ft_mem* and ft_strlen optimization

6 files changed, 148 insertions, 50 deletions

diff --git a/src/mem/ft_memccpy.c b/src/mem/ft_memccpy.c
index 8ce656a..0f52242 100644
--- a/src/mem/ft_memccpy.c
+++ b/src/mem/ft_memccpy.c
@@ -12,20 +12,56 @@
 
 #include "libft.h"
 
+#define HIMAGIC 0x8080808080808080L
+#define LOMAGIC 0x0101010101010101L
+
 void	*ft_memccpy(void *dest, const void *src, int c, size_t n)
 {
-	size_t		i;
-	t_ftbyte	*cast_dest;
-	t_ftbyte	*cast_src;
+	uint64_t	buf;
+	uint64_t	lw;
 
-	cast_dest = (t_ftbyte*)dest;
-	cast_src = (t_ftbyte*)src;
-	i = -1;
-	while (++i < n)
+	if (dest == src)
+		return (dest);
+	c = (uint8_t)c;
+	while ((n & 0b111) != 0)
+	{
+		*(uint8_t*)dest = *(uint8_t*)src;
+		if (*(uint8_t*)dest == c)
+			return ((uint8_t*)dest + 1);
+		src++;
+		dest++;
+		n--;
+	}
+	buf = (uint64_t)c | (uint64_t)c << 8 | (uint64_t)c << 16
+		| (uint64_t)c << 24 | (uint64_t)c << 32 | (uint64_t)c << 40
+		| (uint64_t)c << 48 | (uint64_t)c << 56;
+	while (n > 0)
 	{
-		cast_dest[i] = cast_src[i];
-		if (cast_dest[i] == (unsigned char)c)
-			return (cast_dest + i + 1);
+		lw = *(uint64_t*)src ^ buf;
+		if ((lw - LOMAGIC) & ~lw & HIMAGIC)
+		{
+			if ( (((uint8_t*)dest)[0] = ((uint8_t*)src)[0]) == (uint8_t)c)
+				return ((uint8_t*)dest + 1);
+			if ( (((uint8_t*)dest)[1] = ((uint8_t*)src)[1]) == (uint8_t)c)
+				return ((uint8_t*)dest + 2);
+			if ( (((uint8_t*)dest)[2] = ((uint8_t*)src)[2]) == (uint8_t)c)
+				return ((uint8_t*)dest + 3);
+			if ( (((uint8_t*)dest)[3] = ((uint8_t*)src)[3]) == (uint8_t)c)
+				return ((uint8_t*)dest + 4);
+			if ( (((uint8_t*)dest)[4] = ((uint8_t*)src)[4]) == (uint8_t)c)
+				return ((uint8_t*)dest + 5);
+			if ( (((uint8_t*)dest)[5] = ((uint8_t*)src)[5]) == (uint8_t)c)
+				return ((uint8_t*)dest + 6);
+			if ( (((uint8_t*)dest)[6] = ((uint8_t*)src)[6]) == (uint8_t)c)
+				return ((uint8_t*)dest + 7);
+			if ( (((uint8_t*)dest)[7] = ((uint8_t*)src)[7]) == (uint8_t)c)
+				return ((uint8_t*)dest + 8);
+		}
+		else
+			*(uint64_t*)dest = *(uint64_t*)src;
+		n -= 8;
+		dest += 8;
+		src += 8;
 	}
 	return (NULL);
 }
diff --git a/src/mem/ft_memchr.c b/src/mem/ft_memchr.c
index 4fd8689..54780fe 100644
--- a/src/mem/ft_memchr.c
+++ b/src/mem/ft_memchr.c
@@ -12,15 +12,45 @@
 
 #include "libft.h"
 
+/*
+** Determining if a long word contain byte n
+**
+** xor all bytes with n, then check for zero byte like in ft_strlen.
+*/
+
+#define HIMAGIC 0x8080808080808080L
+#define LOMAGIC 0x0101010101010101L
+
 void	*ft_memchr(const void *s, int c, size_t n)
 {
-	size_t		i;
-	t_ftbyte	*cast_s;
+	uint64_t	buf;
+	uint64_t	lw;
 
