diff --git a/ypsu.cpp b/ypsu.cpp
index 0f778d7..0da4382 100644
--- a/ypsu.cpp
+++ b/ypsu.cpp
@@ -185,18 +185,16 @@ static inline void mormord_sort_impl(uint32_t *a, int n) noexcept {
 	/*       (because radix value 3 is not found in input)   */
 	uint32_t prev = 0;
 	uint32_t reali = 0;
-	#pragma GCC unroll 16
+	#pragma GCC unroll 256
 	for(int i = 0; i < 256; ++i) {
 		radics[i] += prev;
-		if(radics[i] != 0) {
-			real_radics[reali] = prev;
-			real_radics[reali + 1] = radics[i];
-			reali += 2;
-		}
+		bool act =(radics[i] != 0);
+		real_radics[reali] = act ? prev : real_radics[reali];
+		real_radics[reali + 1] = act ? radics[i] : real_radics[reali + 1];
+		reali += 2 * act;
 		prev = radics[i];
 	}
 
-
 	// Inplace swap - own ideas + some ideas based-on "famous" ct-swap (for branchless / more ILP):
 	//	void ct-swap(bool secret, uint64_t a[], uint64_t b[], size_t len) {
 	//		uint64_t mask = ~((uint64_t)secret - 1); // 1->111....111; 0->000....000