xref: /freebsd/contrib/bearssl/src/hash/ghash_ctmul32.c (revision af23369a6deaaeb612ab266eb88b8bb8d560c322)
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  * This implementation uses 32-bit multiplications, and only the low
29  * 32 bits for each multiplication result. This is meant primarily for
30  * the ARM Cortex M0 and M0+, whose multiplication opcode does not yield
31  * the upper 32 bits; but it might also be useful on architectures where
32  * access to the upper 32 bits requires use of specific registers that
33  * create contention (e.g. on i386, "mul" necessarily outputs the result
34  * in edx:eax, while "imul" can use any registers but is limited to the
35  * low 32 bits).
36  *
37  * The implementation trick that is used here is bit-reversing (bit 0
38  * is swapped with bit 31, bit 1 with bit 30, and so on). In GF(2)[X],
39  * for all values x and y, we have:
40  *    rev32(x) * rev32(y) = rev64(x * y)
41  * In other words, if we bit-reverse (over 32 bits) the operands, then we
42  * bit-reverse (over 64 bits) the result.
43  */
44 
45 /*
46  * Multiplication in GF(2)[X], truncated to its low 32 bits.
47  */
48 static inline uint32_t
49 bmul32(uint32_t x, uint32_t y)
50 {
51 	uint32_t x0, x1, x2, x3;
52 	uint32_t y0, y1, y2, y3;
53 	uint32_t z0, z1, z2, z3;
54 
55 	x0 = x & (uint32_t)0x11111111;
56 	x1 = x & (uint32_t)0x22222222;
57 	x2 = x & (uint32_t)0x44444444;
58 	x3 = x & (uint32_t)0x88888888;
59 	y0 = y & (uint32_t)0x11111111;
60 	y1 = y & (uint32_t)0x22222222;
61 	y2 = y & (uint32_t)0x44444444;
62 	y3 = y & (uint32_t)0x88888888;
63 	z0 = (x0 * y0) ^ (x1 * y3) ^ (x2 * y2) ^ (x3 * y1);
64 	z1 = (x0 * y1) ^ (x1 * y0) ^ (x2 * y3) ^ (x3 * y2);
65 	z2 = (x0 * y2) ^ (x1 * y1) ^ (x2 * y0) ^ (x3 * y3);
66 	z3 = (x0 * y3) ^ (x1 * y2) ^ (x2 * y1) ^ (x3 * y0);
67 	z0 &= (uint32_t)0x11111111;
68 	z1 &= (uint32_t)0x22222222;
69 	z2 &= (uint32_t)0x44444444;
70 	z3 &= (uint32_t)0x88888888;
71 	return z0 | z1 | z2 | z3;
72 }
73 
74 /*
75  * Bit-reverse a 32-bit word.
76  */
77 static uint32_t
78 rev32(uint32_t x)
79 {
80 #define RMS(m, s)   do { \
81 		x = ((x & (uint32_t)(m)) << (s)) \
82 			| ((x >> (s)) & (uint32_t)(m)); \
83 	} while (0)
84 
85 	RMS(0x55555555, 1);
86 	RMS(0x33333333, 2);
87 	RMS(0x0F0F0F0F, 4);
88 	RMS(0x00FF00FF, 8);
89 	return (x << 16) | (x >> 16);
90 
91 #undef RMS
92 }
93 
94 /* see bearssl_hash.h */
95 void
96 br_ghash_ctmul32(void *y, const void *h, const void *data, size_t len)
97 {
98 	/*
99 	 * This implementation is similar to br_ghash_ctmul() except
100 	 * that we have to do the multiplication twice, with the
101 	 * "normal" and "bit reversed" operands. Hence we end up with
102 	 * eighteen 32-bit multiplications instead of nine.
103 	 */
104 
105 	const unsigned char *buf, *hb;
106 	unsigned char *yb;
107 	uint32_t yw[4];
108 	uint32_t hw[4], hwr[4];
109 
110 	buf = data;
111 	yb = y;
112 	hb = h;
113 	yw[3] = br_dec32be(yb);
114 	yw[2] = br_dec32be(yb + 4);
115 	yw[1] = br_dec32be(yb + 8);
116 	yw[0] = br_dec32be(yb + 12);
117 	hw[3] = br_dec32be(hb);
118 	hw[2] = br_dec32be(hb + 4);
119 	hw[1] = br_dec32be(hb + 8);
120 	hw[0] = br_dec32be(hb + 12);
121 	hwr[3] = rev32(hw[3]);
122 	hwr[2] = rev32(hw[2]);
123 	hwr[1] = rev32(hw[1]);
124 	hwr[0] = rev32(hw[0]);
125 	while (len > 0) {
126 		const unsigned char *src;
127 		unsigned char tmp[16];
128 		int i;
129 		uint32_t a[18], b[18], c[18];
130 		uint32_t d0, d1, d2, d3, d4, d5, d6, d7;
131 		uint32_t zw[8];
132 
133 		if (len >= 16) {
134 			src = buf;
135 			buf += 16;
136 			len -= 16;
137 		} else {
138 			memcpy(tmp, buf, len);
139 			memset(tmp + len, 0, (sizeof tmp) - len);
140 			src = tmp;
141 			len = 0;
142 		}
143 		yw[3] ^= br_dec32be(src);
144 		yw[2] ^= br_dec32be(src + 4);
145 		yw[1] ^= br_dec32be(src + 8);
146 		yw[0] ^= br_dec32be(src + 12);
147 
148 		/*
149 		 * We are using Karatsuba: the 128x128 multiplication is
150 		 * reduced to three 64x64 multiplications, hence nine
151 		 * 32x32 multiplications. With the bit-reversal trick,
152 		 * we have to perform 18 32x32 multiplications.
