replace_dist_arrays_in_io: fix code style, replace only distributed arrays

dvm/fdvm/trunk/fdvm/acc.cpp

@@ -5338,6 +5338,18 @@ SgStatement *doIfThenConstrForLoop_GPU(SgExpression *ref, SgStatement *endhost,
     return(ifst);
 }
 
+SgExpression *ReductionPrivateVariables()
+{
+    reduction_operation_list *rl;
+    SgExpression *red_vars=NULL;
+    for (rl = red_struct_list; rl; rl = rl->next)
+    {  
+        red_vars =  AddListToList(red_vars, new SgExprListExp(*new SgVarRefExp(rl->redvar)));   
+        if (rl->locvar) 
+            red_vars = AddListToList(red_vars, new SgExprListExp(*new SgVarRefExp(rl->locvar)));   
+    }
+    return red_vars;
+}
 
 SgExpression * TranslateReductionToOpenmp(SgExpression *reduction_clause)  /* OpenMP */
 {
@@ -6003,20 +6015,20 @@ SgStatement *Create_Host_Loop_Subroutine_Main (SgSymbol *sHostProc)
 
 SgStatement *Create_Host_Loop_Subroutine(SgSymbol *sHostProc, int dependency)
 {
-    SgStatement *stmt = NULL, *st_end = NULL, *st_hedr = NULL, *cur = NULL, *last_decl = NULL, *ass = NULL;
+    SgStatement *stmt = NULL, *st_end = NULL, *st_hedr = NULL, *cur = NULL, *last_decl = NULL, *ass = NULL;
     SgStatement *alloc = NULL, *red_init_first = NULL;
     SgStatement *paralleldo = NULL;
     SgStatement *firstdopar = NULL;
     SgExprListExp *parallellist = NULL;
     SgExprListExp *omp_dolist = NULL;
-    SgExprListExp *omp_perflist = NULL;
+    SgExprListExp *omp_perflist = NULL;
     SgExpression *ae, *arg_list = NULL, *el = NULL, *de = NULL, *tail = NULL, *baseMem_list = NULL, *omp_red_vars=NULL;
     SgSymbol *s_loop_ref = NULL, *sarg = NULL, *h_first = NULL, *hl = NULL;
     SgSymbol *s_lgsc = NULL; /* OpenMP */
     SgVarRefExp *v_lgsc = NULL;   /* OpenMP */
     SgSymbol *s = NULL, *s_low_bound = NULL, *s_high_bound = NULL, *s_step = NULL;
     symb_list *sl = NULL;
-    SgType *tdvm = NULL;
+    SgType *tdvm = NULL;
     int ln, lrank, addopenmp, number_of_reductions = 0;
     char *name;
     tail = NULL;
@@ -6185,7 +6197,7 @@ SgStatement *Create_Host_Loop_Subroutine(SgSymbol *sHostProc, int dependency)
     parallellist = new SgExprListExp(*new SgExpression(OMP_NUM_THREADS, v_lgsc, NULL, NULL)); /* OpenMP */
 
     // create reduction variables declarations and 
-    // generate 'loop_red_init' and 'loop_red_post' function calls
+    // generate 'loop_red_init' and 'loop_red_post' function calls
      
