initial commit

5 files changed, 727 insertions, 0 deletions

diff --git a/src/aes128.asm b/src/aes128.asm
new file mode 100644
index 0000000..dfa7a7d
--- /dev/null
+++ b/src/aes128.asm
@@ -0,0 +1,129 @@
+; Copyright 2015 Egor Tensin <Egor.Tensin@gmail.com>
+; This file is licensed under the terms of the MIT License.
+; See LICENSE.txt for details.
+
+.586
+.xmm
+.model flat
+
+.data
+
+align 10h
+key_schedule oword 11 dup(0)
+
+align 10h
+inverted_key_schedule oword 11 dup(0)
+
+.code
+
+@aes128ecb_encrypt@32 proc
+    call expand_keys_128ecb
+    pxor xmm0, [key_schedule]
+    aesenc xmm0, [key_schedule + 10h]
+    aesenc xmm0, [key_schedule + 20h]
+    aesenc xmm0, [key_schedule + 30h]
+    aesenc xmm0, [key_schedule + 40h]
+    aesenc xmm0, [key_schedule + 50h]
+    aesenc xmm0, [key_schedule + 60h]
+    aesenc xmm0, [key_schedule + 70h]
+    aesenc xmm0, [key_schedule + 80h]
+    aesenc xmm0, [key_schedule + 90h]
+    aesenclast xmm0, [key_schedule + 0A0h]
+    ret
+@aes128ecb_encrypt@32 endp
+
+@aes128ecb_decrypt@32 proc
+    call expand_keys_128ecb
+    pxor xmm0, [inverted_key_schedule]
+    aesdec xmm0, [inverted_key_schedule + 10h]
+    aesdec xmm0, [inverted_key_schedule + 20h]
+    aesdec xmm0, [inverted_key_schedule + 30h]
+    aesdec xmm0, [inverted_key_schedule + 40h]
+    aesdec xmm0, [inverted_key_schedule + 50h]
+    aesdec xmm0, [inverted_key_schedule + 60h]
+    aesdec xmm0, [inverted_key_schedule + 70h]
+    aesdec xmm0, [inverted_key_schedule + 80h]
+    aesdec xmm0, [inverted_key_schedule + 90h]
+    aesdeclast xmm0, [inverted_key_schedule + 0A0h]
+    ret
+@aes128ecb_decrypt@32 endp
+
+expand_keys_128ecb proc
+    lea ecx, [key_schedule + 10h]
+    movdqa [key_schedule], xmm1
+
+    aeskeygenassist xmm7, xmm1, 01h
+    call gen_round_key
+    aeskeygenassist xmm7, xmm1, 02h
+    call gen_round_key
+    aeskeygenassist xmm7, xmm1, 04h
+    call gen_round_key
+    aeskeygenassist xmm7, xmm1, 08h
+    call gen_round_key
+    aeskeygenassist xmm7, xmm1, 10h
+    call gen_round_key
+    aeskeygenassist xmm7, xmm1, 20h
+    call gen_round_key
+    aeskeygenassist xmm7, xmm1, 40h
+    call gen_round_key
+    aeskeygenassist xmm7, xmm1, 80h
+    call gen_round_key
+    aeskeygenassist xmm7, xmm1, 1Bh
+    call gen_round_key
+    aeskeygenassist xmm7, xmm1, 36h
+    call gen_round_key
+
+    call invert_key_schedule
+    ret
+
+gen_round_key:
+    movdqa xmm6, xmm1    ; xmm6 = key_schedule[i]
+                         ; xmm6 = x3 x2 x1 x0
+
+    pslldq xmm6, 4     ; xmm6 = x2 x1 x0 0
+    pxor xmm1, xmm6    ; xmm1 = (x3 x2) (x2 x1) (x1 x0) x0
+    pslldq xmm6, 4     ; xmm6 = x1 x0 0 0
+    pxor xmm1, xmm6    ; xmm1 = (x3 x2 x1) (x2 x1 x0) (x1 x0) x0
+    pslldq xmm6, 4     ; xmm6 = x0 0 0 0
