Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* multi_arith.h: multi-precision integer arithmetic functions, needed |
| 2 | to do extended-precision floating point. |
| 3 | |
| 4 | (c) 1998 David Huggins-Daines. |
| 5 | |
| 6 | Somewhat based on arch/alpha/math-emu/ieee-math.c, which is (c) |
| 7 | David Mosberger-Tang. |
| 8 | |
| 9 | You may copy, modify, and redistribute this file under the terms of |
| 10 | the GNU General Public License, version 2, or any later version, at |
| 11 | your convenience. */ |
| 12 | |
| 13 | /* Note: |
| 14 | |
| 15 | These are not general multi-precision math routines. Rather, they |
| 16 | implement the subset of integer arithmetic that we need in order to |
| 17 | multiply, divide, and normalize 128-bit unsigned mantissae. */ |
| 18 | |
| 19 | #ifndef MULTI_ARITH_H |
| 20 | #define MULTI_ARITH_H |
| 21 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 22 | static inline void fp_denormalize(struct fp_ext *reg, unsigned int cnt) |
| 23 | { |
| 24 | reg->exp += cnt; |
| 25 | |
| 26 | switch (cnt) { |
| 27 | case 0 ... 8: |
| 28 | reg->lowmant = reg->mant.m32[1] << (8 - cnt); |
| 29 | reg->mant.m32[1] = (reg->mant.m32[1] >> cnt) | |
| 30 | (reg->mant.m32[0] << (32 - cnt)); |
| 31 | reg->mant.m32[0] = reg->mant.m32[0] >> cnt; |
| 32 | break; |
| 33 | case 9 ... 32: |
| 34 | reg->lowmant = reg->mant.m32[1] >> (cnt - 8); |
| 35 | if (reg->mant.m32[1] << (40 - cnt)) |
| 36 | reg->lowmant |= 1; |
| 37 | reg->mant.m32[1] = (reg->mant.m32[1] >> cnt) | |
| 38 | (reg->mant.m32[0] << (32 - cnt)); |
| 39 | reg->mant.m32[0] = reg->mant.m32[0] >> cnt; |
| 40 | break; |
| 41 | case 33 ... 39: |
| 42 | asm volatile ("bfextu %1{%2,#8},%0" : "=d" (reg->lowmant) |
| 43 | : "m" (reg->mant.m32[0]), "d" (64 - cnt)); |
| 44 | if (reg->mant.m32[1] << (40 - cnt)) |
| 45 | reg->lowmant |= 1; |
| 46 | reg->mant.m32[1] = reg->mant.m32[0] >> (cnt - 32); |
| 47 | reg->mant.m32[0] = 0; |
| 48 | break; |
| 49 | case 40 ... 71: |
| 50 | reg->lowmant = reg->mant.m32[0] >> (cnt - 40); |
| 51 | if ((reg->mant.m32[0] << (72 - cnt)) || reg->mant.m32[1]) |
| 52 | reg->lowmant |= 1; |
| 53 | reg->mant.m32[1] = reg->mant.m32[0] >> (cnt - 32); |
| 54 | reg->mant.m32[0] = 0; |
| 55 | break; |
| 56 | default: |
| 57 | reg->lowmant = reg->mant.m32[0] || reg->mant.m32[1]; |
| 58 | reg->mant.m32[0] = 0; |
| 59 | reg->mant.m32[1] = 0; |
| 60 | break; |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | static inline int fp_overnormalize(struct fp_ext *reg) |
| 65 | { |
| 66 | int shift; |
| 67 | |
| 68 | if (reg->mant.m32[0]) { |
| 69 | asm ("bfffo %1{#0,#32},%0" : "=d" (shift) : "dm" (reg->mant.m32[0])); |
| 70 | reg->mant.m32[0] = (reg->mant.m32[0] << shift) | (reg->mant.m32[1] >> (32 - shift)); |
| 71 | reg->mant.m32[1] = (reg->mant.m32[1] << shift); |
| 72 | } else { |
| 73 | asm ("bfffo %1{#0,#32},%0" : "=d" (shift) : "dm" (reg->mant.m32[1])); |
