Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
Add preliminary support for MPI_Pack, MPI_Pack_size, and MPI_Unpack.
[simgrid.git] / src / smpi / smpi_mpi_dt.c
index 5fb0a1d..54eecc2 100644 (file)
@@ -1,7 +1,7 @@
 /* smpi_mpi_dt.c -- MPI primitives to handle datatypes                        */
 /* FIXME: a very incomplete implementation                                    */
 
-/* Copyright (c) 2009, 2010. The SimGrid Team.
+/* Copyright (c) 2009-2014. The SimGrid Team.
  * All rights reserved.                                                     */
 
 /* This program is free software; you can redistribute it and/or modify it
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
-
+#include <limits.h>
 #include "private.h"
 #include "smpi_mpi_dt_private.h"
 #include "mc/mc.h"
+#include "xbt/replay.h"
 #include "simgrid/modelchecker.h"
 
 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi_dt, smpi,
                                 "Logging specific to SMPI (datatype)");
 
+#define INTSIZEDCHAR (sizeof(int)*CHAR_BIT-1)/3 + 3 
+xbt_dict_t smpi_type_keyvals = NULL;
+int type_keyval_id=0;//avoid collisions
+
 #define CREATE_MPI_DATATYPE(name, type)       \
   static s_smpi_mpi_datatype_t mpi_##name = { \
+    (char*) # name,                                   \
     sizeof(type),  /* size */                 \
     0,             /*was 1 has_subtype*/             \
     0,             /* lb */                   \
     sizeof(type),  /* ub = lb + size */       \
     DT_FLAG_BASIC,  /* flags */              \
-    NULL           /* pointer on extended struct*/ \
+    NULL,           /* attributes */         \
+    NULL,           /* pointer on extended struct*/ \
   };                                          \
 MPI_Datatype name = &mpi_##name;
 
 #define CREATE_MPI_DATATYPE_NULL(name)       \
   static s_smpi_mpi_datatype_t mpi_##name = { \
+    (char*) # name,                                   \
     0,  /* size */                 \
     0,             /*was 1 has_subtype*/             \
     0,             /* lb */                   \
     0,  /* ub = lb + size */       \
     DT_FLAG_BASIC,  /* flags */              \
+    NULL,           /* attributes */         \
     NULL           /* pointer on extended struct*/ \
   };                                          \
 MPI_Datatype name = &mpi_##name;
@@ -50,6 +59,10 @@ typedef struct {
   float value;
   float index;
 } float_float;
+typedef struct {
+  long value;
+  long index;
+} long_long;
 typedef struct {
   double value;
   double index;
@@ -74,7 +87,10 @@ typedef struct {
   long double value;
   int index;
 } long_double_int;
-
+typedef struct {
+  int64_t value;
+  int64_t index;
+} integer128_t;
 // Predefined data types
 CREATE_MPI_DATATYPE(MPI_CHAR, char);
 CREATE_MPI_DATATYPE(MPI_SHORT, short);
@@ -92,6 +108,7 @@ CREATE_MPI_DATATYPE(MPI_DOUBLE, double);
 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE, long double);
 CREATE_MPI_DATATYPE(MPI_WCHAR, wchar_t);
 CREATE_MPI_DATATYPE(MPI_C_BOOL, _Bool);
+CREATE_MPI_DATATYPE(MPI_BYTE, int8_t);
 CREATE_MPI_DATATYPE(MPI_INT8_T, int8_t);
 CREATE_MPI_DATATYPE(MPI_INT16_T, int16_t);
 CREATE_MPI_DATATYPE(MPI_INT32_T, int32_t);
@@ -113,23 +130,41 @@ CREATE_MPI_DATATYPE(MPI_SHORT_INT, short_int);
 CREATE_MPI_DATATYPE(MPI_2INT, int_int);
 CREATE_MPI_DATATYPE(MPI_2FLOAT, float_float);
 CREATE_MPI_DATATYPE(MPI_2DOUBLE, double_double);
+CREATE_MPI_DATATYPE(MPI_2LONG, long_long);
+
+CREATE_MPI_DATATYPE(MPI_REAL, float);
+CREATE_MPI_DATATYPE(MPI_REAL4, float);
+CREATE_MPI_DATATYPE(MPI_REAL8, float);
+CREATE_MPI_DATATYPE(MPI_REAL16, double);
+CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX8);
+CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX16);
+CREATE_MPI_DATATYPE_NULL(MPI_COMPLEX32);
+CREATE_MPI_DATATYPE(MPI_INTEGER1, int);
+CREATE_MPI_DATATYPE(MPI_INTEGER2, int16_t);
+CREATE_MPI_DATATYPE(MPI_INTEGER4, int32_t);
+CREATE_MPI_DATATYPE(MPI_INTEGER8, int64_t);
+CREATE_MPI_DATATYPE(MPI_INTEGER16, integer128_t);
 
 CREATE_MPI_DATATYPE(MPI_LONG_DOUBLE_INT, long_double_int);
 
 CREATE_MPI_DATATYPE_NULL(MPI_UB);
 CREATE_MPI_DATATYPE_NULL(MPI_LB);
-CREATE_MPI_DATATYPE_NULL(MPI_PACKED);
+CREATE_MPI_DATATYPE(MPI_PACKED, char);
 // Internal use only
 CREATE_MPI_DATATYPE(MPI_PTR, void*);
 