-	cast_s = (t_ftbyte*)s;
-	i = -1;
-	while (++i < n)
-		if (cast_s[i] == (unsigned char)c)
-			return (cast_s + i);
+	c = (uint8_t)c;
+	while (((uint64_t)s & 0b111) != 0)
+	{
+		n--;
+		if (*(uint8_t*)s == (uint8_t)c)
+			return ((uint8_t*)s);
+		s++;
+	}
+	buf = (uint64_t)c | (uint64_t)c << 8 | (uint64_t)c << 16
+		| (uint64_t)c << 24 | (uint64_t)c << 32 | (uint64_t)c << 40
+		| (uint64_t)c << 48 | (uint64_t)c << 56;
+	while (n >= 8)
+	{
+		lw = *(uint64_t*)s ^ buf;
+		if ((lw - LOMAGIC) & ~lw & HIMAGIC)
+			break;
+		n -= 8;
+		s += 8;
+	}
+	while (n > 0)
+	{
+		if (*(uint8_t*)s == (uint8_t)c)
+			return ((uint8_t*)s);
+		n--;
+		s++;
+	}
 	return (NULL);
 }
diff --git a/src/mem/ft_memcmp.c b/src/mem/ft_memcmp.c
index 233d796..c61ca9a 100644
--- a/src/mem/ft_memcmp.c
+++ b/src/mem/ft_memcmp.c
@@ -11,20 +11,42 @@
 /* ************************************************************************** */
 
 #include "libft.h"
+#include "libft_mem.h"
 
 int	ft_memcmp(const void *s1, const void *s2, size_t n)
 {
-	size_t		i;
-	t_ftbyte	*cast_s1;
-	t_ftbyte	*cast_s2;
-
-	cast_s1 = (t_ftbyte*)s1;
-	cast_s2 = (t_ftbyte*)s2;
-	if (n == 0)
-		return (0);
-	i = -1;
-	while (++i < n)
-		if (cast_s1[i] != cast_s2[i])
-			return (cast_s1[i] - cast_s2[i]);
+	while ((n & 0b111) != 0)
+	{
+		n--;
+		if (*(uint8_t*)s1 != *(uint8_t*)s2)
+			return (*(uint8_t*)s1 - *(uint8_t*)s2);
+		s1++;
+		s2++;
+	}
+	while (n > 0)
+	{
+		if (*(uint64_t*)s1 != *(uint64_t*)s2)
+		{
+			if (((uint8_t*)s1)[0] != ((uint8_t*)s2)[0])
+				return (((uint8_t*)s1)[0] - ((uint8_t*)s2)[0]);
+			if (((uint8_t*)s1)[1] != ((uint8_t*)s2)[1])
+				return (((uint8_t*)s1)[1] - ((uint8_t*)s2)[1]);
+			if (((uint8_t*)s1)[2] != ((uint8_t*)s2)[2])
+				return (((uint8_t*)s1)[2] - ((uint8_t*)s2)[2]);
+			if (((uint8_t*)s1)[3] != ((uint8_t*)s2)[3])
+				return (((uint8_t*)s1)[3] - ((uint8_t*)s2)[3]);
+			if (((uint8_t*)s1)[4] != ((uint8_t*)s2)[4])
+				return (((uint8_t*)s1)[4] - ((uint8_t*)s2)[4]);
+			if (((uint8_t*)s1)[5] != ((uint8_t*)s2)[5])
+				return (((uint8_t*)s1)[5] - ((uint8_t*)s2)[5]);
+			if (((uint8_t*)s1)[6] != ((uint8_t*)s2)[6])
+				return (((uint8_t*)s1)[6] - ((uint8_t*)s2)[6]);
+			if (((uint8_t*)s1)[7] != ((uint8_t*)s2)[7])
+				return (((uint8_t*)s1)[7] - ((uint8_t*)s2)[7]);
+		}
+		n -= 8;
+		s1 += 8;
+		s2 += 8;
+	}
 	return (0);
 }
diff --git a/src/mem/ft_memcpy.c b/src/mem/ft_memcpy.c
index d0ef008..1f84bfd 100644
--- a/src/mem/ft_memcpy.c
+++ b/src/mem/ft_memcpy.c
@@ -11,18 +11,19 @@
 /* ************************************************************************** */
 