+    pxor xmm1, xmm6    ; xmm1 = (x3 x2 x1 x0) (x2 x1 x0) (x1 x0) x0
+
+    pshufd xmm7, xmm7, 0FFh
+    pxor xmm1, xmm7
+
+    movdqa [ecx], xmm1
+    add ecx, 10h
+    ret
+
+invert_key_schedule:
+    movdqa xmm7, [key_schedule]
+    movdqa xmm6, [key_schedule + 0A0h]
+    movdqa [inverted_key_schedule], xmm6
+    movdqa [inverted_key_schedule + 0A0h], xmm7
+
+    aesimc xmm7, [key_schedule + 10h]
+    aesimc xmm6, [key_schedule + 90h]
+    movdqa [inverted_key_schedule + 10h], xmm6
+    movdqa [inverted_key_schedule + 90h], xmm7
+
+    aesimc xmm7, [key_schedule + 20h]
+    aesimc xmm6, [key_schedule + 80h]
+    movdqa [inverted_key_schedule + 20h], xmm6
+    movdqa [inverted_key_schedule + 80h], xmm7
+
+    aesimc xmm7, [key_schedule + 30h]
+    aesimc xmm6, [key_schedule + 70h]
+    movdqa [inverted_key_schedule + 30h], xmm6
+    movdqa [inverted_key_schedule + 70h], xmm7
+
+    aesimc xmm7, [key_schedule + 40h]
+    aesimc xmm6, [key_schedule + 60h]
+    movdqa [inverted_key_schedule + 40h], xmm6
+    movdqa [inverted_key_schedule + 60h], xmm7
+
+    aesimc xmm7, [key_schedule + 50h]
+    movdqa [inverted_key_schedule + 50h], xmm7
+
+    ret
+expand_keys_128ecb endp
+
+end
diff --git a/src/aes192.asm b/src/aes192.asm
new file mode 100644
index 0000000..84d3a3a
--- /dev/null
+++ b/src/aes192.asm
@@ -0,0 +1,196 @@
+; Copyright 2015 Egor Tensin <Egor.Tensin@gmail.com>
+; This file is licensed under the terms of the MIT License.
+; See LICENSE.txt for details.
+
+.586
+.xmm
+.model flat
+
+.data
+
+align 10h
+key_schedule oword 13 dup(0)
+
+align 10h
+inverted_key_schedule oword 13 dup(0)
+
+.code
+
+@aes192ecb_encrypt@48 proc
+    call expand_keys_192ecb
+    pxor xmm0, [key_schedule]
+    aesenc xmm0, [key_schedule + 10h]
+    aesenc xmm0, [key_schedule + 20h]
+    aesenc xmm0, [key_schedule + 30h]
+    aesenc xmm0, [key_schedule + 40h]
+    aesenc xmm0, [key_schedule + 50h]
+    aesenc xmm0, [key_schedule + 60h]
+    aesenc xmm0, [key_schedule + 70h]
+    aesenc xmm0, [key_schedule + 80h]
+    aesenc xmm0, [key_schedule + 90h]
+    aesenc xmm0, [key_schedule + 0A0h]
+    aesenc xmm0, [key_schedule + 0B0h]
+    aesenclast xmm0, [key_schedule + 0C0h]
+    ret
+@aes192ecb_encrypt@48 endp
+
+@aes192ecb_decrypt@48 proc
+    call expand_keys_192ecb
+    pxor xmm0, [inverted_key_schedule]
+    aesdec xmm0, [inverted_key_schedule + 10h]
+    aesdec xmm0, [inverted_key_schedule + 20h]
+    aesdec xmm0, [inverted_key_schedule + 30h]
+    aesdec xmm0, [inverted_key_schedule + 40h]
+    aesdec xmm0, [inverted_key_schedule + 50h]
+    aesdec xmm0, [inverted_key_schedule + 60h]
+    aesdec xmm0, [inverted_key_schedule + 70h]
+    aesdec xmm0, [inverted_key_schedule + 80h]