+/** Check if the datatype is usable for communications
+ */
+int is_datatype_valid(MPI_Datatype datatype) {
+    return datatype != MPI_DATATYPE_NULL
+        && (datatype->flags & DT_FLAG_COMMITED);
+}
 
 size_t smpi_datatype_size(MPI_Datatype datatype)
 {
   return datatype->size;
 }
 
-
-
 MPI_Aint smpi_datatype_lb(MPI_Datatype datatype)
 {
   return datatype->lb;
@@ -140,23 +175,77 @@ MPI_Aint smpi_datatype_ub(MPI_Datatype datatype)
   return datatype->ub;
 }
 
+int smpi_datatype_dup(MPI_Datatype datatype, MPI_Datatype* new_t)
+{
+  int ret=MPI_SUCCESS;
+  *new_t= xbt_new(s_smpi_mpi_datatype_t,1);
+  memcpy(*new_t, datatype, sizeof(s_smpi_mpi_datatype_t));
+  if (datatype->has_subtype){
+    //FIXME: may copy too much information.
+    (*new_t)->substruct=xbt_malloc(sizeof(s_smpi_mpi_struct_t));
+    memcpy((*new_t)->substruct, datatype->substruct, sizeof(s_smpi_mpi_struct_t));
+  }
+  if(datatype->name)
+    (*new_t)->name = strdup(datatype->name);
+  if(datatype->attributes !=NULL){
+      (*new_t)->attributes=xbt_dict_new();
+      xbt_dict_cursor_t cursor = NULL;
+      int *key;
+      int flag;
+      void* value_in;
+      void* value_out;
+      xbt_dict_foreach(datatype->attributes, cursor, key, value_in){
+        smpi_type_key_elem elem = xbt_dict_get_or_null(smpi_type_keyvals, (const char*)key);
+        if(elem && elem->copy_fn!=MPI_NULL_COPY_FN){
+          ret = elem->copy_fn(datatype, atoi((const char*)key), NULL, value_in, &value_out, &flag );
+          if(ret!=MPI_SUCCESS){
+            *new_t=MPI_DATATYPE_NULL;
+            return ret;
+          }
+          if(flag)
+            xbt_dict_set((*new_t)->attributes, (const char*)key,value_out, NULL);
+        }
+      }
+    }
+  return ret;
+}
+
 int smpi_datatype_extent(MPI_Datatype datatype, MPI_Aint * lb,
                          MPI_Aint * extent)
 {
+  if(datatype == MPI_DATATYPE_NULL){
+    *lb=0;
+    *extent=0;
+    return MPI_SUCCESS;
+  }
   *lb = datatype->lb;
   *extent = datatype->ub - datatype->lb;
   return MPI_SUCCESS;
 }
 
 MPI_Aint smpi_datatype_get_extent(MPI_Datatype datatype){
+  if(datatype == MPI_DATATYPE_NULL){
+    return 0;
+  }
   return datatype->ub - datatype->lb;
 }
 
+void smpi_datatype_get_name(MPI_Datatype datatype, char* name, int* length){
+  *length = strlen(datatype->name);
+  strcpy(name, datatype->name);
+}
+
+void smpi_datatype_set_name(MPI_Datatype datatype, char* name){
+  datatype->name = strdup(name);;
+}
+
 int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
                        void *recvbuf, int recvcount, MPI_Datatype recvtype)
 {
   int count;
-
+  if(smpi_privatize_global_variables){
+    smpi_switch_data_segment(smpi_process_index());
+  }
   /* First check if we really have something to do */
   if (recvcount > 0 && recvbuf != sendbuf) {
     /* FIXME: treat packed cases */
@@ -165,17 +254,17 @@ int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
     count = sendcount < recvcount ? sendcount : recvcount;
 
     if(sendtype->has_subtype == 0 && recvtype->has_subtype == 0) {
-      memcpy(recvbuf, sendbuf, count);
+      if(!smpi_process_get_replaying()) memcpy(recvbuf, sendbuf, count);
     }
     else if (sendtype->has_subtype == 0)
     {
       s_smpi_subtype_t *subtype =  recvtype->substruct;
-      subtype->unserialize( sendbuf, recvbuf,1, subtype);
+      subtype->unserialize( sendbuf, recvbuf, recvcount/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
     }
     else if (recvtype->has_subtype == 0)
     {
       s_smpi_subtype_t *subtype =  sendtype->substruct;
-      subtype->serialize(sendbuf, recvbuf,1, subtype);
+      subtype->serialize(sendbuf, recvbuf, sendcount/smpi_datatype_size(sendtype), subtype);
     }else{
       s_smpi_subtype_t *subtype =  sendtype->substruct;
 
