Potential bug in broadcast sends on CMA builds

[ritvikrao]

In src/arch/util/machine-common-core.C, line 650, in function CmiInterSendNetworkFunc, there is no return statement here:

#if CMK_USE_CMA
        if(cma_reg_msg && partition == CmiMyPartition() && CmiPeOnSamePhysicalNode(CmiMyPe(), destPE)) {
          if(CMI_CMA_MSGTYPE(msg) == CMK_REG_NO_CMA_MSG && cma_min_threshold <= size && size <= cma_max_threshold) {
            CmiSendMessageCma(&msg, &size); // size & msg are modififed
          }
        }
#endif 

This means that send to PEs on the same physical node will use (from line 692):

LrtsSendFunc(destNode, CmiGetPeGlobal(destPE,partition), size, msg, mode)

[stwhite91]

This is not a bug. If you look at the definition of CmiSendMessageCma() you can see that that routine only updates the buffer pointer to include some CMA metadata. That updated buffer pointer is then passed to LrtsSendFunc which does the actual message transmission.