+    aesdec xmm0, [inverted_key_schedule + 90h]
+    aesdec xmm0, [inverted_key_schedule + 0A0h]
+    aesdec xmm0, [inverted_key_schedule + 0B0h]
+    aesdeclast xmm0, [inverted_key_schedule + 0C0h]
+    ret
+@aes192ecb_decrypt@48 endp
+
+expand_keys_192ecb proc
+    ; key = k0 k1 k2 k3 k4 k5
+    ; xmm1 = k0 k1 k2 k3
+    ; xmm2 =  0  0 k5 k4
+
+    ; w[0] = k0 k1 k2 k3
+    ; w[1] = k4 k5  -  -
+
+    ; i = 6
+    ; while (i < 52):
+    ;     temp = w[i - 1]
+    ;     if (i % 6 == 0):
+    ;         temp = SubWord(RotWord(w[i - 1])) * Rcon
+    ;     w[i] = w[i - 6] * temp
+    ;     i = i + 1
+
+    ; w[6] = SubWord(RotWord(w[5])) * Rcon * w[0]
+    ; w[7] = w[6] * w[1]
+    ;      = SubWord(RotWord(w[5])) * Rcon * w[0] * w[1]
+    ; w[8] = w[7] * w[2]
+    ;      = SubWord(RotWord(w[5])) * Rcon * w[0] * w[1] * w[2]
+    ; w[9] = w[8] * w[3]
+    ;      = SubWord(RotWord(w[5])) * Rcon * w[0] * w[1] * w[2] * w[3]
+    ; w[10] = w[9] * w[4]
+    ;       = SubWord(RotWord(w[5])) * Rcon * w[0] * w[1] * w[2] * w[3] * w[4]
+    ; w[11] = w[10] * w[5]
+    ;       = SubWord(RotWord(w[5])) * Rcon * w[0] * w[1] * w[2] * w[3] * w[4] * w[5]
+
+    movdqa [key_schedule], xmm1
+    movdqa [key_schedule + 10h], xmm2
+
+    lea ecx, [key_schedule + 18h]
+    aeskeygenassist xmm7, xmm2, 1
+    call gen_round_key
+    aeskeygenassist xmm7, xmm2, 2
+    call gen_round_key
+    aeskeygenassist xmm7, xmm2, 4
+    call gen_round_key
+    aeskeygenassist xmm7, xmm2, 8
+    call gen_round_key
+    aeskeygenassist xmm7, xmm2, 10h
+    call gen_round_key
+    aeskeygenassist xmm7, xmm2, 20h
+    call gen_round_key
+    aeskeygenassist xmm7, xmm2, 40h
+    call gen_round_key
+    aeskeygenassist xmm7, xmm2, 80h
+    call gen_round_key
+
+    call invert_key_schedule
+    ret
+
+gen_round_key:
+    ; xmm1 = x3 x2 x1 x0
+    ; xmm2 =  -  - x5 x4
+    ; xmm7 = RotWord(SubWord(-)) xor Rcon
+    ;        SubWord(-)
+    ;        RotWord(SubWord(x5)) xor Rcon
+    ;        SubWord(x5)
+    movdqa xmm6, xmm1
+
+    pslldq xmm6, 4     ; xmm6 = x2 x1 x0 0
+    pxor xmm1, xmm6    ; xmm1 = (x3 * x2) (x1 * x2) (x1 * x0) x0
+    pslldq xmm6, 4     ; xmm6 = x1 x0 0 0
+    pxor xmm1, xmm6    ; xmm1 = (x3 * x2 * x1) (x1 * x2 * x0) (x1 * x0) x0
+    pslldq xmm6, 4     ; xmm6 = x0 0 0 0
+    pxor xmm1, xmm6    ; xmm1 = (x3 * x2 * x1 * x0) (x1 * x2 * x0) (x1 * x0) x0
+
+    pshufd xmm7, xmm7, 55h    ; xmm7 = RotWord(SubWord(x5)) * Rcon
+                              ;        RotWord(SubWord(x5)) * Rcon
+                              ;        RotWord(SubWord(x5)) * Rcon
+                              ;        RotWord(SubWord(x5)) * Rcon
+
+    pxor xmm1, xmm7    ; xmm1 = RotWord(SubWord(x5)) * Rcon * x3 * x2 * x1 * x0
+                       ;        RotWord(SubWord(x5)) * Rcon * x2 * x1 * x0