 #include "libft.h"
+#include "libft_mem.h"
 
 void	*ft_memcpy(void *dest, const void *src, size_t n)
 {
-	long int		*long_dest;
-	const long int	*long_src;
+	uint64_t		*long_dest;
+	const uint64_t	*long_src;
 
 	if (dest == src)
 		return (dest);
 	while (n % 8 > 0)
 	{
 		n--;
-		((t_ftbyte*)dest)[n] = ((t_ftbyte*)src)[n];
+		((uint8_t*)dest)[n] = ((uint8_t*)src)[n];
 	}
 	long_dest = dest;
 	long_src = src;
diff --git a/src/mem/ft_memmove.c b/src/mem/ft_memmove.c
index 2f794fd..142b761 100644
--- a/src/mem/ft_memmove.c
+++ b/src/mem/ft_memmove.c
@@ -11,11 +11,12 @@
 /* ************************************************************************** */
 
 #include "libft.h"
+#include "libft_mem.h"
 
 void	*ft_memmove(void *dst, const void *src, size_t len)
 {
-	long int		*long_dst;
-	const long int	*long_src;
+	uint64_t		*dst64;
+	const uint64_t	*src64;
 	void			*dst_copy;
 
 	if (dst >= src)
@@ -24,12 +25,12 @@ void	*ft_memmove(void *dst, const void *src, size_t len)
 	while (len % 8 > 0)
 	{
 		len--;
-		*(t_ftbyte*)dst++ = *(t_ftbyte*)src++;
+		*(uint8_t*)dst++ = *(uint8_t*)src++;
 	}
-	long_dst = dst;
-	long_src = src;
+	dst64 = dst;
+	src64 = src;
 	len /= 8;
 	while (len-- > 0)
-		*long_dst++ = *long_src++;
+		*dst64++ = *src64++;
 	return (dst_copy);
 }
diff --git a/src/mem/ft_memset.c b/src/mem/ft_memset.c
index 89f53ff..ce471a5 100644
--- a/src/mem/ft_memset.c
+++ b/src/mem/ft_memset.c
@@ -11,21 +11,29 @@
 /* ************************************************************************** */
 
 #include "libft.h"
+#include "libft_mem.h"
 
 void	*ft_memset(void *s, int c, size_t n)
 {
-	long int	buf;
-	long int	*long_s;
+	uint64_t	buf;
+	void *cpy;
 
-	c = (unsigned char)c;
-	while (n % 8 > 0)
-		*((t_ftbyte*)s + --n) = c;
-	buf = (long int)c | (long int)c << 8 | (long int)c << 16
-			| (long int)c << 24 | (long int)c << 32 | (long int)c << 40
-			| (long int)c << 48 | (long int)c << 56;
-	n /= 8;
-	long_s = s;
+	cpy = s;
+	c = (uint8_t)c;
+	buf = (uint64_t)c | (uint64_t)c << 8 | (uint64_t)c << 16
+		| (uint64_t)c << 24 | (uint64_t)c << 32 | (uint64_t)c << 40
+		| (uint64_t)c << 48 | (uint64_t)c << 56;
+	while (n > 8)
+	{
+		*(uint64_t*)s = buf;
+		n -= 8;
+		s += 8;
+	}
 	while (n > 0)
-		long_s[--n] = buf;
-	return (s);
+	{
+		*(uint8_t*)s = c;
+		s++;
+		n--;
+	}
+	return (cpy);
 }