vector: Timsort all of the things

Drop the GLibc implementation of Merge Sort and replace it with Timsort.

The algorithm has been tuned to work on arrays of pointers (void **),
so there's no longer a need to abstract the byte-width of each element
in the array.

All the comparison callbacks now take pointers-to-elements, not
pointers-to-pointers, so there's now one less level of dereferencing.

E.g.

	 int index_cmp(const void *a, const void *b)
	 {
	-	const git_index_entry *entry_a = *(const git_index_entry **)(a);
	+	const git_index_entry *entry_a = (const git_index_entry *)(a);

The result is up to a 40% speed-up when sorting vectors. Memory usage
remains lineal.

A new `bsearch` implementation has been added, whose callback also
supplies pointer-to-elements, to uniform the Vector API again.

1 files changed, 380 insertions, 0 deletions

diff --git a/src/tsort.c b/src/tsort.c
new file mode 100644
index 000000000..eda9a2fd2
--- /dev/null
+++ b/src/tsort.c
@@ -0,0 +1,380 @@
+#include <stdio.h>
+#include <string.h>
+#include <stdint.h>
+#include <stdlib.h>
+
+/**
+ * An array-of-pointers implementation of Python's Timsort
+ * Based on code by Christopher Swenson under the MIT license
+ *
+ * Copyright (c) 2010 Christopher Swenson
+ * Copyright (c) 2011 Vicent Marti
+ */
+
+#ifndef MAX
+#	define MAX(x,y) (((x) > (y) ? (x) : (y)))
+#endif
+
+#ifndef MIN
+#	define MIN(x,y) (((x) < (y) ? (x) : (y)))
+#endif
+
+#if defined(__GNUC__)
+#	define CLZ(x) __builtin_clz(x)
+#elif defined(_MSV_VER)
+#	define CLZ(x) _CountLeadingZeros(x)
+#else
+int CLZ(int32_t x)
+{
+    int e = 31;
+
+    if (!x) return 32;
+    if (x&0xFFFF0000)   { e -=16; x >>=16; }
+    if (x&0x0000FF00)   { e -= 8; x >>= 8; }
+    if (x&0x000000F0)   { e -= 4; x >>= 4; }
+    if (x&0x0000000C)   { e -= 2; x >>= 2; }
+    if (x&0x00000002)   { e -= 1; }
+    return e;
+}
+#endif
+
+typedef int (*cmp_ptr_t)(const void *, const void *);
+
+static int binsearch(void **dst, const void *x, size_t size, cmp_ptr_t cmp)
+{
+	int l, c, r;
+	void *lx, *cx, *rx;
+
+	l = 0;
+	r = size - 1;
+	c = r >> 1;
+	lx = dst[l];
+
+	/* check for beginning conditions */
+	if (cmp(x, lx) < 0)
+		return 0;
+
+	else if (cmp(x, lx) == 0) {
+		int i = 1;
+		while (cmp(x, dst[i]) == 0)
+			i++;
+		return i;
+	}
+
+	rx = dst[r];
+	/* guaranteed not to be >= rx */
+	cx = dst[c];
+	while (1) {
+		const int val = cmp(x, cx);
+		if (val < 0) {
+			if (c - l <= 1) return c;
+			r = c;
+			rx = cx;
+		} else if (val > 0) {
+			if (r - c <= 1) return c + 1;
+			l = c;
+			lx = cx;
+		} else {
+			do {
+				cx = dst[++c];
+			} while (cmp(x, cx) == 0);
+			return c;
+		}
+		c = l + ((r - l) >> 1);
+		cx = dst[c];
+	}
+}
+
+/* Binary insertion sort, but knowing that the first "start" entries are sorted.  Used in timsort. */
+static void bisort(void **dst, size_t start, size_t size, cmp_ptr_t cmp)
+{
+	size_t i;
+
+	for (i = start; i < size; i++) {
+		int j;
+		/* If this entry is already correct, just move along */
+		if (cmp(dst[i - 1], dst[i]) <= 0)
+			continue;
+
+		/* Else we need to find the right place, shift everything over, and squeeze in */
+		void *x = dst[i];
+		int location = binsearch(dst, x, i, cmp);
