Improve ethread atomics

The ethread atomics API now also provide double word size atomics.
Double word size atomics are implemented using native atomic
instructions on x86 (when the cmpxchg8b instruction is available)
and on x86_64 (when the cmpxchg16b instruction is available). On
other hardware where 32-bit atomics or word size atomics are
available, an optimized fallback is used; otherwise, a spinlock,
or a mutex based fallback is used.

The ethread library now performs runtime tests for presence of
hardware features, such as for example SSE2 instructions, instead
of requiring this to be determined at compile time.

There are now functions implementing each atomic operation with the
following implied memory barrier semantics: none, read, write,
acquire, release, and full. Some of the operation-barrier
combinations aren't especially useful. But instead of filtering
useful ones out, and potentially miss a useful one, we implement
them all.

A much smaller set of functionality for native atomics are required
to be implemented than before. More or less only cmpxchg and a
membar macro are required to be implemented for each atomic size.
Other functions will automatically be constructed from these. It is,
of course, often wise to implement more that this if possible from a
performance perspective.

4 files changed, 288 insertions, 171 deletions

diff --git a/erts/include/internal/gcc/ethr_atomic.h b/erts/include/internal/gcc/ethr_atomic.h
index 16935084b1..f598f8537b 100644
--- a/erts/include/internal/gcc/ethr_atomic.h
+++ b/erts/include/internal/gcc/ethr_atomic.h
@@ -1,7 +1,7 @@
 /*
  * %CopyrightBegin%
  *
- * Copyright Ericsson AB 2010. All Rights Reserved.
+ * Copyright Ericsson AB 2010-2011. All Rights Reserved.
  *
  * The contents of this file are subject to the Erlang Public License,
  * Version 1.1, (the "License"); you may not use this file except in
@@ -25,11 +25,15 @@
 #undef ETHR_INCLUDE_ATOMIC_IMPL__
 #if !defined(ETHR_GCC_ATOMIC32_H__) && defined(ETHR_ATOMIC_WANT_32BIT_IMPL__)
 #define ETHR_GCC_ATOMIC32_H__
-#define ETHR_INCLUDE_ATOMIC_IMPL__ 4
+#if defined(ETHR_HAVE___SYNC_VAL_COMPARE_AND_SWAP32)
+#  define ETHR_INCLUDE_ATOMIC_IMPL__ 4
+#endif
 #undef ETHR_ATOMIC_WANT_32BIT_IMPL__
 #elif !defined(ETHR_GCC_ATOMIC64_H__) && defined(ETHR_ATOMIC_WANT_64BIT_IMPL__)
 #define ETHR_GCC_ATOMIC64_H__
-#define ETHR_INCLUDE_ATOMIC_IMPL__ 8
+#if defined(ETHR_HAVE___SYNC_VAL_COMPARE_AND_SWAP64)
+#  define ETHR_INCLUDE_ATOMIC_IMPL__ 8
+#endif
 #undef ETHR_ATOMIC_WANT_64BIT_IMPL__
 #endif
 
