draft of parallel regions merging pass

CMakeLists.txt

@@ -166,6 +166,8 @@ set(PARALLEL_REG src/ParallelizationRegions/ParRegions.cpp
 set(ARRAY_PROP src/ArrayConstantPropagation/propagation.cpp
 	src/ArrayConstantPropagation/propagation.h	
 )
+set(MERGE_COPY_ARRAYS src/ParallelizationRegions/parse_merge_dirs.cpp
+         src/ParallelizationRegions/parse_merge_dirs.h)
 
 set(TR_DEAD_CODE src/Transformations/DeadCodeRemoving/dead_code.cpp
         src/Transformations/DeadCodeRemoving/dead_code.h)
@@ -423,6 +425,7 @@ set(SOURCE_EXE
     ${LOOP_ANALYZER}
     ${TRANSFORMS}
     ${PARALLEL_REG}
+    ${MERGE_COPY_ARRAYS}
     ${PRIV}
     ${ARRAY_PROP}
     ${FDVM}
@@ -475,6 +478,7 @@ source_group (GraphCall FILES ${GR_CALL})
 source_group (GraphLoop FILES ${GR_LOOP})
 source_group (LoopAnalyzer FILES ${LOOP_ANALYZER})
 source_group (ParallelizationRegions FILES ${PARALLEL_REG})
+source_group (MergeCopyArrays FILES ${MERGE_COPY_ARRAYS})
 source_group (PrivateAnalyzer FILES ${PRIV})      
 source_group (ArrayConstantPropagation FILES ${ARRAY_PROP})      
 source_group (FDVM_Compiler FILES ${FDVM})

src/ArrayConstantPropagation/propagation.cpp

@@ -10,7 +10,6 @@
 using namespace std;
 
 static SgStatement* declPlace = NULL;
-static unordered_set<SgStatement*> changed;;
 
 static bool CheckConstIndexes(SgExpression* exp)
 {
@@ -37,13 +36,13 @@ static bool CheckConstIndexes(SgExpression* exp)
 
 static SgExpression* CreateVar(int& variableNumber, SgType* type)
 {
-    string varName = "tmp_prop_var";
+    string varName = "__tmp_prop_var";
     string name = varName + std::to_string(variableNumber) + "__";
     variableNumber++;
 
     SgSymbol* varSymbol = new SgSymbol(VARIABLE_NAME, name.c_str(), *type, *declPlace->controlParent());
 
-    const string commonBlockName = "propagation_common__";
+    const string commonBlockName = "__propagation_common__";
 
     SgStatement* funcStart = declPlace->controlParent();
     SgStatement* commonStat = NULL;
@@ -170,7 +169,8 @@ static SgExpression* CreateVar(int& variableNumber, SgType* type)
 
         commonList->setLhs(varList);
     }
-    return  new SgExpression(VAR_REF, NULL, NULL, varSymbol, type->copyPtr());
+
+    return new SgExpression(VAR_REF, NULL, NULL, varSymbol, type->copyPtr());
 }
 
 static void TransformRightPart(SgStatement* st, SgExpression* exp, unordered_map<string, SgExpression*>& arrayToVariable, int& variableNumber) 
@@ -227,8 +227,6 @@ static void TransformLeftPart(SgStatement* st, SgExpression* exp, unordered_map<
 {
     if (exp->symbol()->type()->variant() == T_STRING)
         return;
-    if (changed.find(st) != changed.end())
-        return;
     string expUnparsed = exp->unparse();
     if (arrayToVariable.find(expUnparsed) == arrayToVariable.end() && exp->symbol()->type()->baseType())
     {
@@ -236,62 +234,12 @@ static void TransformLeftPart(SgStatement* st, SgExpression* exp, unordered_map<
     }
     SgStatement* newStatement = new SgStatement(ASSIGN_STAT, NULL, NULL, arrayToVariable[expUnparsed]->copyPtr(), st->expr(1)->copyPtr(), NULL);
 
