DM: add wrappers to gcc built-in atomic operations

This patch introduces wrappers to the built-in atomic operations provided by
gcc.

There are two sets of built-in atomic functions available. Before gcc 4.7.0 only
the __sync built-in functions are available, while since 4.7.0 a new set of
built-in functions with the __atomic prefix is introduced as a
replacement. Since the __sync functions will eventually be deprecated, the
__atomic ones are preferred whenever available.

The interfaces provided are listed below, mostly following the naming of the
underlying built-in functions which explain themselves.

    atomic_load
    atomic_store
    atomic_xchg
    atomic_cmpxchg

    atomic_add_fetch
    atomic_sub_fetch
    atomic_and_fetch
    atomic_xor_fetch
    atomic_or_fetch
    atomic_nand_fetch

    atomic_fetch_add
    atomic_fetch_sub
    atomic_fetch_and
    atomic_fetch_xor
    atomic_fetch_or
    atomic_fetch_nand

    atomic_test_and_set
    atomic_clear
    atomic_thread_fence
    atomic_signal_fence

Tracked-On: #875
Signed-off-by: Junjie Mao <junjie.mao@intel.com>
Acked-by: Yu Wang <yu1.wang@intel.com>

devicemodel/include/atomic.h

@@ -0,0 +1,124 @@
+/*
+ * Copyright (C) 2018 Intel Corporation. All rights reserved.
+ *
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+
+#ifndef _ATOMIC_H_
+#define _ATOMIC_H_
+
+/* Test for GCC >= 4.7.0 */
+#if ((__GNUC__ > 4) || (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7)))
+
+/* Since GCC 4.7.0, the __atomic builtins are introduced as a replacement of the
+ * __sync ones. The original __sync builtins maps to their __atomic counter-part
+ * using the __ATOMIC_SEQ_CST model and will be eventually deprecated. */
+
+#define atomic_load(ptr)			\
+	__atomic_load_n(ptr, __ATOMIC_SEQ_CST)
+
+#define atomic_store(ptr, val)				\
+	__atomic_store_n(ptr, val, __ATOMIC_SEQ_CST)
+
+#define atomic_xchg(ptr, val)				\
+	__atomic_exchange_n(ptr, val, __ATOMIC_SEQ_CST)
+
+/* Note: expected should also be a pointer. */
+#define atomic_cmpxchg(ptr, expected, desired)				\
+	__atomic_compare_exchange_n(ptr, expected, desired,		\
+				false, __ATOMIC_SEQ_CST, __ATOMIC_ACQUIRE)
+
+#define atomic_add_fetch(ptr, val)			\
+	__atomic_add_fetch(ptr, val, __ATOMIC_SEQ_CST)
+#define atomic_sub_fetch(ptr, val)			\
+	__atomic_sub_fetch(ptr, val, __ATOMIC_SEQ_CST)
+#define atomic_and_fetch(ptr, val)			\
+	__atomic_and_fetch(ptr, val, __ATOMIC_SEQ_CST)
+#define atomic_xor_fetch(ptr, val)			\
+	__atomic_xor_fetch(ptr, val, __ATOMIC_SEQ_CST)
+#define atomic_or_fetch(ptr, val)			\
+	__atomic_or_fetch(ptr, val, __ATOMIC_SEQ_CST)
+#define atomic_nand_fetch(ptr, val)			\
+	__atomic_nand_fetch(ptr, val, __ATOMIC_SEQ_CST)
+
+#define atomic_fetch_add(ptr, val)			\
+	__atomic_fetch_add(ptr, val, __ATOMIC_SEQ_CST)
+#define atomic_fetch_sub(ptr, val)			\
+	__atomic_fetch_sub(ptr, val, __ATOMIC_SEQ_CST)
+#define atomic_fetch_and(ptr, val)			\
+	__atomic_fetch_and(ptr, val, __ATOMIC_SEQ_CST)
+#define atomic_fetch_xor(ptr, val)			\
+	__atomic_fetch_xor(ptr, val, __ATOMIC_SEQ_CST)
+#define atomic_fetch_or(ptr, val)			\
+	__atomic_fetch_or(ptr, val, __ATOMIC_SEQ_CST)
+#define atomic_fetch_nand(ptr, val)			\
+	__atomic_fetch_nand(ptr, val, __ATOMIC_SEQ_CST)
+
+#define atomic_test_and_set(ptr)		\
+	__atomic_test_and_set(ptr, __ATOMIC_SEQ_CST)
+#define atomic_clear(ptr)			\
+	__atomic_clear(ptr, __ATOMIC_SEQ_CST)
+#define atomic_thread_fence()			\
+	__atomic_thread_fence(__ATOMIC_SEQ_CST)
+#define atomic_signal_fence()			\
+	__atomic_signal_fence(__ATOMIC_SEQ_CST)
+
+#else  /* not GCC >= 4.7.0 */
+
+/* __sync builtins do not have load/store interfaces. Use add_fetch and xchg to
+ * mimic their functinality.
+ *
+ * Also note that __sync_lock_test_and_set is rather an atomic exchange
+ * operation per GCC manual on the __sync builtins.
+ */
+#define atomic_load(ptr)			\
+	__sync_add_and_fetch(ptr, 0)
+
+#define atomic_store(ptr, val)			\
+	(void)(__sync_lock_test_and_set(ptr, val))
+
+#define atomic_xchg(ptr, val)			\
+	__sync_lock_test_and_set(ptr, val)
+
+/* Note: expected should also be a pointer. */
+#define atomic_cmpxchg(ptr, expected, desired)			\
+	__sync_bool_compare_and_swap(ptr, (*(expected)), desired)
+
+#define atomic_add_fetch(ptr, val)		\
+	__sync_add_and_fetch(ptr, val)
+#define atomic_sub_fetch(ptr, val)		\
+	__sync_sub_and_fetch(ptr, val)
+#define atomic_and_fetch(ptr, val)		\
+	__sync_and_and_fetch(ptr, val)
+#define atomic_xor_fetch(ptr, val)		\
+	__sync_xor_and_fetch(ptr, val)
+#define atomic_or_fetch(ptr, val)		\
+	__sync_or_and_fetch(ptr, val)
+#define atomic_nand_fetch(ptr, val)		\
+	__sync_nand_and_fetch(ptr, val)
+
+#define atomic_fetch_add(ptr, val)		\
+	__sync_fetch_and_add(ptr, val)
+#define atomic_fetch_sub(ptr, val)		\
+	__sync_fetch_and_sub(ptr, val)
+#define atomic_fetch_and(ptr, val)		\
+	__sync_fetch_and_and(ptr, val)
+#define atomic_fetch_xor(ptr, val)		\
+	__sync_fetch_and_xor(ptr, val)
+#define atomic_fetch_or(ptr, val)		\
+	__sync_fetch_and_or(ptr, val)
+#define atomic_fetch_nand(ptr, val)		\
+	__sync_fetch_and_nand(ptr, val)
+
+#define atomic_test_and_set(ptr)		\
+	(bool)(__sync_lock_test_and_set(ptr, 1))
+#define atomic_clear(ptr)			\
+	__sync_lock_release(ptr)
+#define atomic_thread_fence()			\
+	__sync_synchronize()
+#define atomic_signal_fence()			\
+	__sync_synchronize()
+
+#endif /* GCC >= 4.7.0 */
+
+#endif /* _ATOMIC_H_ */