@@ -184,7 +273,7 @@ int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
 
       subtype->serialize( sendbuf, buf_tmp,count/smpi_datatype_size(sendtype), subtype);
       subtype =  recvtype->substruct;
-      subtype->unserialize( buf_tmp, recvbuf,count/smpi_datatype_size(recvtype), subtype);
+      subtype->unserialize( buf_tmp, recvbuf,count/smpi_datatype_size(recvtype), subtype, MPI_REPLACE);
 
       free(buf_tmp);
     }
@@ -204,7 +293,7 @@ int smpi_datatype_copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
  */
 void serialize_vector( const void *noncontiguous_vector,
                        void *contiguous_vector,
-                       size_t count,
+                       int count,
                        void *type)
 {
   s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
@@ -241,8 +330,9 @@ void serialize_vector( const void *noncontiguous_vector,
  */
 void unserialize_vector( const void *contiguous_vector,
                          void *noncontiguous_vector,
-                         size_t count,
-                         void *type)
+                         int count,
+                         void *type,
+                         MPI_Op op)
 {
   s_smpi_mpi_vector_t* type_c = (s_smpi_mpi_vector_t*)type;
   int i;
@@ -252,13 +342,16 @@ void unserialize_vector( const void *contiguous_vector,
 
   for (i = 0; i < type_c->block_count * count; i++) {
     if (type_c->old_type->has_subtype == 0)
-      memcpy(noncontiguous_vector_char,
-             contiguous_vector_char, type_c->block_length * type_c->size_oldtype);
+      smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length,
+          &type_c->old_type);
+     /* memcpy(noncontiguous_vector_char,
+             contiguous_vector_char, type_c->block_length * type_c->size_oldtype);*/
     else
       ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char,
                                                                      noncontiguous_vector_char,
                                                                      type_c->block_length,
-                                                                     type_c->old_type->substruct);
+                                                                     type_c->old_type->substruct,
+                                                                     op);
     contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
     if((i+1)%type_c->block_count ==0)
     noncontiguous_vector_char += type_c->block_length*smpi_datatype_get_extent(type_c->old_type);
@@ -285,6 +378,7 @@ s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride,
   new_t->block_stride = block_stride;
   new_t->block_length = block_length;
   new_t->block_count = block_count;
+  smpi_datatype_use(old_type);
   new_t->old_type = old_type;
   new_t->size_oldtype = size_oldtype;
   return new_t;
@@ -293,6 +387,7 @@ s_smpi_mpi_vector_t* smpi_datatype_vector_create( int block_stride,
 void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int has_subtype,
                           void *struct_type, int flags){
   MPI_Datatype new_t= xbt_new(s_smpi_mpi_datatype_t,1);
+  new_t->name = NULL;
   new_t->size = size;
   new_t->has_subtype = size>0? has_subtype:0;
   new_t->lb = lb;
@@ -300,13 +395,27 @@ void smpi_datatype_create(MPI_Datatype* new_type, int size,int lb, int ub, int h
   new_t->flags = flags;
   new_t->substruct = struct_type;
   new_t->in_use=0;
+  new_t->attributes=NULL;
   *new_type = new_t;
 
+#ifdef HAVE_MC
   if(MC_is_active())
     MC_ignore(&(new_t->in_use), sizeof(new_t->in_use));
+#endif
 }
 
 void smpi_datatype_free(MPI_Datatype* type){
+  if((*type)->attributes !=NULL){
+      xbt_dict_cursor_t cursor = NULL;
+      int* key;
+      void * value;
+      int flag;
+      xbt_dict_foreach((*type)->attributes, cursor, key, value){
+        smpi_type_key_elem elem = xbt_dict_get_or_null(smpi_type_keyvals, (const char*)key);
+        if(elem &&  elem->delete_fn)
+          elem->delete_fn(*type, atoi((const char*)key), value, &flag);
+      }
+  }
 
   if((*type)->flags & DT_FLAG_PREDEFINED)return;
 
@@ -320,15 +429,20 @@ void smpi_datatype_free(MPI_Datatype* type){
     ((s_smpi_subtype_t *)(*type)->substruct)->subtype_free(type);  
     xbt_free((*type)->substruct);
   }
+  if ((*type)->name != NULL){
+    xbt_free((*type)->name);
+  }
   xbt_free(*type);
-
+  *type = MPI_DATATYPE_NULL;
 }
 
 void smpi_datatype_use(MPI_Datatype type){
   if(type)type->in_use++;
-  
+
+#ifdef HAVE_MC
   if(MC_is_active())
     MC_ignore(&(type->in_use), sizeof(type->in_use));
+#endif
 }
 
 
@@ -336,8 +450,10 @@ void smpi_datatype_unuse(MPI_Datatype type){
   if(type && type->in_use-- == 0 && (type->flags & DT_FLAG_DESTROYED))
     smpi_datatype_free(&type);
 
+#ifdef HAVE_MC
   if(MC_is_active())
     MC_ignore(&(type->in_use), sizeof(type->in_use));
+#endif
 }
 
 
@@ -359,7 +475,7 @@ Contiguous Implementation
  */
 void serialize_contiguous( const void *noncontiguous_hvector,
                        void *contiguous_hvector,
-                       size_t count,
+                       int count,
                        void *type)
 {
   s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
@@ -379,18 +495,22 @@ void serialize_contiguous( const void *noncontiguous_hvector,
  */
 void unserialize_contiguous( const void *contiguous_vector,
                          void *noncontiguous_vector,
-                         size_t count,
-                         void *type)
+                         int count,
+                         void *type,
+                         MPI_Op op)
 {
   s_smpi_mpi_contiguous_t* type_c = (s_smpi_mpi_contiguous_t*)type;
   char* contiguous_vector_char = (char*)contiguous_vector;
   char* noncontiguous_vector_char = (char*)noncontiguous_vector+type_c->lb;
-
-  memcpy(noncontiguous_vector_char,
-           contiguous_vector_char, count*  type_c->block_count * type_c->size_oldtype);
+  int n= count* type_c->block_count;
+  smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &n,
+            &type_c->old_type);
+       /*memcpy(noncontiguous_vector_char,
+           contiguous_vector_char, count*  type_c->block_count * type_c->size_oldtype);*/
 }
 
 void free_contiguous(MPI_Datatype* d){
+  smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
 }
 