-   newStatement->setFileId(st->getFileId());
+    newStatement->setFileId(st->getFileId());
     newStatement->setProject(st->getProject()); 
 
     newStatement->setlineNumber(getNextNegativeLineNumber());
     newStatement->setLocalLineNumber(st->lineNumber());
     st->insertStmtBefore(*newStatement, *st->controlParent());
-    changed.insert(st);
-}
-
-static void TransformBorder(SgStatement* st, SgExpression* exp, unordered_map<string, SgExpression*>& arrayToVariable, int& variableNumber)
-{
-    SgStatement* firstStatement = declPlace->lexPrev();
-    st = st->lexPrev();
-    string array = exp->unparse();
-    arrayToVariable[array] = CreateVar(variableNumber, exp->symbol()->type()->baseType());
-    while (st != firstStatement)
-    {
-        if (st->variant() == ASSIGN_STAT && arrayToVariable.find(st->expr(0)->unparse()) != arrayToVariable.end())
-        {
-            if (st->expr(1))
-            {
-                TransformRightPart(st, st->expr(1), arrayToVariable, variableNumber);
-            }
-            if (st->expr(0) && st->expr(0)->variant() == ARRAY_REF && CheckConstIndexes(st->expr(0)->lhs()) && arrayToVariable.find(st->expr(0)->unparse()) != arrayToVariable.end())
-            {
-                TransformLeftPart(st, st->expr(0), arrayToVariable, variableNumber);
-            }
-        }
-        st = st->lexPrev();
-    }
-}
-
-static void CheckVariable(SgStatement* st, SgExpression* exp, unordered_map<string, SgExpression*>& arrayToVariable, int& variableNumber)
-{
-    SgStatement* firstStatement = declPlace->lexPrev();
-    st = st->lexPrev();
-    string varName = exp->unparse();
-    while (st != firstStatement)
-    {
-        if (st->variant() == ASSIGN_STAT && st->expr(0)->symbol() == exp->symbol())
-        {
-            TransformRightPart(st, st->expr(1), arrayToVariable, variableNumber);
-        }
-        if (st->variant() == ASSIGN_STAT && arrayToVariable.find(st->expr(0)->unparse()) != arrayToVariable.end())
-        {
-            if (st->expr(1))
-            {
-                TransformRightPart(st, st->expr(1), arrayToVariable, variableNumber);
-            }
-            if (st->expr(0) && st->expr(0)->variant() == ARRAY_REF && CheckConstIndexes(st->expr(0)->lhs()) && arrayToVariable.find(st->expr(0)->unparse()) != arrayToVariable.end())
-            {
-                TransformLeftPart(st, st->expr(0), arrayToVariable, variableNumber);
-            }
-        }
-        st = st->lexPrev();
-    }
 }
 
 void ArrayConstantPropagation(SgProject& project) 
@@ -314,29 +262,40 @@ void ArrayConstantPropagation(SgProject& project)
 
             for (; st != lastNode; st = st->lexNext())
             {
-                if (st->variant() == FOR_NODE)
+                if (st->variant() == ASSIGN_STAT)
+                {
+                    if (st->expr(1))
+                    {
+                        TransformRightPart(st, st->expr(1), arrayToVariable, variableNumber);
+                    }
+                    if (st->expr(0) && st->expr(0)->variant() == ARRAY_REF && CheckConstIndexes(st->expr(0)->lhs()))
+                    {
+                        TransformLeftPart(st, st->expr(0), arrayToVariable, variableNumber);
+                    }
+                }
+                else if (st->variant() == FOR_NODE)
                 {
                     SgExpression* lowerBound = st->expr(0)->lhs();
                     SgExpression* upperBound = st->expr(0)->rhs();
                     string lowerBoundUnparsed = lowerBound->unparse(), upperBoundUnparsed = upperBound->unparse();
                     if (upperBound->variant() == ARRAY_REF && upperBound->symbol()->type()->baseType() && CheckConstIndexes(upperBound->lhs()))
                     {
-                        TransformBorder(st, upperBound, arrayToVariable, variableNumber);
+                        if (arrayToVariable.find(upperBoundUnparsed) == arrayToVariable.end())
+                        {
+                            arrayToVariable[upperBoundUnparsed] = CreateVar(variableNumber, upperBound->symbol()->type()->baseType());
+                        }
                         st->expr(0)->setRhs(arrayToVariable[upperBoundUnparsed]->copyPtr());
                     }
-                    else if (upperBound->variant() == VAR_REF)
-                        CheckVariable(st, upperBound, arrayToVariable, variableNumber);
-
                     if (lowerBound->variant() == ARRAY_REF && lowerBound->symbol()->type()->baseType() && CheckConstIndexes(lowerBound->lhs()))
                     {
-                        TransformBorder(st, lowerBound, arrayToVariable, variableNumber);
+                        if (arrayToVariable.find(lowerBoundUnparsed) == arrayToVariable.end())
+                        {
+                            arrayToVariable[lowerBoundUnparsed] = CreateVar(variableNumber, lowerBound->symbol()->type()->baseType());
+                        }
                         st->expr(0)->setLhs(arrayToVariable[lowerBoundUnparsed]->copyPtr());
                     }
-                    else if (lowerBound->variant() == VAR_REF)
-                        CheckVariable(st, lowerBound, arrayToVariable, variableNumber);
                 }
             }
-            cout << file->functions(i)->unparse() << endl;
         }
     }
 }

