Small code cleanups in SSE2 Serpent

2 files changed, 195 insertions, 192 deletions

diff --git a/src/block/serpent_sse2/serp_sse2.cpp b/src/block/serpent_sse2/serp_sse2.cpp
index ea937c95a..5ce7d8f47 100644
--- a/src/block/serpent_sse2/serp_sse2.cpp
+++ b/src/block/serpent_sse2/serp_sse2.cpp
@@ -65,6 +65,8 @@ void serpent_encrypt_4(const byte in[64],
                        byte out[64],
                        const u32bit keys_32[132])
    {
+   const __m128i all_ones = _mm_set1_epi8(0xFF);
+
    const __m128i* keys = (const __m128i*)(keys_32);
    __m128i* out_mm = (__m128i*)(out);
    __m128i* in_mm = (__m128i*)(in);
@@ -84,6 +86,7 @@ void serpent_encrypt_4(const byte in[64],
    key_xor( 5,B0,B1,B2,B3); SBoxE6(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor( 6,B0,B1,B2,B3); SBoxE7(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor( 7,B0,B1,B2,B3); SBoxE8(B0,B1,B2,B3); transform(B0,B1,B2,B3);
+
    key_xor( 8,B0,B1,B2,B3); SBoxE1(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor( 9,B0,B1,B2,B3); SBoxE2(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor(10,B0,B1,B2,B3); SBoxE3(B0,B1,B2,B3); transform(B0,B1,B2,B3);
@@ -92,6 +95,7 @@ void serpent_encrypt_4(const byte in[64],
    key_xor(13,B0,B1,B2,B3); SBoxE6(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor(14,B0,B1,B2,B3); SBoxE7(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor(15,B0,B1,B2,B3); SBoxE8(B0,B1,B2,B3); transform(B0,B1,B2,B3);
+
    key_xor(16,B0,B1,B2,B3); SBoxE1(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor(17,B0,B1,B2,B3); SBoxE2(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor(18,B0,B1,B2,B3); SBoxE3(B0,B1,B2,B3); transform(B0,B1,B2,B3);
@@ -100,6 +104,7 @@ void serpent_encrypt_4(const byte in[64],
    key_xor(21,B0,B1,B2,B3); SBoxE6(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor(22,B0,B1,B2,B3); SBoxE7(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor(23,B0,B1,B2,B3); SBoxE8(B0,B1,B2,B3); transform(B0,B1,B2,B3);
+
    key_xor(24,B0,B1,B2,B3); SBoxE1(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor(25,B0,B1,B2,B3); SBoxE2(B0,B1,B2,B3); transform(B0,B1,B2,B3);
    key_xor(26,B0,B1,B2,B3); SBoxE3(B0,B1,B2,B3); transform(B0,B1,B2,B3);
@@ -127,7 +132,6 @@ void Serpent_SSE2::encrypt_n(const byte in[], byte out[], u32bit blocks) const
    while(blocks >= 4)
       {
       serpent_encrypt_4(in, out, this->round_key);
-      //Serpent::encrypt_n(in, out, 4);
       in += 4 * BLOCK_SIZE;
       out += 4 * BLOCK_SIZE;
       blocks -= 4;
diff --git a/src/block/serpent_sse2/serp_sse2_sbox.h b/src/block/serpent_sse2/serp_sse2_sbox.h
index bc8678a89..1660643ad 100644
--- a/src/block/serpent_sse2/serp_sse2_sbox.h
+++ b/src/block/serpent_sse2/serp_sse2_sbox.h
@@ -8,211 +8,210 @@
 #ifndef SERPENT_SSE2_SBOXES_H__
 #define SERPENT_SSE2_SBOXES_H__
 