 /*
@@ -411,6 +531,7 @@ s_smpi_mpi_contiguous_t* smpi_datatype_contiguous_create( MPI_Aint lb,
   new_t->block_count = block_count;
   new_t->old_type = old_type;
   new_t->size_oldtype = size_oldtype;
+  smpi_datatype_use(old_type);
   return new_t;
 }
 
@@ -478,6 +599,7 @@ int smpi_datatype_vector(int count, int blocklen, int stride, MPI_Datatype old_t
 }
 
 void free_vector(MPI_Datatype* d){
+  smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
 }
 
 /*
@@ -496,7 +618,7 @@ Hvector Implementation - Vector with stride in bytes
  */
 void serialize_hvector( const void *noncontiguous_hvector,
                        void *contiguous_hvector,
-                       size_t count,
+                       int count,
                        void *type)
 {
   s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
@@ -532,8 +654,9 @@ void serialize_hvector( const void *noncontiguous_hvector,
  */
 void unserialize_hvector( const void *contiguous_vector,
                          void *noncontiguous_vector,
-                         size_t count,
-                         void *type)
+                         int count,
+                         void *type,
+                         MPI_Op op)
 {
   s_smpi_mpi_hvector_t* type_c = (s_smpi_mpi_hvector_t*)type;
   int i;
@@ -543,13 +666,16 @@ void unserialize_hvector( const void *contiguous_vector,
 
   for (i = 0; i < type_c->block_count * count; i++) {
     if (type_c->old_type->has_subtype == 0)
-      memcpy(noncontiguous_vector_char,
-           contiguous_vector_char, type_c->block_length * type_c->size_oldtype);
+      smpi_op_apply(op, contiguous_vector_char, noncontiguous_vector_char, &type_c->block_length,
+                  &type_c->old_type);
+             /*memcpy(noncontiguous_vector_char,
+           contiguous_vector_char, type_c->block_length * type_c->size_oldtype);*/
     else
       ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_vector_char,
                                                                      noncontiguous_vector_char,
                                                                      type_c->block_length,
-                                                                     type_c->old_type->substruct);
+                                                                     type_c->old_type->substruct,
+                                                                     op);
     contiguous_vector_char += type_c->block_length*type_c->size_oldtype;
     if((i+1)%type_c->block_count ==0)
     noncontiguous_vector_char += type_c->block_length*type_c->size_oldtype;
@@ -578,11 +704,13 @@ s_smpi_mpi_hvector_t* smpi_datatype_hvector_create( MPI_Aint block_stride,
   new_t->block_count = block_count;
   new_t->old_type = old_type;
   new_t->size_oldtype = size_oldtype;
+  smpi_datatype_use(old_type);
   return new_t;
 }
 
 //do nothing for vector types
 void free_hvector(MPI_Datatype* d){
+  smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*d)->substruct)->old_type);
 }
 
 int smpi_datatype_hvector(int count, int blocklen, MPI_Aint stride, MPI_Datatype old_type, MPI_Datatype* new_type)
@@ -636,7 +764,7 @@ Indexed Implementation
  */
 void serialize_indexed( const void *noncontiguous_indexed,
                        void *contiguous_indexed,
-                       size_t count,
+                       int count,
                        void *type)
 {
   s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
@@ -673,8 +801,9 @@ void serialize_indexed( const void *noncontiguous_indexed,
  */
 void unserialize_indexed( const void *contiguous_indexed,
                          void *noncontiguous_indexed,
-                         size_t count,
-                         void *type)
+                         int count,
+                         void *type,
+                         MPI_Op op)
 {
 
   s_smpi_mpi_indexed_t* type_c = (s_smpi_mpi_indexed_t*)type;
@@ -684,13 +813,16 @@ void unserialize_indexed( const void *contiguous_indexed,
   for(j=0; j<count;j++){
     for (i = 0; i < type_c->block_count; i++) {
       if (type_c->old_type->has_subtype == 0)
-        memcpy(noncontiguous_indexed_char ,
-             contiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
+        smpi_op_apply(op, contiguous_indexed_char, noncontiguous_indexed_char, &type_c->block_lengths[i],
+                    &type_c->old_type);
+               /*memcpy(noncontiguous_indexed_char ,
+             contiguous_indexed_char, type_c->block_lengths[i] * type_c->size_oldtype);*/
       else
         ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_indexed_char,
                                                                        noncontiguous_indexed_char,
                                                                        type_c->block_lengths[i],
-                                                                       type_c->old_type->substruct);
+                                                                       type_c->old_type->substruct,
+                                                                       op);
 
       contiguous_indexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
       if (i<type_c->block_count-1)
@@ -704,6 +836,7 @@ void unserialize_indexed( const void *contiguous_indexed,
 void free_indexed(MPI_Datatype* type){
   xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_lengths);
   xbt_free(((s_smpi_mpi_indexed_t *)(*type)->substruct)->block_indices);
+  smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
 }
 
