+ <CustomBuild Include="Gost89_x64.asm">

+ <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">true</ExcludedFromBuild>

+ <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">echo %(Filename)%(Extension) &amp; nasm.exe -Xvc -f win64 -Ox -o "$(TargetDir)\%(Filename).obj" -l "$(TargetDir)\%(Filename).lst" "%(FullPath)"

+</Command>

+ <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)\%(Filename).obj;%(Outputs)</Outputs>

+ <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">true</ExcludedFromBuild>

+ <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">echo %(Filename)%(Extension) &amp; nasm.exe -Xvc -f win64 -Ox -o "$(TargetDir)\%(Filename).obj" -l "$(TargetDir)\%(Filename).lst" "%(FullPath)"

+</Command>

+ <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)\%(Filename).obj;%(Outputs)</Outputs>

+ </CustomBuild>

+ <ClCompile Include="GostCipher.c" />

+ <ClInclude Include="GostCipher.h" />

+ <ClCompile Include="GostCipher.c">

+ <Filter>Source Files</Filter>

+ </ClCompile>

+ <ClInclude Include="GostCipher.h">

+ <Filter>Header Files</Filter>

+ </ClInclude>

+ <CustomBuild Include="Gost89_x64.asm">

+ <Filter>Source Files</Filter>

+ </CustomBuild>

+/** @file

+GOST89 implementation

+

+Copyright (c) 2016. Disk Cryptography Services for EFI (DCS), Alex Kolotnikov

+

+This program and the accompanying materials

+are licensed and made available under the terms and conditions

+of the Apache License, Version 2.0. The full text of the license may be found at

+https://opensource.org/licenses/Apache-2.0

+

+Dynamic SBOX idea is from GostCrypt project. Copyright (c) 2008-2011 TrueCrypt Developers Association

+**/

+

+

+

+#include "GostCipher.h"

+#include "Streebog.h"

+#include "common\Tcdefs.h"

+#include "cpu.h"

+

+#if defined(CIPHER_GOST89)

+

+// Crypto Pro

+byte S_CryptoPro[8][16] = {

+ {0x1,0x3,0xA,0x9,0x5,0xB,0x4,0xF,0x8,0x6,0x7,0xE,0xD,0x0,0x2,0xC},

+ {0xD,0xE,0x4,0x1,0x7,0x0,0x5,0xA,0x3,0xC,0x8,0xF,0x6,0x2,0x9,0xB},

+ {0x7,0x6,0x2,0x4,0xD,0x9,0xF,0x0,0xA,0x1,0x5,0xB,0x8,0xE,0xC,0x3},

+ {0x7,0x6,0x4,0xB,0x9,0xC,0x2,0xA,0x1,0x8,0x0,0xE,0xF,0xD,0x3,0x5},

+ {0x4,0xA,0x7,0xC,0x0,0xF,0x2,0x8,0xE,0x1,0x6,0x5,0xD,0xB,0x9,0x3},

+ {0x7,0xF,0xC,0xE,0x9,0x4,0x1,0x0,0x3,0xB,0x5,0x2,0x6,0xA,0x8,0xD},

+ {0x5,0xF,0x4,0x0,0x2,0xD,0xB,0x9,0x1,0x7,0x6,0x3,0xC,0xE,0xA,0x8},

+ {0xA,0x4,0x5,0x6,0x8,0x1,0x3,0x7,0xD,0xC,0xE,0x0,0x9,0x2,0xB,0xF}

+ };

+

+// TC26

+byte S_TC26[8][16] =

+{

+ { 0xc, 0x4, 0x6, 0x2, 0xa, 0x5, 0xb, 0x9, 0xe, 0x8, 0xd, 0x7, 0x0, 0x3, 0xf, 0x1 },

+ { 0x6, 0x8, 0x2, 0x3, 0x9, 0xa, 0x5, 0xc, 0x1, 0xe, 0x4, 0x7, 0xb, 0xd, 0x0, 0xf },

+ { 0xb, 0x3, 0x5, 0x8, 0x2, 0xf, 0xa, 0xd, 0xe, 0x1, 0x7, 0x4, 0xc, 0x9, 0x6, 0x0 },