@@ -45,58 +49,38 @@
 #  define ETHR_READ_AND_SET_WITHOUT_SYNC_OP__ 1
 #endif
 
-#if defined(__x86_64__) || (defined(__i386__) \
-			    && !defined(ETHR_PRE_PENTIUM4_COMPAT))
-#  define ETHR_READ_ACQB_AND_SET_RELB_COMPILER_BARRIER_ONLY__ 1
-#else
-#  define ETHR_READ_ACQB_AND_SET_RELB_COMPILER_BARRIER_ONLY__ 0
-#endif
-
-/*
- * According to the documentation this is what we want:
- *   #define ETHR_MEMORY_BARRIER __sync_synchronize()
- * However, __sync_synchronize() is known to erroneously be
- * a noop on at least some platforms with some gcc versions.
- * This has suposedly been fixed in some gcc version, but we
- * don't know from which version. Therefore, we only use
- * it when it has been verified to work. Otherwise
- * we use a workaround.
- */
-#if defined(__mips__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 2))
-/* __sync_synchronize() has been verified to work here */
-#define ETHR_MEMORY_BARRIER __sync_synchronize()
-#define ETHR_READ_DEPEND_MEMORY_BARRIER __sync_synchronize()
-#elif defined(__x86_64__) || (defined(__i386__) \
-			      && !defined(ETHR_PRE_PENTIUM4_COMPAT))
-/* Use fence instructions directly instead of workaround */
-#define ETHR_MEMORY_BARRIER __asm__ __volatile__("mfence" : : : "memory")
-#define ETHR_WRITE_MEMORY_BARRIER __asm__ __volatile__("sfence" : : : "memory")
-#define ETHR_READ_MEMORY_BARRIER __asm__ __volatile__("lfence" : : : "memory")
-#define ETHR_READ_DEPEND_MEMORY_BARRIER __asm__ __volatile__("" : : : "memory")
-#else
-/* Workaround */
-#define ETHR_MEMORY_BARRIER \
-do { \
-    volatile ethr_sint32_t x___ = 0; \
-    (void) __sync_val_compare_and_swap(&x___, (ethr_sint32_t) 0, (ethr_sint32_t) 1); \
-} while (0)
-#define ETHR_READ_DEPEND_MEMORY_BARRIER ETHR_MEMORY_BARRIER
-#endif
-
-#define ETHR_COMPILER_BARRIER __asm__ __volatile__("" : : : "memory")
-
 #endif /* ETHR_GCC_ATOMIC_COMMON__ */
 
 #if ETHR_INCLUDE_ATOMIC_IMPL__ == 4
 #define ETHR_HAVE_NATIVE_ATOMIC32 1
+#define ETHR_NATIVE_ATOMIC32_IMPL "gcc"
 #define ETHR_NATMC_FUNC__(X) ethr_native_atomic32_ ## X
 #define ETHR_ATMC_T__ ethr_native_atomic32_t
 #define ETHR_AINT_T__ ethr_sint32_t
+#if defined(ETHR_HAVE___SYNC_ADD_AND_FETCH32)
+#  define ETHR_HAVE___SYNC_ADD_AND_FETCH
+#endif
+#if defined(ETHR_HAVE___SYNC_FETCH_AND_AND32)
+#  define ETHR_HAVE___SYNC_FETCH_AND_AND
+#endif 
+#if defined(ETHR_HAVE___SYNC_FETCH_AND_OR32)
+#  define ETHR_HAVE___SYNC_FETCH_AND_OR
+#endif
 #elif ETHR_INCLUDE_ATOMIC_IMPL__ == 8
 #define ETHR_HAVE_NATIVE_ATOMIC64 1
+#define ETHR_NATIVE_ATOMIC64_IMPL "gcc"
 #define ETHR_NATMC_FUNC__(X) ethr_native_atomic64_ ## X
 #define ETHR_ATMC_T__ ethr_native_atomic64_t
 #define ETHR_AINT_T__ ethr_sint64_t
+#if defined(ETHR_HAVE___SYNC_ADD_AND_FETCH64)
+#  define ETHR_HAVE___SYNC_ADD_AND_FETCH
+#endif
+#if defined(ETHR_HAVE___SYNC_FETCH_AND_AND64)
+#  define ETHR_HAVE___SYNC_FETCH_AND_AND
+#endif 
+#if defined(ETHR_HAVE___SYNC_FETCH_AND_OR64)
+#  define ETHR_HAVE___SYNC_FETCH_AND_OR
+#endif
 #else
 #error "Unsupported integer size"
 #endif
@@ -108,183 +92,120 @@ typedef struct {
 
 #if defined(ETHR_TRY_INLINE_FUNCS) || defined(ETHR_ATOMIC_IMPL__)
 
+#if ETHR_INCLUDE_ATOMIC_IMPL__ == 4
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC32_ADDR 1
+#else
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC64_ADDR 1
+#endif
+
 static ETHR_INLINE ETHR_AINT_T__ *
 ETHR_NATMC_FUNC__(addr)(ETHR_ATMC_T__ *var)
 {
     return (ETHR_AINT_T__ *) &var->counter;
 }
 
