| 1 |
/* Copyright (C) 2007-2025 Free Software Foundation, Inc. |
| 2 |
|
| 3 |
This file is part of GCC. |
| 4 |
|
| 5 |
GCC is free software; you can redistribute it and/or modify it under |
| 6 |
the terms of the GNU General Public License as published by the Free |
| 7 |
Software Foundation; either version 3, or (at your option) any later |
| 8 |
version. |
| 9 |
|
| 10 |
GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
| 11 |
WARRANTY; without even the implied warranty of MERCHANTABILITY or |
| 12 |
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
| 13 |
for more details. |
| 14 |
|
| 15 |
Under Section 7 of GPL version 3, you are granted additional |
| 16 |
permissions described in the GCC Runtime Library Exception, version |
| 17 |
3.1, as published by the Free Software Foundation. |
| 18 |
|
| 19 |
You should have received a copy of the GNU General Public License and |
| 20 |
a copy of the GCC Runtime Library Exception along with this program; |
| 21 |
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see |
| 22 |
<http://www.gnu.org/licenses/>. */ |
| 23 |
|
| 24 |
#ifndef _SQRT_MACROS_H_ |
| 25 |
#define _SQRT_MACROS_H_ |
| 26 |
|
| 27 |
#define FENCE __fence |
| 28 |
|
| 29 |
#if DOUBLE_EXTENDED_ON |
| 30 |
|
| 31 |
extern BINARY80 SQRT80 (BINARY80); |
| 32 |
|
| 33 |
|
| 34 |
__BID_INLINE__ UINT64 |
| 35 |
short_sqrt128 (UINT128 A10) { |
| 36 |
BINARY80 lx, ly, l64; |
| 37 |
int_float f64; |
| 38 |
|
| 39 |
// 2^64 |
| 40 |
f64.i = 0x5f800000; |
| 41 |
l64 = (BINARY80) f64.d; |
| 42 |
lx = (BINARY80) A10.w[1] * l64 + (BINARY80) A10.w[0]; |
| 43 |
ly = SQRT80 (lx); |
| 44 |
return (UINT64) ly; |
| 45 |
} |
| 46 |
|
| 47 |
|
| 48 |
__BID_INLINE__ void |
| 49 |
long_sqrt128 (UINT128 * pCS, UINT256 C256) { |
| 50 |
UINT256 C4; |
| 51 |
UINT128 CS; |
| 52 |
UINT64 X; |
| 53 |
SINT64 SE; |
| 54 |
BINARY80 l64, lm64, l128, lxL, lx, ly, lS, lSH, lSL, lE, l3, l2, |
| 55 |
l1, l0, lp, lCl; |
| 56 |
int_float fx, f64, fm64; |
| 57 |
int *ple = (int *) &lx; |
| 58 |
|
| 59 |
// 2^64 |
| 60 |
f64.i = 0x5f800000; |
| 61 |
l64 = (BINARY80) f64.d; |
| 62 |
|
| 63 |
l128 = l64 * l64; |
| 64 |
lx = l3 = (BINARY80) C256.w[3] * l64 * l128; |
| 65 |
l2 = (BINARY80) C256.w[2] * l128; |
| 66 |
lx = FENCE (lx + l2); |
| 67 |
l1 = (BINARY80) C256.w[1] * l64; |
| 68 |
lx = FENCE (lx + l1); |
| 69 |
l0 = (BINARY80) C256.w[0]; |
| 70 |
lx = FENCE (lx + l0); |
| 71 |
// sqrt(C256) |
| 72 |
lS = SQRT80 (lx); |
| 73 |
|
| 74 |
// get coefficient |
| 75 |
// 2^(-64) |
| 76 |
fm64.i = 0x1f800000; |
| 77 |
