forked from wolfSSL/wolfssl
remove unique aes last round Te table, pre fetch Te tables during software aes encrypt
This commit is contained in:
@ -507,7 +507,7 @@ static const word32 rcon[] = {
|
||||
/* for 128-bit blocks, Rijndael never uses more than 10 rcon values */
|
||||
};
|
||||
|
||||
static const word32 Te[5][256] = {
|
||||
static const word32 Te[4][256] = {
|
||||
{
|
||||
0xc66363a5U, 0xf87c7c84U, 0xee777799U, 0xf67b7b8dU,
|
||||
0xfff2f20dU, 0xd66b6bbdU, 0xde6f6fb1U, 0x91c5c554U,
|
||||
@ -771,72 +771,6 @@ static const word32 Te[5][256] = {
|
||||
0xbfbfda65U, 0xe6e631d7U, 0x4242c684U, 0x6868b8d0U,
|
||||
0x4141c382U, 0x9999b029U, 0x2d2d775aU, 0x0f0f111eU,
|
||||
0xb0b0cb7bU, 0x5454fca8U, 0xbbbbd66dU, 0x16163a2cU,
|
||||
},
|
||||
{
|
||||
0x63636363U, 0x7c7c7c7cU, 0x77777777U, 0x7b7b7b7bU,
|
||||
0xf2f2f2f2U, 0x6b6b6b6bU, 0x6f6f6f6fU, 0xc5c5c5c5U,
|
||||
0x30303030U, 0x01010101U, 0x67676767U, 0x2b2b2b2bU,
|
||||
0xfefefefeU, 0xd7d7d7d7U, 0xababababU, 0x76767676U,
|
||||
0xcacacacaU, 0x82828282U, 0xc9c9c9c9U, 0x7d7d7d7dU,
|
||||
0xfafafafaU, 0x59595959U, 0x47474747U, 0xf0f0f0f0U,
|
||||
0xadadadadU, 0xd4d4d4d4U, 0xa2a2a2a2U, 0xafafafafU,
|
||||
0x9c9c9c9cU, 0xa4a4a4a4U, 0x72727272U, 0xc0c0c0c0U,
|
||||
0xb7b7b7b7U, 0xfdfdfdfdU, 0x93939393U, 0x26262626U,
|
||||
0x36363636U, 0x3f3f3f3fU, 0xf7f7f7f7U, 0xccccccccU,
|
||||
0x34343434U, 0xa5a5a5a5U, 0xe5e5e5e5U, 0xf1f1f1f1U,
|
||||
0x71717171U, 0xd8d8d8d8U, 0x31313131U, 0x15151515U,
|
||||
0x04040404U, 0xc7c7c7c7U, 0x23232323U, 0xc3c3c3c3U,
|
||||
0x18181818U, 0x96969696U, 0x05050505U, 0x9a9a9a9aU,
|
||||
0x07070707U, 0x12121212U, 0x80808080U, 0xe2e2e2e2U,
|
||||
0xebebebebU, 0x27272727U, 0xb2b2b2b2U, 0x75757575U,
|
||||
0x09090909U, 0x83838383U, 0x2c2c2c2cU, 0x1a1a1a1aU,
|
||||
0x1b1b1b1bU, 0x6e6e6e6eU, 0x5a5a5a5aU, 0xa0a0a0a0U,
|
||||
0x52525252U, 0x3b3b3b3bU, 0xd6d6d6d6U, 0xb3b3b3b3U,
|
||||
0x29292929U, 0xe3e3e3e3U, 0x2f2f2f2fU, 0x84848484U,
|
||||
0x53535353U, 0xd1d1d1d1U, 0x00000000U, 0xededededU,
|
||||
