partial fix in multi-threading relevant code

src/include/dynamic_operation.cc.h

@@ -50,9 +50,7 @@ namespace MultiArrayTools
           mL(std::make_tuple(*mProto.mOp,mOp), std::make_tuple(inds...),
              std::make_tuple(mMa), std::make_tuple(mProto.mOp->assign( mOp, mkMIndex(inds...) )),
              std::array<size_t,1>({1}), std::array<size_t,1>({0}))
-    {
+    {}
-        *mMa = 0;
-    }
 
     template <typename T, class Operation, class... Ranges>
     OpH<OperationRoot<T,Ranges...>> DynamicOuterOp<T,Operation,Ranges...>::get(const DExtT& pos) const

src/include/dynamic_operation.h

@@ -75,11 +75,15 @@ namespace MultiArrayTools
 	std::shared_ptr<MultiArray<T,Ranges...>> mMa;
         OpH<OperationRoot<T,Ranges...>> mProto;
 
+
+	
 	typedef ILoop<std::tuple<OperationRoot<T,Ranges...>,Operation>,
 		      std::tuple<std::shared_ptr<typename Ranges::IndexType>...>,
 		      std::tuple<std::shared_ptr<MultiArray<T,Ranges...>>>,
             std::tuple<decltype(mProto.mOp->assign( mOp, mkMIndex(std::shared_ptr<typename Ranges::IndexType>()...) ))>> LoopT;
 	
+	
+	
 	mutable LoopT mL;
 	
     public:

src/include/multi_array_operation.h

@@ -131,6 +131,13 @@ namespace MultiArrayTools
         AssignmentExpr2(T* dataPtr, const Target& tar, const OpClass& sec);
         AssignmentExpr2(const AssignmentExpr2& in) = default;
         AssignmentExpr2(AssignmentExpr2&& in) = default;
+        AssignmentExpr2& operator=(const AssignmentExpr2& in) = default;
+        AssignmentExpr2& operator=(AssignmentExpr2&& in) = default;
+
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const override final
+	{
+	    return std::make_shared<AssignmentExpr2<T,Target,OpClass,OIA>>(*this);
+	}
 	
         inline void operator()(size_t start = 0); 
         inline void operator()(size_t start, ExtType last);
@@ -160,6 +167,13 @@ namespace MultiArrayTools
         GetExpr(const OpClass& sec, const NextExpr& nexpr);
         GetExpr(const GetExpr& in) = default;
         GetExpr(GetExpr&& in) = default;
+        GetExpr& operator=(const GetExpr& in) = default;
+        GetExpr& operator=(GetExpr&& in) = default;
+	    
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const override final
+	{
+	    return std::make_shared<GetExpr<OpClass,NextExpr>>(*this);
+	}
 
         inline void operator()(size_t start = 0);
         inline void operator()(size_t start, ExtType last);
@@ -201,6 +215,13 @@ namespace MultiArrayTools
         AddExpr(T* dataPtr, const Target& tar, const OpClass& sec);
         AddExpr(const AddExpr& in) = default;
         AddExpr(AddExpr&& in) = default;
+        AddExpr& operator=(const AddExpr& in) = default;
+        AddExpr& operator=(AddExpr&& in) = default;
+	    
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const override final
+	{
+	    return std::make_shared<AddExpr<T,Target,OpClass,OIA>>(*this);
+	}
 
         inline void operator()(size_t start = 0); 
         inline void operator()(size_t start, ExtType last);

src/include/xfor/iloop.h

@@ -126,6 +126,11 @@ namespace MultiArrayHelper
               const std::array<size_t,LTpSize>& umpos, const std::array<size_t,VarTpSize>& setzero) :
             mOpTp(opTp), mIndTp(indTp), mVarTp(varTp), mLTp(lTp), mUmpos(umpos), mSetzero(setzero) {}
         
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const override final
+	{
+	    return std::make_shared<ILoop<OpTp,IndTp,VarTp,LTp>>(*this);
+	}
+
 	inline size_t operator()(size_t mpos, ExtType pos)
         {
             NN<VarTpSize-1>::zero(mVarTp, mSetzero);
@@ -229,7 +234,7 @@ namespace MultiArrayHelper
     // if instance copied to different thread, the "copy" will be newly created from this function
     // -> ensures that there is NO SHARED WORKSPACE
     template <class CF>
-    class PILoop
+    class PILoop : public ExpressionBase
     {
     private:
         size_t mThreadId = 0;
@@ -268,10 +273,25 @@ namespace MultiArrayHelper
         }
 