-static ETHR_INLINE void
-ETHR_NATMC_FUNC__(set)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ value)
-{
 #if ETHR_READ_AND_SET_WITHOUT_SYNC_OP__
-    var->counter = value;
+
+#if ETHR_INCLUDE_ATOMIC_IMPL__ == 4
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC32_SET 1
 #else
-    /*
-     * Unfortunately no __sync_store() or similar exist in the gcc atomic
-     * op interface. We therefore have to simulate it this way...
-     */
-    ETHR_AINT_T__ act = 0, exp;
-    do {
-	exp = act;
-	act = __sync_val_compare_and_swap(&var->counter, exp, value);
-    } while (act != exp);
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC64_SET 1
 #endif
-}
 
 static ETHR_INLINE void
-ETHR_NATMC_FUNC__(init)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ value)
+ETHR_NATMC_FUNC__(set)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ value)
 {
-    ETHR_NATMC_FUNC__(set)(var, value);
+    var->counter = value;
 }
 
-static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(read)(ETHR_ATMC_T__ *var)
-{
+#endif /* ETHR_READ_AND_SET_WITHOUT_SYNC_OP__ */
+
 #if ETHR_READ_AND_SET_WITHOUT_SYNC_OP__
-    return var->counter;
+
+#if ETHR_INCLUDE_ATOMIC_IMPL__ == 4
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC32_READ 1
 #else
-    /*
-     * Unfortunately no __sync_fetch() or similar exist in the gcc atomic
-     * op interface. We therefore have to simulate it this way...
-     */
-    return __sync_add_and_fetch(&var->counter, (ETHR_AINT_T__) 0);
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC64_READ 1
 #endif
-}
-
-static ETHR_INLINE void
-ETHR_NATMC_FUNC__(add)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ incr)
-{
-    (void) __sync_add_and_fetch(&var->counter, incr);
-}
 
 static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(add_return)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ incr)
+ETHR_NATMC_FUNC__(read)(ETHR_ATMC_T__ *var)
 {
-    return __sync_add_and_fetch(&var->counter, incr);
+    return var->counter;
 }
 
-static ETHR_INLINE void
-ETHR_NATMC_FUNC__(inc)(ETHR_ATMC_T__ *var)
-{
-    (void) __sync_add_and_fetch(&var->counter, (ETHR_AINT_T__) 1);
-}
+#endif /* ETHR_READ_AND_SET_WITHOUT_SYNC_OP__ */
 
-static ETHR_INLINE void
-ETHR_NATMC_FUNC__(dec)(ETHR_ATMC_T__ *var)
-{
-    (void) __sync_sub_and_fetch(&var->counter, (ETHR_AINT_T__) 1);
-}
+#if defined(ETHR_HAVE___SYNC_ADD_AND_FETCH)
 
-static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(inc_return)(ETHR_ATMC_T__ *var)
-{
-    return __sync_add_and_fetch(&var->counter, (ETHR_AINT_T__) 1);
-}
+#if ETHR_INCLUDE_ATOMIC_IMPL__ == 4
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC32_ADD_RETURN_MB 1
+#else
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC64_ADD_RETURN_MB 1
+#endif
 
 static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(dec_return)(ETHR_ATMC_T__ *var)
+ETHR_NATMC_FUNC__(add_return_mb)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ incr)
 {
-    return __sync_sub_and_fetch(&var->counter, (ETHR_AINT_T__) 1);
+    return __sync_add_and_fetch(&var->counter, incr);
 }
 
-static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(and_retold)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ mask)
-{
-    return __sync_fetch_and_and(&var->counter, mask);
-}
+#endif
 
-static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(or_retold)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ mask)
-{
-    return (ETHR_AINT_T__) __sync_fetch_and_or(&var->counter, mask);
-}
+#if defined(ETHR_HAVE___SYNC_FETCH_AND_AND)
 