lm64 = (BINARY80) fm64.d; |
| 78 |
CS.w[1] = (UINT64) (lS * lm64); |
| 79 |
CS.w[0] = (UINT64) (lS - (BINARY80) CS.w[1] * l64); |
| 80 |
|
| 81 |
/////////////////////////////////////// |
| 82 |
// CAUTION! |
| 83 |
// little endian code only |
| 84 |
// add solution for big endian |
| 85 |
////////////////////////////////////// |
| 86 |
lSH = lS; |
| 87 |
*((UINT64 *) & lSH) &= 0xffffffff00000000ull; |
| 88 |
|
| 89 |
// correction for C256 rounding |
| 90 |
lCl = FENCE (l3 - lx); |
| 91 |
lCl = FENCE (lCl + l2); |
| 92 |
lCl = FENCE (lCl + l1); |
| 93 |
lCl = FENCE (lCl + l0); |
| 94 |
|
| 95 |
lSL = lS - lSH; |
| 96 |
|
| 97 |
////////////////////////////////////////// |
| 98 |
// Watch for compiler re-ordering |
| 99 |
// |
| 100 |
///////////////////////////////////////// |
| 101 |
// C256-S^2 |
| 102 |
lxL = FENCE (lx - lSH * lSH); |
| 103 |
lp = lSH * lSL; |
| 104 |
lp += lp; |
| 105 |
lxL = FENCE (lxL - lp); |
| 106 |
lSL *= lSL; |
| 107 |
lxL = FENCE (lxL - lSL); |
| 108 |
lCl += lxL; |
| 109 |
|
| 110 |
// correction term |
| 111 |
lE = lCl / (lS + lS); |
| 112 |
|
| 113 |
// get low part of coefficient |
| 114 |
X = CS.w[0]; |
| 115 |
if (lCl >= 0) { |
| 116 |
SE = (SINT64) (lE); |
| 117 |
CS.w[0] += SE; |
| 118 |
if (CS.w[0] < X) |
| 119 |
CS.w[1]++; |
| 120 |
} else { |
| 121 |
SE = (SINT64) (-lE); |
| 122 |
CS.w[0] -= SE; |
| 123 |
if (CS.w[0] > X) |
| 124 |
CS.w[1]--; |
| 125 |
} |
| 126 |
|
| 127 |
pCS->w[0] = CS.w[0]; |
| 128 |
pCS->w[1] = CS.w[1]; |
| 129 |
} |
| 130 |
|
| 131 |
#else |
| 132 |
|
| 133 |
extern double sqrt (double); |
| 134 |
|
| 135 |
__BID_INLINE__ UINT64 |
| 136 |
short_sqrt128 (UINT128 A10) { |
| 137 |
UINT256 ARS, ARS0, AE0, AE, S; |
| 138 |
|
| 139 |
UINT64 MY, ES, CY; |
| 140 |
double lx, l64; |
| 141 |
int_double f64, ly; |
| 142 |
int ey, k; |
| 143 |
|
| 144 |
// 2^64 |
| 145 |
f64.i = 0x43f0000000000000ull; |
| 146 |
l64 = f64.d; |
| 147 |
lx = (double) A10.w[1] * l64 + (double) A10.w[0]; |
| 148 |
ly.d = 1.0 / sqrt (lx); |
| 149 |
|
| 150 |
MY = (ly.i & 0x000fffffffffffffull) | 0x0010000000000000ull; |
| 151 |
ey = 0x3ff - (ly.i >> 52); |
| 152 |
|
| 153 |
// A10*RS^2 |
| 154 |
__mul_64x128_to_192 (ARS0, MY, A10); |
| 155 |
__mul_64x192_to_256 (ARS, MY, ARS0); |
| 156 |
|
| 157 |
// shr by 2*ey+40, to get a 64-bit value |
| 158 |
k = (ey << 1) + 104 - 64; |
| 159 |
if (k >= 128) { |
| 160 |
if (k > 128) |
| 161 |
ES = (ARS.w[2] >> (k - 128)) | (ARS.w[3] << (192 - k)); |
| 162 |
else |
| 163 |
ES = ARS.w[2]; |
| 164 |
} else { |
| 165 |
if (k >= 64) { |
| 166 |
ARS.w[0] = ARS.w[1]; |
