Merge remote-tracking branch 'remotes/nccl/master'

[ROCm/rccl commit: 8e73a2ad60]

projects/rccl/src/init.cc

@@ -127,7 +127,13 @@ ncclResult_t ncclGetUniqueId(ncclUniqueId* out) {
 }
 
 // Prevent compiler from optimizing out these operations
-void __attribute__((optimize("O0"))) commPoison(ncclComm_t comm) {
+#ifdef __clang__
+#define NCCL_NO_OPTIMIZE __attribute__((optnone))
+#else
+#define NCCL_NO_OPTIMIZE __attribute__((optimize("O0")))
+#endif
+
+void NCCL_NO_OPTIMIZE commPoison(ncclComm_t comm) {
   comm->rank = comm->cudaDev = comm->busId = comm->nRanks = -1;
 }
 
@@ -179,6 +185,8 @@ void *ncclCommThreadMain(void *arg) {
 }
 #endif
 
+#undef NCCL_NO_OPTIMIZE
+
 static ncclResult_t commFree(ncclComm_t comm) {
   if (comm == NULL)
     return ncclSuccess;