static void
aesni_crypt_xts_block8(int rounds, const __m128i *key_schedule, __m128i *tweak,
const uint8_t *from, uint8_t *to, int do_encrypt)
{
__m128i tmptweak;
__m128i a, b, c, d, e, f, g, h;
__m128i tweaks[8];
__m128i tmp[8];
__m128i *top;
const __m128i *fromp;
tmptweak = *tweak;
/*
* unroll the loop. This lets gcc put values directly in the
* register and saves memory accesses.
*/
fromp = (const __m128i *)from;
#define PREPINP(v, pos) \
do { \
tweaks[(pos)] = tmptweak; \
(v) = _mm_loadu_si128(&fromp[pos]) ^ \
tmptweak; \
tmptweak = xts_crank_lfsr(tmptweak); \
} while (0)
PREPINP(a, 0);
PREPINP(b, 1);
PREPINP(c, 2);
PREPINP(d, 3);
PREPINP(e, 4);
PREPINP(f, 5);
PREPINP(g, 6);
PREPINP(h, 7);
*tweak = tmptweak;
if (do_encrypt)
aesni_enc8(rounds - 1, key_schedule, a, b, c, d, e, f, g, h,
tmp);
else
aesni_dec8(rounds - 1, key_schedule, a, b, c, d, e, f, g, h,
tmp);
top = (__m128i *)to;
_mm_storeu_si128(&top[0], tmp[0] ^ tweaks[0]);
_mm_storeu_si128(&top[1], tmp[1] ^ tweaks[1]);
_mm_storeu_si128(&top[2], tmp[2] ^ tweaks[2]);
_mm_storeu_si128(&top[3], tmp[3] ^ tweaks[3]);
_mm_storeu_si128(&top[4], tmp[4] ^ tweaks[4]);
_mm_storeu_si128(&top[5], tmp[5] ^ tweaks[5]);
_mm_storeu_si128(&top[6], tmp[6] ^ tweaks[6]);
_mm_storeu_si128(&top[7], tmp[7] ^ tweaks[7]);
}
static void
aesni_crypt_xts_block8(int rounds, const void *key_schedule, __m128i *tweak,
const __m128i *from, __m128i *to, int do_encrypt)
{
__m128i tmptweak;
__m128i a, b, c, d, e, f, g, h;
__m128i tweaks[8];
__m128i tmp[8];
tmptweak = *tweak;
/*
* unroll the loop. This lets gcc put values directly in the
* register and saves memory accesses.
*/
#define PREPINP(v, pos) \
do { \
tweaks[(pos)] = tmptweak; \
(v) = from[(pos)] ^ tmptweak; \
tmptweak = xts_crank_lfsr(tmptweak); \
} while (0)
PREPINP(a, 0);
PREPINP(b, 1);
PREPINP(c, 2);
PREPINP(d, 3);
PREPINP(e, 4);
PREPINP(f, 5);
PREPINP(g, 6);
PREPINP(h, 7);
*tweak = tmptweak;
if (do_encrypt)
aesni_enc8(rounds - 1, key_schedule, a, b, c, d, e, f, g, h,
tmp);
else
aesni_dec8(rounds - 1, key_schedule, a, b, c, d, e, f, g, h,
tmp);
to[0] = tmp[0] ^ tweaks[0];
to[1] = tmp[1] ^ tweaks[1];
to[2] = tmp[2] ^ tweaks[2];
to[3] = tmp[3] ^ tweaks[3];
to[4] = tmp[4] ^ tweaks[4];
to[5] = tmp[5] ^ tweaks[5];
to[6] = tmp[6] ^ tweaks[6];
to[7] = tmp[7] ^ tweaks[7];
}
