Crypto++  8.2
Free C++ class library of cryptographic schemes
donna_sse.cpp
1 // donna_sse.cpp - written and placed in public domain by Jeffrey Walton
2 // This is a integration of Andrew Moon's public domain code.
3 // Also see https://github.com/floodyberry/curve25519-donna.
4 
5 // This is a integration of Andrew Moon's public domain code. The port was
6 // clean, but it has one potential problem. The original code is C and relies
7 // upon unions. Accessing the inactive union member is undefined behavior in
8 // C++. That means copying the array into packedelem8.u is OK; but then using
9 // packedelem8.v in a calcualtion is UB. Fortunately most (all?) compilers
10 // take pity on C++ developers and compile the code. We will have to keep an
11 // eye on things or rewrite significant portions of this code.
12 
13 // If needed, see Moon's commit "Go back to ignoring 256th bit [sic]",
14 // https://github.com/floodyberry/curve25519-donna/commit/57a683d18721a658
15 
16 #include "pch.h"
17 
18 #include "config.h"
19 #include "donna.h"
20 #include "secblock.h"
21 #include "misc.h"
22 
23 // The data is aligned, but Clang issues warning based on type
24 // and not the actual alignment of the variable and data.
25 #if CRYPTOPP_GCC_DIAGNOSTIC_AVAILABLE
26 # pragma GCC diagnostic ignored "-Wcast-align"
27 # pragma GCC diagnostic ignored "-Wunused-function"
28 #endif
29 
30 // Squash MS LNK4221 and libtool warnings
31 extern const char DONNA_SSE_FNAME[] = __FILE__;
32 
33 #if (CRYPTOPP_CURVE25519_SSE2)
34 
35 #include "donna_sse.h"
36 
37 ANONYMOUS_NAMESPACE_BEGIN
38 
39 using CryptoPP::byte;
40 using CryptoPP::word32;
41 using CryptoPP::sword32;
42 using CryptoPP::word64;
43 using CryptoPP::sword64;
44 using CryptoPP::GetBlock;
46 
47 // Bring in all the symbols from the SSE header
48 using namespace CryptoPP::Donna::ArchSSE;
49 
50 /* Copy a bignum to another: out = in */
51 inline void
52 curve25519_copy(bignum25519 out, const bignum25519 in) {
53  xmmi x0,x1,x2;
54  x0 = _mm_load_si128((xmmi*)in + 0);
55  x1 = _mm_load_si128((xmmi*)in + 1);
56  x2 = _mm_load_si128((xmmi*)in + 2);
57  _mm_store_si128((xmmi*)out + 0, x0);
58  _mm_store_si128((xmmi*)out + 1, x1);
59  _mm_store_si128((xmmi*)out + 2, x2);
60 }
61 
62 /* Take a little-endian, 32-byte number and expand it into polynomial form */
63 inline void
64 curve25519_expand(bignum25519 out, const byte in[32]) {
65  word32 x0,x1,x2,x3,x4,x5,x6,x7;
66 
67  x0 = *(word32 *)(in + 0);
68  x1 = *(word32 *)(in + 4);
69  x2 = *(word32 *)(in + 8);
70  x3 = *(word32 *)(in + 12);
71  x4 = *(word32 *)(in + 16);
72  x5 = *(word32 *)(in + 20);
73  x6 = *(word32 *)(in + 24);
74  x7 = *(word32 *)(in + 28);
75 
76  out[0] = ( x0 ) & reduce_mask_26;
77  out[1] = ((((word64)x1 << 32) | x0) >> 26) & reduce_mask_25;
78  out[2] = ((((word64)x2 << 32) | x1) >> 19) & reduce_mask_26;
79  out[3] = ((((word64)x3 << 32) | x2) >> 13) & reduce_mask_25;
80  out[4] = (( x3) >> 6) & reduce_mask_26;
81  out[5] = ( x4 ) & reduce_mask_25;
82  out[6] = ((((word64)x5 << 32) | x4) >> 25) & reduce_mask_26;
83  out[7] = ((((word64)x6 << 32) | x5) >> 19) & reduce_mask_25;
84  out[8] = ((((word64)x7 << 32) | x6) >> 12) & reduce_mask_26;
85  out[9] = (( x7) >> 6) & reduce_mask_25; /* ignore the top bit */
86 
87  out[10] = 0;
88  out[11] = 0;
89 }
90 
91 /* Take a fully reduced polynomial form number and contract it into a
92  * little-endian, 32-byte array
93  */
94 inline void
95 curve25519_contract(byte out[32], const bignum25519 in) {
96  ALIGN(16) bignum25519 f;
97 
98  curve25519_copy(f, in);
99 
100  #define carry_pass() \
101  f[1] += f[0] >> 26; f[0] &= reduce_mask_26; \
102  f[2] += f[1] >> 25; f[1] &= reduce_mask_25; \
103  f[3] += f[2] >> 26; f[2] &= reduce_mask_26; \
104  f[4] += f[3] >> 25; f[3] &= reduce_mask_25; \
105  f[5] += f[4] >> 26; f[4] &= reduce_mask_26; \
106  f[6] += f[5] >> 25; f[5] &= reduce_mask_25; \
107  f[7] += f[6] >> 26; f[6] &= reduce_mask_26; \
108  f[8] += f[7] >> 25; f[7] &= reduce_mask_25; \
109  f[9] += f[8] >> 26; f[8] &= reduce_mask_26;
110 
111  #define carry_pass_full() \
112  carry_pass() \
113  f[0] += 19 * (f[9] >> 25); f[9] &= reduce_mask_25;
114 
115  #define carry_pass_final() \
116  carry_pass() \
117  f[9] &= reduce_mask_25;
118 
119  carry_pass_full()
120  carry_pass_full()
121 
122  /* now t is between 0 and 2^255-1, properly carried. */
123  /* case 1: between 0 and 2^255-20. case 2: between 2^255-19 and 2^255-1. */
124  f[0] += 19;
125  carry_pass_full()
126 
127  /* now between 19 and 2^255-1 in both cases, and offset by 19. */
128  f[0] += (1 << 26) - 19;
129  f[1] += (1 << 25) - 1;
130  f[2] += (1 << 26) - 1;
131  f[3] += (1 << 25) - 1;
132  f[4] += (1 << 26) - 1;
133  f[5] += (1 << 25) - 1;
134  f[6] += (1 << 26) - 1;
135  f[7] += (1 << 25) - 1;
136  f[8] += (1 << 26) - 1;
137  f[9] += (1 << 25) - 1;
138 
139  /* now between 2^255 and 2^256-20, and offset by 2^255. */
140  carry_pass_final()
141 
142  #undef carry_pass
143  #undef carry_full
144  #undef carry_final
145 
146  *(word32 *)(out + 0) = ((f[0] ) | (f[1] << 26));
147  *(word32 *)(out + 4) = ((f[1] >> 6) | (f[2] << 19));
148  *(word32 *)(out + 8) = ((f[2] >> 13) | (f[3] << 13));
149  *(word32 *)(out + 12) = ((f[3] >> 19) | (f[4] << 6));
150  *(word32 *)(out + 16) = ((f[5] ) | (f[6] << 25));
151  *(word32 *)(out + 20) = ((f[6] >> 7) | (f[7] << 19));
152  *(word32 *)(out + 24) = ((f[7] >> 13) | (f[8] << 12));
153  *(word32 *)(out + 28) = ((f[8] >> 20) | (f[9] << 6));
154 }
155 
156 /*
157  * Maybe swap the contents of two felem arrays (@a and @b), each 5 elements
158  * long. Perform the swap iff @swap is non-zero.
159  */
160 inline void
161 curve25519_swap_conditional(bignum25519 a, bignum25519 b, word32 iswap) {
162  const word32 swap = (word32)(-(sword32)iswap);
163  xmmi a0,a1,a2,b0,b1,b2,x0,x1,x2;
164  xmmi mask = _mm_cvtsi32_si128(swap);
165  mask = _mm_shuffle_epi32(mask, 0);
166  a0 = _mm_load_si128((xmmi *)a + 0);
167  a1 = _mm_load_si128((xmmi *)a + 1);
168  a2 = _mm_load_si128((xmmi *)a + 2);
169  b0 = _mm_load_si128((xmmi *)b + 0);
170  b1 = _mm_load_si128((xmmi *)b + 1);
171  b2 = _mm_load_si128((xmmi *)b + 2);
172  b0 = _mm_xor_si128(a0, b0);
173  b1 = _mm_xor_si128(a1, b1);
174  b2 = _mm_xor_si128(a2, b2);
175  x0 = _mm_and_si128(b0, mask);
176  x1 = _mm_and_si128(b1, mask);
177  x2 = _mm_and_si128(b2, mask);
178  x0 = _mm_xor_si128(x0, a0);
179  x1 = _mm_xor_si128(x1, a1);
180  x2 = _mm_xor_si128(x2, a2);
181  a0 = _mm_xor_si128(x0, b0);
182  a1 = _mm_xor_si128(x1, b1);
183  a2 = _mm_xor_si128(x2, b2);
184  _mm_store_si128((xmmi *)a + 0, x0);
185  _mm_store_si128((xmmi *)a + 1, x1);
186  _mm_store_si128((xmmi *)a + 2, x2);
187  _mm_store_si128((xmmi *)b + 0, a0);
188  _mm_store_si128((xmmi *)b + 1, a1);
189  _mm_store_si128((xmmi *)b + 2, a2);
190 }
191 
192 /* interleave two bignums */
193 inline void
194 curve25519_tangle32(packedelem32 *out, const bignum25519 x, const bignum25519 z) {
195  xmmi x0,x1,x2,z0,z1,z2;
196 
197  x0 = _mm_load_si128((xmmi *)(x + 0));
198  x1 = _mm_load_si128((xmmi *)(x + 4));
199  x2 = _mm_load_si128((xmmi *)(x + 8));
200  z0 = _mm_load_si128((xmmi *)(z + 0));
201  z1 = _mm_load_si128((xmmi *)(z + 4));
202  z2 = _mm_load_si128((xmmi *)(z + 8));
203 
204  out[0].v = _mm_unpacklo_epi32(x0, z0);
205  out[1].v = _mm_unpackhi_epi32(x0, z0);
206  out[2].v = _mm_unpacklo_epi32(x1, z1);
207  out[3].v = _mm_unpackhi_epi32(x1, z1);
208  out[4].v = _mm_unpacklo_epi32(x2, z2);
209 }
210 
211 /* split a packed bignum in to it's two parts */
212 inline void
213 curve25519_untangle64(bignum25519 x, bignum25519 z, const packedelem64 *in) {
214  _mm_store_si128((xmmi *)(x + 0), _mm_unpacklo_epi64(_mm_unpacklo_epi32(in[0].v, in[1].v), _mm_unpacklo_epi32(in[2].v, in[3].v)));
215  _mm_store_si128((xmmi *)(x + 4), _mm_unpacklo_epi64(_mm_unpacklo_epi32(in[4].v, in[5].v), _mm_unpacklo_epi32(in[6].v, in[7].v)));
