fix algorythm

src/PrivateAnalyzer/private_arrays_search.cpp

@@ -1,3 +1,4 @@
+#include <algorithm>
 #include <map>
 #include <unordered_set>
 #include <unordered_map>
@@ -16,6 +17,8 @@
 
 using namespace std;
 
+static unordered_set<Region*> collapsed;
+
 static void RemoveEmptyPoints(ArrayAccessingIndexes& container)
 {
     ArrayAccessingIndexes resultContainer;
@@ -49,41 +52,81 @@ static void Collapse(Region* region)
     if (region->getBasickBlocks().empty())
         return;
 
-    for (auto& [arrayName, arrayRanges] : region->getHeader()->array_out)
+    bool firstRegion = true;
+    for (Region* basickBlock : region->getBasickBlocks())
     {
-        for (Region* byBlock : region->getBasickBlocks())
+        if (basickBlock->getNextRegions().empty())
         {
-            AccessingSet intersection = byBlock->array_def[arrayName].Intersect(arrayRanges);
+            if (firstRegion)
-            region->array_def[arrayName] = region->array_def[arrayName].Union(intersection);
+            {
+                region->array_def = basickBlock->array_out;
+                firstRegion = false;
+            }
+            else
+            {
+                unordered_set<string> toErease;
+                for (auto& [arrayName, arrayRanges] : region->array_def)
+                {
+                    if (basickBlock->array_out.find(arrayName) != basickBlock->array_out.end())
+                        arrayRanges = arrayRanges.Intersect(basickBlock->array_out[arrayName]);
+                    else
+                    {
+                        arrayRanges = AccessingSet();
+                        toErease.insert(arrayName);
+                    }
+                }
+                for (string arrayName : toErease)
+                    region->array_def.erase(arrayName);
+            }
         }
     }
 
+    RegionInstruction instruction;
+    instruction.def = move(region->array_def);
+
+
     for (auto& byBlock : region->getBasickBlocks())
     {
-        for (auto& [arrayName, arrayRanges] : byBlock->array_use)
+        for (auto& instruction : byBlock->instructions)
         {
-            AccessingSet diff = byBlock->array_use[arrayName].Diff(byBlock->array_in[arrayName]);
+            for (auto& [arrayName, _] : instruction.use)
+            {
+                AccessingSet diff = instruction.use[arrayName].Diff(instruction.in[arrayName]);
                 region->array_use[arrayName] = region->array_use[arrayName].Union(diff);
             }
         }
+    }
 
-    ArrayAccessingIndexes useUnion;
+    ArrayAccessingIndexes useUnionB;
     for (auto& byBlock : region->getBasickBlocks())
-        for (auto& [arrayName, arrayRanges] : byBlock->array_use)
+        for (auto& instruction : byBlock->instructions)
-            useUnion[arrayName] = useUnion[arrayName].Union(byBlock->array_use[arrayName]);
+            for (auto& [arrayName, _] : instruction.use)
+                useUnionB[arrayName] = useUnionB[arrayName].Union(instruction.use[arrayName]);
 
-    region->array_priv = region->array_use;
+    for (auto& [arrayName, _] : useUnionB)
+        region->array_priv[arrayName] = useUnionB[arrayName].Diff(region->array_use[arrayName]);
+
+    instruction.use = move(region->array_use);
 
     for (Region* prevBlock : region->getHeader()->getPrevRegions())
+    {
         prevBlock->replaceInNextRegions(region, region->getHeader());
+        region->addPrevRegion(prevBlock);
+    }
 
     for (Region* nextBlock : region->getHeader()->getNextRegions())
+    {
         nextBlock->replaceInPrevRegions(region, region->getHeader());
+        region->addNextRegion(nextBlock);
+    }
+    region->instructions.push_back(instruction);
+
 }
 
 static void SolveDataFlowIteratively(Region* DFG)
 {
-    unordered_set<Region*> worklist(DFG->getBasickBlocks());
+    auto blocks = DFG->getBasickBlocks();
+    std::unordered_set<Region*> worklist(blocks.begin(), blocks.end());
     do
     {
         Region* b = *worklist.begin();
@@ -101,7 +144,7 @@ static void SolveDataFlowIteratively(Region* DFG)
                 if (prevBlock->array_out.empty())
                 {
                     newIn.clear();
-                    continue;
+                    break;
                 }
 
