draft of parallel regions merging pass

CMakeLists.txt

@@ -163,6 +163,12 @@ set(PARALLEL_REG src/ParallelizationRegions/ParRegions.cpp
         src/ParallelizationRegions/resolve_par_reg_conflicts.cpp
         src/ParallelizationRegions/resolve_par_reg_conflicts.h)
 
+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)
 set(TR_CP src/Transformations/CheckPoints/checkpoints.cpp
@@ -419,7 +425,9 @@ set(SOURCE_EXE
     ${LOOP_ANALYZER}
     ${TRANSFORMS}
     ${PARALLEL_REG}
+    ${MERGE_COPY_ARRAYS}
     ${PRIV}
+    ${ARRAY_PROP}
     ${FDVM}
     ${OMEGA}
     ${UTILS}
@@ -470,7 +478,9 @@ 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})
 source_group (SageExtension FILES ${OMEGA})
 source_group (Utils FILES ${UTILS})

src/ArrayConstantPropagation/propagation.cpp

@@ -0,0 +1,301 @@
+#include "propagation.h"
+
+#include "../Utils/SgUtils.h"
+
+#include <iostream>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+using namespace std;
+
+static SgStatement* declPlace = NULL;
+
+static bool CheckConstIndexes(SgExpression* exp)
+{
+    if (!exp)
+    {
+        return true;
+    }
+    SgExpression* lhs = exp->lhs();
+    SgExpression* rhs = exp->rhs();
+    do
+    {
+        if (lhs->variant() != INT_VAL)
+        {
+            return false;
+        }
+        if (rhs)
+        {
+            lhs = rhs->lhs();
+            rhs = rhs->rhs();
+        }
+    } while (rhs);
+    return true;
+}
+
+static SgExpression* CreateVar(int& variableNumber, SgType* type)
+{
+    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__";
+
+    SgStatement* funcStart = declPlace->controlParent();
+    SgStatement* commonStat = NULL;
+    SgExpression* commonList = NULL;
+
+    SgStatement* funcEnd = funcStart->lastNodeOfStmt();
+    SgStatement* current = funcStart->lexNext();
+
+    while (current != funcEnd && current)
+    {
+        if (current->variant() == COMM_STAT)
+        {
+            for (SgExpression* exp = current->expr(0); exp; exp = exp->rhs())
+            {
+                if (exp->variant() == COMM_LIST)
+                {
+                    string existingName = exp->symbol() ?
+                        string(exp->symbol()->identifier()) :
+                        string("spf_unnamed");
+                    if (existingName == commonBlockName)
+                    {
+                        commonStat = current;
+                        commonList = exp;
+                        break;
+                    }
+                }
+            }
+            if (commonStat)
+                break;
+        }
+        current = current->lexNext();
+    }
+
+    vector<SgExpression*> varRefs;
+    if (commonList)
+    {
+        SgExpression* varList = commonList->lhs();
+        if (varList)
+        {
+            auto extractSymbol = [](SgExpression* exp) -> SgSymbol* {
+                if (!exp) 
+                    return NULL;
+                if (exp->symbol()) 
+                    return exp->symbol();
+                if (exp->lhs() && exp->lhs()->symbol())
+                    return exp->lhs()->symbol();
+                return NULL;
+                };
+            if (varList->variant() == EXPR_LIST)
+            {
+                for (SgExpression* exp = varList; exp; exp = exp->rhs())
+                {
+                    SgExpression* varExp = exp->lhs();
+                    SgSymbol* sym = extractSymbol(varExp);
+                    if (sym)
+                    {
+                        varRefs.push_back(new SgVarRefExp(sym));
+                    }
+                }
+            }
+            else
+            {
+                for (SgExpression* varExp = varList; varExp; varExp = varExp->rhs())
+                {
+                    SgSymbol* sym = extractSymbol(varExp);
+                    if (sym)
+                    {
+                        varRefs.push_back(new SgVarRefExp(sym));
+                    }
+                }
+            }
+        }
+    }
+
+    if (!commonList)
+    {
+        current = funcStart->lexNext();
+        while (current != funcEnd && current)
+        {
+            if (current->variant() == COMM_STAT)
+            {
+                commonStat = current;
+                break;
+            }
+            current = current->lexNext();
+        }
+
+        SgSymbol* commonSymbol = new SgSymbol(COMMON_NAME, commonBlockName.c_str());
+        commonList = new SgExpression(COMM_LIST, NULL, NULL, commonSymbol);
+
+        if (commonStat)
+        {
+            SgExpression* lastCommList = commonStat->expr(0);
+            if (lastCommList)
+            {
+                while (lastCommList->rhs())
+                    lastCommList = lastCommList->rhs();
+                lastCommList->setRhs(commonList);
+            }
+            else
+            {
+                commonStat->setExpression(0, commonList);
+            }
+        }
+        else
+        {
+            commonStat = new SgStatement(COMM_STAT);
+            commonStat->setFileName(declPlace->fileName());
+            commonStat->setFileId(declPlace->getFileId());
+            commonStat->setProject(declPlace->getProject());