         PILoop(const CF& cf) : mThreadId(omp_get_thread_num()), mCF(cf), mL(mCF()) {}
+	/*
+	virtual void ensureThread() override final
+	{
+	    VCHECK(omp_get_thread_num());
+	    //if(static_cast<int>(mThreadId) != omp_get_thread_num()){
+	    mThreadId = omp_get_thread_num();
+	    mL = mCF();
+	    //}
+	}
+	*/
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const override final
+	{
+	    return std::make_shared<PILoop<CF>>(*this);
+	}
 
         inline size_t operator()(size_t mpos, ExtType pos)
         {
-            return mL(mpos, pos);
+	    mL(mpos, pos);
+	    return 0;
         }
 
         auto rootSteps(std::intptr_t i = 0) const
@@ -284,6 +304,28 @@ namespace MultiArrayHelper
         {
             return mL.var();
         }
+
+        virtual void operator()(size_t mlast, DExt last) override final
+        {
+            operator()(mlast, std::dynamic_pointer_cast<ExtT<ExtType>>(last)->ext());
+        }
+
+        virtual void operator()(size_t mlast = 0) override final
+        {
+            ExtType last = rootSteps();
+            last.zero();
+            operator()(mlast, last);
+        }
+        
+        virtual DExt dRootSteps(std::intptr_t iPtrNum = 0) const override final
+        {
+            return std::make_shared<ExtT<ExtType>>(rootSteps(iPtrNum));
+        }
+        
+        virtual DExt dExtension() const override final
+        {
+            return nullptr; //!!
+        }
     };
 
 }

src/include/xfor/xfor.h

@@ -112,8 +112,8 @@ namespace MultiArrayHelper
         DExtTX() { mDExt = std::make_shared<ExtT<None>>(); }
         DExtTX(const DExtTX& in) : mDExt(in.mDExt->deepCopy()), mNext(in.mNext) {}
         DExtTX(DExtTX&& in) : mDExt(in.mDExt->deepCopy()), mNext(in.mNext) {}
-        DExtTX& operator=(const DExtTX& in) { mNext = in.mNext; mDExt = in.mDExt->deepCopy(); }
+        DExtTX& operator=(const DExtTX& in) { mNext = in.mNext; mDExt = in.mDExt->deepCopy(); return *this; }
-        DExtTX& operator=(DExtTX&& in) { mNext = in.mNext; mDExt = in.mDExt->deepCopy(); }
+        DExtTX& operator=(DExtTX&& in) { mNext = in.mNext; mDExt = in.mDExt->deepCopy(); return *this; }
         explicit DExtTX(const DExt& in) : mDExt(in) {}
         /*
 	template <class Y>
@@ -173,6 +173,8 @@ namespace MultiArrayHelper
 	ExpressionBase& operator=(const ExpressionBase& in) = default;
 	ExpressionBase& operator=(ExpressionBase&& in) = default;
 
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const = 0;
+	
 	virtual void operator()(size_t mlast, DExt last) = 0;
 	virtual void operator()(size_t mlast = 0) = 0;
 
@@ -264,6 +266,10 @@ namespace MultiArrayHelper
 	SingleExpression(const IndexClass* indPtr,
 			 Expr expr);
 
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const override final
+	{
+	    return std::make_shared<SingleExpression<IndexClass,Expr>>(*this);
+	}
 
 	inline void operator()(size_t mlast, DExt last) override final;
 	inline void operator()(size_t mlast, ExtType last);
@@ -313,6 +319,10 @@ namespace MultiArrayHelper
 	SubExpr(const IndexClass* indPtr, std::intptr_t siptr,
                 const vector<size_t>* subset, Expr expr);
 
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const override final
+	{
+	    return std::make_shared<SubExpr<IndexClass,Expr>>(*this);
+	}
 
 	inline void operator()(size_t mlast, DExt last) override final;
 	inline void operator()(size_t mlast, ExtType last) ;
@@ -362,6 +372,11 @@ namespace MultiArrayHelper
 	For(const IndexClass* indPtr,
 	    size_t step, Expr expr);
 
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const override final
+	{
+	    return std::make_shared<For<IndexClass,Expr,FT>>(*this);
+	}
+
 	inline void operator()(size_t mlast, DExt last) override final;
 	inline void operator()(size_t mlast, ExtType last) ;
 	inline void operator()(size_t mlast = 0) override final;
@@ -410,6 +425,11 @@ namespace MultiArrayHelper
 	PFor(const IndexClass* indPtr,
 	    size_t step, Expr expr);
 
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const override final
+	{
+	    return std::make_shared<PFor<IndexClass,Expr>>(*this);
+	}
+
 	inline void operator()(size_t mlast, DExt last) override final;
 	inline void operator()(size_t mlast, ExtType last) ;
 	inline void operator()(size_t mlast = 0) override final;
@@ -431,8 +451,9 @@ namespace MultiArrayHelper
     class DynamicExpression : public ExpressionBase
     {
     private:
-	DynamicExpression() = default;
+	DynamicExpression() : mThreadId(omp_get_thread_num()) {}
 