216  _mm_store_si128((xmmi *)(x + 8), _mm_unpacklo_epi32(in[8].v, in[9].v) );
217  _mm_store_si128((xmmi *)(z + 0), _mm_unpacklo_epi64(_mm_unpackhi_epi32(in[0].v, in[1].v), _mm_unpackhi_epi32(in[2].v, in[3].v)));
218  _mm_store_si128((xmmi *)(z + 4), _mm_unpacklo_epi64(_mm_unpackhi_epi32(in[4].v, in[5].v), _mm_unpackhi_epi32(in[6].v, in[7].v)));
219  _mm_store_si128((xmmi *)(z + 8), _mm_unpackhi_epi32(in[8].v, in[9].v) );
220 }
221 
222 /* add two packed bignums */
223 inline void
224 curve25519_add_packed32(packedelem32 *out, const packedelem32 *r, const packedelem32 *s) {
225  out[0].v = _mm_add_epi32(r[0].v, s[0].v);
226  out[1].v = _mm_add_epi32(r[1].v, s[1].v);
227  out[2].v = _mm_add_epi32(r[2].v, s[2].v);
228  out[3].v = _mm_add_epi32(r[3].v, s[3].v);
229  out[4].v = _mm_add_epi32(r[4].v, s[4].v);
230 }
231 
232 /* subtract two packed bignums */
233 inline void
234 curve25519_sub_packed32(packedelem32 *out, const packedelem32 *r, const packedelem32 *s) {
235  xmmi r0,r1,r2,r3,r4;
236  xmmi s0,s1,s2,s3;
237  xmmi c1,c2;
238 
239  r0 = _mm_add_epi32(r[0].v, packed32zeromodp0.v);
240  r1 = _mm_add_epi32(r[1].v, packed32zeromodp1.v);
241  r2 = _mm_add_epi32(r[2].v, packed32zeromodp1.v);
242  r3 = _mm_add_epi32(r[3].v, packed32zeromodp1.v);
243  r4 = _mm_add_epi32(r[4].v, packed32zeromodp1.v);
244  r0 = _mm_sub_epi32(r0, s[0].v); /* 00 11 */
245  r1 = _mm_sub_epi32(r1, s[1].v); /* 22 33 */
246  r2 = _mm_sub_epi32(r2, s[2].v); /* 44 55 */
247  r3 = _mm_sub_epi32(r3, s[3].v); /* 66 77 */
248  r4 = _mm_sub_epi32(r4, s[4].v); /* 88 99 */
249 
250  s0 = _mm_unpacklo_epi64(r0, r2); /* 00 44 */
251  s1 = _mm_unpackhi_epi64(r0, r2); /* 11 55 */
252  s2 = _mm_unpacklo_epi64(r1, r3); /* 22 66 */
253  s3 = _mm_unpackhi_epi64(r1, r3); /* 33 77 */
254 
255  c1 = _mm_srli_epi32(s0, 26); c2 = _mm_srli_epi32(s2, 26); s0 = _mm_and_si128(s0, packedmask26262626.v); s2 = _mm_and_si128(s2, packedmask26262626.v); s1 = _mm_add_epi32(s1, c1); s3 = _mm_add_epi32(s3, c2);
256  c1 = _mm_srli_epi32(s1, 25); c2 = _mm_srli_epi32(s3, 25); s1 = _mm_and_si128(s1, packedmask25252525.v); s3 = _mm_and_si128(s3, packedmask25252525.v); s2 = _mm_add_epi32(s2, c1); r4 = _mm_add_epi32(r4, _mm_srli_si128(c2, 8)); s0 = _mm_add_epi32(s0, _mm_slli_si128(c2, 8));
257 
258  out[0].v = _mm_unpacklo_epi64(s0, s1); /* 00 11 */
259  out[1].v = _mm_unpacklo_epi64(s2, s3); /* 22 33 */
260  out[2].v = _mm_unpackhi_epi64(s0, s1); /* 44 55 */
261  out[3].v = _mm_unpackhi_epi64(s2, s3); /* 66 77 */
262  out[4].v = r4; /* 88 99 */
263 }
264 
265 /* multiply two packed bignums */
266 inline void
267 curve25519_mul_packed64(packedelem64 *out, const packedelem64 *r, const packedelem64 *s) {
268  xmmi r1,r2,r3,r4,r5,r6,r7,r8,r9;
269  xmmi r1_2,r3_2,r5_2,r7_2,r9_2;
270  xmmi c1,c2;
271 
272  out[0].v = _mm_mul_epu32(r[0].v, s[0].v);
273  out[1].v = _mm_add_epi64(_mm_mul_epu32(r[0].v, s[1].v), _mm_mul_epu32(r[1].v, s[0].v));
274  r1_2 = _mm_slli_epi32(r[1].v, 1);
275  out[2].v = _mm_add_epi64(_mm_mul_epu32(r[0].v, s[2].v), _mm_add_epi64(_mm_mul_epu32(r1_2 , s[1].v), _mm_mul_epu32(r[2].v, s[0].v)));
276  out[3].v = _mm_add_epi64(_mm_mul_epu32(r[0].v, s[3].v), _mm_add_epi64(_mm_mul_epu32(r[1].v, s[2].v), _mm_add_epi64(_mm_mul_epu32(r[2].v, s[1].v), _mm_mul_epu32(r[3].v, s[0].v))));
277  r3_2 = _mm_slli_epi32(r[3].v, 1);
278  out[4].v = _mm_add_epi64(_mm_mul_epu32(r[0].v, s[4].v), _mm_add_epi64(_mm_mul_epu32(r1_2 , s[3].v), _mm_add_epi64(_mm_mul_epu32(r[2].v, s[2].v), _mm_add_epi64(_mm_mul_epu32(r3_2 , s[1].v), _mm_mul_epu32(r[4].v, s[0].v)))));
279  out[5].v = _mm_add_epi64(_mm_mul_epu32(r[0].v, s[5].v), _mm_add_epi64(_mm_mul_epu32(r[1].v, s[4].v), _mm_add_epi64(_mm_mul_epu32(r[2].v, s[3].v), _mm_add_epi64(_mm_mul_epu32(r[3].v, s[2].v), _mm_add_epi64(_mm_mul_epu32(r[4].v, s[1].v), _mm_mul_epu32(r[5].v, s[0].v))))));
280  r5_2 = _mm_slli_epi32(r[5].v, 1);
281  out[6].v = _mm_add_epi64(_mm_mul_epu32(r[0].v, s[6].v), _mm_add_epi64(_mm_mul_epu32(r1_2 , s[5].v), _mm_add_epi64(_mm_mul_epu32(r[2].v, s[4].v), _mm_add_epi64(_mm_mul_epu32(r3_2 , s[3].v), _mm_add_epi64(_mm_mul_epu32(r[4].v, s[2].v), _mm_add_epi64(_mm_mul_epu32(r5_2 , s[1].v), _mm_mul_epu32(r[6].v, s[0].v)))))));
282  out[7].v = _mm_add_epi64(_mm_mul_epu32(r[0].v, s[7].v), _mm_add_epi64(_mm_mul_epu32(r[1].v, s[6].v), _mm_add_epi64(_mm_mul_epu32(r[2].v, s[5].v), _mm_add_epi64(_mm_mul_epu32(r[3].v, s[4].v), _mm_add_epi64(_mm_mul_epu32(r[4].v, s[3].v), _mm_add_epi64(_mm_mul_epu32(r[5].v, s[2].v), _mm_add_epi64(_mm_mul_epu32(r[6].v, s[1].v), _mm_mul_epu32(r[7].v , s[0].v))))))));
283  r7_2 = _mm_slli_epi32(r[7].v, 1);
284  out[8].v = _mm_add_epi64(_mm_mul_epu32(r[0].v, s[8].v), _mm_add_epi64(_mm_mul_epu32(r1_2 , s[7].v), _mm_add_epi64(_mm_mul_epu32(r[2].v, s[6].v), _mm_add_epi64(_mm_mul_epu32(r3_2 , s[5].v), _mm_add_epi64(_mm_mul_epu32(r[4].v, s[4].v), _mm_add_epi64(_mm_mul_epu32(r5_2 , s[3].v), _mm_add_epi64(_mm_mul_epu32(r[6].v, s[2].v), _mm_add_epi64(_mm_mul_epu32(r7_2 , s[1].v), _mm_mul_epu32(r[8].v, s[0].v)))))))));
285  out[9].v = _mm_add_epi64(_mm_mul_epu32(r[0].v, s[9].v), _mm_add_epi64(_mm_mul_epu32(r[1].v, s[8].v), _mm_add_epi64(_mm_mul_epu32(r[2].v, s[7].v), _mm_add_epi64(_mm_mul_epu32(r[3].v, s[6].v), _mm_add_epi64(_mm_mul_epu32(r[4].v, s[5].v), _mm_add_epi64(_mm_mul_epu32(r[5].v, s[4].v), _mm_add_epi64(_mm_mul_epu32(r[6].v, s[3].v), _mm_add_epi64(_mm_mul_epu32(r[7].v, s[2].v), _mm_add_epi64(_mm_mul_epu32(r[8].v, s[1].v), _mm_mul_epu32(r[9].v, s[0].v))))))))));
286 
287  r1 = _mm_mul_epu32(r[1].v, packednineteen.v);
288  r2 = _mm_mul_epu32(r[2].v, packednineteen.v);
289  r1_2 = _mm_slli_epi32(r1, 1);
290  r3 = _mm_mul_epu32(r[3].v, packednineteen.v);
291  r4 = _mm_mul_epu32(r[4].v, packednineteen.v);
292  r3_2 = _mm_slli_epi32(r3, 1);
293  r5 = _mm_mul_epu32(r[5].v, packednineteen.v);
294  r6 = _mm_mul_epu32(r[6].v, packednineteen.v);
295  r5_2 = _mm_slli_epi32(r5, 1);
296  r7 = _mm_mul_epu32(r[7].v, packednineteen.v);
297  r8 = _mm_mul_epu32(r[8].v, packednineteen.v);
298  r7_2 = _mm_slli_epi32(r7, 1);
299  r9 = _mm_mul_epu32(r[9].v, packednineteen.v);
300  r9_2 = _mm_slli_epi32(r9, 1);
301 
302  out[0].v = _mm_add_epi64(out[0].v, _mm_add_epi64(_mm_mul_epu32(r9_2, s[1].v), _mm_add_epi64(_mm_mul_epu32(r8, s[2].v), _mm_add_epi64(_mm_mul_epu32(r7_2, s[3].v), _mm_add_epi64(_mm_mul_epu32(r6, s[4].v), _mm_add_epi64(_mm_mul_epu32(r5_2, s[5].v), _mm_add_epi64(_mm_mul_epu32(r4, s[6].v), _mm_add_epi64(_mm_mul_epu32(r3_2, s[7].v), _mm_add_epi64(_mm_mul_epu32(r2, s[8].v), _mm_mul_epu32(r1_2, s[9].v))))))))));
303  out[1].v = _mm_add_epi64(out[1].v, _mm_add_epi64(_mm_mul_epu32(r9 , s[2].v), _mm_add_epi64(_mm_mul_epu32(r8, s[3].v), _mm_add_epi64(_mm_mul_epu32(r7 , s[4].v), _mm_add_epi64(_mm_mul_epu32(r6, s[5].v), _mm_add_epi64(_mm_mul_epu32(r5 , s[6].v), _mm_add_epi64(_mm_mul_epu32(r4, s[7].v), _mm_add_epi64(_mm_mul_epu32(r3 , s[8].v), _mm_mul_epu32(r2, s[9].v)))))))));
304  out[2].v = _mm_add_epi64(out[2].v, _mm_add_epi64(_mm_mul_epu32(r9_2, s[3].v), _mm_add_epi64(_mm_mul_epu32(r8, s[4].v), _mm_add_epi64(_mm_mul_epu32(r7_2, s[5].v), _mm_add_epi64(_mm_mul_epu32(r6, s[6].v), _mm_add_epi64(_mm_mul_epu32(r5_2, s[7].v), _mm_add_epi64(_mm_mul_epu32(r4, s[8].v), _mm_mul_epu32(r3_2, s[9].v))))))));
305  out[3].v = _mm_add_epi64(out[3].v, _mm_add_epi64(_mm_mul_epu32(r9 , s[4].v), _mm_add_epi64(_mm_mul_epu32(r8, s[5].v), _mm_add_epi64(_mm_mul_epu32(r7 , s[6].v), _mm_add_epi64(_mm_mul_epu32(r6, s[7].v), _mm_add_epi64(_mm_mul_epu32(r5 , s[8].v), _mm_mul_epu32(r4, s[9].v)))))));
306  out[4].v = _mm_add_epi64(out[4].v, _mm_add_epi64(_mm_mul_epu32(r9_2, s[5].v), _mm_add_epi64(_mm_mul_epu32(r8, s[6].v), _mm_add_epi64(_mm_mul_epu32(r7_2, s[7].v), _mm_add_epi64(_mm_mul_epu32(r6, s[8].v), _mm_mul_epu32(r5_2, s[9].v))))));
307  out[5].v = _mm_add_epi64(out[5].v, _mm_add_epi64(_mm_mul_epu32(r9 , s[6].v), _mm_add_epi64(_mm_mul_epu32(r8, s[7].v), _mm_add_epi64(_mm_mul_epu32(r7 , s[8].v), _mm_mul_epu32(r6, s[9].v)))));
