1 /* 2 * QEMU float support 3 * 4 * The code in this source file is derived from release 2a of the SoftFloat 5 * IEC/IEEE Floating-point Arithmetic Package. Those parts of the code (and 6 * some later contributions) are provided under that license, as detailed below. 7 * It has subsequently been modified by contributors to the QEMU Project, 8 * so some portions are provided under: 9 * the SoftFloat-2a license 10 * the BSD license 11 * GPL-v2-or-later 12 * 13 * Any future contributions to this file after December 1st 2014 will be 14 * taken to be licensed under the Softfloat-2a license unless specifically 15 * indicated otherwise. 16 */ 17 18 /* 19 =============================================================================== 20 This C source file is part of the SoftFloat IEC/IEEE Floating-point 21 Arithmetic Package, Release 2a. 22 23 Written by John R. Hauser. This work was made possible in part by the 24 International Computer Science Institute, located at Suite 600, 1947 Center 25 Street, Berkeley, California 94704. Funding was partially provided by the 26 National Science Foundation under grant MIP-9311980. The original version 27 of this code was written as part of a project to build a fixed-point vector 28 processor in collaboration with the University of California at Berkeley, 29 overseen by Profs. Nelson Morgan and John Wawrzynek. More information 30 is available through the Web page `http://HTTP.CS.Berkeley.EDU/~jhauser/ 31 arithmetic/SoftFloat.html'. 32 33 THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE. Although reasonable effort 34 has been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT 35 TIMES RESULT IN INCORRECT BEHAVIOR. USE OF THIS SOFTWARE IS RESTRICTED TO 36 PERSONS AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ANY 37 AND ALL LOSSES, COSTS, OR OTHER PROBLEMS ARISING FROM ITS USE. 38 39 Derivative works are acceptable, even for commercial purposes, so long as 40 (1) they include prominent notice that the work is derivative, and (2) they 41 include prominent notice akin to these four paragraphs for those parts of 42 this code that are retained. 43 44 =============================================================================== 45 */ 46 47 /* BSD licensing: 48 * Copyright (c) 2006, Fabrice Bellard 49 * All rights reserved. 50 * 51 * Redistribution and use in source and binary forms, with or without 52 * modification, are permitted provided that the following conditions are met: 53 * 54 * 1. Redistributions of source code must retain the above copyright notice, 55 * this list of conditions and the following disclaimer. 56 * 57 * 2. Redistributions in binary form must reproduce the above copyright notice, 58 * this list of conditions and the following disclaimer in the documentation 59 * and/or other materials provided with the distribution. 60 * 61 * 3. Neither the name of the copyright holder nor the names of its contributors 62 * may be used to endorse or promote products derived from this software without 63 * specific prior written permission. 64 * 65 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" 66 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 67 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 68 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE 69 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 70 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 71 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 72 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 73 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 74 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF 75 * THE POSSIBILITY OF SUCH DAMAGE. 76 */ 77 78 /* Portions of this work are licensed under the terms of the GNU GPL, 79 * version 2 or later. See the COPYING file in the top-level directory. 80 */ 81 82 /* softfloat (and in particular the code in softfloat-specialize.h) is 83 * target-dependent and needs the TARGET_* macros. 84 */ 85 #include "qemu/osdep.h" 86 #include <math.h> 87 #include "qemu/bitops.h" 88 #include "fpu/softfloat.h" 89 90 /* We only need stdlib for abort() */ 91 92 /*---------------------------------------------------------------------------- 93 | Primitive arithmetic functions, including multi-word arithmetic, and 94 | division and square root approximations. (Can be specialized to target if 95 | desired.) 96 *----------------------------------------------------------------------------*/ 97 #include "fpu/softfloat-macros.h" 98 99 /* 100 * Hardfloat 101 * 102 * Fast emulation of guest FP instructions is challenging for two reasons. 103 * First, FP instruction semantics are similar but not identical, particularly 104 * when handling NaNs. Second, emulating at reasonable speed the guest FP 105 * exception flags is not trivial: reading the host's flags register with a 106 * feclearexcept & fetestexcept pair is slow [slightly slower than soft-fp], 107 * and trapping on every FP exception is not fast nor pleasant to work with. 108 * 109 * We address these challenges by leveraging the host FPU for a subset of the 110 * operations. To do this we expand on the idea presented in this paper: 111 * 112 * Guo, Yu-Chuan, et al. "Translating the ARM Neon and VFP instructions in a 113 * binary translator." Software: Practice and Experience 46.12 (2016):1591-1615. 