add memmem function (gnu extension)

based on strstr. passes gnulib tests and a few quick checks of my own.

include/string.h

@@ -95,6 +95,7 @@ int strcasecmp_l (const char *, const char *, locale_t);
 int strncasecmp_l (const char *, const char *, size_t, locale_t);
 char *strchrnul(const char *, int);
 char *strcasestr(const char *, const char *);
+void *memmem(const void *, size_t, const void *, size_t);
 void *memrchr(const void *, int, size_t);
 void *mempcpy(void *, const void *, size_t);
 #ifndef __cplusplus

src/string/memmem.c

+#define _GNU_SOURCE
+#include <string.h>
+#include <stdlib.h>
+#include <stdint.h>
+
+static char *twobyte_memmem(const unsigned char *h, size_t k, const unsigned char *n)
+{
+	uint16_t nw = n[0]<<8 | n[1], hw = h[0]<<8 | h[1];
+	for (h++, k--; k; k--, hw = hw<<8 | *++h)
+		if (hw == nw) return (char *)h-1;
+	return 0;
+}
+
+static char *threebyte_memmem(const unsigned char *h, size_t k, const unsigned char *n)
+{
+	uint32_t nw = n[0]<<24 | n[1]<<16 | n[2]<<8;
+	uint32_t hw = h[0]<<24 | h[1]<<16 | h[2]<<8;
+	for (h+=2, k-=2; k; k--, hw = (hw|*++h)<<8)
+		if (hw == nw) return (char *)h-2;
+	return 0;
+}
+
+static char *fourbyte_memmem(const unsigned char *h, size_t k, const unsigned char *n)
+{
+	uint32_t nw = n[0]<<24 | n[1]<<16 | n[2]<<8 | n[3];
+	uint32_t hw = h[0]<<24 | h[1]<<16 | h[2]<<8 | h[3];
+	for (h+=3, k-=3; k; k--, hw = hw<<8 | *++h)
+		if (hw == nw) return (char *)h-3;
+	return 0;
+}
+
+#define MAX(a,b) ((a)>(b)?(a):(b))
+#define MIN(a,b) ((a)<(b)?(a):(b))
+
+#define BITOP(a,b,op) \
+ ((a)[(size_t)(b)/(8*sizeof *(a))] op (size_t)1<<((size_t)(b)%(8*sizeof *(a))))
+
+static char *twoway_memmem(const unsigned char *h, const unsigned char *z, const unsigned char *n, size_t l)
+{
+	size_t i, ip, jp, k, p, ms, p0, mem, mem0;
+	size_t byteset[32 / sizeof(size_t)] = { 0 };
+	size_t shift[256];
+
+	/* Computing length of needle and fill shift table */
+	for (i=0; i<l; i++)
+		BITOP(byteset, n[i], |=), shift[n[i]] = i+1;
+
+	/* Compute maximal suffix */
+	ip = -1; jp = 0; k = p = 1;
+	while (jp+k<l) {
+		if (n[ip+k] == n[jp+k]) {
+			if (k == p) {
+				jp += p;
+				k = 1;
+			} else k++;
+		} else if (n[ip+k] > n[jp+k]) {
+			jp += k;
+			k = 1;
+			p = jp - ip;
+		} else {
+			ip = jp++;
+			k = p = 1;
+		}
+	}
+	ms = ip;
+	p0 = p;
+
+	/* And with the opposite comparison */
+	ip = -1; jp = 0; k = p = 1;
+	while (jp+k<l) {
+		if (n[ip+k] == n[jp+k]) {
+			if (k == p) {
+				jp += p;
+				k = 1;
+			} else k++;
+		} else if (n[ip+k] < n[jp+k]) {
+			jp += k;
+			k = 1;
+			p = jp - ip;
+		} else {
+			ip = jp++;
+			k = p = 1;
+		}
+	}
+	if (ip+1 > ms+1) ms = ip;
+	else p = p0;
+
+	/* Periodic needle? */
+	if (memcmp(n, n+p, ms+1)) {
+		mem0 = 0;
+		p = MAX(ms, l-ms-1) + 1;
+	} else mem0 = l-p;
+	mem = 0;
+
+	/* Search loop */
+	for (;;) {
+		/* If remainder of haystack is shorter than needle, done */
+		if (z-h < l) return 0;
+
+		/* Check last byte first; advance by shift on mismatch */
+		if (BITOP(byteset, h[l-1], &)) {
+			k = l-shift[h[l-1]];
+			if (k) {
+				if (mem0 && mem && k < p) k = l-p;
+				h += k;
+				mem = 0;
+				continue;
+			}
+		} else {
+			h += l;
+			mem = 0;
+			continue;
+		}
+
+		/* Compare right half */
+		for (k=MAX(ms+1,mem); n[k] && n[k] == h[k]; k++);
+		if (n[k]) {
+			h += k-ms;
+			mem = 0;
+			continue;
+		}
+		/* Compare left half */
+		for (k=ms+1; k>mem && n[k-1] == h[k-1]; k--);
+		if (k == mem) return (char *)h;
+		h += p;
+		mem = mem0;
+	}
+}
+
+void *memmem(const void *h0, size_t k, const void *n0, size_t l)
+{
+	const unsigned char *h = h0, *n = n0;
+
+	/* Return immediately on empty needle */
+	if (!l) return (void *)h;
+
+	/* Return immediately when needle is longer than haystack */
+	if (k<l) return 0;
+
+	/* Use faster algorithms for short needles */
+	h = memchr(h0, *n, k);
+	if (!h || l==1) return (void *)h;
+	if (l==2) return twobyte_memmem(h, k, n);
+	if (l==3) return threebyte_memmem(h, k, n);
+	if (l==4) return fourbyte_memmem(h, k, n);
+
+	return twoway_memmem(h, h+k, n, l);
+}