308  out[6].v = _mm_add_epi64(out[6].v, _mm_add_epi64(_mm_mul_epu32(r9_2, s[7].v), _mm_add_epi64(_mm_mul_epu32(r8, s[8].v), _mm_mul_epu32(r7_2, s[9].v))));
309  out[7].v = _mm_add_epi64(out[7].v, _mm_add_epi64(_mm_mul_epu32(r9 , s[8].v), _mm_mul_epu32(r8, s[9].v)));
310  out[8].v = _mm_add_epi64(out[8].v, _mm_mul_epu32(r9_2, s[9].v));
311 
312  c1 = _mm_srli_epi64(out[0].v, 26); c2 = _mm_srli_epi64(out[4].v, 26); out[0].v = _mm_and_si128(out[0].v, packedmask26.v); out[4].v = _mm_and_si128(out[4].v, packedmask26.v); out[1].v = _mm_add_epi64(out[1].v, c1); out[5].v = _mm_add_epi64(out[5].v, c2);
313  c1 = _mm_srli_epi64(out[1].v, 25); c2 = _mm_srli_epi64(out[5].v, 25); out[1].v = _mm_and_si128(out[1].v, packedmask25.v); out[5].v = _mm_and_si128(out[5].v, packedmask25.v); out[2].v = _mm_add_epi64(out[2].v, c1); out[6].v = _mm_add_epi64(out[6].v, c2);
314  c1 = _mm_srli_epi64(out[2].v, 26); c2 = _mm_srli_epi64(out[6].v, 26); out[2].v = _mm_and_si128(out[2].v, packedmask26.v); out[6].v = _mm_and_si128(out[6].v, packedmask26.v); out[3].v = _mm_add_epi64(out[3].v, c1); out[7].v = _mm_add_epi64(out[7].v, c2);
315  c1 = _mm_srli_epi64(out[3].v, 25); c2 = _mm_srli_epi64(out[7].v, 25); out[3].v = _mm_and_si128(out[3].v, packedmask25.v); out[7].v = _mm_and_si128(out[7].v, packedmask25.v); out[4].v = _mm_add_epi64(out[4].v, c1); out[8].v = _mm_add_epi64(out[8].v, c2);
316  c2 = _mm_srli_epi64(out[8].v, 26); out[8].v = _mm_and_si128(out[8].v, packedmask26.v); out[9].v = _mm_add_epi64(out[9].v, c2);
317  c2 = _mm_srli_epi64(out[9].v, 25); out[9].v = _mm_and_si128(out[9].v, packedmask25.v); out[0].v = _mm_add_epi64(out[0].v, _mm_mul_epu32(c2, packednineteen.v));
318  c1 = _mm_srli_epi64(out[0].v, 26); c2 = _mm_srli_epi64(out[4].v, 26); out[0].v = _mm_and_si128(out[0].v, packedmask26.v); out[4].v = _mm_and_si128(out[4].v, packedmask26.v); out[1].v = _mm_add_epi64(out[1].v, c1); out[5].v = _mm_add_epi64(out[5].v, c2);
319 }
320 
321 /* multiply a bignum */
322 void
323 curve25519_mul(bignum25519 out, const bignum25519 r, const bignum25519 s) {
324  xmmi m01,m23,m45,m67,m89;
325  xmmi m0123,m4567;
326  xmmi s0123,s4567;
327  xmmi s01,s23,s45,s67,s89;
328  xmmi s12,s34,s56,s78,s9;
329  xmmi r0,r2,r4,r6,r8;
330  xmmi r1,r3,r5,r7,r9;
331  xmmi r119,r219,r319,r419,r519,r619,r719,r819,r919;
332  xmmi c1,c2,c3;
333 
334  s0123 = _mm_load_si128((xmmi*)s + 0);
335  s01 = _mm_shuffle_epi32(s0123,_MM_SHUFFLE(3,1,2,0));
336  s12 = _mm_shuffle_epi32(s0123, _MM_SHUFFLE(2,2,1,1));
337  s23 = _mm_shuffle_epi32(s0123,_MM_SHUFFLE(3,3,2,2));
338  s4567 = _mm_load_si128((xmmi*)s + 1);
339  s34 = _mm_unpacklo_epi64(_mm_srli_si128(s0123,12),s4567);
340  s45 = _mm_shuffle_epi32(s4567,_MM_SHUFFLE(3,1,2,0));
341  s56 = _mm_shuffle_epi32(s4567, _MM_SHUFFLE(2,2,1,1));
342  s67 = _mm_shuffle_epi32(s4567,_MM_SHUFFLE(3,3,2,2));
343  s89 = _mm_load_si128((xmmi*)s + 2);
344  s78 = _mm_unpacklo_epi64(_mm_srli_si128(s4567,12),s89);
345  s89 = _mm_shuffle_epi32(s89,_MM_SHUFFLE(3,1,2,0));
346  s9 = _mm_shuffle_epi32(s89, _MM_SHUFFLE(3,3,2,2));
347 
348  r0 = _mm_load_si128((xmmi*)r + 0);
349  r1 = _mm_shuffle_epi32(r0, _MM_SHUFFLE(1,1,1,1));
350  r1 = _mm_add_epi64(r1, _mm_and_si128(r1, sse2_top64bitmask.v));
351  r2 = _mm_shuffle_epi32(r0, _MM_SHUFFLE(2,2,2,2));
352  r3 = _mm_shuffle_epi32(r0, _MM_SHUFFLE(3,3,3,3));
353  r3 = _mm_add_epi64(r3, _mm_and_si128(r3, sse2_top64bitmask.v));
354  r0 = _mm_shuffle_epi32(r0, _MM_SHUFFLE(0,0,0,0));
355  r4 = _mm_load_si128((xmmi*)r + 1);
356  r5 = _mm_shuffle_epi32(r4, _MM_SHUFFLE(1,1,1,1));
357  r5 = _mm_add_epi64(r5, _mm_and_si128(r5, sse2_top64bitmask.v));
358  r6 = _mm_shuffle_epi32(r4, _MM_SHUFFLE(2,2,2,2));
359  r7 = _mm_shuffle_epi32(r4, _MM_SHUFFLE(3,3,3,3));
360  r7 = _mm_add_epi64(r7, _mm_and_si128(r7, sse2_top64bitmask.v));
361  r4 = _mm_shuffle_epi32(r4, _MM_SHUFFLE(0,0,0,0));
362  r8 = _mm_load_si128((xmmi*)r + 2);
363  r9 = _mm_shuffle_epi32(r8, _MM_SHUFFLE(3,1,3,1));
364  r9 = _mm_add_epi64(r9, _mm_and_si128(r9, sse2_top64bitmask.v));
365  r8 = _mm_shuffle_epi32(r8, _MM_SHUFFLE(3,0,3,0));
366 
367  m01 = _mm_mul_epu32(r1,s01);
368  m23 = _mm_mul_epu32(r1,s23);
369  m45 = _mm_mul_epu32(r1,s45);
370  m67 = _mm_mul_epu32(r1,s67);
371  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r3,s01));
372  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r3,s23));
373  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r3,s45));
374  m89 = _mm_mul_epu32(r1,s89);
375  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r5,s01));
376  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r5,s23));
377  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r3,s67));
378  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r7,s01));
379  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r5,s45));
380  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r7,s23));
381  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r9,s01));
382 
383  /* shift up */
384  m89 = _mm_unpackhi_epi64(m67,_mm_slli_si128(m89,8));
385  m67 = _mm_unpackhi_epi64(m45,_mm_slli_si128(m67,8));
386  m45 = _mm_unpackhi_epi64(m23,_mm_slli_si128(m45,8));
387  m23 = _mm_unpackhi_epi64(m01,_mm_slli_si128(m23,8));
388  m01 = _mm_unpackhi_epi64(_mm_setzero_si128(),_mm_slli_si128(m01,8));
389 
390  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r0,s01));
391  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r0,s23));
392  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r0,s45));
393  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r0,s67));
394  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r2,s01));
395  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r2,s23));
396  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r4,s23));
397  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r0,s89));
398  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r4,s01));
399  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r2,s45));
400  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r2,s67));
401  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r6,s01));
402  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r4,s45));
403  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r6,s23));
404  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r8,s01));
405 
406  r219 = _mm_mul_epu32(r2, packednineteen.v);
407  r419 = _mm_mul_epu32(r4, packednineteen.v);
408  r619 = _mm_mul_epu32(r6, packednineteen.v);
409  r819 = _mm_mul_epu32(r8, packednineteen.v);
410  r119 = _mm_shuffle_epi32(r1,_MM_SHUFFLE(0,0,2,2)); r119 = _mm_mul_epu32(r119, packednineteen.v);
411  r319 = _mm_shuffle_epi32(r3,_MM_SHUFFLE(0,0,2,2)); r319 = _mm_mul_epu32(r319, packednineteen.v);
412  r519 = _mm_shuffle_epi32(r5,_MM_SHUFFLE(0,0,2,2)); r519 = _mm_mul_epu32(r519, packednineteen.v);
413  r719 = _mm_shuffle_epi32(r7,_MM_SHUFFLE(0,0,2,2)); r719 = _mm_mul_epu32(r719, packednineteen.v);
414  r919 = _mm_shuffle_epi32(r9,_MM_SHUFFLE(0,0,2,2)); r919 = _mm_mul_epu32(r919, packednineteen.v);
415 
416  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r919,s12));
417  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r919,s34));
418  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r919,s56));
419  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r919,s78));
420  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r719,s34));
421  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r719,s56));
422  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r719,s78));
423  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r719,s9));
424  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r519,s56));
425  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r519,s78));
426  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r519,s9));
427  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r819,s89));
428  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r319,s78));