+		for (j = i - 1; j >= location; j--) {
+			dst[j + 1] = dst[j];
+		}
+		dst[location] = x;
+	}
+}
+
+
+/* timsort implementation, based on timsort.txt */
+struct tsort_run {
+	ssize_t start;
+	ssize_t length;
+};
+
+struct tsort_store {
+	size_t alloc;
+	cmp_ptr_t cmp;
+	void **storage;
+};
+
+static void reverse_elements(void **dst, int start, int end)
+{
+	while (start < end) {
+		void *tmp = dst[start];
+		dst[start] = dst[end];
+		dst[end] = tmp;
+
+		start++;
+		end--;
+	}
+}
+
+static int count_run(void **dst, ssize_t start, ssize_t size, struct tsort_store *store)
+{
+	ssize_t curr = start + 2;
+
+	if (size - start == 1)
+		return 1;
+
+	if (start >= size - 2) {
+		if (store->cmp(dst[size - 2], dst[size - 1]) > 0) {
+			void *tmp = dst[size - 1];
+			dst[size - 1] = dst[size - 2];
+			dst[size - 2] = tmp;
+		}
+
+		return 2;
+	}
+
+	if (store->cmp(dst[start], dst[start + 1]) <= 0) {
+		while (curr < size - 1 && store->cmp(dst[curr - 1], dst[curr]) <= 0)
+			curr++;
+
+		return curr - start;
+	} else {
+		while (curr < size - 1 && store->cmp(dst[curr - 1], dst[curr]) > 0)
+			curr++;
+
+		/* reverse in-place */
+		reverse_elements(dst, start, curr - 1);
+		return curr - start;
+	}
+}
+
+static int compute_minrun(size_t n)
+{
+	int r = 0;
+	while (n >= 64) {
+		r |= n & 1;
+		n >>= 1;
+	}
+	return n + r;
+}
+
+static int check_invariant(struct tsort_run *stack, int stack_curr)
+{
+	if (stack_curr < 2)
+		return 1;
+
+	else if (stack_curr == 2) {
+		const int A = stack[stack_curr - 2].length;
+		const int B = stack[stack_curr - 1].length;
+		return (A > B);
+	} else {
+		const int A = stack[stack_curr - 3].length;
+		const int B = stack[stack_curr - 2].length;
+		const int C = stack[stack_curr - 1].length;
+		return !((A <= B + C) || (B <= C));
+	}
+}
+
+static int resize(struct tsort_store *store, size_t new_size)
+{
+	if (store->alloc < new_size) {
+		void **tempstore = realloc(store->storage, new_size * sizeof(void *));
+
+		/**
+		 * Do not propagate on OOM; this will abort the sort and
+		 * leave the array unsorted, but no error code will be
+		 * raised
+		 */
+		if (tempstore == NULL)
+			return -1;
+
+		store->storage = tempstore;
+		store->alloc = new_size;
+	}
+
+	return 0;
+}
+
+static void merge(void **dst, const struct tsort_run *stack, int stack_curr, struct tsort_store *store)
+{
+	const ssize_t A = stack[stack_curr - 2].length;
+	const ssize_t B = stack[stack_curr - 1].length;
+	const ssize_t curr = stack[stack_curr - 2].start;
+
+	void **storage;
+	ssize_t i, j, k;
+
+	if (resize(store, MIN(A, B)) < 0)
+		return;
+
+	storage =  store->storage;
+
+	/* left merge */
+	if (A < B) {
+		memcpy(storage, &dst[curr], A * sizeof(void *));
+		i = 0;
+		j = curr + A;
+
+		for (k = curr; k < curr + A + B; k++) {
+			if ((i < A) && (j < curr + A + B)) {
+				if (store->cmp(storage[i], dst[j]) <= 0)
+					dst[k] = storage[i++];
+				else
+					dst[k] = dst[j++];
+			} else if (i < A) {
+				dst[k] = storage[i++];
+			} else
+				dst[k] = dst[j++];
+		}
+	} else {
+		memcpy(storage, &dst[curr + A], B * sizeof(void *));
+		i = B - 1;
+		j = curr + A - 1;
+
+		for (k = curr + A + B - 1; k >= curr; k--) {