+ { 0xc, 0x8, 0x2, 0x1, 0xd, 0x4, 0xf, 0x6, 0x7, 0x0, 0xa, 0x5, 0x3, 0xe, 0x9, 0xb },

+ { 0x7, 0xf, 0x5, 0xa, 0x8, 0x1, 0x6, 0xd, 0x0, 0x9, 0x3, 0xe, 0xb, 0x4, 0x2, 0xc },

+ { 0x5, 0xd, 0xf, 0x6, 0x9, 0x2, 0xc, 0xa, 0xb, 0x7, 0x8, 0x1, 0x4, 0x3, 0xe, 0x0 },

+ { 0x8, 0xe, 0x2, 0x5, 0x6, 0x9, 0x1, 0xc, 0xf, 0x4, 0xb, 0x0, 0xd, 0xa, 0x3, 0x7 },

+ { 0x1, 0x7, 0xe, 0xd, 0x0, 0x5, 0x8, 0x3, 0x4, 0xf, 0xa, 0x6, 0x9, 0xc, 0xb, 0x2 },

+};

+

+void gost_prepare_kds(gost_kds* kds) {

+ uint32 i;

+ // Build substitution tables.

+ for (i = 0; i < 256; ++i) {

+ uint32 p;

+ p = kds->sbox[7][i >> 4] << 4 | kds->sbox[6][i & 15];

+ p = p << 24; p = p << 11 | p >> 21;

+ kds->sbox_cvt[i] = p; // S87

+

+ p = kds->sbox[5][i >> 4] << 4 | kds->sbox[4][i & 15];

+ p = p << 16; p = p << 11 | p >> 21;

+ kds->sbox_cvt[256 + i] = p; // S65

+

+ p = kds->sbox[3][i >> 4] << 4 | kds->sbox[2][i & 15];

+ p = p << 8; p = p << 11 | p >> 21;

+ kds->sbox_cvt[256 * 2 + i] = p; // S43

+

+ p = kds->sbox[1][i >> 4] << 4 | kds->sbox[0][i & 15];

+ p = p << 11 | p >> 21;

+ kds->sbox_cvt[256 * 3 + i] = p; // S21

+ }

+}

+

+

+#ifdef GOST_DYNAMIC_SBOXES

+static void xor_s_box(byte s_box[8][16], byte *seed)

+{

+ int i;

+ for (i = 0; i < 16; i++)

+ {

+ s_box[1][i] ^= (seed[ (i * 4) + 0 ] ) & 0xF;

+ s_box[2][i] ^= (seed[ (i * 4) + 0 ]>>4) & 0xF;

+ s_box[3][i] ^= (seed[ (i * 4) + 1 ] ) & 0xF;

+ s_box[4][i] ^= (seed[ (i * 4) + 1 ]>>4) & 0xF;

+ s_box[5][i] ^= (seed[ (i * 4) + 2 ] ) & 0xF;

+ s_box[6][i] ^= (seed[ (i * 4) + 2 ]>>4) & 0xF;

+ s_box[7][i] ^= (seed[ (i * 4) + 3 ] ) & 0xF;

+ s_box[8][i] ^= (seed[ (i * 4) + 3 ]>>4) & 0xF;

+ }

+}

+#endif

+

+void gost_set_key(byte *key, gost_kds *ks)

+{

+#ifdef GOST_DYNAMIC_SBOXES

+ STREEBOG_CTX sctx;

+ byte sbox_seed[64];

+#if defined (DEVICE_DRIVER) && !defined (_WIN64)

+ KFLOATING_SAVE floatingPointState;

+ NTSTATUS saveStatus = STATUS_SUCCESS;

+ if (HasSSE2() || HasSSE41())

+ saveStatus = KeSaveFloatingPointState (&floatingPointState);

+#endif

+#endif

+

+ memcpy(ks->key, key, GOST_KEYSIZE);

+ memcpy(ks->sbox, S_TC26, sizeof(ks->sbox));

+

+#ifdef GOST_DYNAMIC_SBOXES

+ //Generate pseudorandom data based on the key

+ STREEBOG_init(&sctx);

+ STREEBOG_add(&sctx, key, 32);

