ARM Shanghai IoT Team (Internal) / newMiniTLS-GPL

change some parameters in the library to meet the needs of the website httpbin.org

Fork of MiniTLS-GPL by Donatien Garnier

math/sqr/fp_sqr_comba.c@5:95f70ebfe61f, 2015-02-06 (annotated)

Committer:
shiyilei
Date:
Fri Feb 06 06:17:33 2015 +0000
Revision:
5:95f70ebfe61f
Parent:
0:35aa5be3b78d 
change some parameters in the library to meet the needs of httpbin.org

Who changed what in which revision?

UserRevisionLine numberNew contents of line
MiniTLS 0:35aa5be3b78d 1 /*
MiniTLS 0:35aa5be3b78d 2 *
MiniTLS 0:35aa5be3b78d 3 * This project is meant to fill in where LibTomMath
MiniTLS 0:35aa5be3b78d 4 * falls short. That is speed ;-)
MiniTLS 0:35aa5be3b78d 5 *
MiniTLS 0:35aa5be3b78d 6 * This project is public domain and free for all purposes.
MiniTLS 0:35aa5be3b78d 7 *
MiniTLS 0:35aa5be3b78d 8 * Tom St Denis, tomstdenis@gmail.com
MiniTLS 0:35aa5be3b78d 9 */
MiniTLS 0:35aa5be3b78d 10 #include <tfm.h>
MiniTLS 0:35aa5be3b78d 11
MiniTLS 0:35aa5be3b78d 12 #if defined(TFM_PRESCOTT) && defined(TFM_SSE2)
MiniTLS 0:35aa5be3b78d 13 #undef TFM_SSE2
MiniTLS 0:35aa5be3b78d 14 #define TFM_X86
MiniTLS 0:35aa5be3b78d 15 #endif
MiniTLS 0:35aa5be3b78d 16
MiniTLS 0:35aa5be3b78d 17 #if defined(TFM_X86)
MiniTLS 0:35aa5be3b78d 18
MiniTLS 0:35aa5be3b78d 19 /* x86-32 optimized */
MiniTLS 0:35aa5be3b78d 20
MiniTLS 0:35aa5be3b78d 21 #define COMBA_START
MiniTLS 0:35aa5be3b78d 22
MiniTLS 0:35aa5be3b78d 23 #define CLEAR_CARRY \
MiniTLS 0:35aa5be3b78d 24 c0 = c1 = c2 = 0;
MiniTLS 0:35aa5be3b78d 25
MiniTLS 0:35aa5be3b78d 26 #define COMBA_STORE(x) \
MiniTLS 0:35aa5be3b78d 27 x = c0;
MiniTLS 0:35aa5be3b78d 28
MiniTLS 0:35aa5be3b78d 29 #define COMBA_STORE2(x) \
MiniTLS 0:35aa5be3b78d 30 x = c1;
MiniTLS 0:35aa5be3b78d 31
MiniTLS 0:35aa5be3b78d 32 #define CARRY_FORWARD \
MiniTLS 0:35aa5be3b78d 33 do { c0 = c1; c1 = c2; c2 = 0; } while (0);
MiniTLS 0:35aa5be3b78d 34
MiniTLS 0:35aa5be3b78d 35 #define COMBA_FINI
MiniTLS 0:35aa5be3b78d 36
MiniTLS 0:35aa5be3b78d 37 #define SQRADD(i, j) \
MiniTLS 0:35aa5be3b78d 38 asm( \
MiniTLS 0:35aa5be3b78d 39 "movl %6,%%eax \n\t" \
MiniTLS 0:35aa5be3b78d 40 "mull %%eax \n\t" \
MiniTLS 0:35aa5be3b78d 41 "addl %%eax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 42 "adcl %%edx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 43 "adcl $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 44 :"=r"(c0), "=r"(c1), "=r"(c2): "0"(c0), "1"(c1), "2"(c2), "m"(i) :"%eax","%edx","%cc");
MiniTLS 0:35aa5be3b78d 45
MiniTLS 0:35aa5be3b78d 46 #define SQRADD2(i, j) \
MiniTLS 0:35aa5be3b78d 47 asm( \
MiniTLS 0:35aa5be3b78d 48 "movl %6,%%eax \n\t" \
MiniTLS 0:35aa5be3b78d 49 "mull %7 \n\t" \
MiniTLS 0:35aa5be3b78d 50 "addl %%eax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 51 "adcl %%edx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 52 "adcl $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 53 "addl %%eax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 54 "adcl %%edx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 55 "adcl $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 56 :"=r"(c0), "=r"(c1), "=r"(c2): "0"(c0), "1"(c1), "2"(c2), "m"(i), "m"(j) :"%eax","%edx","%cc");
MiniTLS 0:35aa5be3b78d 57
MiniTLS 0:35aa5be3b78d 58 #define SQRADDSC(i, j) \
MiniTLS 0:35aa5be3b78d 59 asm( \
MiniTLS 0:35aa5be3b78d 60 "movl %6,%%eax \n\t" \
MiniTLS 0:35aa5be3b78d 61 "mull %7 \n\t" \
MiniTLS 0:35aa5be3b78d 62 "movl %%eax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 63 "movl %%edx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 64 "xorl %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 65 :"=r"(sc0), "=r"(sc1), "=r"(sc2): "0"(sc0), "1"(sc1), "2"(sc2), "g"(i), "g"(j) :"%eax","%edx","%cc");
MiniTLS 0:35aa5be3b78d 66
MiniTLS 0:35aa5be3b78d 67 #define SQRADDAC(i, j) \
MiniTLS 0:35aa5be3b78d 68 asm( \
MiniTLS 0:35aa5be3b78d 69 "movl %6,%%eax \n\t" \
MiniTLS 0:35aa5be3b78d 70 "mull %7 \n\t" \
MiniTLS 0:35aa5be3b78d 71 "addl %%eax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 72 "adcl %%edx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 73 "adcl $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 74 :"=r"(sc0), "=r"(sc1), "=r"(sc2): "0"(sc0), "1"(sc1), "2"(sc2), "g"(i), "g"(j) :"%eax","%edx","%cc");
MiniTLS 0:35aa5be3b78d 75
MiniTLS 0:35aa5be3b78d 76 #define SQRADDDB \
MiniTLS 0:35aa5be3b78d 77 asm( \