0x20202020U, 0xfcfcfcfcU, 0xb1b1b1b1U, 0x5b5b5b5bU,
|
||||
0x6a6a6a6aU, 0xcbcbcbcbU, 0xbebebebeU, 0x39393939U,
|
||||
0x4a4a4a4aU, 0x4c4c4c4cU, 0x58585858U, 0xcfcfcfcfU,
|
||||
0xd0d0d0d0U, 0xefefefefU, 0xaaaaaaaaU, 0xfbfbfbfbU,
|
||||
0x43434343U, 0x4d4d4d4dU, 0x33333333U, 0x85858585U,
|
||||
0x45454545U, 0xf9f9f9f9U, 0x02020202U, 0x7f7f7f7fU,
|
||||
0x50505050U, 0x3c3c3c3cU, 0x9f9f9f9fU, 0xa8a8a8a8U,
|
||||
0x51515151U, 0xa3a3a3a3U, 0x40404040U, 0x8f8f8f8fU,
|
||||
0x92929292U, 0x9d9d9d9dU, 0x38383838U, 0xf5f5f5f5U,
|
||||
0xbcbcbcbcU, 0xb6b6b6b6U, 0xdadadadaU, 0x21212121U,
|
||||
0x10101010U, 0xffffffffU, 0xf3f3f3f3U, 0xd2d2d2d2U,
|
||||
0xcdcdcdcdU, 0x0c0c0c0cU, 0x13131313U, 0xececececU,
|
||||
0x5f5f5f5fU, 0x97979797U, 0x44444444U, 0x17171717U,
|
||||
0xc4c4c4c4U, 0xa7a7a7a7U, 0x7e7e7e7eU, 0x3d3d3d3dU,
|
||||
0x64646464U, 0x5d5d5d5dU, 0x19191919U, 0x73737373U,
|
||||
0x60606060U, 0x81818181U, 0x4f4f4f4fU, 0xdcdcdcdcU,
|
||||
0x22222222U, 0x2a2a2a2aU, 0x90909090U, 0x88888888U,
|
||||
0x46464646U, 0xeeeeeeeeU, 0xb8b8b8b8U, 0x14141414U,
|
||||
0xdedededeU, 0x5e5e5e5eU, 0x0b0b0b0bU, 0xdbdbdbdbU,
|
||||
0xe0e0e0e0U, 0x32323232U, 0x3a3a3a3aU, 0x0a0a0a0aU,
|
||||
0x49494949U, 0x06060606U, 0x24242424U, 0x5c5c5c5cU,
|
||||
0xc2c2c2c2U, 0xd3d3d3d3U, 0xacacacacU, 0x62626262U,
|
||||
0x91919191U, 0x95959595U, 0xe4e4e4e4U, 0x79797979U,
|
||||
0xe7e7e7e7U, 0xc8c8c8c8U, 0x37373737U, 0x6d6d6d6dU,
|
||||
0x8d8d8d8dU, 0xd5d5d5d5U, 0x4e4e4e4eU, 0xa9a9a9a9U,
|
||||
0x6c6c6c6cU, 0x56565656U, 0xf4f4f4f4U, 0xeaeaeaeaU,
|
||||
0x65656565U, 0x7a7a7a7aU, 0xaeaeaeaeU, 0x08080808U,
|
||||
0xbabababaU, 0x78787878U, 0x25252525U, 0x2e2e2e2eU,
|
||||
0x1c1c1c1cU, 0xa6a6a6a6U, 0xb4b4b4b4U, 0xc6c6c6c6U,
|
||||
0xe8e8e8e8U, 0xddddddddU, 0x74747474U, 0x1f1f1f1fU,
|
||||
0x4b4b4b4bU, 0xbdbdbdbdU, 0x8b8b8b8bU, 0x8a8a8a8aU,
|
||||
0x70707070U, 0x3e3e3e3eU, 0xb5b5b5b5U, 0x66666666U,
|
||||
0x48484848U, 0x03030303U, 0xf6f6f6f6U, 0x0e0e0e0eU,
|
||||
0x61616161U, 0x35353535U, 0x57575757U, 0xb9b9b9b9U,
|
||||
0x86868686U, 0xc1c1c1c1U, 0x1d1d1d1dU, 0x9e9e9e9eU,