+            commonStat->setlineNumber(getNextNegativeLineNumber());
+            commonStat->setExpression(0, commonList);
+
+            declPlace->insertStmtBefore(*commonStat, *declPlace->controlParent());
+        }
+
+    }
+    varRefs.push_back(new SgVarRefExp(varSymbol));
+
+    if (varRefs.size() > 0)
+    {
+        std::reverse(varRefs.begin(), varRefs.end());
+        SgExpression* varList = makeExprList(varRefs, false);
+
+        commonList->setLhs(varList);
+    }
+
+    return new SgExpression(VAR_REF, NULL, NULL, varSymbol, type->copyPtr());
+}
+
+static void TransformRightPart(SgStatement* st, SgExpression* exp, unordered_map<string, SgExpression*>& arrayToVariable, int& variableNumber) 
+{
+    if (!exp)
+    {
+        return;
+    }
+
+    vector<SgExpression*> subnodes = { exp->lhs(), exp->rhs() };
+
+    string expUnparsed;
+    SgExpression* toAdd = NULL;
+    if (exp->variant() == ARRAY_REF && CheckConstIndexes(exp->lhs()))
+    {
+        cout << st->unparse() << endl;
+        if (arrayToVariable.find(expUnparsed) == arrayToVariable.end() && exp->symbol()->type()->baseType())
+        {
+            arrayToVariable[expUnparsed] = CreateVar(variableNumber, exp->symbol()->type()->baseType());  
+        }
+        st->setExpression(1, arrayToVariable[expUnparsed]->copyPtr());
+        return;
+    }
+    for (int i = 0; i < 2; i++)
+    {
+        if (subnodes[i] && subnodes[i]->variant() == ARRAY_REF && subnodes[i]->symbol()->type()->baseType() && CheckConstIndexes(subnodes[i]->lhs()))
+        {
+            expUnparsed = subnodes[i]->unparse();
+            if (arrayToVariable.find(expUnparsed) == arrayToVariable.end())
+            {                
+                arrayToVariable[expUnparsed] = CreateVar(variableNumber, subnodes[i]->symbol()->type()->baseType());;
+            }
+            toAdd = arrayToVariable[expUnparsed]->copyPtr();
+            if (toAdd)
+            {
+                if (i == 0)
+                {
+                    exp->setLhs(toAdd);
+                }
+                else
+                {
+                    exp->setRhs(toAdd);
+                }
+            }
+        }
+        else
+        {
+            TransformRightPart(st, subnodes[i], arrayToVariable, variableNumber);
+        }
+    }
+}
+
+static void TransformLeftPart(SgStatement* st, SgExpression* exp, unordered_map<string, SgExpression*>& arrayToVariable, int& variableNumber)
+{
+    if (exp->symbol()->type()->variant() == T_STRING)
+        return;
+    string expUnparsed = exp->unparse();
+    if (arrayToVariable.find(expUnparsed) == arrayToVariable.end() && exp->symbol()->type()->baseType())
+    {
+        arrayToVariable[expUnparsed] = CreateVar(variableNumber, exp->symbol()->type()->baseType());
+    }
+    SgStatement* newStatement = new SgStatement(ASSIGN_STAT, NULL, NULL, arrayToVariable[expUnparsed]->copyPtr(), st->expr(1)->copyPtr(), NULL);
+
+    newStatement->setFileId(st->getFileId());
+    newStatement->setProject(st->getProject()); 
+
+    newStatement->setlineNumber(getNextNegativeLineNumber());
+    newStatement->setLocalLineNumber(st->lineNumber());
+    st->insertStmtBefore(*newStatement, *st->controlParent());
+}
+
+void ArrayConstantPropagation(SgProject& project) 
+{
+    unordered_map<string, SgExpression*> arrayToVariable;
+    int variableNumber = 0;
+    for (int i = 0; i < project.numberOfFiles(); i++) 
+    {
+        SgFile* file = &(project.file(i));
+
+        if (!file)
+            continue;
+
+        const int funcNum = file->numberOfFunctions();
+        for (int i = 0; i < funcNum; ++i)
+        {
+            SgStatement* st = file->functions(i);
+            declPlace = st->lexNext();
+            SgStatement* lastNode = st->lastNodeOfStmt();
+
+            for (; st != lastNode; st = st->lexNext())
+            {
+                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()))
+                    {
+                        if (arrayToVariable.find(upperBoundUnparsed) == arrayToVariable.end())
+                        {
+                            arrayToVariable[upperBoundUnparsed] = CreateVar(variableNumber, upperBound->symbol()->type()->baseType());
+                        }
+                        st->expr(0)->setRhs(arrayToVariable[upperBoundUnparsed]->copyPtr());
+                    }
+                    if (lowerBound->variant() == ARRAY_REF && lowerBound->symbol()->type()->baseType() && CheckConstIndexes(lowerBound->lhs()))
+                    {
+                        if (arrayToVariable.find(lowerBoundUnparsed) == arrayToVariable.end())
+                        {
+                            arrayToVariable[lowerBoundUnparsed] = CreateVar(variableNumber, lowerBound->symbol()->type()->baseType());
+                        }
+                        st->expr(0)->setLhs(arrayToVariable[lowerBoundUnparsed]->copyPtr());
+                    }
+                }
+            }
+        }
+    }
+}

src/ArrayConstantPropagation/propagation.h

@@ -0,0 +1,4 @@
+#pragma once
+#include "../Utils/SgUtils.h"
+
+void ArrayConstantPropagation(SgProject& project);

