Makefile |  4 ++--
 nnn.c    | 19 +++++++++----------

diff --git a/Makefile b/Makefile
index 03dfd77ce0dd75bc060d971936e2cdc2a2991c93..e8224d8006784430d0bab8995526f689e5d695c5 100644
--- a/Makefile
+++ b/Makefile
@@ -3,7 +3,7 @@
 PREFIX ?= /usr/local
 MANPREFIX = $(PREFIX)/share/man
 
-CFLAGS += -Wall -Wextra -Wno-unused-parameter
+CFLAGS += -O3 -Wall -Wextra -Wno-unused-parameter
 LDLIBS = -lreadline
 
 ifeq ($(shell pkg-config ncursesw && echo 1),1)
@@ -23,7 +23,7 @@
 $(SRC): nnn.h
 
 $(BIN): $(SRC)
-	$(CC) -O2 $(CFLAGS) -o $@ $^ $(LDFLAGS) $(LDLIBS)
+	$(CC) $(CFLAGS) -o $@ $^ $(LDFLAGS) $(LDLIBS)
 	strip $@
 
 debug: $(SRC)
diff --git a/nnn.c b/nnn.c
index fb1d5b4ea152260be9fcbdc41501c0ff36165206..721ef80bac4afe1231379ecae329181f57db285d 100644
--- a/nnn.c
+++ b/nnn.c
@@ -357,11 +357,12 @@  */
 static size_t
 xstrlcpy(char *dest, const char *src, size_t n)
 {
+	static ulong *s, *d;
 	static size_t len, blocks;
 	static const uint lsize = sizeof(ulong);
 	static const uint _WSHIFT = (sizeof(ulong) == 8) ? 3 : 2;
 
-	if (!src || !dest)
+	if (!src || !dest || !n)
 		return 0;
 
 	len = xstrlen(src) + 1;
@@ -371,17 +372,15 @@ 	else if (len > n)
 		/* Save total number of bytes to copy in len */
 		len = n;
 
-	if (n >= lsize) {
+	/*
+	 * To enable -O3 ensure src and dest are 16-byte aligned
+	 * More info: http://www.felixcloutier.com/x86/MOVDQA.html
+	 */
+	if ((n >= lsize) && !((ulong)src & (ulong)dest & 0xF)) {
+		s = (ulong *)src;
+		d = (ulong *)dest;
 		blocks = n >> _WSHIFT;
 		n -= (blocks << _WSHIFT);
-	} else
-		blocks = 0;
-
-	if (blocks) {
-		static ulong *s, *d;
-
-		s = (ulong *)src;
-		d = (ulong *)dest;
 
 		while (blocks) {
 			*d = *s;