1 | #ifndef NTL_mach_desc__H |
---|
2 | #define NTL_mach_desc__H |
---|
3 | |
---|
4 | |
---|
5 | #define NTL_BITS_PER_LONG (32) |
---|
6 | #define NTL_MAX_LONG (2147483647L) |
---|
7 | #define NTL_MAX_INT (2147483647) |
---|
8 | #define NTL_BITS_PER_INT (32) |
---|
9 | #define NTL_BITS_PER_SIZE_T (32) |
---|
10 | #define NTL_ARITH_RIGHT_SHIFT (1) |
---|
11 | #define NTL_NBITS_MAX (30) |
---|
12 | #define NTL_DOUBLE_PRECISION (53) |
---|
13 | #define NTL_FDOUBLE_PRECISION (((double)(1L<<30))*((double)(1L<<22))) |
---|
14 | #define NTL_QUAD_FLOAT_SPLIT ((((double)(1L<<27)))+1.0) |
---|
15 | #define NTL_EXT_DOUBLE (1) |
---|
16 | #define NTL_SINGLE_MUL_OK (1) |
---|
17 | #define NTL_DOUBLES_LOW_HIGH (1) |
---|
18 | |
---|
19 | |
---|
20 | |
---|
21 | |
---|
22 | #define NTL_BB_MUL_CODE0 \ |
---|
23 | _ntl_ulong hi, lo, t;\ |
---|
24 | _ntl_ulong A[8];\ |
---|
25 | A[0] = 0;\ |
---|
26 | A[1] = a;\ |
---|
27 | A[2] = A[1] << 1;\ |
---|
28 | A[3] = A[2] ^ A[1];\ |
---|
29 | A[4] = A[2] << 1;\ |
---|
30 | A[5] = A[4] ^ A[1];\ |
---|
31 | A[6] = A[3] << 1;\ |
---|
32 | A[7] = A[6] ^ A[1];\ |
---|
33 | lo = A[b & 7]; t = A[(b >> 3) & 7]; hi = t >> 29; lo ^= t << 3;\ |
---|
34 | t = A[(b >> 6) & 7]; hi ^= t >> 26; lo ^= t << 6;\ |
---|
35 | t = A[(b >> 9) & 7]; hi ^= t >> 23; lo ^= t << 9;\ |
---|
36 | t = A[(b >> 12) & 7]; hi ^= t >> 20; lo ^= t << 12;\ |
---|
37 | t = A[(b >> 15) & 7]; hi ^= t >> 17; lo ^= t << 15;\ |
---|
38 | t = A[(b >> 18) & 7]; hi ^= t >> 14; lo ^= t << 18;\ |
---|
39 | t = A[(b >> 21) & 7]; hi ^= t >> 11; lo ^= t << 21;\ |
---|
40 | t = A[(b >> 24) & 7]; hi ^= t >> 8; lo ^= t << 24;\ |
---|
41 | t = A[(b >> 27) & 7]; hi ^= t >> 5; lo ^= t << 27;\ |
---|
42 | t = A[b >> 30]; hi ^= t >> 2; lo ^= t << 30;\ |
---|
43 | if (a >> 31) hi ^= ((b & 0xb6db6db6UL) >> 1);\ |
---|
44 | if ((a >> 30) & 1) hi ^= ((b & 0x24924924UL) >> 2);\ |
---|
45 | c[0] = lo; c[1] = hi;\ |
---|
46 | |
---|
47 | |
---|
48 | |
---|
49 | |
---|
50 | |
---|
51 | #define NTL_BB_MUL_CODE1 \ |
---|
52 | long i;\ |
---|
53 | _ntl_ulong carry = 0, b;\ |
---|
54 | _ntl_ulong hi, lo, t;\ |
---|
55 | _ntl_ulong A[16];\ |
---|
56 | A[0] = 0;\ |
---|
57 | A[1] = a;\ |
---|
58 | A[2] = A[1] << 1;\ |
---|
59 | A[3] = A[2] ^ A[1];\ |
---|
60 | A[4] = A[2] << 1;\ |
---|
61 | A[5] = A[4] ^ A[1];\ |
---|
62 | A[6] = A[3] << 1;\ |
---|
63 | A[7] = A[6] ^ A[1];\ |
---|
64 | A[8] = A[4] << 1;\ |
---|
65 | A[9] = A[8] ^ A[1];\ |
---|
66 | A[10] = A[5] << 1;\ |
---|
67 | A[11] = A[10] ^ A[1];\ |
---|
68 | A[12] = A[6] << 1;\ |
---|
69 | A[13] = A[12] ^ A[1];\ |
---|
70 | A[14] = A[7] << 1;\ |
---|
71 | A[15] = A[14] ^ A[1];\ |
---|
72 | for (i = 0; i < sb; i++) {\ |
---|
73 | b = bp[i];\ |
---|
74 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 28; lo ^= t << 4;\ |
---|
75 | t = A[(b >> 8) & 15]; hi ^= t >> 24; lo ^= t << 8;\ |
---|
76 | t = A[(b >> 12) & 15]; hi ^= t >> 20; lo ^= t << 12;\ |
