1 // sm4_simd.cpp - written and placed in the public domain by
2 // Markku-Juhani O. Saarinen and Jeffrey Walton
3 //
4 // This source file uses intrinsics and built-ins to gain access to
5 // AESNI, ARM NEON and ARMv8a, and Power7 Altivec instructions. A separate
6 // source file is needed because additional CXXFLAGS are required to enable
7 // the appropriate instructions sets in some build configurations.
8 //
9 // AES-NI based on Markku-Juhani O. Saarinen work at https://github.com/mjosaarinen/sm4ni.
10 //
11 // ARMv8 is upcoming.
12
13 #include "pch.h"
14 #include "config.h"
15
16 #include "sm4.h"
17 #include "misc.h"
18
19 // Uncomment for benchmarking C++ against SSE.
20 // Do so in both simon.cpp and simon_simd.cpp.
21 // #undef CRYPTOPP_AESNI_AVAILABLE
22
23 #if (CRYPTOPP_AESNI_AVAILABLE)
24 # include "adv_simd.h"
25 # include <emmintrin.h>
26 # include <tmmintrin.h>
27 # include <wmmintrin.h>
28 #endif
29
30 // Squash MS LNK4221 and libtool warnings
31 extern const char SM4_SIMD_FNAME[] = __FILE__;
32
33 ANONYMOUS_NAMESPACE_BEGIN
34
35 using CryptoPP::word32;
36
37 #if (CRYPTOPP_AESNI_AVAILABLE)
38
39 template <unsigned int R>
ShiftLeft(const __m128i & val)40 inline __m128i ShiftLeft(const __m128i& val)
41 {
42 return _mm_slli_epi32(val, R);
43 }
44
45 template <unsigned int R>
ShiftRight(const __m128i & val)46 inline __m128i ShiftRight(const __m128i& val)
47 {
48 return _mm_srli_epi32(val, R);
49 }
50
51 template <unsigned int R>
ShiftLeft64(const __m128i & val)52 inline __m128i ShiftLeft64(const __m128i& val)
53 {
54 return _mm_slli_epi64(val, R);
55 }
56
57 template <unsigned int R>
ShiftRight64(const __m128i & val)58 inline __m128i ShiftRight64(const __m128i& val)
59 {
60 return _mm_srli_epi64(val, R);
61 }
62
63 template <unsigned int R>
RotateLeft(const __m128i & val)64 inline __m128i RotateLeft(const __m128i& val)
65 {
66 return _mm_or_si128(
67 _mm_slli_epi32(val, R), _mm_srli_epi32(val, 32-R));
68 }
69
70 template <unsigned int R>
RotateRight(const __m128i & val)71 inline __m128i RotateRight(const __m128i& val)
72 {
73 return _mm_or_si128(
74 _mm_slli_epi32(val, 32-R), _mm_srli_epi32(val, R));
75 }
76
77 template <>
RotateLeft(const __m128i & val)78 inline __m128i RotateLeft<8>(const __m128i& val)
79 {
80 const __m128i r08 = _mm_set_epi32(0x0E0D0C0F, 0x0A09080B, 0x06050407, 0x02010003);
81 return _mm_shuffle_epi8(val, r08);
82 }
83
84 template <>
RotateLeft(const __m128i & val)85 inline __m128i RotateLeft<16>(const __m128i& val)
86 {
87 const __m128i mask = _mm_set_epi32(0x0D0C0F0E, 0x09080B0A, 0x05040706, 0x01000302);
88 return _mm_shuffle_epi8(val, mask);
89 }
90
91 template <>
RotateLeft(const __m128i & val)92 inline __m128i RotateLeft<24>(const __m128i& val)
93 {
94 const __m128i mask = _mm_set_epi32(0x0C0F0E0D, 0x080B0A09, 0x04070605, 0x00030201);
95 return _mm_shuffle_epi8(val, mask);
96 }
97
98 /// \brief Unpack XMM words
99 /// \tparam IDX the element from each XMM word
100 /// \param a the first XMM word
101 /// \param b the second XMM word
102 /// \param c the third XMM word
103 /// \param d the fourth XMM word
104 /// \details UnpackXMM selects the IDX element from a, b, c, d and returns a concatenation
105 /// equivalent to <tt>a[IDX] || b[IDX] || c[IDX] || d[IDX]</tt>.