     //looking through the reduction list
     if (red_list)
@@ -6193,10 +6205,9 @@ SgStatement *Create_Host_Loop_Subroutine(SgSymbol *sHostProc, int dependency)
         int nr;
         SgExpression *ev, *ered, *er, *red;
         SgSymbol *loc_var;
-        reduction_operation_list *rl;
-
-        red = TranslateReductionToOpenmp(&red_list->copy());   /* OpenMP */
-        if (red != NULL) parallellist->append(*red);  /* OpenMP */
+        reduction_operation_list *rl;
+        red = TranslateReductionToOpenmp(&red_list->copy());           /* OpenMP */
+        if(red != NULL) parallellist->append(*red);  /* OpenMP */ 
         else  omp_red_vars = ReductionPrivateVariables(); /*MAXLOC/MINLOC*/ /* OpenMP */          
         for (rl = red_struct_list,nr = 1; rl; rl = rl->next, nr++)
         {   
@@ -6209,10 +6220,12 @@ SgStatement *Create_Host_Loop_Subroutine(SgSymbol *sHostProc, int dependency)
             // generate loop_red_init and loop_red_post function calls 
             stmt = LoopRedInit_HH(s_loop_ref, nr, sred, rl->locvar);
             cur->insertStmtAfter(*stmt, *st_hedr);
+            cur = stmt;
             if (nr == 1) red_init_first = stmt;
             stmt = LoopRedPost_HH(s_loop_ref, nr, sred, rl->locvar);
             st_end->insertStmtBefore(*stmt, *st_hedr);
 
+        }
         number_of_reductions = nr; /* OpenMP */
     }
 
@@ -6251,7 +6264,7 @@ SgStatement *Create_Host_Loop_Subroutine(SgSymbol *sHostProc, int dependency)
     if (!options.isOn(O_HOST))
         DeclareArrayCoefficients(st_hedr);
 
-    //   <private_variables>
+    //   <private_variables>
     if ((addopenmp == 1) && (private_list != NULL)) parallellist->append(*new SgExpression(OMP_PRIVATE, &(private_list->copy()), NULL, NULL));  /* OpenMP */
     for (el = private_list; el; el = el->rhs())
     {
@@ -6273,7 +6286,7 @@ SgStatement *Create_Host_Loop_Subroutine(SgSymbol *sHostProc, int dependency)
             else indexes = new SgExprListExp(*el->lhs());      /* OpenMP */
         }   /* OpenMP */
     }
-
+
     if ((addopenmp == 1) && (indexes != NULL)) parallellist->append(*new SgExpression(OMP_PRIVATE, AddListToList(indexes,omp_red_vars), NULL, NULL));  /* OpenMP */
 
     // create dummy argument declarations
@@ -6324,35 +6337,37 @@ SgStatement *Create_Host_Loop_Subroutine(SgSymbol *sHostProc, int dependency)
     if (dependency == 0) firstdopar = stmt = first_do_par->extractStmt();
     else firstdopar = stmt = first_do_par->copyPtr();
     cur->insertStmtAfter(*stmt, *st_hedr); 
-
 