| 167 |
ARS.w[1] = ARS.w[2]; |
| 168 |
k -= 64; |
| 169 |
} |
| 170 |
if (k) { |
| 171 |
__shr_128 (ARS, ARS, k); |
| 172 |
} |
| 173 |
ES = ARS.w[0]; |
| 174 |
} |
| 175 |
|
| 176 |
ES = ((SINT64) ES) >> 1; |
| 177 |
|
| 178 |
if (((SINT64) ES) < 0) { |
| 179 |
ES = -ES; |
| 180 |
|
| 181 |
// A*RS*eps (scaled by 2^64) |
| 182 |
__mul_64x192_to_256 (AE0, ES, ARS0); |
| 183 |
|
| 184 |
AE.w[0] = AE0.w[1]; |
| 185 |
AE.w[1] = AE0.w[2]; |
| 186 |
AE.w[2] = AE0.w[3]; |
| 187 |
|
| 188 |
__add_carry_out (S.w[0], CY, ARS0.w[0], AE.w[0]); |
| 189 |
__add_carry_in_out (S.w[1], CY, ARS0.w[1], AE.w[1], CY); |
| 190 |
S.w[2] = ARS0.w[2] + AE.w[2] + CY; |
| 191 |
} else { |
| 192 |
// A*RS*eps (scaled by 2^64) |
| 193 |
__mul_64x192_to_256 (AE0, ES, ARS0); |
| 194 |
|
| 195 |
AE.w[0] = AE0.w[1]; |
| 196 |
AE.w[1] = AE0.w[2]; |
| 197 |
AE.w[2] = AE0.w[3]; |
| 198 |
|
| 199 |
__sub_borrow_out (S.w[0], CY, ARS0.w[0], AE.w[0]); |
| 200 |
__sub_borrow_in_out (S.w[1], CY, ARS0.w[1], AE.w[1], CY); |
| 201 |
S.w[2] = ARS0.w[2] - AE.w[2] - CY; |
| 202 |
} |
| 203 |
|
| 204 |
k = ey + 51; |
| 205 |
|
| 206 |
if (k >= 64) { |
| 207 |
if (k >= 128) { |
| 208 |
S.w[0] = S.w[2]; |
| 209 |
S.w[1] = 0; |
| 210 |
k -= 128; |
| 211 |
} else { |
| 212 |
S.w[0] = S.w[1]; |
| 213 |
S.w[1] = S.w[2]; |
| 214 |
} |
| 215 |
k -= 64; |
| 216 |
} |
| 217 |
if (k) { |
| 218 |
__shr_128 (S, S, k); |
| 219 |
} |
| 220 |
|
| 221 |
|
| 222 |
return (UINT64) ((S.w[0] + 1) >> 1); |
| 223 |
|
| 224 |
} |
| 225 |
|
| 226 |
|
| 227 |
|
| 228 |
__BID_INLINE__ void |
| 229 |
long_sqrt128 (UINT128 * pCS, UINT256 C256) { |
| 230 |
UINT512 ARS0, ARS; |
| 231 |
UINT256 ARS00, AE, AE2, S; |
| 232 |
UINT128 ES, ES2, ARS1; |
| 233 |
UINT64 ES32, CY, MY; |
| 234 |
double l64, l128, lx, l2, l1, l0; |
| 235 |
int_double f64, ly; |
| 236 |
int ey, k, k2; |
| 237 |
|
| 238 |
// 2^64 |
| 239 |
f64.i = 0x43f0000000000000ull; |
| 240 |
l64 = f64.d; |
| 241 |
|
| 242 |
l128 = l64 * l64; |
| 243 |
lx = (double) C256.w[3] * l64 * l128; |
| 244 |
l2 = (double) C256.w[2] * l128; |
| 245 |
lx = FENCE (lx + l2); |
| 246 |
l1 = (double) C256.w[1] * l64; |
| 247 |
lx = FENCE (lx + l1); |
| 248 |
l0 = (double) C256.w[0]; |
| 249 |
lx = FENCE (lx + l0); |
| 250 |
// sqrt(C256) |
| 251 |
ly.d = 1.0 / sqrt (lx); |
| 252 |
|
| 253 |
MY = (ly.i & 0x000fffffffffffffull) | 0x0010000000000000ull; |
| 254 |
ey = 0x3ff - (ly.i >> 52); |
| 255 |
|
| 256 |
// A10*RS^2, scaled by 2^(2*ey+104) |
| 257 |
__mul_64x256_to_320 (ARS0, MY, C256); |
| 258 |
__mul_64x320_to_384 (ARS, MY, ARS0); |
| 259 |
