Charm++ parallel programming system

["Sameer Kumar12" <sameer_km AT in.ibm.com>]

Hi,

Can we integrate the following CKLOOP fix? The current implementation has no memory fence calls and heavily relies on the processor not reordering memory stores (e.g. in an X86 processors).

diff --git a/src/libs/ck-libs/ckloop/CkLoop.C b/src/libs/ck-libs/ckloop/CkLoop.C
index a9efec7..bcb45f9 100644
--- a/src/libs/ck-libs/ckloop/CkLoop.C
+++ b/src/libs/ck-libs/ckloop/CkLoop.C
@@ -563,6 +563,7 @@ void CurLoopInfo::stealWork() {
}

fnPtr(first, last, redBufs[nextChunkId], paramNum, param);
+ CmiMemoryWriteFence(); //Wait for shared data to be updated.
execTimes++;
nextChunkId = getNextChunkIdx();
}
diff --git a/src/libs/ck-libs/ckloop/CkLoop.h b/src/libs/ck-libs/ckloop/CkLoop.h
index 80ab125..950a22e 100644
--- a/src/libs/ck-libs/ckloop/CkLoop.h
+++ b/src/libs/ck-libs/ckloop/CkLoop.h
@@ -83,7 +83,8 @@ public:
void waitLoopDone(int sync) {
//while(!__sync_bool_compare_and_swap(&finishFlag, numChunks, 0));
if (sync) while (finishFlag!=numChunks);
- //finishFlag = 0;
+ CmiMemoryReadFence();
+ //finishFlag = 0;
CmiLock(loop_info_inited_lock);
inited = 0;
CmiUnlock(loop_info_inited_lock);
@@ -117,11 +118,14 @@ public:
#endif
#else
__sync_add_and_fetch(&finishFlag, counter);
+ CmiMemoryWriteFence(); //optional..
#endif
}

int isFree() {
- return finishFlag == numChunks;
+ int fin = finishFlag == numChunks;
+ CmiMemoryReadFence();
+ return fin;
}

void **getRedBufs() {