                 for (const auto& [arrayName, accessSet] : prevBlock->array_out)
@@ -137,21 +180,153 @@ static void SolveDataFlowIteratively(Region* DFG)
             b->array_out = newOut;
         else
             worklist.erase(b);
+    } while (!worklist.empty());
+}
+
+static void SolveForBasickBlock(Region* block)
+{
+    ArrayAccessingIndexes newIn;
+    bool flagFirst = true;
+    for (Region* prevBlock : block->getPrevRegions())
+    {
+        if (flagFirst)
+        {
+            newIn = prevBlock->array_out;
+            flagFirst = false;
+        }
+        else
+        {
+            if (prevBlock->array_out.empty())
+            {
+                newIn.clear();
+                break;
+            }
+
+            for (const auto& [arrayName, accessSet] : prevBlock->array_out)
+            {
+                if (newIn.find(arrayName) != newIn.end())
+                    newIn[arrayName] = newIn[arrayName].Intersect(accessSet);
+                else
+                    newIn[arrayName] = AccessingSet();
+            }
+        }
+    }
+
+    if (block->instructions.empty())
+        block->instructions.push_back(RegionInstruction());
+
+    block->instructions[0].in = move(newIn);
+
+    for (int i = 0; i < block->instructions.size(); i++)
+    {
+        auto& instruction = block->instructions[i];
+
+        if (i > 0)
+            instruction.in = block->instructions[i - 1].out;
+
+        ArrayAccessingIndexes newOut;
+        if (instruction.def.empty())
+            newOut = instruction.in;
+        else if (instruction.in.empty())
+            newOut = instruction.def;
+        else
+        {
+            for (auto& [arrayName, accessSet] : instruction.def)
+            {
+                if (instruction.in.find(arrayName) != instruction.in.end())
+                    newOut[arrayName] = instruction.def[arrayName].Union(instruction.in[arrayName]);
+                else
+                    newOut[arrayName] = accessSet;
+            }
+            for (auto& [arrayName, accessSet] : instruction.in)
+            {
+                if (newOut.find(arrayName) == newOut.end())
+                {
+                    newOut[arrayName] = accessSet;
+                }
+            }
+        }
+
+        instruction.out = move(newOut);
+    }
+    if (!block->instructions.empty())
+        block->array_out = block->instructions.back().out;
+}
+
+static void SolveDataFlowTopologically(Region* DFG)
+{
+    for (Region* b : DFG->getBasickBlocks())
+    {
+        collapsed.insert(b);
+        SolveForBasickBlock(b);
     }
-    while (!worklist.empty());
 }
 
 static void SolveDataFlow(Region* DFG)
 {
     if (!DFG)
         return;
-
-    SolveDataFlowIteratively(DFG);
     for (Region* subRegion : DFG->getSubRegions())
+    {
         SolveDataFlow(subRegion);
+        DFG->addBasickBlocks(subRegion);
+    }
+    vector<Region*>& blocks = DFG->getBasickBlocks();
+    auto pos = remove_if(blocks.begin(), blocks.end(), [](Region* r) { return collapsed.find(r) != collapsed.end(); });
+    blocks.erase(pos, blocks.end());
+    TopologySort(DFG->getBasickBlocks(), DFG->getHeader());
+    SolveDataFlowTopologically(DFG);
     Collapse(DFG);
 }
 
+static bool getArrayDeclaredDimensions(SgArrayRefExp* arrayRef, vector<uint64_t>& declaredDims)
+{
+    declaredDims.clear();
+    if (!arrayRef || !arrayRef->symbol() || !isSgArrayType(arrayRef->symbol()->type()))
+        return false;
+    SgArrayType* arrayType = (SgArrayType*)arrayRef->symbol()->type();
+    int dimCount = arrayType->dimension();
+    for (int i = 0; i < dimCount; i++)
+    {
+        SgExpression* sizeExpr = arrayType->sizeInDim(i);
+        SgConstantSymb* constValSymb = isSgConstantSymb(sizeExpr->symbol());
+        string strDimLength;
+        if (sizeExpr && sizeExpr->variant() == INT_VAL)
+            strDimLength = sizeExpr->unparse();
+        else if (constValSymb)
+            strDimLength = constValSymb->constantValue()->unparse();
+        else
+            return false;
+
+        if (strDimLength == "0")
+            return false;
+        declaredDims.push_back((uint64_t)stoi(strDimLength));
+    }
+    return true;
+}
+
+static bool CheckDimensionLength(const AccessingSet& array)
+{
+    if (array.GetElements().empty())
+        return false;
+    size_t dimCount = array.GetElements()[0].size();
+    SgArrayRefExp* arrayRef = array.GetElements()[0][0].array;
+    if (!arrayRef)
+        return false;
+    vector<uint64_t> declaredDims(dimCount);
+    if (!getArrayDeclaredDimensions(arrayRef, declaredDims))
+        return false;
+    vector<ArrayDimension> testArray(dimCount);
+    for (size_t i = 0; i < dimCount; i++)
+    {
+        testArray[i] = { 1, 1, declaredDims[i], nullptr };
+    }
+    AccessingSet diff = AccessingSet({ testArray }).Diff(array);
+
+    return diff.GetElements().empty();
+}
+
+
 static void AddPrivateArraysToLoop(LoopGraph* loop, const ArrayAccessingIndexes& privates, set<SgStatement*>& insertedPrivates)
 {
     SgStatement* spfStat = new SgStatement(SPF_ANALYSIS_DIR);
@@ -161,6 +336,8 @@ static void AddPrivateArraysToLoop(LoopGraph* loop, const ArrayAccessingIndexes&
     set<SgSymbol*> arraysToInsert;
     for (const auto& [_, accessingSet] : privates)
     {
+        if (!CheckDimensionLength(accessingSet))
+            continue;
         for (const auto& arrayElement : accessingSet.GetElements())
         {
             if (arrayElement.empty())
@@ -188,12 +365,12 @@ static void AddPrivateArraysToLoop(LoopGraph* loop, const ArrayAccessingIndexes&
         toAdd->setLhs(new SgVarRefExp(elem));
     }
 
