Index: sys/crypto/aesni/aesni_ghash.c
===================================================================
--- sys/crypto/aesni/aesni_ghash.c
+++ sys/crypto/aesni/aesni_ghash.c
@@ -534,7 +534,7 @@
 	const unsigned char *tag, uint32_t nbytes, uint32_t abytes, int ibytes,
 	const unsigned char *key, int nr)
 {
-	int i, j ,k;
+	int i, j, k, resid;
 	__m128i tmp1, tmp2, tmp3, tmp4;
 	__m128i tmp5, tmp6, tmp7, tmp8;
 	__m128i H, H2, H3, H4, Y, T;
@@ -790,7 +790,8 @@
 		_mm_storeu_si128(&((__m128i*)out)[k], tmp1);
 	}
 	//If remains one incomplete block
-	if (nbytes%16) {
+	resid = nbytes % 16;
+	if (resid != 0) {
 		tmp1 = _mm_shuffle_epi8(ctr1, BSWAP_EPI64);
 		tmp1 = _mm_xor_si128(tmp1, KEY[0]);
 		for (j=1; j<nr-1; j+=2) {
@@ -799,10 +800,11 @@
 		}
 		tmp1 = _mm_aesenc_si128(tmp1, KEY[nr-1]);
 		tmp1 = _mm_aesenclast_si128(tmp1, KEY[nr]);
-		tmp1 = _mm_xor_si128(tmp1,
-		    _mm_loadu_si128(&((const __m128i *)in)[k]));
+		last_block = _mm_setzero_si128();
+		memcpy(&last_block, &((const __m128i *)in)[k], resid);
+		tmp1 = _mm_xor_si128(tmp1, last_block);
 		last_block = tmp1;
-		for (j=0; j<nbytes%16; j++)
+		for (j=0; j < resid != 0; j++)
 			out[k*16+j] = ((unsigned char*)&last_block)[j];
 	}
 	return 1; //when sucessfull returns 1