Merge branch 'master' of git+ssh://scm.gforge.inria.fr//gitroot/simgrid/simgrid

[simgrid.git] / src / smpi / smpi_pmpi.c

diff --git a/src/smpi/smpi_pmpi.c b/src/smpi/smpi_pmpi.c
index 69de4f4..d7fed42 100644 (file)
--- a/src/smpi/smpi_pmpi.c
+++ b/src/smpi/smpi_pmpi.c
@@ -27,6 +27,7 @@ void TRACE_smpi_set_category(const char *category)
 int PMPI_Init(int *argc, char ***argv)
 {
   smpi_process_init(argc, argv);
+  smpi_process_mark_as_initialized();
 #ifdef HAVE_TRACING
   int rank = smpi_process_index();
   TRACE_smpi_init(rank);
@@ -82,7 +83,7 @@ int PMPI_Init_thread(int *argc, char ***argv, int required, int *provided)
 
 int PMPI_Query_thread(int *provided)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (provided == NULL) {
@@ -97,7 +98,7 @@ int PMPI_Query_thread(int *provided)
 
 int PMPI_Is_thread_main(int *flag)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (flag == NULL) {
@@ -140,7 +141,7 @@ double PMPI_Wtick(void)
 
 int PMPI_Address(void *location, MPI_Aint * address)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (!address) {
@@ -160,7 +161,7 @@ int PMPI_Get_address(void *location, MPI_Aint * address)
 
 int PMPI_Type_free(MPI_Datatype * datatype)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (!datatype) {
@@ -175,7 +176,7 @@ int PMPI_Type_free(MPI_Datatype * datatype)
 
 int PMPI_Type_size(MPI_Datatype datatype, int *size)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (datatype == MPI_DATATYPE_NULL) {
@@ -192,7 +193,7 @@ int PMPI_Type_size(MPI_Datatype datatype, int *size)
 
 int PMPI_Type_get_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (datatype == MPI_DATATYPE_NULL) {
@@ -213,7 +214,7 @@ int PMPI_Type_get_true_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * e
 
 int PMPI_Type_extent(MPI_Datatype datatype, MPI_Aint * extent)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (datatype == MPI_DATATYPE_NULL) {
@@ -230,7 +231,7 @@ int PMPI_Type_extent(MPI_Datatype datatype, MPI_Aint * extent)
 
 int PMPI_Type_lb(MPI_Datatype datatype, MPI_Aint * disp)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (datatype == MPI_DATATYPE_NULL) {
@@ -247,7 +248,7 @@ int PMPI_Type_lb(MPI_Datatype datatype, MPI_Aint * disp)
 
 int PMPI_Type_ub(MPI_Datatype datatype, MPI_Aint * disp)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (datatype == MPI_DATATYPE_NULL) {
@@ -264,7 +265,7 @@ int PMPI_Type_ub(MPI_Datatype datatype, MPI_Aint * disp)
 
 int PMPI_Op_create(MPI_User_function * function, int commute, MPI_Op * op)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (function == NULL || op == NULL) {
@@ -279,7 +280,7 @@ int PMPI_Op_create(MPI_User_function * function, int commute, MPI_Op * op)
 
 int PMPI_Op_free(MPI_Op * op)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (op == NULL) {
@@ -297,13 +298,12 @@ int PMPI_Op_free(MPI_Op * op)
 
 int PMPI_Group_free(MPI_Group * group)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (group == NULL) {
     retval = MPI_ERR_ARG;
   } else {
-    if(*group!= smpi_comm_group(MPI_COMM_WORLD))// do not free the group of the comm_world
     smpi_group_destroy(*group);
     *group = MPI_GROUP_NULL;
     retval = MPI_SUCCESS;
@@ -314,7 +314,7 @@ int PMPI_Group_free(MPI_Group * group)
 
 int PMPI_Group_size(MPI_Group group, int *size)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (group == MPI_GROUP_NULL) {
@@ -331,7 +331,7 @@ int PMPI_Group_size(MPI_Group group, int *size)
 
 int PMPI_Group_rank(MPI_Group group, int *rank)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (group == MPI_GROUP_NULL) {
@@ -370,7 +370,7 @@ int PMPI_Group_translate_ranks(MPI_Group group1, int n, int *ranks1,
 
