diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0342bcc..d3d5c2a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -2,7 +2,7 @@ cmake_minimum_required(VERSION 2.8)
 
 project(multi_array)
 
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Werror -std=c++11 -Wpedantic -O3 -g")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Werror -std=c++11 -Wpedantic -O3 -g -march=native")
 
 enable_testing()
 
diff --git a/src/include/mbase_def.h b/src/include/mbase_def.h
index f3fb0bb..6d569c5 100644
--- a/src/include/mbase_def.h
+++ b/src/include/mbase_def.h
@@ -51,6 +51,14 @@ namespace MultiArrayTools
     // multi_array_operation.h
     template <typename T, class Op, class IndexType>
     class Contraction;
+
+    // slice.h
+    template <typename T, class... SRanges>
+    class Slice;
+
+    // slice.h
+    template <typename T, class... SRanges>
+    class SliceDef;
 }
 
 #endif
diff --git a/src/include/pack_num.h b/src/include/pack_num.h
index eed17c4..21d85dd 100644
--- a/src/include/pack_num.h
+++ b/src/include/pack_num.h
@@ -51,7 +51,7 @@ namespace MultiArrayHelper
 	}
 
 	template <class OpTuple, class Expr>
-	static auto mkLoop( const OpTuple& ot, Expr&& exp )
+	static auto mkLoop( const OpTuple& ot, Expr exp )
 	    -> decltype(std::get<N>(ot).loop( PackNum<N-1>::mkLoop(ot,exp) ))
 	{
 	    return std::get<N>(ot).loop( PackNum<N-1>::mkLoop(ot,exp) );
@@ -92,7 +92,7 @@ namespace MultiArrayHelper
 	}
 
 	template <class OpTuple, class Expr>
-	static auto mkLoop( const OpTuple& ot, Expr&& exp )
+	static auto mkLoop( const OpTuple& ot, Expr exp )
 	    -> decltype(std::get<0>(ot).loop( exp ))
 	{
 	    return std::get<0>(ot).loop( exp );
diff --git a/src/include/slice.h b/src/include/slice.h
index cf8d0d7..ec6ff8f 100644
--- a/src/include/slice.h
+++ b/src/include/slice.h
@@ -35,11 +35,26 @@ namespace MultiArrayTools
 
 	template <class... MARanges>
 	auto define(const std::shared_ptr<typename SRanges::IType>&... inds)
-	    -> SliceDef<T,OperationRoot<T,MARanges...> >;
+	    -> SliceDef<T,MARanges...>;
 	
     private:
 	T* mData;
     };
+
+
+    template <typename T, class... SRanges>
+    class SliceDef
+    {
+    private:
+	SliceDef() = default;
+	
+    public:
+	SliceDef(Slice<T,SRanges...>& sl,
+		 const std::shared_ptr<typename SRanges::IndexType>&... inds);
+
+	template <class... ORanges>
+	SliceDef& operator=(OperationRoot<T,ORanges...>& op);
+    };
     
 } // end namespace MultiArrayTools
 
@@ -122,8 +137,14 @@ namespace MultiArrayTools
 	//i = std::get<sizeof...(SRanges)>(mBlockSizes);
 	return i.setData(data());
     }
-
+    /*
+    SliceDef<T,SRanges...>::
+    SliceDef(Slice<T,SRanges...>& sl,
+	     const std::shared_ptr<typename SRanges::IndexType>&... inds);
     
+    template <class... ORanges>
+    SliceDef<T,SRanges...>& SliceDef<T,SRanges...>::operator=(OperationRoot<T,ORanges...>& op);
+    */
 
 } // end namespace MultiArrayTools
 
diff --git a/src/tests/op_perf_test.cc b/src/tests/op_perf_test.cc
index b08775e..f42ecfd 100644
--- a/src/tests/op_perf_test.cc
+++ b/src/tests/op_perf_test.cc
@@ -174,24 +174,28 @@ namespace {
 		  << std::endl;
 	
 	std::vector<double> vres(4*4);
-
-	std::clock_t begin2 = std::clock();
-	for(size_t i = 0; i != 1000; ++i){
 	for(size_t d = 0; d != 4; ++d){
 	    for(size_t p = 0; p != 4; ++p){
 		const size_t tidx = d*4 + p;
 		vres[tidx] = 0.;
-		for(size_t a = 0; a != 4; ++a){
-		    for(size_t b = 0; b != 4; ++b){
-			for(size_t c = 0; c != 4; ++c){
-			    const size_t sidx = d*4*4*4*4*4*4*4 + a*5*4*4*4*4*4 + b*5*4*4*4 + + c*5*4  + p;
-			    vres[tidx] += data[sidx];
+	    }
+	}
+	std::clock_t begin2 = std::clock();
+	for(size_t i = 0; i != 1000; ++i){
+	    for(size_t a = 0; a != 4; ++a){
+		for(size_t b = 0; b != 4; ++b){
+		    for(size_t c = 0; c != 4; ++c){
+			for(size_t d = 0; d != 4; ++d){
+			    for(size_t p = 0; p != 4; ++p){
+				const size_t tidx = d*4 + p;
+				const size_t sidx = d*4*4*4*4*4*4*4 + a*5*4*4*4*4*4 + b*5*4*4*4 + + c*5*4  + p;
+				vres[tidx] += data[sidx];
+			    }
 			}
 		    }
 		}
 	    }
 	}
-	}
 	std::clock_t end2 = std::clock();
 
 	assert( xround(res1.at(mkts(0,0))) == xround(vres[0]) );