src/aes*.asm -> src/asm/, src/aes*.c -> src/c/

3 files changed, 395 insertions, 0 deletions

diff --git a/src/c/aes128.c b/src/c/aes128.c
new file mode 100644
index 0000000..d4c609d
--- /dev/null
+++ b/src/c/aes128.c
@@ -0,0 +1,99 @@
+/**
+ * \file
+ * \author Egor Tensin <Egor.Tensin@gmail.com>
+ * \date 2015
+ * \copyright This file is licensed under the terms of the MIT License.
+ *            See LICENSE.txt for details.
+ */
+
+#include <aesni/all.h>
+
+#include <emmintrin.h>
+#include <wmmintrin.h>
+
+AesBlock128 __fastcall raw_aes128_encrypt_block(
+    AesBlock128 plain,
+    Aes128KeySchedule* key_schedule)
+{
+    plain = _mm_xor_si128(plain, key_schedule->keys[0]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[1]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[2]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[3]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[4]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[5]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[6]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[7]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[8]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[9]);
+    return _mm_aesenclast_si128(plain, key_schedule->keys[10]);
+}
+
+AesBlock128 __fastcall raw_aes128_decrypt_block(
+    AesBlock128 cipher,
+    Aes128KeySchedule* inverted_schedule)
+{
+    cipher = _mm_xor_si128(cipher, inverted_schedule->keys[0]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[1]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[2]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[3]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[4]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[5]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[6]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[7]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[8]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[9]);
+    return _mm_aesdeclast_si128(cipher, inverted_schedule->keys[10]);
+}
+
+static AesBlock128 __fastcall aes128_keygen_assist(
+    AesBlock128 prev,
+    AesBlock128 hwgen)
+{
+    AesBlock128 tmp = prev;
+
+    tmp = _mm_slli_si128(tmp, 4);
+    prev = _mm_xor_si128(prev, tmp);
+    tmp = _mm_slli_si128(tmp, 4);
+    prev = _mm_xor_si128(prev, tmp);
+    tmp = _mm_slli_si128(tmp, 4);
+    prev = _mm_xor_si128(prev, tmp);
+
+    hwgen = _mm_shuffle_epi32(hwgen, 0xff);
+    prev = _mm_xor_si128(prev, hwgen);
+
+    return prev;
+}
+
+void __fastcall raw_aes128_expand_key_schedule(
+    AesBlock128 key,
+    Aes128KeySchedule* key_schedule)
+{
+    AesBlock128 prev = key_schedule->keys[0] = key;
+    prev = key_schedule->keys[1] = aes128_keygen_assist(prev, _mm_aeskeygenassist_si128(prev, 0x01));
+    prev = key_schedule->keys[2] = aes128_keygen_assist(prev, _mm_aeskeygenassist_si128(prev, 0x02));
+    prev = key_schedule->keys[3] = aes128_keygen_assist(prev, _mm_aeskeygenassist_si128(prev, 0x04));
+    prev = key_schedule->keys[4] = aes128_keygen_assist(prev, _mm_aeskeygenassist_si128(prev, 0x08));
+    prev = key_schedule->keys[5] = aes128_keygen_assist(prev, _mm_aeskeygenassist_si128(prev, 0x10));
+    prev = key_schedule->keys[6] = aes128_keygen_assist(prev, _mm_aeskeygenassist_si128(prev, 0x20));
+    prev = key_schedule->keys[7] = aes128_keygen_assist(prev, _mm_aeskeygenassist_si128(prev, 0x40));
+    prev = key_schedule->keys[8] = aes128_keygen_assist(prev, _mm_aeskeygenassist_si128(prev, 0x80));
+    prev = key_schedule->keys[9] = aes128_keygen_assist(prev, _mm_aeskeygenassist_si128(prev, 0x1b));
+    prev = key_schedule->keys[10] = aes128_keygen_assist(prev, _mm_aeskeygenassist_si128(prev, 0x36));
+}
+
+void __fastcall raw_aes128_invert_key_schedule(
+    Aes128KeySchedule* key_schedule,
+    Aes128KeySchedule* inverted_schedule)
+{
+    inverted_schedule->keys[0] = key_schedule->keys[10];
+    inverted_schedule->keys[1] = _mm_aesimc_si128(key_schedule->keys[9]);
+    inverted_schedule->keys[2] = _mm_aesimc_si128(key_schedule->keys[8]);
+    inverted_schedule->keys[3] = _mm_aesimc_si128(key_schedule->keys[7]);
+    inverted_schedule->keys[4] = _mm_aesimc_si128(key_schedule->keys[6]);
+    inverted_schedule->keys[5] = _mm_aesimc_si128(key_schedule->keys[5]);
+    inverted_schedule->keys[6] = _mm_aesimc_si128(key_schedule->keys[4]);
+    inverted_schedule->keys[7] = _mm_aesimc_si128(key_schedule->keys[3]);
+    inverted_schedule->keys[8] = _mm_aesimc_si128(key_schedule->keys[2]);
+    inverted_schedule->keys[9] = _mm_aesimc_si128(key_schedule->keys[1]);
+    inverted_schedule->keys[10] = key_schedule->keys[0];
+}
diff --git a/src/c/aes192.c b/src/c/aes192.c
new file mode 100644
index 0000000..fec8f06
--- /dev/null
+++ b/src/c/aes192.c
@@ -0,0 +1,134 @@
+/**
+ * \file
+ * \author Egor Tensin <Egor.Tensin@gmail.com>
+ * \date 2015
+ * \copyright This file is licensed under the terms of the MIT License.
