uatomic/x86: Remove redundant memory barriers
[urcu.git] / include / urcu / uatomic / x86.h
index c02c96d88af8bcf4beaedb36475daacf6a741a46..616eee9be3ca60b4be62c8a10cadb0aff2d1ea8d 100644 (file)
@@ -1,21 +1,16 @@
+// SPDX-FileCopyrightText: 1991-1994 by Xerox Corporation.  All rights reserved.
+// SPDX-FileCopyrightText: 1996-1999 by Silicon Graphics.  All rights reserved.
+// SPDX-FileCopyrightText: 1999-2004 Hewlett-Packard Development Company, L.P.
+// SPDX-FileCopyrightText: 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+//
+// SPDX-License-Identifier: LicenseRef-Boehm-GC
+
 #ifndef _URCU_ARCH_UATOMIC_X86_H
 #define _URCU_ARCH_UATOMIC_X86_H
 
+#include <stdlib.h>            /* For abort(3). */
+
 /*
- * Copyright (c) 1991-1994 by Xerox Corporation.  All rights reserved.
- * Copyright (c) 1996-1999 by Silicon Graphics.  All rights reserved.
- * Copyright (c) 1999-2004 Hewlett-Packard Development Company, L.P.
- * Copyright (c) 2009      Mathieu Desnoyers
- *
- * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
- * OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
- *
- * Permission is hereby granted to use or copy this program
- * for any purpose,  provided the above notices are retained on all copies.
- * Permission to modify the code and to distribute modified code is granted,
- * provided the above notices are retained, and a notice that the code was
- * modified is included with the above copyright notice.
- *
  * Code inspired from libuatomic_ops-1.2, inherited in part from the
  * Boehm-Demers-Weiser conservative garbage collector.
  */
@@ -41,10 +36,16 @@ extern "C" {
  * containing an array of char of the specified size. This allows passing the
  * @addr arguments of the following inline functions as "m" and "+m" operands
  * to the assembly. The @size parameter should be a constant to support
- * compilers such as clang which do not support VLA.
+ * compilers such as clang which do not support VLA. Create typedefs because
+ * C++ does not allow types be defined in casts.
  */
 
-#define __hp(size, x)  ((struct { char v[size]; } *)(x))
+typedef struct { char v[1]; } __hp_1;
+typedef struct { char v[2]; } __hp_2;
+typedef struct { char v[4]; } __hp_4;
+typedef struct { char v[8]; } __hp_8;
+
+#define __hp(size, x)  ((__hp_##size *)(x))
 
 #define _uatomic_set(addr, v)  ((void) CMM_STORE_SHARED(*(addr), (v)))
 
@@ -523,7 +524,7 @@ void __uatomic_dec(void *addr, int len)
 
 #define _uatomic_dec(addr)     (__uatomic_dec((addr), sizeof(*(addr))))
 
-#if ((CAA_BITS_PER_LONG != 64) && defined(URCU_ARCH_I386))
+#ifdef URCU_ARCH_X86_NO_CAS
 
 /* For backwards compat */
 #define CONFIG_RCU_COMPAT_ARCH 1
@@ -631,6 +632,474 @@ extern unsigned long _compat_uatomic_add_return(void *addr,
 #define cmm_smp_mb__before_uatomic_dec()       cmm_barrier()
 #define cmm_smp_mb__after_uatomic_dec()                cmm_barrier()
 
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_read_mo(enum cmm_memorder mo)
+{
+       /*
+        * A SMP barrier is not necessary for CMM_SEQ_CST because, only a
+        * previous store can be reordered with the load.  However, emitting the
+        * memory barrier after the store is sufficient to prevent reordering
+        * between the two.  This follows toolchains decision of emitting the
+        * memory fence on the stores instead of the loads.
+        *
+        * A compiler barrier is necessary because the underlying operation does
+        * not clobber the registers.
+        */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               cmm_barrier();
+               break;
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       default:
+               abort();
+               break;
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_read_mo(enum cmm_memorder mo)
+{
+       /*
+        * A SMP barrier is not necessary for CMM_SEQ_CST because following
+        * loads and stores cannot be reordered with the load.
+        *
+        * A SMP barrier is however necessary for CMM_SEQ_CST_FENCE to respect
+        * the memory model, since the underlying operation does not have a lock
+        * prefix.
+        *
+        * A compiler barrier is necessary because the underlying operation does
+        * not clobber the registers.
+        */
+       switch (mo) {
+       case CMM_SEQ_CST_FENCE:
+               cmm_smp_mb();
+               break;
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_SEQ_CST:
+               cmm_barrier();
+               break;
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       default:
+               abort();
+               break;
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_set_mo(enum cmm_memorder mo)
+{
+       /*
+        * A SMP barrier is not necessary for CMM_SEQ_CST because the store can
+        * only be reodered with later loads
+        *
+        * A compiler barrier is necessary because the underlying operation does
+        * not clobber the registers.
+        */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               cmm_barrier();
+               break;
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       default:
+               abort();
+               break;
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_set_mo(enum cmm_memorder mo)
+{
+       /*
+        * A SMP barrier is necessary for CMM_SEQ_CST because the store can be
+        * reorded with later loads.  Since no memory barrier is being emitted
+        * before loads, one has to be emitted after the store.  This follows
+        * toolchains decision of emitting the memory fence on the stores instead
+        * of the loads.
+        *
+        * A SMP barrier is necessary for CMM_SEQ_CST_FENCE to respect the
+        * memory model, since the underlying store does not have a lock prefix.
+        *
+        * A compiler barrier is necessary because the underlying operation does
+        * not clobber the registers.
+        */
+       switch (mo) {
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               cmm_smp_mb();
+               break;
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_RELEASE:
+               cmm_barrier();
+               break;
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       default:
+               abort();
+               break;
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_xchg_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_xchg has implicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_xchg_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_xchg has implicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_cmpxchg_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_cmpxchg has implicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_cmpxchg_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_cmpxchg has implicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_and_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_and has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_and_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_and has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_or_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_or has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_or_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_or has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_add_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_add has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_add_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_add has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_sub_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_sub has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_sub_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_sub has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_inc_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_inc has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_inc_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_inc has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_dec_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_dec has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_dec_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_dec has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_add_return_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_add_return has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_add_return_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_add_return has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__before_uatomic_sub_return_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_sub_return has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+static inline void _cmm_compat_c11_smp_mb__after_uatomic_sub_return_mo(enum cmm_memorder mo)
+{
+       /* NOP. uatomic_sub_return has explicit lock prefix. */
+       switch (mo) {
+       case CMM_RELAXED:       /* Fall-through */
+       case CMM_ACQUIRE:       /* Fall-through */
+       case CMM_CONSUME:       /* Fall-through */
+       case CMM_RELEASE:       /* Fall-through */
+       case CMM_ACQ_REL:       /* Fall-through */
+       case CMM_SEQ_CST:       /* Fall-through */
+       case CMM_SEQ_CST_FENCE:
+               break;
+       default:
+               abort();
+       }
+}
+
+#define _cmm_compat_c11_smp_mb__before_mo(operation, mo)                       \
+       do {                                                    \
+               _cmm_compat_c11_smp_mb__before_ ## operation ## _mo (mo);       \
+       } while (0)
+
+#define _cmm_compat_c11_smp_mb__after_mo(operation, mo)                        \
+       do {                                                    \
+               _cmm_compat_c11_smp_mb__after_ ## operation ## _mo (mo);        \
+       } while (0)
+
+
 #ifdef __cplusplus
 }
 #endif
This page took 0.027738 seconds and 4 git commands to generate.