 int PMPI_Group_compare(MPI_Group group1, MPI_Group group2, int *result)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (group1 == MPI_GROUP_NULL || group2 == MPI_GROUP_NULL) {
@@ -419,7 +419,6 @@ int PMPI_Group_union(MPI_Group group1, MPI_Group group2,
         smpi_group_set_mapping(*newgroup, proc2, i);
       }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -459,7 +458,6 @@ int PMPI_Group_intersection(MPI_Group group1, MPI_Group group2,
         }
       }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -496,7 +494,6 @@ int PMPI_Group_difference(MPI_Group group1, MPI_Group group2, MPI_Group * newgro
         }
       }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -517,6 +514,11 @@ int PMPI_Group_incl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
       *newgroup = MPI_GROUP_EMPTY;
     } else if (n == smpi_group_size(group)) {
       *newgroup = group;
+      if(group!= smpi_comm_group(MPI_COMM_WORLD)
+                && group != MPI_GROUP_NULL
+                && group != smpi_comm_group(MPI_COMM_SELF)
+                && group != MPI_GROUP_EMPTY)
+      smpi_group_use(group);
     } else {
       *newgroup = smpi_group_new(n);
       for (i = 0; i < n; i++) {
@@ -524,7 +526,6 @@ int PMPI_Group_incl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
         smpi_group_set_mapping(*newgroup, index, i);
       }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -543,6 +544,11 @@ int PMPI_Group_excl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
   } else {
     if (n == 0) {
       *newgroup = group;
+      if(group!= smpi_comm_group(MPI_COMM_WORLD)
+                && group != MPI_GROUP_NULL
+                && group != smpi_comm_group(MPI_COMM_SELF)
+                && group != MPI_GROUP_EMPTY)
+      smpi_group_use(group);
     } else if (n == smpi_group_size(group)) {
       *newgroup = MPI_GROUP_EMPTY;
     } else {
@@ -567,7 +573,6 @@ int PMPI_Group_excl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
 
       xbt_free(to_exclude);
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -626,7 +631,6 @@ int PMPI_Group_range_incl(MPI_Group group, int n, int ranges[][3],
         }
       }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -646,6 +650,11 @@ int PMPI_Group_range_excl(MPI_Group group, int n, int ranges[][3],
   } else {
     if (n == 0) {
       *newgroup = group;
+      if(group!= smpi_comm_group(MPI_COMM_WORLD)
+                && group != MPI_GROUP_NULL
+                && group != smpi_comm_group(MPI_COMM_SELF)
+                && group != MPI_GROUP_EMPTY)
+      smpi_group_use(group);
     } else {
       size = smpi_group_size(group);
       for (i = 0; i < n; i++) {
@@ -699,7 +708,6 @@ int PMPI_Group_range_excl(MPI_Group group, int n, int ranges[][3],
         }
       }
     }
-    smpi_group_use(*newgroup);
 
     retval = MPI_SUCCESS;
   }
@@ -709,7 +717,7 @@ int PMPI_Group_range_excl(MPI_Group group, int n, int ranges[][3],
 
 int PMPI_Comm_rank(MPI_Comm comm, int *rank)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (comm == MPI_COMM_NULL) {
@@ -726,7 +734,7 @@ int PMPI_Comm_rank(MPI_Comm comm, int *rank)
 
 int PMPI_Comm_size(MPI_Comm comm, int *size)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (comm == MPI_COMM_NULL) {
@@ -743,7 +751,7 @@ int PMPI_Comm_size(MPI_Comm comm, int *size)
 
 int PMPI_Comm_get_name (MPI_Comm comm, char* name, int* len)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (comm == MPI_COMM_NULL)  {
@@ -760,7 +768,7 @@ int PMPI_Comm_get_name (MPI_Comm comm, char* name, int* len)
 
 int PMPI_Comm_group(MPI_Comm comm, MPI_Group * group)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (comm == MPI_COMM_NULL) {
@@ -769,6 +777,11 @@ int PMPI_Comm_group(MPI_Comm comm, MPI_Group * group)
     retval = MPI_ERR_ARG;
   } else {
     *group = smpi_comm_group(comm);
+    if(*group!= smpi_comm_group(MPI_COMM_WORLD)
+              && *group != MPI_GROUP_NULL
+              && *group != smpi_comm_group(MPI_COMM_SELF)
+              && *group != MPI_GROUP_EMPTY)
+    smpi_group_use(*group);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -777,7 +790,7 @@ int PMPI_Comm_group(MPI_Comm comm, MPI_Group * group)
 