 /*
@@ -729,6 +862,7 @@ s_smpi_mpi_indexed_t* smpi_datatype_indexed_create( int* block_lengths,
     new_t->block_indices[i]=block_indices[i];
   }
   new_t->block_count = block_count;
+  smpi_datatype_use(old_type);
   new_t->old_type = old_type;
   new_t->size_oldtype = size_oldtype;
   return new_t;
@@ -799,7 +933,7 @@ Hindexed Implementation - Indexed with indices in bytes
  */
 void serialize_hindexed( const void *noncontiguous_hindexed,
                        void *contiguous_hindexed,
-                       size_t count,
+                       int count,
                        void *type)
 {
   s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
@@ -835,8 +969,9 @@ void serialize_hindexed( const void *noncontiguous_hindexed,
  */
 void unserialize_hindexed( const void *contiguous_hindexed,
                          void *noncontiguous_hindexed,
-                         size_t count,
-                         void *type)
+                         int count,
+                         void *type,
+                         MPI_Op op)
 {
   s_smpi_mpi_hindexed_t* type_c = (s_smpi_mpi_hindexed_t*)type;
   int i,j;
@@ -846,13 +981,16 @@ void unserialize_hindexed( const void *contiguous_hindexed,
   for(j=0; j<count;j++){
     for (i = 0; i < type_c->block_count; i++) {
       if (type_c->old_type->has_subtype == 0)
-        memcpy(noncontiguous_hindexed_char,
-               contiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);
+        smpi_op_apply(op, contiguous_hindexed_char, noncontiguous_hindexed_char, &type_c->block_lengths[i],
+                            &type_c->old_type);
+                       /*memcpy(noncontiguous_hindexed_char,
+               contiguous_hindexed_char, type_c->block_lengths[i] * type_c->size_oldtype);*/
       else
         ((s_smpi_subtype_t*)type_c->old_type->substruct)->unserialize( contiguous_hindexed_char,
                                                                        noncontiguous_hindexed_char,
                                                                        type_c->block_lengths[i],
-                                                                       type_c->old_type->substruct);
+                                                                       type_c->old_type->substruct,
+                                                                       op);
 
       contiguous_hindexed_char += type_c->block_lengths[i]*type_c->size_oldtype;
       if (i<type_c->block_count-1)noncontiguous_hindexed_char = (char*)noncontiguous_hindexed + type_c->block_indices[i+1];
@@ -865,6 +1003,7 @@ void unserialize_hindexed( const void *contiguous_hindexed,
 void free_hindexed(MPI_Datatype* type){
   xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_lengths);
   xbt_free(((s_smpi_mpi_hindexed_t *)(*type)->substruct)->block_indices);
+  smpi_datatype_unuse(((s_smpi_mpi_indexed_t *)(*type)->substruct)->old_type);
 }
 
 /*
@@ -960,7 +1099,7 @@ struct Implementation - Indexed with indices in bytes
  */
 void serialize_struct( const void *noncontiguous_struct,
                        void *contiguous_struct,
-                       size_t count,
+                       int count,
                        void *type)
 {
   s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type;
@@ -997,8 +1136,9 @@ void serialize_struct( const void *noncontiguous_struct,
  */
 void unserialize_struct( const void *contiguous_struct,
                          void *noncontiguous_struct,
-                         size_t count,
-                         void *type)
+                         int count,
+                         void *type,
+                         MPI_Op op)
 {
   s_smpi_mpi_struct_t* type_c = (s_smpi_mpi_struct_t*)type;
   int i,j;
@@ -1008,13 +1148,16 @@ void unserialize_struct( const void *contiguous_struct,
   for(j=0; j<count;j++){
     for (i = 0; i < type_c->block_count; i++) {
       if (type_c->old_types[i]->has_subtype == 0)
-        memcpy(noncontiguous_struct_char,
-             contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));
+        smpi_op_apply(op, contiguous_struct_char, noncontiguous_struct_char, &type_c->block_lengths[i],
+           & type_c->old_types[i]);
+                       /*memcpy(noncontiguous_struct_char,
+             contiguous_struct_char, type_c->block_lengths[i] * smpi_datatype_size(type_c->old_types[i]));*/
       else
         ((s_smpi_subtype_t*)type_c->old_types[i]->substruct)->unserialize( contiguous_struct_char,
                                                                            noncontiguous_struct_char,
                                                                            type_c->block_lengths[i],
-                                                                           type_c->old_types[i]->substruct);
+                                                                           type_c->old_types[i]->substruct,
+                                                                           op);
 
       contiguous_struct_char += type_c->block_lengths[i]*smpi_datatype_size(type_c->old_types[i]);
       if (i<type_c->block_count-1)noncontiguous_struct_char =  (char*)noncontiguous_struct + type_c->block_indices[i+1];
@@ -1028,6 +1171,9 @@ void unserialize_struct( const void *contiguous_struct,
 void free_struct(MPI_Datatype* type){
   xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_lengths);
   xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->block_indices);
+  int i=0;
+  for (i = 0; i < ((s_smpi_mpi_struct_t *)(*type)->substruct)->block_count; i++)
+    smpi_datatype_unuse(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types[i]);
   xbt_free(((s_smpi_mpi_struct_t *)(*type)->substruct)->old_types);
 }
 