src/ParallelizationRegions/parse_merge_dirs.cpp

@@ -0,0 +1,567 @@
+#include "parse_merge_dirs.h"
+
+#include <unordered_set>
+
+using std::map;
+using std::pair;
+using std::string;
+using std::unordered_map;
+using std::unordered_set;
+using std::vector;
+
+static void parseMergeDirective(const char *comment,
+                                vector<pair<string, string>> &parsed_mapping)
+{
+    while (comment)
+    {
+        auto *line_end = strchr(comment, '\n');
+
+        static const char prefix[] = "!!SPF TRANSFORM(MERGE_ARRAYS(";
+        static const auto compare_chars = sizeof(prefix) - 1;
+
+        if (strncasecmp(comment, prefix, compare_chars) == 0)
+        {
+            auto *pair_start = comment + compare_chars;
+            auto *comma = strchr(pair_start, ',');
+            if (comma)
+            {
+                auto *close_br = strchr(comma + 1, ')');
+                if (close_br)
+                {
+                    parsed_mapping.emplace_back(
+                        string(pair_start, comma - pair_start),
+                        string(comma + 1, close_br - comma - 1));
+                }
+            }
+        }
+
+        comment = line_end;
+
+        if (comment)
+            comment++;
+    }
+}
+
+static string getNonDefaultRegion(DIST::Array *a)
+{
+    string result;
+
+    if (!a)
+        return result;
+
+    for (const auto &reg_name : a->GetRegionsName())
+    {
+        if (reg_name != "default")
+        {
+            if (!result.empty())
+                printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+
+            result = reg_name;
+        }
+    }
+
+    return result;
+}
+
+static bool hasSameSizes(DIST::Array *a, DIST::Array *b)
+{
+    for (auto *array : {a, b})
+    {
+        for (const auto &p : array->GetSizes())
+        {
+            if (p.first < 0 || p.second < 0)
+                return false;
+        }
+    }
+
+    return a->GetSizes() == b->GetSizes() && a->GetTypeSize() == b->GetTypeSize();
+}
+
+static bool checkSimilarTemplates(vector<ParallelRegion *> &regions,
+                                  const unordered_map<string, string> &new_region_mapping)
+{
+    // new region -> old regions
+    unordered_map<string, unordered_set<string>> new_region_inverse_mapping;
+    for (const auto &p : new_region_mapping)
+        new_region_inverse_mapping[p.second].insert(p.first);
+
+    for (const auto &new_reg : new_region_inverse_mapping)
+    {
+        DIST::Array *template_array = nullptr;
+        string first_reg_name;
+
+        for (const auto &old_region_name : new_reg.second)
+        {
+            auto *old_reg = getRegionByName(regions, old_region_name);
+
+            if (!old_reg)
+                printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+
+            const auto &distr_rules = old_reg->GetDataDir().GetDistrRules();
+            if (distr_rules.size() != 1)
+                printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+
+            auto *current_template = distr_rules.front().first;
+
+            if (template_array)
+            {
+                if (!hasSameSizes(template_array, current_template))
+                {
+                    __spf_print(1, "Templates of %s and %s has different sizes\n",
+                                first_reg_name.c_str(),
+                                old_region_name.c_str());
+
+                    return false;
+                }
+                // else everything OK
+            }
+            else
+            {
+                template_array = current_template;
+                first_reg_name = old_region_name;
+            }
+        }
+    }
+
+    return true;
+}
+
+static bool hasSameAlignment(const std::unordered_set<const AlignRule *> &align_a,
+                             const std::unordered_set<const AlignRule *> &align_b)
+{
+    if (align_a.size() != 1 || align_b.size() != 1)
+        return false;
+
+    const auto *rule_a = *align_a.begin();
+    const auto *rule_b = *align_b.begin();
+
+    if (rule_a->alignRule != rule_b->alignRule)
+        return false;
+
+    return true;
+}
+
+static void printExpr(SgExpression *e, string pad)
+{
+    if (!e)
+        return;
+
+    __spf_print(1, "%s%d: %s\n", pad.c_str(), e->variant(), e->unparse());
+
+    printExpr(e->lhs(), pad + "  ");
+    printExpr(e->rhs(), pad + "  ");
+}
+
+static pair<vector<SgStatement *>, SgSymbol *> generateDeclaration(const string &array_name, const string &common_block_name,
+                                                                   const vector<pair<int, int>> &sizes, SgType *type, SgStatement *scope)