---|
77 | t = A[(b >> 16) & 15]; hi ^= t >> 16; lo ^= t << 16;\ |
---|
78 | t = A[(b >> 20) & 15]; hi ^= t >> 12; lo ^= t << 20;\ |
---|
79 | t = A[(b >> 24) & 15]; hi ^= t >> 8; lo ^= t << 24;\ |
---|
80 | t = A[b >> 28]; hi ^= t >> 4; lo ^= t << 28;\ |
---|
81 | if (a >> 31) hi ^= ((b & 0xeeeeeeeeUL) >> 1);\ |
---|
82 | if ((a >> 30) & 1) hi ^= ((b & 0xccccccccUL) >> 2);\ |
---|
83 | if ((a >> 29) & 1) hi ^= ((b & 0x88888888UL) >> 3);\ |
---|
84 | cp[i] = carry ^ lo; carry = hi;\ |
---|
85 | }\ |
---|
86 | cp[sb] = carry;\ |
---|
87 | |
---|
88 | |
---|
89 | |
---|
90 | |
---|
91 | |
---|
92 | #define NTL_BB_MUL_CODE2 \ |
---|
93 | long i;\ |
---|
94 | _ntl_ulong carry = 0, b;\ |
---|
95 | _ntl_ulong hi, lo, t;\ |
---|
96 | _ntl_ulong A[16];\ |
---|
97 | A[0] = 0;\ |
---|
98 | A[1] = a;\ |
---|
99 | A[2] = A[1] << 1;\ |
---|
100 | A[3] = A[2] ^ A[1];\ |
---|
101 | A[4] = A[2] << 1;\ |
---|
102 | A[5] = A[4] ^ A[1];\ |
---|
103 | A[6] = A[3] << 1;\ |
---|
104 | A[7] = A[6] ^ A[1];\ |
---|
105 | A[8] = A[4] << 1;\ |
---|
106 | A[9] = A[8] ^ A[1];\ |
---|
107 | A[10] = A[5] << 1;\ |
---|
108 | A[11] = A[10] ^ A[1];\ |
---|
109 | A[12] = A[6] << 1;\ |
---|
110 | A[13] = A[12] ^ A[1];\ |
---|
111 | A[14] = A[7] << 1;\ |
---|
112 | A[15] = A[14] ^ A[1];\ |
---|
113 | for (i = 0; i < sb; i++) {\ |
---|
114 | b = bp[i];\ |
---|
115 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 28; lo ^= t << 4;\ |
---|
116 | t = A[(b >> 8) & 15]; hi ^= t >> 24; lo ^= t << 8;\ |
---|
117 | t = A[(b >> 12) & 15]; hi ^= t >> 20; lo ^= t << 12;\ |
---|
118 | t = A[(b >> 16) & 15]; hi ^= t >> 16; lo ^= t << 16;\ |
---|
119 | t = A[(b >> 20) & 15]; hi ^= t >> 12; lo ^= t << 20;\ |
---|
120 | t = A[(b >> 24) & 15]; hi ^= t >> 8; lo ^= t << 24;\ |
---|
121 | t = A[b >> 28]; hi ^= t >> 4; lo ^= t << 28;\ |
---|
122 | if (a >> 31) hi ^= ((b & 0xeeeeeeeeUL) >> 1);\ |
---|
123 | if ((a >> 30) & 1) hi ^= ((b & 0xccccccccUL) >> 2);\ |
---|
124 | if ((a >> 29) & 1) hi ^= ((b & 0x88888888UL) >> 3);\ |
---|
125 | cp[i] ^= (carry ^ lo); carry = hi;\ |
---|
126 | }\ |
---|
127 | cp[sb] ^= carry;\ |
---|
128 | |
---|
129 | |
---|
130 | |
---|
131 | |
---|
132 | |
---|
133 | #define NTL_SHORT_BB_MUL_CODE1 \ |
---|
134 | long i;\ |
---|
135 | _ntl_ulong carry = 0, b;\ |
---|
136 | _ntl_ulong hi, lo, t;\ |
---|
137 | _ntl_ulong A[16];\ |
---|
138 | A[0] = 0;\ |
---|
139 | A[1] = a;\ |
---|
140 | A[2] = A[1] << 1;\ |
---|
141 | A[3] = A[2] ^ A[1];\ |
---|
142 | A[4] = A[2] << 1;\ |
---|
143 | A[5] = A[4] ^ A[1];\ |
---|
144 | A[6] = A[3] << 1;\ |
---|
145 | A[7] = A[6] ^ A[1];\ |
---|
146 | A[8] = A[4] << 1;\ |
---|
147 | A[9] = A[8] ^ A[1];\ |
---|
148 | A[10] = A[5] << 1;\ |
---|
149 | A[11] = A[10] ^ A[1];\ |
---|
150 | A[12] = A[6] << 1;\ |
---|
151 | A[13] = A[12] ^ A[1];\ |
---|
152 | A[14] = A[7] << 1;\ |
---|
153 | A[15] = A[14] ^ A[1];\ |
---|