-    if (arraysToInsert.size() == 0)
+    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)
 {

src/PrivateAnalyzer/range_structures.cpp

@@ -132,8 +132,10 @@ static vector<ArrayDimension> ElementsIntersection(const vector<ArrayDimension>&
 static vector<vector<ArrayDimension>> ElementsDifference(const vector<ArrayDimension>& firstElement,
     const vector<ArrayDimension>& secondElement)
 {
-    if (firstElement.empty() || secondElement.empty())
+    if (firstElement.empty())
         return {};
+    if (secondElement.empty())
+        return { firstElement };
 
     vector<ArrayDimension> intersection = ElementsIntersection(firstElement, secondElement);
     vector<vector<ArrayDimension>> result;
@@ -188,7 +190,7 @@ bool AccessingSet::ContainsElement(const vector<ArrayDimension>& element) const
 {
     vector<vector<ArrayDimension>> tails;
     FindUncovered(element, tails);
-    return !tails.empty();
+    return tails.empty();
 }
 
 void AccessingSet::FindCoveredBy(const vector<ArrayDimension>& element, vector<vector<ArrayDimension>>& result) const
@@ -252,13 +254,15 @@ AccessingSet AccessingSet::Diff(const AccessingSet& secondSet) const
         return *this;
 
     AccessingSet intersection = this->Intersect(secondSet);
-    AccessingSet uncovered = *this;
+    vector<vector<ArrayDimension>> uncovered;
-    vector<vector<ArrayDimension>> result;
+    for (const auto& element : allElements)
-    for (const auto& element : intersection.GetElements())
     {
         vector<vector<ArrayDimension>> current_uncovered;
-        uncovered.FindUncovered(element, current_uncovered);
+        intersection.FindUncovered(element, current_uncovered);
-        uncovered = AccessingSet(current_uncovered);
+        uncovered.insert(uncovered.end(),
+            std::move_iterator(current_uncovered.begin()),
+            std::move_iterator(current_uncovered.end())
+        );
     }
     return uncovered;
 }

src/PrivateAnalyzer/region.cpp

@@ -1,3 +1,4 @@
+#include <algorithm>
 #include <vector>
 #include <map>
 #include <unordered_set>
@@ -9,6 +10,7 @@
 #include "range_structures.h"
 #include "region.h"
 
+#include "..\Transformations\ExpressionSubstitution\expr_transform.h"
 #include "SgUtils.h"
 
 using namespace std;
@@ -93,7 +95,7 @@ static string FindIndexName(int pos, SAPFOR::BasicBlock* block, map<string, Loop
     return "";
 }
 
-static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAccessingIndexes& def, ArrayAccessingIndexes& use) {
+static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAccessingIndexes& def, ArrayAccessingIndexes& use, Region* region) {
     auto instructions = block->getInstructions();
     map<string, LoopGraph*> loopForIndex;
     BuildLoopIndex(loopForIndex, loop);
@@ -123,7 +125,11 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
                     }
 
                     if (point.size() == dimCount)
+                    {
                         def[instruction->getInstruction()->getResult()->getValue()] = AccessingSet({ point });
+                        RegionInstruction regionInstruction;
+                        regionInstruction.def[instruction->getInstruction()->getResult()->getValue()] = AccessingSet({ point });
+                    }
                 }
             }
         }
@@ -148,7 +154,21 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
             auto* ref = isSgArrayRefExp(instruction->getInstruction()->getExpression());
             int fillCount = 0;
 
