diff --git a/threepass_xbit.h b/threepass_xbit.h
index 4298664..5c87d27 100644
--- a/threepass_xbit.h
+++ b/threepass_xbit.h
@@ -55,9 +55,9 @@ static inline void threepass_xb(uint32_t *a, uint32_t *buf, int n) noexcept {
 	}
 
 	/* Count prefix sums - try as much ILP as possible because bigger arrays than usual! */
-	uint32_t prev1 = 0;
-	uint32_t prev2 = 0;
-	uint32_t prev3 = 0;
+	int prev1 = -1;
+	int prev2 = -1;
+	int prev3 = -1;
 	uint32_t common = min3u32_xb(
 		(1 << TPBX1),
 		(1 << TPBX2),
@@ -89,47 +89,37 @@ static inline void threepass_xb(uint32_t *a, uint32_t *buf, int n) noexcept {
 		prev3 = bucket3[j];
 	}
 
-	// Bottom digit a->buf
-	// right-to-left to ensure already sorted digits order we keep for iterations
-	#pragma GCC unroll 48
-	for(uint32_t i = n; i > 0; --i) {
-		// Prefetch caches
-		//__builtin_prefetch(&a[i-8]);
-		// Get num and its new offset / location
-		auto num = a[i - 1];
-		auto bkeyni = (num >> shr3) & mask3;
-		auto offset = --bucket3[bkeyni];
+	/* Move elements (the 3 pass): This utilizes frewr trick to overwrite keys (elements) by ror-ing them */
+	/* This rewrites the key into its rotated state to save some pipeline stalls. (3 rotates happen and in the end its ok) */
+	/* right-to-left to ensure already sorted digits order we keep for iterations */
 
-		// Add to the proper target location
-		buf[offset] = num;
+	/* Bottom digit a->buf */
+	#pragma GCC unroll 80
+	for(uint32_t i = n; i > 0; --i) {
+		uint32_t num = a[i - 1];
+		/* save rotated into the masked, bucketed loc */
+		buf[bucket3[num & mask3]--] = (num >> TPBX3) | (num << (32 - TPBX3));
+		//__builtin_prefetch(&buf[bucket3[num & mask3] - 2]);
 	}
-	// Mid digit buf->a
-	// right-to-left to ensure already sorted digits order we keep for iterations
-	#pragma GCC unroll 48
-	for(uint32_t i = n; i > 0; --i) {
-		// Prefetch caches
-		//__builtin_prefetch(&buf[i-8]);
-		// Get num and its new offset / location
-		auto num = buf[i - 1];
-		auto bkeyni = (num >> shr2) & mask2;
-		auto offset = --bucket2[bkeyni];
 
-		// Add to the proper target location
-		a[offset] = num;
+	/* Mid digit buf->a */
+	#pragma GCC unroll 80
+	for(uint32_t i = n; i > 0; --i) {
+		uint32_t num = buf[i - 1];
+		/* save rotated into the masked, bucketed loc */
+		a[bucket2[num & mask2]--] = (num >> TPBX2) | (num << (32 - TPBX2));
+		//__builtin_prefetch(&a[bucket2[num & mask2] - 2]);
 	}
-	// Top digit a->buf
-	// right-to-left to ensure already sorted digits order we keep for iterations
-	#pragma GCC unroll 48
-	for(uint32_t i = n; i > 0; --i) {
-		// Prefetch caches
-		// __builtin_prefetch(&a[i-16]);
-		// Get num and its new offset / location
-		auto num = a[i - 1];
-		auto bkeyni = (num >> shr1) & mask1;
-		auto offset = --bucket1[bkeyni];
 
-		// Add to the proper target location
-		buf[offset] = num;
+	/* Top digit a->buf */
+	#pragma GCC unroll 80
+	for(uint32_t i = n; i > 0; --i) {
+		uint32_t num = a[i - 1];
+		/* Reconstruct the original key in this element by where its stuff is stored */
+		constexpr int rot = TPBX1 + (32 - TPBX1 - TPBX2 - TPBX3);
+		/* save rotated into the masked, bucketed loc */
+		buf[bucket1[num & mask1]--] = (num >> rot) | (num << (32 - rot));
+		//__builtin_prefetch(&buf[bucket1[num & mask1] - 2]);
 	}
 }