429  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r319,s9));
430  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r619,s89));
431  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r919,s9));
432  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r819,s23));
433  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r819,s45));
434  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r819,s67));
435  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r619,s45));
436  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r619,s67));
437  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r419,s67));
438  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r419,s89));
439  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r219,s89));
440  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r119,s9));
441 
442  r0 = _mm_unpacklo_epi64(m01, m45);
443  r1 = _mm_unpackhi_epi64(m01, m45);
444  r2 = _mm_unpacklo_epi64(m23, m67);
445  r3 = _mm_unpackhi_epi64(m23, m67);
446  r4 = _mm_unpacklo_epi64(m89, m89);
447  r5 = _mm_unpackhi_epi64(m89, m89);
448 
449  c1 = _mm_srli_epi64(r0, 26); c2 = _mm_srli_epi64(r2, 26); r0 = _mm_and_si128(r0, packedmask26.v); r2 = _mm_and_si128(r2, packedmask26.v); r1 = _mm_add_epi64(r1, c1); r3 = _mm_add_epi64(r3, c2);
450  c1 = _mm_srli_epi64(r1, 25); c2 = _mm_srli_epi64(r3, 25); r1 = _mm_and_si128(r1, packedmask25.v); r3 = _mm_and_si128(r3, packedmask25.v); r2 = _mm_add_epi64(r2, c1); r4 = _mm_add_epi64(r4, c2); c3 = _mm_slli_si128(c2, 8);
451  c1 = _mm_srli_epi64(r4, 26); r4 = _mm_and_si128(r4, packedmask26.v); r5 = _mm_add_epi64(r5, c1);
452  c1 = _mm_srli_epi64(r5, 25); r5 = _mm_and_si128(r5, packedmask25.v); r0 = _mm_add_epi64(r0, _mm_unpackhi_epi64(_mm_mul_epu32(c1, packednineteen.v), c3));
453  c1 = _mm_srli_epi64(r0, 26); c2 = _mm_srli_epi64(r2, 26); r0 = _mm_and_si128(r0, packedmask26.v); r2 = _mm_and_si128(r2, packedmask26.v); r1 = _mm_add_epi64(r1, c1); r3 = _mm_add_epi64(r3, c2);
454 
455  m0123 = _mm_unpacklo_epi32(r0, r1);
456  m4567 = _mm_unpackhi_epi32(r0, r1);
457  m0123 = _mm_unpacklo_epi64(m0123, _mm_unpacklo_epi32(r2, r3));
458  m4567 = _mm_unpacklo_epi64(m4567, _mm_unpackhi_epi32(r2, r3));
459  m89 = _mm_unpackhi_epi32(r4, r5);
460 
461  _mm_store_si128((xmmi*)out + 0, m0123);
462  _mm_store_si128((xmmi*)out + 1, m4567);
463  _mm_store_si128((xmmi*)out + 2, m89);
464 }
465 
466 typedef struct bignum25519mulprecomp_t {
467  xmmi r0,r2,r4,r6,r8;
468  xmmi r1,r3,r5,r7,r9;
469  xmmi r119,r219,r319,r419,r519,r619,r719,r819,r919;
470 } bignum25519mulprecomp;
471 
472 /* precompute a constant to multiply by */
473 inline void
474 curve25519_mul_precompute(bignum25519mulprecomp *pre, const bignum25519 r) {
475  pre->r0 = _mm_load_si128((xmmi*)r + 0);
476  pre->r1 = _mm_shuffle_epi32(pre->r0, _MM_SHUFFLE(1,1,1,1));
477  pre->r1 = _mm_add_epi64(pre->r1, _mm_and_si128(pre->r1, sse2_top64bitmask.v));
478  pre->r2 = _mm_shuffle_epi32(pre->r0, _MM_SHUFFLE(2,2,2,2));
479  pre->r3 = _mm_shuffle_epi32(pre->r0, _MM_SHUFFLE(3,3,3,3));
480  pre->r3 = _mm_add_epi64(pre->r3, _mm_and_si128(pre->r3, sse2_top64bitmask.v));
481  pre->r0 = _mm_shuffle_epi32(pre->r0, _MM_SHUFFLE(0,0,0,0));
482  pre->r4 = _mm_load_si128((xmmi*)r + 1);
483  pre->r5 = _mm_shuffle_epi32(pre->r4, _MM_SHUFFLE(1,1,1,1));
484  pre->r5 = _mm_add_epi64(pre->r5, _mm_and_si128(pre->r5, sse2_top64bitmask.v));
485  pre->r6 = _mm_shuffle_epi32(pre->r4, _MM_SHUFFLE(2,2,2,2));
486  pre->r7 = _mm_shuffle_epi32(pre->r4, _MM_SHUFFLE(3,3,3,3));
487  pre->r7 = _mm_add_epi64(pre->r7, _mm_and_si128(pre->r7, sse2_top64bitmask.v));
488  pre->r4 = _mm_shuffle_epi32(pre->r4, _MM_SHUFFLE(0,0,0,0));
489  pre->r8 = _mm_load_si128((xmmi*)r + 2);
490  pre->r9 = _mm_shuffle_epi32(pre->r8, _MM_SHUFFLE(3,1,3,1));
491  pre->r9 = _mm_add_epi64(pre->r9, _mm_and_si128(pre->r9, sse2_top64bitmask.v));
492  pre->r8 = _mm_shuffle_epi32(pre->r8, _MM_SHUFFLE(3,0,3,0));
493 
494  pre->r219 = _mm_mul_epu32(pre->r2, packednineteen.v);
495  pre->r419 = _mm_mul_epu32(pre->r4, packednineteen.v);
496  pre->r619 = _mm_mul_epu32(pre->r6, packednineteen.v);
497  pre->r819 = _mm_mul_epu32(pre->r8, packednineteen.v);
498  pre->r119 = _mm_shuffle_epi32(pre->r1,_MM_SHUFFLE(0,0,2,2)); pre->r119 = _mm_mul_epu32(pre->r119, packednineteen.v);
499  pre->r319 = _mm_shuffle_epi32(pre->r3,_MM_SHUFFLE(0,0,2,2)); pre->r319 = _mm_mul_epu32(pre->r319, packednineteen.v);
500  pre->r519 = _mm_shuffle_epi32(pre->r5,_MM_SHUFFLE(0,0,2,2)); pre->r519 = _mm_mul_epu32(pre->r519, packednineteen.v);
501  pre->r719 = _mm_shuffle_epi32(pre->r7,_MM_SHUFFLE(0,0,2,2)); pre->r719 = _mm_mul_epu32(pre->r719, packednineteen.v);
502  pre->r919 = _mm_shuffle_epi32(pre->r9,_MM_SHUFFLE(0,0,2,2)); pre->r919 = _mm_mul_epu32(pre->r919, packednineteen.v);
503 }
504 
505 
506 /* multiply a bignum by a pre-computed constant */
507 inline void
508 curve25519_mul_precomputed(bignum25519 out, const bignum25519 s, const bignum25519mulprecomp *r) {
509  xmmi m01,m23,m45,m67,m89;
510  xmmi m0123,m4567;
511  xmmi s0123,s4567;
512  xmmi s01,s23,s45,s67,s89;
513  xmmi s12,s34,s56,s78,s9;
514  xmmi r0,r1,r2,r3,r4,r5;
515  xmmi c1,c2,c3;
516 
517  s0123 = _mm_load_si128((xmmi*)s + 0);
518  s01 = _mm_shuffle_epi32(s0123,_MM_SHUFFLE(3,1,2,0));
519  s12 = _mm_shuffle_epi32(s0123, _MM_SHUFFLE(2,2,1,1));
520  s23 = _mm_shuffle_epi32(s0123,_MM_SHUFFLE(3,3,2,2));
521  s4567 = _mm_load_si128((xmmi*)s + 1);
522  s34 = _mm_unpacklo_epi64(_mm_srli_si128(s0123,12),s4567);
523  s45 = _mm_shuffle_epi32(s4567,_MM_SHUFFLE(3,1,2,0));
524  s56 = _mm_shuffle_epi32(s4567, _MM_SHUFFLE(2,2,1,1));
525  s67 = _mm_shuffle_epi32(s4567,_MM_SHUFFLE(3,3,2,2));
526  s89 = _mm_load_si128((xmmi*)s + 2);
527  s78 = _mm_unpacklo_epi64(_mm_srli_si128(s4567,12),s89);
528  s89 = _mm_shuffle_epi32(s89,_MM_SHUFFLE(3,1,2,0));
529  s9 = _mm_shuffle_epi32(s89, _MM_SHUFFLE(3,3,2,2));
530 
531  m01 = _mm_mul_epu32(r->r1,s01);
532  m23 = _mm_mul_epu32(r->r1,s23);
533  m45 = _mm_mul_epu32(r->r1,s45);
534  m67 = _mm_mul_epu32(r->r1,s67);
535  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r->r3,s01));
536  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r->r3,s23));
537  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r->r3,s45));
538  m89 = _mm_mul_epu32(r->r1,s89);
539  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r->r5,s01));
540  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r->r5,s23));
541  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r->r3,s67));
542  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r->r7,s01));
543  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r->r5,s45));
544  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r->r7,s23));
545  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r->r9,s01));
546 
547  /* shift up */
548  m89 = _mm_unpackhi_epi64(m67,_mm_slli_si128(m89,8));
549  m67 = _mm_unpackhi_epi64(m45,_mm_slli_si128(m67,8));
550  m45 = _mm_unpackhi_epi64(m23,_mm_slli_si128(m45,8));
551  m23 = _mm_unpackhi_epi64(m01,_mm_slli_si128(m23,8));
552  m01 = _mm_unpackhi_epi64(_mm_setzero_si128(),_mm_slli_si128(m01,8));
553 
554  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r->r0,s01));
555  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r->r0,s23));
556  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r->r0,s45));
557  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r->r0,s67));
558  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r->r2,s01));
559  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r->r2,s23));
560  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r->r4,s23));
561  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r->r0,s89));
562  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r->r4,s01));
563  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r->r2,s45));
564  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r->r2,s67));
565  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r->r6,s01));