+ STREEBOG_finalize(&sctx, sbox_seed);

+

+#if defined (DEVICE_DRIVER) && !defined (_WIN64)

+ if (NT_SUCCESS (saveStatus) && (HasSSE2() || HasSSE41()))

+ KeRestoreFloatingPointState (&floatingPointState);

+#endif

+

+ xor_s_box(ks->sbox, sbox_seed);

+#endif

+

+ gost_prepare_kds(ks);

+}

+

+static uint32 f(uint32 v, uint32* sbox){

+ byte* x =(byte*) &v;

+ /* Do substitutions */

+ return sbox[x[3]] | sbox[256 + x[2]] | sbox[256*2 + x[1]] | sbox[256*3 + x[0]];

+}

+

+void gost_encrypt_block(uint64 in_, uint64* out_, gost_kds* kds) {

+ uint32* in = (uint32*)&in_;

+ uint32* out = (uint32*)out_;

+ uint32* key = (uint32*)kds->key;

+ uint32* sbox = kds->sbox_cvt;

+

+ // As named in the GOST

+ uint32 n1 = in[0];

+ uint32 n2 = in[1];

+

+ n2 ^= f(n1+key[0], sbox);

+ n1 ^= f(n2+key[1], sbox);

+ n2 ^= f(n1+key[2], sbox);

+ n1 ^= f(n2+key[3], sbox);

+ n2 ^= f(n1+key[4], sbox);

+ n1 ^= f(n2+key[5], sbox);

+ n2 ^= f(n1+key[6], sbox);

+ n1 ^= f(n2+key[7], sbox);

+

+ n2 ^= f(n1+key[0], sbox);

+ n1 ^= f(n2+key[1], sbox);

+ n2 ^= f(n1+key[2], sbox);

+ n1 ^= f(n2+key[3], sbox);

+ n2 ^= f(n1+key[4], sbox);

+ n1 ^= f(n2+key[5], sbox);

+ n2 ^= f(n1+key[6], sbox);

+ n1 ^= f(n2+key[7], sbox);

+

+ n2 ^= f(n1+key[0], sbox);

+ n1 ^= f(n2+key[1], sbox);

+ n2 ^= f(n1+key[2], sbox);

+ n1 ^= f(n2+key[3], sbox);

+ n2 ^= f(n1+key[4], sbox);

+ n1 ^= f(n2+key[5], sbox);

+ n2 ^= f(n1+key[6], sbox);

+ n1 ^= f(n2+key[7], sbox);

+

+ n2 ^= f(n1+key[7], sbox);

+ n1 ^= f(n2+key[6], sbox);

+ n2 ^= f(n1+key[5], sbox);

+ n1 ^= f(n2+key[4], sbox);

+ n2 ^= f(n1+key[3], sbox);

+ n1 ^= f(n2+key[2], sbox);

+ n2 ^= f(n1+key[1], sbox);

+ n1 ^= f(n2+key[0], sbox);

+

+ // There is no swap after the last round

+ out[0] = n2;

+ out[1] = n1;

+}

+

+void gost_decrypt_block(uint64 in_, uint64* out_, gost_kds* kds) {

+ uint32* in = (uint32*)&in_;

+ uint32* out = (uint32*)out_;

+ uint32* key = (uint32*)kds->key;

+ uint32* sbox = kds->sbox_cvt;

+

+ // As named in the GOST

+ uint32 n1 = in[0];

+ uint32 n2 = in[1];

+

+ n2 ^= f(n1+key[0], sbox);

+ n1 ^= f(n2+key[1], sbox);

+ n2 ^= f(n1+key[2], sbox);

+ n1 ^= f(n2+key[3], sbox);

+ n2 ^= f(n1+key[4], sbox);

+ n1 ^= f(n2+key[5], sbox);

+ n2 ^= f(n1+key[6], sbox);

+ n1 ^= f(n2+key[7], sbox);

+

+ n2 ^= f(n1+key[7], sbox);

+ n1 ^= f(n2+key[6], sbox);

+ n2 ^= f(n1+key[5], sbox);

+ n1 ^= f(n2+key[4], sbox);