+			if ((i >= 0) && (j >= curr)) {
+				if (store->cmp(dst[j], storage[i]) > 0)
+					dst[k] = dst[j--];
+				else
+					dst[k] = storage[i--];
+			} else if (i >= 0)
+				dst[k] = storage[i--];
+			else
+				dst[k] = dst[j--];
+		}
+	}
+}
+
+static ssize_t collapse(void **dst, struct tsort_run *stack, ssize_t stack_curr, struct tsort_store *store, ssize_t size)
+{
+	ssize_t A, B, C;
+
+	while (1) {
+		/* if the stack only has one thing on it, we are done with the collapse */
+		if (stack_curr <= 1)
+			break;
+
+		/* if this is the last merge, just do it */
+		if ((stack_curr == 2) && (stack[0].length + stack[1].length == size)) {
+			merge(dst, stack, stack_curr, store);
+			stack[0].length += stack[1].length;
+			stack_curr--;
+			break;
+		}
+
+		/* check if the invariant is off for a stack of 2 elements */
+		else if ((stack_curr == 2) && (stack[0].length <= stack[1].length)) {
+			merge(dst, stack, stack_curr, store);
+			stack[0].length += stack[1].length;
+			stack_curr--;
+			break;
+		}
+		else if (stack_curr == 2)
+			break;
+
+		A = stack[stack_curr - 3].length;
+		B = stack[stack_curr - 2].length;
+		C = stack[stack_curr - 1].length;
+
+		/* check first invariant */
+		if (A <= B + C) {
+			if (A < C) {
+				merge(dst, stack, stack_curr - 1, store);
+				stack[stack_curr - 3].length += stack[stack_curr - 2].length;
+				stack[stack_curr - 2] = stack[stack_curr - 1];
+				stack_curr--;
+			} else {
+				merge(dst, stack, stack_curr, store);
+				stack[stack_curr - 2].length += stack[stack_curr - 1].length;
+				stack_curr--;
+			}
+		} else if (B <= C) {
+			merge(dst, stack, stack_curr, store);
+			stack[stack_curr - 2].length += stack[stack_curr - 1].length;
+			stack_curr--;
+		} else
+			break;
+	}
+
+	return stack_curr;
+}
+
+#define PUSH_NEXT() do {\
+	len = count_run(dst, curr, size, store);\
+	run = minrun;\
+	if (run < minrun) run = minrun;\
+	if (run > (ssize_t)size - curr) run = size - curr;\
+	if (run > len) {\
+		bisort(&dst[curr], len, run, cmp);\
+		len = run;\
+	}\
+	run_stack[stack_curr++] = (struct tsort_run) {curr, len};\
+	curr += len;\
+	if (curr == (ssize_t)size) {\
+		/* finish up */ \
+		while (stack_curr > 1) { \
+			merge(dst, run_stack, stack_curr, store); \
+			run_stack[stack_curr - 2].length += run_stack[stack_curr - 1].length; \
+			stack_curr--; \
+		} \
+		if (store->storage != NULL) {\
+			free(store->storage);\
+			store->storage = NULL;\
+		}\
+		return;\
+	}\
+}\
+while (0)
+
+
+void git__tsort(void **dst, size_t size, cmp_ptr_t cmp)
+{
+	struct tsort_store _store, *store = &_store;
+	struct tsort_run run_stack[128];
+
+	ssize_t stack_curr = 0;
+	ssize_t len, run;
+	ssize_t curr = 0;
+	ssize_t minrun;
+
+	if (size < 64) {
+		bisort(dst, 1, size, cmp);
+		return;
+	}
+
+	/* compute the minimum run length */
+	minrun = compute_minrun(size);
+
+	/* temporary storage for merges */
+	store->alloc = 0;
+	store->storage = NULL;
+	store->cmp = cmp;
+
+	PUSH_NEXT();
+	PUSH_NEXT();
+	PUSH_NEXT();
+
+	while (1) {
+		if (!check_invariant(run_stack, stack_curr)) {
+			stack_curr = collapse(dst, run_stack, stack_curr, store, size);
+			continue;
+		}
+
+		PUSH_NEXT();
+	}
+}