154 | for (i = 0; i < sb; i++) {\ |
---|
155 | b = bp[i];\ |
---|
156 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 28; lo ^= t << 4;\ |
---|
157 | t = A[(b >> 8) & 15]; hi ^= t >> 24; lo ^= t << 8;\ |
---|
158 | t = A[(b >> 12) & 15]; hi ^= t >> 20; lo ^= t << 12;\ |
---|
159 | t = A[(b >> 16) & 15]; hi ^= t >> 16; lo ^= t << 16;\ |
---|
160 | t = A[(b >> 20) & 15]; hi ^= t >> 12; lo ^= t << 20;\ |
---|
161 | t = A[(b >> 24) & 15]; hi ^= t >> 8; lo ^= t << 24;\ |
---|
162 | t = A[b >> 28]; hi ^= t >> 4; lo ^= t << 28;\ |
---|
163 | cp[i] = carry ^ lo; carry = hi;\ |
---|
164 | }\ |
---|
165 | cp[sb] = carry;\ |
---|
166 | |
---|
167 | |
---|
168 | |
---|
169 | |
---|
170 | |
---|
171 | #define NTL_HALF_BB_MUL_CODE0 \ |
---|
172 | _ntl_ulong hi, lo, t;\ |
---|
173 | _ntl_ulong A[4];\ |
---|
174 | A[0] = 0;\ |
---|
175 | A[1] = a;\ |
---|
176 | A[2] = A[1] << 1;\ |
---|
177 | A[3] = A[2] ^ A[1];\ |
---|
178 | lo = A[b & 3]; t = A[(b >> 2) & 3]; hi = t >> 30; lo ^= t << 2;\ |
---|
179 | t = A[(b >> 4) & 3]; hi ^= t >> 28; lo ^= t << 4;\ |
---|
180 | t = A[(b >> 6) & 3]; hi ^= t >> 26; lo ^= t << 6;\ |
---|
181 | t = A[(b >> 8) & 3]; hi ^= t >> 24; lo ^= t << 8;\ |
---|
182 | t = A[(b >> 10) & 3]; hi ^= t >> 22; lo ^= t << 10;\ |
---|
183 | t = A[(b >> 12) & 3]; hi ^= t >> 20; lo ^= t << 12;\ |
---|
184 | t = A[b >> 14]; hi ^= t >> 18; lo ^= t << 14;\ |
---|
185 | if (a >> 31) hi ^= ((b & 0xaaaaUL) >> 1);\ |
---|
186 | c[0] = lo; c[1] = hi;\ |
---|
187 | |
---|
188 | |
---|
189 | |
---|
190 | |
---|
191 | |
---|
192 | #define NTL_ALT_BB_MUL_CODE0 \ |
---|
193 | _ntl_ulong A[8];\ |
---|
194 | A[0] = 0;\ |
---|
195 | A[1] = a;\ |
---|
196 | A[2] = A[1] << 1;\ |
---|
197 | A[3] = A[2] ^ A[1];\ |
---|
198 | A[4] = A[2] << 1;\ |
---|
199 | A[5] = A[4] ^ A[1];\ |
---|
200 | A[6] = A[3] << 1;\ |
---|
201 | A[7] = A[6] ^ A[1];\ |
---|
202 | const _ntl_ulong t3 = A[(b >> 3) & 7]; \ |
---|
203 | const _ntl_ulong t6 = A[(b >> 6) & 7]; \ |
---|
204 | const _ntl_ulong t9 = A[(b >> 9) & 7]; \ |
---|
205 | const _ntl_ulong t12 = A[(b >> 12) & 7]; \ |
---|
206 | const _ntl_ulong t15 = A[(b >> 15) & 7]; \ |
---|
207 | const _ntl_ulong t18 = A[(b >> 18) & 7]; \ |
---|
208 | const _ntl_ulong t21 = A[(b >> 21) & 7]; \ |
---|
209 | const _ntl_ulong t24 = A[(b >> 24) & 7]; \ |
---|
210 | const _ntl_ulong t27 = A[(b >> 27) & 7]; \ |
---|
211 | const _ntl_ulong t30 = A[b >> 30]; \ |
---|
212 | const _ntl_ulong lo = A[b & 7] \ |
---|
213 | ^ (t3 << 3)\ |
---|
214 | ^ (t6 << 6)\ |
---|
215 | ^ (t9 << 9)\ |
---|
216 | ^ (t12 << 12)\ |
---|
217 | ^ (t15 << 15)\ |
---|
218 | ^ (t18 << 18)\ |
---|
219 | ^ (t21 << 21)\ |
---|
220 | ^ (t24 << 24)\ |
---|
221 | ^ (t27 << 27)\ |
---|
222 | ^ (t30 << 30);\ |
---|
223 | const _ntl_ulong hi = (t3 >> 29)\ |
---|
224 | ^ (t6 >> 26)\ |
---|
225 | ^ (t9 >> 23)\ |
---|
226 | ^ (t12 >> 20)\ |
---|
227 | ^ (t15 >> 17)\ |
---|