+ n2 ^= f(n1+key[3], sbox);

+ n1 ^= f(n2+key[2], sbox);

+ n2 ^= f(n1+key[1], sbox);

+ n1 ^= f(n2+key[0], sbox);

+

+ n2 ^= f(n1+key[7], sbox);

+ n1 ^= f(n2+key[6], sbox);

+ n2 ^= f(n1+key[5], sbox);

+ n1 ^= f(n2+key[4], sbox);

+ n2 ^= f(n1+key[3], sbox);

+ n1 ^= f(n2+key[2], sbox);

+ n2 ^= f(n1+key[1], sbox);

+ n1 ^= f(n2+key[0], sbox);

+

+ n2 ^= f(n1+key[7], sbox);

+ n1 ^= f(n2+key[6], sbox);

+ n2 ^= f(n1+key[5], sbox);

+ n1 ^= f(n2+key[4], sbox);

+ n2 ^= f(n1+key[3], sbox);

+ n1 ^= f(n2+key[2], sbox);

+ n2 ^= f(n1+key[1], sbox);

+ n1 ^= f(n2+key[0], sbox);

+

+ out[0] = n2;

+ out[1] = n1;

+}

+

+#if defined(_M_AMD64)

+void gost_encrypt_128_CBC_asm(byte *in, byte *out, gost_kds *ks, uint64 count);

+void gost_decrypt_128_CBC_asm(byte *in, byte *out, gost_kds *ks, uint64 count);

+#endif

+

+void gost_encrypt(byte *in, byte *out, gost_kds *ks, int count) {

+#if defined(_M_AMD64)

+ gost_encrypt_128_CBC_asm(in, out, ks, (uint64)count);

+#else

+ while (count > 0) {

+ // encrypt two blocks in CBC mode

+ gost_encrypt_block(*((uint64*)in), (uint64*)out, ks);

+ *((gst_udword*)(out + 8)) = *((gst_udword*)(in + 8)) ^ *((gst_udword*)(out));

+ *((gst_udword*)(out + 12)) = *((gst_udword*)(in + 12)) ^ *((gst_udword*)(out + 4));

+ gost_encrypt_block(*((uint64*)(out + 8)), (uint64*)(out + 8), ks);

+ count--;

+ in += 16;

+ out += 16;

+ }

+#endif

+}

+

+void gost_decrypt(byte *in, byte *out, gost_kds *ks, int count) {

+#if defined(_M_AMD64)

+ gost_decrypt_128_CBC_asm(in, out, ks, (uint64)count);

+#else

+ while (count > 0) {

+ // decrypt two blocks in CBC mode

+ gost_decrypt_block(*((uint64*)(in + 8)), (uint64*)(out + 8), ks);

+ *((gst_udword*)(out + 8)) ^= *((gst_udword*)(in));;

+ *((gst_udword*)(out + 12)) ^= *((gst_udword*)(in + 4));;

+ gost_decrypt_block(*((uint64*)(in)), (uint64*)(out), ks);

+ count--;

+ in += 16;

+ out += 16;

+ }

+#endif

+}

+

+#endif \ No newline at end of file

+

+/*

+ Copyright (c) 2008-2011 TrueCrypt Developers Association. All rights reserved.

+

+ Governed by the TrueCrypt License 3.0 the full text of which is contained in

+ the file License.txt included in TrueCrypt binary and source code distribution

+ packages.

+*/

+

+

+

+#ifndef GOST_CIPHER_H

+#define GOST_CIPHER_H

+

+//In unsigned chars

+#define GOST_KEYSIZE 32

+#define GOST_BLOCKSIZE 8

+#define GOST_SBOX_SIZE 16

+

+//Production setting, but can be turned off to compare the algorithm with other implementations

+#define CIPHER_GOST89

+#define GOST_DYNAMIC_SBOXES

+

+#if defined(CIPHER_GOST89)

+

+#ifndef rotl32

+#define rotl32(b, shift) ((b << shift) | (b >> (32 - shift)))

+#endif

+

+typedef unsigned char byte;

+#ifdef GST_WINDOWS_BOOT

+typedef int gst_word;

+typedef long gst_dword;

