From d16505a297c25d0fc1e38955b5d62ab13a27334b Mon Sep 17 00:00:00 2001
From: Richard Thier <magosit@outlook.hu>
Date: Thu, 11 Apr 2024 19:00:52 +0200
Subject: [PATCH] mormordsort got template recursion for 33% speedup (I think
 it still has 2x maybe)

---
 ypsu.cpp | 36 ++++++++++++++++++++----------------
 1 file changed, 20 insertions(+), 16 deletions(-)

diff --git a/ypsu.cpp b/ypsu.cpp
index abc6112..81d579f 100644
--- a/ypsu.cpp
+++ b/ypsu.cpp
@@ -144,10 +144,12 @@ void twopass(uint32_t *a, int n) {
 //     --index
 // különben
 //     ++pivot_index
-static inline uint32_t morgrab(uint32_t elem, uint32_t j) noexcept {
+template<int j>
+static inline uint32_t morgrab(uint32_t elem) noexcept {
 	return (elem >> (8 * j)) & 0xff;
 }
-static inline void mormord_sort_impl(uint32_t *a, int n, int j) noexcept {
+template<int j>
+static inline void mormord_sort_impl(uint32_t *a, int n) noexcept {
 	/* Preparation */
 	uint32_t radics[256] = {0};
 	uint32_t radics2[256] = {0};
@@ -158,12 +160,13 @@ static inline void mormord_sort_impl(uint32_t *a, int n, int j) noexcept {
 	/* We can go both down and upwards here to increase ILP or even do SSE2 */
 	uint32_t k1 = 0;
 	uint32_t k2 = (n - 1);
+	#pragma GCC unroll 64
 	for(k1 = 0; k1 < k2; ++k1, --k2) {
-		++radics[morgrab(a[k1], j)];
-		++radics2[morgrab(a[k2], j)];
+		++radics[morgrab<j>(a[k1])];
+		++radics2[morgrab<j>(a[k2])];
 	}
 	if(k1 == k2) {
-		++radics[morgrab(a[k1], j)];
+		++radics[morgrab<j>(a[k1])];
 	}
 	for(int i = 0; i < 256; ++i) {
 		radics[i] += radics2[i];
@@ -179,6 +182,7 @@ static inline void mormord_sort_impl(uint32_t *a, int n, int j) noexcept {
 	/*       (because radix value 3 is not found in input)   */
 	uint32_t prev = 0;
 	uint32_t reali = 0;
+	#pragma GCC unroll 16
 	for(int i = 0; i < 256; ++i) {
 		if(radics[i] != 0) {
 			radics[i] += prev;
@@ -195,7 +199,7 @@ static inline void mormord_sort_impl(uint32_t *a, int n, int j) noexcept {
 	// Inplace swap
 	uint32_t pivoti = 0;
 	while(pivoti < n) {
-		uint32_t radixval = morgrab(a[pivoti], j);
+		uint32_t radixval = morgrab<j>(a[pivoti]);
 		uint32_t targeti = radics[radixval] - 1;
 		if(targeti > pivoti) {
 			// swap
@@ -211,20 +215,20 @@ static inline void mormord_sort_impl(uint32_t *a, int n, int j) noexcept {
 	}
 
 	// Ends recursion
-	if(j == 0) return;
-
-	// Recursion
-	for(int i = 0; i < reali; i += 2) {
-		/* inclusive */
-		uint32_t from = real_radics[i];
-		/* non-inclusive */
-		uint32_t to = real_radics[i + 1];
-		mormord_sort_impl(&a[from], (to - (from)), j - 1);
+	if constexpr (j != 0) {
+		// Recursion
+		for(int i = 0; i < reali; i += 2) {
+			/* inclusive */
+			uint32_t from = real_radics[i];
+			/* non-inclusive */
+			uint32_t to = real_radics[i + 1];
+			mormord_sort_impl<j - 1>(&a[from], (to - (from)));
+		}
 	}
 }
 static inline void mormord_sort(uint32_t *a, int n) noexcept {
 	assert(n * uint32_t(sizeof(a[0])) <= INT_MAX);
-	mormord_sort_impl(a, n, 3);
+	mormord_sort_impl<3>(a, n);
 }
 
 void fourpass(uint32_t *a, int n) {