diff --git a/magyarsort.h b/magyarsort.h
index 66d6aee..8394fe9 100644
--- a/magyarsort.h
+++ b/magyarsort.h
@@ -242,10 +242,10 @@ namespace MagyarSort {
 	/** Recursive Functor: no class should be generated I think (compiler should be smart) */
 	template<int DIGIT, typename COUNTER_TYP>
 	struct RadixMagic : public RadixMagic<DIGIT - 1, COUNTER_TYP> {
-		inline __attribute__((always_inline)) RadixMagic(COUNTER_TYP *radics, uint32_t *from, uint32_t *to, COUNTER_TYP size) noexcept
-				: RadixMagic<DIGIT - 1, COUNTER_TYP>(radics, from, to, size) {
+		inline __attribute__((always_inline)) RadixMagic(bool &swapped, COUNTER_TYP *radics, uint32_t *from, uint32_t *to, COUNTER_TYP size) noexcept
+				: RadixMagic<DIGIT - 1, COUNTER_TYP>(swapped, radics, from, to, size) {
 			// Tricky: see (**)
-			if(this->swapped) { // never true for DIGIT 0
+			if(swapped) { // never true for DIGIT 0, see (***)
 				std::swap(from, to);
 			}
 
@@ -274,14 +274,13 @@ namespace MagyarSort {
 			//debugArr(to, size);
 
 			// (**) Only swaps pointers above in the child class constructor IF NEEDED :-)
-			this->swapped = !this->swapped;
+			swapped = !swapped;
 		}
 	};
 	/** Ends template recursion */
 	template<typename COUNTER_TYP>
 	struct RadixMagic<-1, COUNTER_TYP> {
-		bool swapped = false;
-		inline RadixMagic(COUNTER_TYP *radics, uint32_t *&from, uint32_t *&to, COUNTER_TYP size) noexcept { }
+		inline RadixMagic(bool swapped, COUNTER_TYP *radics, uint32_t *&from, uint32_t *&to, COUNTER_TYP size) noexcept {}
 	};
 
 	/* SORT */
@@ -361,13 +360,15 @@ namespace MagyarSort {
 
 		uint32_t *from = arr;
 		uint32_t *to = &arc[0];
+		static thread_local bool swapped;
+		swapped = false; // must be separate line
 
-		RadixMagic<DIGITS - 1, COUNTER_TYP> r(radics, from, to, size);
+		RadixMagic<DIGITS - 1, COUNTER_TYP> r(swapped, radics, from, to, size);
 
 		// With an other API we could spare this copy if we can delete original arr and return ptr or something...
 		// I am fine with this... this is not my main idea anyways, just little ILP tweak to regular radix sort
 		//if(to != arr) // <- logically, but bad they are already swapped here!!! BEWARE
-		if(r.swapped) { // <- in reality this is what we want because of last swap happened anyways!
+		if(swapped) { // <- in reality this is what we want because of last swap happened anyways!
 			memcpy(arr, to, size);
 		}
 	}