-static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(cmpxchg)(ETHR_ATMC_T__ *var,
-			   ETHR_AINT_T__ new,
-			   ETHR_AINT_T__ old)
-{
-    return __sync_val_compare_and_swap(&var->counter, old, new);
-}
+#if ETHR_INCLUDE_ATOMIC_IMPL__ == 4
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC32_AND_RETOLD_MB 1
+#else
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC64_AND_RETOLD_MB 1
+#endif
 
 static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(xchg)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ new)
+ETHR_NATMC_FUNC__(and_retold_mb)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ mask)
 {
-    ETHR_AINT_T__ exp, act = 0;
-    do {
-	exp = act;
-	act = __sync_val_compare_and_swap(&var->counter, exp, new);
-    } while (act != exp);
-    return act;
+    return __sync_fetch_and_and(&var->counter, mask);
 }
 
-/*
- * Atomic ops with at least specified barriers.
- */
-
-static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(read_acqb)(ETHR_ATMC_T__ *var)
-{
-#if ETHR_READ_ACQB_AND_SET_RELB_COMPILER_BARRIER_ONLY__
-    ETHR_AINT_T__ val = var->counter;
-    ETHR_COMPILER_BARRIER;
-    return val;
-#else
-    return __sync_add_and_fetch(&var->counter, (ETHR_AINT_T__) 0);
 #endif
-}
 
-static ETHR_INLINE void
-ETHR_NATMC_FUNC__(set_relb)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ i)
-{
-#if ETHR_READ_ACQB_AND_SET_RELB_COMPILER_BARRIER_ONLY__
-    ETHR_COMPILER_BARRIER;
-    var->counter = i;
+#if defined(ETHR_HAVE___SYNC_FETCH_AND_OR)
+
+#if ETHR_INCLUDE_ATOMIC_IMPL__ == 4
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC32_OR_RETOLD_MB 1
 #else
-    (void) ETHR_NATMC_FUNC__(xchg)(var, i);
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC64_OR_RETOLD_MB 1
 #endif
-}
 
 static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(inc_return_acqb)(ETHR_ATMC_T__ *var)
-{
-    return ETHR_NATMC_FUNC__(inc_return)(var);
-}
-
-static ETHR_INLINE void
-ETHR_NATMC_FUNC__(dec_relb)(ETHR_ATMC_T__ *var)
+ETHR_NATMC_FUNC__(or_retold_mb)(ETHR_ATMC_T__ *var, ETHR_AINT_T__ mask)
 {
-    ETHR_NATMC_FUNC__(dec)(var);
+    return (ETHR_AINT_T__) __sync_fetch_and_or(&var->counter, mask);
 }
 
-static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(dec_return_relb)(ETHR_ATMC_T__ *var)
-{
-    return ETHR_NATMC_FUNC__(dec_return)(var);
-}
+#endif
 
-static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(cmpxchg_acqb)(ETHR_ATMC_T__ *var,
-				ETHR_AINT_T__ new,
-				ETHR_AINT_T__ old)
-{
-    return ETHR_NATMC_FUNC__(cmpxchg)(var, new, old);
-}
+#if ETHR_INCLUDE_ATOMIC_IMPL__ == 4
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC32_CMPXCHG_MB 1
+#else
+#  define ETHR_HAVE_ETHR_NATIVE_ATOMIC64_CMPXCHG_MB 1
+#endif
 
 static ETHR_INLINE ETHR_AINT_T__
-ETHR_NATMC_FUNC__(cmpxchg_relb)(ETHR_ATMC_T__ *var,
-				ETHR_AINT_T__ new,
-				ETHR_AINT_T__ old)
+ETHR_NATMC_FUNC__(cmpxchg_mb)(ETHR_ATMC_T__ *var,
+			      ETHR_AINT_T__ new,
+			      ETHR_AINT_T__ old)
 {
-    return ETHR_NATMC_FUNC__(cmpxchg)(var, new, old);
+    return __sync_val_compare_and_swap(&var->counter, old, new);
 }
 
