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

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

diff --git a/src/smpi/smpi_pmpi.cpp b/src/smpi/smpi_pmpi.cpp
index eb4e490..0cb559c 100644 (file)
--- a/src/smpi/smpi_pmpi.cpp
+++ b/src/smpi/smpi_pmpi.cpp
@@ -1,10 +1,9 @@
-
-/* Copyright (c) 2007-2015. The SimGrid Team.
- * All rights reserved.                                                     */
+/* Copyright (c) 2007-2016. The SimGrid Team. All rights reserved.          */

 
 /* This program is free software; you can redistribute it and/or modify it
  * under the terms of the license (GNU LGPL) which comes with this package. */
 
 
 /* This program is free software; you can redistribute it and/or modify it
  * under the terms of the license (GNU LGPL) which comes with this package. */
 

+#include <simgrid/s4u/host.hpp>

 #include <xbt/ex.hpp>
 
 #include "private.h"
 #include <xbt/ex.hpp>
 
 #include "private.h"


@@ -23,6 +22,7 @@ void TRACE_smpi_set_category(const char *category)
 }
 
 /* PMPI User level calls */
 }
 
 /* PMPI User level calls */

+extern "C" { // Obviously, the C MPI interface should use the C linkage

 
 int PMPI_Init(int *argc, char ***argv)
 {
 
 int PMPI_Init(int *argc, char ***argv)
 {


@@ -47,7 +47,7 @@ int PMPI_Init(int *argc, char ***argv)
   return MPI_SUCCESS;
 }
 
   return MPI_SUCCESS;
 }
 

-int PMPI_Finalize(void)
+int PMPI_Finalize()

 {
   smpi_bench_end();
   int rank = smpi_process_index();
 {
   smpi_bench_end();
   int rank = smpi_process_index();


@@ -128,13 +128,13 @@ int PMPI_Abort(MPI_Comm comm, int errorcode)
   return MPI_SUCCESS;
 }
 
   return MPI_SUCCESS;
 }
 

-double PMPI_Wtime(void)
+double PMPI_Wtime()

 {
   return smpi_mpi_wtime();
 }
 
 extern double sg_maxmin_precision;
 {
   return smpi_mpi_wtime();
 }
 
 extern double sg_maxmin_precision;

-double PMPI_Wtick(void)
+double PMPI_Wtick()

 {
   return sg_maxmin_precision;
 }
 {
   return sg_maxmin_precision;
 }


@@ -185,6 +185,21 @@ int PMPI_Type_size(MPI_Datatype datatype, int *size)
   return retval;
 }
 
   return retval;
 }
 

+int PMPI_Type_size_x(MPI_Datatype datatype, MPI_Count *size)
+{
+  int retval = 0;
+
+  if (datatype == MPI_DATATYPE_NULL) {
+    retval = MPI_ERR_TYPE;
+  } else if (size == nullptr) {
+    retval = MPI_ERR_ARG;
+  } else {
+    *size = static_cast<MPI_Count>(smpi_datatype_size(datatype));
+    retval = MPI_SUCCESS;
+  }
+  return retval;
+}
+

 int PMPI_Type_get_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent)
 {
   int retval = 0;
 int PMPI_Type_get_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent)
 {
   int retval = 0;


@@ -500,7 +515,7 @@ int PMPI_Group_excl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
   } else {
     if (n == 0) {
       *newgroup = group;
   } else {
     if (n == 0) {
       *newgroup = group;

-      if(group!= smpi_comm_group(MPI_COMM_WORLD) && group != MPI_GROUP_NULL
+      if (group != smpi_comm_group(MPI_COMM_WORLD)

                 && group != smpi_comm_group(MPI_COMM_SELF) && group != MPI_GROUP_EMPTY)
       smpi_group_use(group);
     } else if (n == smpi_group_size(group)) {
                 && group != smpi_comm_group(MPI_COMM_SELF) && group != MPI_GROUP_EMPTY)
       smpi_group_use(group);
     } else if (n == smpi_group_size(group)) {


@@ -603,7 +618,7 @@ int PMPI_Group_range_excl(MPI_Group group, int n, int ranges[][3], MPI_Group * n
   } else {
     if (n == 0) {
       *newgroup = group;
   } else {
     if (n == 0) {
       *newgroup = group;

-      if(group!= smpi_comm_group(MPI_COMM_WORLD) && group != MPI_GROUP_NULL
+      if (group!= smpi_comm_group(MPI_COMM_WORLD)

                 && group != smpi_comm_group(MPI_COMM_SELF) && group != MPI_GROUP_EMPTY)
       smpi_group_use(group);
     } else {
                 && group != smpi_comm_group(MPI_COMM_SELF) && group != MPI_GROUP_EMPTY)
       smpi_group_use(group);
     } else {