  
-    if (addopenmp == 1) { /* OpenMP */
-		SgCallStmt *stDvmhstat = NULL;
-		SgStatement *omp_do = new SgStatement(OMP_DO_DIR); /* OpenMP */
-		SgStatement *omp_parallel = new SgStatement(OMP_PARALLEL_DIR);  /* OpenMP */
-		SgStatement *omp_endparallel = new SgStatement(OMP_END_PARALLEL_DIR); /* OpenMP */
-		SgStatement *omp_enddo = new SgStatement(OMP_END_DO_DIR); /* OpenMP */
-		SgForStmt *stdo = isSgForStmt(firstdopar); /* OpenMP */
-		SgStatement *lastdo=LastStatementOfDoNest(stdo);
-		cur->insertStmtAfter(*omp_parallel, *st_hedr); /* OpenMP */
-		if (omp_perf) {/* OpenMP */
-			stDvmhstat = new SgCallStmt(*fdvm[OMP_STAT_BP],*omp_perflist);/* OpenMP */
-			stDvmhstat->setlineNumber(-1);/* OpenMP */
-			cur->insertStmtAfter(*stDvmhstat, *st_hedr); /* OpenMP */	
-		}
-		lastdo->insertStmtAfter(*omp_endparallel); /* OpenMP */
-		if (omp_perf) {/* OpenMP */
-			stDvmhstat = new SgCallStmt(*fdvm[OMP_STAT_AL],*omp_perflist);/* OpenMP */
-			stDvmhstat->setlineNumber(-1);/* OpenMP */
-			lastdo->insertStmtAfter(*stDvmhstat);/* OpenMP */
-		}/* OpenMP */
-		omp_parallel->setExpression(0, *parallellist);/* OpenMP */
-		omp_do->setExpression(0, *omp_dolist);/* OpenMP */
-		omp_enddo->setExpression(0, *new SgExprListExp(*new SgExpression(OMP_NOWAIT)));  /* OpenMP */
+    if (addopenmp == 1) { /* OpenMP */
+	SgCallStmt *stDvmhstat = NULL;
+	SgStatement *omp_do = new SgStatement(OMP_DO_DIR); /* OpenMP */
+	SgStatement *omp_parallel = new SgStatement(OMP_PARALLEL_DIR);  /* OpenMP */
+	SgStatement *omp_endparallel = new SgStatement(OMP_END_PARALLEL_DIR); /* OpenMP */
+	SgStatement *omp_enddo = new SgStatement(OMP_END_DO_DIR); /* OpenMP */
+	SgForStmt *stdo = isSgForStmt(firstdopar); /* OpenMP */
+	SgStatement *lastdo=LastStatementOfDoNest(stdo);
+        cur->insertStmtAfter(*omp_parallel, *st_hedr); /* OpenMP */		
+        if (omp_perf) {/* OpenMP */
+	    stDvmhstat = new SgCallStmt(*fdvm[OMP_STAT_BP],*omp_perflist);/* OpenMP */
+	    stDvmhstat->setlineNumber(-1);/* OpenMP */
+	    cur->insertStmtAfter(*stDvmhstat, *st_hedr); /* OpenMP */	
+	}
+        if (omp_red_vars) /* MINLOC/MAXLOC */   /* OpenMP */
+            st_end->insertStmtBefore(*omp_endparallel,*st_hedr); /* OpenMP */
+        else
+            lastdo->insertStmtAfter(*omp_endparallel,*st_hedr); /* OpenMP */
+	if (omp_perf) {/* OpenMP */
+            stDvmhstat = new SgCallStmt(*fdvm[OMP_STAT_AL],*omp_perflist);/* OpenMP */
+	    stDvmhstat->setlineNumber(-1);/* OpenMP */
+	    lastdo->insertStmtAfter(*stDvmhstat);/* OpenMP */
+	}/* OpenMP */
+	omp_parallel->setExpression(0, *parallellist);/* OpenMP */
+	omp_do->setExpression(0, *omp_dolist);/* OpenMP */
+	omp_enddo->setExpression(0, *new SgExprListExp(*new SgExpression(OMP_NOWAIT)));  /* OpenMP */
 	ass = new SgAssignStmt(*v_lgsc, *LoopGetSlotCount_HH(s_loop_ref)); /* OpenMP */
-        if (!dependency) {
-		omp_parallel->insertStmtAfter(*omp_do); /* OpenMP */           
+        if (!dependency) {
+	    omp_parallel->insertStmtAfter(*omp_do); /* OpenMP */           
             lastdo->insertStmtAfter(*omp_enddo); /* OpenMP */ 
 	} else if (isSgForStmt(firstdopar->lexNext())) {  /* OpenMP */
             int step = 1;  /* OpenMP */
@@ -6515,6 +6530,17 @@ SgStatement *Create_Host_Loop_Subroutine(SgSymbol *sHostProc, int dependency)
             if (alloc != NULL) cur->insertStmtAfter(*alloc, *st_hedr); /* OpenMP */
             ass->setlineNumber(-1); /* OpenMP */
         }  /* OpenMP */
+	cur->insertStmtAfter(*ass, *st_hedr); /* OpenMP */
+        if (omp_red_vars) {   /* OpenMP */
+            //transfer of reduction initialization statements in case of maxloc/minloc
+            int i;           /* OpenMP */
+            SgStatement *from = red_init_first->lexPrev();  /* OpenMP */
+            cur = omp_parallel;                             /* OpenMP */
+            for (i=number_of_reductions-1; i; i--) {        /* OpenMP */
+                stmt = from->lexNext()->extractStmt();      /* OpenMP */
+                cur->insertStmtAfter(*stmt);                /* OpenMP */
+                cur = stmt;                                 /* OpenMP */
+            }                                               /* OpenMP */
         }                                                   /* OpenMP */
 	if (omp_perf) {/* OpenMP */	
 		stDvmhstat = new SgCallStmt(*fdvm[OMP_STAT_BL],*omp_perflist);/* OpenMP */
@@ -6523,6 +6549,7 @@ SgStatement *Create_Host_Loop_Subroutine(SgSymbol *sHostProc, int dependency)
 		stDvmhstat = new SgCallStmt(*fdvm[OMP_STAT_AP],*omp_perflist);/* OpenMP */
 		stDvmhstat->setlineNumber(-1);/* OpenMP */
 		omp_endparallel->insertStmtAfter(*stDvmhstat);/* OpenMP */	
+	}/* OpenMP */
        