| 74 | reg->mant.m32[0] = (reg->mant.m32[1] << shift); |
| 75 | reg->mant.m32[1] = 0; |
| 76 | shift += 32; |
| 77 | } |
| 78 | |
| 79 | return shift; |
| 80 | } |
| 81 | |
| 82 | static inline int fp_addmant(struct fp_ext *dest, struct fp_ext *src) |
| 83 | { |
| 84 | int carry; |
| 85 | |
| 86 | /* we assume here, gcc only insert move and a clr instr */ |
| 87 | asm volatile ("add.b %1,%0" : "=d,g" (dest->lowmant) |
| 88 | : "g,d" (src->lowmant), "0,0" (dest->lowmant)); |
| 89 | asm volatile ("addx.l %1,%0" : "=d" (dest->mant.m32[1]) |
| 90 | : "d" (src->mant.m32[1]), "0" (dest->mant.m32[1])); |
| 91 | asm volatile ("addx.l %1,%0" : "=d" (dest->mant.m32[0]) |
| 92 | : "d" (src->mant.m32[0]), "0" (dest->mant.m32[0])); |
| 93 | asm volatile ("addx.l %0,%0" : "=d" (carry) : "0" (0)); |
| 94 | |
| 95 | return carry; |
| 96 | } |
| 97 | |
| 98 | static inline int fp_addcarry(struct fp_ext *reg) |
| 99 | { |
| 100 | if (++reg->exp == 0x7fff) { |
| 101 | if (reg->mant.m64) |
| 102 | fp_set_sr(FPSR_EXC_INEX2); |
| 103 | reg->mant.m64 = 0; |
| 104 | fp_set_sr(FPSR_EXC_OVFL); |
| 105 | return 0; |
| 106 | } |
| 107 | reg->lowmant = (reg->mant.m32[1] << 7) | (reg->lowmant ? 1 : 0); |
| 108 | reg->mant.m32[1] = (reg->mant.m32[1] >> 1) | |
| 109 | (reg->mant.m32[0] << 31); |
| 110 | reg->mant.m32[0] = (reg->mant.m32[0] >> 1) | 0x80000000; |
| 111 | |
| 112 | return 1; |
| 113 | } |
| 114 | |
| 115 | static inline void fp_submant(struct fp_ext *dest, struct fp_ext *src1, |
| 116 | struct fp_ext *src2) |
| 117 | { |
| 118 | /* we assume here, gcc only insert move and a clr instr */ |
| 119 | asm volatile ("sub.b %1,%0" : "=d,g" (dest->lowmant) |
| 120 | : "g,d" (src2->lowmant), "0,0" (src1->lowmant)); |
| 121 | asm volatile ("subx.l %1,%0" : "=d" (dest->mant.m32[1]) |
| 122 | : "d" (src2->mant.m32[1]), "0" (src1->mant.m32[1])); |
| 123 | asm volatile ("subx.l %1,%0" : "=d" (dest->mant.m32[0]) |
| 124 | : "d" (src2->mant.m32[0]), "0" (src1->mant.m32[0])); |
| 125 | } |
| 126 | |
| 127 | #define fp_mul64(desth, destl, src1, src2) ({ \ |
| 128 | asm ("mulu.l %2,%1:%0" : "=d" (destl), "=d" (desth) \ |
Al Viro | 7f26333 | 2006-01-12 01:06:19 -0800 | [diff] [blame] | 129 | : "dm" (src1), "0" (src2)); \ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 130 | }) |
| 131 | #define fp_div64(quot, rem, srch, srcl, div) \ |
| 132 | asm ("divu.l %2,%1:%0" : "=d" (quot), "=d" (rem) \ |
| 133 | : "dm" (div), "1" (srch), "0" (srcl)) |
| 134 | #define fp_add64(dest1, dest2, src1, src2) ({ \ |
| 135 | asm ("add.l %1,%0" : "=d,dm" (dest2) \ |
| 136 | : "dm,d" (src2), "0,0" (dest2)); \ |
| 137 | asm ("addx.l %1,%0" : "=d" (dest1) \ |
| 138 | : "d" (src1), "0" (dest1)); \ |
| 139 | }) |
| 140 | #define fp_addx96(dest, src) ({ \ |
| 141 | /* we assume here, gcc only insert move and a clr instr */ \ |
| 142 | asm volatile ("add.l %1,%0" : "=d,g" (dest->m32[2]) \ |
