Skip to content

Commit

Permalink
Improve wasm performance
Browse files Browse the repository at this point in the history
  • Loading branch information
@jedisct1
jedisct1 committed Sep 11, 2024
1 parent be69e9c commit 564dca6
Showing 1 changed file with 58 additions and 39 deletions.
97 changes: 58 additions & 39 deletions src/common/softaes.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,8 @@
# define SOFTAES_STRIDE 16
#endif

uint32_t _aes_lut[256] __attribute__((visibility("hidden"))) = {
#ifdef FAVOR_PERFORMANCE
static const uint32_t LUT[1024] __attribute__((visibility("hidden"))) = {
0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6, 0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591,
0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56, 0x19fefee7, 0x62d7d7b5, 0xe6abab4d, 0x9a7676ec,
0x45caca8f, 0x9d82821f, 0x40c9c989, 0x877d7dfa, 0x15fafaef, 0xeb5959b2, 0xc947478e, 0x0bf0f0fb,
Expand Down Expand Up @@ -46,13 +47,7 @@ uint32_t _aes_lut[256] __attribute__((visibility("hidden"))) = {
0x38e1e1d9, 0x13f8f8eb, 0xb398982b, 0x33111122, 0xbb6969d2, 0x70d9d9a9, 0x898e8e07, 0xa7949433,
0xb69b9b2d, 0x221e1e3c, 0x92878715, 0x20e9e9c9, 0x49cece87, 0xff5555aa, 0x78282850, 0x7adfdfa5,
0x8f8c8c03, 0xf8a1a159, 0x80898909, 0x170d0d1a, 0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0,
0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e, 0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c
};

static const uint32_t* const LUT = _aes_lut;

#ifdef FAVOR_PERFORMANCE
uint32_t _aes_lut_s1[256] __attribute__((visibility("hidden"))) = {
0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e, 0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c,
0x6363c6a5, 0x7c7cf884, 0x7777ee99, 0x7b7bf68d, 0xf2f2ff0d, 0x6b6bd6bd, 0x6f6fdeb1, 0xc5c59154,
0x30306050, 0x01010203, 0x6767cea9, 0x2b2b567d, 0xfefee719, 0xd7d7b562, 0xabab4de6, 0x7676ec9a,
0xcaca8f45, 0x82821f9d, 0xc9c98940, 0x7d7dfa87, 0xfafaef15, 0x5959b2eb, 0x47478ec9, 0xf0f0fb0b,
Expand Down Expand Up @@ -84,10 +79,7 @@ uint32_t _aes_lut_s1[256] __attribute__((visibility("hidden"))) = {
0xe1e1d938, 0xf8f8eb13, 0x98982bb3, 0x11112233, 0x6969d2bb, 0xd9d9a970, 0x8e8e0789, 0x949433a7,
0x9b9b2db6, 0x1e1e3c22, 0x87871592, 0xe9e9c920, 0xcece8749, 0x5555aaff, 0x28285078, 0xdfdfa57a,
0x8c8c038f, 0xa1a159f8, 0x89890980, 0x0d0d1a17, 0xbfbf65da, 0xe6e6d731, 0x424284c6, 0x6868d0b8,
0x414182c3, 0x999929b0, 0x2d2d5a77, 0x0f0f1e11, 0xb0b07bcb, 0x5454a8fc, 0xbbbb6dd6, 0x16162c3a
};

uint32_t _aes_lut_s2[256] __attribute__((visibility("hidden"))) = {
0x414182c3, 0x999929b0, 0x2d2d5a77, 0x0f0f1e11, 0xb0b07bcb, 0x5454a8fc, 0xbbbb6dd6, 0x16162c3a,
0x63c6a563, 0x7cf8847c, 0x77ee9977, 0x7bf68d7b, 0xf2ff0df2, 0x6bd6bd6b, 0x6fdeb16f, 0xc59154c5,
0x30605030, 0x01020301, 0x67cea967, 0x2b567d2b, 0xfee719fe, 0xd7b562d7, 0xab4de6ab, 0x76ec9a76,
0xca8f45ca, 0x821f9d82, 0xc98940c9, 0x7dfa877d, 0xfaef15fa, 0x59b2eb59, 0x478ec947, 0xf0fb0bf0,
Expand Down Expand Up @@ -119,10 +111,7 @@ uint32_t _aes_lut_s2[256] __attribute__((visibility("hidden"))) = {
0xe1d938e1, 0xf8eb13f8, 0x982bb398, 0x11223311, 0x69d2bb69, 0xd9a970d9, 0x8e07898e, 0x9433a794,
0x9b2db69b, 0x1e3c221e, 0x87159287, 0xe9c920e9, 0xce8749ce, 0x55aaff55, 0x28507828, 0xdfa57adf,
0x8c038f8c, 0xa159f8a1, 0x89098089, 0x0d1a170d, 0xbf65dabf, 0xe6d731e6, 0x4284c642, 0x68d0b868,
0x4182c341, 0x9929b099, 0x2d5a772d, 0x0f1e110f, 0xb07bcbb0, 0x54a8fc54, 0xbb6dd6bb, 0x162c3a16
};