+{
+    auto *array_symbol = new SgSymbol(VARIABLE_NAME, array_name.c_str(), new SgType(T_ARRAY), scope);
+
+    auto *decl = new SgDeclarationStatement(VAR_DECL);
+
+    decl->setExpression(1, new SgTypeExp(*type));
+
+    SgExpression *subs = new SgExprListExp();
+
+    auto *array_ref = new SgArrayRefExp(*array_symbol, *subs);
+
+    for (int i = 0; i < sizes.size(); i++)
+    {
+        const auto &p = sizes[i];
+        auto *d = new SgExpression(DDOT, new SgValueExp(p.first), new SgValueExp(p.second));
+        subs->setLhs(d);
+
+        if (i + 1 < sizes.size())
+        {
+            subs->setRhs(new SgExprListExp());
+            subs = subs->rhs();
+        }
+    }
+
+    decl->setExpression(0, array_ref);
+
+    auto comm = new SgStatement(COMM_STAT);
+
+    comm->setExpression(0, new SgExpression(COMM_LIST,
+                                            new SgVarRefExp(array_symbol),
+                                            NULL,
+                                            new SgSymbol(COMMON_NAME, common_block_name.c_str())));
+
+    return {{decl, comm}, array_symbol};
+}
+
+SgSymbol *insertDeclIfNeeded(const string &array_name,
+                             const string &common_block_name,
+                             DIST::Array *example_array,
+                             FuncInfo *dest,
+                             unordered_map<FuncInfo *, unordered_map<string, SgSymbol *>> &inserted_arrays)
+{
+    if (SgFile::switchToFile(dest->fileName) == -1)
+        printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+
+    auto &by_func = inserted_arrays[dest];
+    auto it = by_func.find(array_name);
+
+    if (it != by_func.end())
+        return it->second;
+
+    SgStatement *st = dest->funcPointer;
+
+    auto *end = st->lastNodeOfStmt();
+
+    st = st->lexNext();
+
+    while (st != end && !isSgExecutableStatement(st))
+    {
+        st = st->lexNext();
+    }
+
+    auto generated = generateDeclaration(array_name, common_block_name,
+                                         example_array->GetSizes(),
+                                         SgTypeInt(), dest->funcPointer);
+    for (auto *new_stmt : generated.first)
+        st->insertStmtBefore(*new_stmt, *dest->funcPointer);
+
+    by_func[array_name] = generated.second;
+
+    return generated.second;
+}
+
+static pair<string, string> createNewArray(DIST::Array *example_array, const string &base_name,
+                                           const map<string, vector<FuncInfo *>> &allFuncInfo,
+                                           unordered_map<FuncInfo *, unordered_map<string, SgSymbol *>> &inserted_arrays)
+{
+    auto common_block_name = base_name + "_merge_cb";
+    auto array_name = base_name;
+
+    for (const auto &by_file : allFuncInfo)
+    {
+        for (auto *func_info : by_file.second)
+        {
+            if (func_info->isMain)
+            {
+                insertDeclIfNeeded(
+                    array_name,
+                    common_block_name,
+                    example_array,
+                    func_info,
+                    inserted_arrays);
+            }
+        }
+    }
+
+    return std::make_pair(array_name, common_block_name);
+}
+
+static void replaceArrayRec(SgExpression *e,
+                            const unordered_set<string> &arrays_to_replace,
+                            SgSymbol **func_symbol_hint,
+                            const pair<string, string> &replace_by,
+                            DIST::Array *example_array,
+                            FuncInfo *func,
+                            unordered_map<FuncInfo *, unordered_map<string, SgSymbol *>> &inserted_arrays)
+{
+    if (!e)
+        return;
+
+    if (isArrayRef(e) && arrays_to_replace.find(e->symbol()->identifier()) != arrays_to_replace.end())
+    {
+        if (!(*func_symbol_hint))
+        {
+            *func_symbol_hint = insertDeclIfNeeded(
+                replace_by.first, replace_by.second,
+                example_array,
+                func,
+                inserted_arrays);
+        }
+
+        e->setSymbol(*func_symbol_hint);
+    }
+
+    replaceArrayRec(
+        e->lhs(),
+        arrays_to_replace,
+        func_symbol_hint,
+        replace_by,
+        example_array,
+        func,
+        inserted_arrays);
+
+    replaceArrayRec(
+        e->rhs(),
+        arrays_to_replace,
+        func_symbol_hint,
+        replace_by,
+        example_array,
+        func,
+        inserted_arrays);
+}
+
+static void replaceRegion(SgStatement* st, const unordered_map<string, string> &new_region_mapping)
