Here is an optimized strlen function for i586+ processors, it works by
testing 4 bytes at a time and averages 1.8 clock cycles per byte (scas =
4n).
I made the patch apply to string-486.h, but it might belong in string.h
too, since string.h is for i686 processors.
Brad Proctor
--------------79EE614C4209BCC66F83B923
Content-Type: text/plain; charset=us-ascii;
name="strlen.diff"
Content-Transfer-Encoding: 7bit
Content-Disposition: inline;
filename="strlen.diff"
diff -urN linux-2.3.24/include/asm-i386/string-486.h linux/include/asm-i486/string-486.h
--- linux-2.3.24/include/asm-i386/string-486.h Wed Oct 27 20:34:41 1999
+++ linux/include/asm-i386/string-486.h Thu Oct 28 15:22:08 1999
@@ -326,21 +326,34 @@
extern inline size_t strlen(const char * s)
{
/*
- * slightly slower on a 486, but with better chances of
- * register allocation
+ * This function works by testing 4 bytes at a time
+ * and averages 1.8 clock cycles per byte.
*/
-register char dummy, *tmp= (char *)s;
-__asm__ __volatile__(
- "\n1:\t"
- "movb\t(%0),%1\n\t"
- "incl\t%0\n\t"
- "testb\t%1,%1\n\t"
- "jne\t1b"
- :"=r" (tmp),"=q" (dummy)
- :"0" (s)
- : "memory" );
-return (tmp-s-1);
+register int __res;
+__asm__ __volatile__ (
+ "addl %0,%2\n\t"
+ "movl (%0),%%ebx\n\t"
+ "addl $4,%0\n\t"
+ ".align 2\n\t"
+ "1:leal -0x1010101(%%ebx),%%ecx\n\t"
+ "xorl $-1,%%ebx\n\t"
+ "andl %%ebx,%%ecx\n\t"
+ "movl (%0),%%ebx\n\t"
+ "addl $4,%0\n\t"
+ "andl $0x80808080,%%ecx\n\t"
+ "jz 1b\n\t"
+ "testl $0x8080,%%ecx\n\t"
+ "jnz 1f\n\t"
+ "shrl $16,%%ecx\n\t"
+ "addl $2,%0\n\t"
+ "1:shlb $1,%%cl\n\t"
+ "sbbl %2,%0\n\t"
+ : "=r" (__res)
+ : "0" (s), "r" (7)
+ : "ebx", "ecx");
+return (__res);
}
+
/* Added by Gertjan van Wingerde to make minix and sysv module work */
#define __HAVE_ARCH_STRNLEN
--------------79EE614C4209BCC66F83B923--
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/