114 * 115 * The idea is thus to leverage the host FPU to (1) compute FP operations 116 * and (2) identify whether FP exceptions occurred while avoiding 117 * expensive exception flag register accesses. 118 * 119 * An important optimization shown in the paper is that given that exception 120 * flags are rarely cleared by the guest, we can avoid recomputing some flags. 121 * This is particularly useful for the inexact flag, which is very frequently 122 * raised in floating-point workloads. 123 * 124 * We optimize the code further by deferring to soft-fp whenever FP exception 125 * detection might get hairy. Two examples: (1) when at least one operand is 126 * denormal/inf/NaN; (2) when operands are not guaranteed to lead to a 0 result 127 * and the result is < the minimum normal. 128 */ 129 #define GEN_INPUT_FLUSH__NOCHECK(name, soft_t) \ 130 static inline void name(soft_t *a, float_status *s) \ 131 { \ 132 if (unlikely(soft_t ## _is_denormal(*a))) { \ 133 *a = soft_t ## _set_sign(soft_t ## _zero, \ 134 soft_t ## _is_neg(*a)); \ 135 float_raise(float_flag_input_denormal, s); \ 136 } \ 137 } 138 139 GEN_INPUT_FLUSH__NOCHECK(float32_input_flush__nocheck, float32) 140 GEN_INPUT_FLUSH__NOCHECK(float64_input_flush__nocheck, float64) 141 #undef GEN_INPUT_FLUSH__NOCHECK 142 143 #define GEN_INPUT_FLUSH1(name, soft_t) \ 144 static inline void name(soft_t *a, float_status *s) \ 145 { \ 146 if (likely(!s->flush_inputs_to_zero)) { \ 147 return; \ 148 } \ 149 soft_t ## _input_flush__nocheck(a, s); \ 150 } 151 152 GEN_INPUT_FLUSH1(float32_input_flush1, float32) 153 GEN_INPUT_FLUSH1(float64_input_flush1, float64) 154 #undef GEN_INPUT_FLUSH1 155 156 #define GEN_INPUT_FLUSH2(name, soft_t) \ 157 static inline void name(soft_t *a, soft_t *b, float_status *s) \ 158 { \ 159 if (likely(!s->flush_inputs_to_zero)) { \ 160 return; \ 161 } \ 162 soft_t ## _input_flush__nocheck(a, s); \ 163 soft_t ## _input_flush__nocheck(b, s); \ 164 } 165 166 GEN_INPUT_FLUSH2(float32_input_flush2, float32) 167 GEN_INPUT_FLUSH2(float64_input_flush2, float64) 168 #undef GEN_INPUT_FLUSH2 169 170 #define GEN_INPUT_FLUSH3(name, soft_t) \ 171 static inline void name(soft_t *a, soft_t *b, soft_t *c, float_status *s) \ 172 { \ 173 if (likely(!s->flush_inputs_to_zero)) { \ 174 return; \ 175 } \ 176 soft_t ## _input_flush__nocheck(a, s); \ 177 soft_t ## _input_flush__nocheck(b, s); \ 178 soft_t ## _input_flush__nocheck(c, s); \ 179 } 180 181 GEN_INPUT_FLUSH3(float32_input_flush3, float32) 182 GEN_INPUT_FLUSH3(float64_input_flush3, float64) 183 #undef GEN_INPUT_FLUSH3 184 185 /* 186 * Choose whether to use fpclassify or float32/64_* primitives in the generated 187 * hardfloat functions. Each combination of number of inputs and float size 188 * gets its own value. 189 */ 190 #if defined(__x86_64__) 191 # define QEMU_HARDFLOAT_1F32_USE_FP 0 192 # define QEMU_HARDFLOAT_1F64_USE_FP 1 193 # define QEMU_HARDFLOAT_2F32_USE_FP 0 194 # define QEMU_HARDFLOAT_2F64_USE_FP 1 195 # define QEMU_HARDFLOAT_3F32_USE_FP 0 196 # define QEMU_HARDFLOAT_3F64_USE_FP 1 197 #else 198 # define QEMU_HARDFLOAT_1F32_USE_FP 0 199 # define QEMU_HARDFLOAT_1F64_USE_FP 0 200 # define QEMU_HARDFLOAT_2F32_USE_FP 0 201 # define QEMU_HARDFLOAT_2F64_USE_FP 0 202 # define QEMU_HARDFLOAT_3F32_USE_FP 0 203 # define QEMU_HARDFLOAT_3F64_USE_FP 0 204 #endif 205 206 /* 207 * QEMU_HARDFLOAT_USE_ISINF chooses whether to use isinf() over 208 * float{32,64}_is_infinity when !USE_FP. 209 * On x86_64/aarch64, using the former over the latter can yield a ~6% speedup. 210 * On power64 however, using isinf() reduces fp-bench performance by up to 50%. 211 */ 212 #if defined(__x86_64__) || defined(__aarch64__) 213 # define QEMU_HARDFLOAT_USE_ISINF 1 214 #else 215 # define QEMU_HARDFLOAT_USE_ISINF 0 216 #endif 217 218 /* 219 * Some targets clear the FP flags before most FP operations. This prevents 220 * the use of hardfloat, since hardfloat relies on the inexact flag being 221 * already set. 222 */ 223 #if defined(TARGET_PPC) || defined(__FAST_MATH__) 224 # if defined(__FAST_MATH__) 225 # warning disabling hardfloat due to -ffast-math: hardfloat requires an exact \ 226 IEEE implementation 227 # endif 228 # define QEMU_NO_HARDFLOAT 1 229 # define QEMU_SOFTFLOAT_ATTR QEMU_FLATTEN 230 #else 231 # define QEMU_NO_HARDFLOAT 0 232 # define QEMU_SOFTFLOAT_ATTR QEMU_FLATTEN __attribute__((noinline)) 233 #endif 234 235 static inline bool can_use_fpu(const float_status *s) 236 { 237 if (QEMU_NO_HARDFLOAT) { 238 return false; 239 } 240 return likely(s->float_exception_flags & float_flag_inexact && 241 s->float_rounding_mode == float_round_nearest_even); 242 } 243 244 /* 245 * Hardfloat generation functions. Each operation can have two flavors: 246 * either using softfloat primitives (e.g. float32_is_zero_or_normal) for 247 * most condition checks, or native ones (e.g. fpclassify). 248 * 249 * The flavor is chosen by the callers. Instead of using macros, we rely on the 250 * compiler to propagate constants and inline everything into the callers. 