schwab: for-loop instead of while - does not seems like speedup

schwab_sort.h

@@ -1,14 +1,19 @@
-#ifndef SWAB_SORT_H
-#define SWAB_SORT_H
+#ifndef SCHWAB_SORT_H
+#define SCHWAB_SORT_H
 
 /* A fast quicksort-like new alg created in Csolnok, Hungary with:
  * - 4-way partitioning with 0..5 copies (not swaps) per elem per run
  * - ensured O(log2(n)) worst recursion depth
  *
- * LICENCE: CC-BY, 2025 May 08
+ * LICENCE: CC-BY, 2025 May 08-09
  * Author: Richárd István Thier (also author of the Magyarsort)
  */
 
+/** 3-way optimization for smallrange and const - 0 turns this off! */
+#ifndef SCHWAB_DELTA_THRESHOLD
+#define SCHWAB_DELTA_THRESHOLD 32
+#endif /* SCHWAB_DELTA_THRESHOLD */
+
 typedef uint32_t sch_rand_state;
 
 /** Create rand state for schwab_sort using a seed - can give 0 if uninterested */
@@ -29,6 +34,71 @@ static inline uint32_t schwab_pick_pivot(sch_rand_state *state, uint32_t len) {
 	return (uint32_t)(((uint64_t)rand * len) >> 32);
 }
 
+/** Swap operation */
+static inline void schwab_swap(uint32_t *a, uint32_t *b) {
+	uint32_t t = *a;
+	*a = *b;
+	*b = t;
+}
+
+/**
+ * 3-way partitioning, in middle all the pivot elements.
+ *
+ * Single-pass version 2.0, taken from qsort.h
+ *
+ * @param array The array to partition
+ * @param low From when. (inclusive)
+ * @param high Until when. (inclusive too!)
+ * @param pivotval This value is used to partition the array.
+ * @param plo OUT: until this, more processing might needed.
+ * @param phi OUT: from this, more processing might needed.
+ */
+static inline void schwab_partition3sp2(
+		uint32_t *array,
+		int low,
+		int high,
+		uint32_t pivotval,
+		int *plo,
+		int *phi) {
+
+	/* Invariant for left: index until smaller (than pivot) elements lay */
+	int il = (low - 1);
+	/* Invariant for right: index until (from top) bigger elements lay */
+	int ir = (high + 1);
+	/* Indices from where we swap left and right into "is" (and sometimes swap among here too) */
+	int jl = low;
+	int jr = high;
+
+	while(jl <= jr) {
+		/* Handle left and find wrongly placed element */
+		while((array[jl] <= pivotval) && (jl <= jr)) {
+			int isNonPivot = (array[jl] != pivotval);
+			int nonSameIndex = (il + 1 != jl);
+			if(isNonPivot & nonSameIndex)
+				schwab_swap(&array[il + 1], &array[jl]);
+			il += isNonPivot;
+			++jl;
+		}
+
+		/* Handle right and find wrongly placed element */
+		while((array[jr] >= pivotval) && (jl <= jr)) {
+			int isNonPivot = (array[jr] != pivotval);
+			int nonSameIndex = (ir - 1 != jr);
+			if(isNonPivot & nonSameIndex)
+					schwab_swap(&array[ir - 1], &array[jr]);
+			ir -= isNonPivot;
+			--jr;
+		}
+
+		/* Swap the two found elements that are wrongly placed */
+		if(jl < jr) schwab_swap(&array[jl], &array[jr]);
+	}
+
+	/* Output the partition points */
+	*plo = il + 1; /* XXX: changed from qsort.h to +1 here! */
+	*phi = ir;
+}
+
 /**
  * 4-way partitioning
  *
@@ -41,12 +111,13 @@ static inline uint32_t schwab_pick_pivot(sch_rand_state *state, uint32_t len) {
  * @param arr The array to partition
  * @param low Inclusive smallest index.
  * @param high Inclusive highest index.
- * @param plo IN-OUT: input low pivot, output index until elements <= low pivot.
+ * @param plo IN-OUT: input low pivot, output - see "Results:"
  * @param kmid IN: The mid spliting value (like a pivot value, but can be imaginary nonexistent)
- * @param pmid OUT: output index until elements <= mid pivot.
- * @param phi IN-OUT: input high pivot, output index until elements <= high pivot.
+ * @param pmid OUT: output - see "Results:"
+ * @param phi IN-OUT: input high pivot, output - see "Results:"
+ * @returns 1 if there is need to process the mid two blocks! Otherwise 0.
  */
-static inline void schwab_partition(
+static inline int schwab_partition(
 		uint32_t *arr,
 		int low,
 		int high,
@@ -59,18 +130,28 @@ static inline void schwab_partition(
 	uint32_t klo = arr[*plo];
 	uint32_t khi = arr[*phi];
 