     }   /* OpenMP */
 
@@ -7943,6 +7970,21 @@ SgSymbol *RedVariableSymbolInKernel(SgSymbol *s, SgExpression *dimSizeArgs, SgEx
 
     return(soff);
 }
+
+SgSymbol *LocRedVariableSymbolInKernel(reduction_operation_list *rsl)
+{
+     SgType *declT;
+
+     if (isSgArrayType(rsl->locvar->type()))
+     {
+         SgArrayType *arrT = new SgArrayType(*C_Type(rsl->locvar->type()));
+         arrT->addDimension(new SgValueExp(rsl->number));
+         declT = arrT;
+     }
+     else
+         declT = C_Type(rsl->locvar->type());
+     return (new SgVariableSymb(rsl->locvar->identifier(), *declT, *kernel_st));
+}
 
 SgSymbol *SymbolInKernel(SgSymbol *s)
 {
@@ -8399,7 +8441,7 @@ void MakeDeclarationsForKernelGpuO1(SgSymbol *red_count_symb, SgType *idxTypeInK
     // declare do_variables
     DeclareDoVars();
 
-    // declare private(local in kernel) variables
+    // declare private(local in kernel) variables
     DeclarePrivateVars(idxTypeInKernel);
 
     // declare dummy arguments:
@@ -9539,7 +9581,7 @@ void MakeDeclarationsForKernel(SgSymbol *red_count_symb, SgType *idxTypeInKernel
     // declare do_variables
     DeclareDoVars();
 
-    // declare private(local in kernel) variables    
+    // declare private(local in kernel) variables    
     DeclarePrivateVars(idxTypeInKernel);
 
     // declare dummy arguments:
@@ -9601,7 +9643,7 @@ void MakeDeclarationsForKernel_On_C(SgType *idxTypeInKernel)
     // declare do_variables
     DeclareDoVars(idxTypeInKernel); 
 
-    // declare private(local in kernel) variables 
+    // declare private(local in kernel) variables 
     DeclarePrivateVars(idxTypeInKernel);
 
     // declare  variables, used in loop and passed by reference:
@@ -9791,6 +9833,11 @@ void DeclareInternalPrivateVars()
     }
 }
 