+                       ;        RotWord(SubWord(x5)) * Rcon * x1 * x0
+                       ;        RotWord(SubWord(x5)) * Rcon * x0
+
+    movq qword ptr [ecx], xmm1
+    add ecx, 8
+
+    pshufd xmm7, xmm1, 0FFh    ; xmm7 = -
+                               ;        -
+                               ;        RotWord(SubWord(x5)) * Rcon * x3 * x2 * x1 * x0
+                               ;        RotWord(SubWord(x5)) * Rcon * x3 * x2 * x1 * x0
+    pxor xmm7, xmm2           ; xmm7 = -
+                              ;        -
+                              ;        RotWord(SubWord(x5)) * Rcon * x5 * x3 * x2 * x1 * x0
+                              ;        RotWord(SubWord(x5)) * Rcon * x4 * x3 * x2 * x1 * x0
+    pslldq xmm2, 4            ; xmm2 = - k5 k4 0
+    pxor xmm7, xmm2           ; xmm7 = -
+                              ;        -
+                              ;        RotWord(SubWord(x5)) * Rcon * x5 * x4 * x3 * x2 * x1 * x0
+                              ;        RotWord(SubWord(x5)) * Rcon * x4 * x3 * x2 * x1 * x0
+
+    movq xmm2, xmm7
+    pslldq xmm7, 8
+    movdqa xmm6, xmm1
+    psrldq xmm6, 8
+    por xmm7, xmm6
+    movdqu [ecx], xmm7
+    add ecx, 10h
+    ret
+
+invert_key_schedule:
+    movdqa xmm7, [key_schedule]
+    movdqa xmm6, [key_schedule + 0C0h]
+    movdqa [inverted_key_schedule], xmm6
+    movdqa [inverted_key_schedule + 0C0h], xmm7
+
+    aesimc xmm7, [key_schedule + 10h]
+    aesimc xmm6, [key_schedule + 0B0h]
+    movdqa [inverted_key_schedule + 10h], xmm6
+    movdqa [inverted_key_schedule + 0B0h], xmm7
+
+    aesimc xmm7, [key_schedule + 20h]
+    aesimc xmm6, [key_schedule + 0A0h]
+    movdqa [inverted_key_schedule +  20h], xmm6
+    movdqa [inverted_key_schedule + 0A0h], xmm7
+
+    aesimc xmm7, [key_schedule + 30h]
+    aesimc xmm6, [key_schedule + 90h]
+    movdqa [inverted_key_schedule + 30h], xmm6
+    movdqa [inverted_key_schedule + 90h], xmm7
+
+    aesimc xmm7, [key_schedule + 40h]
+    aesimc xmm6, [key_schedule + 80h]
+    movdqa [inverted_key_schedule + 40h], xmm6
+    movdqa [inverted_key_schedule + 80h], xmm7
+
+    aesimc xmm7, [key_schedule + 50h]
+    aesimc xmm6, [key_schedule + 70h]
+    movdqa [inverted_key_schedule + 50h], xmm6
+    movdqa [inverted_key_schedule + 70h], xmm7
+
+    aesimc xmm7, [key_schedule + 60h]
+    movdqa [inverted_key_schedule + 60h], xmm7
+    
+    ret
+expand_keys_192ecb endp
+
+end
diff --git a/src/aes256cbc.asm b/src/aes256cbc.asm
new file mode 100644
index 0000000..3446d31
--- /dev/null
+++ b/src/aes256cbc.asm
@@ -0,0 +1,183 @@
+; Copyright 2015 Egor Tensin <Egor.Tensin@gmail.com>
+; This file is licensed under the terms of the MIT License.
+; See LICENSE.txt for details.