+ *            See LICENSE.txt for details.
+ */
+
+#include <aesni/all.h>
+
+#include <emmintrin.h>
+#include <wmmintrin.h>
+
+AesBlock128 __fastcall raw_aes192_encrypt_block(
+    AesBlock128 plain,
+    Aes192KeySchedule* key_schedule)
+{
+    plain = _mm_xor_si128(plain, key_schedule->keys[0]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[1]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[2]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[3]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[4]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[5]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[6]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[7]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[8]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[9]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[10]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[11]);
+    return _mm_aesenclast_si128(plain, key_schedule->keys[12]);
+}
+
+AesBlock128 __fastcall raw_aes192_decrypt_block(
+    AesBlock128 cipher,
+    Aes192KeySchedule* inverted_schedule)
+{
+    cipher = _mm_xor_si128(cipher, inverted_schedule->keys[0]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[1]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[2]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[3]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[4]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[5]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[6]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[7]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[8]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[9]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[10]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[11]);
+    return _mm_aesdeclast_si128(cipher, inverted_schedule->keys[12]);
+}
+
+static void __fastcall aes192_keygen_assist(
+    AesBlock128* prev_lo,
+    AesBlock128* prev_hi,
+    AesBlock128 hwgen)
+{
+    AesBlock128 tmp = *prev_lo;
+
+    tmp = _mm_slli_si128(tmp, 4);
+    *prev_lo = _mm_xor_si128(*prev_lo, tmp);
+    tmp = _mm_slli_si128(tmp, 4);
+    *prev_lo = _mm_xor_si128(*prev_lo, tmp);
+    tmp = _mm_slli_si128(tmp, 4);
+    *prev_lo = _mm_xor_si128(*prev_lo, tmp);
+
+    hwgen = _mm_shuffle_epi32(hwgen, 0x55);
+    *prev_lo = _mm_xor_si128(*prev_lo, hwgen);
+
+    tmp = _mm_shuffle_epi32(*prev_hi, 0xf3);
+    *prev_hi = _mm_xor_si128(*prev_hi, tmp);
+
+    tmp = _mm_shuffle_epi32(*prev_lo, 0xff);
+    tmp = _mm_srli_si128(tmp, 8);
+    *prev_hi = _mm_xor_si128(*prev_hi, tmp);
+}
+
+void __fastcall raw_aes192_expand_key_schedule(
+    AesBlock128 key_lo,
+    AesBlock128 key_hi,
+    Aes192KeySchedule* key_schedule)
+{
+    key_schedule->keys[0] = key_lo;
+    key_schedule->keys[1] = key_hi;
+
+    aes192_keygen_assist(&key_lo, &key_hi, _mm_aeskeygenassist_si128(key_hi, 0x01));
+    key_schedule->keys[1] = _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(key_schedule->keys[1]), _mm_castsi128_pd(key_lo), 0));
+    key_schedule->keys[2] = _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(key_lo), _mm_castsi128_pd(key_hi), 1));