251 * 252 * We only generate functions for operations with two inputs, since only 253 * these are common enough to justify consolidating them into common code. 254 */ 255 256 typedef union { 257 float32 s; 258 float h; 259 } union_float32; 260 261 typedef union { 262 float64 s; 263 double h; 264 } union_float64; 265 266 typedef bool (*f32_check_fn)(union_float32 a, union_float32 b); 267 typedef bool (*f64_check_fn)(union_float64 a, union_float64 b); 268 269 typedef float32 (*soft_f32_op2_fn)(float32 a, float32 b, float_status *s); 270 typedef float64 (*soft_f64_op2_fn)(float64 a, float64 b, float_status *s); 271 typedef float (*hard_f32_op2_fn)(float a, float b); 272 typedef double (*hard_f64_op2_fn)(double a, double b); 273 274 /* 2-input is-zero-or-normal */ 275 static inline bool f32_is_zon2(union_float32 a, union_float32 b) 276 { 277 if (QEMU_HARDFLOAT_2F32_USE_FP) { 278 /* 279 * Not using a temp variable for consecutive fpclassify calls ends up 280 * generating faster code. 281 */ 282 return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) && 283 (fpclassify(b.h) == FP_NORMAL || fpclassify(b.h) == FP_ZERO); 284 } 285 return float32_is_zero_or_normal(a.s) && 286 float32_is_zero_or_normal(b.s); 287 } 288 289 static inline bool f64_is_zon2(union_float64 a, union_float64 b) 290 { 291 if (QEMU_HARDFLOAT_2F64_USE_FP) { 292 return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) && 293 (fpclassify(b.h) == FP_NORMAL || fpclassify(b.h) == FP_ZERO); 294 } 295 return float64_is_zero_or_normal(a.s) && 296 float64_is_zero_or_normal(b.s); 297 } 298 299 /* 3-input is-zero-or-normal */ 300 static inline 301 bool f32_is_zon3(union_float32 a, union_float32 b, union_float32 c) 302 { 303 if (QEMU_HARDFLOAT_3F32_USE_FP) { 304 return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) && 305 (fpclassify(b.h) == FP_NORMAL || fpclassify(b.h) == FP_ZERO) && 306 (fpclassify(c.h) == FP_NORMAL || fpclassify(c.h) == FP_ZERO); 307 } 308 return float32_is_zero_or_normal(a.s) && 309 float32_is_zero_or_normal(b.s) && 310 float32_is_zero_or_normal(c.s); 311 } 312 313 static inline 314 bool f64_is_zon3(union_float64 a, union_float64 b, union_float64 c) 315 { 316 if (QEMU_HARDFLOAT_3F64_USE_FP) { 317 return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) && 318 (fpclassify(b.h) == FP_NORMAL || fpclassify(b.h) == FP_ZERO) && 319 (fpclassify(c.h) == FP_NORMAL || fpclassify(c.h) == FP_ZERO); 320 } 321 return float64_is_zero_or_normal(a.s) && 322 float64_is_zero_or_normal(b.s) && 323 float64_is_zero_or_normal(c.s); 324 } 325 326 static inline bool f32_is_inf(union_float32 a) 327 { 328 if (QEMU_HARDFLOAT_USE_ISINF) { 329 return isinf(a.h); 330 } 331 return float32_is_infinity(a.s); 332 } 333 334 static inline bool f64_is_inf(union_float64 a) 335 { 336 if (QEMU_HARDFLOAT_USE_ISINF) { 337 return isinf(a.h); 338 } 339 return float64_is_infinity(a.s); 340 } 341 342 static inline float32 343 float32_gen2(float32 xa, float32 xb, float_status *s, 344 hard_f32_op2_fn hard, soft_f32_op2_fn soft, 345 f32_check_fn pre, f32_check_fn post) 346 { 347 union_float32 ua, ub, ur; 348 349 ua.s = xa; 350 ub.s = xb; 351 352 if (unlikely(!can_use_fpu(s))) { 353 goto soft; 354 } 355 356 float32_input_flush2(&ua.s, &ub.s, s); 357 if (unlikely(!pre(ua, ub))) { 358 goto soft; 359 } 360 361 ur.h = hard(ua.h, ub.h); 362 if (unlikely(f32_is_inf(ur))) { 363 float_raise(float_flag_overflow, s); 364 } else if (unlikely(fabsf(ur.h) <= FLT_MIN) && post(ua, ub)) { 365 goto soft; 366 } 367 return ur.s; 368 369 soft: 370 return soft(ua.s, ub.s, s); 371 } 372 373 static inline float64 374 float64_gen2(float64 xa, float64 xb, float_status *s, 375 hard_f64_op2_fn hard, soft_f64_op2_fn soft, 376 f64_check_fn pre, f64_check_fn post) 377 { 378 union_float64 ua, ub, ur; 379 380 ua.s = xa; 381 ub.s = xb; 382 383 if (unlikely(!can_use_fpu(s))) { 384 goto soft; 385 } 386 387 float64_input_flush2(&ua.s, &ub.s, s); 388 if (unlikely(!pre(ua, ub))) { 389 goto soft; 390 } 391 392 ur.h = hard(ua.h, ub.h); 393 if (unlikely(f64_is_inf(ur))) { 394 float_raise(float_flag_overflow, s); 395 } else if (unlikely(fabs(ur.h) <= DBL_MIN) && post(ua, ub)) { 396 goto soft; 397 } 398 return ur.s; 399 400 soft: 401 return soft(ua.s, ub.s, s); 402 } 403 404 /* 405 * Classify a floating point number. Everything above float_class_qnan 406 * is a NaN so cls >= float_class_qnan is any NaN. 407 */ 408 409 typedef enum __attribute__ ((__packed__)) { 410 float_class_unclassified, 411 float_class_zero, 412 float_class_normal, 413 float_class_inf, 414 float_class_qnan, /* all NaNs from here */ 415 float_class_snan, 416 } FloatClass; 417 418 #define float_cmask(bit) (1u << (bit)) 419 420 enum { 421 float_cmask_zero = float_cmask(float_class_zero), 422 float_cmask_normal = float_cmask(float_class_normal), 423 float_cmask_inf = float_cmask(float_class_inf), 424 float_cmask_qnan = float_cmask(float_class_qnan), 425 float_cmask_snan = float_cmask(float_class_snan), 426 427 float_cmask_infzero = float_cmask_zero | float_cmask_inf, 428 float_cmask_anynan = float_cmask_qnan | float_cmask_snan, 429 }; 430 431 /* Flags for parts_minmax. */ 432 enum { 433 /* Set for minimum; clear for maximum. */ 434 minmax_ismin = 1, 435 /* Set for the IEEE 754-2008 minNum() and maxNum() operations. */ 436 minmax_isnum = 2, 437 /* Set for the IEEE 754-2008 minNumMag() and minNumMag() operations. */ 438 minmax_ismag = 4, 439 /* 440 * Set for the IEEE 754-2019 minimumNumber() and maximumNumber() 441 * operations. 