+{
+    if (!st)
+        return;
+
+    if(isSPF_stat(st) && st->variant() == SPF_PARALLEL_REG_DIR)
+    {
+        auto it = new_region_mapping.find(st->symbol()->identifier());
+
+        if (it != new_region_mapping.end())
+            st->setSymbol(*(new SgSymbol(CONST_NAME, it->second.c_str())));
+    }
+}
+
+void mergeCopyArrays(vector<ParallelRegion *> &regions, const map<string, vector<FuncInfo *>> &allFuncInfo)
+{
+    for (const auto *region : regions)
+    {
+        __spf_print(1, "region %s\n", region->GetName().c_str());
+
+        const auto &dirs = region->GetDataDir();
+
+        __spf_print(1, "  distr rules: %d\n", dirs.distrRules.size());
+
+        const auto &currentVariant = region->GetCurrentVariant();
+
+        int distr_idx = 0;
+        for (const auto &distr : dirs.distrRules)
+        {
+            const auto &dist_rule = distr.second.back().distRule;
+
+            string sizes;
+            for (const auto &p : distr.first->GetSizes())
+            {
+                if (!sizes.empty())
+                    sizes.push_back(',');
+
+                sizes += std::to_string(p.first) + ":" + std::to_string(p.second);
+            }
+
+            __spf_print(1, "      DIST %s(%s)", distr.first->GetName().c_str(), sizes.c_str());
+            for (const auto &dim : dist_rule)
+                __spf_print(1, " %c", dim == dist::BLOCK ? 'B' : '*');
+
+            __spf_print(1, "\n");
+            distr_idx++;
+        }
+
+        __spf_print(1, "  align rules: %d\n", dirs.alignRules.size());
+
+        for (const auto &align : dirs.alignRules)
+        {
+            string sub_a, sub_b;
+
+            int i = 0;
+            for (const auto coefs : align.alignRule)
+            {
+                if (!sub_a.empty())
+                    sub_a.push_back(',');
+
+                sub_a += std::to_string(coefs.first) + "*i" +
+                         std::to_string(i) + "+" + std::to_string(coefs.second);
+
+                i++;
+            }
+
+            for (const auto coefs : align.alignRuleWith)
+            {
+                if (!sub_b.empty())
+                    sub_b.push_back(',');
+
+                sub_b += std::to_string(coefs.second.first) + "*i" +
+                         std::to_string(coefs.first) + "+" + std::to_string(coefs.second.second);
+            }
+
+            __spf_print(1, "      ALIGN %s(%s) WITH %s(%s)\n",
+                        align.alignArray->GetName().c_str(), sub_a.c_str(),
+                        align.alignWith->GetName().c_str(), sub_b.c_str());
+        }
+    }
+
+    // parse directives
+
+    // new array name -> current arrays
+    unordered_map<string, unordered_set<DIST::Array *>> arrays_to_merge;
+    unordered_map<DIST::Array *, unordered_set<const AlignRule *>> array_alignment;
+
+    for (const auto &by_file : allFuncInfo)
+    {
+        const auto current_file_name = by_file.first;
+
+        if (SgFile::switchToFile(current_file_name) == -1)
+            printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+
+        for (auto *func_info : by_file.second)
+        {
+            SgStatement *curr_stmt = func_info->funcPointer;
+            if (!curr_stmt)
+                continue;
+
+            auto *stmt_end = curr_stmt->lastDeclaration();
+            if (!stmt_end)
+                continue;
+
+            stmt_end = stmt_end->lexNext();
+
+            for (; curr_stmt && curr_stmt != stmt_end; curr_stmt = curr_stmt->lexNext())
+            {
+                if (curr_stmt->comments())
+                {
+                    vector<pair<string, string>> parsed_mapping;
+                    parseMergeDirective(curr_stmt->comments(), parsed_mapping);
+
+                    for (const auto &p : parsed_mapping)
+                    {
+                        auto *found_array = getArrayFromDeclarated(curr_stmt, p.first);
+                        if (found_array)
+                        {
+                            arrays_to_merge[p.second].insert(found_array);
+                            array_alignment[found_array] = {};
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    // find alignment rules for array
+
+    for (const auto *region : regions)
+    {
+        const auto &dirs = region->GetDataDir();
+
+        for (const auto &align : dirs.alignRules)
+        {
+            auto it = array_alignment.find(align.alignArray);
+
+            if (it != array_alignment.end())
+                it->second.insert(&align);
+        }
+    }
+
+    // old region -> new region
+    unordered_map<string, string> new_region_mapping;
+