-            while (!index_vars.empty() && !refPos.empty())
+            vector<pair<int, int>> coeffsForDims;
+            int subs = ref->numberOfSubscripts();
+            for (int i = 0; ref && i < ref->numberOfSubscripts(); ++i)
+            {
+                const vector<int*>& coeffs = getAttributes<SgExpression*, int*>(ref->subscript(i), set<int>{ INT_VAL });
+                if (coeffs.size() == 1)
+                {
+                    const pair<int, int> coef(coeffs[0][0], coeffs[0][1]);
+                    coeffsForDims.push_back(coef);
+                }
+
+            }
+            coeffsForDims = { coeffsForDims.rbegin(), coeffsForDims.rend() };
+
+            while (!index_vars.empty() && !refPos.empty() && !coeffsForDims.empty())
             {
                 auto var = index_vars.back();
                 int currentVarPos = refPos.back();
@@ -180,7 +200,7 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
                             return -1;
                     }
 
-                    uint64_t start = currentLoop->startVal;
+                    uint64_t start = coeffsForDims.back().second * currentLoop->startVal + coeffsForDims.back().first;
                     uint64_t step = currentLoop->stepVal;
                     uint64_t iters = currentLoop->calculatedCountOfIters;
                     current_dim = { start, step, iters, ref };
@@ -193,14 +213,29 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
                 }
                 index_vars.pop_back();
                 refPos.pop_back();
+                coeffsForDims.pop_back();
             }
 
             if (fillCount == accessPoint.size())
             {
+                RegionInstruction instruction;
                 if (operation == SAPFOR::CFG_OP::STORE)
+                {
                     def[array_name].Insert(accessPoint);
+                    instruction.def[array_name] = { { accessPoint } };
+                }
                 else
+                {
+                    instruction.use[array_name] = { { accessPoint } };
+                    if (def.find(array_name) == def.end())
                         use[array_name].Insert(accessPoint);
+                    else
+                    {
+                        AccessingSet element({ accessPoint });
+                        use[array_name] = use[array_name].Union(element.Diff(def[array_name]));
+                    }
+                }
+                region->instructions.push_back(instruction);
             }
         }
     }
@@ -208,6 +243,41 @@ static int GetDefUseArray(SAPFOR::BasicBlock* block, LoopGraph* loop, ArrayAcces
 
 }
 
+static void RemoveHeaderConnection(SAPFOR::BasicBlock* header, const unordered_set<SAPFOR::BasicBlock*>& blockSet, unordered_map<SAPFOR::BasicBlock*, Region*>& bbToRegion)
+{
+    for (SAPFOR::BasicBlock* block : blockSet)
+    {
+        bool isCycleBlock = false;
+        for (SAPFOR::BasicBlock* prevBlock : block->getPrev())
+            isCycleBlock = isCycleBlock || (blockSet.find(prevBlock) != blockSet.end());
+
+        if (isCycleBlock)
+        {
+            bbToRegion[block]->removeNextRegion(bbToRegion[header]);
+            bbToRegion[header]->removePrevRegion(bbToRegion[block]);
+        }
+    }
+}
+
+static void DFS(Region* block, vector<Region*>& result, unordered_set<Region*> cycleBlocks)
+{
+    for (Region* nextBlock : block->getNextRegions())
+    {
+        if (cycleBlocks.find(nextBlock) != cycleBlocks.end())
+            DFS(nextBlock, result, cycleBlocks);
+    }
+    result.push_back(block);
+}
+
+void TopologySort(std::vector<Region*>& basikBlocks, Region* header)
+{
+    vector<Region*> result;
+    unordered_set<Region*> cycleBlocks(basikBlocks.begin(), basikBlocks.end());
+    DFS(header, result, cycleBlocks);
+    reverse(result.begin(), result.end());
+    basikBlocks = result;
+}
+
 static void SetConnections(unordered_map<SAPFOR::BasicBlock*, Region*>& bbToRegion, const unordered_set<SAPFOR::BasicBlock*>& blockSet)
 {
     for (SAPFOR::BasicBlock* block : blockSet)
@@ -222,10 +292,11 @@ static void SetConnections(unordered_map<SAPFOR::BasicBlock*, Region*>& bbToRegi
     }
 }
 
