draft of parallel regions merging pass

CMakeLists.txt

@@ -163,6 +163,9 @@ set(PARALLEL_REG src/ParallelizationRegions/ParRegions.cpp
         src/ParallelizationRegions/resolve_par_reg_conflicts.cpp
         src/ParallelizationRegions/resolve_par_reg_conflicts.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
@@ -419,6 +422,7 @@ set(SOURCE_EXE
     ${LOOP_ANALYZER}
     ${TRANSFORMS}
     ${PARALLEL_REG}
+    ${MERGE_COPY_ARRAYS}
     ${PRIV}
     ${FDVM}
     ${OMEGA}
@@ -470,6 +474,7 @@ 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 (FDVM_Compiler FILES ${FDVM})
 source_group (SageExtension FILES ${OMEGA})

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

@@ -12,6 +12,7 @@
 #include "SgUtils.h"
 #include "graph_loops.h"
 #include "CFGraph/CFGraph.h"
+#include "utils.h"
 
 using namespace std;
 
@@ -121,21 +122,95 @@ static void SolveDataFlow(Region* DFG)
     Collapse(DFG);
 }
 
-map<LoopGraph*, ArrayAccessingIndexes>  FindPrivateArrays(map<string, vector<LoopGraph*>> &loopGraph, map<FuncInfo*, vector<SAPFOR::BasicBlock*>>& FullIR)
+unsigned long long CalculateLength(const AccessingSet& array) 
+{
+    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());
+    spfStat->setFileName(loop->loop->fileName());
+    SgExpression* toAdd = new SgExpression(EXPR_LIST, new SgExpression(ACC_PRIVATE_OP), NULL, NULL);
+    set<SgSymbol*> arraysToInsert;
+    for (const auto& [_, accessingSet] : privates)
+    {
+        for (const auto& arrayElement : accessingSet.GetElements())
+        {
+            if (arrayElement.empty())
+                continue;
+            arraysToInsert.insert(arrayElement[0].array->symbol());
+        }
+    }
+
+    spfStat->setExpression(0, *toAdd);
+    toAdd = toAdd->lhs();
+    bool first = true;
+    for (auto& elem : arraysToInsert)
+    {
+        if (first)
+        {
+            toAdd->setLhs(new SgExpression(EXPR_LIST));
+            toAdd = toAdd->lhs();
+            first = false;
+        }
+        else
+        {
+            toAdd->setRhs(new SgExpression(EXPR_LIST));
+            toAdd = toAdd->rhs();
+        }
+        toAdd->setLhs(new SgVarRefExp(elem));
+    }
+    
+    if (arraysToInsert.size() == 0)
+        printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+
+    loop->loop->insertStmtBefore(*spfStat, *loop->loop->controlParent());
+    insertedPrivates.insert(spfStat);
+}
+
+void FindPrivateArrays(map<string, vector<LoopGraph*>> &loopGraph, map<FuncInfo*, vector<SAPFOR::BasicBlock*>>& FullIR, set<SgStatement*> &insertedPrivates)
 {  
     map<LoopGraph*, ArrayAccessingIndexes> result;
-    for (const auto& [loopName, loops] : loopGraph)
+    for (const auto& [fileName, loops] : loopGraph)
     {
+        SgFile::switchToFile(fileName);
         for (const auto& loop : loops)
         {
+            SgStatement* search_func = loop->loop->GetOriginal();
+
+            while (search_func && (!isSgProgHedrStmt(search_func)))
+                search_func = search_func->controlParent();
+
             for (const auto& [funcInfo, blocks]: FullIR)
+            {   
+                if (funcInfo->fileName == fileName && funcInfo->funcPointer->GetOriginal() == search_func)
                 {
                     Region* loopRegion = new Region(loop, blocks);
+                    if (loopRegion->getBasickBlocks().size() <= 1)
+                    {
+                        delete(loopRegion);
+                        continue;
+                    }
                     SolveDataFlow(loopRegion);
                     result[loop] = loopRegion->array_priv;
                     delete(loopRegion);
                 }
             }
+
+            if (result.find(loop) != result.end() && !result[loop].empty())
+                AddPrivateArraysToLoop(loop, result[loop], insertedPrivates);
+        }
     }
-    return result;
 }

src/PrivateAnalyzer/private_arrays_search.h

@@ -2,11 +2,12 @@
 
 #include <vector>
 #include <map>
+#include <set>
 #include <unordered_set>
 
 #include "range_structures.h"
 #include "graph_loops.h"
 #include "CFGraph/CFGraph.h"
 