+
+.586
+.xmm
+.model flat
+
+.data
+
+align 10h
+key_schedule oword 15 dup(0)
+
+align 10h
+inverse_key_schedule oword 15 dup(0)
+
+.code
+
+@aes256cbc_encrypt@52 proc
+    call expand_keys_256cbc
+    pxor xmm0, [ecx]
+    pxor xmm0, [key_schedule]
+    aesenc xmm0, [key_schedule + 10h]
+    aesenc xmm0, [key_schedule + 20h]
+    aesenc xmm0, [key_schedule + 30h]
+    aesenc xmm0, [key_schedule + 40h]
+    aesenc xmm0, [key_schedule + 50h]
+    aesenc xmm0, [key_schedule + 60h]
+    aesenc xmm0, [key_schedule + 70h]
+    aesenc xmm0, [key_schedule + 80h]
+    aesenc xmm0, [key_schedule + 90h]
+    aesenc xmm0, [key_schedule + 0A0h]
+    aesenc xmm0, [key_schedule + 0B0h]
+    aesenc xmm0, [key_schedule + 0C0h]
+    aesenc xmm0, [key_schedule + 0D0h]
+    aesenclast xmm0, [key_schedule + 0E0h]
+    ret
+@aes256cbc_encrypt@52 endp
+
+expand_keys_256cbc proc
+    lea edx, [key_schedule + 20h]
+    movdqa [key_schedule], xmm1
+    movdqa [key_schedule + 10h], xmm2
+
+    aeskeygenassist xmm7, xmm2, 1h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 2h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 4h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 8h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 10h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 20h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 40h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    call invert_key_schedule
+    ret
+
+gen_round_key:
+    movdqa xmm6, xmm1
+
+    pslldq xmm6, 4
+    pxor xmm1, xmm6
+    pslldq xmm6, 4
+    pxor xmm1, xmm6
+    pslldq xmm6, 4
+    pxor xmm1, xmm6
+
+    pxor xmm1, xmm7
+
+    movdqa [edx], xmm1
+    add edx, 10h
+
+    movdqa xmm7, xmm1
+    movdqa xmm1, xmm2
+    movdqa xmm2, xmm7
+    ret
+
+invert_key_schedule:
+    movdqa xmm7, [key_schedule]
+    movdqa xmm6, [key_schedule + 0E0h]
+    movdqa [inverse_key_schedule], xmm6
+    movdqa [inverse_key_schedule + 0E0h], xmm7
+
+    aesimc xmm7, [key_schedule + 10h]
+    aesimc xmm6, [key_schedule + 0D0h]
+    movdqa [inverse_key_schedule + 10h], xmm6
+    movdqa [inverse_key_schedule + 0D0h], xmm7
+
+    aesimc xmm7, [key_schedule + 20h]
+    aesimc xmm6, [key_schedule + 0C0h]
+    movdqa [inverse_key_schedule + 20h], xmm6
+    movdqa [inverse_key_schedule + 0C0h], xmm7
+
+    aesimc xmm7, [key_schedule + 30h]
+    aesimc xmm6, [key_schedule + 0B0h]
+    movdqa [inverse_key_schedule + 30h], xmm6
+    movdqa [inverse_key_schedule + 0B0h], xmm7
+
+    aesimc xmm7, [key_schedule + 40h]
+    aesimc xmm6, [key_schedule + 0A0h]
+    movdqa [inverse_key_schedule + 40h], xmm6
+    movdqa [inverse_key_schedule + 0A0h], xmm7
+
+    aesimc xmm7, [key_schedule + 50h]
+    aesimc xmm6, [key_schedule + 90h]
+    movdqa [inverse_key_schedule + 50h], xmm6
+    movdqa [inverse_key_schedule + 90h], xmm7
+
+    aesimc xmm7, [key_schedule + 60h]
+    aesimc xmm6, [key_schedule + 80h]
+    movdqa [inverse_key_schedule + 60h], xmm6
+    movdqa [inverse_key_schedule + 80h], xmm7
+
+    aesimc xmm7, [key_schedule + 70h]