+typedef unsigned int gst_uword;

+typedef unsigned long gst_udword;

+#else

+typedef short gst_word;

+typedef int gst_dword;

+typedef unsigned short gst_uword;

+typedef unsigned int gst_udword;

+#endif

+

+typedef struct gost_kds

+{

+ byte key[32];

+ gst_udword sbox_cvt[256 * 4];

+ byte sbox[8][16];

+} gost_kds;

+

+#define GOST_KS (sizeof(gost_kds))

+

+void gost_encrypt(byte *in, byte *out, gost_kds *ks, int count);

+void gost_decrypt(byte *in, byte *out, gost_kds *ks, int count);

+void gost_set_key(byte *key, gost_kds *ks);

+

+#else

+#define GOST_KS (0)

+#endif

+

+

+#endif

+"$(OBJ_PATH)\$(O)\gost89_$(TC_ARCH).obj": gost89_$(TC_ARCH).asm

+ nasm.exe $(TC_ASFLAGS) -o "$@" -l "$(OBJ_PATH)\$(O)\gost89_$(TC_ARCH).lst" gost89_$(TC_ARCH).asm 2>$(TC_ASM_ERR_LOG)

+

+ "$(OBJ_PATH)\$(O)\Aes_hw_cpu.obj" \

+ "$(OBJ_PATH)\$(O)\gost89_$(TC_ARCH).obj"

+ gost89_$(TC_ARCH).asm \

+ GostCipher.c \

+;

+; GOST89 implementation x64

+;

+; Copyright (c) 2016. Disk Cryptography Services for EFI (DCS), Alex Kolotnikov

+;

+; This program and the accompanying materials

+; are licensed and made available under the terms and conditions

+; of the Apache License, Version 2.0. The full text of the license may be found at

+; https://opensource.org/licenses/Apache-2.0

+;

+; Some ideas from article https://xakep.ru/2013/10/19/shifrovanie-gost-28147-89/

+;

+

+[section .bss align=16]

+

+;///////////////////////////////////////////////////////////////////

+;// Win64 registers to save

+;///////////////////////////////////////////////////////////////////

+%macro SaveRegs 0

+ sub rsp,8*8+10*16

+ mov [rsp], rbx

+ mov [rsp+8], rbp

+ mov [rsp+8*2], rdi

+ mov [rsp+8*3], rsi

+ mov [rsp+8*4], r12

+ mov [rsp+8*5], r13

+ mov [rsp+8*6], r14

+ mov [rsp+8*7], r15

+%endmacro

+

+%macro RestoreRegs 0

+ mov rbx, [rsp]

+ mov rbp, [rsp+8]

+ mov rdi, [rsp+8*2]

+ mov rsi, [rsp+8*3]

+ mov r12, [rsp+8*4]

+ mov r13, [rsp+8*5]

+ mov r14, [rsp+8*6]

+ mov r15, [rsp+8*7]

+ add rsp,8*8+10*16

+%endmacro

+

+[section .text align=16]

+;///////////////////////////////////////////////////////////////////

+;// Crypting 2 blocks

+;///////////////////////////////////////////////////////////////////

+%macro gost_round2 2 ; 1 - pos1, 2 - pos2

+ ; 1st

+ ; 1-2 byte

+ add ecx, r13d ; add key

+ movzx edi, cl

+ movzx esi, ch

+ xor r14d, dword [r8 + 32 + 256*3*4 + rdi*4]

+ xor r14d, dword [r8 + 32 + 256*2*4 + rsi*4]

+ shr ecx, 16

+ ; 3-4 áàéò

+ movzx edi, cl

+ xor r14d, dword [r8 + 32 + 256*4 + rdi*4]

+ movzx esi, ch

+ xor r14d, dword [r8 + 32 + rsi*4]

+ mov edx, [r8 + %1*4] ; read key for second step

+

+ ; 2nd

+ ; 1-2 byte

+ add eax, r10d ; read key

+ movzx r15d, al

+ movzx ebp, ah

+ xor r11d, dword [r8 + 32 + 256*3*4 + r15*4]

+ xor r11d, dword [r8 + 32 + 256*2*4 + rbp*4]