|
||||
0xe1e1e1e1U, 0xf8f8f8f8U, 0x98989898U, 0x11111111U,
|
||||
0x69696969U, 0xd9d9d9d9U, 0x8e8e8e8eU, 0x94949494U,
|
||||
0x9b9b9b9bU, 0x1e1e1e1eU, 0x87878787U, 0xe9e9e9e9U,
|
||||
0xcecececeU, 0x55555555U, 0x28282828U, 0xdfdfdfdfU,
|
||||
0x8c8c8c8cU, 0xa1a1a1a1U, 0x89898989U, 0x0d0d0d0dU,
|
||||
0xbfbfbfbfU, 0xe6e6e6e6U, 0x42424242U, 0x68686868U,
|
||||
0x41414141U, 0x99999999U, 0x2d2d2d2dU, 0x0f0f0f0fU,
|
||||
0xb0b0b0b0U, 0x54545454U, 0xbbbbbbbbU, 0x16161616U,
|
||||
}
|
||||
};
|
||||
|
||||
@ -1337,6 +1271,32 @@ static int AES_set_decrypt_key(const unsigned char* userKey, const int bits,
|
||||
#if defined(HAVE_AES_CBC) || defined(WOLFSSL_AES_DIRECT) ||\
|
||||
defined(HAVE_AESGCM)
|
||||
|
||||
|
||||
#ifndef WC_CACHE_LINE_SZ
|
||||
#if defined(__x86_64__) || defined(_M_X64) || (__ILP32__ >= 1)
|
||||
#define WC_CACHE_LINE_SZ 64
|
||||
#else
|
||||
/* default cache line size */
|
||||
#define WC_CACHE_LINE_SZ 32
|
||||
#endif
|
||||
#endif
|
||||
|
||||
|
||||
/* load 4 Te Tables into cache by cache line stride */
|
||||
static INLINE word32 PreFetchTe(void)
|
||||
{
|
||||
word32 x = 0;
|
||||
int i,j;
|
||||
|
||||
for (i = 0; i < 4; i++) {
|
||||
for (j = 0; j < 256; j += WC_CACHE_LINE_SZ) {
|
||||
x &= Te[i][j];
|
||||
}
|
||||
}
|
||||
return x;
|
||||
}
|
||||
|
||||
|
||||
static void wc_AesEncrypt(Aes* aes, const byte* inBlock, byte* outBlock)
|
||||
{
|
||||
word32 s0, s1, s2, s3;
|
||||
@ -1411,6 +1371,8 @@ static void wc_AesEncrypt(Aes* aes, const byte* inBlock, byte* outBlock)
|
||||
s2 ^= rk[2];
|
||||
s3 ^= rk[3];
|
||||
|
||||
s0 |= PreFetchTe();
|
||||
|
||||
/*
|
||||
* Nr - 1 full rounds:
|
||||
*/
|
||||
@ -1478,28 +1440,28 @@ static void wc_AesEncrypt(Aes* aes, const byte* inBlock, byte* outBlock)
|
||||
*/
|
||||
|
||||
s0 =
|
||||
(Te[4][GETBYTE(t0, 3)] & 0xff000000) ^
|
||||
(Te[4][GETBYTE(t1, 2)] & 0x00ff0000) ^
|
||||
(Te[4][GETBYTE(t2, 1)] & 0x0000ff00) ^
|
||||
(Te[4][GETBYTE(t3, 0)] & 0x000000ff) ^
|
||||
(Te[2][GETBYTE(t0, 3)] & 0xff000000) ^
|
||||
