diff --git a/src/rawSHA1_ng_fmt.c b/src/rawSHA1_ng_fmt.c
index 17dd484..93e3c09 100644
--- a/src/rawSHA1_ng_fmt.c
+++ b/src/rawSHA1_ng_fmt.c
@@ -98,6 +98,15 @@
     E   = _mm_add_epi32(E, _mm_roti_epi32(A, 5));                                                       \
 } while (false)
 
+#ifdef __XOP__
+#define R3(W, A, B, C, D, E) do {                                                                       \
+    E   = _mm_add_epi32(E, K);                                                                          \
+    E   = _mm_add_epi32(E, _mm_xor_si128(_mm_cmov_si128(D, B, C), _mm_andnot_si128(D, B))); \
+    E   = _mm_add_epi32(E, W);                                                                          \
+    B   = _mm_roti_epi32(B, 30);                                                                        \
+    E   = _mm_add_epi32(E, _mm_roti_epi32(A, 5));                                                       \
+} while (false)
+#else
 #define R3(W, A, B, C, D, E) do {                                                                       \
     E   = _mm_add_epi32(E, K);                                                                          \
     E   = _mm_add_epi32(E, _mm_or_si128(_mm_and_si128(_mm_or_si128(B, D), C), _mm_and_si128(B, D)));    \
@@ -105,6 +114,7 @@
     B   = _mm_roti_epi32(B, 30);                                                                        \
     E   = _mm_add_epi32(E, _mm_roti_epi32(A, 5));                                                       \
 } while (false)
+#endif
 
 #define _MM_TRANSPOSE4_EPI32(R0, R1, R2, R3) do {                                                       \
     __m128i T0, T1, T2, T3;                                                                             \
diff --git a/src/sse-intrinsics.c b/src/sse-intrinsics.c
index 3880f6d..7aee2eb 100644
--- a/src/sse-intrinsics.c
+++ b/src/sse-intrinsics.c
@@ -620,11 +620,18 @@ void SSEmd4body(__m128i* data, unsigned int * out, int init)
 	SHA1_PARA_DO(i) tmp[i] = _mm_xor_si128((y[i]),(z[i])); \
 	SHA1_PARA_DO(i) tmp[i] = _mm_xor_si128((tmp[i]),(x[i]));
 
+#ifdef __XOP__
+#define SHA1_H(x,y,z) \
+	SHA1_PARA_DO(i) tmp[i] = _mm_cmov_si128((x[i]),(y[i]),(z[i])); \
+	SHA1_PARA_DO(i) tmp2[i] = _mm_andnot_si128((x[i]),(y[i])); \
+	SHA1_PARA_DO(i) tmp[i] = _mm_xor_si128((tmp[i]),(tmp2[i]));
+#else
 #define SHA1_H(x,y,z) \
 	SHA1_PARA_DO(i) tmp[i] = _mm_and_si128((x[i]),(y[i])); \
 	SHA1_PARA_DO(i) tmp2[i] = _mm_or_si128((x[i]),(y[i])); \
 	SHA1_PARA_DO(i) tmp2[i] = _mm_and_si128((tmp2[i]),(z[i])); \
-	SHA1_PARA_DO(i) tmp[i] = _mm_or_si128((tmp[i]),(tmp2[i])); \
+	SHA1_PARA_DO(i) tmp[i] = _mm_or_si128((tmp[i]),(tmp2[i]));
+#endif
 
 #define SHA1_I(x,y,z) SHA1_G(x,y,z)