),0, UC_ORDER_CST, UC_ORDER_CST)
+
+#define UC_MEM_BARRIER() __c11_atomic_thread_fence(UC_ORDER_CST)
+
+#define UC_COMPILER_BARRIER() __asm__ __volatile__("": : :"memory")
+
+/* gcc 4.7 or later */
+#elif __GNUC_ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 7)
+
+#define UC_ATOMIC(type) type
+
+#define UC_ATOMIC_SUB(ptr, val) __atomic_fetch_sub((ptr), (val), __ATOMIC_SEQ_CST)
+
+#define UC_ATOMIC_ADD(ptr, val) __atomic_fetch_add((ptr), (val), __ATOMIC_SEQ_CST)
+
+#define UC_ATOMIC_DEC(ptr) UC_ATOMIC_SUB((ptr), 1)
+
+#define UC_ATOMIC_INC(ptr) UC_ATOMIC_ADD((ptr), 1)
+
+#define UC_ATOMIC_GET(ptr) __atomic_load_n((ptr), __ATOMIC_SEQ_CST)
+
+#define UC_ATOMIC_CAS(ptr, oldval, newval) __atomic_compare_exchange((ptr), (oldval), (newval),0, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)
+
+#define UC_MEM_BARRIER() __atomic_thread_fence(__ATOMIC_SEQ_CST)
+
+#define UC_COMPILER_BARRIER() __asm__ __volatile__("": : :"memory")
+
+
+/* Earlier gcc versions */
+#elif __GNUC__
+
+#define UC_ATOMIC(type) type
+
+#define UC_ATOMIC_SUB(ptr, val) __sync_fetch_and_sub((ptr), (val))
+
+#define UC_ATOMIC_ADD(ptr, val) __sync_fetch_and_add((ptr), (val))
+
+#define UC_ATOMIC_DEC(ptr) UC_ATOMIC_SUB((ptr), 1)
+
+#define UC_ATOMIC_INC(ptr) UC_ATOMIC_ADD((ptr), 1)
+
+#define UC_ATOMIC_GET(ptr) UC_ATOMIC_ADD((ptr), 0)
+
+#define UC_ATOMIC_CAS(ptr, oldval, newval) __sync_val_compare_and_swap((ptr), (oldval), (newval))
+
+#define UC_MEM_BARRIER() __sync_synchronize()
+
 #define UC_COMPILER_BARRIER() __asm__ __volatile__("": : :"memory")
 
 #else