566  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r->r4,s45));
567  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r->r6,s23));
568  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r->r8,s01));
569  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r->r919,s12));
570  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r->r919,s34));
571  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r->r919,s56));
572  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r->r919,s78));
573  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r->r719,s34));
574  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r->r719,s56));
575  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r->r719,s78));
576  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r->r719,s9));
577  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r->r519,s56));
578  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r->r519,s78));
579  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r->r519,s9));
580  m67 = _mm_add_epi64(m67,_mm_mul_epu32(r->r819,s89));
581  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r->r319,s78));
582  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r->r319,s9));
583  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r->r619,s89));
584  m89 = _mm_add_epi64(m89,_mm_mul_epu32(r->r919,s9));
585  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r->r819,s23));
586  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r->r819,s45));
587  m45 = _mm_add_epi64(m45,_mm_mul_epu32(r->r819,s67));
588  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r->r619,s45));
589  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r->r619,s67));
590  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r->r419,s67));
591  m23 = _mm_add_epi64(m23,_mm_mul_epu32(r->r419,s89));
592  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r->r219,s89));
593  m01 = _mm_add_epi64(m01,_mm_mul_epu32(r->r119,s9));
594 
595  r0 = _mm_unpacklo_epi64(m01, m45);
596  r1 = _mm_unpackhi_epi64(m01, m45);
597  r2 = _mm_unpacklo_epi64(m23, m67);
598  r3 = _mm_unpackhi_epi64(m23, m67);
599  r4 = _mm_unpacklo_epi64(m89, m89);
600  r5 = _mm_unpackhi_epi64(m89, m89);
601 
602  c1 = _mm_srli_epi64(r0, 26); c2 = _mm_srli_epi64(r2, 26); r0 = _mm_and_si128(r0, packedmask26.v); r2 = _mm_and_si128(r2, packedmask26.v); r1 = _mm_add_epi64(r1, c1); r3 = _mm_add_epi64(r3, c2);
603  c1 = _mm_srli_epi64(r1, 25); c2 = _mm_srli_epi64(r3, 25); r1 = _mm_and_si128(r1, packedmask25.v); r3 = _mm_and_si128(r3, packedmask25.v); r2 = _mm_add_epi64(r2, c1); r4 = _mm_add_epi64(r4, c2); c3 = _mm_slli_si128(c2, 8);
604  c1 = _mm_srli_epi64(r4, 26); r4 = _mm_and_si128(r4, packedmask26.v); r5 = _mm_add_epi64(r5, c1);
605  c1 = _mm_srli_epi64(r5, 25); r5 = _mm_and_si128(r5, packedmask25.v); r0 = _mm_add_epi64(r0, _mm_unpackhi_epi64(_mm_mul_epu32(c1, packednineteen.v), c3));
606  c1 = _mm_srli_epi64(r0, 26); c2 = _mm_srli_epi64(r2, 26); r0 = _mm_and_si128(r0, packedmask26.v); r2 = _mm_and_si128(r2, packedmask26.v); r1 = _mm_add_epi64(r1, c1); r3 = _mm_add_epi64(r3, c2);
607 
608  m0123 = _mm_unpacklo_epi32(r0, r1);
609  m4567 = _mm_unpackhi_epi32(r0, r1);
610  m0123 = _mm_unpacklo_epi64(m0123, _mm_unpacklo_epi32(r2, r3));
611  m4567 = _mm_unpacklo_epi64(m4567, _mm_unpackhi_epi32(r2, r3));
612  m89 = _mm_unpackhi_epi32(r4, r5);
613 
614  _mm_store_si128((xmmi*)out + 0, m0123);
615  _mm_store_si128((xmmi*)out + 1, m4567);
616  _mm_store_si128((xmmi*)out + 2, m89);
617 }
618 
619 /* square a bignum 'count' times */
620 #define curve25519_square(r,x) curve25519_square_times(r,x,1)
621 
622 void
623 curve25519_square_times(bignum25519 r, const bignum25519 in, int count) {
624  xmmi m01,m23,m45,m67,m89;
625  xmmi r0,r1,r2,r3,r4,r5,r6,r7,r8,r9;
626  xmmi r0a,r1a,r2a,r3a,r7a,r9a;
627  xmmi r0123,r4567;
628  xmmi r01,r23,r45,r67,r6x,r89,r8x;
629  xmmi r12,r34,r56,r78,r9x;
630  xmmi r5619;
631  xmmi c1,c2,c3;
632 
633  r0123 = _mm_load_si128((xmmi*)in + 0);
634  r01 = _mm_shuffle_epi32(r0123,_MM_SHUFFLE(3,1,2,0));
635  r23 = _mm_shuffle_epi32(r0123,_MM_SHUFFLE(3,3,2,2));
636  r4567 = _mm_load_si128((xmmi*)in + 1);
637  r45 = _mm_shuffle_epi32(r4567,_MM_SHUFFLE(3,1,2,0));
638  r67 = _mm_shuffle_epi32(r4567,_MM_SHUFFLE(3,3,2,2));
639  r89 = _mm_load_si128((xmmi*)in + 2);
640  r89 = _mm_shuffle_epi32(r89,_MM_SHUFFLE(3,1,2,0));
641 
642  do {
643  r12 = _mm_unpackhi_epi64(r01, _mm_slli_si128(r23, 8));
644  r0 = _mm_shuffle_epi32(r01, _MM_SHUFFLE(0,0,0,0));
645  r0 = _mm_add_epi64(r0, _mm_and_si128(r0, sse2_top64bitmask.v));
646  r0a = _mm_shuffle_epi32(r0,_MM_SHUFFLE(3,2,1,2));
647  r1 = _mm_shuffle_epi32(r01, _MM_SHUFFLE(2,2,2,2));
648  r2 = _mm_shuffle_epi32(r23, _MM_SHUFFLE(0,0,0,0));
649  r2 = _mm_add_epi64(r2, _mm_and_si128(r2, sse2_top64bitmask.v));
650  r2a = _mm_shuffle_epi32(r2,_MM_SHUFFLE(3,2,1,2));
651  r3 = _mm_shuffle_epi32(r23, _MM_SHUFFLE(2,2,2,2));
652  r34 = _mm_unpackhi_epi64(r23, _mm_slli_si128(r45, 8));
653  r4 = _mm_shuffle_epi32(r45, _MM_SHUFFLE(0,0,0,0));
654  r4 = _mm_add_epi64(r4, _mm_and_si128(r4, sse2_top64bitmask.v));
655  r56 = _mm_unpackhi_epi64(r45, _mm_slli_si128(r67, 8));
656  r5619 = _mm_mul_epu32(r56, packednineteen.v);
657  r5 = _mm_shuffle_epi32(r5619, _MM_SHUFFLE(1,1,1,0));
658  r6 = _mm_shuffle_epi32(r5619, _MM_SHUFFLE(3,2,3,2));
659  r78 = _mm_unpackhi_epi64(r67, _mm_slli_si128(r89, 8));
660  r6x = _mm_unpacklo_epi64(r67, _mm_setzero_si128());
661  r7 = _mm_shuffle_epi32(r67, _MM_SHUFFLE(2,2,2,2));
662  r7 = _mm_mul_epu32(r7, packed3819.v);
663  r7a = _mm_shuffle_epi32(r7, _MM_SHUFFLE(3,3,3,2));
664  r8x = _mm_unpacklo_epi64(r89, _mm_setzero_si128());
665  r8 = _mm_shuffle_epi32(r89, _MM_SHUFFLE(0,0,0,0));
666  r8 = _mm_mul_epu32(r8, packednineteen.v);
667  r9 = _mm_shuffle_epi32(r89, _MM_SHUFFLE(2,2,2,2));
668  r9x = _mm_slli_epi32(_mm_shuffle_epi32(r89, _MM_SHUFFLE(3,3,3,2)), 1);
669  r9 = _mm_mul_epu32(r9, packed3819.v);
670  r9a = _mm_shuffle_epi32(r9, _MM_SHUFFLE(2,2,2,2));
671 
672  m01 = _mm_mul_epu32(r01, r0);
673  m23 = _mm_mul_epu32(r23, r0a);
674  m45 = _mm_mul_epu32(r45, r0a);
675  m45 = _mm_add_epi64(m45, _mm_mul_epu32(r23, r2));
676  r23 = _mm_slli_epi32(r23, 1);
677  m67 = _mm_mul_epu32(r67, r0a);
678  m67 = _mm_add_epi64(m67, _mm_mul_epu32(r45, r2a));
679  m89 = _mm_mul_epu32(r89, r0a);
680  m89 = _mm_add_epi64(m89, _mm_mul_epu32(r67, r2a));
681  r67 = _mm_slli_epi32(r67, 1);
682  m89 = _mm_add_epi64(m89, _mm_mul_epu32(r45, r4));
683  r45 = _mm_slli_epi32(r45, 1);
684 
685  r1 = _mm_slli_epi32(r1, 1);
686  r3 = _mm_slli_epi32(r3, 1);
687  r1a = _mm_add_epi64(r1, _mm_and_si128(r1, sse2_bot64bitmask.v));
688  r3a = _mm_add_epi64(r3, _mm_and_si128(r3, sse2_bot64bitmask.v));
689 
690  m23 = _mm_add_epi64(m23, _mm_mul_epu32(r12, r1));
691  m45 = _mm_add_epi64(m45, _mm_mul_epu32(r34, r1a));
692  m67 = _mm_add_epi64(m67, _mm_mul_epu32(r56, r1a));
693  m67 = _mm_add_epi64(m67, _mm_mul_epu32(r34, r3));
694  r34 = _mm_slli_epi32(r34, 1);
695  m89 = _mm_add_epi64(m89, _mm_mul_epu32(r78, r1a));
696  r78 = _mm_slli_epi32(r78, 1);
697  m89 = _mm_add_epi64(m89, _mm_mul_epu32(r56, r3a));
698  r56 = _mm_slli_epi32(r56, 1);
699 
700  m01 = _mm_add_epi64(m01, _mm_mul_epu32(_mm_slli_epi32(r12, 1), r9));
701  m01 = _mm_add_epi64(m01, _mm_mul_epu32(r34, r7));
702  m23 = _mm_add_epi64(m23, _mm_mul_epu32(r34, r9));
703  m01 = _mm_add_epi64(m01, _mm_mul_epu32(r56, r5));
704  m23 = _mm_add_epi64(m23, _mm_mul_epu32(r56, r7));
705  m45 = _mm_add_epi64(m45, _mm_mul_epu32(r56, r9));
706  m01 = _mm_add_epi64(m01, _mm_mul_epu32(r23, r8));
707  m01 = _mm_add_epi64(m01, _mm_mul_epu32(r45, r6));
708  m23 = _mm_add_epi64(m23, _mm_mul_epu32(r45, r8));
709  m23 = _mm_add_epi64(m23, _mm_mul_epu32(r6x, r6));
710  m45 = _mm_add_epi64(m45, _mm_mul_epu32(r78, r7a));
711  m67 = _mm_add_epi64(m67, _mm_mul_epu32(r78, r9));
712  m45 = _mm_add_epi64(m45, _mm_mul_epu32(r67, r8));