+    movdqa [inverse_key_schedule + 70h], xmm7
+
+    ret
+expand_keys_256cbc endp
+
+@aes256cbc_decrypt@52 proc
+    call expand_keys_256cbc
+    pxor xmm0, [inverse_key_schedule]
+    aesdec xmm0, [inverse_key_schedule + 10h]
+    aesdec xmm0, [inverse_key_schedule + 20h]
+    aesdec xmm0, [inverse_key_schedule + 30h]
+    aesdec xmm0, [inverse_key_schedule + 40h]
+    aesdec xmm0, [inverse_key_schedule + 50h]
+    aesdec xmm0, [inverse_key_schedule + 60h]
+    aesdec xmm0, [inverse_key_schedule + 70h]
+    aesdec xmm0, [inverse_key_schedule + 80h]
+    aesdec xmm0, [inverse_key_schedule + 90h]
+    aesdec xmm0, [inverse_key_schedule + 0A0h]
+    aesdec xmm0, [inverse_key_schedule + 0B0h]
+    aesdec xmm0, [inverse_key_schedule + 0C0h]
+    aesdec xmm0, [inverse_key_schedule + 0D0h]
+    aesdeclast xmm0, [inverse_key_schedule + 0E0h]    
+    pxor xmm0, [ecx]
+    ret
+@aes256cbc_decrypt@52 endp
+
+end
diff --git a/src/aes256ecb.asm b/src/aes256ecb.asm
new file mode 100644
index 0000000..4246dee
--- /dev/null
+++ b/src/aes256ecb.asm
@@ -0,0 +1,181 @@
+; Copyright 2015 Egor Tensin <Egor.Tensin@gmail.com>
+; This file is licensed under the terms of the MIT License.
+; See LICENSE.txt for details.
+
+.586
+.xmm
+.model flat
+
+.data
+
+align 10h
+key_schedule oword 15 dup(0)
+
+align 10h
+inverse_key_schedule oword 15 dup(0)
+
+.code
+
+@aes256ecb_encrypt@48 proc
+    call expand_keys_256ecb
+    pxor xmm0, [key_schedule]
+    aesenc xmm0, [key_schedule + 10h]
+    aesenc xmm0, [key_schedule + 20h]
+    aesenc xmm0, [key_schedule + 30h]
+    aesenc xmm0, [key_schedule + 40h]
+    aesenc xmm0, [key_schedule + 50h]
+    aesenc xmm0, [key_schedule + 60h]
+    aesenc xmm0, [key_schedule + 70h]
+    aesenc xmm0, [key_schedule + 80h]
+    aesenc xmm0, [key_schedule + 90h]
+    aesenc xmm0, [key_schedule + 0A0h]
+    aesenc xmm0, [key_schedule + 0B0h]
+    aesenc xmm0, [key_schedule + 0C0h]
+    aesenc xmm0, [key_schedule + 0D0h]
+    aesenclast xmm0, [key_schedule + 0E0h]
+    ret
+@aes256ecb_encrypt@48 endp
+
+expand_keys_256ecb proc
+    lea edx, [key_schedule + 20h]
+    movdqa [key_schedule], xmm1
+    movdqa [key_schedule + 10h], xmm2
+
+    aeskeygenassist xmm7, xmm2, 1h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 2h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 4h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 8h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 10h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 20h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 0
+    pshufd xmm7, xmm7, 0AAh
+    call gen_round_key
+
+    aeskeygenassist xmm7, xmm2, 40h
+    pshufd xmm7, xmm7, 0FFh
+    call gen_round_key
+
+    call invert_key_schedule
+    ret
+
+gen_round_key:
+    movdqa xmm6, xmm1
+
+    pslldq xmm6, 4
+    pxor xmm1, xmm6
+    pslldq xmm6, 4
+    pxor xmm1, xmm6
+    pslldq xmm6, 4
+    pxor xmm1, xmm6
+
+    pxor xmm1, xmm7
+
+    movdqa [edx], xmm1
+    add edx, 10h
+
+    movdqa xmm7, xmm1
+    movdqa xmm1, xmm2