| 143 | : "g,d" (temp.m32[1]), "0,0" (dest->m32[2])); \ |
| 144 | asm volatile ("addx.l %1,%0" : "=d" (dest->m32[1]) \ |
| 145 | : "d" (temp.m32[0]), "0" (dest->m32[1])); \ |
| 146 | asm volatile ("addx.l %1,%0" : "=d" (dest->m32[0]) \ |
| 147 | : "d" (0), "0" (dest->m32[0])); \ |
| 148 | }) |
| 149 | #define fp_sub64(dest, src) ({ \ |
| 150 | asm ("sub.l %1,%0" : "=d,dm" (dest.m32[1]) \ |
| 151 | : "dm,d" (src.m32[1]), "0,0" (dest.m32[1])); \ |
| 152 | asm ("subx.l %1,%0" : "=d" (dest.m32[0]) \ |
| 153 | : "d" (src.m32[0]), "0" (dest.m32[0])); \ |
| 154 | }) |
| 155 | #define fp_sub96c(dest, srch, srcm, srcl) ({ \ |
| 156 | char carry; \ |
| 157 | asm ("sub.l %1,%0" : "=d,dm" (dest.m32[2]) \ |
| 158 | : "dm,d" (srcl), "0,0" (dest.m32[2])); \ |
| 159 | asm ("subx.l %1,%0" : "=d" (dest.m32[1]) \ |
| 160 | : "d" (srcm), "0" (dest.m32[1])); \ |
| 161 | asm ("subx.l %2,%1; scs %0" : "=d" (carry), "=d" (dest.m32[0]) \ |
| 162 | : "d" (srch), "1" (dest.m32[0])); \ |
| 163 | carry; \ |
| 164 | }) |
| 165 | |
| 166 | static inline void fp_multiplymant(union fp_mant128 *dest, struct fp_ext *src1, |
| 167 | struct fp_ext *src2) |
| 168 | { |
| 169 | union fp_mant64 temp; |
| 170 | |
| 171 | fp_mul64(dest->m32[0], dest->m32[1], src1->mant.m32[0], src2->mant.m32[0]); |
| 172 | fp_mul64(dest->m32[2], dest->m32[3], src1->mant.m32[1], src2->mant.m32[1]); |
| 173 | |
| 174 | fp_mul64(temp.m32[0], temp.m32[1], src1->mant.m32[0], src2->mant.m32[1]); |
| 175 | fp_addx96(dest, temp); |
| 176 | |
| 177 | fp_mul64(temp.m32[0], temp.m32[1], src1->mant.m32[1], src2->mant.m32[0]); |
| 178 | fp_addx96(dest, temp); |
| 179 | } |
| 180 | |
| 181 | static inline void fp_dividemant(union fp_mant128 *dest, struct fp_ext *src, |
| 182 | struct fp_ext *div) |
| 183 | { |
| 184 | union fp_mant128 tmp; |
| 185 | union fp_mant64 tmp64; |
| 186 | unsigned long *mantp = dest->m32; |
| 187 | unsigned long fix, rem, first, dummy; |
| 188 | int i; |
| 189 | |
| 190 | /* the algorithm below requires dest to be smaller than div, |
| 191 | but both have the high bit set */ |
| 192 | if (src->mant.m64 >= div->mant.m64) { |
| 193 | fp_sub64(src->mant, div->mant); |
| 194 | *mantp = 1; |
| 195 | } else |
| 196 | *mantp = 0; |
| 197 | mantp++; |
| 198 | |
| 199 | /* basic idea behind this algorithm: we can't divide two 64bit numbers |
| 200 | (AB/CD) directly, but we can calculate AB/C0, but this means this |
| 201 | quotient is off by C0/CD, so we have to multiply the first result |
| 202 | to fix the result, after that we have nearly the correct result |
| 203 | and only a few corrections are needed. */ |
| 204 | |
| 205 | /* C0/CD can be precalculated, but it's an 64bit division again, but |
| 206 | we can make it a bit easier, by dividing first through C so we get |
| 207 | 10/1D and now only a single shift and the value fits into 32bit. */ |
| 208 | fix = 0x80000000; |