src/Distribution/CreateDistributionDirs.cpp

@@ -274,7 +274,7 @@ static void convertTrees(const map<DIST::Array*, int> &treesIn, map<int, vector<
 static DIST::Array* findBestInEqual(vector<DIST::Array*> &arrays, DIST::GraphCSR<int, double, attrType> &reducedG, DIST::Arrays<int> &allArrays)
 {
     DIST::Array *retVal = NULL;
-    vector<vector<attrType>> coefsByDims;
+    vector<vector<attrType>> coeffsByDims;
     for (auto &array : arrays)
     {
         vector<int> verts;
@@ -285,7 +285,7 @@ static DIST::Array* findBestInEqual(vector<DIST::Array*> &arrays, DIST::GraphCSR
         {
             retVal = array;
             for (auto &V : verts)
-                coefsByDims.push_back(reducedG.GetAllAttributes(V));
+                coeffsByDims.push_back(reducedG.GetAllAttributes(V));
         }
         else
         {
@@ -294,11 +294,11 @@ static DIST::Array* findBestInEqual(vector<DIST::Array*> &arrays, DIST::GraphCSR
                 toCmp.push_back(reducedG.GetAllAttributes(V));
             for (int z = 0; z < toCmp.size(); ++z)
             {
-                if (toCmp[z].size() && coefsByDims[z].size())
+                if (toCmp[z].size() && coeffsByDims[z].size())
                 {
-                    if (toCmp[z].back().first.first > coefsByDims[z].back().first.first)
+                    if (toCmp[z].back().first.first > coeffsByDims[z].back().first.first)
                     {
-                        coefsByDims = toCmp;
+                        coeffsByDims = toCmp;
                         retVal = array;
                         break;
                     }

src/Distribution/DvmhDirective.cpp

@@ -45,10 +45,10 @@ static bool findArrayRefAndCheck(SgExpression *ex, const DIST::Array* currArray,
                 int countOfShadows = 0;
                 for (int i = 0; i < ref->numberOfSubscripts(); ++i)
                 {
-                    const vector<int*> &coefs = getAttributes<SgExpression*, int*>(ref->subscript(i), set<int>{ INT_VAL });
+                    const vector<int*> &coeffs = getAttributes<SgExpression*, int*>(ref->subscript(i), set<int>{ INT_VAL });
-                    if (coefs.size() == 1)
+                    if (coeffs.size() == 1)
                     {
-                        const pair<int, int> coef(coefs[0][0], coefs[0][1]);
+                        const pair<int, int> coef(coeffs[0][0], coeffs[0][1]);
                         auto it = shiftsByAccess[i].find(coef);
                         if (it != shiftsByAccess[i].end())
                             if (it->second != 0)

src/Distribution/DvmhDirectiveBase.cpp

@@ -364,11 +364,11 @@ static inline string calculateShifts(DIST::GraphCSR<int, double, attrType> &redu
             {
                 if (sharedMemoryParallelization)
                 {
-                    for (auto& coefs : currReadOp->first[k].coefficients)
+                    for (auto& coeffs : currReadOp->first[k].coefficients)
                     {
-                        auto currAccess = coefs.first;
+                        auto currAccess = coeffs.first;
 
-                        const int currShift = coefs.first.second;
+                        const int currShift = coeffs.first.second;
 
                         auto itFound = shiftsByAccess[k].find(currAccess);
                         if (itFound == shiftsByAccess[k].end())
@@ -393,9 +393,9 @@ static inline string calculateShifts(DIST::GraphCSR<int, double, attrType> &redu
                             int minShift = 9999999;
                             int maxShift = -9999999;
 
-                            for (auto &coefs : currReadOp->first[k].coefficients)
+                            for (auto &coeffs : currReadOp->first[k].coefficients)
                             {
-                                auto currAccess = coefs.first;
+                                auto currAccess = coeffs.first;
                                 auto result = DIST::Fx(currAccess, currRuleShadow);
 
                                 if (result.first == loopRule.first)
@@ -417,7 +417,7 @@ static inline string calculateShifts(DIST::GraphCSR<int, double, attrType> &redu
                                     auto it = remoteRegularReads.find(calcForArray);
                                     if (it == remoteRegularReads.end())
                                         it = remoteRegularReads.insert(it, make_pair(calcForArray, vector<ArrayOp>(calcForArray->GetDimSize())));
-                                    it->second[k].coefficients.insert(coefs);
+                                    it->second[k].coefficients.insert(coeffs);
                                 }
                             }
 

src/GraphLoop/graph_loops_base.cpp

@@ -1137,9 +1137,9 @@ static bool isMapped(const vector<ArrayOp> &allOps)
     bool mapped = false;
     for (auto &ops : allOps)
     {
-        for (auto &coefs : ops.coefficients)
+        for (auto &coeffs : ops.coefficients)
         {
-            if (coefs.first.first != 0)
+            if (coeffs.first.first != 0)
             {
                 mapped = true;
                 break;

src/LoopAnalyzer/loop_analyzer.cpp

@@ -247,7 +247,7 @@ static vector<int> matchSubscriptToLoopSymbols(const vector<SgForStmt*> &parentL
         }
     }
 
-    pair<int, int> coefs = pair<int, int>(0, 0);
+    pair<int, int> coeffs = pair<int, int>(0, 0);
     // more than one loop symbol in subscription
     if (countOfSymbols > 1)
     {
@@ -326,16 +326,16 @@ static vector<int> matchSubscriptToLoopSymbols(const vector<SgForStmt*> &parentL
         {
             if (subscr->symbol()->id() == (parentLoops[position]->doName())->id())
             {
-                coefs.first = 1;
+                coeffs.first = 1;
                 needToCacl = false;
             }
         }
 
         if (needToCacl)
-            getCoefsOfSubscript(coefs, subscr, parentLoops[position]->doName());
+            getCoefsOfSubscript(coeffs, subscr, parentLoops[position]->doName());
-        __spf_print(PRINT_ARRAY_ARCS, " <%d %d> ", coefs.first, coefs.second);
+        __spf_print(PRINT_ARRAY_ARCS, " <%d %d> ", coeffs.first, coeffs.second);
 