+    movdqa xmm2, xmm7
+    ret
+
+invert_key_schedule:
+    movdqa xmm7, [key_schedule       ]
+    movdqa xmm6, [key_schedule + 0E0h]
+    movdqa [inverse_key_schedule       ], xmm6
+    movdqa [inverse_key_schedule + 0E0h], xmm7
+
+    aesimc xmm7, [key_schedule +  10h]
+    aesimc xmm6, [key_schedule + 0D0h]
+    movdqa [inverse_key_schedule +  10h], xmm6
+    movdqa [inverse_key_schedule + 0D0h], xmm7
+
+    aesimc xmm7, [key_schedule +  20h]
+    aesimc xmm6, [key_schedule + 0C0h]
+    movdqa [inverse_key_schedule +  20h], xmm6
+    movdqa [inverse_key_schedule + 0C0h], xmm7
+
+    aesimc xmm7, [key_schedule +  30h]
+    aesimc xmm6, [key_schedule + 0B0h]
+    movdqa [inverse_key_schedule +  30h], xmm6
+    movdqa [inverse_key_schedule + 0B0h], xmm7
+
+    aesimc xmm7, [key_schedule +  40h]
+    aesimc xmm6, [key_schedule + 0A0h]
+    movdqa [inverse_key_schedule +  40h], xmm6
+    movdqa [inverse_key_schedule + 0A0h], xmm7
+
+    aesimc xmm7, [key_schedule + 50h]
+    aesimc xmm6, [key_schedule + 90h]
+    movdqa [inverse_key_schedule + 50h], xmm6
+    movdqa [inverse_key_schedule + 90h], xmm7
+
+    aesimc xmm7, [key_schedule + 60h]
+    aesimc xmm6, [key_schedule + 80h]
+    movdqa [inverse_key_schedule + 60h], xmm6
+    movdqa [inverse_key_schedule + 80h], xmm7
+
+    aesimc xmm7, [key_schedule + 70h]
+    movdqa [inverse_key_schedule + 70h], xmm7
+
+    ret
+expand_keys_256ecb endp
+
+@aes256ecb_decrypt@48 proc
+    call expand_keys_256ecb
+    pxor xmm0, [inverse_key_schedule]
+    aesdec xmm0, [inverse_key_schedule + 10h]
+    aesdec xmm0, [inverse_key_schedule + 20h]
+    aesdec xmm0, [inverse_key_schedule + 30h]
+    aesdec xmm0, [inverse_key_schedule + 40h]
+    aesdec xmm0, [inverse_key_schedule + 50h]
+    aesdec xmm0, [inverse_key_schedule + 60h]
+    aesdec xmm0, [inverse_key_schedule + 70h]
+    aesdec xmm0, [inverse_key_schedule + 80h]
+    aesdec xmm0, [inverse_key_schedule + 90h]
+    aesdec xmm0, [inverse_key_schedule + 0A0h]
+    aesdec xmm0, [inverse_key_schedule + 0B0h]
+    aesdec xmm0, [inverse_key_schedule + 0C0h]
+    aesdec xmm0, [inverse_key_schedule + 0D0h]
+    aesdeclast xmm0, [inverse_key_schedule + 0E0h]    
+    ret
+@aes256ecb_decrypt@48 endp
+
+end
diff --git a/src/common.c b/src/common.c
new file mode 100644
index 0000000..92f5e24
--- /dev/null
+++ b/src/common.c
@@ -0,0 +1,38 @@
+/**
+ * \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 <intrin.h>
+
+#include <stdio.h>
+
+AesBlock make_aes_block(int highest, int high, int low, int lowest)
+{
+    return _mm_set_epi32(highest, high, low, lowest);
+}
+
+AesState aes_block_to_state(AesBlock block)
+{
+    AesState state;
+    _mm_storeu_si128((__m128i*) &state.bytes, block);
+    return state;
+}
+
+void print_aes_block(AesBlock block)
+{
+    int i, j;
+    AesState state = aes_block_to_state(block);
+
+    for (i = 0; i < 4; ++i)
+    {
+        for (j = 0; j < 3; ++j)
+            printf("%02x ", state.bytes[j][i]);
+        printf("%02x\n", state.bytes[3][i]);
+    }
+}