xref: /freebsd/contrib/bearssl/src/hash/ghash_ctmul.c (revision cab6a39d7b343596a5823e65c0f7b426551ec22d)
1 /*
2  * Copyright (c) 2016 Thomas Pornin <pornin@bolet.org>
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining
5  * a copy of this software and associated documentation files (the
6  * "Software"), to deal in the Software without restriction, including
7  * without limitation the rights to use, copy, modify, merge, publish,
8  * distribute, sublicense, and/or sell copies of the Software, and to
9  * permit persons to whom the Software is furnished to do so, subject to
10  * the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be
13  * included in all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
16  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
17  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
18  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
19  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
20  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
21  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 
25 #include "inner.h"
26 
27 /*
28  * We compute "carryless multiplications" through normal integer
29  * multiplications, masking out enough bits to create "holes" in which
30  * carries may expand without altering our bits; we really use 8 data
31  * bits per 32-bit word, spaced every fourth bit. Accumulated carries
32  * may not exceed 8 in total, which fits in 4 bits.
33  *
34  * It would be possible to use a 3-bit spacing, allowing two operands,
35  * one with 7 non-zero data bits, the other one with 10 or 11 non-zero
36  * data bits; this asymmetric splitting makes the overall code more
37  * complex with thresholds and exceptions, and does not appear to be
38  * worth the effort.
39  */
40 
41 /*
42  * We cannot really autodetect whether multiplications are "slow" or
43  * not. A typical example is the ARM Cortex M0+, which exists in two
44  * versions: one with a 1-cycle multiplication opcode, the other with
45  * a 32-cycle multiplication opcode. They both use exactly the same
46  * architecture and ABI, and cannot be distinguished from each other
47  * at compile-time.
48  *
49  * Since most modern CPU (even embedded CPU) still have fast
50  * multiplications, we use the "fast mul" code by default.
51  */
52 
53 #if BR_SLOW_MUL
54 
55 /*
56  * This implementation uses Karatsuba-like reduction to make fewer
57  * integer multiplications (9 instead of 16), at the expense of extra
58  * logical operations (XOR, shifts...). On modern x86 CPU that offer
59  * fast, pipelined multiplications, this code is about twice slower than
60  * the simpler code with 16 multiplications. This tendency may be
61  * reversed on low-end platforms with expensive multiplications.
62  */
63 
64 #define MUL32(h, l, x, y)   do { \
65 		uint64_t mul32tmp = MUL(x, y); \
66 		(h) = (uint32_t)(mul32tmp >> 32); \
67 		(l) = (uint32_t)mul32tmp; \
68 	} while (0)
69 
70 static inline void
71 bmul(uint32_t *hi, uint32_t *lo, uint32_t x, uint32_t y)
72 {
73 	uint32_t x0, x1, x2, x3;
74 	uint32_t y0, y1, y2, y3;
75 	uint32_t a0, a1, a2, a3, a4, a5, a6, a7, a8;
76 	uint32_t b0, b1, b2, b3, b4, b5, b6, b7, b8;
77 
78 	x0 = x & (uint32_t)0x11111111;
79 	x1 = x & (uint32_t)0x22222222;
80 	x2 = x & (uint32_t)0x44444444;
81 	x3 = x & (uint32_t)0x88888888;
82 	y0 = y & (uint32_t)0x11111111;
83 	y1 = y & (uint32_t)0x22222222;
84 	y2 = y & (uint32_t)0x44444444;
85 	y3 = y & (uint32_t)0x88888888;
86 
87 	/*
88 	 * (x0+W*x1)*(y0+W*y1) -> a0:b0
89 	 * (x2+W*x3)*(y2+W*y3) -> a3:b3
90 	 * ((x0+x2)+W*(x1+x3))*((y0+y2)+W*(y1+y3)) -> a6:b6
91 	 */
92 	a0 = x0;
93 	b0 = y0;
94 	a1 = x1 >> 1;