-        if (coefs.first == 0) // && coefs.second == 0)
+        if (coeffs.first == 0) // && coeffs.second == 0)
         {
             if (currRegime == REMOTE_ACC)
             {
@@ -346,7 +346,7 @@ static vector<int> matchSubscriptToLoopSymbols(const vector<SgForStmt*> &parentL
             {                
                 const pair<bool, string> &arrayRefString = constructArrayRefForPrint(arrayRef, dimNum, origSubscr);
                 __spf_print(1, "WARN: can not calculate index expression for array ref '%s' at line %d\n", arrayRefString.second.c_str(), currLine);
-                addInfoToVectors(loopInfo, parentLoops[position], currOrigArrayS, dimNum, coefs, UNREC_OP, numOfSubscriptions, currentW);
+                addInfoToVectors(loopInfo, parentLoops[position], currOrigArrayS, dimNum, coeffs, UNREC_OP, numOfSubscriptions, currentW);
                 if (side == LEFT)
                     allPositions.clear();
 
@@ -371,19 +371,19 @@ static vector<int> matchSubscriptToLoopSymbols(const vector<SgForStmt*> &parentL
                         currOp.resize(numOfSubscriptions);
 
                     //add only uniq 
-                    auto itAdd = currOp[dimNum].coefficients.find(coefs);
+                    auto itAdd = currOp[dimNum].coefficients.find(coeffs);
                     if (itAdd == currOp[dimNum].coefficients.end())
-                        itAdd = currOp[dimNum].coefficients.insert(itAdd, make_pair(coefs, currentW));
+                        itAdd = currOp[dimNum].coefficients.insert(itAdd, make_pair(coeffs, currentW));
                 }
 
-                if (coefs.first < 0)
+                if (coeffs.first < 0)
                     addInfoToMap(loopInfo, parentLoops[position], currOrigArrayS, arrayRef, dimNum, REMOTE_TRUE, currLine, numOfSubscriptions);
                 else
                     //if we found regular access to array - set it false
                     addInfoToMap(loopInfo, parentLoops[position], currOrigArrayS, arrayRef, dimNum, REMOTE_FALSE, currLine, numOfSubscriptions);
             }
 
-            if (coefs.first < 0 && sharedMemoryParallelization == 0)
+            if (coeffs.first < 0 && sharedMemoryParallelization == 0)
             {
                 if (currRegime == DATA_DISTR)
                 {
@@ -402,15 +402,15 @@ static vector<int> matchSubscriptToLoopSymbols(const vector<SgForStmt*> &parentL
                     if (side == LEFT)
                         allPositions.clear();
                     else
-                        addInfoToVectors(loopInfo, parentLoops[position], currOrigArrayS, dimNum, coefs, UNREC_OP, numOfSubscriptions, currentW);
+                        addInfoToVectors(loopInfo, parentLoops[position], currOrigArrayS, dimNum, coeffs, UNREC_OP, numOfSubscriptions, currentW);
                 }
             }
             else
             {
                 if (side == LEFT)
-                    addInfoToVectors(loopInfo, parentLoops[position], currOrigArrayS, dimNum, coefs, WRITE_OP, numOfSubscriptions, currentW);
+                    addInfoToVectors(loopInfo, parentLoops[position], currOrigArrayS, dimNum, coeffs, WRITE_OP, numOfSubscriptions, currentW);
                 else
-                    addInfoToVectors(loopInfo, parentLoops[position], currOrigArrayS, dimNum, coefs, READ_OP, numOfSubscriptions, currentW);
+                    addInfoToVectors(loopInfo, parentLoops[position], currOrigArrayS, dimNum, coeffs, READ_OP, numOfSubscriptions, currentW);
             }
         }
     }
@@ -418,13 +418,13 @@ static vector<int> matchSubscriptToLoopSymbols(const vector<SgForStmt*> &parentL
     if (currRegime == ARRAY_ACC_CORNER)
     {
         int *valueSubs = new int[2];
-        valueSubs[0] = coefs.first;
+        valueSubs[0] = coeffs.first;
-        valueSubs[1] = coefs.second;
+        valueSubs[1] = coeffs.second;
 #ifdef __SPF
         addToCollection(__LINE__, __FILE__, valueSubs, 2);
 #endif
-        const vector<int*> &coefs = getAttributes<SgExpression*, int*>(subscr, set<int>{ INT_VAL });
+        const vector<int*> &coeffs = getAttributes<SgExpression*, int*>(subscr, set<int>{ INT_VAL });
-        if (coefs.size() == 0)
+        if (coeffs.size() == 0)
         {
             subscr->addAttribute(INT_VAL, valueSubs, sizeof(int*));
             if (position != -1 && allPositions.size() == 1 && position < parentLoops.size())

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/Predictor/PredictScheme.cpp

@@ -326,7 +326,7 @@ static json parseAlign(const map<DIST::Array*, int>& byPos, SgSymbol* srcArr, Sg
         list = list->rhs();
     }
 