-std::map<LoopGraph*, ArrayAccessingIndexes> FindPrivateArrays(std::map<std::string, std::vector<LoopGraph*>>& loopGraph, std::map<FuncInfo*, std::vector<SAPFOR::BasicBlock*>>& FullIR);
+void FindPrivateArrays(std::map<std::string, std::vector<LoopGraph*>>& loopGraph, std::map<FuncInfo*, std::vector<SAPFOR::BasicBlock*>>& FullIR, std::set<SgStatement*>& insertedPrivates);
 std::pair<SAPFOR::BasicBlock*, std::unordered_set<SAPFOR::BasicBlock*>> GetBasicBlocksForLoop(const LoopGraph* loop, const std::vector<SAPFOR::BasicBlock*> blocks);

src/PrivateAnalyzer/range_structures.cpp

@@ -47,7 +47,7 @@ static ArrayDimension* DimensionIntersection(const ArrayDimension& dim1, const A
     
     uint64_t start3 = dim1.start + x0 * dim1.step;
     uint64_t step3 = c * dim1.step;
-    ArrayDimension* result = new(ArrayDimension){ start3, step3, tMax + 1 };
+    ArrayDimension* result = new(ArrayDimension){ start3, step3, tMax + 1 , dim1.array};
     return result;
 }
 
@@ -61,7 +61,7 @@ static vector<ArrayDimension> DimensionDifference(const ArrayDimension& dim1, co
     vector<ArrayDimension> result;
     /* add the part before intersection */
     if (dim1.start < intersection->start)    
-        result.push_back({ dim1.start, dim1.step, (intersection->start - dim1.start) / dim1.step });
+        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;
@@ -73,7 +73,7 @@ static vector<ArrayDimension> DimensionDifference(const ArrayDimension& dim1, co
         {
             if (i - start > 1)
             {
-                result.push_back({ dim1.start + (start + 1) * dim1.step, dim1.step, i - start - 1 });
+                result.push_back({ dim1.start + (start + 1) * dim1.step, dim1.step, i - start - 1, dim1.array });
                 start = i;
             }
             interValue += intersection->step;
@@ -85,7 +85,7 @@ static vector<ArrayDimension> DimensionDifference(const ArrayDimension& dim1, co
         /* first value after intersection */
         uint64_t right_start = intersection->start + intersection->step * (intersection->tripCount - 1) + dim1.step;
         uint64_t tripCount = (dim1.start + dim1.step * dim1.tripCount - right_start) / dim1.step;
-        result.push_back({ right_start, dim1.step, tripCount });
+        result.push_back({ right_start, dim1.step, tripCount, dim1.array });
     }
     delete(intersection);
     return result;

src/PrivateAnalyzer/range_structures.h

@@ -6,9 +6,12 @@
 #include <string>
 #include <cstdint>
 
+#include "SgUtils.h"
+
 struct ArrayDimension
 {
     uint64_t start, step, tripCount;
+    SgArrayRefExp* array;
 };
 
 class AccessingSet {

src/PrivateAnalyzer/region.cpp

@@ -139,17 +139,16 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
 
             }
 
-            if(coefsForDims.empty())
-                printInternalError(convertFileName(__FILE__).c_str(), __LINE__);
+            int fillCount = 0;
 
-            while (!index_vars.empty())
+            while (!index_vars.empty() && !refPos.empty() && !coefsForDims.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 };
+                    current_dim = { stoul(var->getValue()), 1, 1, ref};
                 else
                 {
                     string name, full_name = var->getValue();
@@ -177,21 +176,25 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
 
                     uint64_t start = currentLoop->startVal * currentCoefs.first + currentCoefs.second;
                     uint64_t step = currentCoefs.first;
-                    current_dim = { start, step, (uint64_t)currentLoop->calculatedCountOfIters };
+                    current_dim = { start, step, (uint64_t)currentLoop->calculatedCountOfIters, ref };
                 }
 
                 accessPoint[n - index_vars.size()] = current_dim;
+                fillCount++;
                 index_vars.pop_back();
                 refPos.pop_back();
                 coefsForDims.pop_back();
             }
 
+            if (fillCount == accessPoint.size())
+            {
                 if (operation == SAPFOR::CFG_OP::STORE)
                     def[array_name].Insert(accessPoint);
                 else
                     use[array_name].Insert(accessPoint);
             }
         }
+    }
     return 0;
 
 }

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"
@@ -279,6 +280,7 @@ static string unparseProjectIfNeed(SgFile* file, const int curr_regime, const bo
             for (SgStatement* st = file->firstStatement(); st; st = st->lexNext())
                 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);
 
             for (auto& elem : toDel)
@@ -1019,8 +1021,6 @@ static bool runAnalysis(SgProject &project, const int curr_regime, const bool ne
                 if(func->funcPointer->variant() != ENTRY_STAT)
                     countOfTransform += removeDeadCode(func->funcPointer, allFuncInfo, commonBlocks);
         }
-        else if (curr_regime == FIND_PRIVATE_ARRAYS)
-            FindPrivateArrays(loopGraph, fullIR);
         else if (curr_regime == TEST_PASS)
         {
         //test pass
@@ -1916,6 +1916,12 @@ static bool runAnalysis(SgProject &project, const int curr_regime, const bool ne
     }
     else if (curr_regime == TRANSFORM_ASSUMED_SIZE_PARAMETERS)
         transformAssumedSizeParameters(allFuncInfo);
+    else if (curr_regime == FIND_PRIVATE_ARRAYS_ANALYSIS)
+        FindPrivateArrays(loopGraph, fullIR, insertedPrivates);
+    else if (curr_regime == MERGE_COPY_ARRAYS)
+    {
+        mergeCopyArrays(parallelRegions, allFuncInfo);
+    }
 
