Merge pull request #269 from insertinterestingnamehere/fences

Adjust Fences on X86 to Correct Performance Regression
sandialabs · Sep 3, 2024 · 1aa1958 · 1aa1958
2 parents cb0c514 + 65ec77a
commit 1aa1958
Show file tree

Hide file tree

Showing 2 changed files with 12 additions and 0 deletions.
diff --git a/include/qthread/qthread.h b/include/qthread/qthread.h
@@ -104,6 +104,10 @@ using std::memory_order_relaxed;
 #define Q_ALIGNED(x) __attribute__((aligned(x)))
 #endif
 
+#if QTHREAD_ASSEMBLY_ARCH == QTHREAD_AMD64
+#define QTHREAD_SWAPS_IMPLY_ACQ_REL_FENCES
+#endif
+
 Q_STARTCXX /* */
 /* NOTE!!!!!!!!!!!
  * Reads and writes operate on aligned_t-size segments of memory.

diff --git a/src/feb.c b/src/feb.c
@@ -1127,7 +1127,9 @@ int API_FUNC qthread_writeEF(aligned_t *restrict dest,
       &me->thread_state, QTHREAD_STATE_FEB_BLOCKED, memory_order_relaxed);
     me->rdata->blockedon.addr = m;
     QTHREAD_WAIT_TIMER_START();
+#ifndef QTHREAD_SWAPS_IMPLY_ACQ_REL_FENCES
     MACHINE_FENCE;
+#endif
     qthread_back_to_master(me);
     QTHREAD_WAIT_TIMER_STOP(me, febwait);
 #ifdef QTHREAD_USE_EUREKAS
@@ -1300,7 +1302,9 @@ int API_FUNC qthread_writeFF(aligned_t *restrict dest,
       &me->thread_state, QTHREAD_STATE_FEB_BLOCKED, memory_order_relaxed);
     me->rdata->blockedon.addr = m;
     QTHREAD_WAIT_TIMER_START();
+#ifndef QTHREAD_SWAPS_IMPLY_ACQ_REL_FENCES
     MACHINE_FENCE;
+#endif
     qthread_back_to_master(me);
     QTHREAD_WAIT_TIMER_STOP(me, febwait);
 #ifdef QTHREAD_USE_EUREKAS
@@ -1414,7 +1418,9 @@ int API_FUNC qthread_readFF(aligned_t *restrict dest,
     me->rdata->blockedon.addr = m;
     QTHREAD_WAIT_TIMER_START();
     qthread_back_to_master(me);
+#ifndef QTHREAD_SWAPS_IMPLY_ACQ_REL_FENCES
     MACHINE_FENCE;
+#endif
     QTHREAD_WAIT_TIMER_STOP(me, febwait);
 #ifdef QTHREAD_USE_EUREKAS
     qt_eureka_check(0);
@@ -1600,7 +1606,9 @@ int API_FUNC qthread_readFE(aligned_t *restrict dest,
     me->rdata->blockedon.addr = m;
     QTHREAD_WAIT_TIMER_START();
     qthread_back_to_master(me);
+#ifndef QTHREAD_SWAPS_IMPLY_ACQ_REL_FENCES
     MACHINE_FENCE;
+#endif
     QTHREAD_WAIT_TIMER_STOP(me, febwait);
 #ifdef QTHREAD_USE_EUREKAS
     qt_eureka_check(0);