-    vector<pair<int, int>> coefs(srcSymbs.size());
+    vector<pair<int, int>> coeffs(srcSymbs.size());
     list = listTgt;
     while (list)
     {
@@ -337,8 +337,8 @@ static json parseAlign(const map<DIST::Array*, int>& byPos, SgSymbol* srcArr, Sg
             has = recSymbolFind(exp, srcSymbs[z].first, VAR_REF);
             if (has)
             {
-                getCoefsOfSubscript(coefs[z], exp, srcSymbs[z].second);
+                getCoefsOfSubscript(coeffs[z], exp, srcSymbs[z].second);
-                if (coefs[z].first == 0)
+                if (coeffs[z].first == 0)
                     printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
                 break;
             }
@@ -346,14 +346,14 @@ static json parseAlign(const map<DIST::Array*, int>& byPos, SgSymbol* srcArr, Sg
         list = list->rhs();
     }
 
-    for (int z = 0; z < coefs.size(); ++z)
+    for (int z = 0; z < coeffs.size(); ++z)
     {
-        if (coefs[z].first == 0)
+        if (coeffs[z].first == 0)
             continue;
-        if (coefs[z].second)
+        if (coeffs[z].second)
-            align["rules"].push_back({ z, coefs[z].first });
+            align["rules"].push_back({ z, coeffs[z].first });
         else
-            align["rules"].push_back({ z, coefs[z].first, coefs[z].second });
+            align["rules"].push_back({ z, coeffs[z].first, coeffs[z].second });
     }
     return align;
 }

src/PrivateAnalyzer/private_arrays_search.cpp

@@ -16,6 +16,34 @@
 
 using namespace std;
 
+static void RemoveEmptyPoints(ArrayAccessingIndexes& container)
+{
+    ArrayAccessingIndexes resultContainer;
+    unordered_set<string> toRemove;
+
+    for (auto& [arrayName, accessingSet] : container)
+    {
+        vector<vector<ArrayDimension>> points;
+        for (auto& arrayPoint : accessingSet.GetElements())
+        {
+            if (!arrayPoint.empty())
+                points.push_back(arrayPoint);
+        }
+
+        if (points.size() < accessingSet.GetElements().size() && !points.empty())
+            resultContainer[arrayName] = points;
+
+        if (points.empty())
+            toRemove.insert(arrayName);
+    }
+
+    for (const string& name : toRemove)
+        container.erase(name);
+
+    for (auto& [arrayName, accessingSet] : resultContainer)
+        container[arrayName] = accessingSet;
+}
+
 static void Collapse(Region* region)
 {
     if (region->getBasickBlocks().empty())
@@ -38,13 +66,13 @@ static void Collapse(Region* region)
             region->array_use[arrayName] = region->array_use[arrayName].Union(diff);
         }
     }
+
     ArrayAccessingIndexes useUnion;
     for (auto& byBlock : region->getBasickBlocks())
         for (auto& [arrayName, arrayRanges] : byBlock->array_use)
             useUnion[arrayName] = useUnion[arrayName].Union(byBlock->array_use[arrayName]);
 
-    for (auto& [arrayName, arrayRanges] : useUnion)
+    region->array_priv = region->array_use;
-        region->array_priv[arrayName] = useUnion[arrayName].Diff(region->array_use[arrayName]);
     
     for (Region* prevBlock : region->getHeader()->getPrevRegions()) 
         prevBlock->replaceInNextRegions(region, region->getHeader());
@@ -75,6 +103,7 @@ static void SolveDataFlowIteratively(Region* DFG)
                     newIn.clear();
                     continue;
                 }
+
                 for (const auto& [arrayName, accessSet] : prevBlock->array_out)
                 {
                     if (newIn.find(arrayName) != newIn.end())
@@ -87,6 +116,7 @@ static void SolveDataFlowIteratively(Region* DFG)
 
         b->array_in = move(newIn);
         ArrayAccessingIndexes newOut;
+
         if (b->array_def.empty())        
             newOut = b->array_in;
         else if (b->array_in.empty())
@@ -122,22 +152,7 @@ static void SolveDataFlow(Region* DFG)
     Collapse(DFG);
 }
 
-unsigned long long CalculateLength(const AccessingSet& array) 
+static void AddPrivateArraysToLoop(LoopGraph* loop, const ArrayAccessingIndexes& privates, set<SgStatement*>& insertedPrivates)
-{
-    if (array.GetElements().empty())
-        return 0;
-    unsigned long long result = 1;
-    for (const auto& range : array.GetElements()) 
-    {
-        for (const auto& dim : range)
-        {
-            result *= (dim.step * dim.tripCount);
-        }
-    }
-    return result;
-}
-
-void AddPrivateArraysToLoop(LoopGraph* loop, const ArrayAccessingIndexes& privates, set<SgStatement*>& insertedPrivates)
 {
     SgStatement* spfStat = new SgStatement(SPF_ANALYSIS_DIR);
     spfStat->setlineNumber(loop->loop->lineNumber());
@@ -188,6 +203,8 @@ void FindPrivateArrays(map<string, vector<LoopGraph*>> &loopGraph, map<FuncInfo*
         SgFile::switchToFile(fileName);
         for (const auto& loop : loops)
         {
+            if (!loop->isFor())
+                continue;
             SgStatement* search_func = loop->loop->GetOriginal();
 
             while (search_func && (!isSgProgHedrStmt(search_func)))
@@ -204,6 +221,7 @@ void FindPrivateArrays(map<string, vector<LoopGraph*>> &loopGraph, map<FuncInfo*
                         continue;
                     }
                     SolveDataFlow(loopRegion);
+                    RemoveEmptyPoints(loopRegion->array_priv);
                     result[loop] = loopRegion->array_priv;
                     delete(loopRegion);
                 }

src/PrivateAnalyzer/range_structures.cpp

@@ -64,18 +64,13 @@ static vector<ArrayDimension> DimensionDifference(const ArrayDimension& dim1, co
         result.push_back({ dim1.start, dim1.step, (intersection->start - dim1.start) / dim1.step, dim1.array});
     