228 | ^ (t18 >> 14)\ |
---|
229 | ^ (t21 >> 11)\ |
---|
230 | ^ (t24 >> 8)\ |
---|
231 | ^ (t27 >> 5)\ |
---|
232 | ^ (t30 >> 2)\ |
---|
233 | ^ (((b & 0xb6db6db6UL) >> 1) & (-(a >> 31)))\ |
---|
234 | ^ (((b & 0x24924924UL) >> 2) & (-((a >> 30) & 1UL)));\ |
---|
235 | c[0] = lo; c[1] = hi;\ |
---|
236 | |
---|
237 | |
---|
238 | |
---|
239 | |
---|
240 | |
---|
241 | #define NTL_ALT_BB_MUL_CODE1 \ |
---|
242 | long i;\ |
---|
243 | _ntl_ulong carry = 0;\ |
---|
244 | _ntl_ulong A[16];\ |
---|
245 | A[0] = 0;\ |
---|
246 | A[1] = a;\ |
---|
247 | A[2] = A[1] << 1;\ |
---|
248 | A[3] = A[2] ^ A[1];\ |
---|
249 | A[4] = A[2] << 1;\ |
---|
250 | A[5] = A[4] ^ A[1];\ |
---|
251 | A[6] = A[3] << 1;\ |
---|
252 | A[7] = A[6] ^ A[1];\ |
---|
253 | A[8] = A[4] << 1;\ |
---|
254 | A[9] = A[8] ^ A[1];\ |
---|
255 | A[10] = A[5] << 1;\ |
---|
256 | A[11] = A[10] ^ A[1];\ |
---|
257 | A[12] = A[6] << 1;\ |
---|
258 | A[13] = A[12] ^ A[1];\ |
---|
259 | A[14] = A[7] << 1;\ |
---|
260 | A[15] = A[14] ^ A[1];\ |
---|
261 | for (i = 0; i < sb; i++) {\ |
---|
262 | const _ntl_ulong b = bp[i];\ |
---|
263 | const _ntl_ulong t4 = A[(b >> 4) & 15]; \ |
---|
264 | const _ntl_ulong t8 = A[(b >> 8) & 15]; \ |
---|
265 | const _ntl_ulong t12 = A[(b >> 12) & 15]; \ |
---|
266 | const _ntl_ulong t16 = A[(b >> 16) & 15]; \ |
---|
267 | const _ntl_ulong t20 = A[(b >> 20) & 15]; \ |
---|
268 | const _ntl_ulong t24 = A[(b >> 24) & 15]; \ |
---|
269 | const _ntl_ulong t28 = A[b >> 28]; \ |
---|
270 | const _ntl_ulong lo = A[b & 15] \ |
---|
271 | ^ (t4 << 4)\ |
---|
272 | ^ (t8 << 8)\ |
---|
273 | ^ (t12 << 12)\ |
---|
274 | ^ (t16 << 16)\ |
---|
275 | ^ (t20 << 20)\ |
---|
276 | ^ (t24 << 24)\ |
---|
277 | ^ (t28 << 28);\ |
---|
278 | const _ntl_ulong hi = (t4 >> 28)\ |
---|
279 | ^ (t8 >> 24)\ |
---|
280 | ^ (t12 >> 20)\ |
---|
281 | ^ (t16 >> 16)\ |
---|
282 | ^ (t20 >> 12)\ |
---|
283 | ^ (t24 >> 8)\ |
---|
284 | ^ (t28 >> 4)\ |
---|
285 | ^ (((b & 0xeeeeeeeeUL) >> 1) & (-(a >> 31)))\ |
---|
286 | ^ (((b & 0xccccccccUL) >> 2) & (-((a >> 30) & 1UL)))\ |
---|
287 | ^ (((b & 0x88888888UL) >> 3) & (-((a >> 29) & 1UL)));\ |
---|
288 | cp[i] = carry ^ lo; carry = hi;\ |
---|
289 | }\ |
---|
290 | cp[sb] = carry;\ |
---|
291 | |
---|
292 | |
---|
293 | |
---|
294 | |
---|
295 | |
---|
296 | #define NTL_ALT_BB_MUL_CODE2 \ |
---|
297 | long i;\ |
---|
298 | _ntl_ulong carry = 0;\ |
---|
299 | _ntl_ulong A[16];\ |
---|
300 | A[0] = 0;\ |
---|
301 | A[1] = a;\ |
---|
302 | A[2] = A[1] << 1;\ |
---|
303 | A[3] = A[2] ^ A[1];\ |
---|
304 | A[4] = A[2] << 1;\ |
---|
305 | A[5] = A[4] ^ A[1];\ |
---|
306 | A[6] = A[3] << 1;\ |
---|
307 | A[7] = A[6] ^ A[1];\ |
---|
308 | A[8] = A[4] << 1;\ |
---|
309 | A[9] = A[8] ^ A[1];\ |
---|
310 | A[10] = A[5] << 1;\ |
---|
311 | A[11] = A[10] ^ A[1];\ |
---|
312 | A[12] = A[6] << 1;\ |
---|
313 | A[13] = A[12] ^ A[1];\ |
---|
314 | A[14] = A[7] << 1;\ |