-#endif
+#endif /* ETHR_TRY_INLINE_FUNCS */
 
 #undef ETHR_NATMC_FUNC__
 #undef ETHR_ATMC_T__
 #undef ETHR_AINT_T__
 #undef ETHR_AINT_SUFFIX__
+#undef ETHR_HAVE___SYNC_ADD_AND_FETCH
+#undef ETHR_HAVE___SYNC_FETCH_AND_AND
+#undef ETHR_HAVE___SYNC_FETCH_AND_OR
 
 #endif
diff --git a/erts/include/internal/gcc/ethr_dw_atomic.h b/erts/include/internal/gcc/ethr_dw_atomic.h
new file mode 100644
index 0000000000..6736f9c547
--- /dev/null
+++ b/erts/include/internal/gcc/ethr_dw_atomic.h
@@ -0,0 +1,115 @@
+/*
+ * %CopyrightBegin%
+ *
+ * Copyright Ericsson AB 2011. All Rights Reserved.
+ *
+ * The contents of this file are subject to the Erlang Public License,
+ * Version 1.1, (the "License"); you may not use this file except in
+ * compliance with the License. You should have received a copy of the
+ * Erlang Public License along with this software. If not, it can be
+ * retrieved online at http://www.erlang.org/.
+ *
+ * Software distributed under the License is distributed on an "AS IS"
+ * basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See
+ * the License for the specific language governing rights and limitations
+ * under the License.
+ *
+ * %CopyrightEnd%
+ */
+
+/*
+ * Description: Native double word atomics using gcc's builtins
+ * Author: Rickard Green
+ */
+
+#undef ETHR_INCLUDE_DW_ATOMIC_IMPL__
+#ifndef ETHR_GCC_DW_ATOMIC_H__
+#  define ETHR_GCC_DW_ATOMIC_H__
+#  if ((ETHR_SIZEOF_PTR == 4 \
+        && defined(ETHR_HAVE___SYNC_VAL_COMPARE_AND_SWAP64)) \
+       || (ETHR_SIZEOF_PTR == 8 \
+           && defined(ETHR_HAVE___SYNC_VAL_COMPARE_AND_SWAP128) \
+	   && defined(ETHR_HAVE_INT128_T)))
+#    define ETHR_INCLUDE_DW_ATOMIC_IMPL__
+#  endif
+#endif
+
+#ifdef ETHR_INCLUDE_DW_ATOMIC_IMPL__
+#  define ETHR_HAVE_NATIVE_SU_DW_ATOMIC
+#  define ETHR_NATIVE_DW_ATOMIC_IMPL "gcc"
+
+#  if defined(__i386__) || defined(__x86_64__)
+/*
+ * If ETHR_RTCHK_USE_NATIVE_DW_ATOMIC_IMPL__ is defined, it will be used
+ * at runtime in order to determine if native or fallback implementation
+ * should be used.
+ */
+#    define ETHR_RTCHK_USE_NATIVE_DW_ATOMIC_IMPL__ \
+       ETHR_X86_RUNTIME_CONF_HAVE_DW_CMPXCHG__
+#  endif
+
+#  if ETHR_SIZEOF_PTR == 4
+#    define ETHR_DW_NATMC_ALIGN_MASK__ 0x7
+#    define ETHR_NATIVE_SU_DW_SINT_T ethr_sint64_t
+#  elif ETHR_SIZEOF_PTR == 8
+#    define ETHR_DW_NATMC_ALIGN_MASK__ 0xf
+#    define ETHR_NATIVE_SU_DW_SINT_T ethr_sint128_t
+#  endif
+
+typedef volatile ETHR_NATIVE_SU_DW_SINT_T * ethr_native_dw_ptr_t;
+
+/*
+ * We need 16 byte aligned memory in 64-bit mode, and 8 byte aligned
+ * memory in 32-bit mode. 16 byte aligned malloc in 64-bit mode is
+ * not common, and at least some glibc malloc implementations
+ * only 4 byte align in 32-bit mode.
+ *
+ * This code assumes 8 byte aligned memory in 64-bit mode, and 4 byte
+ * aligned memory in 32-bit mode. A malloc implementation that does
+ * not adhere to these alignment requirements is seriously broken,
+ * and we wont bother trying to work around it.
+ *
+ * Since memory alignment may be off by one word we need to align at
+ * runtime. We, therefore, need an extra word allocated.
+ */
+#define ETHR_DW_NATMC_MEM__(VAR) \
+   (&var->c[(int) ((ethr_uint_t) &(VAR)->c[0]) & ETHR_DW_NATMC_ALIGN_MASK__])
+typedef union {
+    volatile ETHR_NATIVE_SU_DW_SINT_T dw_sint;
+    volatile ethr_sint_t sint[3];
+    volatile char c[ETHR_SIZEOF_PTR*3];
+} ethr_native_dw_atomic_t;
+
+#if defined(ETHR_TRY_INLINE_FUNCS) || defined(ETHR_ATOMIC_IMPL__)
+
+#  ifdef ETHR_DEBUG
+#    define ETHR_DW_DBG_ALIGNED__(PTR) \
+       ETHR_ASSERT((((ethr_uint_t) (PTR)) & ETHR_DW_NATMC_ALIGN_MASK__) == 0);
+#  else
+#    define ETHR_DW_DBG_ALIGNED__(PTR)
+#  endif
+
+#define ETHR_HAVE_ETHR_NATIVE_DW_ATOMIC_ADDR
+static ETHR_INLINE ethr_sint_t *
+ethr_native_dw_atomic_addr(ethr_native_dw_atomic_t *var)