@@ -839,6 +854,23 @@ int PMPI_Comm_split(MPI_Comm comm, int color, int key, MPI_Comm* comm_out)
   return retval;
 }
 
   return retval;
 }
 

+int PMPI_Comm_create_group(MPI_Comm comm, MPI_Group group, int, MPI_Comm* comm_out)
+{
+  int retval = 0;
+  smpi_bench_end();
+
+  if (comm_out == nullptr) {
+    retval = MPI_ERR_ARG;
+  } else if (comm == MPI_COMM_NULL) {
+    retval = MPI_ERR_COMM;
+  } else {
+    retval = MPI_Comm_create(comm, group, comm_out);
+  }
+  smpi_bench_begin();
+
+  return retval;
+}
+

 int PMPI_Send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm, MPI_Request * request)
 {
   int retval = 0;
 int PMPI_Send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm, MPI_Request * request)
 {
   int retval = 0;


@@ -1047,7 +1079,7 @@ int PMPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MP
       dt_size_send = smpi_datatype_size(datatype);
     extra->send_size = count*dt_size_send;
     TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);
       dt_size_send = smpi_datatype_size(datatype);
     extra->send_size = count*dt_size_send;
     TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);

-    TRACE_smpi_send(rank, rank, dst_traced, count*smpi_datatype_size(datatype));
+    TRACE_smpi_send(rank, rank, dst_traced, tag, count*smpi_datatype_size(datatype));

 
     *request = smpi_mpi_isend(buf, count, datatype, dst, tag, comm);
     retval = MPI_SUCCESS;
 
     *request = smpi_mpi_isend(buf, count, datatype, dst, tag, comm);
     retval = MPI_SUCCESS;


@@ -1097,7 +1129,7 @@ int PMPI_Issend(void* buf, int count, MPI_Datatype datatype, int dst, int tag, M
       dt_size_send = smpi_datatype_size(datatype);
     extra->send_size = count*dt_size_send;
     TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);
       dt_size_send = smpi_datatype_size(datatype);
     extra->send_size = count*dt_size_send;
     TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);

-    TRACE_smpi_send(rank, rank, dst_traced, count*smpi_datatype_size(datatype));
+    TRACE_smpi_send(rank, rank, dst_traced, tag, count*smpi_datatype_size(datatype));

 
     *request = smpi_mpi_issend(buf, count, datatype, dst, tag, comm);
     retval = MPI_SUCCESS;
 
     *request = smpi_mpi_issend(buf, count, datatype, dst, tag, comm);
     retval = MPI_SUCCESS;


@@ -1153,7 +1185,7 @@ int PMPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI
   if(status!=MPI_STATUS_IGNORE){
     src_traced = smpi_group_index(smpi_comm_group(comm), status->MPI_SOURCE);
     if (!TRACE_smpi_view_internals()) {
   if(status!=MPI_STATUS_IGNORE){
     src_traced = smpi_group_index(smpi_comm_group(comm), status->MPI_SOURCE);
     if (!TRACE_smpi_view_internals()) {

-      TRACE_smpi_recv(rank, src_traced, rank);
+      TRACE_smpi_recv(rank, src_traced, rank, tag);

     }
   }
   TRACE_smpi_ptp_out(rank, src_traced, rank, __FUNCTION__);
     }
   }
   TRACE_smpi_ptp_out(rank, src_traced, rank, __FUNCTION__);


@@ -1197,7 +1229,7 @@ int PMPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI
     extra->send_size = count*dt_size_send;
     TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);
     if (!TRACE_smpi_view_internals()) {
     extra->send_size = count*dt_size_send;
     TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);
     if (!TRACE_smpi_view_internals()) {

-      TRACE_smpi_send(rank, rank, dst_traced,count*smpi_datatype_size(datatype));
+      TRACE_smpi_send(rank, rank, dst_traced, tag,count*smpi_datatype_size(datatype));

     }
 
     smpi_mpi_send(buf, count, datatype, dst, tag, comm);
     }
 
     smpi_mpi_send(buf, count, datatype, dst, tag, comm);


@@ -1242,7 +1274,7 @@ int PMPI_Ssend(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MP
     }
     extra->send_size = count*dt_size_send;
     TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);
     }
     extra->send_size = count*dt_size_send;
     TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, extra);

-    TRACE_smpi_send(rank, rank, dst_traced,count*smpi_datatype_size(datatype));
+    TRACE_smpi_send(rank, rank, dst_traced, tag,count*smpi_datatype_size(datatype));

   
     smpi_mpi_ssend(buf, count, datatype, dst, tag, comm);
     retval = MPI_SUCCESS;
   
     smpi_mpi_ssend(buf, count, datatype, dst, tag, comm);
     retval = MPI_SUCCESS;


@@ -1300,14 +1332,14 @@ int PMPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, int dst,
   extra->recv_size = recvcount*dt_size_recv;
 
   TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__, extra);
   extra->recv_size = recvcount*dt_size_recv;
 
   TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__, extra);

-  TRACE_smpi_send(rank, rank, dst_traced,sendcount*smpi_datatype_size(sendtype));
+  TRACE_smpi_send(rank, rank, dst_traced, sendtag,sendcount*smpi_datatype_size(sendtype));

 
     smpi_mpi_sendrecv(sendbuf, sendcount, sendtype, dst, sendtag, recvbuf,
                       recvcount, recvtype, src, recvtag, comm, status);
     retval = MPI_SUCCESS;
 
   TRACE_smpi_ptp_out(rank, src_traced, dst_traced, __FUNCTION__);
 
     smpi_mpi_sendrecv(sendbuf, sendcount, sendtype, dst, sendtag, recvbuf,
                       recvcount, recvtype, src, recvtag, comm, status);
     retval = MPI_SUCCESS;
 
   TRACE_smpi_ptp_out(rank, src_traced, dst_traced, __FUNCTION__);

-  TRACE_smpi_recv(rank, src_traced, rank);
+  TRACE_smpi_recv(rank, src_traced, rank, recvtag);

   }
 
   smpi_bench_begin();
   }
 
   smpi_bench_begin();


@@ -1347,7 +1379,7 @@ int PMPI_Test(MPI_Request * request, int *flag, MPI_Status * status)
     smpi_empty_status(status);
     retval = MPI_SUCCESS;
   } else {
     smpi_empty_status(status);
     retval = MPI_SUCCESS;
   } else {

-    int rank = (request!=nullptr && (*request)->comm != MPI_COMM_NULL) ? smpi_process_index() : -1;
+    int rank = ((*request)->comm != MPI_COMM_NULL) ? smpi_process_index() : -1;

 
     instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
     extra->type = TRACING_TEST;
 
     instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
     extra->type = TRACING_TEST;


@@ -1451,6 +1483,7 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
 
     int src_traced = (*request)->src;
     int dst_traced = (*request)->dst;
 
     int src_traced = (*request)->src;
     int dst_traced = (*request)->dst;

+    int tag_traced= (*request)->tag;

     MPI_Comm comm = (*request)->comm;
     int is_wait_for_receive = (*request)->recv;
     instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
     MPI_Comm comm = (*request)->comm;
     int is_wait_for_receive = (*request)->recv;
     instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);


@@ -1467,7 +1500,7 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
         src_traced = (status!=MPI_STATUS_IGNORE) ?
           smpi_group_rank(smpi_comm_group(comm), status->MPI_SOURCE) :
           src_traced;
         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_recv(rank, src_traced, dst_traced, tag_traced);

     }
   }
 
     }
   }
 


@@ -1483,21 +1516,21 @@ int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * sta
   smpi_bench_end();
   //save requests information for tracing
   int i;
   smpi_bench_end();
   //save requests information for tracing
   int i;

-  int *srcs = nullptr, *dsts = nullptr, *recvs = nullptr;
-  MPI_Comm* comms = nullptr;
+  typedef struct {
+    int src;
+    int dst;
+    int recv;
+    int tag;
+    MPI_Comm comm;
+  } savedvalstype;
+  savedvalstype* savedvals=nullptr;

   if(count>0){
   if(count>0){

-    srcs = xbt_new0(int, count);
-    dsts = xbt_new0(int, count);
-    recvs = xbt_new0(int, count);
-    comms = xbt_new0(MPI_Comm, count);
+    savedvals = xbt_new0(savedvalstype, count);

   }
   for (i = 0; i < count; i++) {
     MPI_Request req = requests[i];      //already received requests are no longer valid
     if (req) {
   }
   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;
+      savedvals[i]=(savedvalstype){req->src, req->dst, req->recv, req->tag, req->comm};

     }
   }
   int rank_traced = smpi_process_index();
     }
   }
   int rank_traced = smpi_process_index();


@@ -1509,23 +1542,20 @@ int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * sta
   *index = smpi_mpi_waitany(count, requests, status);
 
   if(*index!=MPI_UNDEFINED){
   *index = smpi_mpi_waitany(count, requests, status);
 
   if(*index!=MPI_UNDEFINED){

-    int src_traced = srcs[*index];
+    int src_traced = savedvals[*index].src;

     //the src may not have been known at the beginning of the recv (MPI_ANY_SOURCE)
     //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];
+    int dst_traced = savedvals[*index].dst;
+    int is_wait_for_receive = savedvals[*index].recv;

     if (is_wait_for_receive) {
     if (is_wait_for_receive) {

-      if(srcs[*index]==MPI_ANY_SOURCE)
+      if(savedvals[*index].src==MPI_ANY_SOURCE)

         src_traced = (status!=MPI_STATUSES_IGNORE) ?
         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);
+                      smpi_group_rank(smpi_comm_group(savedvals[*index].comm), status->MPI_SOURCE) : savedvals[*index].src;
+      TRACE_smpi_recv(rank_traced, src_traced, dst_traced, savedvals[*index].tag);