+	size_t mThreadId;
 	std::shared_ptr<ExpressionBase> mNext;
 
     public:
@@ -440,10 +461,29 @@ namespace MultiArrayHelper
 	static constexpr size_t LAYER = 0;
 	static constexpr size_t SIZE = 0;
 
-	DynamicExpression(const DynamicExpression& in) = default;
+	DynamicExpression(const DynamicExpression& in) :
-	DynamicExpression(DynamicExpression&& in) = default;
+	    mThreadId(omp_get_thread_num()),
-	DynamicExpression& operator=(const DynamicExpression& in) = default;
+	    mNext( (static_cast<int>(in.mThreadId) == omp_get_thread_num()) ?
-	DynamicExpression& operator=(DynamicExpression&& in) = default;
+		   in.mNext : in.mNext->deepCopy()) {}
+	DynamicExpression(DynamicExpression&& in) :
+	    mThreadId(omp_get_thread_num()),
+	    mNext( (static_cast<int>(in.mThreadId) == omp_get_thread_num()) ?
+		   in.mNext : in.mNext->deepCopy()) {}
+	DynamicExpression& operator=(const DynamicExpression& in)
+	{
+	    mThreadId = omp_get_thread_num();
+	    mNext = (static_cast<int>(in.mThreadId) == omp_get_thread_num()) ?
+		in.mNext : in.mNext->deepCopy();
+	    return *this;
+	}
+	
+	DynamicExpression& operator=(DynamicExpression&& in)
+	{
+	    mThreadId = omp_get_thread_num();
+	    mNext = (static_cast<int>(in.mThreadId) == omp_get_thread_num()) ?
+		in.mNext : in.mNext->deepCopy();
+	    return *this;
+	}
 	
 	DynamicExpression(const std::shared_ptr<ExpressionBase>& next) :
 	    mNext(next)
@@ -457,6 +497,11 @@ namespace MultiArrayHelper
         template <class Expr>
         DynamicExpression(Expr ex) : mNext( std::make_shared<Expr>(ex) ) {}
         
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const override final
+	{
+	    return std::make_shared<DynamicExpression>(*this);
+	}
+
 	inline void operator()(size_t mlast, DExt last) override final;
         inline void operator()(size_t mlast, DExtT last) { (*this)(mlast,last.get()); }
 	inline void operator()(size_t mlast = 0) override final;
@@ -495,6 +540,11 @@ namespace MultiArrayHelper
 
 	ExpressionHolder(DynamicExpression expr);
 	
+	virtual std::shared_ptr<ExpressionBase> deepCopy() const override final
+	{
+	    return std::make_shared<ExpressionHolder<Expr>>(*this);
+	}
+
 	inline void operator()(size_t mlast, DExt last) override final;
 	inline void operator()(size_t mlast, ExtType last) ;
 	inline void operator()(size_t mlast = 0) override final;

src/tests/op4_unit_test.cc

@@ -137,15 +137,21 @@ namespace
         resx2(i1,di4) = mkDynOp(ma1(i1,di1) * ma2(i1,di2));
 	resx3(i1,di4) = mkDynOp(mkDynOp(ma1(i1,di1)) * mkDynOp(ma2(i1,di2)));
 
-	auto op1 = mkDynOutOp((ma1(i1,di1) * ma2(i1,di2)), ci4_1, ci4_2);
+	//auto op1 = mkDynOutOp((ma1(i1,di1) * ma2(i1,di2)), ci4_1, ci4_2);
+	//auto opr = resx4(i1,di4);
+	auto op1x = (ma1(i1,di1) * ma2(i1,di2));
 	auto opr = resx4(i1,di4);
 	
-	auto loop = mkILoop(std::make_tuple(opr,op1,*op1.data()->mOp), std::make_tuple(ci4_1, ci4_2),
+	auto loop = mkPILoop
+	    ( [&op1x,&opr,&xx,this](){
+		auto op1 = mkDynOutOp(op1x, ci4_1, ci4_2);
+		return mkGetExpr(op1,mkILoop(std::make_tuple(opr,op1,*op1.data()->mOp), std::make_tuple(ci4_1, ci4_2),
 					     std::make_tuple(xx),
 					     std::make_tuple(opr.assign( *op1.data()->mOp, mkMIndex(ci4_1, ci4_2) )),
-			    std::array<size_t,1>({1}), std::array<size_t,1>({0}));
+					     std::array<size_t,1>({1}), std::array<size_t,1>({0}))); } );
 
-        mi->ifor(1, mkGetExpr(op1,loop))();
+	//loop.dummy();
+        mi->pifor(1,loop)();
 	
 	auto i2_1 = imap.at("i2_1");
 	auto i2_2 = imap.at("i2_2");