MiniTLS 0:35aa5be3b78d 78 "addl %6,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 79 "adcl %7,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 80 "adcl %8,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 81 "addl %6,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 82 "adcl %7,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 83 "adcl %8,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 84 :"=r"(c0), "=r"(c1), "=r"(c2) : "0"(c0), "1"(c1), "2"(c2), "r"(sc0), "r"(sc1), "r"(sc2) : "%cc");
MiniTLS 0:35aa5be3b78d 85
MiniTLS 0:35aa5be3b78d 86 #elif defined(TFM_X86_64)
MiniTLS 0:35aa5be3b78d 87 /* x86-64 optimized */
MiniTLS 0:35aa5be3b78d 88
MiniTLS 0:35aa5be3b78d 89 #define COMBA_START
MiniTLS 0:35aa5be3b78d 90
MiniTLS 0:35aa5be3b78d 91 #define CLEAR_CARRY \
MiniTLS 0:35aa5be3b78d 92 c0 = c1 = c2 = 0;
MiniTLS 0:35aa5be3b78d 93
MiniTLS 0:35aa5be3b78d 94 #define COMBA_STORE(x) \
MiniTLS 0:35aa5be3b78d 95 x = c0;
MiniTLS 0:35aa5be3b78d 96
MiniTLS 0:35aa5be3b78d 97 #define COMBA_STORE2(x) \
MiniTLS 0:35aa5be3b78d 98 x = c1;
MiniTLS 0:35aa5be3b78d 99
MiniTLS 0:35aa5be3b78d 100 #define CARRY_FORWARD \
MiniTLS 0:35aa5be3b78d 101 do { c0 = c1; c1 = c2; c2 = 0; } while (0);
MiniTLS 0:35aa5be3b78d 102
MiniTLS 0:35aa5be3b78d 103 #define COMBA_FINI
MiniTLS 0:35aa5be3b78d 104
MiniTLS 0:35aa5be3b78d 105 #define SQRADD(i, j) \
MiniTLS 0:35aa5be3b78d 106 asm( \
MiniTLS 0:35aa5be3b78d 107 "movq %6,%%rax \n\t" \
MiniTLS 0:35aa5be3b78d 108 "mulq %%rax \n\t" \
MiniTLS 0:35aa5be3b78d 109 "addq %%rax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 110 "adcq %%rdx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 111 "adcq $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 112 :"=r"(c0), "=r"(c1), "=r"(c2): "0"(c0), "1"(c1), "2"(c2), "g"(i) :"%rax","%rdx","%cc");
MiniTLS 0:35aa5be3b78d 113
MiniTLS 0:35aa5be3b78d 114 #define SQRADD2(i, j) \
MiniTLS 0:35aa5be3b78d 115 asm( \
MiniTLS 0:35aa5be3b78d 116 "movq %6,%%rax \n\t" \
MiniTLS 0:35aa5be3b78d 117 "mulq %7 \n\t" \
MiniTLS 0:35aa5be3b78d 118 "addq %%rax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 119 "adcq %%rdx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 120 "adcq $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 121 "addq %%rax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 122 "adcq %%rdx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 123 "adcq $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 124 :"=r"(c0), "=r"(c1), "=r"(c2): "0"(c0), "1"(c1), "2"(c2), "g"(i), "g"(j) :"%rax","%rdx","%cc");
MiniTLS 0:35aa5be3b78d 125
MiniTLS 0:35aa5be3b78d 126 #define SQRADDSC(i, j) \
MiniTLS 0:35aa5be3b78d 127 asm( \
MiniTLS 0:35aa5be3b78d 128 "movq %6,%%rax \n\t" \
MiniTLS 0:35aa5be3b78d 129 "mulq %7 \n\t" \
MiniTLS 0:35aa5be3b78d 130 "movq %%rax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 131 "movq %%rdx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 132 "xorq %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 133 :"=r"(sc0), "=r"(sc1), "=r"(sc2): "0"(sc0), "1"(sc1), "2"(sc2), "g"(i), "g"(j) :"%rax","%rdx","%cc");
MiniTLS 0:35aa5be3b78d 134
MiniTLS 0:35aa5be3b78d 135 #define SQRADDAC(i, j) \
MiniTLS 0:35aa5be3b78d 136 asm( \
MiniTLS 0:35aa5be3b78d 137 "movq %6,%%rax \n\t" \
MiniTLS 0:35aa5be3b78d 138 "mulq %7 \n\t" \
MiniTLS 0:35aa5be3b78d 139 "addq %%rax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 140 "adcq %%rdx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 141 "adcq $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 142 :"=r"(sc0), "=r"(sc1), "=r"(sc2): "0"(sc0), "1"(sc1), "2"(sc2), "g"(i), "g"(j) :"%rax","%rdx","%cc");
MiniTLS 0:35aa5be3b78d 143
MiniTLS 0:35aa5be3b78d 144 #define SQRADDDB \
MiniTLS 0:35aa5be3b78d 145 asm( \
MiniTLS 0:35aa5be3b78d 146 "addq %6,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 147 "adcq %7,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 148 "adcq %8,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 149 "addq %6,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 150 "adcq %7,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 151 "adcq %8,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 152 :"=r"(c0), "=r"(c1), "=r"(c2) : "0"(c0), "1"(c1), "2"(c2), "r"(sc0), "r"(sc1), "r"(sc2) : "%cc");
MiniTLS 0:35aa5be3b78d 153
MiniTLS 0:35aa5be3b78d 154 #elif defined(TFM_SSE2)
MiniTLS 0:35aa5be3b78d 155
MiniTLS 0:35aa5be3b78d 156 /* SSE2 Optimized */
MiniTLS 0:35aa5be3b78d 157 #define COMBA_START
MiniTLS 0:35aa5be3b78d 158
MiniTLS 0:35aa5be3b78d 159 #define CLEAR_CARRY \
MiniTLS 0:35aa5be3b78d 160 c0 = c1 = c2 = 0;
MiniTLS 0:35aa5be3b78d 161
MiniTLS 0:35aa5be3b78d 162 #define COMBA_STORE(x) \
MiniTLS 0:35aa5be3b78d 163 x = c0;
MiniTLS 0:35aa5be3b78d 164
MiniTLS 0:35aa5be3b78d 165 #define COMBA_STORE2(x) \