     }
     TRACE_smpi_ptp_out(rank_traced, src_traced, dst_traced, __FUNCTION__);
     }
     TRACE_smpi_ptp_out(rank_traced, src_traced, dst_traced, __FUNCTION__);

-    xbt_free(srcs);
-    xbt_free(dsts);
-    xbt_free(recvs);
-    xbt_free(comms);
-

   }
   }

+  xbt_free(savedvals);
+

   smpi_bench_begin();
   return MPI_SUCCESS;
 }
   smpi_bench_begin();
   return MPI_SUCCESS;
 }


@@ -1535,22 +1565,22 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
   smpi_bench_end();
   //save information from requests
   int i;
   smpi_bench_end();
   //save information from requests
   int i;

-  int *srcs = xbt_new0(int, count);
-  int *dsts = xbt_new0(int, count);
-  int *recvs = xbt_new0(int, count);
-  int *valid = xbt_new0(int, count);
-  MPI_Comm *comms = xbt_new0(MPI_Comm, count);
+  typedef struct {
+    int src;
+    int dst;
+    int recv;
+    int tag;
+    int valid;
+    MPI_Comm comm;
+  } savedvalstype;
+  savedvalstype* savedvals=xbt_new0(savedvalstype, count);

 
   for (i = 0; i < count; i++) {
     MPI_Request req = requests[i];
     if(req!=MPI_REQUEST_NULL){
 
   for (i = 0; i < count; i++) {
     MPI_Request req = requests[i];
     if(req!=MPI_REQUEST_NULL){

-      srcs[i] = req->src;
-      dsts[i] = req->dst;
-      recvs[i] = req->recv;
-      comms[i] = req->comm;
-      valid[i]=1;;
+      savedvals[i]=(savedvalstype){req->src, req->dst, req->recv, req->tag, 1, req->comm};

     }else{
     }else{

-      valid[i]=0;
+      savedvals[i].valid=0;

     }
   }
   int rank_traced = smpi_process_index();
     }
   }
   int rank_traced = smpi_process_index();


@@ -1562,25 +1592,21 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
   int retval = smpi_mpi_waitall(count, requests, status);
 
   for (i = 0; i < count; i++) {
   int retval = smpi_mpi_waitall(count, requests, status);
 
   for (i = 0; i < count; i++) {

-    if(valid[i]){
+    if(savedvals[i].valid){

     //the src may not have been known at the beginning of the recv (MPI_ANY_SOURCE)
     //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];
+      int src_traced = savedvals[i].src;
+      int dst_traced = savedvals[i].dst;
+      int is_wait_for_receive = savedvals[i].recv;

       if (is_wait_for_receive) {
         if(src_traced==MPI_ANY_SOURCE)
         src_traced = (status!=MPI_STATUSES_IGNORE) ?
       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);
+                          smpi_group_rank(smpi_comm_group(savedvals[i].comm), status[i].MPI_SOURCE) : savedvals[i].src;
+        TRACE_smpi_recv(rank_traced, src_traced, dst_traced,savedvals[i].tag);

       }
     }
   }
   TRACE_smpi_ptp_out(rank_traced, -1, -1, __FUNCTION__);
       }
     }
   }
   TRACE_smpi_ptp_out(rank_traced, -1, -1, __FUNCTION__);

-  xbt_free(srcs);
-  xbt_free(dsts);
-  xbt_free(recvs);
-  xbt_free(valid);
-  xbt_free(comms);
+  xbt_free(savedvals);

 
   smpi_bench_begin();
   return retval;
 
   smpi_bench_begin();
   return retval;


@@ -1661,15 +1687,15 @@ int PMPI_Barrier(MPI_Comm comm)
   if (comm == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;
   } else {
   if (comm == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;
   } else {

-  int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
-  instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
-  extra->type = TRACING_BARRIER;
-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__, extra);
+    int rank               = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
+    instr_extra_data extra = xbt_new0(s_instr_extra_data_t, 1);
+    extra->type            = TRACING_BARRIER;
+    TRACE_smpi_collective_in(rank, -1, __FUNCTION__, extra);

 
 

-  mpi_coll_barrier_fun(comm);
-  retval = MPI_SUCCESS;
+    mpi_coll_barrier_fun(comm);
+    retval = MPI_SUCCESS;

 
 

-  TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+    TRACE_smpi_collective_out(rank, -1, __FUNCTION__);

   }
 
   smpi_bench_begin();
   }
 
   smpi_bench_begin();


@@ -2138,11 +2164,18 @@ int PMPI_Exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
   if(known==0)
     dt_size_send = smpi_datatype_size(datatype);
   extra->send_size = count*dt_size_send;
   if(known==0)
     dt_size_send = smpi_datatype_size(datatype);
   extra->send_size = count*dt_size_send;