+
+    aes192_keygen_assist(&key_lo, &key_hi, _mm_aeskeygenassist_si128(key_hi, 0x02));
+    key_schedule->keys[3] = key_lo;
+    key_schedule->keys[4] = key_hi;
+
+    aes192_keygen_assist(&key_lo, &key_hi, _mm_aeskeygenassist_si128(key_hi, 0x04));
+    key_schedule->keys[4] = _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(key_schedule->keys[4]), _mm_castsi128_pd(key_lo), 0));
+    key_schedule->keys[5] = _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(key_lo), _mm_castsi128_pd(key_hi), 1));
+
+    aes192_keygen_assist(&key_lo, &key_hi, _mm_aeskeygenassist_si128(key_hi, 0x08));
+    key_schedule->keys[6] = key_lo;
+    key_schedule->keys[7] = key_hi;
+
+    aes192_keygen_assist(&key_lo, &key_hi, _mm_aeskeygenassist_si128(key_hi, 0x10));
+    key_schedule->keys[7] = _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(key_schedule->keys[7]), _mm_castsi128_pd(key_lo), 0));
+    key_schedule->keys[8] = _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(key_lo), _mm_castsi128_pd(key_hi), 1));
+
+    aes192_keygen_assist(&key_lo, &key_hi, _mm_aeskeygenassist_si128(key_hi, 0x20));
+    key_schedule->keys[9] = key_lo;
+    key_schedule->keys[10] = key_hi;
+
+    aes192_keygen_assist(&key_lo, &key_hi, _mm_aeskeygenassist_si128(key_hi, 0x40));
+    key_schedule->keys[10] = _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(key_schedule->keys[10]), _mm_castsi128_pd(key_lo), 0));
+    key_schedule->keys[11] = _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(key_lo), _mm_castsi128_pd(key_hi), 1));
+
+    aes192_keygen_assist(&key_lo, &key_hi, _mm_aeskeygenassist_si128(key_hi, 0x80));
+    key_schedule->keys[12] = key_lo;
+}
+
+void __fastcall raw_aes192_invert_key_schedule(
+    Aes192KeySchedule* key_schedule,
+    Aes192KeySchedule* inverted_schedule)
+{
+    inverted_schedule->keys[0] = key_schedule->keys[12];
+    inverted_schedule->keys[1] = _mm_aesimc_si128(key_schedule->keys[11]);
+    inverted_schedule->keys[2] = _mm_aesimc_si128(key_schedule->keys[10]);
+    inverted_schedule->keys[3] = _mm_aesimc_si128(key_schedule->keys[9]);
+    inverted_schedule->keys[4] = _mm_aesimc_si128(key_schedule->keys[8]);
+    inverted_schedule->keys[5] = _mm_aesimc_si128(key_schedule->keys[7]);
+    inverted_schedule->keys[6] = _mm_aesimc_si128(key_schedule->keys[6]);
+    inverted_schedule->keys[7] = _mm_aesimc_si128(key_schedule->keys[5]);
+    inverted_schedule->keys[8] = _mm_aesimc_si128(key_schedule->keys[4]);
+    inverted_schedule->keys[9] = _mm_aesimc_si128(key_schedule->keys[3]);
+    inverted_schedule->keys[10] = _mm_aesimc_si128(key_schedule->keys[2]);
+    inverted_schedule->keys[11] = _mm_aesimc_si128(key_schedule->keys[1]);
+    inverted_schedule->keys[12] = key_schedule->keys[0];
+}
diff --git a/src/c/aes256.c b/src/c/aes256.c
new file mode 100644
index 0000000..be4f783
--- /dev/null
+++ b/src/c/aes256.c
@@ -0,0 +1,162 @@
+/**
+ * \file
+ * \author Egor Tensin <Egor.Tensin@gmail.com>
+ * \date 2015
+ * \copyright This file is licensed under the terms of the MIT License.
+ *            See LICENSE.txt for details.