-#define SBoxE1(B0, B1, B2, B3)                          \
-   do {                                                 \
-      B3 = _mm_xor_si128(B3, B0);                       \
-      __m128i B4 = B1;                                  \
-      B1 = _mm_and_si128(B1, B3);                       \
-      B4 = _mm_xor_si128(B4, B2);                       \
-      B1 = _mm_xor_si128(B1, B0);                       \
-      B0 = _mm_or_si128(B0, B3);                        \
-      B0 = _mm_xor_si128(B0, B4);                       \
-      B4 = _mm_xor_si128(B4, B3);                       \
-      B3 = _mm_xor_si128(B3, B2);                       \
-      B2 = _mm_or_si128(B2, B1);                        \
-      B2 = _mm_xor_si128(B2, B4);                       \
-      B4 = _mm_andnot_si128(B4, _mm_set1_epi8(0xFF));   \
-      B4 = _mm_or_si128(B4, B1);                        \
-      B1 = _mm_xor_si128(B1, B3);                       \
-      B1 = _mm_xor_si128(B1, B4);                       \
-      B3 = _mm_or_si128(B3, B0);                        \
-      B1 = _mm_xor_si128(B1, B3);                       \
-      B4 = _mm_xor_si128(B4, B3);                       \
-      B3 = B0;                                          \
-      B0 = B1;                                          \
-      B1 = B4;                                          \
+#define SBoxE1(B0, B1, B2, B3)                    \
+   do {                                           \
+      B3 = _mm_xor_si128(B3, B0);                 \
+      __m128i B4 = B1;                            \
+      B1 = _mm_and_si128(B1, B3);                 \
+      B4 = _mm_xor_si128(B4, B2);                 \
+      B1 = _mm_xor_si128(B1, B0);                 \
+      B0 = _mm_or_si128(B0, B3);                  \
+      B0 = _mm_xor_si128(B0, B4);                 \
+      B4 = _mm_xor_si128(B4, B3);                 \
+      B3 = _mm_xor_si128(B3, B2);                 \
+      B2 = _mm_or_si128(B2, B1);                  \
+      B2 = _mm_xor_si128(B2, B4);                 \
+      B4 = _mm_xor_si128(B4, all_ones);           \
+      B4 = _mm_or_si128(B4, B1);                  \
+      B1 = _mm_xor_si128(B1, B3);                 \
+      B1 = _mm_xor_si128(B1, B4);                 \
+      B3 = _mm_or_si128(B3, B0);                  \
+      B1 = _mm_xor_si128(B1, B3);                 \
+      B4 = _mm_xor_si128(B4, B3);                 \
+      B3 = B0;                                    \
+      B0 = B1;                                    \
+      B1 = B4;                                    \
    } while(0);
 
-#define SBoxE2(B0, B1, B2, B3)                          \
-   do {                                                 \
-      B0 = _mm_andnot_si128(B0, _mm_set1_epi8(0xFF));   \
-      B2 = _mm_andnot_si128(B2, _mm_set1_epi8(0xFF));   \
-      __m128i B4 = B0;                                  \
-      B0 = _mm_and_si128(B0, B1);                       \
-      B2 = _mm_xor_si128(B2, B0);                       \
-      B0 = _mm_or_si128(B0, B3);                        \
-      B3 = _mm_xor_si128(B3, B2);                       \
-      B1 = _mm_xor_si128(B1, B0);                       \
-      B0 = _mm_xor_si128(B0, B4);                       \
-      B4 = _mm_or_si128(B4, B1);                        \
-      B1 = _mm_xor_si128(B1, B3);                       \
-      B2 = _mm_or_si128(B2, B0);                        \
-      B2 = _mm_and_si128(B2, B4);                       \
-      B0 = _mm_xor_si128(B0, B1);                       \
-      B1 = _mm_and_si128(B1, B2);                       \
-      B1 = _mm_xor_si128(B1, B0);                       \
-      B0 = _mm_and_si128(B0, B2);                       \
-      B4 = _mm_xor_si128(B4, B0);                       \
-      B0 = B2;                                          \
-      B2 = B3;                                          \
-      B3 = B1;                                          \
-      B1 = B4;                                          \
+#define SBoxE2(B0, B1, B2, B3)                    \
+   do {                                           \
+      B0 = _mm_xor_si128(B0, all_ones);           \
+      B2 = _mm_xor_si128(B2, all_ones);           \
+      __m128i B4 = B0;                            \
+      B0 = _mm_and_si128(B0, B1);                 \
+      B2 = _mm_xor_si128(B2, B0);                 \
+      B0 = _mm_or_si128(B0, B3);                  \
+      B3 = _mm_xor_si128(B3, B2);                 \
+      B1 = _mm_xor_si128(B1, B0);                 \
+      B0 = _mm_xor_si128(B0, B4);                 \
+      B4 = _mm_or_si128(B4, B1);                  \
+      B1 = _mm_xor_si128(B1, B3);                 \
+      B2 = _mm_or_si128(B2, B0);                  \
+      B2 = _mm_and_si128(B2, B4);                 \
+      B0 = _mm_xor_si128(B0, B1);                 \
+      B1 = _mm_and_si128(B1, B2);                 \
+      B1 = _mm_xor_si128(B1, B0);                 \
+      B0 = _mm_and_si128(B0, B2);                 \
+      B4 = _mm_xor_si128(B4, B0);                 \
+      B0 = B2;                                    \
+      B2 = B3;                                    \
+      B3 = B1;                                    \
+      B1 = B4;                                    \
    } while(0);
 
