1 
2 #define BLAKE2_USE_SSSE3
3 #define BLAKE2_USE_SSE41
4 
5 #include <stdint.h>
6 #include <string.h>
7 
8 #include "blake2.h"
9 #include "private/common.h"
10 #include "private/sse2_64_32.h"
11 
12 #if defined(HAVE_EMMINTRIN_H) && defined(HAVE_TMMINTRIN_H) && \
13     defined(HAVE_SMMINTRIN_H)
14 
15 # ifdef __GNUC__
16 #  pragma GCC target("sse2")
17 #  pragma GCC target("ssse3")
18 #  pragma GCC target("sse4.1")
19 # endif
20 
21 # include <emmintrin.h>
22 # include <smmintrin.h>
23 # include <tmmintrin.h>
24 
25 # include "blake2b-compress-sse41.h"
26 
27 CRYPTO_ALIGN(64)
28 static const uint64_t blake2b_IV[8] = {
29     0x6a09e667f3bcc908ULL, 0xbb67ae8584caa73bULL, 0x3c6ef372fe94f82bULL,
30     0xa54ff53a5f1d36f1ULL, 0x510e527fade682d1ULL, 0x9b05688c2b3e6c1fULL,
31     0x1f83d9abfb41bd6bULL, 0x5be0cd19137e2179ULL
32 };
33 
34 int
35 blake2b_compress_sse41(blake2b_state *S,
36                        const uint8_t  block[BLAKE2B_BLOCKBYTES])
37 {
38     __m128i       row1l, row1h;
39     __m128i       row2l, row2h;
40     __m128i       row3l, row3h;
41     __m128i       row4l, row4h;
42     __m128i       b0, b1;
43     __m128i       t0, t1;
44     const __m128i r16 =
45         _mm_setr_epi8(2, 3, 4, 5, 6, 7, 0, 1, 10, 11, 12, 13, 14, 15, 8, 9);
46     const __m128i r24 =
47         _mm_setr_epi8(3, 4, 5, 6, 7, 0, 1, 2, 11, 12, 13, 14, 15, 8, 9, 10);
48     const __m128i m0 = LOADU(block + 00);
49     const __m128i m1 = LOADU(block + 16);
50     const __m128i m2 = LOADU(block + 32);
51     const __m128i m3 = LOADU(block + 48);
52     const __m128i m4 = LOADU(block + 64);
53     const __m128i m5 = LOADU(block + 80);
54     const __m128i m6 = LOADU(block + 96);
55     const __m128i m7 = LOADU(block + 112);
56     row1l            = LOADU(&S->h[0]);
57     row1h            = LOADU(&S->h[2]);
58     row2l            = LOADU(&S->h[4]);
59     row2h            = LOADU(&S->h[6]);
60     row3l            = LOADU(&blake2b_IV[0]);
61     row3h            = LOADU(&blake2b_IV[2]);
62     row4l            = _mm_xor_si128(LOADU(&blake2b_IV[4]), LOADU(&S->t[0]));
63     row4h            = _mm_xor_si128(LOADU(&blake2b_IV[6]), LOADU(&S->f[0]));
64     ROUND(0);
65     ROUND(1);
66     ROUND(2);
67     ROUND(3);
68     ROUND(4);
69     ROUND(5);
70     ROUND(6);
71     ROUND(7);
72     ROUND(8);
73     ROUND(9);
74     ROUND(10);
75     ROUND(11);
76     row1l = _mm_xor_si128(row3l, row1l);
77     row1h = _mm_xor_si128(row3h, row1h);
78     STOREU(&S->h[0], _mm_xor_si128(LOADU(&S->h[0]), row1l));
79     STOREU(&S->h[2], _mm_xor_si128(LOADU(&S->h[2]), row1h));
80     row2l = _mm_xor_si128(row4l, row2l);
81     row2h = _mm_xor_si128(row4h, row2h);
82     STOREU(&S->h[4], _mm_xor_si128(LOADU(&S->h[4]), row2l));
83     STOREU(&S->h[6], _mm_xor_si128(LOADU(&S->h[6]), row2h));
84     return 0;
85 }
86 
87 #endif
88