---|
315 | A[15] = A[14] ^ A[1];\ |
---|
316 | for (i = 0; i < sb; i++) {\ |
---|
317 | const _ntl_ulong b = bp[i];\ |
---|
318 | const _ntl_ulong t4 = A[(b >> 4) & 15]; \ |
---|
319 | const _ntl_ulong t8 = A[(b >> 8) & 15]; \ |
---|
320 | const _ntl_ulong t12 = A[(b >> 12) & 15]; \ |
---|
321 | const _ntl_ulong t16 = A[(b >> 16) & 15]; \ |
---|
322 | const _ntl_ulong t20 = A[(b >> 20) & 15]; \ |
---|
323 | const _ntl_ulong t24 = A[(b >> 24) & 15]; \ |
---|
324 | const _ntl_ulong t28 = A[b >> 28]; \ |
---|
325 | const _ntl_ulong lo = A[b & 15] \ |
---|
326 | ^ (t4 << 4)\ |
---|
327 | ^ (t8 << 8)\ |
---|
328 | ^ (t12 << 12)\ |
---|
329 | ^ (t16 << 16)\ |
---|
330 | ^ (t20 << 20)\ |
---|
331 | ^ (t24 << 24)\ |
---|
332 | ^ (t28 << 28);\ |
---|
333 | const _ntl_ulong hi = (t4 >> 28)\ |
---|
334 | ^ (t8 >> 24)\ |
---|
335 | ^ (t12 >> 20)\ |
---|
336 | ^ (t16 >> 16)\ |
---|
337 | ^ (t20 >> 12)\ |
---|
338 | ^ (t24 >> 8)\ |
---|
339 | ^ (t28 >> 4)\ |
---|
340 | ^ (((b & 0xeeeeeeeeUL) >> 1) & (-(a >> 31)))\ |
---|
341 | ^ (((b & 0xccccccccUL) >> 2) & (-((a >> 30) & 1UL)))\ |
---|
342 | ^ (((b & 0x88888888UL) >> 3) & (-((a >> 29) & 1UL)));\ |
---|
343 | cp[i] ^= (carry ^ lo); carry = hi;\ |
---|
344 | }\ |
---|
345 | cp[sb] ^= carry;\ |
---|
346 | |
---|
347 | |
---|
348 | |
---|
349 | |
---|
350 | |
---|
351 | #define NTL_ALT_SHORT_BB_MUL_CODE1 \ |
---|
352 | long i;\ |
---|
353 | _ntl_ulong carry = 0;\ |
---|
354 | _ntl_ulong A[16];\ |
---|
355 | A[0] = 0;\ |
---|
356 | A[1] = a;\ |
---|
357 | A[2] = A[1] << 1;\ |
---|
358 | A[3] = A[2] ^ A[1];\ |
---|
359 | A[4] = A[2] << 1;\ |
---|
360 | A[5] = A[4] ^ A[1];\ |
---|
361 | A[6] = A[3] << 1;\ |
---|
362 | A[7] = A[6] ^ A[1];\ |
---|
363 | A[8] = A[4] << 1;\ |
---|
364 | A[9] = A[8] ^ A[1];\ |
---|
365 | A[10] = A[5] << 1;\ |
---|
366 | A[11] = A[10] ^ A[1];\ |
---|
367 | A[12] = A[6] << 1;\ |
---|
368 | A[13] = A[12] ^ A[1];\ |
---|
369 | A[14] = A[7] << 1;\ |
---|
370 | A[15] = A[14] ^ A[1];\ |
---|
371 | for (i = 0; i < sb; i++) {\ |
---|
372 | const _ntl_ulong b = bp[i];\ |
---|
373 | const _ntl_ulong t4 = A[(b >> 4) & 15]; \ |
---|
374 | const _ntl_ulong t8 = A[(b >> 8) & 15]; \ |
---|
375 | const _ntl_ulong t12 = A[(b >> 12) & 15]; \ |
---|
376 | const _ntl_ulong t16 = A[(b >> 16) & 15]; \ |
---|
377 | const _ntl_ulong t20 = A[(b >> 20) & 15]; \ |
---|
378 | const _ntl_ulong t24 = A[(b >> 24) & 15]; \ |
---|
379 | const _ntl_ulong t28 = A[b >> 28]; \ |
---|
380 | const _ntl_ulong lo = A[b & 15] \ |
---|
381 | ^ (t4 << 4)\ |
---|
382 | ^ (t8 << 8)\ |
---|
383 | ^ (t12 << 12)\ |
---|
384 | ^ (t16 << 16)\ |
---|
385 | ^ (t20 << 20)\ |
---|
386 | ^ (t24 << 24)\ |
---|
387 | ^ (t28 << 28);\ |
---|
388 | const _ntl_ulong hi = (t4 >> 28)\ |
---|
389 | ^ (t8 >> 24)\ |
---|
390 | ^ (t12 >> 20)\ |
---|
391 | ^ (t16 >> 16)\ |
---|
392 | ^ (t20 >> 12)\ |
---|
393 | ^ (t24 >> 8)\ |
---|