@@ -1053,6 +1199,7 @@ s_smpi_mpi_struct_t* smpi_datatype_struct_create( int* block_lengths,
     new_t->block_lengths[i]=block_lengths[i];
     new_t->block_indices[i]=block_indices[i];
     new_t->old_types[i]=old_types[i];
+    smpi_datatype_use(new_t->old_types[i]);
   }
   //new_t->block_lengths = block_lengths;
   //new_t->block_indices = block_indices;
@@ -1137,7 +1284,6 @@ typedef struct s_smpi_mpi_op {
 #define BXOR_OP(a, b) (b) ^= (a)
 #define MAXLOC_OP(a, b)  (b) = (a.value) < (b.value) ? (b) : (a)
 #define MINLOC_OP(a, b)  (b) = (a.value) < (b.value) ? (a) : (b)
-//TODO : MINLOC & MAXLOC
 
 #define APPLY_FUNC(a, b, length, type, func) \
 {                                          \
@@ -1166,6 +1312,8 @@ static void max_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, unsigned int, MAX_OP);
   } else if (*datatype == MPI_UNSIGNED_LONG) {
     APPLY_FUNC(a, b, length, unsigned long, MAX_OP);
+  } else if (*datatype == MPI_UNSIGNED_CHAR) {
+    APPLY_FUNC(a, b, length, unsigned char, MAX_OP);
   } else if (*datatype == MPI_FLOAT) {
     APPLY_FUNC(a, b, length, float, MAX_OP);
   } else if (*datatype == MPI_DOUBLE) {
@@ -1192,6 +1340,8 @@ static void min_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, unsigned int, MIN_OP);
   } else if (*datatype == MPI_UNSIGNED_LONG) {
     APPLY_FUNC(a, b, length, unsigned long, MIN_OP);
+  } else if (*datatype == MPI_UNSIGNED_CHAR) {
+    APPLY_FUNC(a, b, length, unsigned char, MIN_OP);
   } else if (*datatype == MPI_FLOAT) {
     APPLY_FUNC(a, b, length, float, MIN_OP);
   } else if (*datatype == MPI_DOUBLE) {
@@ -1218,6 +1368,8 @@ static void sum_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, unsigned int, SUM_OP);
   } else if (*datatype == MPI_UNSIGNED_LONG) {
     APPLY_FUNC(a, b, length, unsigned long, SUM_OP);
+  } else if (*datatype == MPI_UNSIGNED_CHAR) {
+    APPLY_FUNC(a, b, length, unsigned char, SUM_OP);
   } else if (*datatype == MPI_FLOAT) {
     APPLY_FUNC(a, b, length, float, SUM_OP);
   } else if (*datatype == MPI_DOUBLE) {
@@ -1250,6 +1402,8 @@ static void prod_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, unsigned int, PROD_OP);
   } else if (*datatype == MPI_UNSIGNED_LONG) {
     APPLY_FUNC(a, b, length, unsigned long, PROD_OP);
+  } else if (*datatype == MPI_UNSIGNED_CHAR) {
+    APPLY_FUNC(a, b, length, unsigned char, PROD_OP);
   } else if (*datatype == MPI_FLOAT) {
     APPLY_FUNC(a, b, length, float, PROD_OP);
   } else if (*datatype == MPI_DOUBLE) {
@@ -1282,6 +1436,8 @@ static void land_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, unsigned int, LAND_OP);
   } else if (*datatype == MPI_UNSIGNED_LONG) {
     APPLY_FUNC(a, b, length, unsigned long, LAND_OP);
+  } else if (*datatype == MPI_UNSIGNED_CHAR) {
+    APPLY_FUNC(a, b, length, unsigned char, LAND_OP);
   } else if (*datatype == MPI_C_BOOL) {
     APPLY_FUNC(a, b, length, _Bool, LAND_OP);
   }
@@ -1304,6 +1460,8 @@ static void lor_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, unsigned int, LOR_OP);
   } else if (*datatype == MPI_UNSIGNED_LONG) {
     APPLY_FUNC(a, b, length, unsigned long, LOR_OP);
+  } else if (*datatype == MPI_UNSIGNED_CHAR) {