 int PMPI_Comm_compare(MPI_Comm comm1, MPI_Comm comm2, int *result)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (comm1 == MPI_COMM_NULL || comm2 == MPI_COMM_NULL) {
@@ -803,7 +816,7 @@ int PMPI_Comm_compare(MPI_Comm comm1, MPI_Comm comm2, int *result)
 
 int PMPI_Comm_dup(MPI_Comm comm, MPI_Comm * newcomm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (comm == MPI_COMM_NULL) {
@@ -820,7 +833,7 @@ int PMPI_Comm_dup(MPI_Comm comm, MPI_Comm * newcomm)
 
 int PMPI_Comm_create(MPI_Comm comm, MPI_Group group, MPI_Comm * newcomm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (comm == MPI_COMM_NULL) {
@@ -843,7 +856,7 @@ int PMPI_Comm_create(MPI_Comm comm, MPI_Group group, MPI_Comm * newcomm)
 
 int PMPI_Comm_free(MPI_Comm * comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (comm == NULL) {
@@ -862,7 +875,7 @@ int PMPI_Comm_free(MPI_Comm * comm)
 int PMPI_Comm_disconnect(MPI_Comm * comm)
 {
   /* TODO: wait until all communication in comm are done */
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (comm == NULL) {
@@ -880,7 +893,7 @@ int PMPI_Comm_disconnect(MPI_Comm * comm)
 
 int PMPI_Comm_split(MPI_Comm comm, int color, int key, MPI_Comm* comm_out)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (comm_out == NULL) {
@@ -898,7 +911,7 @@ int PMPI_Comm_split(MPI_Comm comm, int color, int key, MPI_Comm* comm_out)
 int PMPI_Send_init(void *buf, int count, MPI_Datatype datatype, int dst,
                   int tag, MPI_Comm comm, MPI_Request * request)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (request == NULL) {
@@ -912,13 +925,14 @@ int PMPI_Send_init(void *buf, int count, MPI_Datatype datatype, int dst,
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
+  if(retval!=MPI_SUCCESS)*request=MPI_REQUEST_NULL;
   return retval;
 }
 
 int PMPI_Recv_init(void *buf, int count, MPI_Datatype datatype, int src,
                   int tag, MPI_Comm comm, MPI_Request * request)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (request == NULL) {
@@ -932,11 +946,12 @@ int PMPI_Recv_init(void *buf, int count, MPI_Datatype datatype, int src,
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
+  if(retval!=MPI_SUCCESS)*request=MPI_REQUEST_NULL;
   return retval;
 }
 
 int PMPI_Ssend_init(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm, MPI_Request* request) {
-  int retval;
+  int retval = 0;
 
     smpi_bench_end();
     if (request == NULL) {
@@ -950,12 +965,13 @@ int PMPI_Ssend_init(void* buf, int count, MPI_Datatype datatype, int dst, int ta
       retval = MPI_SUCCESS;
     }
     smpi_bench_begin();
+    if(retval!=MPI_SUCCESS)*request=MPI_REQUEST_NULL;
     return retval;
 }
 
 int PMPI_Start(MPI_Request * request)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (request == NULL || *request == MPI_REQUEST_NULL) {
@@ -970,7 +986,7 @@ int PMPI_Start(MPI_Request * request)
 
 int PMPI_Startall(int count, MPI_Request * requests)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (requests == NULL) {
@@ -985,7 +1001,7 @@ int PMPI_Startall(int count, MPI_Request * requests)
 
 int PMPI_Request_free(MPI_Request * request)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (*request == MPI_REQUEST_NULL) {
@@ -1002,7 +1018,7 @@ int PMPI_Request_free(MPI_Request * request)
 int PMPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src,
               int tag, MPI_Comm comm, MPI_Request * request)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 
@@ -1041,6 +1057,7 @@ int PMPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src,
   }
 
   smpi_bench_begin();
+  if(retval!=MPI_SUCCESS)*request=MPI_REQUEST_NULL;
   return retval;
 }
 