394 | ^ (t28 >> 4);\ |
---|
395 | cp[i] = carry ^ lo; carry = hi;\ |
---|
396 | }\ |
---|
397 | cp[sb] = carry;\ |
---|
398 | |
---|
399 | |
---|
400 | |
---|
401 | |
---|
402 | |
---|
403 | #define NTL_ALT_HALF_BB_MUL_CODE0 \ |
---|
404 | _ntl_ulong A[4];\ |
---|
405 | A[0] = 0;\ |
---|
406 | A[1] = a;\ |
---|
407 | A[2] = A[1] << 1;\ |
---|
408 | A[3] = A[2] ^ A[1];\ |
---|
409 | const _ntl_ulong t2 = A[(b >> 2) & 3]; \ |
---|
410 | const _ntl_ulong t4 = A[(b >> 4) & 3]; \ |
---|
411 | const _ntl_ulong t6 = A[(b >> 6) & 3]; \ |
---|
412 | const _ntl_ulong t8 = A[(b >> 8) & 3]; \ |
---|
413 | const _ntl_ulong t10 = A[(b >> 10) & 3]; \ |
---|
414 | const _ntl_ulong t12 = A[(b >> 12) & 3]; \ |
---|
415 | const _ntl_ulong t14 = A[b >> 14]; \ |
---|
416 | const _ntl_ulong lo = A[b & 3] \ |
---|
417 | ^ (t2 << 2)\ |
---|
418 | ^ (t4 << 4)\ |
---|
419 | ^ (t6 << 6)\ |
---|
420 | ^ (t8 << 8)\ |
---|
421 | ^ (t10 << 10)\ |
---|
422 | ^ (t12 << 12)\ |
---|
423 | ^ (t14 << 14);\ |
---|
424 | const _ntl_ulong hi = (t2 >> 30)\ |
---|
425 | ^ (t4 >> 28)\ |
---|
426 | ^ (t6 >> 26)\ |
---|
427 | ^ (t8 >> 24)\ |
---|
428 | ^ (t10 >> 22)\ |
---|
429 | ^ (t12 >> 20)\ |
---|
430 | ^ (t14 >> 18)\ |
---|
431 | ^ (((b & 0xaaaaUL) >> 1) & (-(a >> 31)));\ |
---|
432 | c[0] = lo; c[1] = hi;\ |
---|
433 | |
---|
434 | |
---|
435 | |
---|
436 | |
---|
437 | |
---|
438 | #define NTL_ALT1_BB_MUL_CODE0 \ |
---|
439 | _ntl_ulong hi, lo, t;\ |
---|
440 | _ntl_ulong A[8];\ |
---|
441 | A[0] = 0;\ |
---|
442 | A[1] = a;\ |
---|
443 | A[2] = A[1] << 1;\ |
---|
444 | A[3] = A[2] ^ A[1];\ |
---|
445 | A[4] = A[2] << 1;\ |
---|
446 | A[5] = A[4] ^ A[1];\ |
---|
447 | A[6] = A[3] << 1;\ |
---|
448 | A[7] = A[6] ^ A[1];\ |
---|
449 | lo = A[b & 7]; t = A[(b >> 3) & 7]; hi = t >> 29; lo ^= t << 3;\ |
---|
450 | t = A[(b >> 6) & 7]; hi ^= t >> 26; lo ^= t << 6;\ |
---|
451 | t = A[(b >> 9) & 7]; hi ^= t >> 23; lo ^= t << 9;\ |
---|
452 | t = A[(b >> 12) & 7]; hi ^= t >> 20; lo ^= t << 12;\ |
---|
453 | t = A[(b >> 15) & 7]; hi ^= t >> 17; lo ^= t << 15;\ |
---|
454 | t = A[(b >> 18) & 7]; hi ^= t >> 14; lo ^= t << 18;\ |
---|
455 | t = A[(b >> 21) & 7]; hi ^= t >> 11; lo ^= t << 21;\ |
---|
456 | t = A[(b >> 24) & 7]; hi ^= t >> 8; lo ^= t << 24;\ |
---|
457 | t = A[(b >> 27) & 7]; hi ^= t >> 5; lo ^= t << 27;\ |
---|
458 | t = A[b >> 30]; hi ^= t >> 2; lo ^= t << 30;\ |
---|
459 | hi ^= (((b & 0xb6db6db6UL) >> 1) & (-(a >> 31)))\ |
---|
460 | ^ (((b & 0x24924924UL) >> 2) & (-((a >> 30) & 1UL)));\ |
---|
461 | c[0] = lo; c[1] = hi;\ |
---|
462 | |
---|
463 | |
---|
464 | |
---|
465 | |
---|
466 | |
---|
467 | #define NTL_ALT1_BB_MUL_CODE1 \ |
---|
468 | long i;\ |
---|
469 | _ntl_ulong carry = 0, b;\ |
---|
470 | _ntl_ulong hi, lo, t;\ |
---|
471 | _ntl_ulong A[16];\ |
---|
472 | A[0] = 0;\ |
---|
473 | A[1] = a;\ |
---|
474 | A[2] = A[1] << 1;\ |
---|
475 | A[3] = A[2] ^ A[1];\ |
---|