(Te[3][GETBYTE(t1, 2)] & 0x00ff0000) ^
|
||||
(Te[0][GETBYTE(t2, 1)] & 0x0000ff00) ^
|
||||
(Te[1][GETBYTE(t3, 0)] & 0x000000ff) ^
|
||||
rk[0];
|
||||
s1 =
|
||||
(Te[4][GETBYTE(t1, 3)] & 0xff000000) ^
|
||||
(Te[4][GETBYTE(t2, 2)] & 0x00ff0000) ^
|
||||
(Te[4][GETBYTE(t3, 1)] & 0x0000ff00) ^
|
||||
(Te[4][GETBYTE(t0, 0)] & 0x000000ff) ^
|
||||
(Te[2][GETBYTE(t1, 3)] & 0xff000000) ^
|
||||
(Te[3][GETBYTE(t2, 2)] & 0x00ff0000) ^
|
||||
(Te[0][GETBYTE(t3, 1)] & 0x0000ff00) ^
|
||||
(Te[1][GETBYTE(t0, 0)] & 0x000000ff) ^
|
||||
rk[1];
|
||||
s2 =
|
||||
(Te[4][GETBYTE(t2, 3)] & 0xff000000) ^
|
||||
(Te[4][GETBYTE(t3, 2)] & 0x00ff0000) ^
|
||||
(Te[4][GETBYTE(t0, 1)] & 0x0000ff00) ^
|
||||
(Te[4][GETBYTE(t1, 0)] & 0x000000ff) ^
|
||||
(Te[2][GETBYTE(t2, 3)] & 0xff000000) ^
|
||||
(Te[3][GETBYTE(t3, 2)] & 0x00ff0000) ^
|
||||
(Te[0][GETBYTE(t0, 1)] & 0x0000ff00) ^
|
||||
(Te[1][GETBYTE(t1, 0)] & 0x000000ff) ^
|
||||
rk[2];
|
||||
s3 =
|
||||
(Te[4][GETBYTE(t3, 3)] & 0xff000000) ^
|
||||
(Te[4][GETBYTE(t0, 2)] & 0x00ff0000) ^
|
||||
(Te[4][GETBYTE(t1, 1)] & 0x0000ff00) ^
|
||||
(Te[4][GETBYTE(t2, 0)] & 0x000000ff) ^
|
||||
(Te[2][GETBYTE(t3, 3)] & 0xff000000) ^
|
||||
(Te[3][GETBYTE(t0, 2)] & 0x00ff0000) ^
|
||||
(Te[0][GETBYTE(t1, 1)] & 0x0000ff00) ^
|
||||
(Te[1][GETBYTE(t2, 0)] & 0x000000ff) ^
|
||||
rk[3];
|
||||
|
||||
/* write out */
|
||||
@ -1514,6 +1476,7 @@ static void wc_AesEncrypt(Aes* aes, const byte* inBlock, byte* outBlock)
|
||||
XMEMCPY(outBlock + sizeof(s0), &s1, sizeof(s1));
|
||||
XMEMCPY(outBlock + 2 * sizeof(s0), &s2, sizeof(s2));
|
||||
XMEMCPY(outBlock + 3 * sizeof(s0), &s3, sizeof(s3));
|
||||
|
||||
}
|
||||
#endif /* HAVE_AES_CBC || WOLFSSL_AES_DIRECT || HAVE_AESGCM */
|
||||
|
||||
@ -2055,10 +2018,10 @@ static void wc_AesDecrypt(Aes* aes, const byte* inBlock, byte* outBlock)
|
||||
{
|
||||
temp = rk[3];
|
||||
rk[4] = rk[0] ^
|
||||
(Te[4][GETBYTE(temp, 2)] & 0xff000000) ^
|
||||
(Te[4][GETBYTE(temp, 1)] & 0x00ff0000) ^
|
||||
(Te[4][GETBYTE(temp, 0)] & 0x0000ff00) ^
|