@@ -1048,7 +1065,7 @@ int PMPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src,
 int PMPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst,
               int tag, MPI_Comm comm, MPI_Request * request)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (request == NULL) {
@@ -1089,11 +1106,12 @@ int PMPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst,
   }
 
   smpi_bench_begin();
+  if(retval!=MPI_SUCCESS)*request=MPI_REQUEST_NULL;
   return retval;
 }
 
 int PMPI_Issend(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm, MPI_Request* request) {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (request == NULL) {
@@ -1134,13 +1152,14 @@ int PMPI_Issend(void* buf, int count, MPI_Datatype datatype, int dst, int tag, M
   }
 
   smpi_bench_begin();
+  if(retval!=MPI_SUCCESS)*request=MPI_REQUEST_NULL;
   return retval;
 }
 
 int PMPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag,
              MPI_Comm comm, MPI_Status * status)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (comm == MPI_COMM_NULL) {
@@ -1187,7 +1206,7 @@ int PMPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag,
 int PMPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag,
              MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 
@@ -1231,7 +1250,7 @@ int PMPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag,
 
 
 int PMPI_Ssend(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) {
-  int retval;
+  int retval = 0;
 
    smpi_bench_end();
 
@@ -1277,7 +1296,7 @@ int PMPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
                  MPI_Datatype recvtype, int src, int recvtag,
                  MPI_Comm comm, MPI_Status * status)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 
@@ -1333,7 +1352,7 @@ int PMPI_Sendrecv_replace(void *buf, int count, MPI_Datatype datatype,
 {
   //TODO: suboptimal implementation
   void *recvbuf;
-  int retval;
+  int retval = 0;
   if (datatype == MPI_DATATYPE_NULL) {
       retval = MPI_ERR_TYPE;
   } else if (count < 0) {
@@ -1355,7 +1374,7 @@ int PMPI_Sendrecv_replace(void *buf, int count, MPI_Datatype datatype,
 
 int PMPI_Test(MPI_Request * request, int *flag, MPI_Status * status)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (request == MPI_REQUEST_NULL || flag == NULL) {
@@ -1374,7 +1393,7 @@ int PMPI_Test(MPI_Request * request, int *flag, MPI_Status * status)
 int PMPI_Testany(int count, MPI_Request requests[], int *index, int *flag,
                 MPI_Status * status)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (index == NULL || flag == NULL) {
@@ -1389,7 +1408,7 @@ int PMPI_Testany(int count, MPI_Request requests[], int *index, int *flag,
 
 int PMPI_Testall(int count, MPI_Request* requests, int* flag, MPI_Status* statuses)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (flag == NULL) {
@@ -1403,7 +1422,7 @@ int PMPI_Testall(int count, MPI_Request* requests, int* flag, MPI_Status* status
 }
 
 int PMPI_Probe(int source, int tag, MPI_Comm comm, MPI_Status* status) {
-  int retval;
+  int retval = 0;
   smpi_bench_end();
 
   if (status == NULL) {
@@ -1424,7 +1443,7 @@ int PMPI_Probe(int source, int tag, MPI_Comm comm, MPI_Status* status) {
 
 
 int PMPI_Iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status) {
-  int retval;
+  int retval = 0;
   smpi_bench_end();
 
   if (flag == NULL) {
@@ -1434,6 +1453,7 @@ int PMPI_Iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* statu
   } else if (comm == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;
   } else if (source == MPI_PROC_NULL) {
+    *flag=TRUE;
     smpi_empty_status(status);
     status->MPI_SOURCE = MPI_PROC_NULL;
     retval = MPI_SUCCESS;
@@ -1447,7 +1467,7 @@ int PMPI_Iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* statu
 
 int PMPI_Wait(MPI_Request * request, MPI_Status * status)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 
@@ -1465,6 +1485,7 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
 
   int src_traced = (*request)->src;
   int dst_traced = (*request)->dst;
+  MPI_Comm comm = (*request)->comm;
   int is_wait_for_receive = (*request)->recv;
   TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__);
 #endif
@@ -1473,8 +1494,13 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
     retval = MPI_SUCCESS;
 
 #ifdef HAVE_TRACING
+  //the src may not have been known at the beginning of the recv (MPI_ANY_SOURCE)
   TRACE_smpi_ptp_out(rank, src_traced, dst_traced, __FUNCTION__);
   if (is_wait_for_receive) {
+    if(src_traced==MPI_ANY_SOURCE)
+    src_traced = (status!=MPI_STATUS_IGNORE) ?
+                                smpi_group_rank(smpi_comm_group(comm), status->MPI_SOURCE) :
+                                src_traced;
     TRACE_smpi_recv(rank, src_traced, dst_traced);
   }
   TRACE_smpi_computing_in(rank);
@@ -1488,7 +1514,7 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
 