476 | A[4] = A[2] << 1;\ |
---|
477 | A[5] = A[4] ^ A[1];\ |
---|
478 | A[6] = A[3] << 1;\ |
---|
479 | A[7] = A[6] ^ A[1];\ |
---|
480 | A[8] = A[4] << 1;\ |
---|
481 | A[9] = A[8] ^ A[1];\ |
---|
482 | A[10] = A[5] << 1;\ |
---|
483 | A[11] = A[10] ^ A[1];\ |
---|
484 | A[12] = A[6] << 1;\ |
---|
485 | A[13] = A[12] ^ A[1];\ |
---|
486 | A[14] = A[7] << 1;\ |
---|
487 | A[15] = A[14] ^ A[1];\ |
---|
488 | for (i = 0; i < sb; i++) {\ |
---|
489 | b = bp[i];\ |
---|
490 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 28; lo ^= t << 4;\ |
---|
491 | t = A[(b >> 8) & 15]; hi ^= t >> 24; lo ^= t << 8;\ |
---|
492 | t = A[(b >> 12) & 15]; hi ^= t >> 20; lo ^= t << 12;\ |
---|
493 | t = A[(b >> 16) & 15]; hi ^= t >> 16; lo ^= t << 16;\ |
---|
494 | t = A[(b >> 20) & 15]; hi ^= t >> 12; lo ^= t << 20;\ |
---|
495 | t = A[(b >> 24) & 15]; hi ^= t >> 8; lo ^= t << 24;\ |
---|
496 | t = A[b >> 28]; hi ^= t >> 4; lo ^= t << 28;\ |
---|
497 | hi ^= (((b & 0xeeeeeeeeUL) >> 1) & (-(a >> 31)))\ |
---|
498 | ^ (((b & 0xccccccccUL) >> 2) & (-((a >> 30) & 1UL)))\ |
---|
499 | ^ (((b & 0x88888888UL) >> 3) & (-((a >> 29) & 1UL)));\ |
---|
500 | cp[i] = carry ^ lo; carry = hi;\ |
---|
501 | }\ |
---|
502 | cp[sb] = carry;\ |
---|
503 | |
---|
504 | |
---|
505 | |
---|
506 | |
---|
507 | |
---|
508 | #define NTL_ALT1_BB_MUL_CODE2 \ |
---|
509 | long i;\ |
---|
510 | _ntl_ulong carry = 0, b;\ |
---|
511 | _ntl_ulong hi, lo, t;\ |
---|
512 | _ntl_ulong A[16];\ |
---|
513 | A[0] = 0;\ |
---|
514 | A[1] = a;\ |
---|
515 | A[2] = A[1] << 1;\ |
---|
516 | A[3] = A[2] ^ A[1];\ |
---|
517 | A[4] = A[2] << 1;\ |
---|
518 | A[5] = A[4] ^ A[1];\ |
---|
519 | A[6] = A[3] << 1;\ |
---|
520 | A[7] = A[6] ^ A[1];\ |
---|
521 | A[8] = A[4] << 1;\ |
---|
522 | A[9] = A[8] ^ A[1];\ |
---|
523 | A[10] = A[5] << 1;\ |
---|
524 | A[11] = A[10] ^ A[1];\ |
---|
525 | A[12] = A[6] << 1;\ |
---|
526 | A[13] = A[12] ^ A[1];\ |
---|
527 | A[14] = A[7] << 1;\ |
---|
528 | A[15] = A[14] ^ A[1];\ |
---|
529 | for (i = 0; i < sb; i++) {\ |
---|
530 | b = bp[i];\ |
---|
531 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 28; lo ^= t << 4;\ |
---|
532 | t = A[(b >> 8) & 15]; hi ^= t >> 24; lo ^= t << 8;\ |
---|
533 | t = A[(b >> 12) & 15]; hi ^= t >> 20; lo ^= t << 12;\ |
---|
534 | t = A[(b >> 16) & 15]; hi ^= t >> 16; lo ^= t << 16;\ |
---|
535 | t = A[(b >> 20) & 15]; hi ^= t >> 12; lo ^= t << 20;\ |
---|
536 | t = A[(b >> 24) & 15]; hi ^= t >> 8; lo ^= t << 24;\ |
---|
537 | t = A[b >> 28]; hi ^= t >> 4; lo ^= t << 28;\ |
---|
538 | hi ^= (((b & 0xeeeeeeeeUL) >> 1) & (-(a >> 31)))\ |
---|
539 | ^ (((b & 0xccccccccUL) >> 2) & (-((a >> 30) & 1UL)))\ |
---|
540 | ^ (((b & 0x88888888UL) >> 3) & (-((a >> 29) & 1UL)));\ |
---|
541 | cp[i] ^= (carry ^ lo); carry = hi;\ |
---|
542 | }\ |
---|
543 | cp[sb] ^= carry;\ |
---|
544 | |
---|
545 | |
---|
546 | |
---|
547 | |
---|
548 | |
---|