||||
(Te[4][GETBYTE(temp, 3)] & 0x000000ff) ^
|
||||
(Te[2][GETBYTE(temp, 2)] & 0xff000000) ^
|
||||
(Te[3][GETBYTE(temp, 1)] & 0x00ff0000) ^
|
||||
(Te[0][GETBYTE(temp, 0)] & 0x0000ff00) ^
|
||||
(Te[1][GETBYTE(temp, 3)] & 0x000000ff) ^
|
||||
rcon[i];
|
||||
rk[5] = rk[1] ^ rk[4];
|
||||
rk[6] = rk[2] ^ rk[5];
|
||||
@ -2077,10 +2040,10 @@ static void wc_AesDecrypt(Aes* aes, const byte* inBlock, byte* outBlock)
|
||||
{
|
||||
temp = rk[ 5];
|
||||
rk[ 6] = rk[ 0] ^
|
||||
(Te[4][GETBYTE(temp, 2)] & 0xff000000) ^
|
||||
(Te[4][GETBYTE(temp, 1)] & 0x00ff0000) ^
|
||||
(Te[4][GETBYTE(temp, 0)] & 0x0000ff00) ^
|
||||
(Te[4][GETBYTE(temp, 3)] & 0x000000ff) ^
|
||||
(Te[2][GETBYTE(temp, 2)] & 0xff000000) ^
|
||||
(Te[3][GETBYTE(temp, 1)] & 0x00ff0000) ^
|
||||
(Te[0][GETBYTE(temp, 0)] & 0x0000ff00) ^
|
||||
(Te[1][GETBYTE(temp, 3)] & 0x000000ff) ^
|
||||
rcon[i];
|
||||
rk[ 7] = rk[ 1] ^ rk[ 6];
|
||||
rk[ 8] = rk[ 2] ^ rk[ 7];
|
||||
@ -2100,10 +2063,10 @@ static void wc_AesDecrypt(Aes* aes, const byte* inBlock, byte* outBlock)
|
||||
{
|
||||
temp = rk[ 7];
|
||||
rk[ 8] = rk[ 0] ^
|
||||
(Te[4][GETBYTE(temp, 2)] & 0xff000000) ^
|
||||
(Te[4][GETBYTE(temp, 1)] & 0x00ff0000) ^
|
||||
(Te[4][GETBYTE(temp, 0)] & 0x0000ff00) ^
|
||||
(Te[4][GETBYTE(temp, 3)] & 0x000000ff) ^
|
||||
(Te[2][GETBYTE(temp, 2)] & 0xff000000) ^
|
||||
(Te[3][GETBYTE(temp, 1)] & 0x00ff0000) ^
|
||||
(Te[0][GETBYTE(temp, 0)] & 0x0000ff00) ^
|
||||
(Te[1][GETBYTE(temp, 3)] & 0x000000ff) ^
|
||||
rcon[i];
|
||||
rk[ 9] = rk[ 1] ^ rk[ 8];
|
||||
rk[10] = rk[ 2] ^ rk[ 9];
|
||||
@ -2112,10 +2075,10 @@ static void wc_AesDecrypt(Aes* aes, const byte* inBlock, byte* outBlock)
|
||||
break;
|
||||
temp = rk[11];
|
||||
rk[12] = rk[ 4] ^
|
||||
(Te[4][GETBYTE(temp, 3)] & 0xff000000) ^
|
||||
(Te[4][GETBYTE(temp, 2)] & 0x00ff0000) ^
|
||||
(Te[4][GETBYTE(temp, 1)] & 0x0000ff00) ^
|
||||
(Te[4][GETBYTE(temp, 0)] & 0x000000ff);
|
||||
(Te[2][GETBYTE(temp, 3)] & 0xff000000) ^
|
||||
(Te[3][GETBYTE(temp, 2)] & 0x00ff0000) ^
|
||||