+  void * sendtmpbuf = sendbuf;
+  if(sendbuf==MPI_IN_PLACE){
+    sendtmpbuf= static_cast<void*>(xbt_malloc(count*smpi_datatype_size(datatype)));
+    memcpy(sendtmpbuf,recvbuf, count*smpi_datatype_size(datatype));
+  }

   TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 

-  smpi_mpi_exscan(sendbuf, recvbuf, count, datatype, op, comm);
+  smpi_mpi_exscan(sendtmpbuf, recvbuf, count, datatype, op, comm);

     retval = MPI_SUCCESS;
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
     retval = MPI_SUCCESS;
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);

+  if(sendbuf==MPI_IN_PLACE)
+    xbt_free(sendtmpbuf);

   }
 
   smpi_bench_begin();
   }
 
   smpi_bench_begin();


@@ -2176,17 +2209,25 @@ int PMPI_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datat
     dt_size_send = smpi_datatype_size(datatype);
   extra->send_size = 0;
   extra->recvcounts= xbt_new(int, size);
     dt_size_send = smpi_datatype_size(datatype);
   extra->send_size = 0;
   extra->recvcounts= xbt_new(int, size);

-  for(i=0; i< size; i++)//copy data to avoid bad free
+  int totalcount = 0;
+  for(i=0; i< size; i++){//copy data to avoid bad free

     extra->recvcounts[i] = recvcounts[i]*dt_size_send;
     extra->recvcounts[i] = recvcounts[i]*dt_size_send;

-  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
-
+    totalcount+= recvcounts[i];
+  }

   void* sendtmpbuf=sendbuf;
   void* sendtmpbuf=sendbuf;

-    if(sendbuf==MPI_IN_PLACE)
-      sendtmpbuf=recvbuf;
+  if(sendbuf==MPI_IN_PLACE){
+    sendtmpbuf= static_cast<void*>(xbt_malloc(totalcount*smpi_datatype_size(datatype)));
+    memcpy(sendtmpbuf,recvbuf, totalcount*smpi_datatype_size(datatype));
+  }

 
 

-    mpi_coll_reduce_scatter_fun(sendtmpbuf, recvbuf, recvcounts, datatype,  op, comm);
-    retval = MPI_SUCCESS;
+  TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
+
+  mpi_coll_reduce_scatter_fun(sendtmpbuf, recvbuf, recvcounts, datatype,  op, comm);
+  retval = MPI_SUCCESS;

   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);

+
+  if(sendbuf==MPI_IN_PLACE)
+    xbt_free(sendtmpbuf);

   }
 
   smpi_bench_begin();
   }
 
   smpi_bench_begin();


@@ -2208,7 +2249,7 @@ int PMPI_Reduce_scatter_block(void *sendbuf, void *recvbuf, int recvcount,
   } else if (recvcount < 0) {
     retval = MPI_ERR_ARG;
   } else {
   } else if (recvcount < 0) {
     retval = MPI_ERR_ARG;
   } else {

-    int count=smpi_comm_size(comm);
+  int count=smpi_comm_size(comm);

 
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
   instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
 
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
   instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);


@@ -2223,17 +2264,25 @@ int PMPI_Reduce_scatter_block(void *sendbuf, void *recvbuf, int recvcount,
   extra->recvcounts= xbt_new(int, count);
   for(i=0; i< count; i++)//copy data to avoid bad free
     extra->recvcounts[i] = recvcount*dt_size_send;
   extra->recvcounts= xbt_new(int, count);
   for(i=0; i< count; i++)//copy data to avoid bad free
     extra->recvcounts[i] = recvcount*dt_size_send;

+  void* sendtmpbuf=sendbuf;
+  if(sendbuf==MPI_IN_PLACE){
+    sendtmpbuf= static_cast<void*>(xbt_malloc(recvcount*count*smpi_datatype_size(datatype)));
+    memcpy(sendtmpbuf,recvbuf, recvcount*count*smpi_datatype_size(datatype));
+  }

 
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 
 
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 

-  int* recvcounts=static_cast<int*>(xbt_malloc(count));
-    for (i=0; i<count;i++)
-      recvcounts[i]=recvcount;
-    mpi_coll_reduce_scatter_fun(sendbuf, recvbuf, recvcounts, datatype,  op, comm);
-    xbt_free(recvcounts);
-    retval = MPI_SUCCESS;
+  int* recvcounts=static_cast<int*>(xbt_malloc(count*sizeof(int)));
+  for (i=0; i<count;i++)
+    recvcounts[i]=recvcount;
+  mpi_coll_reduce_scatter_fun(sendtmpbuf, recvbuf, recvcounts, datatype,  op, comm);
+  xbt_free(recvcounts);
+  retval = MPI_SUCCESS;