+    APPLY_FUNC(a, b, length, unsigned char, LOR_OP);
   } else if (*datatype == MPI_C_BOOL) {
     APPLY_FUNC(a, b, length, _Bool, LOR_OP);
   }
@@ -1326,6 +1484,8 @@ static void lxor_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, unsigned int, LXOR_OP);
   } else if (*datatype == MPI_UNSIGNED_LONG) {
     APPLY_FUNC(a, b, length, unsigned long, LXOR_OP);
+  } else if (*datatype == MPI_UNSIGNED_CHAR) {
+    APPLY_FUNC(a, b, length, unsigned char, LXOR_OP);
   } else if (*datatype == MPI_C_BOOL) {
     APPLY_FUNC(a, b, length, _Bool, LXOR_OP);
   }
@@ -1336,8 +1496,7 @@ static void band_func(void *a, void *b, int *length,
 {
   if (*datatype == MPI_CHAR) {
     APPLY_FUNC(a, b, length, char, BAND_OP);
-  }
-  if (*datatype == MPI_SHORT) {
+  }else if (*datatype == MPI_SHORT) {
     APPLY_FUNC(a, b, length, short, BAND_OP);
   } else if (*datatype == MPI_INT) {
     APPLY_FUNC(a, b, length, int, BAND_OP);
@@ -1349,6 +1508,8 @@ static void band_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, unsigned int, BAND_OP);
   } else if (*datatype == MPI_UNSIGNED_LONG) {
     APPLY_FUNC(a, b, length, unsigned long, BAND_OP);
+  } else if (*datatype == MPI_UNSIGNED_CHAR) {
+    APPLY_FUNC(a, b, length, unsigned char, BAND_OP);
   } else if (*datatype == MPI_BYTE) {
     APPLY_FUNC(a, b, length, uint8_t, BAND_OP);
   }
@@ -1371,6 +1532,8 @@ static void bor_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, unsigned int, BOR_OP);
   } else if (*datatype == MPI_UNSIGNED_LONG) {
     APPLY_FUNC(a, b, length, unsigned long, BOR_OP);
+  } else if (*datatype == MPI_UNSIGNED_CHAR) {
+    APPLY_FUNC(a, b, length, unsigned char, BOR_OP);
   } else if (*datatype == MPI_BYTE) {
     APPLY_FUNC(a, b, length, uint8_t, BOR_OP);
   }
@@ -1393,6 +1556,8 @@ static void bxor_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, unsigned int, BXOR_OP);
   } else if (*datatype == MPI_UNSIGNED_LONG) {
     APPLY_FUNC(a, b, length, unsigned long, BXOR_OP);
+  } else if (*datatype == MPI_UNSIGNED_CHAR) {
+    APPLY_FUNC(a, b, length, unsigned char, BXOR_OP);
   } else if (*datatype == MPI_BYTE) {
     APPLY_FUNC(a, b, length, uint8_t, BXOR_OP);
   }
@@ -1409,6 +1574,8 @@ static void minloc_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, double_int, MINLOC_OP);
   } else if (*datatype == MPI_SHORT_INT) {
     APPLY_FUNC(a, b, length, short_int, MINLOC_OP);
+  } else if (*datatype == MPI_2LONG) {
+    APPLY_FUNC(a, b, length, long_long, MINLOC_OP);
   } else if (*datatype == MPI_2INT) {
     APPLY_FUNC(a, b, length, int_int, MINLOC_OP);
   } else if (*datatype == MPI_LONG_DOUBLE_INT) {
@@ -1431,6 +1598,8 @@ static void maxloc_func(void *a, void *b, int *length,
     APPLY_FUNC(a, b, length, double_int, MAXLOC_OP);
   } else if (*datatype == MPI_SHORT_INT) {
     APPLY_FUNC(a, b, length, short_int, MAXLOC_OP);
+  } else if (*datatype == MPI_2LONG) {
+    APPLY_FUNC(a, b, length, long_long, MAXLOC_OP);
   } else if (*datatype == MPI_2INT) {
     APPLY_FUNC(a, b, length, int_int, MAXLOC_OP);
   } else if (*datatype == MPI_LONG_DOUBLE_INT) {
@@ -1442,6 +1611,11 @@ static void maxloc_func(void *a, void *b, int *length,
   }
 }
 