|
| 260 |
// shr by k=(2*ey+104)-128 |
| 261 |
// expect k is in the range (192, 256) if result in [10^33, 10^34) |
| 262 |
// apply an additional signed shift by 1 at the same time (to get eps=eps0/2) |
| 263 |
k = (ey << 1) + 104 - 128 - 192; |
| 264 |
k2 = 64 - k; |
| 265 |
ES.w[0] = (ARS.w[3] >> (k + 1)) | (ARS.w[4] << (k2 - 1)); |
| 266 |
ES.w[1] = (ARS.w[4] >> k) | (ARS.w[5] << k2); |
| 267 |
ES.w[1] = ((SINT64) ES.w[1]) >> 1; |
| 268 |
|
| 269 |
// A*RS >> 192 (for error term computation) |
| 270 |
ARS1.w[0] = ARS0.w[3]; |
| 271 |
ARS1.w[1] = ARS0.w[4]; |
| 272 |
|
| 273 |
// A*RS>>64 |
| 274 |
ARS00.w[0] = ARS0.w[1]; |
| 275 |
ARS00.w[1] = ARS0.w[2]; |
| 276 |
ARS00.w[2] = ARS0.w[3]; |
| 277 |
ARS00.w[3] = ARS0.w[4]; |
| 278 |
|
| 279 |
if (((SINT64) ES.w[1]) < 0) { |
| 280 |
ES.w[0] = -ES.w[0]; |
| 281 |
ES.w[1] = -ES.w[1]; |
| 282 |
if (ES.w[0]) |
| 283 |
ES.w[1]--; |
| 284 |
|
| 285 |
// A*RS*eps |
| 286 |
__mul_128x128_to_256 (AE, ES, ARS1); |
| 287 |
|
| 288 |
__add_carry_out (S.w[0], CY, ARS00.w[0], AE.w[0]); |
| 289 |
__add_carry_in_out (S.w[1], CY, ARS00.w[1], AE.w[1], CY); |
| 290 |
__add_carry_in_out (S.w[2], CY, ARS00.w[2], AE.w[2], CY); |
| 291 |
S.w[3] = ARS00.w[3] + AE.w[3] + CY; |
| 292 |
} else { |
| 293 |
// A*RS*eps |
| 294 |
__mul_128x128_to_256 (AE, ES, ARS1); |
| 295 |
|
| 296 |
__sub_borrow_out (S.w[0], CY, ARS00.w[0], AE.w[0]); |
| 297 |
__sub_borrow_in_out (S.w[1], CY, ARS00.w[1], AE.w[1], CY); |
| 298 |
__sub_borrow_in_out (S.w[2], CY, ARS00.w[2], AE.w[2], CY); |
| 299 |
S.w[3] = ARS00.w[3] - AE.w[3] - CY; |
| 300 |
} |
| 301 |
|
| 302 |
// 3/2*eps^2, scaled by 2^128 |
| 303 |
ES32 = ES.w[1] + (ES.w[1] >> 1); |
| 304 |
__mul_64x64_to_128 (ES2, ES32, ES.w[1]); |
| 305 |
// A*RS*3/2*eps^2 |
| 306 |
__mul_128x128_to_256 (AE2, ES2, ARS1); |
| 307 |
|
| 308 |
// result, scaled by 2^(ey+52-64) |
| 309 |
__add_carry_out (S.w[0], CY, S.w[0], AE2.w[0]); |
| 310 |
__add_carry_in_out (S.w[1], CY, S.w[1], AE2.w[1], CY); |
| 311 |
__add_carry_in_out (S.w[2], CY, S.w[2], AE2.w[2], CY); |
| 312 |
S.w[3] = S.w[3] + AE2.w[3] + CY; |
| 313 |
|
| 314 |
// k in (0, 64) |
| 315 |
k = ey + 51 - 128; |
| 316 |
k2 = 64 - k; |
| 317 |
S.w[0] = (S.w[1] >> k) | (S.w[2] << k2); |
| 318 |
S.w[1] = (S.w[2] >> k) | (S.w[3] << k2); |
| 319 |
|
| 320 |
// round to nearest |
| 321 |
S.w[0]++; |
| 322 |
if (!S.w[0]) |
| 323 |
S.w[1]++; |
| 324 |
|
| 325 |
pCS->w[0] = (S.w[1] << 63) | (S.w[0] >> 1); |
| 326 |
pCS->w[1] = S.w[1] >> 1; |
| 327 |
|
| 328 |
} |
| 329 |
|
| 330 |
#endif |
| 331 |
#endif |