(Te[0][GETBYTE(temp, 1)] & 0x0000ff00) ^
|
||||
(Te[1][GETBYTE(temp, 0)] & 0x000000ff);
|
||||
rk[13] = rk[ 5] ^ rk[12];
|
||||
rk[14] = rk[ 6] ^ rk[13];
|
||||
rk[15] = rk[ 7] ^ rk[14];
|
||||
@ -2147,25 +2110,25 @@ static void wc_AesDecrypt(Aes* aes, const byte* inBlock, byte* outBlock)
|
||||
for (i = 1; i < aes->rounds; i++) {
|
||||
rk += 4;
|
||||
rk[0] =
|
||||
Td[0][Te[4][GETBYTE(rk[0], 3)] & 0xff] ^
|
||||
Td[1][Te[4][GETBYTE(rk[0], 2)] & 0xff] ^
|
||||
Td[2][Te[4][GETBYTE(rk[0], 1)] & 0xff] ^
|
||||
Td[3][Te[4][GETBYTE(rk[0], 0)] & 0xff];
|
||||
Td[0][Te[1][GETBYTE(rk[0], 3)] & 0xff] ^
|
||||
Td[1][Te[1][GETBYTE(rk[0], 2)] & 0xff] ^
|
||||
Td[2][Te[1][GETBYTE(rk[0], 1)] & 0xff] ^
|
||||
Td[3][Te[1][GETBYTE(rk[0], 0)] & 0xff];
|
||||
rk[1] =
|
||||
Td[0][Te[4][GETBYTE(rk[1], 3)] & 0xff] ^
|
||||
Td[1][Te[4][GETBYTE(rk[1], 2)] & 0xff] ^
|
||||
Td[2][Te[4][GETBYTE(rk[1], 1)] & 0xff] ^
|
||||
Td[3][Te[4][GETBYTE(rk[1], 0)] & 0xff];
|
||||
Td[0][Te[1][GETBYTE(rk[1], 3)] & 0xff] ^
|
||||
Td[1][Te[1][GETBYTE(rk[1], 2)] & 0xff] ^
|
||||
Td[2][Te[1][GETBYTE(rk[1], 1)] & 0xff] ^
|
||||
Td[3][Te[1][GETBYTE(rk[1], 0)] & 0xff];
|
||||
rk[2] =
|
||||
Td[0][Te[4][GETBYTE(rk[2], 3)] & 0xff] ^
|
||||
Td[1][Te[4][GETBYTE(rk[2], 2)] & 0xff] ^
|
||||
Td[2][Te[4][GETBYTE(rk[2], 1)] & 0xff] ^
|
||||
Td[3][Te[4][GETBYTE(rk[2], 0)] & 0xff];
|
||||
Td[0][Te[1][GETBYTE(rk[2], 3)] & 0xff] ^
|
||||
Td[1][Te[1][GETBYTE(rk[2], 2)] & 0xff] ^
|
||||
Td[2][Te[1][GETBYTE(rk[2], 1)] & 0xff] ^
|
||||
Td[3][Te[1][GETBYTE(rk[2], 0)] & 0xff];
|
||||
rk[3] =
|
||||
Td[0][Te[4][GETBYTE(rk[3], 3)] & 0xff] ^
|
||||
Td[1][Te[4][GETBYTE(rk[3], 2)] & 0xff] ^
|
||||
Td[2][Te[4][GETBYTE(rk[3], 1)] & 0xff] ^
|
||||
Td[3][Te[4][GETBYTE(rk[3], 0)] & 0xff];
|
||||
Td[0][Te[1][GETBYTE(rk[3], 3)] & 0xff] ^
|
||||
Td[1][Te[1][GETBYTE(rk[3], 2)] & 0xff] ^
|
||||
Td[2][Te[1][GETBYTE(rk[3], 1)] & 0xff] ^
|
||||
Td[3][Te[1][GETBYTE(rk[3], 0)] & 0xff];
|
||||
}
|
||||
}
|
||||
#endif /* HAVE_AES_DECRYPT */
|
||||
|
Reference in New Issue
Block a user