uint32_t _aes_lut_s3[256] __attribute__((visibility("hidden"))) = {
0x4182c341, 0x9929b099, 0x2d5a772d, 0x0f1e110f, 0xb07bcbb0, 0x54a8fc54, 0xbb6dd6bb, 0x162c3a16,
0xc6a56363, 0xf8847c7c, 0xee997777, 0xf68d7b7b, 0xff0df2f2, 0xd6bd6b6b, 0xdeb16f6f, 0x9154c5c5,
0x60503030, 0x02030101, 0xcea96767, 0x567d2b2b, 0xe719fefe, 0xb562d7d7, 0x4de6abab, 0xec9a7676,
0x8f45caca, 0x1f9d8282, 0x8940c9c9, 0xfa877d7d, 0xef15fafa, 0xb2eb5959, 0x8ec94747, 0xfb0bf0f0,
Expand Down Expand Up @@ -157,43 +146,73 @@ uint32_t _aes_lut_s3[256] __attribute__((visibility("hidden"))) = {
0x82c34141, 0x29b09999, 0x5a772d2d, 0x1e110f0f, 0x7bcbb0b0, 0xa8fc5454, 0x6dd6bbbb, 0x2c3a1616
};

static const uint32_t* const LUT_S1 = _aes_lut_s1;
static const uint32_t* const LUT_S2 = _aes_lut_s2;
static const uint32_t* const LUT_S3 = _aes_lut_s3;

#endif

#ifdef FAVOR_PERFORMANCE
static SoftAesBlock
_encrypt(const uint8_t ix0[4], const uint8_t ix1[4], const uint8_t ix2[4], const uint8_t ix3[4])
{
CRYPTO_ALIGN(64) SoftAesBlock out;

out.w0 = LUT[ix0[0]];
out.w1 = LUT[ix0[1]];
out.w2 = LUT[ix0[2]];
out.w3 = LUT[ix0[3]];
out.w0 = LUT[ix0[0] + 256 * 0];
out.w1 = LUT[ix0[1] + 256 * 0];
out.w2 = LUT[ix0[2] + 256 * 0];
out.w3 = LUT[ix0[3] + 256 * 0];

out.w0 ^= LUT_S1[ix1[0]];
out.w1 ^= LUT_S1[ix1[1]];
out.w2 ^= LUT_S1[ix1[2]];
out.w3 ^= LUT_S1[ix1[3]];
out.w0 ^= LUT[ix1[0] + 256 * 1];
out.w1 ^= LUT[ix1[1] + 256 * 1];
out.w2 ^= LUT[ix1[2] + 256 * 1];
out.w3 ^= LUT[ix1[3] + 256 * 1];

out.w0 ^= LUT_S2[ix2[0]];
out.w1 ^= LUT_S2[ix2[1]];
out.w2 ^= LUT_S2[ix2[2]];
out.w3 ^= LUT_S2[ix2[3]];
out.w0 ^= LUT[ix2[0] + 256 * 2];
out.w1 ^= LUT[ix2[1] + 256 * 2];
out.w2 ^= LUT[ix2[2] + 256 * 2];
out.w3 ^= LUT[ix2[3] + 256 * 2];

out.w0 ^= LUT_S3[ix3[0]];
out.w1 ^= LUT_S3[ix3[1]];
out.w2 ^= LUT_S3[ix3[2]];
out.w3 ^= LUT_S3[ix3[3]];
out.w0 ^= LUT[ix3[0] + 256 * 3];
out.w1 ^= LUT[ix3[1] + 256 * 3];
out.w2 ^= LUT[ix3[2] + 256 * 3];
out.w3 ^= LUT[ix3[3] + 256 * 3];