 
     TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
 
     TRACE_smpi_collective_out(rank, -1, __FUNCTION__);

+
+  if(sendbuf==MPI_IN_PLACE)
+    xbt_free(sendtmpbuf);

   }
 
   smpi_bench_begin();
   }
 
   smpi_bench_begin();


@@ -2249,29 +2298,44 @@ int PMPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype,
 
   if (comm == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;
 
   if (comm == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;

-  } else if (sendtype == MPI_DATATYPE_NULL
+  } else if ((sendbuf != MPI_IN_PLACE && sendtype == MPI_DATATYPE_NULL)

              || recvtype == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else {
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
   instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
   extra->type = TRACING_ALLTOALL;
              || recvtype == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
   } else {
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
   instr_extra_data extra = xbt_new0(s_instr_extra_data_t,1);
   extra->type = TRACING_ALLTOALL;

+
+  void* sendtmpbuf = static_cast<char*>(sendbuf);
+  int sendtmpcount = sendcount;
+  MPI_Datatype sendtmptype = sendtype;
+  if( sendbuf == MPI_IN_PLACE ) {
+    sendtmpbuf = static_cast<void*>(xbt_malloc(recvcount*smpi_comm_size(comm)*smpi_datatype_size(recvtype)));
+    memcpy(sendtmpbuf,recvbuf, recvcount*smpi_comm_size(comm)*smpi_datatype_size(recvtype));
+    sendtmpcount = recvcount;
+    sendtmptype = recvtype;
+  }
+

   int known=0;
   int known=0;

-  extra->datatype1 = encode_datatype(sendtype, &known);
+  extra->datatype1 = encode_datatype(sendtmptype, &known);

   if(known==0)
   if(known==0)

-    extra->send_size = sendcount*smpi_datatype_size(sendtype);
+    extra->send_size = sendtmpcount*smpi_datatype_size(sendtmptype);

   else
   else

-    extra->send_size = sendcount;
+    extra->send_size = sendtmpcount;

   extra->datatype2 = encode_datatype(recvtype, &known);
   if(known==0)
     extra->recv_size = recvcount*smpi_datatype_size(recvtype);
   else
     extra->recv_size = recvcount;
   extra->datatype2 = encode_datatype(recvtype, &known);
   if(known==0)
     extra->recv_size = recvcount*smpi_datatype_size(recvtype);
   else
     extra->recv_size = recvcount;

+

   TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
 

-  retval = mpi_coll_alltoall_fun(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
+  retval = mpi_coll_alltoall_fun(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, comm);

 
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
 
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);

+
+  if( sendbuf == MPI_IN_PLACE ) 
+    xbt_free(sendtmpbuf);

   }
 
   smpi_bench_begin();
   }
 
   smpi_bench_begin();


@@ -2289,7 +2353,7 @@ int PMPI_Alltoallv(void *sendbuf, int *sendcounts, int *senddisps,MPI_Datatype s
     retval = MPI_ERR_COMM;
   } else if (sendtype == MPI_DATATYPE_NULL || recvtype == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;
     retval = MPI_ERR_COMM;
   } else if (sendtype == MPI_DATATYPE_NULL || recvtype == MPI_DATATYPE_NULL) {
     retval = MPI_ERR_TYPE;

-  } else if (sendcounts == nullptr || senddisps == nullptr || recvcounts == nullptr || recvdisps == nullptr) {
+  } else if ((sendbuf!= MPI_IN_PLACE && (sendcounts == nullptr || senddisps == nullptr)) || recvcounts == nullptr || recvdisps == nullptr) {

     retval = MPI_ERR_ARG;
   } else {
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
     retval = MPI_ERR_ARG;
   } else {
   int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;


@@ -2302,27 +2366,51 @@ int PMPI_Alltoallv(void *sendbuf, int *sendcounts, int *senddisps,MPI_Datatype s
   extra->recvcounts= xbt_new(int, size);
   extra->sendcounts= xbt_new(int, size);
   int known=0;
   extra->recvcounts= xbt_new(int, size);
   extra->sendcounts= xbt_new(int, size);
   int known=0;

-  extra->datatype1 = encode_datatype(sendtype, &known);
-  int dt_size_send = 1;
-  if(known==0)
-    dt_size_send = smpi_datatype_size(sendtype);

   int dt_size_recv = 1;
   extra->datatype2 = encode_datatype(recvtype, &known);
   int dt_size_recv = 1;
   extra->datatype2 = encode_datatype(recvtype, &known);

-  if(known==0)
-    dt_size_recv = smpi_datatype_size(recvtype);
+  dt_size_recv = smpi_datatype_size(recvtype);
+
+  void* sendtmpbuf = static_cast<char*>(sendbuf);
+  int * sendtmpcounts = sendcounts;
+  int *sendtmpdisps = senddisps;
+  MPI_Datatype sendtmptype = sendtype;
+  int maxsize=0;

   for(i=0; i< size; i++){//copy data to avoid bad free
   for(i=0; i< size; i++){//copy data to avoid bad free