95 	b1 = y1 >> 1;
96 	a2 = a0 ^ a1;
97 	b2 = b0 ^ b1;
98 	a3 = x2 >> 2;
99 	b3 = y2 >> 2;
100 	a4 = x3 >> 3;
101 	b4 = y3 >> 3;
102 	a5 = a3 ^ a4;
103 	b5 = b3 ^ b4;
104 	a6 = a0 ^ a3;
105 	b6 = b0 ^ b3;
106 	a7 = a1 ^ a4;
107 	b7 = b1 ^ b4;
108 	a8 = a6 ^ a7;
109 	b8 = b6 ^ b7;
110 
111 	MUL32(b0, a0, b0, a0);
112 	MUL32(b1, a1, b1, a1);
113 	MUL32(b2, a2, b2, a2);
114 	MUL32(b3, a3, b3, a3);
115 	MUL32(b4, a4, b4, a4);
116 	MUL32(b5, a5, b5, a5);
117 	MUL32(b6, a6, b6, a6);
118 	MUL32(b7, a7, b7, a7);
119 	MUL32(b8, a8, b8, a8);
120 
121 	a0 &= (uint32_t)0x11111111;
122 	a1 &= (uint32_t)0x11111111;
123 	a2 &= (uint32_t)0x11111111;
124 	a3 &= (uint32_t)0x11111111;
125 	a4 &= (uint32_t)0x11111111;
126 	a5 &= (uint32_t)0x11111111;
127 	a6 &= (uint32_t)0x11111111;
128 	a7 &= (uint32_t)0x11111111;
129 	a8 &= (uint32_t)0x11111111;
130 	b0 &= (uint32_t)0x11111111;
131 	b1 &= (uint32_t)0x11111111;
132 	b2 &= (uint32_t)0x11111111;
133 	b3 &= (uint32_t)0x11111111;
134 	b4 &= (uint32_t)0x11111111;
135 	b5 &= (uint32_t)0x11111111;
136 	b6 &= (uint32_t)0x11111111;
137 	b7 &= (uint32_t)0x11111111;
138 	b8 &= (uint32_t)0x11111111;
139 
140 	a2 ^= a0 ^ a1;
141 	b2 ^= b0 ^ b1;
142 	a0 ^= (a2 << 1) ^ (a1 << 2);
143 	b0 ^= (b2 << 1) ^ (b1 << 2);
144 	a5 ^= a3 ^ a4;
145 	b5 ^= b3 ^ b4;
146 	a3 ^= (a5 << 1) ^ (a4 << 2);
147 	b3 ^= (b5 << 1) ^ (b4 << 2);
148 	a8 ^= a6 ^ a7;
149 	b8 ^= b6 ^ b7;
150 	a6 ^= (a8 << 1) ^ (a7 << 2);
151 	b6 ^= (b8 << 1) ^ (b7 << 2);
152 	a6 ^= a0 ^ a3;
153 	b6 ^= b0 ^ b3;
154 	*lo = a0 ^ (a6 << 2) ^ (a3 << 4);
155 	*hi = b0 ^ (b6 << 2) ^ (b3 << 4) ^ (a6 >> 30) ^ (a3 >> 28);
156 }
157 
158 #else
159 
160 /*
161  * Simple multiplication in GF(2)[X], using 16 integer multiplications.
162  */
163 
164 static inline void
165 bmul(uint32_t *hi, uint32_t *lo, uint32_t x, uint32_t y)
166 {
167 	uint32_t x0, x1, x2, x3;
168 	uint32_t y0, y1, y2, y3;
169 	uint64_t z0, z1, z2, z3;
170 	uint64_t z;
171 
172 	x0 = x & (uint32_t)0x11111111;
173 	x1 = x & (uint32_t)0x22222222;
174 	x2 = x & (uint32_t)0x44444444;
175 	x3 = x & (uint32_t)0x88888888;
176 	y0 = y & (uint32_t)0x11111111;
177 	y1 = y & (uint32_t)0x22222222;
178 	y2 = y & (uint32_t)0x44444444;
179 	y3 = y & (uint32_t)0x88888888;
180 	z0 = MUL(x0, y0) ^ MUL(x1, y3) ^ MUL(x2, y2) ^ MUL(x3, y1);
181 	z1 = MUL(x0, y1) ^ MUL(x1, y0) ^ MUL(x2, y3) ^ MUL(x3, y2);
182 	z2 = MUL(x0, y2) ^ MUL(x1, y1) ^ MUL(x2, y0) ^ MUL(x3, y3);
183 	z3 = MUL(x0, y3) ^ MUL(x1, y2) ^ MUL(x2, y1) ^ MUL(x3, y0);
184 	z0 &= (uint64_t)0x1111111111111111;
185 	z1 &= (uint64_t)0x2222222222222222;
186 	z2 &= (uint64_t)0x4444444444444444;
187 	z3 &= (uint64_t)0x8888888888888888;
188 	z = z0 | z1 | z2 | z3;
189 	*lo = (uint32_t)z;
190 	*hi = (uint32_t)(z >> 32);
191 }
192 
193 #endif
194 
195 /* see bearssl_hash.h */
196 void
197 br_ghash_ctmul(void *y, const void *h, const void *data, size_t len)
198 {
199 	const unsigned char *buf, *hb;
200 	unsigned char *yb;
201 	uint32_t yw[4];
202 	uint32_t hw[4];
203 
204 	/*
205 	 * Throughout the loop we handle the y and h values as arrays
206 	 * of 32-bit words.
207 	 */
208 	buf = data;
209 	yb = y;
210 	hb = h;
211 	yw[3] = br_dec32be(yb);
212 	yw[2] = br_dec32be(yb + 4);
213 	yw[1] = br_dec32be(yb + 8);
214 	yw[0] = br_dec32be(yb + 12);
215 	hw[3] = br_dec32be(hb);
216 	hw[2] = br_dec32be(hb + 4);
217 	hw[1] = br_dec32be(hb + 8);
218 	hw[0] = br_dec32be(hb + 12);
219 	while (len > 0) {
220 		const unsigned char *src;
221 		unsigned char tmp[16];
222 		int i;
223 		uint32_t a[9], b[9], zw[8];
224 		uint32_t c0, c1, c2, c3, d0, d1, d2, d3, e0, e1, e2, e3;
225 
226 		/*
227 		 * Get the next 16-byte block (using zero-padding if
228 		 * necessary).