+ shr eax, 16

+ ; 3-4 áàéò

+ movzx r15d, al

+ xor r11d, dword [r8 + 32 + 256*4 + r15*4]

+ movzx ebp, ah

+ xor r11d, dword [r8 + 32 + rbp*4]

+ mov ebx, [r8 + %1*4] ; read key for second step

+

+ ; second step

+ ; 1st

+ ; 1-2 byte

+ add edx, r14d ; add key

+ movzx edi, dl

+ movzx esi, dh

+ xor r13d, dword [r8 + 32 + 256*3*4 + rdi*4]

+ xor r13d, dword [r8 + 32 + 256*2*4 + rsi*4]

+ shr edx, 16

+ ; 3-4 áàéò

+ movzx edi, dl

+ xor r13d, dword [r8 + 32 + 256*4 + rdi*4]

+ movzx esi, dh

+ xor r13d, dword [r8 + 32 + rsi*4]

+ mov ecx, [r8 + %2*4] ; read key

+

+ ; 2nd

+ ; 1-2 byte

+ add ebx, r11d; ; add key

+ movzx r15d, bl;

+ movzx ebp, bh;

+ xor r10d, dword [r8 + 32 + 256*3*4 + r15*4]

+ xor r10d, dword [r8 + 32 + 256*2*4 + rbp*4]

+ shr ebx, 16

+ ; 3-4 áàéò

+ movzx r15d, bl

+ xor r10d, dword [r8 + 32 + 256*4 + r15*4]

+ movzx ebp, bh

+ xor r10d, dword [r8 + 32 + rbp*4]

+ mov eax, [r8 + %2*4] ; read key

+%endmacro

+

+; input: r8 - &key, rdx - &IN

+; returns: (r13) & (r10)

+GostEncrypt2x64:

+ ; 1st

+ mov r13d, [rdx]

+ mov r14, [rdx]

+ shr r14, 32

+

+ ; 2nd

+ mov r10d, [rdx + 16]

+ mov r11, [rdx + 16]

+ shr r11, 32

+

+ mov ecx, [r8]

+ mov eax, ecx

+

+ gost_round2 1, 2

+ gost_round2 3, 4

+ gost_round2 5, 6

+ gost_round2 7, 0

+

+ gost_round2 1, 2

+ gost_round2 3, 4

+ gost_round2 5, 6

+ gost_round2 7, 0

+

+ gost_round2 1, 2

+ gost_round2 3, 4

+ gost_round2 5, 6

+ gost_round2 7, 7

+

+ gost_round2 6, 5

+ gost_round2 4, 3

+ gost_round2 2, 1

+ gost_round2 0, 0

+

+ shl r13, 32 ; combine

+ or r13, r14

+

+ shl r10, 32 ; combine

+ or r10, r11

+ ret

+

+; input: r8 - &key, rdx - &IN

+; returns: (r13) & (r10)

+GostDecrypt2x64:

+ ; 1st

+ mov r13d, [rdx]

+ mov r14, [rdx]

+ shr r14, 32

+

+ ; 2nd

+ mov r10d, [rdx + 16]

+ mov r11, [rdx + 16]

+ shr r11, 32

+

+ mov ecx, [r8]

+ mov eax, ecx

+

+ gost_round2 1, 2

+ gost_round2 3, 4

+ gost_round2 5, 6

+ gost_round2 7, 7

+

+ gost_round2 6, 5

+ gost_round2 4, 3

+ gost_round2 2, 1

+ gost_round2 0, 7

+

+ gost_round2 6, 5

+ gost_round2 4, 3

+ gost_round2 2, 1

+ gost_round2 0, 7

+

+ gost_round2 6, 5

+ gost_round2 4, 3

+ gost_round2 2, 1

+ gost_round2 0, 0

+

+ shl r13, 32 ; combine

+ or r13, r14

+

+ shl r10, 32 ; combine

+ or r10, r11

+ret

+

+;///////////////////////////////////////////////////////////////////

+;// Crypting 1 block

+;///////////////////////////////////////////////////////////////////

+%macro gost_round1 2 ; 1 - pos1, 2 - pos2

+ ; 1-2 byte

+ add ecx, r13d ; add key

+ movzx edi, cl

+ movzx esi, ch

+ xor r14d, dword [r8 + 32 + 256*3*4 + rdi*4]