+{
+    return (ethr_sint_t *) ETHR_DW_NATMC_MEM__(var);
+}
+
+
+#define ETHR_HAVE_ETHR_NATIVE_SU_DW_ATOMIC_CMPXCHG_MB
+
+static ETHR_INLINE ETHR_NATIVE_SU_DW_SINT_T
+ethr_native_su_dw_atomic_cmpxchg_mb(ethr_native_dw_atomic_t *var,
+				    ETHR_NATIVE_SU_DW_SINT_T new,
+				    ETHR_NATIVE_SU_DW_SINT_T old)
+{
+    ethr_native_dw_ptr_t p = (ethr_native_dw_ptr_t) ETHR_DW_NATMC_MEM__(var);
+    ETHR_DW_DBG_ALIGNED__(p);
+    return __sync_val_compare_and_swap(p, old, new);
+}
+
+#endif /* ETHR_TRY_INLINE_FUNCS */
+
+#endif /* ETHR_GCC_DW_ATOMIC_H__ */
+
diff --git a/erts/include/internal/gcc/ethr_membar.h b/erts/include/internal/gcc/ethr_membar.h
new file mode 100644
index 0000000000..7d428fc68e
--- /dev/null
+++ b/erts/include/internal/gcc/ethr_membar.h
@@ -0,0 +1,73 @@
+/*
+ * %CopyrightBegin%
+ *
+ * Copyright Ericsson AB 2011. All Rights Reserved.
+ *
+ * The contents of this file are subject to the Erlang Public License,
+ * Version 1.1, (the "License"); you may not use this file except in
+ * compliance with the License. You should have received a copy of the
+ * Erlang Public License along with this software. If not, it can be
+ * retrieved online at http://www.erlang.org/.
+ *
+ * Software distributed under the License is distributed on an "AS IS"
+ * basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See
+ * the License for the specific language governing rights and limitations
+ * under the License.
+ *
+ * %CopyrightEnd%
+ */
+
+/*
+ * Description: Memory barriers when using gcc's builtins
+ * Author: Rickard Green
+ */
+
+#ifndef ETHR_GCC_MEMBAR_H__
+#define ETHR_GCC_MEMBAR_H__
+
+#define ETHR_LoadLoad	(1 << 0)
+#define ETHR_LoadStore	(1 << 1)
+#define ETHR_StoreLoad	(1 << 2)
+#define ETHR_StoreStore	(1 << 3)
+
+/*
+ * According to the documentation __sync_synchronize() will
+ * issue a full memory barrier. However, __sync_synchronize()
+ * is known to erroneously be a noop on at least some
+ * platforms with some gcc versions. This has suposedly been
+ * fixed in some gcc version, but we don't know from which
+ * version. Therefore, we only use it when it has been
+ * verified to work. Otherwise we use the workaround
+ * below.
+ */
+
+#if defined(ETHR_HAVE___SYNC_VAL_COMPARE_AND_SWAP32)
+#  define ETHR_MB_T__ ethr_sint32_t
+#elif defined(ETHR_HAVE___SYNC_VAL_COMPARE_AND_SWAP64)
+#  define ETHR_MB_T__ ethr_sint64_t
+#else
+#  error "No __sync_val_compare_and_swap"
+#endif
+#define ETHR_SYNC_SYNCHRONIZE_WORKAROUND__ \
+do { \
+    volatile ETHR_MB_T__ x___ = 0; \
+    (void) __sync_val_compare_and_swap(&x___, (ETHR_MB_T__) 0, (ETHR_MB_T__) 1); \
+} while (0)
+
+#define ETHR_COMPILER_BARRIER __asm__ __volatile__("" : : : "memory")
+
+#if defined(__mips__) && ETHR_AT_LEAST_GCC_VSN__(4, 2, 0)
+#  define ETHR_MEMBAR(B) __sync_synchronize()
+#  define ETHR_READ_DEPEND_MEMORY_BARRIER __sync_synchronize()
+#elif ((defined(__powerpc__) || defined(__ppc__)) \
+       && ETHR_AT_LEAST_GCC_VSN__(4, 1, 2))
+#  define ETHR_MEMBAR(B) __sync_synchronize()
+#else /* Use workaround */
+#  define ETHR_MEMBAR(B) \
+     ETHR_SYNC_SYNCHRONIZE_WORKAROUND__
+#  define ETHR_READ_DEPEND_MEMORY_BARRIER \
+     ETHR_SYNC_SYNCHRONIZE_WORKAROUND__
+#endif
+
+
+#endif /* ETHR_GCC_MEMBAR_H__ */
diff --git a/erts/include/internal/gcc/ethread.h b/erts/include/internal/gcc/ethread.h
index 392a1aa2b2..fcfdc39441 100644
--- a/erts/include/internal/gcc/ethread.h
+++ b/erts/include/internal/gcc/ethread.h
@@ -1,7 +1,7 @@
 /*
  * %CopyrightBegin%
  *
- * Copyright Ericsson AB 2010. All Rights Reserved.
+ * Copyright Ericsson AB 2010-2011. All Rights Reserved.
  *
  * The contents of this file are subject to the Erlang Public License,
  * Version 1.1, (the "License"); you may not use this file except in
@@ -25,16 +25,24 @@
 #ifndef ETHREAD_GCC_H__
 #define ETHREAD_GCC_H__
 