+    // new array -> new region
+    unordered_map<string, string> arrays_new_region_mapping;
+    vector<string> created_region_names;
+
+    for (const auto &by_new_array : arrays_to_merge)
+    {
+        string new_region_name;
+        for (auto *current_array : by_new_array.second)
+        {
+            auto current_array_region = getNonDefaultRegion(current_array);
+            auto it = new_region_mapping.find(current_array_region);
+            if (it != new_region_mapping.end())
+            {
+                if (new_region_name.empty())
+                    new_region_name = it->second;
+                else if (new_region_name != it->second)
+                    printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+            }
+        }
+
+        if (new_region_name.empty())
+        {
+            new_region_name = "merged_reg_" + std::to_string(created_region_names.size());
+            created_region_names.push_back(new_region_name);
+        }
+
+        for (auto *current_array : by_new_array.second)
+        {
+            auto current_array_region = getNonDefaultRegion(current_array);
+            new_region_mapping[current_array_region] = new_region_name;
+        }
+
+        arrays_new_region_mapping[by_new_array.first] = new_region_name;
+    }
+
+    if (!checkSimilarTemplates(regions, new_region_mapping))
+        printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+
+    unordered_map<FuncInfo *, unordered_map<string, SgSymbol *>> inserted_arrays;
+
+    for (const auto &by_dest_array : arrays_to_merge)
+    {
+        const auto &copy_arrays = by_dest_array.second;
+
+        if (copy_arrays.empty())
+            printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+
+        auto *first_element = *copy_arrays.begin();
+        auto first_elem_rules_it = array_alignment.find(first_element);
+
+        if (first_elem_rules_it == array_alignment.end())
+            continue;
+
+        const auto &first_elem_rules = first_elem_rules_it->second;
+
+        for (auto *array_to_merge : copy_arrays)
+        {
+            auto array_rules_it = array_alignment.find(array_to_merge);
+
+            if (array_rules_it == array_alignment.end())
+                printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+
+            const auto &array_rules = array_rules_it->second;
+
+            if (!hasSameSizes(array_to_merge, first_element) || !hasSameAlignment(first_elem_rules, array_rules))
+            {
+                __spf_print(1, "Arrays %s and %s has different sizes or align rules\n",
+                            array_to_merge->GetName().c_str(),
+                            first_element->GetName().c_str());
+
+                printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+            }
+        }
+
+        __spf_print(1, "merge into %s (%s):\n", by_dest_array.first.c_str(), arrays_new_region_mapping[by_dest_array.first].c_str());
+        for (auto *array_to_merge : copy_arrays)
+            __spf_print(1, "%s\n", array_to_merge->GetName().c_str());
+
+        auto created_array_info = createNewArray(first_element, by_dest_array.first, allFuncInfo, inserted_arrays);
+
+        unordered_set<string> arrays_to_replace;
+        for (auto *array_to_merge : copy_arrays)
+            arrays_to_replace.insert(array_to_merge->GetShortName());
+
+        for (const auto &by_file : allFuncInfo)
+        {
+            if (SgFile::switchToFile(by_file.first) == -1)
+                printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+
+            for (auto *func_info : by_file.second)
+            {
+                SgSymbol *func_symbol_hint = nullptr;
+                SgStatement *st = func_info->funcPointer;
+
+                auto *func_end = st->lastNodeOfStmt();
+
+                st = st->lexNext();
+
+                while (st && !isSgExecutableStatement(st) && st != func_end)
+                    st = st->lexNext();
+
+                while (st && st != func_end)
+                {
+                    for (int i = 0; i < 3; i++)
+                    {
+                        replaceArrayRec(
+                            st->expr(i),
+                            arrays_to_replace,
+                            &func_symbol_hint,
+                            created_array_info,
+                            first_element,
+                            func_info,
+                            inserted_arrays);
+                    }
+
+                    replaceRegion(st, new_region_mapping);
+
+                    st = st->lexNext();
+                }
+            }
+        }
+    }
+}

