From: Ivan Maidanski <ivmai@mail.ru>
Date: Mon, 8 Apr 2013 19:52:44 +0000 (+0400)
Subject: Add comment about double-wide load/store on x86_64 (GCC)
X-Git-Tag: libatomic_ops-7_4_0~27
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=ce4a44ca4a1acfb2f6ee44d52993d894dffe7b6f;p=libatomic_ops

Add comment about double-wide load/store on x86_64 (GCC)

* src/atomic_ops/sysdeps/gcc/x86.h: Add comment about AO_double_load
and AO_double_store implementation in 64-bit mode; remove the
corresponding TODO item.
---

diff --git a/src/atomic_ops/sysdeps/gcc/x86.h b/src/atomic_ops/sysdeps/gcc/x86.h
index 9ce3949..62b50c9 100644
--- a/src/atomic_ops/sysdeps/gcc/x86.h
+++ b/src/atomic_ops/sysdeps/gcc/x86.h
@@ -282,7 +282,18 @@ AO_fetch_compare_and_swap_full(volatile AO_t *addr, AO_t old_val,
   }
 # define AO_HAVE_int_fetch_and_add_full
 
-/* TODO: Implement double_load/store. */
+  /* The Intel and AMD Architecture Programmer Manuals state roughly    */
+  /* the following:                                                     */
+  /* - CMPXCHG16B (with a LOCK prefix) can be used to perform 16-byte   */
+  /* atomic accesses in 64-bit mode (with certain alignment             */
+  /* restrictions);                                                     */
+  /* - SSE instructions that access data larger than a quadword (like   */
+  /* MOVDQA) may be implemented using multiple memory accesses;         */
+  /* - LOCK prefix causes an invalid-opcode exception when used with    */
+  /* 128-bit media (SSE) instructions.                                  */
+  /* Thus, currently, the only way to implement lock-free double_load   */
+  /* and double_store on x86_64 is to use CMPXCHG16B (if available).    */
+
 /* TODO: Test some gcc macro to detect presence of cmpxchg16b. */
 
 # ifdef AO_CMPXCHG16B_AVAILABLE