713  m67 = _mm_add_epi64(m67, _mm_mul_epu32(r8x, r8));
714  m89 = _mm_add_epi64(m89, _mm_mul_epu32(r9x, r9a));
715 
716  r0 = _mm_unpacklo_epi64(m01, m45);
717  r1 = _mm_unpackhi_epi64(m01, m45);
718  r2 = _mm_unpacklo_epi64(m23, m67);
719  r3 = _mm_unpackhi_epi64(m23, m67);
720  r4 = _mm_unpacklo_epi64(m89, m89);
721  r5 = _mm_unpackhi_epi64(m89, m89);
722 
723  c1 = _mm_srli_epi64(r0, 26); c2 = _mm_srli_epi64(r2, 26); r0 = _mm_and_si128(r0, packedmask26.v); r2 = _mm_and_si128(r2, packedmask26.v); r1 = _mm_add_epi64(r1, c1); r3 = _mm_add_epi64(r3, c2);
724  c1 = _mm_srli_epi64(r1, 25); c2 = _mm_srli_epi64(r3, 25); r1 = _mm_and_si128(r1, packedmask25.v); r3 = _mm_and_si128(r3, packedmask25.v); r2 = _mm_add_epi64(r2, c1); r4 = _mm_add_epi64(r4, c2); c3 = _mm_slli_si128(c2, 8);
725  c1 = _mm_srli_epi64(r4, 26); r4 = _mm_and_si128(r4, packedmask26.v); r5 = _mm_add_epi64(r5, c1);
726  c1 = _mm_srli_epi64(r5, 25); r5 = _mm_and_si128(r5, packedmask25.v); r0 = _mm_add_epi64(r0, _mm_unpackhi_epi64(_mm_mul_epu32(c1, packednineteen.v), c3));
727  c1 = _mm_srli_epi64(r0, 26); c2 = _mm_srli_epi64(r2, 26); r0 = _mm_and_si128(r0, packedmask26.v); r2 = _mm_and_si128(r2, packedmask26.v); r1 = _mm_add_epi64(r1, c1); r3 = _mm_add_epi64(r3, c2);
728 
729  r01 = _mm_unpacklo_epi64(r0, r1);
730  r45 = _mm_unpackhi_epi64(r0, r1);
731  r23 = _mm_unpacklo_epi64(r2, r3);
732  r67 = _mm_unpackhi_epi64(r2, r3);
733  r89 = _mm_unpackhi_epi64(r4, r5);
734  } while (--count);
735 
736  r0123 = _mm_shuffle_epi32(r23, _MM_SHUFFLE(2,0,3,3));
737  r4567 = _mm_shuffle_epi32(r67, _MM_SHUFFLE(2,0,3,3));
738  r0123 = _mm_or_si128(r0123, _mm_shuffle_epi32(r01, _MM_SHUFFLE(3,3,2,0)));
739  r4567 = _mm_or_si128(r4567, _mm_shuffle_epi32(r45, _MM_SHUFFLE(3,3,2,0)));
740  r89 = _mm_shuffle_epi32(r89, _MM_SHUFFLE(3,3,2,0));
741 
742  _mm_store_si128((xmmi*)r + 0, r0123);
743  _mm_store_si128((xmmi*)r + 1, r4567);
744  _mm_store_si128((xmmi*)r + 2, r89);
745 }
746 
747 /* square two packed bignums */
748 inline void
749 curve25519_square_packed64(packedelem64 *out, const packedelem64 *r) {
750  xmmi r0,r1,r2,r3;
751  xmmi r1_2,r3_2,r4_2,r5_2,r6_2,r7_2;
752  xmmi d5,d6,d7,d8,d9;
753  xmmi c1,c2;
754 
755  r0 = r[0].v;
756  r1 = r[1].v;
757  r2 = r[2].v;
758  r3 = r[3].v;
759 
760  out[0].v = _mm_mul_epu32(r0, r0);
761  r0 = _mm_slli_epi32(r0, 1);
762  out[1].v = _mm_mul_epu32(r0, r1);
763  r1_2 = _mm_slli_epi32(r1, 1);
764  out[2].v = _mm_add_epi64(_mm_mul_epu32(r0, r2 ), _mm_mul_epu32(r1, r1_2));
765  r1 = r1_2;
766  out[3].v = _mm_add_epi64(_mm_mul_epu32(r0, r3 ), _mm_mul_epu32(r1, r2 ));
767  r3_2 = _mm_slli_epi32(r3, 1);
768  out[4].v = _mm_add_epi64(_mm_mul_epu32(r0, r[4].v), _mm_add_epi64(_mm_mul_epu32(r1, r3_2 ), _mm_mul_epu32(r2, r2)));
769  r2 = _mm_slli_epi32(r2, 1);
770  out[5].v = _mm_add_epi64(_mm_mul_epu32(r0, r[5].v), _mm_add_epi64(_mm_mul_epu32(r1, r[4].v), _mm_mul_epu32(r2, r3)));
771  r5_2 = _mm_slli_epi32(r[5].v, 1);
772  out[6].v = _mm_add_epi64(_mm_mul_epu32(r0, r[6].v), _mm_add_epi64(_mm_mul_epu32(r1, r5_2 ), _mm_add_epi64(_mm_mul_epu32(r2, r[4].v), _mm_mul_epu32(r3, r3_2 ))));
773  r3 = r3_2;
774  out[7].v = _mm_add_epi64(_mm_mul_epu32(r0, r[7].v), _mm_add_epi64(_mm_mul_epu32(r1, r[6].v), _mm_add_epi64(_mm_mul_epu32(r2, r[5].v), _mm_mul_epu32(r3, r[4].v))));
775  r7_2 = _mm_slli_epi32(r[7].v, 1);
776  out[8].v = _mm_add_epi64(_mm_mul_epu32(r0, r[8].v), _mm_add_epi64(_mm_mul_epu32(r1, r7_2 ), _mm_add_epi64(_mm_mul_epu32(r2, r[6].v), _mm_add_epi64(_mm_mul_epu32(r3, r5_2 ), _mm_mul_epu32(r[4].v, r[4].v)))));
777  out[9].v = _mm_add_epi64(_mm_mul_epu32(r0, r[9].v), _mm_add_epi64(_mm_mul_epu32(r1, r[8].v), _mm_add_epi64(_mm_mul_epu32(r2, r[7].v), _mm_add_epi64(_mm_mul_epu32(r3, r[6].v), _mm_mul_epu32(r[4].v, r5_2 )))));
778 
779  d5 = _mm_mul_epu32(r[5].v, packedthirtyeight.v);
780  d6 = _mm_mul_epu32(r[6].v, packednineteen.v);
781  d7 = _mm_mul_epu32(r[7].v, packedthirtyeight.v);
782  d8 = _mm_mul_epu32(r[8].v, packednineteen.v);
783  d9 = _mm_mul_epu32(r[9].v, packedthirtyeight.v);
784 
785  r4_2 = _mm_slli_epi32(r[4].v, 1);
786  r6_2 = _mm_slli_epi32(r[6].v, 1);
787  out[0].v = _mm_add_epi64(out[0].v, _mm_add_epi64(_mm_mul_epu32(d9, r1 ), _mm_add_epi64(_mm_mul_epu32(d8, r2 ), _mm_add_epi64(_mm_mul_epu32(d7, r3 ), _mm_add_epi64(_mm_mul_epu32(d6, r4_2), _mm_mul_epu32(d5, r[5].v))))));
788  out[1].v = _mm_add_epi64(out[1].v, _mm_add_epi64(_mm_mul_epu32(d9, _mm_srli_epi32(r2, 1)), _mm_add_epi64(_mm_mul_epu32(d8, r3 ), _mm_add_epi64(_mm_mul_epu32(d7, r[4].v), _mm_mul_epu32(d6, r5_2 )))));
789  out[2].v = _mm_add_epi64(out[2].v, _mm_add_epi64(_mm_mul_epu32(d9, r3 ), _mm_add_epi64(_mm_mul_epu32(d8, r4_2), _mm_add_epi64(_mm_mul_epu32(d7, r5_2 ), _mm_mul_epu32(d6, r[6].v)))));
790  out[3].v = _mm_add_epi64(out[3].v, _mm_add_epi64(_mm_mul_epu32(d9, r[4].v ), _mm_add_epi64(_mm_mul_epu32(d8, r5_2), _mm_mul_epu32(d7, r[6].v))));
791  out[4].v = _mm_add_epi64(out[4].v, _mm_add_epi64(_mm_mul_epu32(d9, r5_2 ), _mm_add_epi64(_mm_mul_epu32(d8, r6_2), _mm_mul_epu32(d7, r[7].v))));
792  out[5].v = _mm_add_epi64(out[5].v, _mm_add_epi64(_mm_mul_epu32(d9, r[6].v ), _mm_mul_epu32(d8, r7_2 )));
793  out[6].v = _mm_add_epi64(out[6].v, _mm_add_epi64(_mm_mul_epu32(d9, r7_2 ), _mm_mul_epu32(d8, r[8].v)));
794  out[7].v = _mm_add_epi64(out[7].v, _mm_mul_epu32(d9, r[8].v));
795  out[8].v = _mm_add_epi64(out[8].v, _mm_mul_epu32(d9, r[9].v));
796 
797  c1 = _mm_srli_epi64(out[0].v, 26); c2 = _mm_srli_epi64(out[4].v, 26); out[0].v = _mm_and_si128(out[0].v, packedmask26.v); out[4].v = _mm_and_si128(out[4].v, packedmask26.v); out[1].v = _mm_add_epi64(out[1].v, c1); out[5].v = _mm_add_epi64(out[5].v, c2);
798  c1 = _mm_srli_epi64(out[1].v, 25); c2 = _mm_srli_epi64(out[5].v, 25); out[1].v = _mm_and_si128(out[1].v, packedmask25.v); out[5].v = _mm_and_si128(out[5].v, packedmask25.v); out[2].v = _mm_add_epi64(out[2].v, c1); out[6].v = _mm_add_epi64(out[6].v, c2);
799  c1 = _mm_srli_epi64(out[2].v, 26); c2 = _mm_srli_epi64(out[6].v, 26); out[2].v = _mm_and_si128(out[2].v, packedmask26.v); out[6].v = _mm_and_si128(out[6].v, packedmask26.v); out[3].v = _mm_add_epi64(out[3].v, c1); out[7].v = _mm_add_epi64(out[7].v, c2);
800  c1 = _mm_srli_epi64(out[3].v, 25); c2 = _mm_srli_epi64(out[7].v, 25); out[3].v = _mm_and_si128(out[3].v, packedmask25.v); out[7].v = _mm_and_si128(out[7].v, packedmask25.v); out[4].v = _mm_add_epi64(out[4].v, c1); out[8].v = _mm_add_epi64(out[8].v, c2);
801  c2 = _mm_srli_epi64(out[8].v, 26); out[8].v = _mm_and_si128(out[8].v, packedmask26.v); out[9].v = _mm_add_epi64(out[9].v, c2);
802  c2 = _mm_srli_epi64(out[9].v, 25); out[9].v = _mm_and_si128(out[9].v, packedmask25.v); out[0].v = _mm_add_epi64(out[0].v, _mm_mul_epu32(c2, packednineteen.v));
803  c1 = _mm_srli_epi64(out[0].v, 26); c2 = _mm_srli_epi64(out[4].v, 26); out[0].v = _mm_and_si128(out[0].v, packedmask26.v); out[4].v = _mm_and_si128(out[4].v, packedmask26.v); out[1].v = _mm_add_epi64(out[1].v, c1); out[5].v = _mm_add_epi64(out[5].v, c2);
804 }
805 
806 /* make [nqx+nqz,nqpqx+nqpqz], [nqpqx-nqpqz,nqx-nqz] from [nqx+nqz,nqpqx+nqpqz], [nqx-nqz,nqpqx-nqpqz] */
807 inline void
808 curve25519_make_nqpq(packedelem64 *primex, packedelem64 *primez, const packedelem32 *pqx, const packedelem32 *pqz) {
809  primex[0].v = _mm_shuffle_epi32(pqx[0].v, _MM_SHUFFLE(1,1,0,0));
810  primex[1].v = _mm_shuffle_epi32(pqx[0].v, _MM_SHUFFLE(3,3,2,2));
811  primex[2].v = _mm_shuffle_epi32(pqx[1].v, _MM_SHUFFLE(1,1,0,0));
812  primex[3].v = _mm_shuffle_epi32(pqx[1].v, _MM_SHUFFLE(3,3,2,2));
813  primex[4].v = _mm_shuffle_epi32(pqx[2].v, _MM_SHUFFLE(1,1,0,0));
814  primex[5].v = _mm_shuffle_epi32(pqx[2].v, _MM_SHUFFLE(3,3,2,2));
815  primex[6].v = _mm_shuffle_epi32(pqx[3].v, _MM_SHUFFLE(1,1,0,0));
816  primex[7].v = _mm_shuffle_epi32(pqx[3].v, _MM_SHUFFLE(3,3,2,2));
817  primex[8].v = _mm_shuffle_epi32(pqx[4].v, _MM_SHUFFLE(1,1,0,0));
818  primex[9].v = _mm_shuffle_epi32(pqx[4].v, _MM_SHUFFLE(3,3,2,2));