| 209 | dummy = div->mant.m32[1] / div->mant.m32[0] + 1; |
| 210 | dummy = (dummy >> 1) | fix; |
| 211 | fp_div64(fix, dummy, fix, 0, dummy); |
| 212 | fix--; |
| 213 | |
| 214 | for (i = 0; i < 3; i++, mantp++) { |
| 215 | if (src->mant.m32[0] == div->mant.m32[0]) { |
| 216 | fp_div64(first, rem, 0, src->mant.m32[1], div->mant.m32[0]); |
| 217 | |
| 218 | fp_mul64(*mantp, dummy, first, fix); |
| 219 | *mantp += fix; |
| 220 | } else { |
| 221 | fp_div64(first, rem, src->mant.m32[0], src->mant.m32[1], div->mant.m32[0]); |
| 222 | |
| 223 | fp_mul64(*mantp, dummy, first, fix); |
| 224 | } |
| 225 | |
| 226 | fp_mul64(tmp.m32[0], tmp.m32[1], div->mant.m32[0], first - *mantp); |
| 227 | fp_add64(tmp.m32[0], tmp.m32[1], 0, rem); |
| 228 | tmp.m32[2] = 0; |
| 229 | |
| 230 | fp_mul64(tmp64.m32[0], tmp64.m32[1], *mantp, div->mant.m32[1]); |
| 231 | fp_sub96c(tmp, 0, tmp64.m32[0], tmp64.m32[1]); |
| 232 | |
| 233 | src->mant.m32[0] = tmp.m32[1]; |
| 234 | src->mant.m32[1] = tmp.m32[2]; |
| 235 | |
| 236 | while (!fp_sub96c(tmp, 0, div->mant.m32[0], div->mant.m32[1])) { |
| 237 | src->mant.m32[0] = tmp.m32[1]; |
| 238 | src->mant.m32[1] = tmp.m32[2]; |
| 239 | *mantp += 1; |
| 240 | } |
| 241 | } |
| 242 | } |
| 243 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 244 | static inline void fp_putmant128(struct fp_ext *dest, union fp_mant128 *src, |
| 245 | int shift) |
| 246 | { |
| 247 | unsigned long tmp; |
| 248 | |
| 249 | switch (shift) { |
| 250 | case 0: |
| 251 | dest->mant.m64 = src->m64[0]; |
| 252 | dest->lowmant = src->m32[2] >> 24; |
| 253 | if (src->m32[3] || (src->m32[2] << 8)) |
| 254 | dest->lowmant |= 1; |
| 255 | break; |
| 256 | case 1: |
| 257 | asm volatile ("lsl.l #1,%0" |
| 258 | : "=d" (tmp) : "0" (src->m32[2])); |
| 259 | asm volatile ("roxl.l #1,%0" |
| 260 | : "=d" (dest->mant.m32[1]) : "0" (src->m32[1])); |
| 261 | asm volatile ("roxl.l #1,%0" |
| 262 | : "=d" (dest->mant.m32[0]) : "0" (src->m32[0])); |
| 263 | dest->lowmant = tmp >> 24; |
| 264 | if (src->m32[3] || (tmp << 8)) |
| 265 | dest->lowmant |= 1; |
| 266 | break; |
| 267 | case 31: |
| 268 | asm volatile ("lsr.l #1,%1; roxr.l #1,%0" |
| 269 | : "=d" (dest->mant.m32[0]) |
| 270 | : "d" (src->m32[0]), "0" (src->m32[1])); |
| 271 | asm volatile ("roxr.l #1,%0" |
| 272 | : "=d" (dest->mant.m32[1]) : "0" (src->m32[2])); |
| 273 | asm volatile ("roxr.l #1,%0" |
| 274 | : "=d" (tmp) : "0" (src->m32[3])); |
| 275 | dest->lowmant = tmp >> 24; |
| 276 | if (src->m32[3] << 7) |
| 277 | dest->lowmant |= 1; |
| 278 | break; |
| 279 | case 32: |
| 280 | dest->mant.m32[0] = src->m32[1]; |
| 281 | dest->mant.m32[1] = src->m32[2]; |
| 282 | dest->lowmant = src->m32[3] >> 24; |
| 283 | if (src->m32[3] << 8) |
| 284 | dest->lowmant |= 1; |
| 285 | break; |
| 286 | } |
| 287 | } |
| 288 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 289 | #endif /* MULTI_ARITH_H */ |