     /* add the parts between intersection steps */
-    uint64_t start = (intersection->start - dim1.start) / dim1.step;
+    if (intersection->step > dim1.step)
-    uint64_t interValue = intersection->start;
-    for (int64_t i = start; dim1.start + i * dim1.step <= intersection->start + intersection->step * (intersection->tripCount - 1); i++)
     {
-        uint64_t centerValue = dim1.start + i * dim1.step;
+        uint64_t start = (intersection->start - dim1.start) / dim1.step;
-        if (centerValue == interValue)
+        uint64_t interValue = intersection->start;
+        for (int64_t i = start; interValue <= intersection->start + intersection->step * (intersection->tripCount - 1); i++)
         {
-            if (i - start > 1)
+            result.push_back({interValue + dim1.step, dim1.step, intersection->step / dim1.step, dim1.array});
-            {
-                result.push_back({ dim1.start + (start + 1) * dim1.step, dim1.step, i - start - 1, dim1.array });
-                start = i;
-            }
             interValue += intersection->step;
         }
     }
@@ -216,6 +211,10 @@ void AccessingSet::Insert(const vector<ArrayDimension>& element)
 }
 
 AccessingSet AccessingSet::Union(const AccessingSet& source) {
+    if (source.GetElements().empty())
+        return *this;
+    if (allElements.empty())
+        return source;
     AccessingSet result;
     for (auto& element : source.GetElements())
         result.Insert(element);

src/PrivateAnalyzer/region.cpp

@@ -4,6 +4,7 @@
 #include<unordered_map>
 #include<string>
 #include <numeric>
+#include <iostream>
 
 #include "range_structures.h"
 #include "region.h"
@@ -104,15 +105,33 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
 
         auto operation = instruction->getInstruction()->getOperation();
         auto type = instruction->getInstruction()->getArg1()->getType();
+        if (operation == SAPFOR::CFG_OP::ASSIGN && instruction->getInstruction()->getResult()->getType() == SAPFOR::CFG_ARG_TYPE::ARRAY)
+        {
+            SgStatement* op = instruction->getInstruction()->getOperator();
+            if (op && op->expr(0) && isArrayRef(op->expr(0)) && op->expr(0)->symbol() && op->expr(0)->type())
+            {
+                if (isSgArrayType(op->expr(0)->symbol()->type()))
+                {
+                    SgArrayType* arrayType = (SgArrayType*)op->expr(0)->symbol()->type();
+                    int dimCount = ((SgArrayType*)op->expr(0)->symbol()->type())->dimension();
+                    vector<ArrayDimension> point;
+                    for (int i = 0; i < dimCount; i++)
+                    {
+                        string strDimLength = arrayType->sizeInDim(i)->unparse();
+                        if (arrayType->sizeInDim(i)->variant() == INT_VAL && strDimLength != "0")
+                            point.push_back({ 1ULL, 1ULL, (uint64_t)stoi(strDimLength), (SgArrayRefExp*)op->expr(0) });
+                    }
+
+                    if (point.size() == dimCount)
+                        def[instruction->getInstruction()->getResult()->getValue()] = AccessingSet({point});
+                }
+            }
+        }
         if ((operation == SAPFOR::CFG_OP::STORE || operation == SAPFOR::CFG_OP::LOAD) && type == SAPFOR::CFG_ARG_TYPE::ARRAY)
         {
             vector<SAPFOR::Argument*> index_vars;
             vector<int> refPos;
-            string array_name;
+            string array_name = instruction->getInstruction()->getArg1()->getValue();            
-            if (operation == SAPFOR::CFG_OP::STORE)
-                array_name = instruction->getInstruction()->getArg1()->getValue();            
-            else
-                array_name = instruction->getInstruction()->getArg2()->getValue();
 
             int j = i - 1;
             while (j >= 0 && instructions[j]->getInstruction()->getOperation() == SAPFOR::CFG_OP::REF)
@@ -127,25 +146,12 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
             vector<ArrayDimension> accessPoint(n);
 
             auto* ref = isSgArrayRefExp(instruction->getInstruction()->getExpression());
-            vector<pair<int, int>> coefsForDims;
-            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);
-                }
-
-            }
-
             int fillCount = 0;
 
-            while (!index_vars.empty() && !refPos.empty() && !coefsForDims.empty())
+            while (!index_vars.empty() && !refPos.empty())
             {
                 auto var = index_vars.back();
                 int currentVarPos = refPos.back();
-                pair<int, int> currentCoefs = coefsForDims.back();
                 ArrayDimension current_dim;
                 if (var->getType() == SAPFOR::CFG_ARG_TYPE::CONST)
                     current_dim = { stoul(var->getValue()), 1, 1, ref};
@@ -174,16 +180,19 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
                             return -1;
                     }
 
-                    uint64_t start = currentLoop->startVal * currentCoefs.first + currentCoefs.second;
+                    uint64_t start = currentLoop->startVal;
-                    uint64_t step = currentCoefs.first;
+                    uint64_t step = currentLoop->stepVal;
-                    current_dim = { start, step, (uint64_t)currentLoop->calculatedCountOfIters, ref };
+                    uint64_t iters = currentLoop->calculatedCountOfIters;
+                    current_dim = { start, step, iters, ref };
                 }
 