819  primez[0].v = _mm_shuffle_epi32(pqz[0].v, _MM_SHUFFLE(0,0,1,1));
820  primez[1].v = _mm_shuffle_epi32(pqz[0].v, _MM_SHUFFLE(2,2,3,3));
821  primez[2].v = _mm_shuffle_epi32(pqz[1].v, _MM_SHUFFLE(0,0,1,1));
822  primez[3].v = _mm_shuffle_epi32(pqz[1].v, _MM_SHUFFLE(2,2,3,3));
823  primez[4].v = _mm_shuffle_epi32(pqz[2].v, _MM_SHUFFLE(0,0,1,1));
824  primez[5].v = _mm_shuffle_epi32(pqz[2].v, _MM_SHUFFLE(2,2,3,3));
825  primez[6].v = _mm_shuffle_epi32(pqz[3].v, _MM_SHUFFLE(0,0,1,1));
826  primez[7].v = _mm_shuffle_epi32(pqz[3].v, _MM_SHUFFLE(2,2,3,3));
827  primez[8].v = _mm_shuffle_epi32(pqz[4].v, _MM_SHUFFLE(0,0,1,1));
828  primez[9].v = _mm_shuffle_epi32(pqz[4].v, _MM_SHUFFLE(2,2,3,3));
829 }
830 
831 /* make [nqx+nqz,nqx-nqz] from [nqx+nqz,nqpqx+nqpqz], [nqx-nqz,nqpqx-nqpqz] */
832 inline void
833 curve25519_make_nq(packedelem64 *nq, const packedelem32 *pqx, const packedelem32 *pqz) {
834  nq[0].v = _mm_unpacklo_epi64(pqx[0].v, pqz[0].v);
835  nq[1].v = _mm_unpackhi_epi64(pqx[0].v, pqz[0].v);
836  nq[2].v = _mm_unpacklo_epi64(pqx[1].v, pqz[1].v);
837  nq[3].v = _mm_unpackhi_epi64(pqx[1].v, pqz[1].v);
838  nq[4].v = _mm_unpacklo_epi64(pqx[2].v, pqz[2].v);
839  nq[5].v = _mm_unpackhi_epi64(pqx[2].v, pqz[2].v);
840  nq[6].v = _mm_unpacklo_epi64(pqx[3].v, pqz[3].v);
841  nq[7].v = _mm_unpackhi_epi64(pqx[3].v, pqz[3].v);
842  nq[8].v = _mm_unpacklo_epi64(pqx[4].v, pqz[4].v);
843  nq[9].v = _mm_unpackhi_epi64(pqx[4].v, pqz[4].v);
844 }
845 
846 /* compute [nqx+nqz,nqx-nqz] from nqx, nqz */
847 inline void
848 curve25519_compute_nq(packedelem64 *nq, const bignum25519 nqx, const bignum25519 nqz) {
849  xmmi x0,x1,x2;
850  xmmi z0,z1,z2;
851  xmmi a0,a1,a2;
852  xmmi s0,s1,s2;
853  xmmi r0,r1;
854  xmmi c1,c2;
855  x0 = _mm_load_si128((xmmi*)nqx + 0);
856  x1 = _mm_load_si128((xmmi*)nqx + 1);
857  x2 = _mm_load_si128((xmmi*)nqx + 2);
858  z0 = _mm_load_si128((xmmi*)nqz + 0);
859  z1 = _mm_load_si128((xmmi*)nqz + 1);
860  z2 = _mm_load_si128((xmmi*)nqz + 2);
861  a0 = _mm_add_epi32(x0, z0);
862  a1 = _mm_add_epi32(x1, z1);
863  a2 = _mm_add_epi32(x2, z2);
864  s0 = _mm_add_epi32(x0, packed2p0.v);
865  s1 = _mm_add_epi32(x1, packed2p1.v);
866  s2 = _mm_add_epi32(x2, packed2p2.v);
867  s0 = _mm_sub_epi32(s0, z0);
868  s1 = _mm_sub_epi32(s1, z1);
869  s2 = _mm_sub_epi32(s2, z2);
870  r0 = _mm_and_si128(_mm_shuffle_epi32(s0, _MM_SHUFFLE(2,2,0,0)), sse2_bot32bitmask.v);
871  r1 = _mm_and_si128(_mm_shuffle_epi32(s0, _MM_SHUFFLE(3,3,1,1)), sse2_bot32bitmask.v);
872  c1 = _mm_srli_epi32(r0, 26);
873  c2 = _mm_srli_epi32(r1, 25);
874  r0 = _mm_and_si128(r0, packedmask26.v);
875  r1 = _mm_and_si128(r1, packedmask25.v);
876  r0 = _mm_add_epi32(r0, _mm_slli_si128(c2, 8));
877  r1 = _mm_add_epi32(r1, c1);
878  s0 = _mm_unpacklo_epi64(_mm_unpacklo_epi32(r0, r1), _mm_unpackhi_epi32(r0, r1));
879  s1 = _mm_add_epi32(s1, _mm_srli_si128(c2, 8));
880  nq[0].v = _mm_unpacklo_epi64(a0, s0);
881  nq[2].v = _mm_unpackhi_epi64(a0, s0);
882  nq[4].v = _mm_unpacklo_epi64(a1, s1);
883  nq[6].v = _mm_unpackhi_epi64(a1, s1);
884  nq[8].v = _mm_unpacklo_epi64(a2, s2);
885  nq[1].v = _mm_shuffle_epi32(nq[0].v, _MM_SHUFFLE(3,3,1,1));
886  nq[3].v = _mm_shuffle_epi32(nq[2].v, _MM_SHUFFLE(3,3,1,1));
887  nq[5].v = _mm_shuffle_epi32(nq[4].v, _MM_SHUFFLE(3,3,1,1));
888  nq[7].v = _mm_shuffle_epi32(nq[6].v, _MM_SHUFFLE(3,3,1,1));
889  nq[9].v = _mm_shuffle_epi32(nq[8].v, _MM_SHUFFLE(3,3,1,1));
890 }
891 
892 
893 /* compute [x+z,x-z] from [x,z] */
894 inline void
895 curve25519_addsub_packed64(packedelem64 *r) {
896  packed32bignum25519 x,z,add,sub;
897 
898  x[0].v = _mm_unpacklo_epi64(r[0].v, r[1].v);
899  z[0].v = _mm_unpackhi_epi64(r[0].v, r[1].v);
900  x[1].v = _mm_unpacklo_epi64(r[2].v, r[3].v);
901  z[1].v = _mm_unpackhi_epi64(r[2].v, r[3].v);
902  x[2].v = _mm_unpacklo_epi64(r[4].v, r[5].v);
903  z[2].v = _mm_unpackhi_epi64(r[4].v, r[5].v);
904  x[3].v = _mm_unpacklo_epi64(r[6].v, r[7].v);
905  z[3].v = _mm_unpackhi_epi64(r[6].v, r[7].v);
906  x[4].v = _mm_unpacklo_epi64(r[8].v, r[9].v);
907  z[4].v = _mm_unpackhi_epi64(r[8].v, r[9].v);
908 
909  curve25519_add_packed32(add, x, z);
910  curve25519_sub_packed32(sub, x, z);
911 
912  r[0].v = _mm_unpacklo_epi64(add[0].v, sub[0].v);
913  r[1].v = _mm_unpackhi_epi64(add[0].v, sub[0].v);
914  r[2].v = _mm_unpacklo_epi64(add[1].v, sub[1].v);
915  r[3].v = _mm_unpackhi_epi64(add[1].v, sub[1].v);
916  r[4].v = _mm_unpacklo_epi64(add[2].v, sub[2].v);
917  r[5].v = _mm_unpackhi_epi64(add[2].v, sub[2].v);
918  r[6].v = _mm_unpacklo_epi64(add[3].v, sub[3].v);
919  r[7].v = _mm_unpackhi_epi64(add[3].v, sub[3].v);
920  r[8].v = _mm_unpacklo_epi64(add[4].v, sub[4].v);
921  r[9].v = _mm_unpackhi_epi64(add[4].v, sub[4].v);
922 }
923 
924 /* compute [x,z] * [121666,121665] */
925 inline void
926 curve25519_121665_packed64(packedelem64 *out, const packedelem64 *in) {
927  xmmi c1,c2;
928 
929  out[0].v = _mm_mul_epu32(in[0].v, packed121666121665.v);
930  out[1].v = _mm_mul_epu32(in[1].v, packed121666121665.v);
931  out[2].v = _mm_mul_epu32(in[2].v, packed121666121665.v);
932  out[3].v = _mm_mul_epu32(in[3].v, packed121666121665.v);
933  out[4].v = _mm_mul_epu32(in[4].v, packed121666121665.v);
934  out[5].v = _mm_mul_epu32(in[5].v, packed121666121665.v);
935  out[6].v = _mm_mul_epu32(in[6].v, packed121666121665.v);
936  out[7].v = _mm_mul_epu32(in[7].v, packed121666121665.v);
937  out[8].v = _mm_mul_epu32(in[8].v, packed121666121665.v);
938  out[9].v = _mm_mul_epu32(in[9].v, packed121666121665.v);
939 
940  c1 = _mm_srli_epi64(out[0].v, 26); c2 = _mm_srli_epi64(out[4].v, 26); out[0].v = _mm_and_si128(out[0].v, packedmask26.v); out[4].v = _mm_and_si128(out[4].v, packedmask26.v); out[1].v = _mm_add_epi64(out[1].v, c1); out[5].v = _mm_add_epi64(out[5].v, c2);
941  c1 = _mm_srli_epi64(out[1].v, 25); c2 = _mm_srli_epi64(out[5].v, 25); out[1].v = _mm_and_si128(out[1].v, packedmask25.v); out[5].v = _mm_and_si128(out[5].v, packedmask25.v); out[2].v = _mm_add_epi64(out[2].v, c1); out[6].v = _mm_add_epi64(out[6].v, c2);
942  c1 = _mm_srli_epi64(out[2].v, 26); c2 = _mm_srli_epi64(out[6].v, 26); out[2].v = _mm_and_si128(out[2].v, packedmask26.v); out[6].v = _mm_and_si128(out[6].v, packedmask26.v); out[3].v = _mm_add_epi64(out[3].v, c1); out[7].v = _mm_add_epi64(out[7].v, c2);
943  c1 = _mm_srli_epi64(out[3].v, 25); c2 = _mm_srli_epi64(out[7].v, 25); out[3].v = _mm_and_si128(out[3].v, packedmask25.v); out[7].v = _mm_and_si128(out[7].v, packedmask25.v); out[4].v = _mm_add_epi64(out[4].v, c1); out[8].v = _mm_add_epi64(out[8].v, c2);
944  c2 = _mm_srli_epi64(out[8].v, 26); out[8].v = _mm_and_si128(out[8].v, packedmask26.v); out[9].v = _mm_add_epi64(out[9].v, c2);
945  c2 = _mm_srli_epi64(out[9].v, 25); out[9].v = _mm_and_si128(out[9].v, packedmask25.v); out[0].v = _mm_add_epi64(out[0].v, _mm_mul_epu32(c2, packednineteen.v));
946  c1 = _mm_srli_epi64(out[0].v, 26); c2 = _mm_srli_epi64(out[4].v, 26); out[0].v = _mm_and_si128(out[0].v, packedmask26.v); out[4].v = _mm_and_si128(out[4].v, packedmask26.v); out[1].v = _mm_add_epi64(out[1].v, c1); out[5].v = _mm_add_epi64(out[5].v, c2);
947 }
948 
949 /* compute [sq.x,sqscalar.x-sqscalar.z] * [sq.z,sq.x-sq.z] */
950 inline void
951 curve25519_final_nq(packedelem64 *nq, const packedelem64 *sq, const packedelem64 *sq121665) {
952  packed32bignum25519 x, z, sub;
953  packed64bignum25519 t, nqa, nqb;
954 
955  x[0].v = _mm_or_si128(_mm_unpacklo_epi64(sq[0].v, sq[1].v), _mm_slli_si128(_mm_unpacklo_epi64(sq121665[0].v, sq121665[1].v), 4));
956  z[0].v = _mm_or_si128(_mm_unpackhi_epi64(sq[0].v, sq[1].v), _mm_slli_si128(_mm_unpackhi_epi64(sq121665[0].v, sq121665[1].v), 4));
957  x[1].v = _mm_or_si128(_mm_unpacklo_epi64(sq[2].v, sq[3].v), _mm_slli_si128(_mm_unpacklo_epi64(sq121665[2].v, sq121665[3].v), 4));
958  z[1].v = _mm_or_si128(_mm_unpackhi_epi64(sq[2].v, sq[3].v), _mm_slli_si128(_mm_unpackhi_epi64(sq121665[2].v, sq121665[3].v), 4));