106 template <unsigned int IDX>
UnpackXMM(const __m128i & a,const __m128i & b,const __m128i & c,const __m128i & d)107 inline __m128i UnpackXMM(const __m128i& a, const __m128i& b, const __m128i& c, const __m128i& d)
108 {
109 // Should not be instantiated
110 CRYPTOPP_UNUSED(a); CRYPTOPP_UNUSED(b);
111 CRYPTOPP_UNUSED(c); CRYPTOPP_UNUSED(d);
112 CRYPTOPP_ASSERT(0);
113 return _mm_setzero_si128();
114 }
115
116 template <>
UnpackXMM(const __m128i & a,const __m128i & b,const __m128i & c,const __m128i & d)117 inline __m128i UnpackXMM<0>(const __m128i& a, const __m128i& b, const __m128i& c, const __m128i& d)
118 {
119 const __m128i r1 = _mm_unpacklo_epi32(a, b);
120 const __m128i r2 = _mm_unpacklo_epi32(c, d);
121 return _mm_unpacklo_epi64(r1, r2);
122 }
123
124 template <>
UnpackXMM(const __m128i & a,const __m128i & b,const __m128i & c,const __m128i & d)125 inline __m128i UnpackXMM<1>(const __m128i& a, const __m128i& b, const __m128i& c, const __m128i& d)
126 {
127 const __m128i r1 = _mm_unpacklo_epi32(a, b);
128 const __m128i r2 = _mm_unpacklo_epi32(c, d);
129 return _mm_unpackhi_epi64(r1, r2);
130 }
131
132 template <>
UnpackXMM(const __m128i & a,const __m128i & b,const __m128i & c,const __m128i & d)133 inline __m128i UnpackXMM<2>(const __m128i& a, const __m128i& b, const __m128i& c, const __m128i& d)
134 {
135 const __m128i r1 = _mm_unpackhi_epi32(a, b);
136 const __m128i r2 = _mm_unpackhi_epi32(c, d);
137 return _mm_unpacklo_epi64(r1, r2);
138 }
139
140 template <>
UnpackXMM(const __m128i & a,const __m128i & b,const __m128i & c,const __m128i & d)141 inline __m128i UnpackXMM<3>(const __m128i& a, const __m128i& b, const __m128i& c, const __m128i& d)
142 {
143 const __m128i r1 = _mm_unpackhi_epi32(a, b);
144 const __m128i r2 = _mm_unpackhi_epi32(c, d);
145 return _mm_unpackhi_epi64(r1, r2);
146 }
147
148 /// \brief Unpack a XMM word
149 /// \tparam IDX the element from each XMM word
150 /// \param v the first XMM word
151 /// \details UnpackXMM selects the IDX element from v and returns a concatenation
152 /// equivalent to <tt>v[IDX] || v[IDX] || v[IDX] || v[IDX]</tt>.
