From 523605e8d841733d7c398131ea50e356b35b88e3 Mon Sep 17 00:00:00 2001
From: Richard Thier <magosit@outlook.hu>
Date: Mon, 29 Sep 2025 18:51:53 +0200
Subject: [PATCH] hand unrolled thiersort3 - I think its slower than gcc
 unrolling and surely more complex so I will revert

---
 threepass_xbit.h | 93 +++++++++++++++++++++++++++++++++++-------------
 1 file changed, 68 insertions(+), 25 deletions(-)

diff --git a/threepass_xbit.h b/threepass_xbit.h
index 4298664..65673e3 100644
--- a/threepass_xbit.h
+++ b/threepass_xbit.h
@@ -63,7 +63,7 @@ static inline void threepass_xb(uint32_t *a, uint32_t *buf, int n) noexcept {
 		(1 << TPBX2),
 		(1 << TPBX3)
 	);
-	int i = 0;
+	uint32_t i = 0;
 	#pragma GCC unroll 8
 	for (; i < common; ++i) {
 		bucket1[i] += prev1;
@@ -74,61 +74,104 @@ static inline void threepass_xb(uint32_t *a, uint32_t *buf, int n) noexcept {
 		prev3 = bucket3[i];
 	}
 	/* Do remaining 1 */
-	for (int j = i; j < (1 << TPBX1); ++j) {
+	for (uint32_t j = i; j < (1 << TPBX1); ++j) {
 		bucket1[j] += prev1;
 		prev1 = bucket1[j];
 	}
 	/* Do remaining 2 */
-	for (int j = i; j< (1 << TPBX2); ++j) {
+	for (uint32_t j = i; j< (1 << TPBX2); ++j) {
 		bucket2[j] += prev2;
 		prev2 = bucket2[j];
 	}
 	/* Do remaining 3 */
-	for (int j = i; j < (1 << TPBX3); ++j) {
+	for (uint32_t j = i; j < (1 << TPBX3); ++j) {
 		bucket3[j] += prev3;
 		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
+	#pragma GCC unroll 13
+	for(i = n; i >= 4; i -= 4) {
+		auto num0 = a[i - 1];
+		auto bkeyni0 = (num0 >> shr3) & mask3;
+		auto offset0 = --bucket3[bkeyni0];
+		buf[offset0] = num0;
+		auto num1 = a[i - 2];
+		auto bkeyni1 = (num1 >> shr3) & mask3;
+		auto offset1 = --bucket3[bkeyni1];
+		buf[offset1] = num1;
+		auto num2 = a[i - 3];
+		auto bkeyni2 = (num2 >> shr3) & mask3;
+		auto offset2 = --bucket3[bkeyni2];
+		buf[offset2] = num2;
+		auto num3 = a[i - 4];
+		auto bkeyni3 = (num3 >> shr3) & mask3;
+		auto offset3 = --bucket3[bkeyni3];
+		buf[offset3] = num3;
+	}
+	#pragma GCC unroll 4
+	for(; i > 0; --i) {
 		auto num = a[i - 1];
 		auto bkeyni = (num >> shr3) & mask3;
 		auto offset = --bucket3[bkeyni];
-
-		// Add to the proper target location
 		buf[offset] = num;
 	}
+
 	// 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
+	#pragma GCC unroll 13
+	for(i = n; i >= 4; i -= 4) {
+		auto num0 = buf[i - 1];
+		auto bkeyni0 = (num0 >> shr2) & mask2;
+		auto offset0 = --bucket2[bkeyni0];
+		a[offset0] = num0;
+		auto num1 = buf[i - 2];
+		auto bkeyni1 = (num1 >> shr2) & mask2;
+		auto offset1 = --bucket2[bkeyni1];
+		a[offset1] = num1;
+		auto num2 = buf[i - 3];
+		auto bkeyni2 = (num2 >> shr2) & mask2;
+		auto offset2 = --bucket2[bkeyni2];
+		a[offset2] = num2;
+		auto num3 = buf[i - 4];
+		auto bkeyni3 = (num3 >> shr2) & mask2;
+		auto offset3 = --bucket2[bkeyni3];
+		a[offset3] = num3;
+	}
+	#pragma GCC unroll 4
+	for(; i > 0; --i) {
 		auto num = buf[i - 1];
 		auto bkeyni = (num >> shr2) & mask2;
 		auto offset = --bucket2[bkeyni];
-
-		// Add to the proper target location
 		a[offset] = num;
 	}
 	// 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
+	#pragma GCC unroll 13
+	for(i = n; i >= 4; i -= 4) {
+		auto num0 = a[i - 1];
+		auto bkeyni0 = (num0 >> shr1) & mask1;
+		auto offset0 = --bucket1[bkeyni0];
+		buf[offset0] = num0;
+		auto num1 = a[i - 2];
+		auto bkeyni1 = (num1 >> shr1) & mask1;
+		auto offset1 = --bucket1[bkeyni1];
+		buf[offset1] = num1;
+		auto num2 = a[i - 3];
+		auto bkeyni2 = (num2 >> shr1) & mask1;
+		auto offset2 = --bucket1[bkeyni2];
+		buf[offset2] = num2;
+		auto num3 = a[i - 4];
+		auto bkeyni3 = (num3 >> shr1) & mask1;
+		auto offset3 = --bucket1[bkeyni3];
+		buf[offset3] = num3;
+	}
+	#pragma GCC unroll 4
+	for(; i > 0; --i) {
 		auto num = a[i - 1];
 		auto bkeyni = (num >> shr1) & mask1;
 		auto offset = --bucket1[bkeyni];
-
-		// Add to the proper target location
 		buf[offset] = num;
 	}
 }