+static void replace_func(void *a, void *b, int *length,
+                        MPI_Datatype * datatype)
+{
+  memcpy(b, a, *length * smpi_datatype_size(*datatype));
+}
 
 #define CREATE_MPI_OP(name, func)                             \
   static s_smpi_mpi_op_t mpi_##name = { &(func) /* func */, TRUE }; \
@@ -1459,6 +1633,8 @@ CREATE_MPI_OP(MPI_BOR, bor_func);
 CREATE_MPI_OP(MPI_BXOR, bxor_func);
 CREATE_MPI_OP(MPI_MAXLOC, maxloc_func);
 CREATE_MPI_OP(MPI_MINLOC, minloc_func);
+CREATE_MPI_OP(MPI_REPLACE, replace_func);
+
 
 MPI_Op smpi_op_new(MPI_User_function * function, int commute)
 {
@@ -1482,5 +1658,136 @@ void smpi_op_destroy(MPI_Op op)
 void smpi_op_apply(MPI_Op op, void *invec, void *inoutvec, int *len,
                    MPI_Datatype * datatype)
 {
+  if(op==MPI_OP_NULL)
+    return;
+
+  if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables
+    XBT_DEBUG("Applying operation, switch to the right data frame ");
+    smpi_switch_data_segment(smpi_process_index());
+  }
+
+  if(!smpi_process_get_replaying())
   op->func(invec, inoutvec, len, datatype);
 }
+
+int smpi_type_attr_delete(MPI_Datatype type, int keyval){
+  char* tmpkey=xbt_malloc(INTSIZEDCHAR);
+  sprintf(tmpkey, "%d", keyval);
+  smpi_type_key_elem elem = xbt_dict_get_or_null(smpi_type_keyvals, (const char*)tmpkey);
+  if(!elem)
+    return MPI_ERR_ARG;
+  if(elem->delete_fn!=MPI_NULL_DELETE_FN){
+    void * value;
+    int flag;
+    if(smpi_type_attr_get(type, keyval, &value, &flag)==MPI_SUCCESS){
+      int ret = elem->delete_fn(type, keyval, value, &flag);
+      if(ret!=MPI_SUCCESS) return ret;
+    }
+  }  
+  if(type->attributes==NULL)
+    return MPI_ERR_ARG;
+
+  xbt_dict_remove(type->attributes, (const char*)tmpkey);
+  xbt_free(tmpkey);
+  return MPI_SUCCESS;
+}
+
+int smpi_type_attr_get(MPI_Datatype type, int keyval, void* attr_value, int* flag){
+  char* tmpkey=xbt_malloc(INTSIZEDCHAR);
+  sprintf(tmpkey, "%d", keyval);
+  smpi_type_key_elem elem = xbt_dict_get_or_null(smpi_type_keyvals, (const char*)tmpkey);
+  if(!elem)
+    return MPI_ERR_ARG;
+  xbt_ex_t ex;
+  if(type->attributes==NULL){
+    *flag=0;
+    return MPI_SUCCESS;
+  }
+  TRY {
+    *(void**)attr_value = xbt_dict_get(type->attributes, (const char*)tmpkey);
+    *flag=1;
+  }
+  CATCH(ex) {
+    *flag=0;
+    xbt_ex_free(ex);
+  }
+  xbt_free(tmpkey);
+  return MPI_SUCCESS;
+}
+
+int smpi_type_attr_put(MPI_Datatype type, int keyval, void* attr_value){
+  if(!smpi_type_keyvals)
+  smpi_type_keyvals = xbt_dict_new();
+  char* tmpkey=xbt_malloc(INTSIZEDCHAR);
+  sprintf(tmpkey, "%d", keyval);
+  smpi_type_key_elem elem = xbt_dict_get_or_null(smpi_type_keyvals, (const char*)tmpkey);
+  if(!elem )
+    return MPI_ERR_ARG;
+  int flag;
+  void* value;
+  smpi_type_attr_get(type, keyval, &value, &flag);
+  if(flag && elem->delete_fn!=MPI_NULL_DELETE_FN){
+    int ret = elem->delete_fn(type, keyval, value, &flag);
+    if(ret!=MPI_SUCCESS) return ret;
+  }
+  if(type->attributes==NULL)
+    type->attributes=xbt_dict_new();
+
+  xbt_dict_set(type->attributes, (const char*)tmpkey, attr_value, NULL);
+  xbt_free(tmpkey);
+  return MPI_SUCCESS;
+}
+
+int smpi_type_keyval_create(MPI_Type_copy_attr_function* copy_fn, MPI_Type_delete_attr_function* delete_fn, int* keyval, void* extra_state){
+
+  if(!smpi_type_keyvals)
+  smpi_type_keyvals = xbt_dict_new();
+  
+  smpi_type_key_elem value = (smpi_type_key_elem) xbt_new0(s_smpi_mpi_type_key_elem_t,1);
+  
+  value->copy_fn=copy_fn;
+  value->delete_fn=delete_fn;
+  
+  *keyval = type_keyval_id;
+  char* tmpkey=xbt_malloc(INTSIZEDCHAR);
+  sprintf(tmpkey, "%d", *keyval);
+  xbt_dict_set(smpi_type_keyvals,(const char*)tmpkey,(void*)value, NULL);
+  type_keyval_id++;
+  xbt_free(tmpkey);
+  return MPI_SUCCESS;
+}
+
+int smpi_type_keyval_free(int* keyval){
+  char* tmpkey=xbt_malloc(INTSIZEDCHAR);
+  sprintf(tmpkey, "%d", *keyval);
+  smpi_type_key_elem elem = xbt_dict_get_or_null(smpi_type_keyvals, (const char*)tmpkey);
+  if(!elem){
+    xbt_free(tmpkey);
+    return MPI_ERR_ARG;
+  }
+  xbt_dict_remove(smpi_type_keyvals, (const char*)tmpkey);
+  xbt_free(elem);
+  xbt_free(tmpkey);
+  return MPI_SUCCESS;
+}
+
+int smpi_mpi_pack(void* inbuf, int incount, MPI_Datatype type, void* outbuf, int outcount, int* position, MPI_Comm comm){
+  size_t size = smpi_datatype_size(type);
+  if (outcount - *position < incount*size)
+    return MPI_ERR_BUFFER;
+  smpi_datatype_copy(inbuf, incount, type,
+                   (char*)outbuf + *position, outcount, MPI_CHAR);
+  *position += incount * size;
+  return MPI_SUCCESS;
+}
+
+int smpi_mpi_unpack(void* inbuf, int insize, int* position, void* outbuf, int outcount, MPI_Datatype type, MPI_Comm comm){
+  size_t size = smpi_datatype_size(type);
+  if (outcount*size> insize)
+    return MPI_ERR_BUFFER;
+  smpi_datatype_copy((char*)inbuf + *position, insize, MPI_CHAR,
+                   outbuf, outcount, type);
+  *position += outcount * size;
+  return MPI_SUCCESS;
+}
+