-                accessPoint[n - index_vars.size()] = current_dim;
+                if (current_dim.start != 0 && current_dim.step != 0 && current_dim.tripCount != 0)
-                fillCount++;
+                {
+                    accessPoint[n - index_vars.size()] = current_dim;
+                    fillCount++;
+                }
                 index_vars.pop_back();
                 refPos.pop_back();
-                coefsForDims.pop_back();
             }
 
             if (fillCount == accessPoint.size())
@@ -230,8 +239,11 @@ static Region* CreateSubRegion(LoopGraph* loop, const vector<SAPFOR::BasicBlock*
             region->addBasickBlocks(bbToRegion.at(block));
 
     for (LoopGraph* childLoop : loop->children)
+    {
+        if (!childLoop->isFor())
+            continue;
         region->addSubRegions(CreateSubRegion(childLoop, Blocks, bbToRegion));
-
+    }
     return region;
 }
 
@@ -250,5 +262,9 @@ Region::Region(LoopGraph* loop, const vector<SAPFOR::BasicBlock*>& Blocks)
     SetConnections(bbToRegion, blockSet);
     //create subRegions
     for (LoopGraph* childLoop : loop->children)
+    {
+        if (!childLoop->isFor())
+            continue;
         subRegions.insert(CreateSubRegion(childLoop, Blocks, bbToRegion));
+    }
 }

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"
@@ -46,6 +47,7 @@
 #include "DynamicAnalysis/gCov_parser_func.h"
 #include "DynamicAnalysis/createParallelRegions.h"
 
+#include "ArrayConstantPropagation/propagation.h"
 #include "DirectiveProcessing/directive_analyzer.h"
 #include "DirectiveProcessing/directive_creator.h"
 #include "DirectiveProcessing/insert_directive.h"
@@ -280,7 +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();
@@ -1917,6 +1919,13 @@ 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);
 
     const float elapsed = duration_cast<milliseconds>(high_resolution_clock::now() - timeForPass).count() / 1000.;
     const float elapsedGlobal = duration_cast<milliseconds>(high_resolution_clock::now() - globalTime).count() / 1000.;
@@ -2370,6 +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,
@@ -187,6 +189,7 @@ enum passes {
     FIND_PRIVATE_ARRAYS,
 
     TRANSFORM_ASSUMED_SIZE_PARAMETERS,
+    ARRAY_PROPAGATION,
 
     TEST_PASS,
     EMPTY_PASS
@@ -270,6 +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";
@@ -376,6 +380,7 @@ static void setPassValues()
     passNames[FIND_PRIVATE_ARRAYS] = "FIND_PRIVATE_ARRAYS";
 
     passNames[TRANSFORM_ASSUMED_SIZE_PARAMETERS] = "TRANSFORM_ASSUMED_SIZE_PARAMETERS";
+    passNames[ARRAY_PROPAGATION] = "ARRAY_PROPAGATION";
 
     passNames[TEST_PASS] = "TEST_PASS";
 }

src/Transformations/LoopCombining/loops_combiner.cpp

@@ -1123,7 +1123,7 @@ static bool hasDependenciesBetweenArrays(LoopGraph* firstLoop, LoopGraph* loop,
         for (int d = 0; d < dimensions; ++d)
         {
             //по измерениям массива отображение на цикл вложенности d
-            vector<set<pair<int, int>>> coefsRead[2], coefsWrite[2];
+            vector<set<pair<int, int>>> coeffsRead[2], coeffsWrite[2];
 
             checkNull(currLoop[0], convertFileName(__FILE__).c_str(), __LINE__);
             checkNull(currLoop[1], convertFileName(__FILE__).c_str(), __LINE__);
@@ -1133,31 +1133,31 @@ static bool hasDependenciesBetweenArrays(LoopGraph* firstLoop, LoopGraph* loop,
                 auto it = currLoop[k]->readOpsForLoop.find(array);
                 if (it != currLoop[k]->readOpsForLoop.end())
                 {
-                    if (coefsRead[k].size() == 0)
+                    if (coeffsRead[k].size() == 0)
-                        coefsRead[k].resize(it->second.size());
+                        coeffsRead[k].resize(it->second.size());
 
                     for (int z = 0; z < it->second.size(); ++z)
                         if (it->second[z].coefficients.size())
                             for (auto& coef : it->second[z].coefficients)
-                                coefsRead[k][z].insert(coef.first);
+                                coeffsRead[k][z].insert(coef.first);
                 }
 
                 auto itW = currLoop[k]->writeOpsForLoop.find(array);
                 if (itW != currLoop[k]->writeOpsForLoop.end())
                 {
-                    if (coefsWrite[k].size() == 0)
+                    if (coeffsWrite[k].size() == 0)
-                        coefsWrite[k].resize(itW->second.size());
+                        coeffsWrite[k].resize(itW->second.size());
 
                     for (int z = 0; z < itW->second.size(); ++z)
                         if (itW->second[z].coefficients.size())
                             for (auto& coef : itW->second[z].coefficients)
-                                coefsWrite[k][z].insert(coef.first);
+                                coeffsWrite[k][z].insert(coef.first);
                 }
             }
 
             //нет записей, значит нет зависимости
             bool nulWrite = true;
-            for (auto& wr : coefsWrite)
+            for (auto& wr : coeffsWrite)
                 for (auto& elem : wr)
                     if (elem.size() != 0)
                         nulWrite = false;
@@ -1168,62 +1168,62 @@ static bool hasDependenciesBetweenArrays(LoopGraph* firstLoop, LoopGraph* loop,
             // если чтение в одном цикле и запись (и наоборот) в другом идут по разным правилам, то пока что это зависимость.
             // здесь можно уточнить.
 