MiniTLS 0:35aa5be3b78d 166 x = c1;
MiniTLS 0:35aa5be3b78d 167
MiniTLS 0:35aa5be3b78d 168 #define CARRY_FORWARD \
MiniTLS 0:35aa5be3b78d 169 do { c0 = c1; c1 = c2; c2 = 0; } while (0);
MiniTLS 0:35aa5be3b78d 170
MiniTLS 0:35aa5be3b78d 171 #define COMBA_FINI \
MiniTLS 0:35aa5be3b78d 172 asm("emms");
MiniTLS 0:35aa5be3b78d 173
MiniTLS 0:35aa5be3b78d 174 #define SQRADD(i, j) \
MiniTLS 0:35aa5be3b78d 175 asm( \
MiniTLS 0:35aa5be3b78d 176 "movd %6,%%mm0 \n\t" \
MiniTLS 0:35aa5be3b78d 177 "pmuludq %%mm0,%%mm0\n\t" \
MiniTLS 0:35aa5be3b78d 178 "movd %%mm0,%%eax \n\t" \
MiniTLS 0:35aa5be3b78d 179 "psrlq $32,%%mm0 \n\t" \
MiniTLS 0:35aa5be3b78d 180 "addl %%eax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 181 "movd %%mm0,%%eax \n\t" \
MiniTLS 0:35aa5be3b78d 182 "adcl %%eax,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 183 "adcl $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 184 :"=r"(c0), "=r"(c1), "=r"(c2): "0"(c0), "1"(c1), "2"(c2), "m"(i) :"%eax","%cc");
MiniTLS 0:35aa5be3b78d 185
MiniTLS 0:35aa5be3b78d 186 #define SQRADD2(i, j) \
MiniTLS 0:35aa5be3b78d 187 asm( \
MiniTLS 0:35aa5be3b78d 188 "movd %6,%%mm0 \n\t" \
MiniTLS 0:35aa5be3b78d 189 "movd %7,%%mm1 \n\t" \
MiniTLS 0:35aa5be3b78d 190 "pmuludq %%mm1,%%mm0\n\t" \
MiniTLS 0:35aa5be3b78d 191 "movd %%mm0,%%eax \n\t" \
MiniTLS 0:35aa5be3b78d 192 "psrlq $32,%%mm0 \n\t" \
MiniTLS 0:35aa5be3b78d 193 "movd %%mm0,%%edx \n\t" \
MiniTLS 0:35aa5be3b78d 194 "addl %%eax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 195 "adcl %%edx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 196 "adcl $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 197 "addl %%eax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 198 "adcl %%edx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 199 "adcl $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 200 :"=r"(c0), "=r"(c1), "=r"(c2): "0"(c0), "1"(c1), "2"(c2), "m"(i), "m"(j) :"%eax","%edx","%cc");
MiniTLS 0:35aa5be3b78d 201
MiniTLS 0:35aa5be3b78d 202 #define SQRADDSC(i, j) \
MiniTLS 0:35aa5be3b78d 203 asm( \
MiniTLS 0:35aa5be3b78d 204 "movd %6,%%mm0 \n\t" \
MiniTLS 0:35aa5be3b78d 205 "movd %7,%%mm1 \n\t" \
MiniTLS 0:35aa5be3b78d 206 "pmuludq %%mm1,%%mm0\n\t" \
MiniTLS 0:35aa5be3b78d 207 "movd %%mm0,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 208 "psrlq $32,%%mm0 \n\t" \
MiniTLS 0:35aa5be3b78d 209 "movd %%mm0,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 210 "xorl %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 211 :"=r"(sc0), "=r"(sc1), "=r"(sc2): "0"(sc0), "1"(sc1), "2"(sc2), "m"(i), "m"(j));
MiniTLS 0:35aa5be3b78d 212
MiniTLS 0:35aa5be3b78d 213 #define SQRADDAC(i, j) \
MiniTLS 0:35aa5be3b78d 214 asm( \
MiniTLS 0:35aa5be3b78d 215 "movd %6,%%mm0 \n\t" \
MiniTLS 0:35aa5be3b78d 216 "movd %7,%%mm1 \n\t" \
MiniTLS 0:35aa5be3b78d 217 "pmuludq %%mm1,%%mm0\n\t" \
MiniTLS 0:35aa5be3b78d 218 "movd %%mm0,%%eax \n\t" \
MiniTLS 0:35aa5be3b78d 219 "psrlq $32,%%mm0 \n\t" \
MiniTLS 0:35aa5be3b78d 220 "movd %%mm0,%%edx \n\t" \
MiniTLS 0:35aa5be3b78d 221 "addl %%eax,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 222 "adcl %%edx,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 223 "adcl $0,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 224 :"=r"(sc0), "=r"(sc1), "=r"(sc2): "0"(sc0), "1"(sc1), "2"(sc2), "m"(i), "m"(j) :"%eax","%edx","%cc");
MiniTLS 0:35aa5be3b78d 225
MiniTLS 0:35aa5be3b78d 226 #define SQRADDDB \
MiniTLS 0:35aa5be3b78d 227 asm( \
MiniTLS 0:35aa5be3b78d 228 "addl %6,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 229 "adcl %7,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 230 "adcl %8,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 231 "addl %6,%0 \n\t" \
MiniTLS 0:35aa5be3b78d 232 "adcl %7,%1 \n\t" \
MiniTLS 0:35aa5be3b78d 233 "adcl %8,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 234 :"=r"(c0), "=r"(c1), "=r"(c2) : "0"(c0), "1"(c1), "2"(c2), "r"(sc0), "r"(sc1), "r"(sc2) : "%cc");
MiniTLS 0:35aa5be3b78d 235
MiniTLS 0:35aa5be3b78d 236 #elif defined(TFM_ARM)
MiniTLS 0:35aa5be3b78d 237
MiniTLS 0:35aa5be3b78d 238 /* ARM code */
MiniTLS 0:35aa5be3b78d 239
MiniTLS 0:35aa5be3b78d 240 #define COMBA_START
MiniTLS 0:35aa5be3b78d 241
MiniTLS 0:35aa5be3b78d 242 #define CLEAR_CARRY \
MiniTLS 0:35aa5be3b78d 243 c0 = c1 = c2 = 0;
MiniTLS 0:35aa5be3b78d 244
MiniTLS 0:35aa5be3b78d 245 #define COMBA_STORE(x) \
MiniTLS 0:35aa5be3b78d 246 x = c0;
MiniTLS 0:35aa5be3b78d 247
MiniTLS 0:35aa5be3b78d 248 #define COMBA_STORE2(x) \
MiniTLS 0:35aa5be3b78d 249 x = c1;
MiniTLS 0:35aa5be3b78d 250
MiniTLS 0:35aa5be3b78d 251 #define CARRY_FORWARD \
