diff --git a/newlib/ChangeLog b/newlib/ChangeLog
index 79722459b..75c922326 100644
--- a/newlib/ChangeLog
+++ b/newlib/ChangeLog
@@ -1,5 +1,8 @@
 2002-01-17  Nick Clifton  <nickc@cambridge.redhat.com>
 
+	* libc/machine/xscale/memset.c (memset): Fix bug when len == 1 and
+	dst was not word aligned.
+
 	* libc/sys/arm/syscalls.c (_sbrk): Return -1 rather than aborting
 	if too much memory is requested.
 
diff --git a/newlib/libc/machine/xscale/memset.c b/newlib/libc/machine/xscale/memset.c
index bfd0f0d47..ad1fc74bc 100644
--- a/newlib/libc/machine/xscale/memset.c
+++ b/newlib/libc/machine/xscale/memset.c
@@ -11,6 +11,7 @@ void *
 memset (void *dst, int c, size_t len)
 {
   int dummy;
+
   asm volatile ("tst	%0, #0x3"
 #ifndef __OPTIMIZE_SIZE__
 "
@@ -24,6 +25,13 @@ memset (void *dst, int c, size_t len)
 	movs	r3, %2
 	sub	%2, %2, #1
 	bne	0b
+# At this point we know that %2 == len == -1 (since the SUB has already taken
+# place).  If we fall through to the 1: label (as the code used to do), the
+# CMP will detect this negative value and branch to the 2: label.  This will
+# test %2 again, but this time against 0.  The test will fail and the loop
+# at 2: will go on for (almost) ever.  Hence the explicit branch to the end
+# of the hand written assembly code.
+	b       4f
 1:
 	cmp	%2, #0x3
 	bls	2f
@@ -63,17 +71,18 @@ memset (void *dst, int c, size_t len)
 2:
 	movs	r3, %2
 	sub	%2, %2, #1
-	beq	1f
+	beq	4f
 0:
 	movs	r3, %2
 	sub	%2, %2, #1
 	strb	%1, [%0], #1
 	bne	0b
-1:"
+4:"
 
        : "=&r" (dummy), "=&r" (c), "=&r" (len)
        : "0" (dst), "1" (c), "2" (len)
        : "memory", "r3", "r4", "r5", "lr");
+
   return dst;
 }