-    extra->send_size += sendcounts[i]*dt_size_send;

     extra->recv_size += recvcounts[i]*dt_size_recv;
     extra->recv_size += recvcounts[i]*dt_size_recv;

-
-    extra->sendcounts[i] = sendcounts[i]*dt_size_send;

     extra->recvcounts[i] = recvcounts[i]*dt_size_recv;
     extra->recvcounts[i] = recvcounts[i]*dt_size_recv;

+    if (((recvdisps[i]+recvcounts[i])*dt_size_recv) > maxsize)
+      maxsize=(recvdisps[i]+recvcounts[i])*dt_size_recv;
+  }
+
+  if( sendbuf == MPI_IN_PLACE ) {
+    sendtmpbuf = static_cast<void*>(xbt_malloc(maxsize));
+    memcpy(sendtmpbuf,recvbuf, maxsize);
+    sendtmpcounts= static_cast<int*>(xbt_malloc(size*sizeof(int)));
+    memcpy(sendtmpcounts,recvcounts, size*sizeof(int));
+    sendtmpdisps= static_cast<int*>(xbt_malloc(size*sizeof(int)));
+    memcpy(sendtmpdisps,recvdisps, size*sizeof(int));
+    sendtmptype=recvtype;
+  }
+
+  extra->datatype1 = encode_datatype(sendtmptype, &known);
+  int dt_size_send = 1;
+  dt_size_send = smpi_datatype_size(sendtmptype);
+
+  for(i=0; i< size; i++){//copy data to avoid bad free
+    extra->send_size += sendtmpcounts[i]*dt_size_send;
+    extra->sendcounts[i] = sendtmpcounts[i]*dt_size_send;

   }
   extra->num_processes = size;
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);
   }
   extra->num_processes = size;
   TRACE_smpi_collective_in(rank, -1, __FUNCTION__,extra);

-
-  retval = mpi_coll_alltoallv_fun(sendbuf, sendcounts, senddisps, sendtype, recvbuf, recvcounts, recvdisps, recvtype,
+  retval = mpi_coll_alltoallv_fun(sendtmpbuf, sendtmpcounts, sendtmpdisps, sendtmptype, recvbuf, recvcounts, recvdisps, recvtype,

                                   comm);
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
                                   comm);
   TRACE_smpi_collective_out(rank, -1, __FUNCTION__);

+
+  if( sendbuf == MPI_IN_PLACE ) {
+    xbt_free(sendtmpbuf);
+    xbt_free(sendtmpcounts);
+    xbt_free(sendtmpdisps);
+  }

   }
 
   smpi_bench_begin();
   }
 
   smpi_bench_begin();


@@ -2334,9 +2422,9 @@ int PMPI_Get_processor_name(char *name, int *resultlen)
 {
   int retval = MPI_SUCCESS;
 
 {
   int retval = MPI_SUCCESS;
 

-  strncpy(name, sg_host_get_name(SIMIX_host_self()),
-          strlen(sg_host_get_name(SIMIX_host_self())) < MPI_MAX_PROCESSOR_NAME - 1 ?
-          strlen(sg_host_get_name(SIMIX_host_self())) +1 : MPI_MAX_PROCESSOR_NAME - 1 );
+  strncpy(name, SIMIX_host_self()->cname(), strlen(SIMIX_host_self()->cname()) < MPI_MAX_PROCESSOR_NAME - 1
+                                                ? strlen(SIMIX_host_self()->cname()) + 1
+                                                : MPI_MAX_PROCESSOR_NAME - 1);

   *resultlen = strlen(name) > MPI_MAX_PROCESSOR_NAME ? MPI_MAX_PROCESSOR_NAME : strlen(name);
 
   return retval;
   *resultlen = strlen(name) > MPI_MAX_PROCESSOR_NAME ? MPI_MAX_PROCESSOR_NAME : strlen(name);
 
   return retval;


@@ -2454,7 +2542,7 @@ int PMPI_Type_create_indexed_block(int count, int blocklength, int* indices, MPI
   } else if (count<0){
     retval = MPI_ERR_COUNT;
   } else {
   } else if (count<0){
     retval = MPI_ERR_COUNT;
   } else {

-    int* blocklens=static_cast<int*>(xbt_malloc(blocklength*count));
+    int* blocklens=static_cast<int*>(xbt_malloc(blocklength*count*sizeof(int)));

     for (i=0; i<count;i++)
       blocklens[i]=blocklength;
     retval = smpi_datatype_indexed(count, blocklens, indices, old_type, new_type);
     for (i=0; i<count;i++)
       blocklens[i]=blocklength;
     retval = smpi_datatype_indexed(count, blocklens, indices, old_type, new_type);