MiniTLS 0:35aa5be3b78d 252 do { c0 = c1; c1 = c2; c2 = 0; } while (0);
MiniTLS 0:35aa5be3b78d 253
MiniTLS 0:35aa5be3b78d 254 #define COMBA_FINI
MiniTLS 0:35aa5be3b78d 255
MiniTLS 0:35aa5be3b78d 256 /* multiplies point i and j, updates carry "c1" and digit c2 */
MiniTLS 0:35aa5be3b78d 257 #define SQRADD(i, j) \
MiniTLS 0:35aa5be3b78d 258 asm( \
MiniTLS 0:35aa5be3b78d 259 " UMULL r0,r1,%6,%6 \n\t" \
MiniTLS 0:35aa5be3b78d 260 " ADDS %0,%0,r0 \n\t" \
MiniTLS 0:35aa5be3b78d 261 " ADCS %1,%1,r1 \n\t" \
MiniTLS 0:35aa5be3b78d 262 " ADC %2,%2,#0 \n\t" \
MiniTLS 0:35aa5be3b78d 263 :"=r"(c0), "=r"(c1), "=r"(c2) : "0"(c0), "1"(c1), "2"(c2), "r"(i) : "r0", "r1", "%cc");
MiniTLS 0:35aa5be3b78d 264
MiniTLS 0:35aa5be3b78d 265 /* for squaring some of the terms are doubled... */
MiniTLS 0:35aa5be3b78d 266 #define SQRADD2(i, j) \
MiniTLS 0:35aa5be3b78d 267 asm( \
MiniTLS 0:35aa5be3b78d 268 " UMULL r0,r1,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 269 " ADDS %0,%0,r0 \n\t" \
MiniTLS 0:35aa5be3b78d 270 " ADCS %1,%1,r1 \n\t" \
MiniTLS 0:35aa5be3b78d 271 " ADC %2,%2,#0 \n\t" \
MiniTLS 0:35aa5be3b78d 272 " ADDS %0,%0,r0 \n\t" \
MiniTLS 0:35aa5be3b78d 273 " ADCS %1,%1,r1 \n\t" \
MiniTLS 0:35aa5be3b78d 274 " ADC %2,%2,#0 \n\t" \
MiniTLS 0:35aa5be3b78d 275 :"=r"(c0), "=r"(c1), "=r"(c2) : "0"(c0), "1"(c1), "2"(c2), "r"(i), "r"(j) : "r0", "r1", "%cc");
MiniTLS 0:35aa5be3b78d 276
MiniTLS 0:35aa5be3b78d 277 #define SQRADDSC(i, j) \
MiniTLS 0:35aa5be3b78d 278 asm( \
MiniTLS 0:35aa5be3b78d 279 " UMULL %0,%1,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 280 " SUB %2,%2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 281 :"=r"(sc0), "=r"(sc1), "=r"(sc2) : "0"(sc0), "1"(sc1), "2"(sc2), "r"(i), "r"(j) : "%cc");
MiniTLS 0:35aa5be3b78d 282
MiniTLS 0:35aa5be3b78d 283 #define SQRADDAC(i, j) \
MiniTLS 0:35aa5be3b78d 284 asm( \
MiniTLS 0:35aa5be3b78d 285 " UMULL r0,r1,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 286 " ADDS %0,%0,r0 \n\t" \
MiniTLS 0:35aa5be3b78d 287 " ADCS %1,%1,r1 \n\t" \
MiniTLS 0:35aa5be3b78d 288 " ADC %2,%2,#0 \n\t" \
MiniTLS 0:35aa5be3b78d 289 :"=r"(sc0), "=r"(sc1), "=r"(sc2) : "0"(sc0), "1"(sc1), "2"(sc2), "r"(i), "r"(j) : "r0", "r1", "%cc");
MiniTLS 0:35aa5be3b78d 290
MiniTLS 0:35aa5be3b78d 291 #define SQRADDDB \
MiniTLS 0:35aa5be3b78d 292 asm( \
MiniTLS 0:35aa5be3b78d 293 " ADDS %0,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 294 " ADCS %1,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 295 " ADC %2,%2,%5 \n\t" \
MiniTLS 0:35aa5be3b78d 296 " ADDS %0,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 297 " ADCS %1,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 298 " ADC %2,%2,%5 \n\t" \
MiniTLS 0:35aa5be3b78d 299 :"=r"(c0), "=r"(c1), "=r"(c2) : "r"(sc0), "r"(sc1), "r"(sc2), "0"(c0), "1"(c1), "2"(c2) : "%cc");
MiniTLS 0:35aa5be3b78d 300
MiniTLS 0:35aa5be3b78d 301 #elif defined(TFM_PPC32)
MiniTLS 0:35aa5be3b78d 302
MiniTLS 0:35aa5be3b78d 303 /* PPC32 */
MiniTLS 0:35aa5be3b78d 304
MiniTLS 0:35aa5be3b78d 305 #define COMBA_START
MiniTLS 0:35aa5be3b78d 306
MiniTLS 0:35aa5be3b78d 307 #define CLEAR_CARRY \
MiniTLS 0:35aa5be3b78d 308 c0 = c1 = c2 = 0;
MiniTLS 0:35aa5be3b78d 309
MiniTLS 0:35aa5be3b78d 310 #define COMBA_STORE(x) \
MiniTLS 0:35aa5be3b78d 311 x = c0;
MiniTLS 0:35aa5be3b78d 312
MiniTLS 0:35aa5be3b78d 313 #define COMBA_STORE2(x) \
MiniTLS 0:35aa5be3b78d 314 x = c1;
MiniTLS 0:35aa5be3b78d 315
MiniTLS 0:35aa5be3b78d 316 #define CARRY_FORWARD \
MiniTLS 0:35aa5be3b78d 317 do { c0 = c1; c1 = c2; c2 = 0; } while (0);
MiniTLS 0:35aa5be3b78d 318
MiniTLS 0:35aa5be3b78d 319 #define COMBA_FINI
MiniTLS 0:35aa5be3b78d 320
MiniTLS 0:35aa5be3b78d 321 /* multiplies point i and j, updates carry "c1" and digit c2 */
MiniTLS 0:35aa5be3b78d 322 #define SQRADD(i, j) \
MiniTLS 0:35aa5be3b78d 323 asm( \
MiniTLS 0:35aa5be3b78d 324 " mullw 16,%6,%6 \n\t" \
MiniTLS 0:35aa5be3b78d 325 " addc %0,%0,16 \n\t" \
MiniTLS 0:35aa5be3b78d 326 " mulhwu 16,%6,%6 \n\t" \
MiniTLS 0:35aa5be3b78d 327 " adde %1,%1,16 \n\t" \
MiniTLS 0:35aa5be3b78d 328 " addze %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 329 :"=r"(c0), "=r"(c1), "=r"(c2):"0"(c0), "1"(c1), "2"(c2), "r"(i):"16","%cc");
MiniTLS 0:35aa5be3b78d 330
MiniTLS 0:35aa5be3b78d 331 /* for squaring some of the terms are doubled... */
MiniTLS 0:35aa5be3b78d 332 #define SQRADD2(i, j) \
MiniTLS 0:35aa5be3b78d 333 asm( \
MiniTLS 0:35aa5be3b78d 334 " mullw 16,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 335 " mulhwu 17,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 336 " addc %0,%0,16 \n\t" \