+void DeclarePrivateVars()
+{
+    DeclarePrivateVars(C_UnsignedLongLongType());
+}
+
 void DeclarePrivateVars(SgType *idxTypeInKernel)
 {
     SgStatement *st = NULL, *st_first=NULL;
@@ -9825,7 +9872,7 @@ void DeclarePrivateVars()
             if (Rank(s)>1)
             {
                 char *name = new char[strlen(s->identifier())+7];
-                sprintf(name, "_%s_dims", s->identifier());
+                sprintf(name, "_%s_dims", s->identifier());
                 s_dims = ArraySymbol(name, idxTypeInKernel, new SgValueExp(Rank(s)-1), kernel_st);
                 SgExpression *einit = new SgExpression(INIT_LIST);
                 SgExpression *elist = NULL;
@@ -9841,7 +9888,7 @@ void DeclarePrivateVars()
                 }
                 else
                 {
-                     for (int i=Rank(s)-1; i; i--)
+                     for (int i=Rank(s)-1; i; i--)
                          elist = AddListToList(elist, new SgExprListExp(*Calculate(ArrayDimSize(s,i))));
                 }   
                 einit->setLhs(elist); 
@@ -10587,18 +10634,7 @@ void ReductionBlockInKernel_On_C_Cuda(SgStatement *stat, SgSymbol *i_var, SgExpr
         }
     }
     else if (rsl->locvar) // maxloc/minloc reduction scalar
-    {
-        SgType *decl;
-        int rank = rsl->number;
-
-        if (rank > 1)
-        {
-            SgArrayType *arrT = new SgArrayType(*C_Type(rsl->locvar->type()));
-            arrT->addDimension(new SgValueExp(rank));
-            decl = arrT;
-        }
-        else
-            decl = C_Type(rsl->locvar->type());
+    {
         newst = Declaration_Statement(LocRedVariableSymbolInKernel(rsl)); //declare location variable
         kernel_st->insertStmtAfter(*newst, *kernel_st);
 
@@ -10615,20 +10651,30 @@ void ReductionBlockInKernel_On_C_Cuda(SgStatement *stat, SgSymbol *i_var, SgExpr
         fun_ref->setRhs(tmplArgs);
 
         stat->insertStmtBefore(*new SgCExpStmt(*fun_ref), *stat->controlParent());
-
+
+        if (across)
+            newst = AssignStatement(new SgArrayRefExp(*rsl->red_grid, *ex), new SgVarRefExp(rsl->redvar));
+        else
+            newst = AssignStatement(new SgArrayRefExp(*rsl->red_grid, *BlockIdxRefExpr("x") * *ex1 + *ex), new SgVarRefExp(rsl->redvar));
 
         if_st->insertStmtAfter(*newst);
 
         if (rsl->number > 1)
         {
             for (int i = 0; i < rsl->number; ++i)
-            {
+            {
+                if (across)
+                    newst = AssignStatement(new SgArrayRefExp(*rsl->loc_grid, *new SgValueExp(rsl->number) * *ex + *new SgValueExp(i)), new SgArrayRefExp(*rsl->locvar, *new SgValueExp(i)));
+                else
                     newst = AssignStatement(new SgArrayRefExp(*rsl->loc_grid, *new SgValueExp(rsl->number) * (*BlockIdxRefExpr("x") * *ex1 + *ex) + *new SgValueExp(i)), new SgArrayRefExp(*rsl->locvar, *new SgValueExp(i)));
                 if_st->lastExecutable()->insertStmtAfter(*newst);
             }
         }
         else
-        {
+        {
+            if (across)
+                newst = AssignStatement(new SgArrayRefExp(*rsl->loc_grid, *ex), new SgVarRefExp(*rsl->locvar));
+            else
                 newst = AssignStatement(new SgArrayRefExp(*rsl->loc_grid, *BlockIdxRefExpr("x") * *ex1 + *ex), new SgVarRefExp(*rsl->locvar));
             if_st->lastExecutable()->insertStmtAfter(*newst);
         }

dvm/fdvm/trunk/fdvm/acc_across.cpp

@@ -1588,6 +1588,7 @@ vector<ArgsForKernel> Create_C_Adapter_Function_Across_OneThread(SgSymbol *sadap
             for (int i = NumberOfCoeffs(sg); i>0; i--)
                 funcCall->addArg(*new SgArrayRefExp(*sg, *new SgValueExp(i)));
         }
+
         if (red_list)
         {
             reduction_operation_list *rsl;
@@ -1613,6 +1614,7 @@ vector<ArgsForKernel> Create_C_Adapter_Function_Across_OneThread(SgSymbol *sadap
                         funcCall->addArg(*new SgArrayRefExp(*s, *new SgValueExp(i)));
                 }
                 s = s->next();
+
                 if (options.isOn(C_CUDA))
                     funcCall->addArg(*new SgVarRefExp(reduction_ptr[i]));
                 else
@@ -1738,8 +1740,8 @@ static inline void insertReductionArgs(SgSymbol **reduction_ptr, SgSymbol **redu
                                        SgFunctionCallExp *funcCallKernel, SgSymbol* numBlocks, int &has_red_array)
 {
     reduction_operation_list *rsl;
-    SgSymbol *s;
-    SgExpression *e;
+    SgSymbol *s = NULL;
+    SgExpression *e = NULL;
 