 int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * status)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -1497,12 +1523,15 @@ int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * sta
   int *srcs = xbt_new(int, count);
   int *dsts = xbt_new(int, count);
   int *recvs = xbt_new(int, count);
+  MPI_Comm *comms = xbt_new(MPI_Comm, count);
+
   for (i = 0; i < count; i++) {
     MPI_Request req = requests[i];      //already received requests are no longer valid
     if (req) {
       srcs[i] = req->src;
       dsts[i] = req->dst;
       recvs[i] = req->recv;
+      comms[i] = req->comm;
     }
   }
   int rank_traced = smpi_process_index();
@@ -1520,15 +1549,22 @@ int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * sta
 #ifdef HAVE_TRACING
   if(*index!=MPI_UNDEFINED){
     int src_traced = srcs[*index];
+    //the src may not have been known at the beginning of the recv (MPI_ANY_SOURCE)
     int dst_traced = dsts[*index];
     int is_wait_for_receive = recvs[*index];
     if (is_wait_for_receive) {
+      if(srcs[*index]==MPI_ANY_SOURCE)
+        src_traced = (status!=MPI_STATUSES_IGNORE) ?
+                      smpi_group_rank(smpi_comm_group(comms[*index]), status->MPI_SOURCE) :
+                      srcs[*index];
       TRACE_smpi_recv(rank_traced, src_traced, dst_traced);
     }
     TRACE_smpi_ptp_out(rank_traced, src_traced, dst_traced, __FUNCTION__);
     xbt_free(srcs);
     xbt_free(dsts);
     xbt_free(recvs);
+    xbt_free(comms);
+
   }
   TRACE_smpi_computing_in(rank_traced);
 #endif
@@ -1546,14 +1582,20 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
   int *srcs = xbt_new(int, count);
   int *dsts = xbt_new(int, count);
   int *recvs = xbt_new(int, count);
-  int valid_count = 0;
+  int *valid = xbt_new(int, count);
+  MPI_Comm *comms = xbt_new(MPI_Comm, count);
+
+  //int valid_count = 0;
   for (i = 0; i < count; i++) {
     MPI_Request req = requests[i];
     if(req!=MPI_REQUEST_NULL){
-      srcs[valid_count] = req->src;
-      dsts[valid_count] = req->dst;
-      recvs[valid_count] = req->recv;
-      valid_count++;
+      srcs[i] = req->src;
+      dsts[i] = req->dst;
+      recvs[i] = req->recv;
+      comms[i] = req->comm;
+      valid[i]=1;;
+    }else{
+      valid[i]=0;
     }
   }
   int rank_traced = smpi_process_index();
@@ -1563,18 +1605,29 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
 #endif
   int retval = smpi_mpi_waitall(count, requests, status);
 #ifdef HAVE_TRACING
-  for (i = 0; i < valid_count; i++) {
-    int src_traced = srcs[i];
-    int dst_traced = dsts[i];
-    int is_wait_for_receive = recvs[i];
-    if (is_wait_for_receive) {
-      TRACE_smpi_recv(rank_traced, src_traced, dst_traced);
+  for (i = 0; i < count; i++) {
+    if(valid[i]){
+    //int src_traced = srcs[*index];
+    //the src may not have been known at the beginning of the recv (MPI_ANY_SOURCE)
+      int src_traced = srcs[i];
+      int dst_traced = dsts[i];
+      int is_wait_for_receive = recvs[i];
+      if (is_wait_for_receive) {
+        if(src_traced==MPI_ANY_SOURCE)
+        src_traced = (status!=MPI_STATUSES_IGNORE) ?
+                          smpi_group_rank(smpi_comm_group(comms[i]), status[i].MPI_SOURCE) :
+                          srcs[i];
+        TRACE_smpi_recv(rank_traced, src_traced, dst_traced);
+      }
     }
   }
   TRACE_smpi_ptp_out(rank_traced, -1, -1, __FUNCTION__);
   xbt_free(srcs);
   xbt_free(dsts);
   xbt_free(recvs);
+  xbt_free(valid);
+  xbt_free(comms);
+
   TRACE_smpi_computing_in(rank_traced);
 #endif
   smpi_bench_begin();
@@ -1584,7 +1637,7 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
 int PMPI_Waitsome(int incount, MPI_Request requests[], int *outcount,
                  int *indices, MPI_Status status[])
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (outcount == NULL) {
@@ -1600,7 +1653,7 @@ int PMPI_Waitsome(int incount, MPI_Request requests[], int *outcount,
 int PMPI_Testsome(int incount, MPI_Request requests[], int* outcount,
                  int* indices, MPI_Status status[])
 {
-  int retval;
+  int retval = 0;
 