442 */ 443 minmax_isnumber = 8, 444 }; 445 446 /* Simple helpers for checking if, or what kind of, NaN we have */ 447 static inline __attribute__((unused)) bool is_nan(FloatClass c) 448 { 449 return unlikely(c >= float_class_qnan); 450 } 451 452 static inline __attribute__((unused)) bool is_snan(FloatClass c) 453 { 454 return c == float_class_snan; 455 } 456 457 static inline __attribute__((unused)) bool is_qnan(FloatClass c) 458 { 459 return c == float_class_qnan; 460 } 461 462 /* 463 * Structure holding all of the decomposed parts of a float. 464 * The exponent is unbiased and the fraction is normalized. 465 * 466 * The fraction words are stored in big-endian word ordering, 467 * so that truncation from a larger format to a smaller format 468 * can be done simply by ignoring subsequent elements. 469 */ 470 471 typedef struct { 472 FloatClass cls; 473 bool sign; 474 int32_t exp; 475 union { 476 /* Routines that know the structure may reference the singular name. */ 477 uint64_t frac; 478 /* 479 * Routines expanded with multiple structures reference "hi" and "lo" 480 * depending on the operation. In FloatParts64, "hi" and "lo" are 481 * both the same word and aliased here. 482 */ 483 uint64_t frac_hi; 484 uint64_t frac_lo; 485 }; 486 } FloatParts64; 487 488 typedef struct { 489 FloatClass cls; 490 bool sign; 491 int32_t exp; 492 uint64_t frac_hi; 493 uint64_t frac_lo; 494 } FloatParts128; 495 496 typedef struct { 497 FloatClass cls; 498 bool sign; 499 int32_t exp; 500 uint64_t frac_hi; 501 uint64_t frac_hm; /* high-middle */ 502 uint64_t frac_lm; /* low-middle */ 503 uint64_t frac_lo; 504 } FloatParts256; 505 506 /* These apply to the most significant word of each FloatPartsN. */ 507 #define DECOMPOSED_BINARY_POINT 63 508 #define DECOMPOSED_IMPLICIT_BIT (1ull << DECOMPOSED_BINARY_POINT) 509 510 /* Structure holding all of the relevant parameters for a format. 511 * exp_size: the size of the exponent field 512 * exp_bias: the offset applied to the exponent field 513 * exp_max: the maximum normalised exponent 514 * frac_size: the size of the fraction field 515 * frac_shift: shift to normalise the fraction with DECOMPOSED_BINARY_POINT 516 * The following are computed based the size of fraction 517 * round_mask: bits below lsb which must be rounded 518 * The following optional modifiers are available: 519 * arm_althp: handle ARM Alternative Half Precision 520 */ 521 typedef struct { 522 int exp_size; 523 int exp_bias; 524 int exp_max; 525 int frac_size; 526 int frac_shift; 527 bool arm_althp; 528 uint64_t round_mask; 529 } FloatFmt; 530 531 /* Expand fields based on the size of exponent and fraction */ 532 #define FLOAT_PARAMS_(E) \ 533 .exp_size = E, \ 534 .exp_bias = ((1 << E) - 1) >> 1, \ 535 .exp_max = (1 << E) - 1 536 537 #define FLOAT_PARAMS(E, F) \ 538 FLOAT_PARAMS_(E), \ 539 .frac_size = F, \ 540 .frac_shift = (-F - 1) & 63, \ 541 .round_mask = (1ull << ((-F - 1) & 63)) - 1 542 543 static const FloatFmt float16_params = { 544 FLOAT_PARAMS(5, 10) 545 }; 546 547 static const FloatFmt float16_params_ahp = { 548 FLOAT_PARAMS(5, 10), 549 .arm_althp = true 550 }; 551 552 static const FloatFmt bfloat16_params = { 553 FLOAT_PARAMS(8, 7) 554 }; 555 556 static const FloatFmt float32_params = { 557 FLOAT_PARAMS(8, 23) 558 }; 559 560 static const FloatFmt float64_params = { 561 FLOAT_PARAMS(11, 52) 562 }; 563 564 static const FloatFmt float128_params = { 565 FLOAT_PARAMS(15, 112) 566 }; 567 568 #define FLOATX80_PARAMS(R) \ 569 FLOAT_PARAMS_(15), \ 570 .frac_size = R == 64 ? 63 : R, \ 571 .frac_shift = 0, \ 572 .round_mask = R == 64 ? -1 : (1ull << ((-R - 1) & 63)) - 1 573 574 static const FloatFmt floatx80_params[3] = { 575 [floatx80_precision_s] = { FLOATX80_PARAMS(23) }, 576 [floatx80_precision_d] = { FLOATX80_PARAMS(52) }, 577 [floatx80_precision_x] = { FLOATX80_PARAMS(64) }, 578 }; 579 580 /* Unpack a float to parts, but do not canonicalize. */ 581 static void unpack_raw64(FloatParts64 *r, const FloatFmt *fmt, uint64_t raw) 582 { 583 const int f_size = fmt->frac_size; 584 const int e_size = fmt->exp_size; 585 586 *r = (FloatParts64) { 587 .cls = float_class_unclassified, 588 .sign = extract64(raw, f_size + e_size, 1), 589 .exp = extract64(raw, f_size, e_size), 590 .frac = extract64(raw, 0, f_size) 591 }; 592 } 593 594 static inline void float16_unpack_raw(FloatParts64 *p, float16 f) 595 { 596 unpack_raw64(p, &float16_params, f); 597 } 598 599 static inline void bfloat16_unpack_raw(FloatParts64 *p, bfloat16 f) 600 { 601 unpack_raw64(p, &bfloat16_params, f); 602 } 603 604 static inline void float32_unpack_raw(FloatParts64 *p, float32 f) 605 { 606 unpack_raw64(p, &float32_params, f); 607 } 608 609 static inline void float64_unpack_raw(FloatParts64 *p, float64 f) 610 { 611 unpack_raw64(p, &float64_params, f); 612 } 613 614 static void floatx80_unpack_raw(FloatParts128 *p, floatx80 f) 615 { 616 *p = (FloatParts128) { 617 .cls = float_class_unclassified, 618 .sign = extract32(f.high, 15, 1), 619 .exp = extract32(f.high, 0, 15), 620 .frac_hi = f.low 621 }; 622 } 623 624 static void float128_unpack_raw(FloatParts128 *p, float128 f) 625 { 626 