diff --git a/sysdeps/i386/i686/multiarch/wcscpy-ssse3.S b/sysdeps/i386/i686/multiarch/wcscpy-ssse3.S
index c9498d6..84d92a8 100644
--- a/sysdeps/i386/i686/multiarch/wcscpy-ssse3.S
+++ b/sysdeps/i386/i686/multiarch/wcscpy-ssse3.S
@@ -49,7 +49,7 @@ ENTRY (__wcscpy_ssse3)
 	jz	L(ExitTail8)
 	cmp	$0, 8(%ecx)
 	jz	L(ExitTail12)
-	cmpb	$0, 12(%ecx)
+	cmp	$0, 12(%ecx)
 	jz	L(ExitTail16)
 
 	PUSH	(%edi)
@@ -204,6 +204,7 @@ L(Aligned64Leave):
 	jnz	L(CopyFrom1To16Bytes)
 
 	mov	$-0x40, %esi
+	movaps	%xmm7, -16(%edx)
 	jmp	L(Aligned64Loop)
 
 	.p2align 4
diff --git a/sysdeps/x86_64/multiarch/wcscpy-ssse3.S b/sysdeps/x86_64/multiarch/wcscpy-ssse3.S
index c7b30db..133700f 100644
--- a/sysdeps/x86_64/multiarch/wcscpy-ssse3.S
+++ b/sysdeps/x86_64/multiarch/wcscpy-ssse3.S
@@ -189,6 +189,7 @@ L(Aligned64Leave):
 	jnz	L(CopyFrom1To16Bytes)
 
 	mov	$-0x40, %rsi
+	movaps	%xmm7, -16(%rdx)
 	jmp	L(Aligned64Loop)
 
 	.p2align 4