[utils/atomic] Fix native atomic intrinsics on buggy ARM64 compilers.

See the comment in atomic.h for details.

diff --git a/mono/utils/atomic.h b/mono/utils/atomic.h
index 4f5f010afd31bd61551347f754fe9f31639b0bae..2459f8835d6b91512916c646eb4f2a364423b724 100755 (executable)
--- a/mono/utils/atomic.h
+++ b/mono/utils/atomic.h
@@ -14,6 +14,7 @@
 
 #include "config.h"
 #include <glib.h>
+#include <mono/utils/mono-membar.h>
 
 /*
 The current Nexus 7 arm-v7a fails with:
@@ -29,7 +30,6 @@ Apple targets have historically being problematic, xcode 4.6 would miscompile th
 #define WIN32_LEAN_AND_MEAN
 #endif
 #include <windows.h>
-#include <mono/utils/mono-membar.h>
 
 /* mingw is missing InterlockedCompareExchange64 () from winbase.h */
 #if HAVE_DECL_INTERLOCKEDCOMPAREEXCHANGE64==0
@@ -180,30 +180,63 @@ static inline void InterlockedWrite16(volatile gint16 *dst, gint16 val)
 /* Prefer GCC atomic ops if the target supports it (see configure.ac). */
 #elif defined(USE_GCC_ATOMIC_OPS)
 
+/*
+ * As of this comment (August 2016), all current Clang versions get atomic
+ * intrinsics on ARM64 wrong. All GCC versions prior to 5.3.0 do, too. The bug
+ * is the same: The compiler developers thought that the acq + rel barriers
+ * that ARM64 load/store instructions can impose are sufficient to provide
+ * sequential consistency semantics. This is not the case:
+ *
+ *     http://lists.infradead.org/pipermail/linux-arm-kernel/2014-February/229588.html
+ *
+ * We work around this bug by inserting full barriers around each atomic
+ * intrinsic if we detect that we're built with a buggy compiler.
+ */
+
+#if defined (HOST_ARM64) && (defined (__clang__) || MONO_GNUC_VERSION < 50300)
+#define WRAP_ATOMIC_INTRINSIC(INTRIN) \
+       ({ \
+               mono_memory_barrier (); \
+               __typeof__ (INTRIN) atomic_ret__ = (INTRIN); \
+               mono_memory_barrier (); \
+               atomic_ret__; \
+       })
+
+#define gcc_sync_val_compare_and_swap(a, b, c) WRAP_ATOMIC_INTRINSIC (__sync_val_compare_and_swap (a, b, c))
+#define gcc_sync_add_and_fetch(a, b) WRAP_ATOMIC_INTRINSIC (__sync_add_and_fetch (a, b))
+#define gcc_sync_sub_and_fetch(a, b) WRAP_ATOMIC_INTRINSIC (__sync_sub_and_fetch (a, b))
+#define gcc_sync_fetch_and_add(a, b) WRAP_ATOMIC_INTRINSIC (__sync_fetch_and_add (a, b))
+#else
+#define gcc_sync_val_compare_and_swap(a, b, c) __sync_val_compare_and_swap (a, b, c)
+#define gcc_sync_add_and_fetch(a, b) __sync_add_and_fetch (a, b)
+#define gcc_sync_sub_and_fetch(a, b) __sync_sub_and_fetch (a, b)
+#define gcc_sync_fetch_and_add(a, b) __sync_fetch_and_add (a, b)
+#endif
+
 static inline gint32 InterlockedCompareExchange(volatile gint32 *dest,
                                                gint32 exch, gint32 comp)
 {
-       return __sync_val_compare_and_swap (dest, comp, exch);
+       return gcc_sync_val_compare_and_swap (dest, comp, exch);
 }
 
 static inline gpointer InterlockedCompareExchangePointer(volatile gpointer *dest, gpointer exch, gpointer comp)
 {
-       return __sync_val_compare_and_swap (dest, comp, exch);
+       return gcc_sync_val_compare_and_swap (dest, comp, exch);
 }
 
 static inline gint32 InterlockedAdd(volatile gint32 *dest, gint32 add)
 {
-       return __sync_add_and_fetch (dest, add);
+       return gcc_sync_add_and_fetch (dest, add);
 }
 
 static inline gint32 InterlockedIncrement(volatile gint32 *val)
 {
-       return __sync_add_and_fetch (val, 1);
+       return gcc_sync_add_and_fetch (val, 1);
 }
 
 static inline gint32 InterlockedDecrement(volatile gint32 *val)
 {
-       return __sync_sub_and_fetch (val, 1);
+       return gcc_sync_sub_and_fetch (val, 1);
 }
 
 static inline gint32 InterlockedExchange(volatile gint32 *val, gint32 new_val)
@@ -211,7 +244,7 @@ static inline gint32 InterlockedExchange(volatile gint32 *val, gint32 new_val)
        gint32 old_val;
        do {
                old_val = *val;
-       } while (__sync_val_compare_and_swap (val, old_val, new_val) != old_val);
+       } while (gcc_sync_val_compare_and_swap (val, old_val, new_val) != old_val);
        return old_val;
 }
 