    smpi_bench_end();
    if (outcount == NULL) {
@@ -1616,7 +1669,7 @@ int PMPI_Testsome(int incount, MPI_Request requests[], int* outcount,
 
 int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -1641,7 +1694,7 @@ int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm c
 
 int PMPI_Barrier(MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -1667,7 +1720,7 @@ int PMPI_Gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
                void *recvbuf, int recvcount, MPI_Datatype recvtype,
                int root, MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -1712,7 +1765,7 @@ int PMPI_Gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
                 void *recvbuf, int *recvcounts, int *displs,
                 MPI_Datatype recvtype, int root, MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -1756,7 +1809,7 @@ int PMPI_Allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
                   void *recvbuf, int recvcount, MPI_Datatype recvtype,
                   MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -1795,7 +1848,7 @@ int PMPI_Allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
                    void *recvbuf, int *recvcounts, int *displs,
                    MPI_Datatype recvtype, MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -1836,7 +1889,7 @@ int PMPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
                 void *recvbuf, int recvcount, MPI_Datatype recvtype,
                 int root, MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -1852,9 +1905,9 @@ int PMPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
              || ((recvbuf !=MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL))) {
     retval = MPI_ERR_TYPE;
   } else {
-
-    if(recvbuf==MPI_IN_PLACE){
-       recvcount=0;
+    if (recvbuf == MPI_IN_PLACE) {
+        recvtype=sendtype;
+        recvcount=sendcount;
     }
     mpi_coll_scatter_fun(sendbuf, sendcount, sendtype, recvbuf, recvcount,
                      recvtype, root, comm);
@@ -1872,7 +1925,7 @@ int PMPI_Scatterv(void *sendbuf, int *sendcounts, int *displs,
                  MPI_Datatype sendtype, void *recvbuf, int recvcount,
                  MPI_Datatype recvtype, int root, MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -1889,11 +1942,10 @@ int PMPI_Scatterv(void *sendbuf, int *sendcounts, int *displs,
              || ((recvbuf !=MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL))) {
     retval = MPI_ERR_TYPE;
   } else {
-
-    if(recvbuf==MPI_IN_PLACE){
-       recvcount=0;
+    if (recvbuf == MPI_IN_PLACE) {
+        recvtype=sendtype;
+        recvcount=sendcounts[smpi_comm_rank(comm)];
     }
-
     smpi_mpi_scatterv(sendbuf, sendcounts, displs, sendtype, recvbuf,
                       recvcount, recvtype, root, comm);
     retval = MPI_SUCCESS;
@@ -1909,7 +1961,7 @@ int PMPI_Scatterv(void *sendbuf, int *sendcounts, int *displs,
 int PMPI_Reduce(void *sendbuf, void *recvbuf, int count,
                MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -1924,17 +1976,7 @@ int PMPI_Reduce(void *sendbuf, void *recvbuf, int count,
     retval = MPI_ERR_ARG;
   } else {
 
-    char* sendtmpbuf = (char*) sendbuf;
-    if( sendbuf == MPI_IN_PLACE ) {
-      sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
-      smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
-    }
-
-    mpi_coll_reduce_fun(sendtmpbuf, recvbuf, count, datatype, op, root, comm);
-
-    if( sendbuf == MPI_IN_PLACE ) {
-      xbt_free(sendtmpbuf);
-    }
+    mpi_coll_reduce_fun(sendbuf, recvbuf, count, datatype, op, root, comm);
 
     retval = MPI_SUCCESS;
   }
@@ -1948,7 +1990,7 @@ int PMPI_Reduce(void *sendbuf, void *recvbuf, int count,
 
 int PMPI_Reduce_local(void *inbuf, void *inoutbuf, int count,
     MPI_Datatype datatype, MPI_Op op){
-  int retval;
+  int retval = 0;
 