const int f_size = float128_params.frac_size - 64; 627 const int e_size = float128_params.exp_size; 628 629 *p = (FloatParts128) { 630 .cls = float_class_unclassified, 631 .sign = extract64(f.high, f_size + e_size, 1), 632 .exp = extract64(f.high, f_size, e_size), 633 .frac_hi = extract64(f.high, 0, f_size), 634 .frac_lo = f.low, 635 }; 636 } 637 638 /* Pack a float from parts, but do not canonicalize. */ 639 static uint64_t pack_raw64(const FloatParts64 *p, const FloatFmt *fmt) 640 { 641 const int f_size = fmt->frac_size; 642 const int e_size = fmt->exp_size; 643 uint64_t ret; 644 645 ret = (uint64_t)p->sign << (f_size + e_size); 646 ret = deposit64(ret, f_size, e_size, p->exp); 647 ret = deposit64(ret, 0, f_size, p->frac); 648 return ret; 649 } 650 651 static inline float16 float16_pack_raw(const FloatParts64 *p) 652 { 653 return make_float16(pack_raw64(p, &float16_params)); 654 } 655 656 static inline bfloat16 bfloat16_pack_raw(const FloatParts64 *p) 657 { 658 return pack_raw64(p, &bfloat16_params); 659 } 660 661 static inline float32 float32_pack_raw(const FloatParts64 *p) 662 { 663 return make_float32(pack_raw64(p, &float32_params)); 664 } 665 666 static inline float64 float64_pack_raw(const FloatParts64 *p) 667 { 668 return make_float64(pack_raw64(p, &float64_params)); 669 } 670 671 static float128 float128_pack_raw(const FloatParts128 *p) 672 { 673 const int f_size = float128_params.frac_size - 64; 674 const int e_size = float128_params.exp_size; 675 uint64_t hi; 676 677 hi = (uint64_t)p->sign << (f_size + e_size); 678 hi = deposit64(hi, f_size, e_size, p->exp); 679 hi = deposit64(hi, 0, f_size, p->frac_hi); 680 return make_float128(hi, p->frac_lo); 681 } 682 683 /*---------------------------------------------------------------------------- 684 | Functions and definitions to determine: (1) whether tininess for underflow 685 | is detected before or after rounding by default, (2) what (if anything) 686 | happens when exceptions are raised, (3) how signaling NaNs are distinguished 687 | from quiet NaNs, (4) the default generated quiet NaNs, and (5) how NaNs 688 | are propagated from function inputs to output. These details are target- 689 | specific. 690 *----------------------------------------------------------------------------*/ 691 #include "softfloat-specialize.c.inc" 692 693 #define PARTS_GENERIC_64_128(NAME, P) \ 694 _Generic((P), FloatParts64 *: parts64_##NAME, \ 695 FloatParts128 *: parts128_##NAME) 696 697 #define PARTS_GENERIC_64_128_256(NAME, P) \ 698 _Generic((P), FloatParts64 *: parts64_##NAME, \ 699 FloatParts128 *: parts128_##NAME, \ 700 FloatParts256 *: parts256_##NAME) 701 702 #define parts_default_nan(P, S) PARTS_GENERIC_64_128(default_nan, P)(P, S) 703 #define parts_silence_nan(P, S) PARTS_GENERIC_64_128(silence_nan, P)(P, S) 704 705 static void parts64_return_nan(FloatParts64 *a, float_status *s); 706 static void parts128_return_nan(FloatParts128 *a, float_status *s); 707 708 #define parts_return_nan(P, S) PARTS_GENERIC_64_128(return_nan, P)(P, S) 709 710 static FloatParts64 *parts64_pick_nan(FloatParts64 *a, FloatParts64 *b, 711 float_status *s); 712 static FloatParts128 *parts128_pick_nan(FloatParts128 *a, FloatParts128 *b, 713 float_status *s); 714 715 #define parts_pick_nan(A, B, S) PARTS_GENERIC_64_128(pick_nan, A)(A, B, S) 716 717 static FloatParts64 *parts64_pick_nan_muladd(FloatParts64 *a, FloatParts64 *b, 718 FloatParts64 *c, float_status *s, 719 int ab_mask, int abc_mask); 720 static FloatParts128 *parts128_pick_nan_muladd(FloatParts128 *a, 721 FloatParts128 *b, 722 FloatParts128 *c, 723 float_status *s, 724 int ab_mask, int abc_mask); 725 726 #define parts_pick_nan_muladd(A, B, C, S, ABM, ABCM) \ 727 PARTS_GENERIC_64_128(pick_nan_muladd, A)(A, B, C, S, ABM, ABCM) 728 729 static void parts64_canonicalize(FloatParts64 *p, float_status *status, 730 const FloatFmt *fmt); 731 static void parts128_canonicalize(FloatParts128 *p, float_status *status, 732 const FloatFmt *fmt); 733 734 #define parts_canonicalize(A, S, F) \ 735 PARTS_GENERIC_64_128(canonicalize, A)(A, S, F) 736 737 static void parts64_uncanon_normal(FloatParts64 *p, float_status *status, 738 const FloatFmt *fmt); 739 static void parts128_uncanon_normal(FloatParts128 *p, float_status *status, 740 const FloatFmt *fmt); 741 742 #define parts_uncanon_normal(A, S, F) \ 743 PARTS_GENERIC_64_128(uncanon_normal, A)(A, S, F) 744 745 static void parts64_uncanon(FloatParts64 *p, float_status *status, 746 const FloatFmt *fmt); 747 static void parts128_uncanon(FloatParts128 *p, float_status *status, 748 const FloatFmt *fmt); 749 750 #define parts_uncanon(A, S, F) \ 751 PARTS_GENERIC_64_128(uncanon, A)(A, S, F) 752 753 static void parts64_add_normal(FloatParts64 *a, FloatParts64 *b); 754 static void parts128_add_normal(FloatParts128 *a, FloatParts128 *b); 755 static void parts256_add_normal(FloatParts256 *a, FloatParts256 *b); 756 757 #define parts_add_normal(A, B) \ 758 PARTS_GENERIC_64_128_256(add_normal, A)(A, B) 759 760 static bool parts64_sub_normal(FloatParts64 *a, FloatParts64 *b); 761 static bool parts128_sub_normal(FloatParts128 *a, FloatParts128 *b); 762 static bool parts256_sub_normal(FloatParts256 *a, FloatParts256 *b); 763 764 #define parts_sub_normal(A, B) \ 