MiniTLS 0:35aa5be3b78d 337 " adde %1,%1,17 \n\t" \
MiniTLS 0:35aa5be3b78d 338 " addze %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 339 " addc %0,%0,16 \n\t" \
MiniTLS 0:35aa5be3b78d 340 " adde %1,%1,17 \n\t" \
MiniTLS 0:35aa5be3b78d 341 " addze %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 342 :"=r"(c0), "=r"(c1), "=r"(c2):"0"(c0), "1"(c1), "2"(c2), "r"(i), "r"(j):"16", "17","%cc");
MiniTLS 0:35aa5be3b78d 343
MiniTLS 0:35aa5be3b78d 344 #define SQRADDSC(i, j) \
MiniTLS 0:35aa5be3b78d 345 asm( \
MiniTLS 0:35aa5be3b78d 346 " mullw %0,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 347 " mulhwu %1,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 348 " xor %2,%2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 349 :"=r"(sc0), "=r"(sc1), "=r"(sc2):"0"(sc0), "1"(sc1), "2"(sc2), "r"(i),"r"(j) : "%cc");
MiniTLS 0:35aa5be3b78d 350
MiniTLS 0:35aa5be3b78d 351 #define SQRADDAC(i, j) \
MiniTLS 0:35aa5be3b78d 352 asm( \
MiniTLS 0:35aa5be3b78d 353 " mullw 16,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 354 " addc %0,%0,16 \n\t" \
MiniTLS 0:35aa5be3b78d 355 " mulhwu 16,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 356 " adde %1,%1,16 \n\t" \
MiniTLS 0:35aa5be3b78d 357 " addze %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 358 :"=r"(sc0), "=r"(sc1), "=r"(sc2):"0"(sc0), "1"(sc1), "2"(sc2), "r"(i), "r"(j):"16", "%cc");
MiniTLS 0:35aa5be3b78d 359
MiniTLS 0:35aa5be3b78d 360 #define SQRADDDB \
MiniTLS 0:35aa5be3b78d 361 asm( \
MiniTLS 0:35aa5be3b78d 362 " addc %0,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 363 " adde %1,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 364 " adde %2,%2,%5 \n\t" \
MiniTLS 0:35aa5be3b78d 365 " addc %0,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 366 " adde %1,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 367 " adde %2,%2,%5 \n\t" \
MiniTLS 0:35aa5be3b78d 368 :"=r"(c0), "=r"(c1), "=r"(c2) : "r"(sc0), "r"(sc1), "r"(sc2), "0"(c0), "1"(c1), "2"(c2) : "%cc");
MiniTLS 0:35aa5be3b78d 369
MiniTLS 0:35aa5be3b78d 370 #elif defined(TFM_PPC64)
MiniTLS 0:35aa5be3b78d 371 /* PPC64 */
MiniTLS 0:35aa5be3b78d 372
MiniTLS 0:35aa5be3b78d 373 #define COMBA_START
MiniTLS 0:35aa5be3b78d 374
MiniTLS 0:35aa5be3b78d 375 #define CLEAR_CARRY \
MiniTLS 0:35aa5be3b78d 376 c0 = c1 = c2 = 0;
MiniTLS 0:35aa5be3b78d 377
MiniTLS 0:35aa5be3b78d 378 #define COMBA_STORE(x) \
MiniTLS 0:35aa5be3b78d 379 x = c0;
MiniTLS 0:35aa5be3b78d 380
MiniTLS 0:35aa5be3b78d 381 #define COMBA_STORE2(x) \
MiniTLS 0:35aa5be3b78d 382 x = c1;
MiniTLS 0:35aa5be3b78d 383
MiniTLS 0:35aa5be3b78d 384 #define CARRY_FORWARD \
MiniTLS 0:35aa5be3b78d 385 do { c0 = c1; c1 = c2; c2 = 0; } while (0);
MiniTLS 0:35aa5be3b78d 386
MiniTLS 0:35aa5be3b78d 387 #define COMBA_FINI
MiniTLS 0:35aa5be3b78d 388
MiniTLS 0:35aa5be3b78d 389 /* multiplies point i and j, updates carry "c1" and digit c2 */
MiniTLS 0:35aa5be3b78d 390 #define SQRADD(i, j) \
MiniTLS 0:35aa5be3b78d 391 asm( \
MiniTLS 0:35aa5be3b78d 392 " mulld r16,%6,%6 \n\t" \
MiniTLS 0:35aa5be3b78d 393 " addc %0,%0,r16 \n\t" \
MiniTLS 0:35aa5be3b78d 394 " mulhdu r16,%6,%6 \n\t" \
MiniTLS 0:35aa5be3b78d 395 " adde %1,%1,r16 \n\t" \
MiniTLS 0:35aa5be3b78d 396 " addze %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 397 :"=r"(c0), "=r"(c1), "=r"(c2):"0"(c0), "1"(c1), "2"(c2), "r"(i):"r16","%cc");
MiniTLS 0:35aa5be3b78d 398
MiniTLS 0:35aa5be3b78d 399 /* for squaring some of the terms are doubled... */
MiniTLS 0:35aa5be3b78d 400 #define SQRADD2(i, j) \
MiniTLS 0:35aa5be3b78d 401 asm( \
MiniTLS 0:35aa5be3b78d 402 " mulld r16,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 403 " mulhdu r17,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 404 " addc %0,%0,r16 \n\t" \
MiniTLS 0:35aa5be3b78d 405 " adde %1,%1,r17 \n\t" \
MiniTLS 0:35aa5be3b78d 406 " addze %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 407 " addc %0,%0,r16 \n\t" \
MiniTLS 0:35aa5be3b78d 408 " adde %1,%1,r17 \n\t" \
MiniTLS 0:35aa5be3b78d 409 " addze %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 410 :"=r"(c0), "=r"(c1), "=r"(c2):"0"(c0), "1"(c1), "2"(c2), "r"(i), "r"(j):"r16", "r17","%cc");
MiniTLS 0:35aa5be3b78d 411
MiniTLS 0:35aa5be3b78d 412 #define SQRADDSC(i, j) \
MiniTLS 0:35aa5be3b78d 413 asm( \
MiniTLS 0:35aa5be3b78d 414 " mulld %0,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 415 " mulhdu %1,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 416 " xor %2,%2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 417 :"=r"(sc0), "=r"(sc1), "=r"(sc2):"0"(sc0), "1"(sc1), "2"(sc2), "r"(i),"r"(j) : "%cc");
MiniTLS 0:35aa5be3b78d 418