     smpi_bench_end();
     if (datatype == MPI_DATATYPE_NULL || op == MPI_OP_NULL) {
@@ -1964,7 +2006,7 @@ int PMPI_Reduce_local(void *inbuf, void *inoutbuf, int count,
 int PMPI_Allreduce(void *sendbuf, void *recvbuf, int count,
                   MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -2006,7 +2048,7 @@ int PMPI_Allreduce(void *sendbuf, void *recvbuf, int count,
 int PMPI_Scan(void *sendbuf, void *recvbuf, int count,
              MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -2034,7 +2076,7 @@ int PMPI_Scan(void *sendbuf, void *recvbuf, int count,
 
 int PMPI_Exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
                 MPI_Op op, MPI_Comm comm){
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -2063,7 +2105,7 @@ int PMPI_Exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
 int PMPI_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
                        MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
   smpi_bench_end();
 #ifdef HAVE_TRACING
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
@@ -2135,7 +2177,7 @@ int PMPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype,
                  void *recvbuf, int recvcount, MPI_Datatype recvtype,
                  MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -2163,7 +2205,7 @@ int PMPI_Alltoallv(void *sendbuf, int *sendcounts, int *senddisps,
                   MPI_Datatype sendtype, void *recvbuf, int *recvcounts,
                   int *recvdisps, MPI_Datatype recvtype, MPI_Comm comm)
 {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
 #ifdef HAVE_TRACING
@@ -2236,7 +2278,7 @@ int PMPI_Get_count(MPI_Status * status, MPI_Datatype datatype, int *count)
 }
 
 int PMPI_Type_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_type) {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
@@ -2251,7 +2293,7 @@ int PMPI_Type_contiguous(int count, MPI_Datatype old_type, MPI_Datatype* new_typ
 }
 
 int PMPI_Type_commit(MPI_Datatype* datatype) {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (datatype == MPI_DATATYPE_NULL) {
@@ -2266,7 +2308,7 @@ int PMPI_Type_commit(MPI_Datatype* datatype) {
 
 
 int PMPI_Type_vector(int count, int blocklen, int stride, MPI_Datatype old_type, MPI_Datatype* new_type) {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
@@ -2281,7 +2323,7 @@ int PMPI_Type_vector(int count, int blocklen, int stride, MPI_Datatype old_type,
 }
 
 int PMPI_Type_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type) {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
@@ -2300,7 +2342,22 @@ int PMPI_Type_create_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datat
 }
 
 int PMPI_Type_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
-  int retval;
+  int retval = 0;
+
+  smpi_bench_end();
+  if (old_type == MPI_DATATYPE_NULL) {
+    retval = MPI_ERR_TYPE;
+  } else if (count<0){
+    retval = MPI_ERR_COUNT;
+  } else {
+    retval = smpi_datatype_indexed(count, blocklens, indices, old_type, new_type);
+  }
+  smpi_bench_begin();
+  return retval;
+}
+
+int PMPI_Type_create_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
+  int retval = 0;
 
   smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
@@ -2334,7 +2391,7 @@ int PMPI_Type_create_indexed_block(int count, int blocklength, int* indices, MPI
 
 
 int PMPI_Type_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (old_type == MPI_DATATYPE_NULL) {
@@ -2348,6 +2405,10 @@ int PMPI_Type_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatyp
   return retval;
 }
 
+int PMPI_Type_create_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
+  return PMPI_Type_hindexed(count, blocklens,indices,old_type,new_type);
+}
+
 int PMPI_Type_create_hindexed_block(int count, int blocklength, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
   int retval,i;
 
@@ -2368,7 +2429,7 @@ int PMPI_Type_create_hindexed_block(int count, int blocklength, MPI_Aint* indice
 
 
 int PMPI_Type_struct(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype* old_types, MPI_Datatype* new_type) {
-  int retval;
+  int retval = 0;
 
   smpi_bench_end();
   if (count<0){
@@ -2393,7 +2454,7 @@ int PMPI_Error_class(int errorcode, int* errorclass) {
 
 
 int PMPI_Initialized(int* flag) {
-   *flag=(smpi_process_data()!=NULL);
+   *flag=smpi_process_initialized();
    return MPI_SUCCESS;
 }
 
@@ -2508,6 +2569,10 @@ int PMPI_Comm_set_errhandler(MPI_Comm comm, MPI_Errhandler errhandler) {
    NOT_YET_IMPLEMENTED
 }
 
+int PMPI_Comm_get_errhandler(MPI_Comm comm, MPI_Errhandler* errhandler) {
+   NOT_YET_IMPLEMENTED
+}
+
 int PMPI_Cancel(MPI_Request* request) {
    NOT_YET_IMPLEMENTED
 }