Index: sys/sys/atomic_common.h
===================================================================
--- sys/sys/atomic_common.h
+++ sys/sys/atomic_common.h
@@ -36,38 +36,97 @@
 #error do not include this header, use machine/atomic.h
 #endif
 
-#define	atomic_load_char(p)	(*(volatile u_char *)(p))
-#define	atomic_load_short(p)	(*(volatile u_short *)(p))
-#define	atomic_load_int(p)	(*(volatile u_int *)(p))
-#define	atomic_load_long(p)	(*(volatile u_long *)(p))
-#define	atomic_load_ptr(p)	(*(volatile __typeof(*p) *)(p))
-#define	atomic_load_8(p)	(*(volatile uint8_t *)(p))
-#define	atomic_load_16(p)	(*(volatile uint16_t *)(p))
-#define	atomic_load_32(p)	(*(volatile uint32_t *)(p))
-#ifdef _LP64
-#define	atomic_load_64(p)	(*(volatile uint64_t *)(p))
-#endif
+#include <sys/cdefs.h>
+#include <sys/types.h>
+
+#define	__atomic_load_char(p)	(*(volatile u_char *)(p))
+#define	__atomic_load_short(p)	(*(volatile u_short *)(p))
+#define	__atomic_load_int(p)	(*(volatile u_int *)(p))
+#define	__atomic_load_long(p)	(*(volatile u_long *)(p))
+#define	__atomic_load_8(p)	(*(volatile uint8_t *)(p))
+#define	__atomic_load_16(p)	(*(volatile uint16_t *)(p))
+#define	__atomic_load_32(p)	(*(volatile uint32_t *)(p))
+#define	__atomic_load_64(p)	(*(volatile uint64_t *)(p))
 
-#define	atomic_store_char(p, v)		\
+#define	__atomic_store_char(p, v)		\
     (*(volatile u_char *)(p) = (u_char)(v))
-#define	atomic_store_short(p, v)		\
+#define	__atomic_store_short(p, v)		\
     (*(volatile u_short *)(p) = (u_short)(v))
-#define	atomic_store_int(p, v)		\
+#define	__atomic_store_int(p, v)		\
     (*(volatile u_int *)(p) = (u_int)(v))
-#define	atomic_store_long(p, v)		\
+#define	__atomic_store_long(p, v)		\
     (*(volatile u_long *)(p) = (u_long)(v))
-#define	atomic_store_ptr(p, v)		\
-    (*(volatile __typeof(*p) *)(p) = (v))
-#define	atomic_store_8(p, v)		\
+#define	__atomic_store_8(p, v)			\
     (*(volatile uint8_t *)(p) = (uint8_t)(v))
-#define	atomic_store_16(p, v)		\
+#define	__atomic_store_16(p, v)			\
     (*(volatile uint16_t *)(p) = (uint16_t)(v))
-#define	atomic_store_32(p, v)		\
+#define	__atomic_store_32(p, v)			\
     (*(volatile uint32_t *)(p) = (uint32_t)(v))
-#ifdef _LP64
-#define	atomic_store_64(p, v)		\
+#define	__atomic_store_64(p, v)			\
     (*(volatile uint64_t *)(p) = (uint64_t)(v))
+
+/*
+ * When _Generic is available, try to provide some type checking.
+ */
+#if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L) ||	\
+    __has_extension(c_generic_selections)
+#define	__atomic_load_generic(p, t, ut, n)	\
+	_Generic(*(p),				\
+	    t: __atomic_load_ ## n (p), ut: __atomic_load_ ## n (p))
+#define	atomic_load_char(p)	__atomic_load_generic(p, char, u_char, char)
+#define	atomic_load_short(p)	__atomic_load_generic(p, short, u_short, short)
+#define	atomic_load_int(p)	__atomic_load_generic(p, int, u_int, int)
+#define	atomic_load_long(p)	__atomic_load_generic(p, long, u_long, long)
+#define	atomic_load_8(p)	__atomic_load_generic(p, int8_t, uint8_t, 8)
+#define	atomic_load_16(p)	__atomic_load_generic(p, int16_t, uint16_t, 16)
+#define	atomic_load_32(p)	__atomic_load_generic(p, int32_t, uint32_t, 32)
+#ifdef __LP64__
+#define	atomic_load_64(p)	__atomic_load_generic(p, int64_t, uint64_t, 64)
+#endif
+#define	__atomic_store_generic(p, v, t, ut, n)	\
+	_Generic(*(p),				\
+	    t: __atomic_store_ ## n (p, v), ut: __atomic_store_ ## n (p, v))
+#define	atomic_store_char(p, v)			\
+    __atomic_store_generic(p, v, char, u_char, char)
+#define	atomic_store_short(p, v)		\
+	__atomic_store_generic(p, v, short, u_short, short)
+#define	atomic_store_int(p, v)			\
+	__atomic_store_generic(p, v, int, u_int, int)
+#define	atomic_store_long(p, v)			\
+	__atomic_store_generic(p, v, long, u_long, long)
+#define	atomic_store_8(p, v)			\
+	__atomic_store_generic(p, v, int8_t, uint8_t, 8)
+#define	atomic_store_16(p, v)			\
+	__atomic_store_generic(p, v, int16_t, uint16_t, 16)
+#define	atomic_store_32(p, v)			\
+	__atomic_store_generic(p, v, int32_t, uint32_t, 32)
+#define	atomic_store_64(p, v)			\
+	__atomic_store_generic(p, v, int64_t, uint64_t, 64)
+#else /* !C11 */
+#define	atomic_load_char(p)	__atomic_load_char(p)
+#define	atomic_load_short(p)	__atomic_load_short(p)
+#define	atomic_load_int(p)	__atomic_load_int(p)
+#define	atomic_load_long(p)	__atomic_load_long(p)
+#define	atomic_load_8(p)	__atomic_load_8(p)
+#define	atomic_load_16(p)	__atomic_load_16(p)
+#define	atomic_load_32(p)	__atomic_load_32(p)
+#ifdef __LP64__
+#define	atomic_load_64(p)	__atomic_load_64(p)
 #endif
+#define	atomic_store_char(p, v)	__atomic_store_char(p, v)
+#define	atomic_store_short(p, v) __atomic_store_short(p, v)
+#define	atomic_store_int(p, v)	__atomic_store_int(p, v)
+#define	atomic_store_long(p, v)	__atomic_store_long(p, v)
+#define	atomic_store_8(p, v)	__atomic_store_8(p, v)
+#define	atomic_store_16(p, v)	__atomic_store_16(p, v)
+#define	atomic_store_32(p, v)	__atomic_store_32(p, v)
+#ifdef __LP64__
+#define	atomic_store_64(p, v)	__atomic_store_64(p, v)
+#endif
+#endif /* C11 */
+
+#define	atomic_load_ptr(p)	(*(volatile __typeof(*p) *)(p))
+#define	atomic_store_ptr(p, v)	(*(volatile __typeof(*p) *)(p) = (v))
 
 /*
  * Currently all architectures provide acquire and release fences on their own,