-static Region* CreateSubRegion(LoopGraph* loop, const vector<SAPFOR::BasicBlock*>& Blocks, const unordered_map<SAPFOR::BasicBlock*, Region*>& bbToRegion)
+static Region* CreateSubRegion(LoopGraph* loop, const vector<SAPFOR::BasicBlock*>& Blocks, unordered_map<SAPFOR::BasicBlock*, Region*>& bbToRegion)
 {
     Region* region = new Region;
     auto [header, blockSet] = GetBasicBlocksForLoop(loop, Blocks);
+    RemoveHeaderConnection(header, blockSet, bbToRegion);
     if (bbToRegion.find(header) != bbToRegion.end())
         region->setHeader(bbToRegion.at(header));
     else
@@ -244,6 +315,7 @@ static Region* CreateSubRegion(LoopGraph* loop, const vector<SAPFOR::BasicBlock*
             continue;
         region->addSubRegions(CreateSubRegion(childLoop, Blocks, bbToRegion));
     }
+    TopologySort(region->getBasickBlocks(), region->getHeader());
     return region;
 }
 
@@ -254,12 +326,13 @@ Region::Region(LoopGraph* loop, const vector<SAPFOR::BasicBlock*>& Blocks)
     for (auto poiner : blockSet)
     {
         bbToRegion[poiner] = new Region(*poiner);
-        this->basickBlocks.insert(bbToRegion[poiner]);
+        this->basickBlocks.push_back(bbToRegion[poiner]);
-        GetDefUseArray(poiner, loop, bbToRegion[poiner]->array_def, bbToRegion[poiner]->array_use);
+        GetDefUseArray(poiner, loop, bbToRegion[poiner]->array_def, bbToRegion[poiner]->array_use, bbToRegion[poiner]);
 
     }
     this->header = bbToRegion[header];
     SetConnections(bbToRegion, blockSet);
+    RemoveHeaderConnection(header, blockSet, bbToRegion);
     //create subRegions
     for (LoopGraph* childLoop : loop->children)
     {
@@ -267,4 +340,5 @@ Region::Region(LoopGraph* loop, const vector<SAPFOR::BasicBlock*>& Blocks)
             continue;
         subRegions.insert(CreateSubRegion(childLoop, Blocks, bbToRegion));
     }
+    TopologySort(basickBlocks, this->header);
 }

src/PrivateAnalyzer/region.h

@@ -8,6 +8,11 @@
 #include "graph_loops.h"
 #include "CFGraph/CFGraph.h"
 
+struct RegionInstruction
+{
+    ArrayAccessingIndexes def, use, in, out;
+};
+
 class Region : public SAPFOR::BasicBlock {
 public:
     Region() { header = nullptr; }
@@ -20,13 +25,25 @@ public:
 
     void setHeader(Region* region) { header = region; }
 
-    std::unordered_set<Region*>& getBasickBlocks() { return basickBlocks; }
+    std::vector<Region*>& getBasickBlocks() { return basickBlocks; }
 
-    void addBasickBlocks(Region* region) { basickBlocks.insert(region); }
+    void addBasickBlocks(Region* region) { basickBlocks.push_back(region); }
 
     const std::unordered_set<Region*>& getPrevRegions() { return prevRegions; }
 
-    std::unordered_set<Region*> getNextRegions() { return nextRegions; }
+    std::unordered_set<Region*>& getNextRegions() { return nextRegions; }
+
+    void removeNextRegion(Region* region)
+    {
+        if (nextRegions.find(region) != nextRegions.end())
+            nextRegions.erase(region);
+    }
+
+    void removePrevRegion(Region* region)
+    {
+        if (prevRegions.find(region) != prevRegions.end())
+            prevRegions.erase(region);
+    }
 
     void addPrevRegion(Region* region) { prevRegions.insert(region); }
 
@@ -48,13 +65,18 @@ public:
 
     void addSubRegions(Region* region) { subRegions.insert(region); }
 
+    std::vector<RegionInstruction> instructions;
+
     ArrayAccessingIndexes array_def, array_use, array_out, array_in, array_priv;
 
 private:
-    std::unordered_set<Region*> subRegions, basickBlocks;
+    std::vector<Region*> basickBlocks;
+    std::unordered_set<Region*> subRegions;
     /*next Region which is BB for current BB Region*/
     std::unordered_set<Region*> nextRegions;
     /*prev Regions which is BBs for current BB Region*/
     std::unordered_set<Region*> prevRegions;
     Region* header;
 };
+
+void TopologySort(std::vector<Region*>& basikBlocks, Region* header);