MiniTLS 0:35aa5be3b78d 419 #define SQRADDAC(i, j) \
MiniTLS 0:35aa5be3b78d 420 asm( \
MiniTLS 0:35aa5be3b78d 421 " mulld r16,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 422 " addc %0,%0,r16 \n\t" \
MiniTLS 0:35aa5be3b78d 423 " mulhdu r16,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 424 " adde %1,%1,r16 \n\t" \
MiniTLS 0:35aa5be3b78d 425 " addze %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 426 :"=r"(sc0), "=r"(sc1), "=r"(sc2):"0"(sc0), "1"(sc1), "2"(sc2), "r"(i), "r"(j):"r16", "%cc");
MiniTLS 0:35aa5be3b78d 427
MiniTLS 0:35aa5be3b78d 428 #define SQRADDDB \
MiniTLS 0:35aa5be3b78d 429 asm( \
MiniTLS 0:35aa5be3b78d 430 " addc %0,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 431 " adde %1,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 432 " adde %2,%2,%5 \n\t" \
MiniTLS 0:35aa5be3b78d 433 " addc %0,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 434 " adde %1,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 435 " adde %2,%2,%5 \n\t" \
MiniTLS 0:35aa5be3b78d 436 :"=r"(c0), "=r"(c1), "=r"(c2) : "r"(sc0), "r"(sc1), "r"(sc2), "0"(c0), "1"(c1), "2"(c2) : "%cc");
MiniTLS 0:35aa5be3b78d 437
MiniTLS 0:35aa5be3b78d 438
MiniTLS 0:35aa5be3b78d 439 #elif defined(TFM_AVR32)
MiniTLS 0:35aa5be3b78d 440
MiniTLS 0:35aa5be3b78d 441 /* AVR32 */
MiniTLS 0:35aa5be3b78d 442
MiniTLS 0:35aa5be3b78d 443 #define COMBA_START
MiniTLS 0:35aa5be3b78d 444
MiniTLS 0:35aa5be3b78d 445 #define CLEAR_CARRY \
MiniTLS 0:35aa5be3b78d 446 c0 = c1 = c2 = 0;
MiniTLS 0:35aa5be3b78d 447
MiniTLS 0:35aa5be3b78d 448 #define COMBA_STORE(x) \
MiniTLS 0:35aa5be3b78d 449 x = c0;
MiniTLS 0:35aa5be3b78d 450
MiniTLS 0:35aa5be3b78d 451 #define COMBA_STORE2(x) \
MiniTLS 0:35aa5be3b78d 452 x = c1;
MiniTLS 0:35aa5be3b78d 453
MiniTLS 0:35aa5be3b78d 454 #define CARRY_FORWARD \
MiniTLS 0:35aa5be3b78d 455 do { c0 = c1; c1 = c2; c2 = 0; } while (0);
MiniTLS 0:35aa5be3b78d 456
MiniTLS 0:35aa5be3b78d 457 #define COMBA_FINI
MiniTLS 0:35aa5be3b78d 458
MiniTLS 0:35aa5be3b78d 459 /* multiplies point i and j, updates carry "c1" and digit c2 */
MiniTLS 0:35aa5be3b78d 460 #define SQRADD(i, j) \
MiniTLS 0:35aa5be3b78d 461 asm( \
MiniTLS 0:35aa5be3b78d 462 " mulu.d r2,%6,%6 \n\t" \
MiniTLS 0:35aa5be3b78d 463 " add %0,%0,r2 \n\t" \
MiniTLS 0:35aa5be3b78d 464 " adc %1,%1,r3 \n\t" \
MiniTLS 0:35aa5be3b78d 465 " acr %2 \n\t" \
MiniTLS 0:35aa5be3b78d 466 :"=r"(c0), "=r"(c1), "=r"(c2):"0"(c0), "1"(c1), "2"(c2), "r"(i):"r2","r3");
MiniTLS 0:35aa5be3b78d 467
MiniTLS 0:35aa5be3b78d 468 /* for squaring some of the terms are doubled... */
MiniTLS 0:35aa5be3b78d 469 #define SQRADD2(i, j) \
MiniTLS 0:35aa5be3b78d 470 asm( \
MiniTLS 0:35aa5be3b78d 471 " mulu.d r2,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 472 " add %0,%0,r2 \n\t" \
MiniTLS 0:35aa5be3b78d 473 " adc %1,%1,r3 \n\t" \
MiniTLS 0:35aa5be3b78d 474 " acr %2, \n\t" \
MiniTLS 0:35aa5be3b78d 475 " add %0,%0,r2 \n\t" \
MiniTLS 0:35aa5be3b78d 476 " adc %1,%1,r3 \n\t" \
MiniTLS 0:35aa5be3b78d 477 " acr %2, \n\t" \
MiniTLS 0:35aa5be3b78d 478 :"=r"(c0), "=r"(c1), "=r"(c2):"0"(c0), "1"(c1), "2"(c2), "r"(i), "r"(j):"r2", "r3");
MiniTLS 0:35aa5be3b78d 479
MiniTLS 0:35aa5be3b78d 480 #define SQRADDSC(i, j) \
MiniTLS 0:35aa5be3b78d 481 asm( \
MiniTLS 0:35aa5be3b78d 482 " mulu.d r2,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 483 " mov %0,r2 \n\t" \
MiniTLS 0:35aa5be3b78d 484 " mov %1,r3 \n\t" \
MiniTLS 0:35aa5be3b78d 485 " eor %2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 486 :"=r"(sc0), "=r"(sc1), "=r"(sc2):"0"(sc0), "1"(sc1), "2"(sc2), "r"(i),"r"(j) : "r2", "r3");
MiniTLS 0:35aa5be3b78d 487
MiniTLS 0:35aa5be3b78d 488 #define SQRADDAC(i, j) \
MiniTLS 0:35aa5be3b78d 489 asm( \
MiniTLS 0:35aa5be3b78d 490 " mulu.d r2,%6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 491 " add %0,%0,r2 \n\t" \
MiniTLS 0:35aa5be3b78d 492 " adc %1,%1,r3 \n\t" \
MiniTLS 0:35aa5be3b78d 493 " acr %2 \n\t" \
MiniTLS 0:35aa5be3b78d 494 :"=r"(sc0), "=r"(sc1), "=r"(sc2):"0"(sc0), "1"(sc1), "2"(sc2), "r"(i), "r"(j):"r2", "r3");
MiniTLS 0:35aa5be3b78d 495
MiniTLS 0:35aa5be3b78d 496 #define SQRADDDB \
MiniTLS 0:35aa5be3b78d 497 asm( \
MiniTLS 0:35aa5be3b78d 498 " add %0,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 499 " adc %1,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 500 " adc %2,%2,%5 \n\t" \
MiniTLS 0:35aa5be3b78d 501 " add %0,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 502 " adc %1,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 503 " adc %2,%2,%5 \n\t" \
MiniTLS 0:35aa5be3b78d 504 :"=r"(c0), "=r"(c1), "=r"(c2) : "r"(sc0), "r"(sc1), "r"(sc2), "0"(c0), "1"(c1), "2"(c2) : "%cc");