     for (rsl = red_struct_list, s = red_first; rsl; rsl = rsl->next)  //s!=s_blocks_info
     {
@@ -1776,13 +1778,16 @@ static inline void insertReductionArgs(SgSymbol **reduction_ptr, SgSymbol **redu
         else
             funcCallKernel->addArg(*new SgCastExp(*C_PointerType(new SgDescriptType(*SgTypeChar(), BIT_SIGNED)), *new SgVarRefExp(reduction_ptr[i])));
 
-        //TODO!!
         if (rsl->locvar)  //MAXLOC,MINLOC
         {
             for (int k = 0; k < rsl->number; ++k)
                 funcCallKernel->addArg(*new SgArrayRefExp(*reduction_loc_symb[i], *new SgValueExp(k)));
             s = s->next();
-            e = new SgCastExp(*C_PointerType(options.isOn(C_CUDA) ? C_Type(rsl->locvar->type()) : new SgDescriptType(*SgTypeChar(), BIT_SIGNED)), *new SgVarRefExp(s));
+
+            if (options.isOn(C_CUDA))
+                e = new SgCastExp(*C_PointerType(C_Type(rsl->locvar->type())), *new SgVarRefExp(reduction_loc_ptr[i]));
+            else
+                e = new SgCastExp(*C_PointerType(new SgDescriptType(*SgTypeChar(), BIT_SIGNED)), *new SgVarRefExp(s));// TODO it like in C_Cuda
             funcCallKernel->addArg(*e);
             s = s->next();
         }
@@ -2442,6 +2447,7 @@ vector<ArgsForKernel> Create_C_Adapter_Function_Across_variants(SgSymbol *sadapt
         stmt = makeSymbolDeclarationWithInit(s, new SgValueExp(0));
         st_hedr->insertStmtAfter(*stmt, *st_hedr);
     }
+
     // create indxs
     for (int i = 0; i < acrossV; ++i)
     {
@@ -4216,7 +4222,7 @@ void MakeDeclarationsForKernel_On_C_Across(SgType *indexType)
     DeclareDoVars(indexType);
 
     // declare private(local in kernel) variables
-    DeclarePrivateVars();
+    DeclarePrivateVars(indexType);
 
     // declare  variables, used in loop and passed by reference:
     // <type> &<name> = *p_<name>;
@@ -4233,7 +4239,7 @@ void MakeDeclarationsForKernelAcross(SgType *indexType)
     DeclareDoVars();
 
     // declare private(local in kernel) variables
-    DeclarePrivateVars();
+    DeclarePrivateVars(indexType);
 
     // declare dummy arguments:
 
@@ -5829,6 +5835,7 @@ SgStatement *CreateLoopKernelAcross(SgSymbol *skernel, ArgsForKernel* argsKer, i
                 flag_func_call = 0; // maxloc
             else if (num == 10)
                 flag_func_call = 0; // minloc
+
             if (flag_func_call == 1)
             {
                 SgFunctionCallExp *funcCall = new SgFunctionCallExp(*createNewFunctionSymbol(str_operation));
@@ -5923,9 +5930,9 @@ SgStatement *CreateLoopKernelAcross(SgSymbol *skernel, ArgsForKernel* argsKer, i
                         locGrid->setType(*new SgArrayType(*tmp_list->loc_grid->type()));
 
