summaryrefslogtreecommitdiff
path: root/libitm
diff options
context:
space:
mode:
authortorvald <torvald@138bc75d-0d04-0410-961f-82ee72b054a4>2012-02-20 20:56:56 +0000
committertorvald <torvald@138bc75d-0d04-0410-961f-82ee72b054a4>2012-02-20 20:56:56 +0000
commit62e1cffcdd1dd7356a2c90ed62d36ef493f96c7b (patch)
treeeecdab57a37716f23f733bcc7296e45e68124beb /libitm
parentdd6f37d576b7d571475b79da9c1e2098ffae1279 (diff)
downloadgcc-62e1cffcdd1dd7356a2c90ed62d36ef493f96c7b.tar.gz
libitm: Optimize memory order requiremens in gl_wt pre_write.
libtim/ * method-gl.cc (gl_wt_dispatch::pre_write): Optimize memory orders. git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@184400 138bc75d-0d04-0410-961f-82ee72b054a4
Diffstat (limited to 'libitm')
-rw-r--r--libitm/ChangeLog4
-rw-r--r--libitm/method-gl.cc17
2 files changed, 18 insertions, 3 deletions
diff --git a/libitm/ChangeLog b/libitm/ChangeLog
index 75410f52ed1..0a06bfec7f4 100644
--- a/libitm/ChangeLog
+++ b/libitm/ChangeLog
@@ -1,5 +1,9 @@
2012-02-20 Torvald Riegel <triegel@redhat.com>
+ * method-gl.cc (gl_wt_dispatch::pre_write): Optimize memory orders.
+
+2012-02-20 Torvald Riegel <triegel@redhat.com>
+
* method-gl.cc (gl_wt_dispatch::memtransfer_static): Add missing undo
for RfW src. Optimize number of calls to gtm_thr.
diff --git a/libitm/method-gl.cc b/libitm/method-gl.cc
index 4a18b62d740..503c19e5921 100644
--- a/libitm/method-gl.cc
+++ b/libitm/method-gl.cc
@@ -104,16 +104,27 @@ protected:
tx->restart(RESTART_VALIDATE_WRITE);
// CAS global orec from our snapshot time to the locked state.
- // We need acq_rel memory order here to synchronize with other loads
- // and modifications of orec.
+ // We need acquire memory order here to synchronize with other
+ // (ownership) releases of the orec. We do not need acq_rel order
+ // because whenever another thread reads from this CAS'
+ // modification, then it will abort anyway and does not rely on
+ // any further happens-before relation to be established.
+ // Also note that unlike in ml_wt's increase of the global time
+ // base (remember that the global orec is used as time base), we do
+ // not need require memory order here because we do not need to make
+ // prior orec acquisitions visible to other threads that try to
+ // extend their snapshot time.
if (!o_gl_mg.orec.compare_exchange_strong (now, gl_mg::set_locked(now),
- memory_order_acq_rel))
+ memory_order_acquire))
tx->restart(RESTART_LOCKED_WRITE);
// We use an explicit fence here to avoid having to use release
// memory order for all subsequent data stores. This fence will
// synchronize with loads of the data with acquire memory order. See
// validate() for why this is necessary.
+ // Adding require memory order to the prior CAS is not sufficient,
+ // at least according to the Batty et al. formalization of the
+ // memory model.
atomic_thread_fence(memory_order_release);
// Set shared_state to new value.