+ */
+
+#include <aesni/all.h>
+
+#include <emmintrin.h>
+#include <wmmintrin.h>
+
+AesBlock128 __fastcall raw_aes256_encrypt_block(
+    AesBlock128 plain,
+    Aes256KeySchedule* key_schedule)
+{
+    plain = _mm_xor_si128(plain, key_schedule->keys[0]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[1]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[2]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[3]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[4]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[5]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[6]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[7]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[8]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[9]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[10]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[11]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[12]);
+    plain = _mm_aesenc_si128(plain, key_schedule->keys[13]);
+    return _mm_aesenclast_si128(plain, key_schedule->keys[14]);
+}
+
+AesBlock128 __fastcall raw_aes256_decrypt_block(
+    AesBlock128 cipher,
+    Aes256KeySchedule* inverted_schedule)
+{
+    cipher = _mm_xor_si128(cipher, inverted_schedule->keys[0]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[1]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[2]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[3]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[4]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[5]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[6]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[7]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[8]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[9]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[10]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[11]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[12]);
+    cipher = _mm_aesdec_si128(cipher, inverted_schedule->keys[13]);
+    return _mm_aesdeclast_si128(cipher, inverted_schedule->keys[14]);
+}
+
+static AesBlock128 __fastcall aes256_keygen_assist(
+    AesBlock128* prev_lo,
+    AesBlock128* prev_hi,
+    AesBlock128 hwgen)
+{
+    AesBlock128 tmp = *prev_lo;
+
+    tmp = _mm_slli_si128(tmp, 4);
+    *prev_lo = _mm_xor_si128(*prev_lo, tmp);
+    tmp = _mm_slli_si128(tmp, 4);
+    *prev_lo = _mm_xor_si128(*prev_lo, tmp);
+    tmp = _mm_slli_si128(tmp, 4);
+    *prev_lo = _mm_xor_si128(*prev_lo, tmp);
+
+    *prev_lo = _mm_xor_si128(*prev_lo, hwgen);
+
+    *prev_hi = _mm_xor_si128(*prev_hi, *prev_lo);
+    *prev_lo = _mm_xor_si128(*prev_lo, *prev_hi);
+    *prev_hi = _mm_xor_si128(*prev_hi, *prev_lo);
+
+    return *prev_hi;
+}
+
+void __fastcall raw_aes256_expand_key_schedule(
+    AesBlock128 key_lo,
+    AesBlock128 key_hi,
+    Aes256KeySchedule* key_schedule)
+{
+    AesBlock128 prev_lo, prev_hi;
+    AesBlock128 hwgen;
+
+    prev_lo = key_schedule->keys[0] = key_lo;
+    prev_hi = key_schedule->keys[1] = key_hi;
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0x01);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xff);
+    key_schedule->keys[2] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xaa);
+    key_schedule->keys[3] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0x02);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xff);
+    key_schedule->keys[4] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xaa);
+    key_schedule->keys[5] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0x04);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xff);
+    key_schedule->keys[6] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xaa);
+    key_schedule->keys[7] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0x08);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xff);
+    key_schedule->keys[8] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xaa);
+    key_schedule->keys[9] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0x10);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xff);
+    key_schedule->keys[10] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xaa);
+    key_schedule->keys[11] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0x20);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xff);
+    key_schedule->keys[12] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xaa);
+    key_schedule->keys[13] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+
+    hwgen = _mm_aeskeygenassist_si128(prev_hi, 0x40);
+    hwgen = _mm_shuffle_epi32(hwgen, 0xff);
+    key_schedule->keys[14] = aes256_keygen_assist(&prev_lo, &prev_hi, hwgen);
+}
+
+void __fastcall raw_aes256_invert_key_schedule(
+    Aes256KeySchedule* key_schedule,
+    Aes256KeySchedule* inverted_schedule)
+{
+    inverted_schedule->keys[0] = key_schedule->keys[14];
+    inverted_schedule->keys[1] = _mm_aesimc_si128(key_schedule->keys[13]);
+    inverted_schedule->keys[2] = _mm_aesimc_si128(key_schedule->keys[12]);
+    inverted_schedule->keys[3] = _mm_aesimc_si128(key_schedule->keys[11]);
+    inverted_schedule->keys[4] = _mm_aesimc_si128(key_schedule->keys[10]);
+    inverted_schedule->keys[5] = _mm_aesimc_si128(key_schedule->keys[9]);
+    inverted_schedule->keys[6] = _mm_aesimc_si128(key_schedule->keys[8]);
+    inverted_schedule->keys[7] = _mm_aesimc_si128(key_schedule->keys[7]);
+    inverted_schedule->keys[8] = _mm_aesimc_si128(key_schedule->keys[6]);
+    inverted_schedule->keys[9] = _mm_aesimc_si128(key_schedule->keys[5]);
+    inverted_schedule->keys[10] = _mm_aesimc_si128(key_schedule->keys[4]);
+    inverted_schedule->keys[11] = _mm_aesimc_si128(key_schedule->keys[3]);
+    inverted_schedule->keys[12] = _mm_aesimc_si128(key_schedule->keys[2]);
+    inverted_schedule->keys[13] = _mm_aesimc_si128(key_schedule->keys[1]);
+    inverted_schedule->keys[14] = key_schedule->keys[0];
+}