MiniTLS 0:35aa5be3b78d 505
MiniTLS 0:35aa5be3b78d 506 #elif defined(TFM_MIPS)
MiniTLS 0:35aa5be3b78d 507
MiniTLS 0:35aa5be3b78d 508 /* MIPS */
MiniTLS 0:35aa5be3b78d 509
MiniTLS 0:35aa5be3b78d 510 #define COMBA_START
MiniTLS 0:35aa5be3b78d 511
MiniTLS 0:35aa5be3b78d 512 #define CLEAR_CARRY \
MiniTLS 0:35aa5be3b78d 513 c0 = c1 = c2 = 0;
MiniTLS 0:35aa5be3b78d 514
MiniTLS 0:35aa5be3b78d 515 #define COMBA_STORE(x) \
MiniTLS 0:35aa5be3b78d 516 x = c0;
MiniTLS 0:35aa5be3b78d 517
MiniTLS 0:35aa5be3b78d 518 #define COMBA_STORE2(x) \
MiniTLS 0:35aa5be3b78d 519 x = c1;
MiniTLS 0:35aa5be3b78d 520
MiniTLS 0:35aa5be3b78d 521 #define CARRY_FORWARD \
MiniTLS 0:35aa5be3b78d 522 do { c0 = c1; c1 = c2; c2 = 0; } while (0);
MiniTLS 0:35aa5be3b78d 523
MiniTLS 0:35aa5be3b78d 524 #define COMBA_FINI
MiniTLS 0:35aa5be3b78d 525
MiniTLS 0:35aa5be3b78d 526 /* multiplies point i and j, updates carry "c1" and digit c2 */
MiniTLS 0:35aa5be3b78d 527 #define SQRADD(i, j) \
MiniTLS 0:35aa5be3b78d 528 asm( \
MiniTLS 0:35aa5be3b78d 529 " multu %6,%6 \n\t" \
MiniTLS 0:35aa5be3b78d 530 " mflo $12 \n\t" \
MiniTLS 0:35aa5be3b78d 531 " mfhi $13 \n\t" \
MiniTLS 0:35aa5be3b78d 532 " addu %0,%0,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 533 " sltu $12,%0,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 534 " addu %1,%1,$13 \n\t" \
MiniTLS 0:35aa5be3b78d 535 " sltu $13,%1,$13 \n\t" \
MiniTLS 0:35aa5be3b78d 536 " addu %1,%1,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 537 " sltu $12,%1,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 538 " addu %2,%2,$13 \n\t" \
MiniTLS 0:35aa5be3b78d 539 " addu %2,%2,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 540 :"=r"(c0), "=r"(c1), "=r"(c2):"0"(c0), "1"(c1), "2"(c2), "r"(i):"$12","$13");
MiniTLS 0:35aa5be3b78d 541
MiniTLS 0:35aa5be3b78d 542 /* for squaring some of the terms are doubled... */
MiniTLS 0:35aa5be3b78d 543 #define SQRADD2(i, j) \
MiniTLS 0:35aa5be3b78d 544 asm( \
MiniTLS 0:35aa5be3b78d 545 " multu %6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 546 " mflo $12 \n\t" \
MiniTLS 0:35aa5be3b78d 547 " mfhi $13 \n\t" \
MiniTLS 0:35aa5be3b78d 548 \
MiniTLS 0:35aa5be3b78d 549 " addu %0,%0,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 550 " sltu $14,%0,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 551 " addu %1,%1,$13 \n\t" \
MiniTLS 0:35aa5be3b78d 552 " sltu $15,%1,$13 \n\t" \
MiniTLS 0:35aa5be3b78d 553 " addu %1,%1,$14 \n\t" \
MiniTLS 0:35aa5be3b78d 554 " sltu $14,%1,$14 \n\t" \
MiniTLS 0:35aa5be3b78d 555 " addu %2,%2,$15 \n\t" \
MiniTLS 0:35aa5be3b78d 556 " addu %2,%2,$14 \n\t" \
MiniTLS 0:35aa5be3b78d 557 \
MiniTLS 0:35aa5be3b78d 558 " addu %0,%0,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 559 " sltu $14,%0,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 560 " addu %1,%1,$13 \n\t" \
MiniTLS 0:35aa5be3b78d 561 " sltu $15,%1,$13 \n\t" \
MiniTLS 0:35aa5be3b78d 562 " addu %1,%1,$14 \n\t" \
MiniTLS 0:35aa5be3b78d 563 " sltu $14,%1,$14 \n\t" \
MiniTLS 0:35aa5be3b78d 564 " addu %2,%2,$15 \n\t" \
MiniTLS 0:35aa5be3b78d 565 " addu %2,%2,$14 \n\t" \
MiniTLS 0:35aa5be3b78d 566 :"=r"(c0), "=r"(c1), "=r"(c2):"0"(c0), "1"(c1), "2"(c2), "r"(i), "r"(j):"$12", "$13", "$14", "$15");
MiniTLS 0:35aa5be3b78d 567
MiniTLS 0:35aa5be3b78d 568 #define SQRADDSC(i, j) \
MiniTLS 0:35aa5be3b78d 569 asm( \
MiniTLS 0:35aa5be3b78d 570 " multu %6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 571 " mflo %0 \n\t" \
MiniTLS 0:35aa5be3b78d 572 " mfhi %1 \n\t" \
MiniTLS 0:35aa5be3b78d 573 " xor %2,%2,%2 \n\t" \
MiniTLS 0:35aa5be3b78d 574 :"=r"(sc0), "=r"(sc1), "=r"(sc2):"0"(sc0), "1"(sc1), "2"(sc2), "r"(i),"r"(j) : "%cc");
MiniTLS 0:35aa5be3b78d 575
MiniTLS 0:35aa5be3b78d 576 #define SQRADDAC(i, j) \
MiniTLS 0:35aa5be3b78d 577 asm( \
MiniTLS 0:35aa5be3b78d 578 " multu %6,%7 \n\t" \
MiniTLS 0:35aa5be3b78d 579 " mflo $12 \n\t" \
MiniTLS 0:35aa5be3b78d 580 " mfhi $13 \n\t" \
MiniTLS 0:35aa5be3b78d 581 " addu %0,%0,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 582 " sltu $12,%0,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 583 " addu %1,%1,$13 \n\t" \
MiniTLS 0:35aa5be3b78d 584 " sltu $13,%1,$13 \n\t" \
MiniTLS 0:35aa5be3b78d 585 " addu %1,%1,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 586 " sltu $12,%1,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 587 " addu %2,%2,$13 \n\t" \
MiniTLS 0:35aa5be3b78d 588 " addu %2,%2,$12 \n\t" \
MiniTLS 0:35aa5be3b78d 589 :"=r"(sc0), "=r"(sc1), "=r"(sc2):"0"(sc0), "1"(sc1), "2"(sc2), "r"(i), "r"(j):"$12", "$13", "$14");