959  x[2].v = _mm_or_si128(_mm_unpacklo_epi64(sq[4].v, sq[5].v), _mm_slli_si128(_mm_unpacklo_epi64(sq121665[4].v, sq121665[5].v), 4));
960  z[2].v = _mm_or_si128(_mm_unpackhi_epi64(sq[4].v, sq[5].v), _mm_slli_si128(_mm_unpackhi_epi64(sq121665[4].v, sq121665[5].v), 4));
961  x[3].v = _mm_or_si128(_mm_unpacklo_epi64(sq[6].v, sq[7].v), _mm_slli_si128(_mm_unpacklo_epi64(sq121665[6].v, sq121665[7].v), 4));
962  z[3].v = _mm_or_si128(_mm_unpackhi_epi64(sq[6].v, sq[7].v), _mm_slli_si128(_mm_unpackhi_epi64(sq121665[6].v, sq121665[7].v), 4));
963  x[4].v = _mm_or_si128(_mm_unpacklo_epi64(sq[8].v, sq[9].v), _mm_slli_si128(_mm_unpacklo_epi64(sq121665[8].v, sq121665[9].v), 4));
964  z[4].v = _mm_or_si128(_mm_unpackhi_epi64(sq[8].v, sq[9].v), _mm_slli_si128(_mm_unpackhi_epi64(sq121665[8].v, sq121665[9].v), 4));
965 
966  curve25519_sub_packed32(sub, x, z);
967 
968  t[0].v = _mm_shuffle_epi32(sub[0].v, _MM_SHUFFLE(1,1,0,0));
969  t[1].v = _mm_shuffle_epi32(sub[0].v, _MM_SHUFFLE(3,3,2,2));
970  t[2].v = _mm_shuffle_epi32(sub[1].v, _MM_SHUFFLE(1,1,0,0));
971  t[3].v = _mm_shuffle_epi32(sub[1].v, _MM_SHUFFLE(3,3,2,2));
972  t[4].v = _mm_shuffle_epi32(sub[2].v, _MM_SHUFFLE(1,1,0,0));
973  t[5].v = _mm_shuffle_epi32(sub[2].v, _MM_SHUFFLE(3,3,2,2));
974  t[6].v = _mm_shuffle_epi32(sub[3].v, _MM_SHUFFLE(1,1,0,0));
975  t[7].v = _mm_shuffle_epi32(sub[3].v, _MM_SHUFFLE(3,3,2,2));
976  t[8].v = _mm_shuffle_epi32(sub[4].v, _MM_SHUFFLE(1,1,0,0));
977  t[9].v = _mm_shuffle_epi32(sub[4].v, _MM_SHUFFLE(3,3,2,2));
978 
979  nqa[0].v = _mm_unpacklo_epi64(sq[0].v, t[0].v);
980  nqb[0].v = _mm_unpackhi_epi64(sq[0].v, t[0].v);
981  nqa[1].v = _mm_unpacklo_epi64(sq[1].v, t[1].v);
982  nqb[1].v = _mm_unpackhi_epi64(sq[1].v, t[1].v);
983  nqa[2].v = _mm_unpacklo_epi64(sq[2].v, t[2].v);
984  nqb[2].v = _mm_unpackhi_epi64(sq[2].v, t[2].v);
985  nqa[3].v = _mm_unpacklo_epi64(sq[3].v, t[3].v);
986  nqb[3].v = _mm_unpackhi_epi64(sq[3].v, t[3].v);
987  nqa[4].v = _mm_unpacklo_epi64(sq[4].v, t[4].v);
988  nqb[4].v = _mm_unpackhi_epi64(sq[4].v, t[4].v);
989  nqa[5].v = _mm_unpacklo_epi64(sq[5].v, t[5].v);
990  nqb[5].v = _mm_unpackhi_epi64(sq[5].v, t[5].v);
991  nqa[6].v = _mm_unpacklo_epi64(sq[6].v, t[6].v);
992  nqb[6].v = _mm_unpackhi_epi64(sq[6].v, t[6].v);
993  nqa[7].v = _mm_unpacklo_epi64(sq[7].v, t[7].v);
994  nqb[7].v = _mm_unpackhi_epi64(sq[7].v, t[7].v);
995  nqa[8].v = _mm_unpacklo_epi64(sq[8].v, t[8].v);
996  nqb[8].v = _mm_unpackhi_epi64(sq[8].v, t[8].v);
997  nqa[9].v = _mm_unpacklo_epi64(sq[9].v, t[9].v);
998  nqb[9].v = _mm_unpackhi_epi64(sq[9].v, t[9].v);
999 
1000  curve25519_mul_packed64(nq, nqa, nqb);
1001 }
1002 
1003 /*
1004  * In: b = 2^5 - 2^0
1005  * Out: b = 2^250 - 2^0
1006  */
1007 void
1008 curve25519_pow_two5mtwo0_two250mtwo0(bignum25519 b) {
1009  ALIGN(16) bignum25519 t0,c;
1010 
1011  /* 2^5 - 2^0 */ /* b */
1012  /* 2^10 - 2^5 */ curve25519_square_times(t0, b, 5);
1013  /* 2^10 - 2^0 */ curve25519_mul(b, t0, b);
1014  /* 2^20 - 2^10 */ curve25519_square_times(t0, b, 10);
1015  /* 2^20 - 2^0 */ curve25519_mul(c, t0, b);
1016  /* 2^40 - 2^20 */ curve25519_square_times(t0, c, 20);
1017  /* 2^40 - 2^0 */ curve25519_mul(t0, t0, c);
1018  /* 2^50 - 2^10 */ curve25519_square_times(t0, t0, 10);
1019  /* 2^50 - 2^0 */ curve25519_mul(b, t0, b);
1020  /* 2^100 - 2^50 */ curve25519_square_times(t0, b, 50);
1021  /* 2^100 - 2^0 */ curve25519_mul(c, t0, b);
1022  /* 2^200 - 2^100 */ curve25519_square_times(t0, c, 100);
1023  /* 2^200 - 2^0 */ curve25519_mul(t0, t0, c);
1024  /* 2^250 - 2^50 */ curve25519_square_times(t0, t0, 50);
1025  /* 2^250 - 2^0 */ curve25519_mul(b, t0, b);
1026 }
1027 
1028 /*
1029  * z^(p - 2) = z(2^255 - 21)
1030  */
1031 void
1032 curve25519_recip(bignum25519 out, const bignum25519 z) {
1033  ALIGN(16) bignum25519 a, t0, b;
1034 
1035  /* 2 */ curve25519_square(a, z); /* a = 2 */
1036  /* 8 */ curve25519_square_times(t0, a, 2);
1037  /* 9 */ curve25519_mul(b, t0, z); /* b = 9 */
1038  /* 11 */ curve25519_mul(a, b, a); /* a = 11 */
1039  /* 22 */ curve25519_square(t0, a);
1040  /* 2^5 - 2^0 = 31 */ curve25519_mul(b, t0, b);
1041  /* 2^250 - 2^0 */ curve25519_pow_two5mtwo0_two250mtwo0(b);
1042  /* 2^255 - 2^5 */ curve25519_square_times(b, b, 5);
1043  /* 2^255 - 21 */ curve25519_mul(out, b, a);
1044 }
1045 
1046 ANONYMOUS_NAMESPACE_END
1047 
1048 NAMESPACE_BEGIN(CryptoPP)
1049 NAMESPACE_BEGIN(Donna)
1050 
1051 int curve25519_mult_SSE2(byte sharedKey[32], const byte secretKey[32], const byte othersKey[32])
1052 {
1054  for (size_t i = 0;i < 32;++i)
1055  e[i] = secretKey[i];
1056  e[0] &= 0xf8; e[31] &= 0x7f; e[31] |= 0x40;
1057 
1058  ALIGN(16) bignum25519 nqx = {1}, nqpqz = {1}, nqz = {0}, nqpqx, zmone;
1059  packed32bignum25519 qx, qz, pqz, pqx;
1060  packed64bignum25519 nq, sq, sqscalar, prime, primex, primez, nqpq;
1061  bignum25519mulprecomp preq;
1062  size_t i=0, bit=0, lastbit=0;
1063 
1064  curve25519_expand(nqpqx, othersKey);
1065  curve25519_mul_precompute(&preq, nqpqx);
1066 
1067  /* do bits 254..3 */
1068  for (i = 254, lastbit=0; i >= 3; i--) {
1069  bit = (e[i/8] >> (i & 7)) & 1;
1070  curve25519_swap_conditional(nqx, nqpqx, (word32)(bit ^ lastbit));
1071  curve25519_swap_conditional(nqz, nqpqz, (word32)(bit ^ lastbit));
1072  lastbit = bit;
1073 
1074  curve25519_tangle32(qx, nqx, nqpqx); /* qx = [nqx,nqpqx] */
1075  curve25519_tangle32(qz, nqz, nqpqz); /* qz = [nqz,nqpqz] */
1076 
1077  curve25519_add_packed32(pqx, qx, qz); /* pqx = [nqx+nqz,nqpqx+nqpqz] */
1078  curve25519_sub_packed32(pqz, qx, qz); /* pqz = [nqx-nqz,nqpqx-nqpqz] */
1079 
1080  curve25519_make_nqpq(primex, primez, pqx, pqz); /* primex = [nqx+nqz,nqpqx+nqpqz], primez = [nqpqx-nqpqz,nqx-nqz] */
1081  curve25519_mul_packed64(prime, primex, primez); /* prime = [nqx+nqz,nqpqx+nqpqz] * [nqpqx-nqpqz,nqx-nqz] */
1082  curve25519_addsub_packed64(prime); /* prime = [prime.x+prime.z,prime.x-prime.z] */
1083  curve25519_square_packed64(nqpq, prime); /* nqpq = prime^2 */
1084  curve25519_untangle64(nqpqx, nqpqz, nqpq);
1085  curve25519_mul_precomputed(nqpqz, nqpqz, &preq); /* nqpqz = nqpqz * q */
1086 
1087  /* (((sq.x-sq.z)*121665)+sq.x) * (sq.x-sq.z) is equivalent to (sq.x*121666-sq.z*121665) * (sq.x-sq.z) */
1088  curve25519_make_nq(nq, pqx, pqz); /* nq = [nqx+nqz,nqx-nqz] */
1089  curve25519_square_packed64(sq, nq); /* sq = nq^2 */
1090  curve25519_121665_packed64(sqscalar, sq); /* sqscalar = sq * [121666,121665] */
1091  curve25519_final_nq(nq, sq, sqscalar); /* nq = [sq.x,sqscalar.x-sqscalar.z] * [sq.z,sq.x-sq.z] */
1092  curve25519_untangle64(nqx, nqz, nq);
1093  };
1094 
1095  /* it's possible to get rid of this swap with the swap in the above loop
1096  at the bottom instead of the top, but compilers seem to optimize better this way */
1097  curve25519_swap_conditional(nqx, nqpqx, (word32)bit);
1098  curve25519_swap_conditional(nqz, nqpqz, (word32)bit);
1099 
1100  /* do bits 2..0 */
1101  for (i = 0; i < 3; i++) {
1102  curve25519_compute_nq(nq, nqx, nqz);
1103  curve25519_square_packed64(sq, nq); /* sq = nq^2 */
1104  curve25519_121665_packed64(sqscalar, sq); /* sqscalar = sq * [121666,121665] */
1105  curve25519_final_nq(nq, sq, sqscalar); /* nq = [sq.x,sqscalar.x-sqscalar.z] * [sq.z,sq.x-sq.z] */
1106  curve25519_untangle64(nqx, nqz, nq);
1107  }
1108 
1109  curve25519_recip(zmone, nqz);
1110  curve25519_mul(nqz, nqx, zmone);
1111  curve25519_contract(sharedKey, nqz);
1112 
1113  return 0;
1114 }
1115 
1116 NAMESPACE_END // Donna
1117 NAMESPACE_END // CryptoPP
1118 
1119 #endif // CRYPTOPP_CURVE25519_SSE2
Utility functions for the Crypto++ library.
EnumToType< ByteOrder, LITTLE_ENDIAN_ORDER > LittleEndian
Provides a constant for LittleEndian.
Definition: cryptlib.h:150
Library configuration file.
Classes and functions for secure memory allocations.
Precompiled header file.
Fixed size stack-based SecBlock.
Definition: secblock.h:1077
Crypto++ library namespace.