-#define SBoxE3(B0, B1, B2, B3)                          \
-   do {                                                 \
-      __m128i B4 = B0;                                  \
-      B0 = _mm_and_si128(B0, B2);                       \
-      B0 = _mm_xor_si128(B0, B3);                       \
-      B2 = _mm_xor_si128(B2, B1);                       \
-      B2 = _mm_xor_si128(B2, B0);                       \
-      B3 = _mm_or_si128(B3, B4);                        \
-      B3 = _mm_xor_si128(B3, B1);                       \
-      B4 = _mm_xor_si128(B4, B2);                       \
-      B1 = B3;                                          \
-      B3 = _mm_or_si128(B3, B4);                        \
-      B3 = _mm_xor_si128(B3, B0);                       \
-      B0 = _mm_and_si128(B0, B1);                       \
-      B4 = _mm_xor_si128(B4, B0);                       \
-      B1 = _mm_xor_si128(B1, B3);                       \
-      B1 = _mm_xor_si128(B1, B4);                       \
-      B4 = _mm_andnot_si128(B4, _mm_set1_epi8(0xFF));   \
-      B0 = B2;                                          \
-      B2 = B1;                                          \
-      B1 = B3;                                          \
-      B3 = B4;                                          \
+#define SBoxE3(B0, B1, B2, B3)                    \
+   do {                                           \
+      __m128i B4 = B0;                            \
+      B0 = _mm_and_si128(B0, B2);                 \
+      B0 = _mm_xor_si128(B0, B3);                 \
+      B2 = _mm_xor_si128(B2, B1);                 \
+      B2 = _mm_xor_si128(B2, B0);                 \
+      B3 = _mm_or_si128(B3, B4);                  \
+      B3 = _mm_xor_si128(B3, B1);                 \
+      B4 = _mm_xor_si128(B4, B2);                 \
+      B1 = B3;                                    \
+      B3 = _mm_or_si128(B3, B4);                  \
+      B3 = _mm_xor_si128(B3, B0);                 \
+      B0 = _mm_and_si128(B0, B1);                 \
+      B4 = _mm_xor_si128(B4, B0);                 \
+      B1 = _mm_xor_si128(B1, B3);                 \
+      B1 = _mm_xor_si128(B1, B4);                 \
+      B4 = _mm_xor_si128(B4, all_ones);           \
+      B0 = B2;                                    \
+      B2 = B1;                                    \
+      B1 = B3;                                    \
+      B3 = B4;                                    \
    } while(0);
 