                         if (options.isOn(C_CUDA))
-                            st = AssignStatement(*new SgArrayRefExp(*locGrid, *new SgValueExp(i), *e1), *new SgArrayRefExp(*loc_var_ref->symbol(), *new SgValueExp(i)));
+                            st = AssignStatement(*new SgArrayRefExp(*locGrid, *new SgValueExp(loc_el_num) * *e1 + *new SgValueExp(i)), *new SgArrayRefExp(*loc_var_ref->symbol(), *new SgValueExp(i)));
                         else
-                            st = AssignStatement(*new SgArrayRefExp(*locGrid, *new SgValueExp(i + 1), *e1), *new SgArrayRefExp(*loc_var_ref->symbol(), *new SgValueExp(i + 1)));
+                            st = AssignStatement(*new SgArrayRefExp(*locGrid, *new SgValueExp(i + 1), *e1), *new SgArrayRefExp(*loc_var_ref->symbol(), *new SgValueExp(i + 1)));//TODO it like in C_Cuda
                         ifSt->insertStmtAfter(*st);
                     }
                 }
@@ -6153,13 +6160,13 @@ void DeclarationOfReductionBlockInKernelAcross(SgExpression *ered, reduction_ope
 
     if (rsl->locvar)
     {
-        newst = Declaration_Statement(rsl->locvar); //declare location variable
+        newst = Declaration_Statement(LocRedVariableSymbolInKernel(rsl)); //declare location variable
         kernel_st->insertStmtAfter(*newst, *kernel_st);
     }
 
     if (rsl->redvar_size > 0)
     {
-        newst = Declaration_Statement(rsl->redvar); //declare reduction variable
+        newst = Declaration_Statement(RedVariableSymbolInKernel(rsl->redvar,NULL,NULL)); //declare reduction variable
         kernel_st->insertStmtAfter(*newst, *kernel_st);
     }
     else if (rsl->redvar_size < 0)
@@ -6168,26 +6175,22 @@ void DeclarationOfReductionBlockInKernelAcross(SgExpression *ered, reduction_ope
         newst = Declaration_Statement(red_var_k); //declare reduction variable
         kernel_st->insertStmtAfter(*newst, *kernel_st);
     }
-    rtype = (rsl->redvar_size >= 0) ? TypeOfRedBlockSymbol(ered) : red_var_k->type();
 
-    s_block = RedBlockSymbolInKernelAcross(red_var, rtype);
-
-    newst = Declaration_Statement(s_block);
-
-    if (options.isOn(C_CUDA))
-        newst->addDeclSpec(BIT_CUDA_SHARED | BIT_EXTERN);
-    else
+    //XXX: shared memory doesnt use in ACROSS by C_Cuda
+    if (!options.isOn(C_CUDA))
     {
+        rtype = (rsl->redvar_size >= 0) ? TypeOfRedBlockSymbol(ered) : red_var_k->type();
+        s_block = RedBlockSymbolInKernelAcross(red_var, rtype);
+        newst = Declaration_Statement(s_block);
         eatr = new SgExprListExp(*new SgExpression(ACC_SHARED_OP));
-        newst->setExpression(2, *eatr);
-    }
+        newst->setExpression(2, *eatr);    
+        kernel_st->insertStmtAfter(*newst, *kernel_st);
 
-    kernel_st->insertStmtAfter(*newst, *kernel_st);
-
-    if (isSgExprListExp(ered->rhs())) //MAXLOC,MINLOC
-    {
-        typedecl = MakeStructDecl(rtype->symbol());
-        kernel_st->insertStmtAfter(*typedecl, *kernel_st);
+        if (isSgExprListExp(ered->rhs())) //MAXLOC,MINLOC
+        {
+            typedecl = MakeStructDecl(rtype->symbol());
+            kernel_st->insertStmtAfter(*typedecl, *kernel_st);
+        } 
     }
 }
 

dvm/fdvm/trunk/fdvm/acc_across_analyzer.cpp

@@ -5,10 +5,12 @@
 using namespace std;
 