src/ParallelizationRegions/parse_merge_dirs.h

@@ -0,0 +1,8 @@
+#include "../GraphCall/graph_calls.h"
+#include "ParRegions.h"
+
+#include <string>
+#include <vector>
+#include <map>
+
+void mergeCopyArrays(std::vector<ParallelRegion*> &regions, const std::map<std::string, std::vector<FuncInfo*>> &allFuncInfo);

src/PrivateAnalyzer/private_arrays_search.cpp

@@ -67,14 +67,12 @@ static void Collapse(Region* region)
         }
     }
 
-    ArrayAccessingIndexes useUnionB;
+    ArrayAccessingIndexes useUnion;
     for (auto& byBlock : region->getBasickBlocks())
-        for (auto& [arrayName, _] : byBlock->array_use)
+        for (auto& [arrayName, arrayRanges] : byBlock->array_use)
-            useUnionB[arrayName] = useUnionB[arrayName].Union(byBlock->array_use[arrayName]);
+            useUnion[arrayName] = useUnion[arrayName].Union(byBlock->array_use[arrayName]);
-
-    for (auto& [arrayName, _] : useUnionB)
-        region->array_priv[arrayName] = useUnionB[arrayName].Diff(region->array_use[arrayName]);
 
+    region->array_priv = region->array_use;
     
     for (Region* prevBlock : region->getHeader()->getPrevRegions()) 
         prevBlock->replaceInNextRegions(region, region->getHeader());

src/PrivateAnalyzer/region.cpp

@@ -9,7 +9,6 @@
 #include "range_structures.h"
 #include "region.h"
 
-#include "..\Transformations\ExpressionSubstitution\expr_transform.h"
 #include "SgUtils.h"
 
 using namespace std;
@@ -132,7 +131,7 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
         {
             vector<SAPFOR::Argument*> index_vars;
             vector<int> refPos;
-            string array_name = instruction->getInstruction()->getArg1()->getValue();
+            string array_name = instruction->getInstruction()->getArg1()->getValue();            
 
             int j = i - 1;
             while (j >= 0 && instructions[j]->getInstruction()->getOperation() == SAPFOR::CFG_OP::REF)
@@ -149,27 +148,13 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
             auto* ref = isSgArrayRefExp(instruction->getInstruction()->getExpression());
             int fillCount = 0;
 
-            vector<pair<int, int>> coefsForDims;
+            while (!index_vars.empty() && !refPos.empty())
-            int subs = ref->numberOfSubscripts();
-            for (int i = 0; ref && i < ref->numberOfSubscripts(); ++i)
-            {
-                const vector<int*>& coefs = getAttributes<SgExpression*, int*>(ref->subscript(i), set<int>{ INT_VAL });
-                if (coefs.size() == 1)
-                {
-                    const pair<int, int> coef(coefs[0][0], coefs[0][1]);
-                    coefsForDims.push_back(coef);
-                }
-
-            }
-            coefsForDims = {coefsForDims.rbegin(), coefsForDims.rend()};
-
-            while (!index_vars.empty() && !refPos.empty() && !coefsForDims.empty())
             {
                 auto var = index_vars.back();
                 int currentVarPos = refPos.back();
                 ArrayDimension current_dim;
                 if (var->getType() == SAPFOR::CFG_ARG_TYPE::CONST)
-                    current_dim = { stoul(var->getValue()), 1, 1, ref };
+                    current_dim = { stoul(var->getValue()), 1, 1, ref};
                 else
                 {
                     string name, full_name = var->getValue();
@@ -195,7 +180,7 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
                             return -1;
                     }
 
-                    uint64_t start = coefsForDims.back().second * currentLoop->startVal + coefsForDims.back().first;
+                    uint64_t start = currentLoop->startVal;
                     uint64_t step = currentLoop->stepVal;
                     uint64_t iters = currentLoop->calculatedCountOfIters;
                     current_dim = { start, step, iters, ref };
@@ -208,7 +193,6 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
                 }
                 index_vars.pop_back();
                 refPos.pop_back();
-                coefsForDims.pop_back();
             }
 
             if (fillCount == accessPoint.size())

src/Sapfor.cpp

@@ -23,6 +23,7 @@
 #include "ParallelizationRegions/ParRegions_func.h"
 #include "ParallelizationRegions/resolve_par_reg_conflicts.h"
 #include "ParallelizationRegions/expand_extract_reg.h"
+#include "ParallelizationRegions/parse_merge_dirs.h"
 