return out;
}

#else

uint32_t _aes_lut[256] __attribute__((visibility("hidden"))) = {
0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6, 0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591,
0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56, 0x19fefee7, 0x62d7d7b5, 0xe6abab4d, 0x9a7676ec,
0x45caca8f, 0x9d82821f, 0x40c9c989, 0x877d7dfa, 0x15fafaef, 0xeb5959b2, 0xc947478e, 0x0bf0f0fb,
0xecadad41, 0x67d4d4b3, 0xfda2a25f, 0xeaafaf45, 0xbf9c9c23, 0xf7a4a453, 0x967272e4, 0x5bc0c09b,
0xc2b7b775, 0x1cfdfde1, 0xae93933d, 0x6a26264c, 0x5a36366c, 0x413f3f7e, 0x02f7f7f5, 0x4fcccc83,
0x5c343468, 0xf4a5a551, 0x34e5e5d1, 0x08f1f1f9, 0x937171e2, 0x73d8d8ab, 0x53313162, 0x3f15152a,
0x0c040408, 0x52c7c795, 0x65232346, 0x5ec3c39d, 0x28181830, 0xa1969637, 0x0f05050a, 0xb59a9a2f,
0x0907070e, 0x36121224, 0x9b80801b, 0x3de2e2df, 0x26ebebcd, 0x6927274e, 0xcdb2b27f, 0x9f7575ea,
0x1b090912, 0x9e83831d, 0x742c2c58, 0x2e1a1a34, 0x2d1b1b36, 0xb26e6edc, 0xee5a5ab4, 0xfba0a05b,
0xf65252a4, 0x4d3b3b76, 0x61d6d6b7, 0xceb3b37d, 0x7b292952, 0x3ee3e3dd, 0x712f2f5e, 0x97848413,
0xf55353a6, 0x68d1d1b9, 0x00000000, 0x2cededc1, 0x60202040, 0x1ffcfce3, 0xc8b1b179, 0xed5b5bb6,
0xbe6a6ad4, 0x46cbcb8d, 0xd9bebe67, 0x4b393972, 0xde4a4a94, 0xd44c4c98, 0xe85858b0, 0x4acfcf85,
0x6bd0d0bb, 0x2aefefc5, 0xe5aaaa4f, 0x16fbfbed, 0xc5434386, 0xd74d4d9a, 0x55333366, 0x94858511,
0xcf45458a, 0x10f9f9e9, 0x06020204, 0x817f7ffe, 0xf05050a0, 0x443c3c78, 0xba9f9f25, 0xe3a8a84b,
0xf35151a2, 0xfea3a35d, 0xc0404080, 0x8a8f8f05, 0xad92923f, 0xbc9d9d21, 0x48383870, 0x04f5f5f1,
0xdfbcbc63, 0xc1b6b677, 0x75dadaaf, 0x63212142, 0x30101020, 0x1affffe5, 0x0ef3f3fd, 0x6dd2d2bf,
0x4ccdcd81, 0x140c0c18, 0x35131326, 0x2fececc3, 0xe15f5fbe, 0xa2979735, 0xcc444488, 0x3917172e,
0x57c4c493, 0xf2a7a755, 0x827e7efc, 0x473d3d7a, 0xac6464c8, 0xe75d5dba, 0x2b191932, 0x957373e6,
0xa06060c0, 0x98818119, 0xd14f4f9e, 0x7fdcdca3, 0x66222244, 0x7e2a2a54, 0xab90903b, 0x8388880b,
0xca46468c, 0x29eeeec7, 0xd3b8b86b, 0x3c141428, 0x79dedea7, 0xe25e5ebc, 0x1d0b0b16, 0x76dbdbad,
0x3be0e0db, 0x56323264, 0x4e3a3a74, 0x1e0a0a14, 0xdb494992, 0x0a06060c, 0x6c242448, 0xe45c5cb8,
0x5dc2c29f, 0x6ed3d3bd, 0xefacac43, 0xa66262c4, 0xa8919139, 0xa4959531, 0x37e4e4d3, 0x8b7979f2,
0x32e7e7d5, 0x43c8c88b, 0x5937376e, 0xb76d6dda, 0x8c8d8d01, 0x64d5d5b1, 0xd24e4e9c, 0xe0a9a949,
0xb46c6cd8, 0xfa5656ac, 0x07f4f4f3, 0x25eaeacf, 0xaf6565ca, 0x8e7a7af4, 0xe9aeae47, 0x18080810,
0xd5baba6f, 0x887878f0, 0x6f25254a, 0x722e2e5c, 0x241c1c38, 0xf1a6a657, 0xc7b4b473, 0x51c6c697,
0x23e8e8cb, 0x7cdddda1, 0x9c7474e8, 0x211f1f3e, 0xdd4b4b96, 0xdcbdbd61, 0x868b8b0d, 0x858a8a0f,
0x907070e0, 0x423e3e7c, 0xc4b5b571, 0xaa6666cc, 0xd8484890, 0x05030306, 0x01f6f6f7, 0x120e0e1c,
0xa36161c2, 0x5f35356a, 0xf95757ae, 0xd0b9b969, 0x91868617, 0x58c1c199, 0x271d1d3a, 0xb99e9e27,
0x38e1e1d9, 0x13f8f8eb, 0xb398982b, 0x33111122, 0xbb6969d2, 0x70d9d9a9, 0x898e8e07, 0xa7949433,
0xb69b9b2d, 0x221e1e3c, 0x92878715, 0x20e9e9c9, 0x49cece87, 0xff5555aa, 0x78282850, 0x7adfdfa5,
0x8f8c8c03, 0xf8a1a159, 0x80898909, 0x170d0d1a, 0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0,
0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e, 0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c
};

static const uint32_t* const LUT = _aes_lut;

static SoftAesBlock
_encrypt(const uint8_t ix0[4], const uint8_t ix1[4], const uint8_t ix2[4], const uint8_t ix3[4])
{
Expand Down

0 comments on commit 564dca6

Please sign in to comment.