-#define SBoxE4(B0, B1, B2, B3)                          \
-   do {                                                 \
-      __m128i B4 = B0;                                  \
-      B0 = _mm_or_si128(B0, B3);                        \
-      B3 = _mm_xor_si128(B3, B1);                       \
-      B1 = _mm_and_si128(B1, B4);                       \
-      B4 = _mm_xor_si128(B4, B2);                       \
-      B2 = _mm_xor_si128(B2, B3);                       \
-      B3 = _mm_and_si128(B3, B0);                       \
-      B4 = _mm_or_si128(B4, B1);                        \
-      B3 = _mm_xor_si128(B3, B4);                       \
-      B0 = _mm_xor_si128(B0, B1);                       \
-      B4 = _mm_and_si128(B4, B0);                       \
-      B1 = _mm_xor_si128(B1, B3);                       \
-      B4 = _mm_xor_si128(B4, B2);                       \
-      B1 = _mm_or_si128(B1, B0);                        \
-      B1 = _mm_xor_si128(B1, B2);                       \
-      B0 = _mm_xor_si128(B0, B3);                       \
-      B2 = B1;                                          \
-      B1 = _mm_or_si128(B1, B3);                        \
-      B1 = _mm_xor_si128(B1, B0);                       \
-      B0 = B1;                                          \
-      B1 = B2;                                          \
-      B2 = B3;                                          \
-      B3 = B4;                                          \
+#define SBoxE4(B0, B1, B2, B3)                    \
+   do {                                           \
+      __m128i B4 = B0;                            \
+      B0 = _mm_or_si128(B0, B3);                  \
+      B3 = _mm_xor_si128(B3, B1);                 \
+      B1 = _mm_and_si128(B1, B4);                 \
+      B4 = _mm_xor_si128(B4, B2);                 \
+      B2 = _mm_xor_si128(B2, B3);                 \
+      B3 = _mm_and_si128(B3, B0);                 \
+      B4 = _mm_or_si128(B4, B1);                  \
+      B3 = _mm_xor_si128(B3, B4);                 \
+      B0 = _mm_xor_si128(B0, B1);                 \
+      B4 = _mm_and_si128(B4, B0);                 \
+      B1 = _mm_xor_si128(B1, B3);                 \
+      B4 = _mm_xor_si128(B4, B2);                 \
+      B1 = _mm_or_si128(B1, B0);                  \
+      B1 = _mm_xor_si128(B1, B2);                 \
+      B0 = _mm_xor_si128(B0, B3);                 \
+      B2 = B1;                                    \
+      B1 = _mm_or_si128(B1, B3);                  \
+      B0 = _mm_xor_si128(B0, B1);                 \
+      B1 = B2;                                    \
+      B2 = B3;                                    \
+      B3 = B4;                                    \
    } while(0);
 
-#define SBoxE5(B0, B1, B2, B3)                          \
-   do {                                                 \
-      B1 = _mm_xor_si128(B1, B3);                       \
-      B3 = _mm_andnot_si128(B3, _mm_set1_epi8(0xFF));   \
-      B2 = _mm_xor_si128(B2, B3);                       \
-      B3 = _mm_xor_si128(B3, B0);                       \
-      __m128i B4 = B1;                                  \
-      B1 = _mm_and_si128(B1, B3);                       \
-      B1 = _mm_xor_si128(B1, B2);                       \
-      B4 = _mm_xor_si128(B4, B3);                       \
-      B0 = _mm_xor_si128(B0, B4);                       \
-      B2 = _mm_and_si128(B2, B4);                       \
-      B2 = _mm_xor_si128(B2, B0);                       \
-      B0 = _mm_and_si128(B0, B1);                       \
-      B3 = _mm_xor_si128(B3, B0);                       \
-      B4 = _mm_or_si128(B4, B1);                        \
-      B4 = _mm_xor_si128(B4, B0);                       \
-      B0 = _mm_or_si128(B0, B3);                        \
-      B0 = _mm_xor_si128(B0, B2);                       \
-      B2 = _mm_and_si128(B2, B3);                       \
-      B0 = _mm_andnot_si128(B0, _mm_set1_epi8(0xFF));   \
-      B4 = _mm_xor_si128(B4, B2);                       \
-      B2 = B0;                                          \
-      B0 = B1;                                          \
-      B1 = B4;                                          \
+#define SBoxE5(B0, B1, B2, B3)                    \
+   do {                                           \
+      B1 = _mm_xor_si128(B1, B3);                 \
+      B3 = _mm_xor_si128(B3, all_ones);           \
+      B2 = _mm_xor_si128(B2, B3);                 \
+      B3 = _mm_xor_si128(B3, B0);                 \
+      __m128i B4 = B1;                            \
+      B1 = _mm_and_si128(B1, B3);                 \
+      B1 = _mm_xor_si128(B1, B2);                 \
+      B4 = _mm_xor_si128(B4, B3);                 \
+      B0 = _mm_xor_si128(B0, B4);                 \
+      B2 = _mm_and_si128(B2, B4);                 \
+      B2 = _mm_xor_si128(B2, B0);                 \
+      B0 = _mm_and_si128(B0, B1);                 \
+      B3 = _mm_xor_si128(B3, B0);                 \
+      B4 = _mm_or_si128(B4, B1);                  \
+      B4 = _mm_xor_si128(B4, B0);                 \
+      B0 = _mm_or_si128(B0, B3);                  \
+      B0 = _mm_xor_si128(B0, B2);                 \
+      B2 = _mm_and_si128(B2, B3);                 \
+      B0 = _mm_xor_si128(B0, all_ones);           \
+      B4 = _mm_xor_si128(B4, B2);                 \
+      B2 = B0;                                    \
+      B0 = B1;                                    \
+      B1 = B4;                                    \
    } while(0);
 