 #include "Distribution/Distribution.h"
 #include "Distribution/GraphCSR.h"
@@ -281,7 +282,7 @@ static string unparseProjectIfNeed(SgFile* file, const int curr_regime, const bo
                 if (isSPF_stat(st)) // except sapfor parallel regions and if attributes dont move
                     if (st->variant() != SPF_PARALLEL_REG_DIR && st->variant() != SPF_END_PARALLEL_REG_DIR)
                         if (insertedPrivates.find(st) == insertedPrivates.end())
-                            toDel.push_back(st);
+                        toDel.push_back(st);
 
             for (auto& elem : toDel)
                 elem->deleteStmt();
@@ -1918,6 +1919,10 @@ static bool runAnalysis(SgProject &project, const int curr_regime, const bool ne
         transformAssumedSizeParameters(allFuncInfo);
     else if (curr_regime == FIND_PRIVATE_ARRAYS_ANALYSIS)
         FindPrivateArrays(loopGraph, fullIR, insertedPrivates);
+    else if (curr_regime == MERGE_COPY_ARRAYS)
+    {
+        mergeCopyArrays(parallelRegions, allFuncInfo);
+    }
 
     else if (curr_regime == ARRAY_PROPAGATION)
         ArrayConstantPropagation(project);
@@ -2374,6 +2379,7 @@ void runPass(const int curr_regime, const char *proj_name, const char *folderNam
     case FIX_COMMON_BLOCKS:
     case TEST_PASS:
     case SET_IMPLICIT_NONE:
+    case MERGE_COPY_ARRAYS:
         runAnalysis(*project, curr_regime, false);
     case SUBST_EXPR_RD_AND_UNPARSE:
     case SUBST_EXPR_AND_UNPARSE:

src/Sapfor.h

@@ -88,6 +88,8 @@ enum passes {
     REMOVE_DVM_INTERVALS,
     VERIFY_DVM_DIRS,
 
+    MERGE_COPY_ARRAYS,
+
     REMOVE_DIST_ARRAYS_FROM_IO,
 
     SUBST_EXPR,
@@ -271,6 +273,7 @@ static void setPassValues()
     passNames[VERIFY_DVM_DIRS] = "VERIFY_DVM_DIRS";
     passNames[REMOVE_DVM_DIRS_TO_COMMENTS] = "REMOVE_DVM_DIRS_TO_COMMENTS";
     passNames[REMOVE_SPF_DIRS] = "REMOVE_SPF_DIRS";
+    passNames[MERGE_COPY_ARRAYS] = "MERGE_COPY_ARRAYS";
     passNames[REMOVE_DIST_ARRAYS_FROM_IO] = "REMOVE_DIST_ARRAYS_FROM_IO";
     passNames[SUBST_EXPR] = "SUBST_EXPR";
     passNames[SUBST_EXPR_RD] = "SUBST_EXPR_RD";

src/Utils/PassManager.h

@@ -319,6 +319,8 @@ void InitPassesDependencies(map<passes, vector<passes>> &passDepsIn, set<passes>
     list({ CALL_GRAPH2, CALL_GRAPH, BUILD_IR, LOOP_GRAPH, LOOP_ANALYZER_DATA_DIST_S2 }) <= Pass(FIND_PRIVATE_ARRAYS_ANALYSIS);
     list({ FIND_PRIVATE_ARRAYS_ANALYSIS, CONVERT_LOOP_TO_ASSIGN, RESTORE_LOOP_FROM_ASSIGN, REVERT_SUBST_EXPR_RD }) <= Pass(FIND_PRIVATE_ARRAYS);
 
+    Pass(CREATE_TEMPLATE_LINKS) <= Pass(MERGE_COPY_ARRAYS);
+
     passesIgnoreStateDone.insert({ CREATE_PARALLEL_DIRS, INSERT_PARALLEL_DIRS, INSERT_SHADOW_DIRS, EXTRACT_PARALLEL_DIRS, 
                                    EXTRACT_SHADOW_DIRS, CREATE_REMOTES, UNPARSE_FILE, REMOVE_AND_CALC_SHADOW,
                                    REVERSE_CREATED_NESTED_LOOPS, PREDICT_SCHEME, CALCULATE_STATS_SCHEME, REVERT_SPF_DIRS, CLEAR_SPF_DIRS, TRANSFORM_SHADOW_IF_FULL,