-#if !defined(ETHR_HAVE_NATIVE_ATOMICS) && defined(ETHR_HAVE_GCC_ATOMIC_OPS)
-#define ETHR_HAVE_NATIVE_ATOMICS 1
+#ifndef ETHR_MEMBAR
+#  include "ethr_membar.h"
+#endif
+
+#if !defined(ETHR_HAVE_NATIVE_ATOMIC32)
+#  define ETHR_ATOMIC_WANT_32BIT_IMPL__
+#  include "ethr_atomic.h"
+#endif
 
-#define ETHR_ATOMIC_WANT_32BIT_IMPL__
-#include "ethr_atomic.h"
-#if ETHR_SIZEOF_PTR == 8
+#if ETHR_SIZEOF_PTR == 8 && !defined(ETHR_HAVE_NATIVE_ATOMIC64)
 #  define ETHR_ATOMIC_WANT_64BIT_IMPL__
 #  include "ethr_atomic.h"
 #endif
 
+#if (!defined(ETHR_HAVE_NATIVE_DW_ATOMIC) \
+     && !(ETHR_SIZEOF_PTR == 4 && defined(ETHR_HAVE_NATIVE_ATOMIC64)) \
+     && !(ETHR_SIZEOF_PTR == 8 && defined(ETHR_HAVE_NATIVE_ATOMIC128)))
+#  include "ethr_dw_atomic.h"
 #endif
 
 #endif