229 		 */
230 		if (len >= 16) {
231 			src = buf;
232 			buf += 16;
233 			len -= 16;
234 		} else {
235 			memcpy(tmp, buf, len);
236 			memset(tmp + len, 0, (sizeof tmp) - len);
237 			src = tmp;
238 			len = 0;
239 		}
240 
241 		/*
242 		 * Decode the block. The GHASH standard mandates
243 		 * big-endian encoding.
244 		 */
245 		yw[3] ^= br_dec32be(src);
246 		yw[2] ^= br_dec32be(src + 4);
247 		yw[1] ^= br_dec32be(src + 8);
248 		yw[0] ^= br_dec32be(src + 12);
249 
250 		/*
251 		 * We multiply two 128-bit field elements. We use
252 		 * Karatsuba to turn that into three 64-bit
253 		 * multiplications, which are themselves done with a
254 		 * total of nine 32-bit multiplications.
255 		 */
256 
257 		/*
258 		 * y[0,1]*h[0,1] -> 0..2
259 		 * y[2,3]*h[2,3] -> 3..5
260 		 * (y[0,1]+y[2,3])*(h[0,1]+h[2,3]) -> 6..8
261 		 */
262 		a[0] = yw[0];
263 		b[0] = hw[0];
264 		a[1] = yw[1];
265 		b[1] = hw[1];
266 		a[2] = a[0] ^ a[1];
267 		b[2] = b[0] ^ b[1];
268 
269 		a[3] = yw[2];
270 		b[3] = hw[2];
271 		a[4] = yw[3];
272 		b[4] = hw[3];
273 		a[5] = a[3] ^ a[4];
274 		b[5] = b[3] ^ b[4];
275 
276 		a[6] = a[0] ^ a[3];
277 		b[6] = b[0] ^ b[3];
278 		a[7] = a[1] ^ a[4];
279 		b[7] = b[1] ^ b[4];
280 		a[8] = a[6] ^ a[7];
281 		b[8] = b[6] ^ b[7];
282 
283 		for (i = 0; i < 9; i ++) {
284 			bmul(&b[i], &a[i], b[i], a[i]);
285 		}
286 
287 		c0 = a[0];
288 		c1 = b[0] ^ a[2] ^ a[0] ^ a[1];
289 		c2 = a[1] ^ b[2] ^ b[0] ^ b[1];
290 		c3 = b[1];
291 		d0 = a[3];
292 		d1 = b[3] ^ a[5] ^ a[3] ^ a[4];
293 		d2 = a[4] ^ b[5] ^ b[3] ^ b[4];
294 		d3 = b[4];
295 		e0 = a[6];
296 		e1 = b[6] ^ a[8] ^ a[6] ^ a[7];
297 		e2 = a[7] ^ b[8] ^ b[6] ^ b[7];
298 		e3 = b[7];
299 
300 		e0 ^= c0 ^ d0;
301 		e1 ^= c1 ^ d1;
302 		e2 ^= c2 ^ d2;
303 		e3 ^= c3 ^ d3;
304 		c2 ^= e0;
305 		c3 ^= e1;
306 		d0 ^= e2;
307 		d1 ^= e3;
308 
309 		/*
310 		 * GHASH specification has the bits "reversed" (most
311 		 * significant is in fact least significant), which does
312 		 * not matter for a carryless multiplication, except that
313 		 * the 255-bit result must be shifted by 1 bit.
314 		 */
315 		zw[0] = c0 << 1;
316 		zw[1] = (c1 << 1) | (c0 >> 31);
317 		zw[2] = (c2 << 1) | (c1 >> 31);
318 		zw[3] = (c3 << 1) | (c2 >> 31);
319 		zw[4] = (d0 << 1) | (c3 >> 31);
320 		zw[5] = (d1 << 1) | (d0 >> 31);
321 		zw[6] = (d2 << 1) | (d1 >> 31);
322 		zw[7] = (d3 << 1) | (d2 >> 31);
323 
324 		/*
325 		 * We now do the reduction modulo the field polynomial
326 		 * to get back to 128 bits.
327 		 */
328 		for (i = 0; i < 4; i ++) {
329 			uint32_t lw;
330 
331 			lw = zw[i];
332 			zw[i + 4] ^= lw ^ (lw >> 1) ^ (lw >> 2) ^ (lw >> 7);
333 			zw[i + 3] ^= (lw << 31) ^ (lw << 30) ^ (lw << 25);
334 		}
335 		memcpy(yw, zw + 4, sizeof yw);
336 	}
337 
338 	/*
339 	 * Encode back the result.
340 	 */
341 	br_enc32be(yb, yw[3]);
342 	br_enc32be(yb + 4, yw[2]);
343 	br_enc32be(yb + 8, yw[1]);
344 	br_enc32be(yb + 12, yw[0]);
345 }
346