MiniTLS 0:35aa5be3b78d 590
MiniTLS 0:35aa5be3b78d 591 #define SQRADDDB \
MiniTLS 0:35aa5be3b78d 592 asm( \
MiniTLS 0:35aa5be3b78d 593 " addu %0,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 594 " sltu $10,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 595 " addu %1,%1,$10 \n\t" \
MiniTLS 0:35aa5be3b78d 596 " sltu $10,%1,$10 \n\t" \
MiniTLS 0:35aa5be3b78d 597 " addu %1,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 598 " sltu $11,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 599 " addu %2,%2,$10 \n\t" \
MiniTLS 0:35aa5be3b78d 600 " addu %2,%2,$11 \n\t" \
MiniTLS 0:35aa5be3b78d 601 " addu %2,%2,%5 \n\t" \
MiniTLS 0:35aa5be3b78d 602 \
MiniTLS 0:35aa5be3b78d 603 " addu %0,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 604 " sltu $10,%0,%3 \n\t" \
MiniTLS 0:35aa5be3b78d 605 " addu %1,%1,$10 \n\t" \
MiniTLS 0:35aa5be3b78d 606 " sltu $10,%1,$10 \n\t" \
MiniTLS 0:35aa5be3b78d 607 " addu %1,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 608 " sltu $11,%1,%4 \n\t" \
MiniTLS 0:35aa5be3b78d 609 " addu %2,%2,$10 \n\t" \
MiniTLS 0:35aa5be3b78d 610 " addu %2,%2,$11 \n\t" \
MiniTLS 0:35aa5be3b78d 611 " addu %2,%2,%5 \n\t" \
MiniTLS 0:35aa5be3b78d 612 :"=r"(c0), "=r"(c1), "=r"(c2) : "r"(sc0), "r"(sc1), "r"(sc2), "0"(c0), "1"(c1), "2"(c2) : "$10", "$11");
MiniTLS 0:35aa5be3b78d 613
MiniTLS 0:35aa5be3b78d 614 #else
MiniTLS 0:35aa5be3b78d 615
MiniTLS 0:35aa5be3b78d 616 #define TFM_ISO
MiniTLS 0:35aa5be3b78d 617
MiniTLS 0:35aa5be3b78d 618 /* ISO C portable code */
MiniTLS 0:35aa5be3b78d 619
MiniTLS 0:35aa5be3b78d 620 #define COMBA_START
MiniTLS 0:35aa5be3b78d 621
MiniTLS 0:35aa5be3b78d 622 #define CLEAR_CARRY \
MiniTLS 0:35aa5be3b78d 623 c0 = c1 = c2 = 0;
MiniTLS 0:35aa5be3b78d 624
MiniTLS 0:35aa5be3b78d 625 #define COMBA_STORE(x) \
MiniTLS 0:35aa5be3b78d 626 x = c0;
MiniTLS 0:35aa5be3b78d 627
MiniTLS 0:35aa5be3b78d 628 #define COMBA_STORE2(x) \
MiniTLS 0:35aa5be3b78d 629 x = c1;
MiniTLS 0:35aa5be3b78d 630
MiniTLS 0:35aa5be3b78d 631 #define CARRY_FORWARD \
MiniTLS 0:35aa5be3b78d 632 do { c0 = c1; c1 = c2; c2 = 0; } while (0);
MiniTLS 0:35aa5be3b78d 633
MiniTLS 0:35aa5be3b78d 634 #define COMBA_FINI
MiniTLS 0:35aa5be3b78d 635
MiniTLS 0:35aa5be3b78d 636 /* multiplies point i and j, updates carry "c1" and digit c2 */
MiniTLS 0:35aa5be3b78d 637 #define SQRADD(i, j) \
MiniTLS 0:35aa5be3b78d 638 do { fp_word t; \
MiniTLS 0:35aa5be3b78d 639 t = c0 + ((fp_word)i) * ((fp_word)j); c0 = t; \
MiniTLS 0:35aa5be3b78d 640 t = c1 + (t >> DIGIT_BIT); c1 = t; c2 += t >> DIGIT_BIT; \
MiniTLS 0:35aa5be3b78d 641 } while (0);
MiniTLS 0:35aa5be3b78d 642
MiniTLS 0:35aa5be3b78d 643
MiniTLS 0:35aa5be3b78d 644 /* for squaring some of the terms are doubled... */
MiniTLS 0:35aa5be3b78d 645 #define SQRADD2(i, j) \
MiniTLS 0:35aa5be3b78d 646 do { fp_word t; \
MiniTLS 0:35aa5be3b78d 647 t = ((fp_word)i) * ((fp_word)j); \
MiniTLS 0:35aa5be3b78d 648 tt = (fp_word)c0 + t; c0 = tt; \
MiniTLS 0:35aa5be3b78d 649 tt = (fp_word)c1 + (tt >> DIGIT_BIT); c1 = tt; c2 += tt >> DIGIT_BIT; \
MiniTLS 0:35aa5be3b78d 650 tt = (fp_word)c0 + t; c0 = tt; \
MiniTLS 0:35aa5be3b78d 651 tt = (fp_word)c1 + (tt >> DIGIT_BIT); c1 = tt; c2 += tt >> DIGIT_BIT; \
MiniTLS 0:35aa5be3b78d 652 } while (0);
MiniTLS 0:35aa5be3b78d 653
MiniTLS 0:35aa5be3b78d 654 #define SQRADDSC(i, j) \
MiniTLS 0:35aa5be3b78d 655 do { fp_word t; \
MiniTLS 0:35aa5be3b78d 656 t = ((fp_word)i) * ((fp_word)j); \
MiniTLS 0:35aa5be3b78d 657 sc0 = (fp_digit)t; sc1 = (t >> DIGIT_BIT); sc2 = 0; \
MiniTLS 0:35aa5be3b78d 658 } while (0);
MiniTLS 0:35aa5be3b78d 659
MiniTLS 0:35aa5be3b78d 660 #define SQRADDAC(i, j) \
MiniTLS 0:35aa5be3b78d 661 do { fp_word t; \
MiniTLS 0:35aa5be3b78d 662 t = sc0 + ((fp_word)i) * ((fp_word)j); sc0 = t; \
MiniTLS 0:35aa5be3b78d 663 t = sc1 + (t >> DIGIT_BIT); sc1 = t; sc2 += t >> DIGIT_BIT; \
MiniTLS 0:35aa5be3b78d 664 } while (0);
MiniTLS 0:35aa5be3b78d 665
MiniTLS 0:35aa5be3b78d 666 #define SQRADDDB \
MiniTLS 0:35aa5be3b78d 667 do { fp_word t; \
MiniTLS 0:35aa5be3b78d 668 t = ((fp_word)sc0) + ((fp_word)sc0) + c0; c0 = t; \
MiniTLS 0:35aa5be3b78d 669 t = ((fp_word)sc1) + ((fp_word)sc1) + c1 + (t >> DIGIT_BIT); c1 = t; \
MiniTLS 0:35aa5be3b78d 670 c2 = c2 + ((fp_word)sc2) + ((fp_word)sc2) + (t >> DIGIT_BIT); \
MiniTLS 0:35aa5be3b78d 671 } while (0);
MiniTLS 0:35aa5be3b78d 672
MiniTLS 0:35aa5be3b78d 673 #endif
MiniTLS 0:35aa5be3b78d 674
MiniTLS 0:35aa5be3b78d 675 /* $Source: /cvs/libtom/tomsfastmath/src/sqr/fp_sqr_comba.c,v $ */
MiniTLS 0:35aa5be3b78d 676 /* $Revision: 1.4 $ */
MiniTLS 0:35aa5be3b78d 677 /* $Date: 2007/03/14 23:47:42 $ */