+ xor r14d, dword [r8 + 32 + 256*2*4 + rsi*4]

+ shr ecx, 16

+ ; 3-4 áàéò

+ movzx edi, cl

+ xor r14d, dword [r8 + 32 + 256*4 + rdi*4]

+ movzx esi, ch

+ xor r14d, dword [r8 + 32 + rsi*4]

+ mov edx, [r8 + %1*4] ; read key for second step

+

+ ; second step

+ ; 1-2 byte

+ add edx, r14d ; add key

+ movzx edi, dl

+ movzx esi, dh

+ xor r13d, dword [r8 + 32 + 256*3*4 + rdi*4]

+ xor r13d, dword [r8 + 32 + 256*2*4 + rsi*4]

+ shr edx, 16

+ ; 3-4 áàéò

+ movzx edi, dl

+ xor r13d, dword [r8 + 32 + 256*4 + rdi*4]

+ movzx esi, dh

+ xor r13d, dword [r8 + 32 + rsi*4]

+ mov ecx, [r8 + %2*4] ; read key

+%endmacro

+

+; input: r8 - &gost_kds rdx - &IN

+; returns: r13

+GostEncrypt1x64:

+ mov r13d, [rdx]

+ mov r14, [rdx]

+ shr r14, 32

+ mov ecx, [r8]

+

+ gost_round1 1, 2

+ gost_round1 3, 4

+ gost_round1 5, 6

+ gost_round1 7, 0

+

+ gost_round1 1, 2

+ gost_round1 3, 4

+ gost_round1 5, 6

+ gost_round1 7, 0

+

+ gost_round1 1, 2

+ gost_round1 3, 4

+ gost_round1 5, 6

+ gost_round1 7, 7

+

+ gost_round1 6, 5

+ gost_round1 4, 3

+ gost_round1 2, 1

+ gost_round1 0, 0

+

+ shl r13, 32 ; combine

+ or r13, r14

+ret

+

+; input: r8 - &gost_kds rdx - IN

+; returns: r13

+GostDecrypt1x64:

+ mov r13d, [rdx]

+ mov r14, [rdx]

+ shr r14, 32

+ mov ecx, [r8]

+

+ gost_round1 1, 2

+ gost_round1 3, 4

+ gost_round1 5, 6

+ gost_round1 7, 7

+

+ gost_round1 6, 5

+ gost_round1 4, 3

+ gost_round1 2, 1

+ gost_round1 0, 7

+

+ gost_round1 6, 5

+ gost_round1 4, 3

+ gost_round1 2, 1

+ gost_round1 0, 7

+

+ gost_round1 6, 5

+ gost_round1 4, 3

+ gost_round1 2, 1

+ gost_round1 0, 0

+

+ shl r13, 32 ; combine

+ or r13, r14

+ret

+

+global gost_encrypt_128_CBC_asm ; gost_encrypt_128_CBC_asm(uint64* out, uint64* in, gost_kds* kds, uint64 count);

+; rcx - &out

+; rdx - &in

+; r8 - &gost_kds

+; r9 - count

+gost_encrypt_128_CBC_asm:

+ SaveRegs ; Saving

+

+ sub rsp, 32

+ mov [rsp], rcx ; Save out addr

+ mov [rsp + 8], rdx ; Save in addr

+ mov [rsp + 16], r8 ; key addr

+

+.do:

+ mov [rsp + 24], r9 ; Save count

+ cmp r9, 2

+ jge .blk2

+ cmp r9, 1

+ jge .blk1

+ jmp .end

+

+; One 128 block encryption

+.blk1:

+ mov rdx, [rsp + 8] ; set in addr

+ call GostEncrypt1x64

+

+ mov rcx, [rsp] ; Restore out

+ mov rdx, [rsp + 8] ; restore in

+

+ mov [rcx], r13

+ mov rax, [rdx + 8]