153 template <unsigned int IDX>
UnpackXMM(const __m128i & v)154 inline __m128i UnpackXMM(const __m128i& v)
155 {
156 // Should not be instantiated
157 CRYPTOPP_UNUSED(v); CRYPTOPP_ASSERT(0);
158 return _mm_setzero_si128();
159 }
160
161 template <>
UnpackXMM(const __m128i & v)162 inline __m128i UnpackXMM<0>(const __m128i& v)
163 {
164 // Splat to all lanes
165 return _mm_shuffle_epi8(v, _mm_set_epi8(3,2,1,0, 3,2,1,0, 3,2,1,0, 3,2,1,0));
166 }
167
168 template <>
UnpackXMM(const __m128i & v)169 inline __m128i UnpackXMM<1>(const __m128i& v)
170 {
171 // Splat to all lanes
172 return _mm_shuffle_epi8(v, _mm_set_epi8(7,6,5,4, 7,6,5,4, 7,6,5,4, 7,6,5,4));
173 }
174
175 template <>
UnpackXMM(const __m128i & v)176 inline __m128i UnpackXMM<2>(const __m128i& v)
177 {
178 // Splat to all lanes
179 return _mm_shuffle_epi8(v, _mm_set_epi8(11,10,9,8, 11,10,9,8, 11,10,9,8, 11,10,9,8));
180 }
181
182 template <>
UnpackXMM(const __m128i & v)183 inline __m128i UnpackXMM<3>(const __m128i& v)
184 {
185 // Splat to all lanes
186 return _mm_shuffle_epi8(v, _mm_set_epi8(15,14,13,12, 15,14,13,12, 15,14,13,12, 15,14,13,12));
187 }
188
189 template <unsigned int IDX>
RepackXMM(const __m128i & a,const __m128i & b,const __m128i & c,const __m128i & d)190 inline __m128i RepackXMM(const __m128i& a, const __m128i& b, const __m128i& c, const __m128i& d)
191 {
192 return UnpackXMM<IDX>(a, b, c, d);
193 }
194
195 template <unsigned int IDX>
RepackXMM(const __m128i & v)196 inline __m128i RepackXMM(const __m128i& v)
197 {
198 return UnpackXMM<IDX>(v);
199 }
200
SM4_Encrypt(__m128i & block0,__m128i & block1,__m128i & block2,__m128i & block3,const word32 * subkeys)201 inline void SM4_Encrypt(__m128i &block0, __m128i &block1,
202 __m128i &block2, __m128i &block3, const word32 *subkeys)
203 {
204 // nibble mask
205 const __m128i c0f = _mm_set_epi32(0x0F0F0F0F, 0x0F0F0F0F, 0x0F0F0F0F, 0x0F0F0F0F);
206
207 // flip all bytes in all 32-bit words
208 const __m128i flp = _mm_set_epi32(0x0C0D0E0F, 0x08090A0B, 0x04050607, 0x00010203);
209
210 // inverse shift rows
211 const __m128i shr = _mm_set_epi32(0x0306090C, 0x0F020508, 0x0B0E0104, 0x070A0D00);
212
213 // Affine transform 1 (low and high hibbles)
214 const __m128i m1l = _mm_set_epi32(0xC7C1B4B2, 0x22245157, 0x9197E2E4, 0x74720701);
215 const __m128i m1h = _mm_set_epi32(0xF052B91B, 0xF95BB012, 0xE240AB09, 0xEB49A200);
216
217 // Affine transform 2 (low and high hibbles)
218 const __m128i m2l = _mm_set_epi32(0xEDD14478, 0x172BBE82, 0x5B67F2CE, 0xA19D0834);
219 const __m128i m2h = _mm_set_epi32(0x11CDBE62, 0xCC1063BF, 0xAE7201DD, 0x73AFDC00);
220
221 __m128i t0 = UnpackXMM<0>(block0, block1, block2, block3);
222 __m128i t1 = UnpackXMM<1>(block0, block1, block2, block3);
223 __m128i t2 = UnpackXMM<2>(block0, block1, block2, block3);
224 __m128i t3 = UnpackXMM<3>(block0, block1, block2, block3);
225
226 t0 = _mm_shuffle_epi8(t0, flp);
227 t1 = _mm_shuffle_epi8(t1, flp);
228 t2 = _mm_shuffle_epi8(t2, flp);
229 t3 = _mm_shuffle_epi8(t3, flp);
230
231 const unsigned int ROUNDS = 32;
232 for (unsigned int i = 0; i < ROUNDS; i++)
233 {
234 const __m128i k = _mm_shuffle_epi32(_mm_castps_si128(
235 _mm_load_ss((const float*)(subkeys+i))), _MM_SHUFFLE(0,0,0,0));