@@ -2490,7 +2578,7 @@ int PMPI_Type_create_hindexed_block(int count, int blocklength, MPI_Aint* indice
   } else if (count<0){
     retval = MPI_ERR_COUNT;
   } else {
   } else if (count<0){
     retval = MPI_ERR_COUNT;
   } else {

-    int* blocklens=(int*)xbt_malloc(blocklength*count);
+    int* blocklens=(int*)xbt_malloc(blocklength*count*sizeof(int));

     for (i=0; i<count;i++)blocklens[i]=blocklength;
     retval = smpi_datatype_hindexed(count, blocklens, indices, old_type, new_type);
     xbt_free(blocklens);
     for (i=0; i<count;i++)blocklens[i]=blocklength;
     retval = smpi_datatype_hindexed(count, blocklens, indices, old_type, new_type);
     xbt_free(blocklens);


@@ -2771,7 +2859,7 @@ int PMPI_Put( void *origin_addr, int origin_count, MPI_Datatype origin_datatype,
     smpi_mpi_win_get_group(win, &group);
     int dst_traced = smpi_group_index(group, target_rank);
     TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, nullptr);
     smpi_mpi_win_get_group(win, &group);
     int dst_traced = smpi_group_index(group, target_rank);
     TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__, nullptr);

-    TRACE_smpi_send(rank, rank, dst_traced, origin_count*smpi_datatype_size(origin_datatype));
+    TRACE_smpi_send(rank, rank, dst_traced, SMPI_RMA_TAG, origin_count*smpi_datatype_size(origin_datatype));

 
     retval = smpi_mpi_put( origin_addr, origin_count, origin_datatype, target_rank, target_disp, target_count,
                            target_datatype, win);
 
     retval = smpi_mpi_put( origin_addr, origin_count, origin_datatype, target_rank, target_disp, target_count,
                            target_datatype, win);


@@ -3120,7 +3208,7 @@ int PMPI_Info_create( MPI_Info *info){
   if (info == nullptr)
     return MPI_ERR_ARG;
   *info = xbt_new(s_smpi_mpi_info_t, 1);
   if (info == nullptr)
     return MPI_ERR_ARG;
   *info = xbt_new(s_smpi_mpi_info_t, 1);

-  (*info)->info_dict= xbt_dict_new_homogeneous(nullptr);
+  (*info)->info_dict= xbt_dict_new_homogeneous(xbt_free_f);

   (*info)->refcount=1;
   return MPI_SUCCESS;
 }
   (*info)->refcount=1;
   return MPI_SUCCESS;
 }


@@ -3129,7 +3217,7 @@ int PMPI_Info_set( MPI_Info info, char *key, char *value){
   if (info == nullptr || key == nullptr || value == nullptr)
     return MPI_ERR_ARG;
 
   if (info == nullptr || key == nullptr || value == nullptr)
     return MPI_ERR_ARG;
 

-  xbt_dict_set(info->info_dict, key, (void*)value, nullptr);
+  xbt_dict_set(info->info_dict, key, xbt_strdup(value), nullptr);

   return MPI_SUCCESS;
 }
 
   return MPI_SUCCESS;
 }
 


@@ -3164,13 +3252,13 @@ int PMPI_Info_dup(MPI_Info info, MPI_Info *newinfo){
   if (info == nullptr || newinfo==nullptr)
     return MPI_ERR_ARG;
   *newinfo = xbt_new(s_smpi_mpi_info_t, 1);
   if (info == nullptr || newinfo==nullptr)
     return MPI_ERR_ARG;
   *newinfo = xbt_new(s_smpi_mpi_info_t, 1);

-  (*newinfo)->info_dict= xbt_dict_new_homogeneous(nullptr);
+  (*newinfo)->info_dict= xbt_dict_new_homogeneous(xbt_free_f);

   (*newinfo)->refcount=1;
   xbt_dict_cursor_t cursor = nullptr;
   int *key;
   void* data;
   xbt_dict_foreach(info->info_dict,cursor,key,data){
   (*newinfo)->refcount=1;
   xbt_dict_cursor_t cursor = nullptr;
   int *key;
   void* data;
   xbt_dict_foreach(info->info_dict,cursor,key,data){

-    xbt_dict_set((*newinfo)->info_dict, reinterpret_cast<char*>(key), data, nullptr);
+    xbt_dict_set((*newinfo)->info_dict, reinterpret_cast<char*>(key), xbt_strdup(reinterpret_cast<char*>(data)), nullptr);

   }
   return MPI_SUCCESS;
 }
   }
   return MPI_SUCCESS;
 }


@@ -3258,3 +3346,4 @@ int PMPI_Pack_size(int incount, MPI_Datatype datatype, MPI_Comm comm, int* size)
   return MPI_SUCCESS;
 }
 
   return MPI_SUCCESS;
 }
 

+} // extern "C"