From: Bart Van Assche <bvanassche@acm.org>
Date: Sat, 11 Jul 2009 14:35:59 +0000 (+0000)
Subject: Clarified an implementation choice.
X-Git-Tag: svn/VALGRIND_3_5_0~418
X-Git-Url: http://git.ipfire.org/gitweb.cgi?a=commitdiff_plain;h=5f83d3907f3a9df999eee3dcdba25362a5c99718;p=thirdparty%2Fvalgrind.git

Clarified an implementation choice.

git-svn-id: svn://svn.valgrind.org/valgrind/trunk@10429
---

diff --git a/drd/drd_load_store.c b/drd/drd_load_store.c
index 6398dac0d9..e123b5c31f 100644
--- a/drd/drd_load_store.c
+++ b/drd/drd_load_store.c
@@ -557,11 +557,14 @@ IRSB* DRD_(instrument)(VgCallbackClosure* const closure,
       case Ist_CAS:
          if (instrument)
          {
-            /* Just treat this as a read of the location.  I believe
-               this is equivalent to the previous logic, which
-               observed bus-lock/unlock Ist_MBEs, and ignored all
-               writes within sections bracketed by bus-lock and
-               bus-unlock annotations. */
+            /*
+             * Treat compare-and-swap as a read. By handling atomic
+             * instructions as read instructions no data races are reported
+             * between conflicting atomic operations nor between atomic
+             * operations and non-atomic reads. Conflicts between atomic
+             * operations and non-atomic write operations are still reported
+             * however.
+             */
             Int    dataSize;
             IRCAS* cas = st->Ist.CAS.details;
             tl_assert(cas->addr != NULL);