+	/* Without this, constant and smallrange is very slooOOoow */
+	if(khi - klo < SCHWAB_DELTA_THRESHOLD) {
+		/* Use three-way which defeats smallrange */
+		/* Outer sort func also optimized for two sides */
+		/* check for size for which recurse which not! */
+		schwab_partition3sp2(arr, low, high, kmid, plo, phi);
+
+		/* No need to process the midle two blocks - all pivot there */
+		return 0;
+	}
+
 	/* [*] Swapping arr[phi]<->arr[high] ensures stop condition later */
 	uint32_t tmphi = arr[*phi];
 	arr[*phi] = arr[high];
 	arr[high] = tmphi;
 
 	/* Aren't inclusive end indices of 4 "blocks" - b0 is smallest vals */
-	int b0 = low, b1 = low, b2 = low, b3 = low;
+	int b0 = low, b1 = low, b2 = low;
 
-	while(b3 < high) {
-		/* This I moved to be first for hot code path for constant / smallrange */
+	for(int b3 = low; b3 < high; ++b3) {
+		/* This I moved to be first to avoid unnecessary curr copy below */
 		if(arr[b3] >= khi) {
-			++b3;
 			continue;
 		}
 
@@ -87,7 +168,7 @@ static inline void schwab_partition(
 			arr[b2] = arr[b1];
 			arr[b1] = arr[b0];
 			arr[b0] = curr;
-			++b0; ++b1; ++b2; ++b3;
+			++b0; ++b1; ++b2;
 			continue;
 		}
 
@@ -95,11 +176,11 @@ static inline void schwab_partition(
 			arr[b3] = arr[b2];
 			arr[b2] = arr[b1];
 			arr[b1] = curr;
-			++b1; ++b2; ++b3;
+			++b1; ++b2;
 		} else {
 			arr[b3] = arr[b2];
 			arr[b2] = curr;
-			++b2; ++b3;
+			++b2;
 		}
 	}
 
@@ -114,7 +195,10 @@ static inline void schwab_partition(
 	/* Handle output vars as per doc comment */
 	*plo = b0;
 	*pmid = b1;
-	*phi = b2; /* Because of: [*] */
+	*phi = b2;
+
+	/* There are mid parts to process */
+	return 1;
 }
 
 /** Swabic-sort its somewhat similar to quicksort but 4-way and tricky */
@@ -144,7 +228,7 @@ static inline void schwab_sort(
 		uint32_t kmid = klo + (khi - klo) / 2;
 
 		int pmid;
-		schwab_partition(array, low, high, &plo, kmid, &pmid, &phi);
+		int needmid = schwab_partition(array, low, high, &plo, kmid, &pmid, &phi);
 
 		/* See where NOT to recurse to avoid worst case stack depth */
 		/* Rem.: These might be "not real" length but we only use them to comparisons */
@@ -160,15 +244,19 @@ static inline void schwab_sort(
 		/*       possible, but had to put loops changes to the end */
 		if(lolen < hilen) {
 			schwab_sort(array, low, plo - 1, state);
-			schwab_sort(array, plo, pmid - 1, state);
-			schwab_sort(array, pmid, phi - 1, state);
+			if(needmid) {
+				schwab_sort(array, plo, pmid - 1, state);
+				schwab_sort(array, pmid, phi - 1, state);
+			}
 
 			low = phi;
 			/* high = high; */
 		} else {
 			schwab_sort(array, phi, high, state);
-			schwab_sort(array, pmid, phi - 1, state);
-			schwab_sort(array, plo, pmid - 1, state);
+			if(needmid) {
+				schwab_sort(array, pmid, phi - 1, state);
+				schwab_sort(array, plo, pmid - 1, state);
+			}
 
 			/* low = low; */
 			high = plo - 1;
@@ -176,4 +264,4 @@ static inline void schwab_sort(
 	}
 }
 
-#endif /* SWAB_SORT_H */
+#endif /* SCHWAB_SORT_H */