2009-03-16 Mark Mitchell <mark@codesourcery.com>

* libc/machine/arm/strlen.c (strlen): Fix defect in Thumb-2 mode.

2009-03-16  Richard Earnshaw  <rearnsha@arm.com>

        * libc/machine/arm/strlen.c (strlen): Correctly detect
        end-of-string.
        * libc/machine/arm/strcpy.c (strcpy): Likewise.
        * libc/machine/arm/strcmp.c (strcmp, strcmp_unaligned): Likewise.

newlib/ChangeLog

@@ -1,3 +1,14 @@
+2009-03-16  Mark Mitchell  <mark@codesourcery.com>
+
+	* libc/machine/arm/strlen.c (strlen): Fix defect in Thumb-2 mode.
+
+2009-03-16  Richard Earnshaw  <rearnsha@arm.com>
+
+	* libc/machine/arm/strlen.c (strlen): Correctly detect
+	end-of-string.
+	* libc/machine/arm/strcpy.c (strcpy): Likewise.
+	* libc/machine/arm/strcmp.c (strcmp, strcmp_unaligned): Likewise.
+
 2009-03-15  Yaakov Selkowitz <yselkowitz@users.sourceforge.net>
 
 	* libc/include/sys/errno.h (ESTRPIPE): Define.

newlib/libc/machine/arm/strcmp.c

@@ -97,7 +97,7 @@ __attribute__((naked)) strcmp (const char* s1, const char* s2)
       "cmp	ip, r3\n\t"
       "itttt	eq\n\t"
       /* check for any zero bytes in first word */
-      "eoreq	r2, r2, ip\n\t"
+      "biceq	r2, r2, ip\n\t"
       "tsteq	r2, "magic2(r4)"\n\t"
       "ldreq	ip, [r0], #4\n\t"
       "ldreq	r3, [r1], #4\n\t"
@@ -171,9 +171,9 @@ strcmp_unaligned(const char* s1, const char* s2)
 	  w2 RSHIFT= shift;						\
 	  break;							\
 	}								\
-      if (__builtin_expect(((w1 - b1) ^ w1) & (b1 << 7), 0))		\
+      if (__builtin_expect(((w1 - b1) & ~w1) & (b1 << 7), 0))		\
 	{								\
-	  if ((((w1 - b1) ^ w1) & (b1 << 7)) & mask)			\
+	  if ((((w1 - b1) & ~w1) & (b1 << 7)) & mask)			\
 	    w2 RSHIFT= shift;						\
 	  else								\
 	    {								\
@@ -284,7 +284,7 @@ strcmp_unaligned(const char* s1, const char* s2)
       "bic	t1, w1, #"MSB"\n\t"
       "cmp	t1, w2, "SHFT2LSB" #8\n\t"
       "sub	r3, w1, b1\n\t"
-      "eor	r3, r3, w1\n\t"
+      "bic	r3, r3, w1\n\t"
       "bne	4f\n\t"
       "ands	r3, r3, b1, lsl #7\n\t"
       "it	eq\n\t"
@@ -320,7 +320,7 @@ strcmp_unaligned(const char* s1, const char* s2)
       SHFT2MSB"	t1, w1, #16\n\t"
       "sub	r3, w1, b1\n\t"
       SHFT2LSB"	t1, t1, #16\n\t"
-      "eor	r3, r3, w1\n\t"
+      "bic	r3, r3, w1\n\t"
       "cmp	t1, w2, "SHFT2LSB" #16\n\t"
       "bne	4f\n\t"
       "ands	r3, r3, b1, lsl #7\n\t"
@@ -356,7 +356,7 @@ strcmp_unaligned(const char* s1, const char* s2)
       "and	t1, w1, #"LSB"\n\t"
       "cmp	t1, w2, "SHFT2LSB" #24\n\t"
       "sub	r3, w1, b1\n\t"
-      "eor	r3, r3, w1\n\t"
+      "bic	r3, r3, w1\n\t"
       "bne	4f\n\t"
       "ands	r3, r3, b1, lsl #7\n\t"
       "it	eq\n\t"

newlib/libc/machine/arm/strcpy.c

@@ -64,7 +64,7 @@ strcpy (char* dst, const char* src)
        "ldr	r3, [r1], #4\n\t"
        "beq	2f\n\t"
        "sub	r2, r3, "magic1(r5)"\n\t"
-       "eors	r2, r2, r3\n\t"
+       "bics	r2, r2, r3\n\t"
        "tst	r2, "magic2(r5)"\n\t"
        "itt	eq\n\t"
        "streq	r3, [ip], #4\n\t"
@@ -78,12 +78,12 @@ strcpy (char* dst, const char* src)
        "optpld	r1, #8\n\t"
        "ldr	r4, [r1], #4\n\t"
        "sub	r2, r3, "magic1(r5)"\n\t"
-       "eors	r2, r2, r3\n\t"
+       "bics	r2, r2, r3\n\t"
        "tst	r2, "magic2(r5)"\n\t"
        "sub	r2, r4, "magic1(r5)"\n\t"
        "bne	1f\n\t"
        "str	r3, [ip], #4\n\t"
-       "eors	r2, r2, r4\n\t"
+       "bics	r2, r2, r4\n\t"
        "tst	r2, "magic2(r5)"\n\t"
        "itt	eq\n\t"
        "ldreq	r3, [r1], #4\n\t"

newlib/libc/machine/arm/strlen.c

@@ -79,16 +79,17 @@ strlen (const char* str)
        "add	ip, len, #4\n\t"
        "mov	ip, ip, asl #3\n\t"
        "mvn	r2, #0\n\t"
-       "it	ne\n\t"
        /* ... are masked out */
 #ifdef __thumb__
+       "itt	ne\n\t"
 # ifdef __ARMEB__
        "lslne	r2, ip\n\t"
 # else
        "lsrne	r2, ip\n\t"
 # endif
-       "orr	data, data, r2\n\t"
+       "orrne	data, data, r2\n\t"
 #else
+       "it	ne\n\t"
 # ifdef __ARMEB__
        "orrne	data, data, r2, lsl ip\n\t"
 # else
@@ -104,13 +105,15 @@ strlen (const char* str)
 #endif
        "orr	ip, ip, ip, lsl #16\n"
 
-	/* This is the main loop.  We subtract one from each byte in the
+	/* This is the main loop.  We subtract one from each byte in
-	   word: the sign bit changes iff the byte was zero.  */
+	   the word: the sign bit changes iff the byte was zero or
+	   0x80 -- we eliminate the latter case by anding the result
+	   with the 1-s complement of the data.  */
        "1:\n\t"
        /* test (data - 0x01010101)  */
        "sub	r2, data, ip\n\t"
-       /* ... ^ data */
+       /* ... & ~data */
-       "eor	r2, r2, data\n\t"
+       "bic	r2, r2, data\n\t"
        /* ... & 0x80808080 == 0? */
        "ands	r2, r2, ip, lsl #7\n\t"
 #ifdef _ISA_ARM_7
@@ -124,8 +127,8 @@ strlen (const char* str)
        /*  test (data - 0x01010101)  */
        "ittt	eq\n\t"
        "subeq	r2, data, ip\n\t"
-       /* ... ^ data */
+       /* ... & ~data */
-       "eoreq	r2, r2, data\n\t"
+       "biceq	r2, r2, data\n\t"
        /* ... & 0x80808080 == 0? */
        "andeqs	r2, r2, ip, lsl #7\n\t"
 #endif