-            const int len = std::max(coefsWrite[0].size(), coefsRead[0].size());
+            const int len = std::max(coeffsWrite[0].size(), coeffsRead[0].size());
             int countW[2] = { 0, 0 };
             int countR[2] = { 0, 0 };
 
             for (int L = 0; L < 2; ++L)
-                for (int z = 0; z < coefsWrite[L].size(); ++z)
+                for (int z = 0; z < coeffsWrite[L].size(); ++z)
-                    countW[L] += (coefsWrite[L][z].size() ? 1 : 0);
+                    countW[L] += (coeffsWrite[L][z].size() ? 1 : 0);
 
             for (int L = 0; L < 2; ++L)
-                for (int z = 0; z < coefsRead[L].size(); ++z)
+                for (int z = 0; z < coeffsRead[L].size(); ++z)
-                    countR[L] += (coefsRead[L][z].size() ? 1 : 0);
+                    countR[L] += (coeffsRead[L][z].size() ? 1 : 0);
 
             for (int p = 0; p < len; ++p)
             {
-                if (coefsWrite[1].size() && coefsWrite[0].size())
+                if (coeffsWrite[1].size() && coeffsWrite[0].size())
-                    if (coefsWrite[0][p].size() != 0 && coefsWrite[1][p].size() != 0)
+                    if (coeffsWrite[0][p].size() != 0 && coeffsWrite[1][p].size() != 0)
-                        if (coefsWrite[0][p] != coefsWrite[1][p])
+                        if (coeffsWrite[0][p] != coeffsWrite[1][p])
                             return true;
 
-                if (coefsRead[1].size() && coefsWrite[0].size())
+                if (coeffsRead[1].size() && coeffsWrite[0].size())
-                    if (coefsWrite[0][p].size() != 0 && coefsRead[1][p].size() != 0)
+                    if (coeffsWrite[0][p].size() != 0 && coeffsRead[1][p].size() != 0)
-                        if (coefsWrite[0][p] != coefsRead[1][p])
+                        if (coeffsWrite[0][p] != coeffsRead[1][p])
                             return true;
 
-                if (coefsWrite[1].size() && coefsRead[0].size())
+                if (coeffsWrite[1].size() && coeffsRead[0].size())
-                    if (coefsWrite[1][p].size() != 0 && coefsRead[0][p].size() != 0)
+                    if (coeffsWrite[1][p].size() != 0 && coeffsRead[0][p].size() != 0)
-                        if (coefsWrite[1][p] != coefsRead[0][p])
+                        if (coeffsWrite[1][p] != coeffsRead[0][p])
                             return true;
                 
 
                 //отображение на разные измерения
-                if (coefsWrite[1].size() && coefsWrite[0].size())
+                if (coeffsWrite[1].size() && coeffsWrite[0].size())
                 {
-                    if (coefsWrite[0][p].size() != 0 && coefsWrite[1][p].size() == 0 && countW[1] ||
+                    if (coeffsWrite[0][p].size() != 0 && coeffsWrite[1][p].size() == 0 && countW[1] ||
-                        coefsWrite[0][p].size() == 0 && coefsWrite[1][p].size() != 0 && countW[0])
+                        coeffsWrite[0][p].size() == 0 && coeffsWrite[1][p].size() != 0 && countW[0])
                         return true;
                 }
 
-                if (coefsRead[1].size() && coefsWrite[0].size())
+                if (coeffsRead[1].size() && coeffsWrite[0].size())
                 {
-                    if (coefsWrite[0][p].size() != 0 && coefsRead[1][p].size() == 0 && countR[1] ||
+                    if (coeffsWrite[0][p].size() != 0 && coeffsRead[1][p].size() == 0 && countR[1] ||
-                        coefsWrite[0][p].size() == 0 && coefsRead[1][p].size() != 0 && countW[0])
+                        coeffsWrite[0][p].size() == 0 && coeffsRead[1][p].size() != 0 && countW[0])
                         return true;
                 }
 
-                if (coefsWrite[1].size() && coefsRead[1].size())
+                if (coeffsWrite[1].size() && coeffsRead[1].size())
                 {
-                    if (coefsWrite[1][p].size() != 0 && coefsRead[0][p].size() == 0 && countR[0] ||
+                    if (coeffsWrite[1][p].size() != 0 && coeffsRead[0][p].size() == 0 && countR[0] ||
-                        coefsWrite[1][p].size() == 0 && coefsRead[0][p].size() != 0 && countW[1])
+                        coeffsWrite[1][p].size() == 0 && coeffsRead[0][p].size() != 0 && countW[1])
                         return true;
                 }
 
                 //где то нет правил отображения вообще, но есть факт его наличия. 
-                if ( ((coefsWrite[0].size() == 0 && coefsRead[0].size() == 0) && (countW[0] == 0 && countR[0] == 0)) 
+                if ( ((coeffsWrite[0].size() == 0 && coeffsRead[0].size() == 0) && (countW[0] == 0 && countR[0] == 0)) 
                     ||
-                    ((coefsWrite[1].size() == 0 && coefsRead[1].size() == 0) && (countW[1] == 0 && countR[1] == 0)) )
+                    ((coeffsWrite[1].size() == 0 && coeffsRead[1].size() == 0) && (countW[1] == 0 && countR[1] == 0)) )
                     return true;
             }
             

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,

src/Utils/version.h

@@ -1,3 +1,3 @@
 #pragma once
 
-#define VERSION_SPF "2448"
+#define VERSION_SPF "2450"