diff --git a/string/test-strncat.c b/string/test-strncat.c
index f7cf7df..2ffa550 100644
--- a/string/test-strncat.c
+++ b/string/test-strncat.c
@@ -94,16 +94,12 @@ do_test (size_t align1, size_t align2, size_t len1, size_t len2,
   size_t i;
   char *s1, *s2;
 
-  align1 &= 7;
-  if (align1 + len1 >= page_size)
+  if (align1 + (len1 > n ? n : len1) >= page_size)
     return;
-  if (align1 + n > page_size)
-    return;
-  align2 &= 7;
-  if (align2 + len1 + len2 >= page_size)
-    return;
-  if (align2 + len1 + n > page_size)
+
+  if (align2 + (len1 > n ? n : len1) + len2 >= page_size)
     return;
+
   s1 = (char *) (buf1 + align1);
   s2 = (char *) (buf2 + align2);
 
@@ -269,6 +265,10 @@ main (void)
     printf ("\t%s", impl->name);
   putchar ('\n');
 
+  do_test (50, 2, 20, 20, -1, 127);
+  do_test (50, 59, 40, 40, -1, 255);
+  do_test (53, 53, 80, 80, -1, 127);
+
   for (n = 2; n <= 2048; n*=4)
     {
       do_test (0, 2, 2, 2, n, 127);
diff --git a/sysdeps/i386/i686/multiarch/strcat-sse2.S b/sysdeps/i386/i686/multiarch/strcat-sse2.S
index 43a6945..b692036 100644
--- a/sysdeps/i386/i686/multiarch/strcat-sse2.S
+++ b/sysdeps/i386/i686/multiarch/strcat-sse2.S
@@ -307,6 +307,10 @@ L(StartStrcpyPart_1):
 	pxor	%xmm0, %xmm0
 	pxor	%xmm1, %xmm1
 
+# ifdef USE_AS_STRNCAT
+	cmp	$48, %ebx
+	ja      L(BigN)
+# endif
 	pcmpeqb	(%esi), %xmm1
 # ifdef USE_AS_STRNCAT
 	add	%ecx, %ebx
@@ -348,7 +352,7 @@ L(Unalign16Both):
 # endif
 	test	%edx, %edx
 	jnz	L(CopyFrom1To16Bytes)
-
+L(Unalign16BothBigN):
 	movaps	16(%esi, %ecx), %xmm3
 	movdqu	%xmm2, (%eax, %ecx)
 	pcmpeqb	%xmm3, %xmm0
@@ -418,7 +422,6 @@ L(Unalign16Both):
 # ifdef USE_AS_STRNCAT
 	lea	128(%ebx, %edx), %ebx
 # endif
-L(Unaligned64Loop):
 	movaps	(%esi), %xmm2
 	movaps	%xmm2, %xmm4
 	movaps	16(%esi), %xmm5
@@ -437,6 +440,7 @@ L(Unaligned64Loop):
 	test	%edx, %edx
 	jnz	L(Unaligned64Leave)
 
+	.p2align 4
 L(Unaligned64Loop_start):
 	add	$64, %eax
 	add	$64, %esi
@@ -489,6 +493,38 @@ L(Unaligned64Leave):
 	add	$48, %eax
 	BRANCH_TO_JMPTBL_ENTRY (L(ExitTable), %edx, 4)
 
+# ifdef USE_AS_STRNCAT
+	.p2align 4
+L(BigN):
+	pcmpeqb	(%esi), %xmm1
+	pmovmskb %xmm1, %edx
+	shr	%cl, %edx
+	test	%edx, %edx
+	jnz	L(CopyFrom1To16BytesTail)
+
+	pcmpeqb	16(%esi), %xmm0
+	pmovmskb %xmm0, %edx
+	test	%edx, %edx
+	jnz	L(CopyFrom1To32Bytes)
+
+	movdqu	(%esi, %ecx), %xmm1   /* copy 16 bytes */
+	movdqu	%xmm1, (%eax)
+	sub	%ecx, %eax
+	sub     $48, %ebx
+	add     %ecx, %ebx
+
+	mov	$16, %ecx
+	movdqa	(%esi, %ecx), %xmm1
+	movaps	16(%esi, %ecx), %xmm2
+	movdqu	%xmm1, (%eax, %ecx)
+	pcmpeqb	%xmm2, %xmm0
+	pmovmskb %xmm0, %edx
+	add	$16, %ecx
+	test	%edx, %edx
+	jnz	L(CopyFrom1To16Bytes)
+	jmp	L(Unalign16BothBigN)
+# endif
+
 /*------------end of main part-------------------------------*/
 
 /* Case1 */
diff --git a/sysdeps/i386/i686/multiarch/strcpy-ssse3.S b/sysdeps/i386/i686/multiarch/strcpy-ssse3.S
index 813dcc1..073856f 100644
--- a/sysdeps/i386/i686/multiarch/strcpy-ssse3.S
+++ b/sysdeps/i386/i686/multiarch/strcpy-ssse3.S
@@ -124,6 +124,7 @@ ENTRY (STRCPY)
 	PUSH	(%esi)
 # ifdef USE_AS_STRNCPY
 	mov	%ecx, %esi
+	sub	$16, %ebx
 	and	$0xf, %esi
 
 /* add 16 bytes ecx_shift to ebx */
@@ -160,7 +161,7 @@ ENTRY (STRCPY)
 /* eax = 0: there isn't end of string from position esi to esi+15 */
 
 # ifdef USE_AS_STRNCPY
-	sub	$32, %ebx
+	sub	$16, %ebx
 	jbe	L(CopyFrom1To16BytesCase2OrCase3)
 # endif
 	test	%eax, %eax