-#define SBoxE6(B0, B1, B2, B3)                          \
-   do {                                                 \
-      B0 = _mm_xor_si128(B0, B1);                       \
-      B1 = _mm_xor_si128(B1, B3);                       \
-      B3 = _mm_andnot_si128(B3, _mm_set1_epi8(0xFF));   \
-      __m128i B4 = B1;                                  \
-      B1 = _mm_and_si128(B1, B0);                       \
-      B2 = _mm_xor_si128(B2, B3);                       \
-      B1 = _mm_xor_si128(B1, B2);                       \
-      B2 = _mm_or_si128(B2, B4);                        \
-      B4 = _mm_xor_si128(B4, B3);                       \
-      B3 = _mm_and_si128(B3, B1);                       \
-      B3 = _mm_xor_si128(B3, B0);                       \
-      B4 = _mm_xor_si128(B4, B1);                       \
-      B4 = _mm_xor_si128(B4, B2);                       \
-      B2 = _mm_xor_si128(B2, B0);                       \
-      B0 = _mm_and_si128(B0, B3);                       \
-      B2 = _mm_andnot_si128(B2, _mm_set1_epi8(0xFF));   \
-      B0 = _mm_xor_si128(B0, B4);                       \
-      B4 = _mm_or_si128(B4, B3);                        \
-      B4 = _mm_xor_si128(B4, B2);                       \
-      B2 = B0;                                          \
-      B0 = B1;                                          \
-      B1 = B3;                                          \
-      B3 = B4;                                          \
+#define SBoxE6(B0, B1, B2, B3)                    \
+   do {                                           \
+      B0 = _mm_xor_si128(B0, B1);                 \
+      B1 = _mm_xor_si128(B1, B3);                 \
+      B3 = _mm_xor_si128(B3, all_ones);           \
+      __m128i B4 = B1;                            \
+      B1 = _mm_and_si128(B1, B0);                 \
+      B2 = _mm_xor_si128(B2, B3);                 \
+      B1 = _mm_xor_si128(B1, B2);                 \
+      B2 = _mm_or_si128(B2, B4);                  \
+      B4 = _mm_xor_si128(B4, B3);                 \
+      B3 = _mm_and_si128(B3, B1);                 \
+      B3 = _mm_xor_si128(B3, B0);                 \
+      B4 = _mm_xor_si128(B4, B1);                 \
+      B4 = _mm_xor_si128(B4, B2);                 \
+      B2 = _mm_xor_si128(B2, B0);                 \
+      B0 = _mm_and_si128(B0, B3);                 \
+      B2 = _mm_xor_si128(B2, all_ones);           \
+      B0 = _mm_xor_si128(B0, B4);                 \
+      B4 = _mm_or_si128(B4, B3);                  \
+      B4 = _mm_xor_si128(B4, B2);                 \
+      B2 = B0;                                    \
+      B0 = B1;                                    \
+      B1 = B3;                                    \
+      B3 = B4;                                    \
    } while(0);
 
-#define SBoxE7(B0, B1, B2, B3)                          \
-   do {                                                 \
-      B2 = _mm_andnot_si128(B2, _mm_set1_epi8(0xFF));   \
-      __m128i B4 = B3;                                  \
-      B3 = _mm_and_si128(B3, B0);                       \
-      B0 = _mm_xor_si128(B0, B4);                       \
-      B3 = _mm_xor_si128(B3, B2);                       \
-      B2 = _mm_or_si128(B2, B4);                        \
-      B1 = _mm_xor_si128(B1, B3);                       \
-      B2 = _mm_xor_si128(B2, B0);                       \
-      B0 = _mm_or_si128(B0, B1);                        \
-      B2 = _mm_xor_si128(B2, B1);                       \
-      B4 = _mm_xor_si128(B4, B0);                       \
-      B0 = _mm_or_si128(B0, B3);                        \
-      B0 = _mm_xor_si128(B0, B2);                       \
-      B4 = _mm_xor_si128(B4, B3);                       \
-      B4 = _mm_xor_si128(B4, B0);                       \
-      B3 = _mm_andnot_si128(B3, _mm_set1_epi8(0xFF));   \
-      B2 = _mm_and_si128(B2, B4);                       \
-      B3 = _mm_xor_si128(B3, B2);                       \
-      B2 = B4;                                          \
+#define SBoxE7(B0, B1, B2, B3)                    \
+   do {                                           \
+      B2 = _mm_xor_si128(B2, all_ones);           \
+      __m128i B4 = B3;                            \
+      B3 = _mm_and_si128(B3, B0);                 \
+      B0 = _mm_xor_si128(B0, B4);                 \
+      B3 = _mm_xor_si128(B3, B2);                 \
+      B2 = _mm_or_si128(B2, B4);                  \
+      B1 = _mm_xor_si128(B1, B3);                 \
+      B2 = _mm_xor_si128(B2, B0);                 \
+      B0 = _mm_or_si128(B0, B1);                  \
+      B2 = _mm_xor_si128(B2, B1);                 \
+      B4 = _mm_xor_si128(B4, B0);                 \
+      B0 = _mm_or_si128(B0, B3);                  \
+      B0 = _mm_xor_si128(B0, B2);                 \
+      B4 = _mm_xor_si128(B4, B3);                 \
+      B4 = _mm_xor_si128(B4, B0);                 \
+      B3 = _mm_xor_si128(B3, all_ones);           \
+      B2 = _mm_and_si128(B2, B4);                 \
+      B3 = _mm_xor_si128(B3, B2);                 \
+      B2 = B4;                                    \
    } while(0);
 