+ xor rax, r13 ; CBC

+

+ add rcx, 8 ;next 8 bytes

+ mov [rcx], rax

+

+ mov rdx, rcx

+ call GostEncrypt1x64

+

+ mov rcx, [rsp] ; Restore out addr

+ mov rdx, [rsp+8] ; Restore in addr

+

+ mov [rcx + 8], r13

+

+ add rcx,16

+ mov [rsp], rcx

+

+ add rdx, 16

+ mov [rsp+8], rdx

+

+ mov r9, [rsp + 24]

+ dec r9

+

+ jmp .do

+

+.blk2:

+ mov rdx, [rsp + 8] ; set in addr

+ call GostEncrypt2x64

+

+ mov rcx, [rsp] ; Restore out

+ mov rdx, [rsp + 8] ; restore in

+

+ mov [rcx], r13

+

+ mov rax, [rdx + 8]

+ xor rax, r13 ; CBC

+

+ mov [rcx + 16], r10

+

+ mov rbx, [rdx + 24]

+ xor rbx, r10 ; CBC

+

+ mov [rcx + 8], rax

+ mov [rcx + 24], rbx

+

+ add rcx, 8 ;next 8 bytes

+

+ mov rdx, rcx

+ call GostEncrypt2x64

+

+ mov rcx, [rsp] ; Restore out addr

+ mov rdx, [rsp+8] ; Restore in addr

+

+ mov [rcx + 8], r13

+ mov [rcx + 24], r10

+

+ add rcx,32

+ mov [rsp], rcx

+

+ add rdx, 32

+ mov [rsp+8], rdx

+

+ mov r9, [rsp + 24]

+ sub r9, 2

+

+ jmp .do

+

+.end:

+ add rsp, 32 ; Load out addr

+ RestoreRegs ; Load

+ret

+

+global gost_decrypt_128_CBC_asm ; gost_decrypt_128_CBC_asm(uint64* out, uint64* in, const gost_kds* kds, uint64 count);

+; rcx - &out

+; rdx - &in

+; r8 - &gost_kds

+; r9 - count

+gost_decrypt_128_CBC_asm:

+ SaveRegs ; Saving

+

+ sub rsp, 32

+ mov [rsp], rdx ; Save out addr

+ mov [rsp+8], rcx ; Save in addr

+ mov [rsp+16], r8 ; key addr

+

+.do:

+ mov [rsp + 24], r9 ; Save count

+ cmp r9, 2

+ jge .blk2

+ cmp r9, 1

+ jge .blk1

+ jmp .end

+

+; One 128 block decryption

+.blk1:

+ add rdx, 8

+ call GostDecrypt1x64

+ mov rcx, [rsp] ; Restore out

+ mov rdx, [rsp + 8] ; Restore in

+ mov rax, [rdx]

+ xor rax, r13 ; CBC

+ mov [rcx + 8], rax

+

+ call GostDecrypt1x64

+

+ mov rcx, [rsp] ; Restore out addr

+ mov rdx, [rsp+8] ; Restore in addr

+

+ mov [rcx], r13

+

+ add rcx,16

+ mov [rsp], rcx

+

+ add rdx, 16

+ mov [rsp+8], rdx

+

+ mov r9, [rsp + 24]

+ dec r9

+

+ jmp .do

+

+.blk2:

+ add rdx, 8

+ call GostDecrypt2x64

+ mov rcx, [rsp] ; Restore out

+ mov rdx, [rsp + 8] ; Restore in

+

+ mov rax, [rdx]

+ xor rax, r13 ; CBC

+ mov [rcx + 8], rax

+

+ mov rbx, [rdx+16]

+ xor rbx, r10 ; CBC

+ mov [rcx + 24], rbx

+

+ call GostDecrypt2x64

+

+ mov rcx, [rsp] ; Restore out addr

+ mov rdx, [rsp+8] ; Restore in addr

+

+ mov [rcx], r13

+ mov [rcx+16], r10

+

+ add rcx,32

+ mov [rsp], rcx

+

+ add rdx,32

+ mov [rsp+8], rdx

+

+ mov r9, [rsp + 24]

+ sub r9, 2

+

+ jmp .do

+

+.end:

+ add rsp, 32 ; Load out addr

+ RestoreRegs ; Load

+ret