765 PARTS_GENERIC_64_128_256(sub_normal, A)(A, B) 766 767 static FloatParts64 *parts64_addsub(FloatParts64 *a, FloatParts64 *b, 768 float_status *s, bool subtract); 769 static FloatParts128 *parts128_addsub(FloatParts128 *a, FloatParts128 *b, 770 float_status *s, bool subtract); 771 772 #define parts_addsub(A, B, S, Z) \ 773 PARTS_GENERIC_64_128(addsub, A)(A, B, S, Z) 774 775 static FloatParts64 *parts64_mul(FloatParts64 *a, FloatParts64 *b, 776 float_status *s); 777 static FloatParts128 *parts128_mul(FloatParts128 *a, FloatParts128 *b, 778 float_status *s); 779 780 #define parts_mul(A, B, S) \ 781 PARTS_GENERIC_64_128(mul, A)(A, B, S) 782 783 static FloatParts64 *parts64_muladd(FloatParts64 *a, FloatParts64 *b, 784 FloatParts64 *c, int flags, 785 float_status *s); 786 static FloatParts128 *parts128_muladd(FloatParts128 *a, FloatParts128 *b, 787 FloatParts128 *c, int flags, 788 float_status *s); 789 790 #define parts_muladd(A, B, C, Z, S) \ 791 PARTS_GENERIC_64_128(muladd, A)(A, B, C, Z, S) 792 793 static FloatParts64 *parts64_div(FloatParts64 *a, FloatParts64 *b, 794 float_status *s); 795 static FloatParts128 *parts128_div(FloatParts128 *a, FloatParts128 *b, 796 float_status *s); 797 798 #define parts_div(A, B, S) \ 799 PARTS_GENERIC_64_128(div, A)(A, B, S) 800 801 static FloatParts64 *parts64_modrem(FloatParts64 *a, FloatParts64 *b, 802 uint64_t *mod_quot, float_status *s); 803 static FloatParts128 *parts128_modrem(FloatParts128 *a, FloatParts128 *b, 804 uint64_t *mod_quot, float_status *s); 805 806 #define parts_modrem(A, B, Q, S) \ 807 PARTS_GENERIC_64_128(modrem, A)(A, B, Q, S) 808 809 static void parts64_sqrt(FloatParts64 *a, float_status *s, const FloatFmt *f); 810 static void parts128_sqrt(FloatParts128 *a, float_status *s, const FloatFmt *f); 811 812 #define parts_sqrt(A, S, F) \ 813 PARTS_GENERIC_64_128(sqrt, A)(A, S, F) 814 815 static bool parts64_round_to_int_normal(FloatParts64 *a, FloatRoundMode rm, 816 int scale, int frac_size); 817 static bool parts128_round_to_int_normal(FloatParts128 *a, FloatRoundMode r, 818 int scale, int frac_size); 819 820 #define parts_round_to_int_normal(A, R, C, F) \ 821 PARTS_GENERIC_64_128(round_to_int_normal, A)(A, R, C, F) 822 823 static void parts64_round_to_int(FloatParts64 *a, FloatRoundMode rm, 824 int scale, float_status *s, 825 const FloatFmt *fmt); 826 static void parts128_round_to_int(FloatParts128 *a, FloatRoundMode r, 827 int scale, float_status *s, 828 const FloatFmt *fmt); 829 830 #define parts_round_to_int(A, R, C, S, F) \ 831 PARTS_GENERIC_64_128(round_to_int, A)(A, R, C, S, F) 832 833 static int64_t parts64_float_to_sint(FloatParts64 *p, FloatRoundMode rmode, 834 int scale, int64_t min, int64_t max, 835 float_status *s); 836 static int64_t parts128_float_to_sint(FloatParts128 *p, FloatRoundMode rmode, 837 int scale, int64_t min, int64_t max, 838 float_status *s); 839 840 #define parts_float_to_sint(P, R, Z, MN, MX, S) \ 841 PARTS_GENERIC_64_128(float_to_sint, P)(P, R, Z, MN, MX, S) 842 843 static uint64_t parts64_float_to_uint(FloatParts64 *p, FloatRoundMode rmode, 844 int scale, uint64_t max, 845 float_status *s); 846 static uint64_t parts128_float_to_uint(FloatParts128 *p, FloatRoundMode rmode, 847 int scale, uint64_t max, 848 float_status *s); 849 850 #define parts_float_to_uint(P, R, Z, M, S) \ 851 PARTS_GENERIC_64_128(float_to_uint, P)(P, R, Z, M, S) 852 853 static void parts64_sint_to_float(FloatParts64 *p, int64_t a, 854 int scale, float_status *s); 855 static void parts128_sint_to_float(FloatParts128 *p, int64_t a, 856 int scale, float_status *s); 857 858 #define parts_sint_to_float(P, I, Z, S) \ 859 PARTS_GENERIC_64_128(sint_to_float, P)(P, I, Z, S) 860 861 static void parts64_uint_to_float(FloatParts64 *p, uint64_t a, 862 int scale, float_status *s); 863 static void parts128_uint_to_float(FloatParts128 *p, uint64_t a, 864 int scale, float_status *s); 865 866 #define parts_uint_to_float(P, I, Z, S) \ 867 PARTS_GENERIC_64_128(uint_to_float, P)(P, I, Z, S) 868 869 static FloatParts64 *parts64_minmax(FloatParts64 *a, FloatParts64 *b, 870 float_status *s, int flags); 871 static FloatParts128 *parts128_minmax(FloatParts128 *a, FloatParts128 *b, 872 float_status *s, int flags); 873 874 #define parts_minmax(A, B, S, F) \ 875 PARTS_GENERIC_64_128(minmax, A)(A, B, S, F) 876 877 static int parts64_compare(FloatParts64 *a, FloatParts64 *b, 878 float_status *s, bool q); 879 static int parts128_compare(FloatParts128 *a, FloatParts128 *b, 880 float_status *s, bool q); 881 882 #define parts_compare(A, B, S, Q) \ 883 PARTS_GENERIC_64_128(compare, A)(A, B, S, Q) 884 885 static void parts64_scalbn(FloatParts64 *a, int n, float_status *s); 886 static void parts128_scalbn(FloatParts128 *a, int n, float_status *s); 887 888 #define parts_scalbn(A, N, S) \ 889 PARTS_GENERIC_64_128(scalbn, A)(A, N, S) 890 891 static void parts64_log2(FloatParts64 *a, float_status *s, const FloatFmt *f); 892 static void parts128_log2(FloatParts128 *a, float_status *s, const FloatFmt *f); 893 894 #define parts_log2(A, S, F) \ 895 PARTS_GENERIC_64_128(log2, A)(A, S, F) 896 897 /* 898 * Helper functions for softfloat-parts.c.inc, per-size operations. 