-#define SBoxE8(B0, B1, B2, B3)                          \
-   do {                                                 \
-      __m128i B4 = B1;                                  \
-      B1 = _mm_or_si128(B1, B2);                        \
-      B1 = _mm_xor_si128(B1, B3);                       \
-      B4 = _mm_xor_si128(B4, B2);                       \
-      B2 = _mm_xor_si128(B2, B1);                       \
-      B3 = _mm_or_si128(B3, B4);                        \
-      B3 = _mm_and_si128(B3, B0);                       \
-      B4 = _mm_xor_si128(B4, B2);                       \
-      B3 = _mm_xor_si128(B3, B1);                       \
-      B1 = _mm_or_si128(B1, B4);                        \
-      B1 = _mm_xor_si128(B1, B0);                       \
-      B0 = _mm_or_si128(B0, B4);                        \
-      B0 = _mm_xor_si128(B0, B2);                       \
-      B1 = _mm_xor_si128(B1, B4);                       \
-      B2 = _mm_xor_si128(B2, B1);                       \
-      B1 = _mm_and_si128(B1, B0);                       \
-      B1 = _mm_xor_si128(B1, B4);                       \
-      B2 = _mm_andnot_si128(B2, _mm_set1_epi8(0xFF));   \
-      B2 = _mm_or_si128(B2, B0);                        \
-      B4 = _mm_xor_si128(B4, B2);                       \
-      B2 = B1;                                          \
-      B1 = B3;                                          \
-      B3 = B0;                                          \
-      B0 = B4;                                          \
+#define SBoxE8(B0, B1, B2, B3)                    \
+   do {                                           \
+      __m128i B4 = B1;                            \
+      B1 = _mm_or_si128(B1, B2);                  \
+      B1 = _mm_xor_si128(B1, B3);                 \
+      B4 = _mm_xor_si128(B4, B2);                 \
+      B2 = _mm_xor_si128(B2, B1);                 \
+      B3 = _mm_or_si128(B3, B4);                  \
+      B3 = _mm_and_si128(B3, B0);                 \
+      B4 = _mm_xor_si128(B4, B2);                 \
+      B3 = _mm_xor_si128(B3, B1);                 \
+      B1 = _mm_or_si128(B1, B4);                  \
+      B1 = _mm_xor_si128(B1, B0);                 \
+      B0 = _mm_or_si128(B0, B4);                  \
+      B0 = _mm_xor_si128(B0, B2);                 \
+      B1 = _mm_xor_si128(B1, B4);                 \
+      B2 = _mm_xor_si128(B2, B1);                 \
+      B1 = _mm_and_si128(B1, B0);                 \
+      B1 = _mm_xor_si128(B1, B4);                 \
+      B2 = _mm_xor_si128(B2, all_ones);           \
+      B2 = _mm_or_si128(B2, B0);                  \
+      B4 = _mm_xor_si128(B4, B2);                 \
+      B2 = B1;                                    \
+      B1 = B3;                                    \
+      B3 = B0;                                    \
+      B0 = B4;                                    \
    } while(0);
 
 #endif