153 		 */
154 
155 		/*
156 		 * y[0,1]*h[0,1] -> 0,1,4
157 		 * y[2,3]*h[2,3] -> 2,3,5
158 		 * (y[0,1]+y[2,3])*(h[0,1]+h[2,3]) -> 6,7,8
159 		 */
160 
161 		a[0] = yw[0];
162 		a[1] = yw[1];
163 		a[2] = yw[2];
164 		a[3] = yw[3];
165 		a[4] = a[0] ^ a[1];
166 		a[5] = a[2] ^ a[3];
167 		a[6] = a[0] ^ a[2];
168 		a[7] = a[1] ^ a[3];
169 		a[8] = a[6] ^ a[7];
170 
171 		a[ 9] = rev32(yw[0]);
172 		a[10] = rev32(yw[1]);
173 		a[11] = rev32(yw[2]);
174 		a[12] = rev32(yw[3]);
175 		a[13] = a[ 9] ^ a[10];
176 		a[14] = a[11] ^ a[12];
177 		a[15] = a[ 9] ^ a[11];
178 		a[16] = a[10] ^ a[12];
179 		a[17] = a[15] ^ a[16];
180 
181 		b[0] = hw[0];
182 		b[1] = hw[1];
183 		b[2] = hw[2];
184 		b[3] = hw[3];
185 		b[4] = b[0] ^ b[1];
186 		b[5] = b[2] ^ b[3];
187 		b[6] = b[0] ^ b[2];
188 		b[7] = b[1] ^ b[3];
189 		b[8] = b[6] ^ b[7];
190 
191 		b[ 9] = hwr[0];
192 		b[10] = hwr[1];
193 		b[11] = hwr[2];
194 		b[12] = hwr[3];
195 		b[13] = b[ 9] ^ b[10];
196 		b[14] = b[11] ^ b[12];
197 		b[15] = b[ 9] ^ b[11];
198 		b[16] = b[10] ^ b[12];
199 		b[17] = b[15] ^ b[16];
200 
201 		for (i = 0; i < 18; i ++) {
202 			c[i] = bmul32(a[i], b[i]);
203 		}
204 
205 		c[4] ^= c[0] ^ c[1];
206 		c[5] ^= c[2] ^ c[3];
207 		c[8] ^= c[6] ^ c[7];
208 
209 		c[13] ^= c[ 9] ^ c[10];
210 		c[14] ^= c[11] ^ c[12];
211 		c[17] ^= c[15] ^ c[16];
212 
213 		/*
214 		 * y[0,1]*h[0,1] -> 0,9^4,1^13,10
215 		 * y[2,3]*h[2,3] -> 2,11^5,3^14,12
216 		 * (y[0,1]+y[2,3])*(h[0,1]+h[2,3]) -> 6,15^8,7^17,16
217 		 */
218 		d0 = c[0];
219 		d1 = c[4] ^ (rev32(c[9]) >> 1);
220 		d2 = c[1] ^ c[0] ^ c[2] ^ c[6] ^ (rev32(c[13]) >> 1);
221 		d3 = c[4] ^ c[5] ^ c[8]
222 			^ (rev32(c[10] ^ c[9] ^ c[11] ^ c[15]) >> 1);
223 		d4 = c[2] ^ c[1] ^ c[3] ^ c[7]
224 			^ (rev32(c[13] ^ c[14] ^ c[17]) >> 1);
225 		d5 = c[5] ^ (rev32(c[11] ^ c[10] ^ c[12] ^ c[16]) >> 1);
226 		d6 = c[3] ^ (rev32(c[14]) >> 1);
227 		d7 = rev32(c[12]) >> 1;
228 
229 		zw[0] = d0 << 1;
230 		zw[1] = (d1 << 1) | (d0 >> 31);
231 		zw[2] = (d2 << 1) | (d1 >> 31);
232 		zw[3] = (d3 << 1) | (d2 >> 31);
233 		zw[4] = (d4 << 1) | (d3 >> 31);
234 		zw[5] = (d5 << 1) | (d4 >> 31);
235 		zw[6] = (d6 << 1) | (d5 >> 31);
236 		zw[7] = (d7 << 1) | (d6 >> 31);
237 
238 		for (i = 0; i < 4; i ++) {
239 			uint32_t lw;
240 
241 			lw = zw[i];
242 			zw[i + 4] ^= lw ^ (lw >> 1) ^ (lw >> 2) ^ (lw >> 7);
243 			zw[i + 3] ^= (lw << 31) ^ (lw << 30) ^ (lw << 25);
244 		}
245 		memcpy(yw, zw + 4, sizeof yw);
246 	}
247 	br_enc32be(yb, yw[3]);
248 	br_enc32be(yb + 4, yw[2]);
249 	br_enc32be(yb + 8, yw[1]);
250 	br_enc32be(yb + 12, yw[0]);
251 }
252