236
237 __m128i x, y;
238 x = _mm_xor_si128(t1, _mm_xor_si128(t2, _mm_xor_si128(t3, k)));
239
240 y = _mm_and_si128(x, c0f); // inner affine
241 y = _mm_shuffle_epi8(m1l, y);
242 x = _mm_and_si128(ShiftRight64<4>(x), c0f);
243 x = _mm_xor_si128(_mm_shuffle_epi8(m1h, x), y);
244
245 x = _mm_shuffle_epi8(x, shr); // inverse MixColumns
246 x = _mm_aesenclast_si128(x, c0f); // AESNI instruction
247
248 y = _mm_andnot_si128(x, c0f); // outer affine
249 y = _mm_shuffle_epi8(m2l, y);
250 x = _mm_and_si128(ShiftRight64<4>(x), c0f);
251 x = _mm_xor_si128(_mm_shuffle_epi8(m2h, x), y);
252
253 // 4 parallel L1 linear transforms
254 y = _mm_xor_si128(x, RotateLeft<8>(x));
255 y = _mm_xor_si128(y, RotateLeft<16>(x));
256 y = _mm_xor_si128(ShiftLeft<2>(y), ShiftRight<30>(y));
257 x = _mm_xor_si128(x, _mm_xor_si128(y, RotateLeft<24>(x)));
258
259 // rotate registers
260 x = _mm_xor_si128(x, t0);
261 t0 = t1; t1 = t2;
262 t2 = t3; t3 = x;
263 }
264
265 t0 = _mm_shuffle_epi8(t0, flp);
266 t1 = _mm_shuffle_epi8(t1, flp);
267 t2 = _mm_shuffle_epi8(t2, flp);
268 t3 = _mm_shuffle_epi8(t3, flp);
269
270 block0 = RepackXMM<0>(t3,t2,t1,t0);
271 block1 = RepackXMM<1>(t3,t2,t1,t0);
272 block2 = RepackXMM<2>(t3,t2,t1,t0);
273 block3 = RepackXMM<3>(t3,t2,t1,t0);
274 }
275
SM4_Enc_4_Blocks(__m128i & block0,__m128i & block1,__m128i & block2,__m128i & block3,const word32 * subkeys,unsigned int)276 inline void SM4_Enc_4_Blocks(__m128i &block0, __m128i &block1,
277 __m128i &block2, __m128i &block3, const word32 *subkeys, unsigned int /*rounds*/)
278 {
279 SM4_Encrypt(block0, block1, block2, block3, subkeys);
280 }
281
SM4_Dec_4_Blocks(__m128i & block0,__m128i & block1,__m128i & block2,__m128i & block3,const word32 * subkeys,unsigned int)282 inline void SM4_Dec_4_Blocks(__m128i &block0, __m128i &block1,
283 __m128i &block2, __m128i &block3, const word32 *subkeys, unsigned int /*rounds*/)
284 {
285 SM4_Encrypt(block0, block1, block2, block3, subkeys);
286 }
287
SM4_Enc_Block(__m128i & block0,const word32 * subkeys,unsigned int)288 inline void SM4_Enc_Block(__m128i &block0,
289 const word32 *subkeys, unsigned int /*rounds*/)
290 {
291 __m128i t1 = _mm_setzero_si128();
292 __m128i t2 = _mm_setzero_si128();
293 __m128i t3 = _mm_setzero_si128();
294
295 SM4_Encrypt(block0, t1, t2, t3, subkeys);
296 }
297
SM4_Dec_Block(__m128i & block0,const word32 * subkeys,unsigned int)298 inline void SM4_Dec_Block(__m128i &block0,
299 const word32 *subkeys, unsigned int /*rounds*/)
300 {
301 __m128i t1 = _mm_setzero_si128();
302 __m128i t2 = _mm_setzero_si128();
303 __m128i t3 = _mm_setzero_si128();
304
305 SM4_Encrypt(block0, t1, t2, t3, subkeys);
306 }
307
308 #endif // CRYPTOPP_AESNI_AVAILABLE
309
310 ANONYMOUS_NAMESPACE_END
311
NAMESPACE_BEGIN(CryptoPP)312 NAMESPACE_BEGIN(CryptoPP)
313
314 #if defined(CRYPTOPP_AESNI_AVAILABLE)
315 size_t SM4_Enc_AdvancedProcessBlocks_AESNI(const word32* subKeys, size_t rounds,
316 const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags)
317 {
318 return AdvancedProcessBlocks128_4x1_SSE(SM4_Enc_Block, SM4_Enc_4_Blocks,
319 subKeys, rounds, inBlocks, xorBlocks, outBlocks, length, flags);
320 }
321 #endif // CRYPTOPP_AESNI_AVAILABLE
322
323 NAMESPACE_END
324