     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.;
@@ -2369,10 +2375,12 @@ 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:
     case REMOVE_DEAD_CODE_AND_UNPARSE:
+    case FIND_PRIVATE_ARRAYS:
         if (folderName)
             runAnalysis(*project, UNPARSE_FILE, true, "", folderName);
         else
@@ -2634,7 +2642,7 @@ int main(int argc, char **argv)
             }
         }
 
-        if (curr_regime == INSERT_PARALLEL_DIRS_NODIST)
+        if (curr_regime == INSERT_PARALLEL_DIRS_NODIST || curr_regime == FIND_PRIVATE_ARRAYS)
         {
             ignoreArrayDistributeState = true;
             sharedMemoryParallelization = 1;

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,
@@ -183,6 +185,7 @@ enum passes {
     SET_IMPLICIT_NONE,
     RENAME_INLCUDES,
 
+    FIND_PRIVATE_ARRAYS_ANALYSIS,
     FIND_PRIVATE_ARRAYS,
 
     TRANSFORM_ASSUMED_SIZE_PARAMETERS,
@@ -269,6 +272,7 @@ 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";
@@ -371,6 +375,7 @@ static void setPassValues()
     passNames[SET_IMPLICIT_NONE] = "SET_IMPLICIT_NONE";
     passNames[RENAME_INLCUDES] = "RENAME_INLCUDES";
     passNames[INSERT_NO_DISTR_FLAGS_FROM_GUI] = "INSERT_NO_DISTR_FLAGS_FROM_GUI";
+    passNames[FIND_PRIVATE_ARRAYS_ANALYSIS] = "FIND_PRIVATE_ARRAYS_ANALYSIS";
     passNames[FIND_PRIVATE_ARRAYS] = "FIND_PRIVATE_ARRAYS";
 
     passNames[TRANSFORM_ASSUMED_SIZE_PARAMETERS] = "TRANSFORM_ASSUMED_SIZE_PARAMETERS";

src/SapforData.h

@@ -175,6 +175,11 @@ std::set<std::tuple<std::string, int, std::string>> parametersOfProject; // [fil
 //for GET_MIN_MAX_BLOCK_DIST
 std::pair<int, int> min_max_block = std::make_pair(-1, -1);
 //
+
+//for FIND_PRIVATE_ARRAYS
+std::set<SgStatement*> insertedPrivates;
+//
+
 const char* passNames[EMPTY_PASS + 1];
 const char* optionNames[EMPTY_OPTION + 1];
 bool passNamesWasInit = false;

src/Utils/PassManager.h

@@ -316,7 +316,10 @@ void InitPassesDependencies(map<passes, vector<passes>> &passDepsIn, set<passes>
 
     list({ VERIFY_INCLUDES, CORRECT_VAR_DECL }) <= Pass(SET_IMPLICIT_NONE);
 
-    list({ CALL_GRAPH2, CALL_GRAPH, BUILD_IR, LOOP_GRAPH, LOOP_ANALYZER_DATA_DIST_S2 }) <= Pass(FIND_PRIVATE_ARRAYS);
+    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,

src/Utils/version.h

@@ -1,3 +1,3 @@
 #pragma once
 
-#define VERSION_SPF "2446"
+#define VERSION_SPF "2448"

src/VisualizerCalls/get_information.cpp

@@ -1793,6 +1793,15 @@ int SPF_RenameIncludes(void*& context, int winHandler, short* options, short* pr
     return simpleTransformPass(RENAME_INLCUDES, options, projName, folderName, output, outputMessage);
 }
 
+int SPF_InsertPrivateArrayDirectives(void*& context, int winHandler, short* options, short* projName, short* folderName, string& output, string& outputMessage)
+{
+    MessageManager::clearCache();
+    MessageManager::setWinHandler(winHandler);
+    ignoreArrayDistributeState = true;
+    sharedMemoryParallelization = 1;
+    return simpleTransformPass(FIND_PRIVATE_ARRAYS, options, projName, folderName, output, outputMessage);
+}
+
 static inline void convertBackSlash(char *str, int strL)
 {
     for (int z = 0; z < strL; ++z)
@@ -2499,6 +2508,8 @@ const wstring Sapfor_RunTransformation(const char* transformName_c, const char*
         retCode = SPF_InsertImplicitNone(context, winHandler, optSh, projSh, fold, output, outputMessage);
     else if (whichRun == "SPF_RenameIncludes")
         retCode = SPF_RenameIncludes(context, winHandler, optSh, projSh, fold, output, outputMessage);
+    else if (whichRun == "SPF_InsertPrivateArrayDirectives")
+        retCode = SPF_InsertPrivateArrayDirectives(context, winHandler, optSh, projSh, fold, output, outputMessage);
     else if (whichRun == "SPF_CreateParallelVariant")
     {
         vector<string> splited;