549 | #define NTL_ALT1_SHORT_BB_MUL_CODE1 \ |
---|
550 | long i;\ |
---|
551 | _ntl_ulong carry = 0, b;\ |
---|
552 | _ntl_ulong hi, lo, t;\ |
---|
553 | _ntl_ulong A[16];\ |
---|
554 | A[0] = 0;\ |
---|
555 | A[1] = a;\ |
---|
556 | A[2] = A[1] << 1;\ |
---|
557 | A[3] = A[2] ^ A[1];\ |
---|
558 | A[4] = A[2] << 1;\ |
---|
559 | A[5] = A[4] ^ A[1];\ |
---|
560 | A[6] = A[3] << 1;\ |
---|
561 | A[7] = A[6] ^ A[1];\ |
---|
562 | A[8] = A[4] << 1;\ |
---|
563 | A[9] = A[8] ^ A[1];\ |
---|
564 | A[10] = A[5] << 1;\ |
---|
565 | A[11] = A[10] ^ A[1];\ |
---|
566 | A[12] = A[6] << 1;\ |
---|
567 | A[13] = A[12] ^ A[1];\ |
---|
568 | A[14] = A[7] << 1;\ |
---|
569 | A[15] = A[14] ^ A[1];\ |
---|
570 | for (i = 0; i < sb; i++) {\ |
---|
571 | b = bp[i];\ |
---|
572 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 28; lo ^= t << 4;\ |
---|
573 | t = A[(b >> 8) & 15]; hi ^= t >> 24; lo ^= t << 8;\ |
---|
574 | t = A[(b >> 12) & 15]; hi ^= t >> 20; lo ^= t << 12;\ |
---|
575 | t = A[(b >> 16) & 15]; hi ^= t >> 16; lo ^= t << 16;\ |
---|
576 | t = A[(b >> 20) & 15]; hi ^= t >> 12; lo ^= t << 20;\ |
---|
577 | t = A[(b >> 24) & 15]; hi ^= t >> 8; lo ^= t << 24;\ |
---|
578 | t = A[b >> 28]; hi ^= t >> 4; lo ^= t << 28;\ |
---|
579 | cp[i] = carry ^ lo; carry = hi;\ |
---|
580 | }\ |
---|
581 | cp[sb] = carry;\ |
---|
582 | |
---|
583 | |
---|
584 | |
---|
585 | |
---|
586 | |
---|
587 | #define NTL_ALT1_HALF_BB_MUL_CODE0 \ |
---|
588 | _ntl_ulong hi, lo, t;\ |
---|
589 | _ntl_ulong A[4];\ |
---|
590 | A[0] = 0;\ |
---|
591 | A[1] = a;\ |
---|
592 | A[2] = A[1] << 1;\ |
---|
593 | A[3] = A[2] ^ A[1];\ |
---|
594 | lo = A[b & 3]; t = A[(b >> 2) & 3]; hi = t >> 30; lo ^= t << 2;\ |
---|
595 | t = A[(b >> 4) & 3]; hi ^= t >> 28; lo ^= t << 4;\ |
---|
596 | t = A[(b >> 6) & 3]; hi ^= t >> 26; lo ^= t << 6;\ |
---|
597 | t = A[(b >> 8) & 3]; hi ^= t >> 24; lo ^= t << 8;\ |
---|
598 | t = A[(b >> 10) & 3]; hi ^= t >> 22; lo ^= t << 10;\ |
---|
599 | t = A[(b >> 12) & 3]; hi ^= t >> 20; lo ^= t << 12;\ |
---|
600 | t = A[b >> 14]; hi ^= t >> 18; lo ^= t << 14;\ |
---|
601 | hi ^= (((b & 0xaaaaUL) >> 1) & (-(a >> 31)));\ |
---|
602 | c[0] = lo; c[1] = hi;\ |
---|
603 | |
---|
604 | |
---|
605 | |
---|
606 | #define NTL_BB_MUL1_BITS (4) |
---|
607 | |
---|
608 | |
---|
609 | |
---|
610 | #define NTL_BB_SQR_CODE \ |
---|
611 | lo=sqrtab[a&255];\ |
---|
612 | lo=lo|(sqrtab[(a>>8)&255]<<16);\ |
---|
613 | hi=sqrtab[(a>>16)&255];\ |
---|
614 | hi=hi|(sqrtab[(a>>24)&255]<<16);\ |
---|
615 | |
---|
616 | |
---|
617 | |
---|
618 | |
---|
619 | #define NTL_BB_REV_CODE (revtab[(a>>0)&255]<<24)\ |
---|
620 | |(revtab[(a>>8)&255]<<16)\ |
---|
621 | |(revtab[(a>>16)&255]<<8)\ |
---|
622 | |(revtab[(a>>24)&255]<<0) |
---|
623 | |
---|
624 | #define NTL_MIN_LONG (-NTL_MAX_LONG - 1L) |
---|
625 | #define NTL_MIN_INT (-NTL_MAX_INT - 1) |
---|
626 | #endif |
---|
627 | |
---|