@@ -221,30 +254,30 @@ static inline gpointer InterlockedExchangePointer(volatile gpointer *val,
        gpointer old_val;
        do {
                old_val = *val;
-       } while (__sync_val_compare_and_swap (val, old_val, new_val) != old_val);
+       } while (gcc_sync_val_compare_and_swap (val, old_val, new_val) != old_val);
        return old_val;
 }
 
 static inline gint32 InterlockedExchangeAdd(volatile gint32 *val, gint32 add)
 {
-       return __sync_fetch_and_add (val, add);
+       return gcc_sync_fetch_and_add (val, add);
 }
 
 static inline gint8 InterlockedRead8(volatile gint8 *src)
 {
        /* Kind of a hack, but GCC doesn't give us anything better, and it's
         * certainly not as bad as using a CAS loop. */
-       return __sync_fetch_and_add (src, 0);
+       return gcc_sync_fetch_and_add (src, 0);
 }
 
 static inline gint16 InterlockedRead16(volatile gint16 *src)
 {
-       return __sync_fetch_and_add (src, 0);
+       return gcc_sync_fetch_and_add (src, 0);
 }
 
 static inline gint32 InterlockedRead(volatile gint32 *src)
 {
-       return __sync_fetch_and_add (src, 0);
+       return gcc_sync_fetch_and_add (src, 0);
 }
 
 static inline void InterlockedWrite8(volatile gint8 *dst, gint8 val)
@@ -253,7 +286,7 @@ static inline void InterlockedWrite8(volatile gint8 *dst, gint8 val)
        gint8 old_val;
        do {
                old_val = *dst;
-       } while (__sync_val_compare_and_swap (dst, old_val, val) != old_val);
+       } while (gcc_sync_val_compare_and_swap (dst, old_val, val) != old_val);
 }
 
 static inline void InterlockedWrite16(volatile gint16 *dst, gint16 val)
@@ -261,7 +294,7 @@ static inline void InterlockedWrite16(volatile gint16 *dst, gint16 val)
        gint16 old_val;
        do {
                old_val = *dst;
-       } while (__sync_val_compare_and_swap (dst, old_val, val) != old_val);
+       } while (gcc_sync_val_compare_and_swap (dst, old_val, val) != old_val);
 }
 
 static inline void InterlockedWrite(volatile gint32 *dst, gint32 val)
@@ -270,7 +303,7 @@ static inline void InterlockedWrite(volatile gint32 *dst, gint32 val)
        gint32 old_val;
        do {
                old_val = *dst;
-       } while (__sync_val_compare_and_swap (dst, old_val, val) != old_val);
+       } while (gcc_sync_val_compare_and_swap (dst, old_val, val) != old_val);
 }
 
 #if defined (TARGET_OSX) || defined (__arm__) || (defined (__mips__) && !defined (__mips64)) || (defined (__powerpc__) && !defined (__powerpc64__)) || (defined (__sparc__) && !defined (__arch64__))
@@ -281,33 +314,33 @@ static inline void InterlockedWrite(volatile gint32 *dst, gint32 val)
 
 static inline gint64 InterlockedCompareExchange64(volatile gint64 *dest, gint64 exch, gint64 comp)
 {
-       return __sync_val_compare_and_swap (dest, comp, exch);
+       return gcc_sync_val_compare_and_swap (dest, comp, exch);
 }
 
 static inline gint64 InterlockedAdd64(volatile gint64 *dest, gint64 add)
 {
-       return __sync_add_and_fetch (dest, add);
+       return gcc_sync_add_and_fetch (dest, add);
 }
 
 static inline gint64 InterlockedIncrement64(volatile gint64 *val)
 {
-       return __sync_add_and_fetch (val, 1);
+       return gcc_sync_add_and_fetch (val, 1);
 }
 
 static inline gint64 InterlockedDecrement64(volatile gint64 *val)
 {
-       return __sync_sub_and_fetch (val, 1);
+       return gcc_sync_sub_and_fetch (val, 1);
 }
 
 static inline gint64 InterlockedExchangeAdd64(volatile gint64 *val, gint64 add)
 {
-       return __sync_fetch_and_add (val, add);
+       return gcc_sync_fetch_and_add (val, add);
 }
 
 static inline gint64 InterlockedRead64(volatile gint64 *src)
 {
        /* Kind of a hack, but GCC doesn't give us anything better. */
-       return __sync_fetch_and_add (src, 0);
+       return gcc_sync_fetch_and_add (src, 0);
 }
 
 #else