New macro AO_PREFER_BUILTIN_ATOMICS to rely on C11 atomics fully (AArch64)

If AO_PREFER_BUILTIN_ATOMICS is defined then inline assembly
implementation is not used.

* src/atomic_ops/sysdeps/gcc/aarch64.h (AO_SKIPATOMIC_double_load,
AO_SKIPATOMIC_double_load_acquire): Define if AO_PREFER_BUILTIN_ATOMICS.
* src/atomic_ops/sysdeps/gcc/aarch64.h (AO_nop_write, AO_double_load,
AO_double_load_acquire, AO_double_store, AO_double_store_release,
AO_double_compare_and_swap, AO_double_compare_and_swap_acquire,
AO_double_compare_and_swap_release, AO_double_compare_and_swap_full):
Do not define if AO_PREFER_BUILTIN_ATOMICS.
* src/atomic_ops/sysdeps/gcc/aarch64.h (AO_nop_write): Add comment;
remove TODO.
* src/atomic_ops/sysdeps/gcc/generic.h: Remove TODO about including
this header for other targets.
* src/atomic_ops/sysdeps/gcc/generic.h (AO_double_load): Do not define
if AO_SKIPATOMIC_double_load.
* src/atomic_ops/sysdeps/gcc/generic.h (AO_double_load_acquire): Do not
define if AO_SKIPATOMIC_double_load_acquire.

2 files changed, 25 insertions, 17 deletions

diff --git a/src/atomic_ops/sysdeps/gcc/aarch64.h b/src/atomic_ops/sysdeps/gcc/aarch64.h
index b7f2ef0..612bc18 100644
--- a/src/atomic_ops/sysdeps/gcc/aarch64.h
+++ b/src/atomic_ops/sysdeps/gcc/aarch64.h
@@ -19,18 +19,23 @@
 
 #include "../standard_ao_double_t.h"
 
-#ifndef AO_UNIPROCESSOR
-  AO_INLINE void
-  AO_nop_write(void)
-  {
-    /* TODO: Use C++11 primitive. */
-    __asm__ __volatile__("dmb ishst" : : : "memory");
-  }
-# define AO_HAVE_nop_write
-#endif
-
-/* TODO: Adjust version check on fixing double-wide AO support in GCC. */
-#if __GNUC__ >= 4
+#ifdef AO_PREFER_BUILTIN_ATOMICS
+  /* As of clang 3.6 (and gcc 5.0), load atomics for double word are    */
+  /* translated to incorrect code lacking STXP (see the note below).    */
+# define AO_SKIPATOMIC_double_load
+# define AO_SKIPATOMIC_double_load_acquire
+#else
+
+  /* As of clang 3.6 (and gcc 4.9), __atomic_thread_fence is always     */
+  /* translated to DMB (which is inefficient for AO_nop_write).         */
+# ifndef AO_UNIPROCESSOR
+    AO_INLINE void
+    AO_nop_write(void)
+    {
+      __asm__ __volatile__("dmb ishst" : : : "memory");
+    }
+#   define AO_HAVE_nop_write
+# endif
 
   AO_INLINE AO_double_t
   AO_double_load(const volatile AO_double_t *addr)
@@ -68,6 +73,9 @@
   }
 # define AO_HAVE_double_load_acquire
 
+  /* As of gcc 5.0, all built-in store and CAS atomics for double       */
+  /* word require -latomic, so use asm-based implementation by default. */
+
   AO_INLINE void
   AO_double_store(volatile AO_double_t *addr, AO_double_t value)
   {
@@ -195,6 +203,7 @@
     return !result;
   }
 # define AO_HAVE_double_compare_and_swap_full
-#endif /* __GNUC__ >= 4 */
+
+#endif /* !AO_PREFER_BUILTIN_ATOMICS */
 
 #include "generic.h"
diff --git a/src/atomic_ops/sysdeps/gcc/generic.h b/src/atomic_ops/sysdeps/gcc/generic.h
index de79edb..1df2914 100644
--- a/src/atomic_ops/sysdeps/gcc/generic.h
+++ b/src/atomic_ops/sysdeps/gcc/generic.h
@@ -19,8 +19,6 @@
 /* For the details, see GNU Manual, chapter 6.52 (Built-in functions    */
 /* for memory model aware atomic operations).                           */
 
-/* TODO: Include this file for other targets if gcc 4.7+ */
-
 #ifdef AO_UNIPROCESSOR
   /* If only a single processor (core) is used, AO_UNIPROCESSOR could   */
   /* be defined by the client to avoid unnecessary memory barrier.      */
@@ -93,7 +91,7 @@
 
 #ifdef AO_HAVE_DOUBLE_PTR_STORAGE
 
-# ifndef AO_HAVE_double_load
+# if !defined(AO_HAVE_double_load) && !defined(AO_SKIPATOMIC_double_load)
     AO_INLINE AO_double_t
     AO_double_load(const volatile AO_double_t *addr)
     {
@@ -105,7 +103,8 @@
 #   define AO_HAVE_double_load
 # endif
 
-# ifndef AO_HAVE_double_load_acquire
+# if !defined(AO_HAVE_double_load_acquire) \
+     && !defined(AO_SKIPATOMIC_double_load_acquire)
     AO_INLINE AO_double_t
     AO_double_load_acquire(const volatile AO_double_t *addr)
     {