 // special storages to avoid recomputing
-map<string, SgExpression*> lhs;
-map<string, SgExpression*> rhs;
-map<SgExpression*, string> unparsedLhs;
-map<SgExpression*, string> unparsedRhs;
+static map<string, SgExpression*> lhs;
+static map<string, SgExpression*> rhs;
+static map<SgExpression*, string> unparsedLhs;
+static map<SgExpression*, string> unparsedRhs;
+
+extern reduction_operation_list* red_struct_list;
 
 template<typename InIt1, typename InIt2, typename OutIt>
 static inline OutIt difference(InIt1 first1, InIt1 last1, InIt2 first2, InIt2 last2, OutIt dest)
@@ -1220,7 +1222,14 @@ void Loop::analyzeInderectAccess()
 Loop::Loop(SgStatement* loop_body, bool enable_opt, bool irreg_access) : 
     irregular_acc_opt(irreg_access), enable_opt(enable_opt), loop_body(loop_body), 
     dimension(0), acrossType(0), acrossDims(NULL), do_irreg_opt(false)
-{    
+{
+    reduction_operation_list* rsl;
+    for (rsl = red_struct_list; rsl; rsl = rsl->next)
+    {
+        if (rsl->locvar)  //MAXLOC,MINLOC
+            redArrays.insert(rsl->locvar);
+    }
+
     lhs.clear();
     rhs.clear();
     unparsedLhs.clear();
@@ -1420,7 +1429,7 @@ void Loop::analyzeAssignments(SgExpression* ex, const int blockIndex)
     else
     {
         SgSymbol* symbol = ex->symbol();
-        if (isSgArrayType(symbol->type()) != NULL)
+        if (isSgArrayType(symbol->type()) != NULL && redArrays.find(symbol) == redArrays.end())
         {
             SgExpression* subscripts = ((SgArrayRefExp*)(ex))->subscripts();
             if (!subscripts)
@@ -1949,7 +1958,14 @@ void Loop::buildCFG()
 }
 
 Loop::Loop(SgStatement* stmt) : do_irreg_opt(false)
-{    
+{
+    reduction_operation_list* rsl;
+    for (rsl = red_struct_list; rsl; rsl = rsl->next)
+    {
+        if (rsl->locvar)  //MAXLOC,MINLOC
+            redArrays.insert(rsl->locvar);
+    }
+
     lhs.clear(); rhs.clear(); unparsedLhs.clear(); unparsedRhs.clear();
     buildCFG();
 }

dvm/fdvm/trunk/include/acc_across_analyzer.h

@@ -104,6 +104,7 @@ private:
     SgStatement* loop_body;
     int dimension;
     std::map<SgSymbol*, Array*> arrays;
+    std::set<SgSymbol*> redArrays;
     int* acrossDims;
     int acrossType;
     std::vector<SgSymbol*> symbols;

dvm/fdvm/trunk/include/dvm.h

@@ -1377,6 +1377,7 @@ int TestOneGroupStatement(SgStatement *stmt);
 int TestOneGroupStatement(SgStatement *stmt);
 void DeclareUsedVars();
 void DeclareInternalPrivateVars();
+void DeclarePrivateVars();
 void DeclarePrivateVars(SgType *idxTypeInKernel);
 void DeclareArrayBases();
 void DeclareArrayCoeffsInKernel(SgType*);
@@ -1451,6 +1452,7 @@ SgExpression * DummyListForPrivateArrays(SgStatement *st_hedr);
 SgExpression * DummyListForPrivateArrays(SgStatement *st_hedr);
 SgExpression *CreatePrivateDummyList();
 char *PointerNameForPrivateArray(SgSymbol *symb);
+void GetMemoryForPrivateArrays(SgSymbol *private_first, SgSymbol *s_loop_ref, int nump, SgStatement *st_end, SgStatement *st_hedr, SgExpression *e_totalThreads);
 SgSymbol *LocRedVariableSymbolInKernel(reduction_operation_list *rsl);
 
 /*   acc_analyzer.cpp   */