899 */ 900 901 #define FRAC_GENERIC_64_128(NAME, P) \ 902 _Generic((P), FloatParts64 *: frac64_##NAME, \ 903 FloatParts128 *: frac128_##NAME) 904 905 #define FRAC_GENERIC_64_128_256(NAME, P) \ 906 _Generic((P), FloatParts64 *: frac64_##NAME, \ 907 FloatParts128 *: frac128_##NAME, \ 908 FloatParts256 *: frac256_##NAME) 909 910 static bool frac64_add(FloatParts64 *r, FloatParts64 *a, FloatParts64 *b) 911 { 912 return uadd64_overflow(a->frac, b->frac, &r->frac); 913 } 914 915 static bool frac128_add(FloatParts128 *r, FloatParts128 *a, FloatParts128 *b) 916 { 917 bool c = 0; 918 r->frac_lo = uadd64_carry(a->frac_lo, b->frac_lo, &c); 919 r->frac_hi = uadd64_carry(a->frac_hi, b->frac_hi, &c); 920 return c; 921 } 922 923 static bool frac256_add(FloatParts256 *r, FloatParts256 *a, FloatParts256 *b) 924 { 925 bool c = 0; 926 r->frac_lo = uadd64_carry(a->frac_lo, b->frac_lo, &c); 927 r->frac_lm = uadd64_carry(a->frac_lm, b->frac_lm, &c); 928 r->frac_hm = uadd64_carry(a->frac_hm, b->frac_hm, &c); 929 r->frac_hi = uadd64_carry(a->frac_hi, b->frac_hi, &c); 930 return c; 931 } 932 933 #define frac_add(R, A, B) FRAC_GENERIC_64_128_256(add, R)(R, A, B) 934 935 static bool frac64_addi(FloatParts64 *r, FloatParts64 *a, uint64_t c) 936 { 937 return uadd64_overflow(a->frac, c, &r->frac); 938 } 939 940 static bool frac128_addi(FloatParts128 *r, FloatParts128 *a, uint64_t c) 941 { 942 c = uadd64_overflow(a->frac_lo, c, &r->frac_lo); 943 return uadd64_overflow(a->frac_hi, c, &r->frac_hi); 944 } 945 946 #define frac_addi(R, A, C) FRAC_GENERIC_64_128(addi, R)(R, A, C) 947 948 static void frac64_allones(FloatParts64 *a) 949 { 950 a->frac = -1; 951 } 952 953 static void frac128_allones(FloatParts128 *a) 954 { 955 a->frac_hi = a->frac_lo = -1; 956 } 957 958 #define frac_allones(A) FRAC_GENERIC_64_128(allones, A)(A) 959 960 static int frac64_cmp(FloatParts64 *a, FloatParts64 *b) 961 { 962 return a->frac == b->frac ? 0 : a->frac < b->frac ? -1 : 1; 963 } 964 965 static int frac128_cmp(FloatParts128 *a, FloatParts128 *b) 966 { 967 uint64_t ta = a->frac_hi, tb = b->frac_hi; 968 if (ta == tb) { 969 ta = a->frac_lo, tb = b->frac_lo; 970 if (ta == tb) { 971 return 0; 972 } 973 } 974 return ta < tb ? -1 : 1; 975 } 976 977 #define frac_cmp(A, B) FRAC_GENERIC_64_128(cmp, A)(A, B) 978 979 static void frac64_clear(FloatParts64 *a) 980 { 981 a->frac = 0; 982 } 983 984 static void frac128_clear(FloatParts128 *a) 985 { 986 a->frac_hi = a->frac_lo = 0; 987 } 988 989 #define frac_clear(A) FRAC_GENERIC_64_128(clear, A)(A) 990 991 static bool frac64_div(FloatParts64 *a, FloatParts64 *b) 992 { 993 uint64_t n1, n0, r, q; 994 bool ret; 995 996 /* 997 * We want a 2*N / N-bit division to produce exactly an N-bit 998 * result, so that we do not lose any precision and so that we 999 * do not have to renormalize afterward. If A.frac < B.frac, 1000 * then division would produce an (N-1)-bit result; shift A left 1001 * by one to produce the an N-bit result, and return true to 1002 * decrement the exponent to match. 1003 * 1004 * The udiv_qrnnd algorithm that we're using requires normalization, 1005 * i.e. the msb of the denominator must be set, which is already true. 1006 */ 1007 ret = a->frac < b->frac; 1008 if (ret) { 1009 n0 = a->frac; 1010 n1 = 0; 1011 } else { 1012 n0 = a->frac >> 1; 1013 n1 = a->frac << 63; 1014 } 1015 q = udiv_qrnnd(&r, n0, n1, b->frac); 1016 1017 /* Set lsb if there is a remainder, to set inexact. */ 1018 a->frac = q | (r != 0); 1019 1020 return ret; 1021 } 1022 1023 static bool frac128_div(FloatParts128 *a, FloatParts128 *b) 1024 { 1025 uint64_t q0, q1, a0, a1, b0, b1; 1026 uint64_t r0, r1, r2, r3, t0, t1, t2, t3; 1027 bool ret = false; 1028 1029 a0 = a->frac_hi, a1 = a->frac_lo; 1030 b0 = b->frac_hi, b1 = b->frac_lo; 1031 1032 ret = lt128(a0, a1, b0, b1); 1033 if (!ret) { 1034 a1 = shr_double(a0, a1, 1); 1035 a0 = a0 >> 1; 1036 } 1037 1038 /* Use 128/64 -> 64 division as estimate for 192/128 -> 128 division. */ 1039 q0 = estimateDiv128To64(a0, a1, b0); 1040 1041 /* 1042 * Estimate is high because B1 was not included (unless B1 == 0). 1043 * Reduce quotient and increase remainder until remainder is non-negative. 1044 * This loop will execute 0 to 2 times. 1045 */ 1046 mul128By64To192(b0, b1, q0, &t0, &t1, &t2); 1047 sub192(a0, a1, 0, t0, t1, t2, &r0, &r1, &r2); 1048 while (r0 != 0) { 1049 q0--; 1050 add192(r0, r1, r2, 0, b0, b1, &r0, &r1, &r2); 1051 } 1052 1053 /* Repeat using the remainder, producing a second word of quotient. */ 1054 q1 = estimateDiv128To64(r1, r2, b0); 1055 mul128By64To192(b0, b1, q1, &t1, &t2, &t3); 1056 sub192(r1, r2, 0, t1, t2, t3, &r1, &r2, &r3); 1057 while (r1 != 0) { 1058 q1--; 1059 add192(r1, r2, r3, 0, b0, b1, &r1, &r2, &r3); 1060 } 1061 1062 /* Any remainder indicates inexact; set sticky bit. */ 1063 q1 |= (r2 | r3) != 0; 1064 1065 a->frac_hi = q0; 1066 a->frac_lo = q1; 1067 return ret; 1068 } 1069 1070 #define frac_div(A, B) FRAC_GENERIC_64_128(div, A)(A, B) 1071 1072 static bool frac64_eqz(FloatParts64 *a) 1073 { 1074 return a->frac == 0; 1075 } 1076 1077 static bool frac128_eqz(FloatParts128 *a) 1078 { 1079 return (a->frac_hi | a->frac_lo) == 0; 1080 } 1081 1082 #define frac_eqz(A) FRAC_GENERIC_64_128(eqz, A)(A) 1083 1084 static void frac64_mulw(FloatParts128 *r, FloatParts64 *a, FloatParts64 *b) 1085 { 1086 mulu64(&r->frac_lo, &r->frac_hi, a->frac, b->frac); 1087 } 1088 1089 static void frac128_mulw(FloatParts256 *r, FloatParts128 *a, FloatParts128 *b) 1090 { 1091 mul128To256(a->frac_hi, a->frac_lo, b->frac_hi, b->frac_lo, 1092 &r->frac_hi, &r->frac_hm, &r->frac_lm, &r->frac_lo); 1093 } 1094 1095 #define frac_mulw(R, A, B) FRAC_GENERIC_64_128(mulw, A)(R, A, B) 1096 1097 static void frac64_neg(FloatParts64 *a) 1098 { 1099 a->frac = -a->frac; 1100 } 1101 1102 static void frac128_neg(FloatParts128 *a) 1103 { 1104 bool c = 0; 1105 a->frac_lo = usub64_borrow(0, a->frac_lo, &c); 1106 a->frac_hi = usub64_borrow(0, a->frac_hi, &c); 1107 } 1108 1109 static void frac256_neg(FloatParts256 *a) 1110 { 1111 bool c = 0; 1112 a->frac_lo = usub64_borrow(0, a->frac_lo, &c); 1113 a->frac_lm = usub64_borrow(0, a->frac_lm, &c); 1114 a->frac_hm = usub64_borrow(0, a->frac_hm, &c); 1115 a->frac_hi = usub64_borrow(0, a->frac_hi, &c); 1116 } 1117 1118 #define frac_neg(A) FRAC_GENERIC_64_128_256(neg, A)(A) 1119 1120 static int frac64_normalize(FloatParts64 *a) 1121 { 1122 if (a->frac) { 1123 int shift = clz64(a->frac); 1124 a->frac <<= shift; 1125 return shift; 1126 } 1127 return 64; 1128 } 1129 1130 static int frac128_normalize(FloatParts128 *a) 1131 { 1132 if (a->frac_hi) { 1133 int shl = clz64(a->frac_hi); 1134 a->frac_hi = shl_double(a->frac_hi, a->frac_lo, shl); 1135 a->frac_lo <<= shl; 1136 return shl; 1137 } else if (a->frac_lo) { 1138 int shl = clz64(a->frac_lo); 1139 a->frac_hi = a->frac_lo << shl; 1140 a->frac_lo = 0; 1141 return shl + 64; 1142 } 1143 return 128; 1144 } 1145 1146 static int frac256_normalize(FloatParts256 *a) 1147 { 1148 uint64_t a0 = a->frac_hi, a1 = a->frac_hm; 1149 uint64_t a2 = a->frac_lm, a3 = a->frac_lo; 1150 int ret, shl; 1151 1152 if (likely(a0)) { 1153 shl = clz64(a0); 1154 if (shl == 0) { 1155 return 0; 1156 } 1157 ret = shl; 1158 } else { 1159 if (a1) { 1160 ret = 64; 1161 a0 = a1, a1 = a2, a2 = a3, a3 = 0; 1162 } else if (a2) { 1163 ret = 128; 1164 a0 = a2, a1 = a3, a2 = 0, a3 = 0; 1165 } else if (a3) { 1166 ret = 192; 1167 a0 = a3, a1 = 0, a2 = 0, a3 = 0; 1168 } else { 1169 ret = 256; 1170 a0 = 0, a1 = 0, a2 = 0, a3 = 0; 1171 goto done; 1172 } 1173 shl = clz64(a0); 1174 if (shl == 0) { 1175 goto done; 1176 } 1177 ret += shl; 1178 } 1179 1180 a0 = shl_double(a0, a1, shl); 1181 a1 = shl_double(a1, a2, shl); 1182 a2 = shl_double(a2, a3, shl); 1183 a3 <<= shl; 1184 1185 done: 1186 a->frac_hi = a0; 1187 a->frac_hm = a1; 1188 a->frac_lm = a2; 1189 a->frac_lo = a3; 1190 return ret; 1191 } 1192 1193 #define frac_normalize(A) FRAC_GENERIC_64_128_256(normalize, A)(A) 1194 1195 static void frac64_modrem(FloatParts64 *a, FloatParts64 *b, uint64_t *mod_quot) 1196 { 1197 uint64_t a0, a1, b0, t0, t1, q, quot; 1198 int exp_diff = a->exp - b->exp; 1199 int shift; 1200 1201 a0 = a->frac; 1202 a1 = 0; 1203 1204 if (exp_diff < -1) { 1205 if (mod_quot) { 1206 *mod_quot = 0; 1207 } 1208 return; 1209 } 1210 if (exp_diff == -1) { 1211 a0 >>= 1; 1212 exp_diff = 0; 1213 } 1214 1215 b0 = b->frac; 1216 quot = q = b0 <= a0; 1217 if (q) { 1218 a0 -= b0; 1219 } 1220 1221 exp_diff -= 64; 1222 while (exp_diff > 0) { 1223 q = estimateDiv128To64(a0, a1, b0); 1224 q = q > 2 ? q - 2 : 0; 1225 mul64To128(b0, q, &t0, &t1); 1226 sub128(a0, a1, t0, t1, &a0, &a1); 1227 shortShift128Left(a0, a1, 62, &a0, &a1); 1228 exp_diff -= 62; 1229 quot = (quot << 62) + q; 1230 } 1231 1232 exp_diff += 64; 1233 if (exp_diff > 0) { 1234 q = estimateDiv128To64(a0, a1, b0); 1235 q = q > 2 ? (q - 2) >> (64 - exp_diff) : 0; 1236 mul64To128(b0, q << (64 - exp_diff), &t0, &t1); 1237 sub128(a0, a1, t0, t1, &a0, &a1); 1238 shortShift128Left(0, b0, 64 - exp_diff, &t0, &t1); 1239 while (le128(t0, t1, a0, a1)) { 1240 ++q; 1241 sub128(a0, a1, t0, t1, &a0, &a1); 1242 } 1243 quot = (exp_diff < 64 ? quot << exp_diff : 0) + q; 1244 } else { 1245 t0 = b0; 1246 t1 = 0; 1247 } 1248 1249 if (mod_quot) { 1250 *mod_quot = quot; 1251 } else { 1252 sub128(t0, t1, a0, a1, &t0, &t1); 1253 if (lt128(t0, t1, a0, a1) || 1254 (eq128(t0, t1, a0, a1) && (q & 1))) { 1255 a0 = t0; 1256 a1 = t1; 1257 a->sign = !a->sign; 1258 } 1259 } 1260 1261 if (likely(a0)) { 1262 shift = clz64(a0); 1263 shortShift128Left(a0, a1, shift, &a0, &a1); 1264 } else if (likely(a1)) { 1265 shift = clz64(a1); 1266 a0 = a1 << shift; 1267 a1 = 0; 1268 shift += 64; 1269 } else { 1270 a->cls = float_class_zero; 1271 return; 1272 } 1273 1274 a->exp = b->exp + exp_diff - shift; 1275 a->frac = a0 | (a1 != 0); 1276 } 1277 1278 static void frac128_modrem(FloatParts128 *a, FloatParts128 *b, 1279 uint64_t *mod_quot) 1280 { 1281 uint64_t a0, a1, a2, b0, b1, t0, t1, t2, q, quot; 1282 int exp_diff = a->exp - b->exp; 1283 int shift; 1284 1285 a0 = a->frac_hi; 1286 a1 = a->frac_lo; 1287 a2 = 0; 1288 1289 if (exp_diff < -1) { 1290 if (mod_quot) { 1291 *mod_quot = 0; 1292 } 1293 return; 1294 } 1295 if (exp_diff == -1) { 1296 shift128Right(a0, a1, 1, &a0, &a1); 1297 exp_diff = 0; 1298 } 1299 1300 b0 = b->frac_hi; 1301 b1 = b->frac_lo; 1302 1303 quot = q = le128(b0, b1, a0, a1); 1304 if (q) { 1305 sub128(a0, a1, b0, b1, &a0, &a1); 1306 } 1307 1308 exp_diff -= 64; 1309 while (exp_diff > 0) { 1310 q = estimateDiv128To64(a0, a1, b0); 1311 q = q > 4 ? q - 4 : 0; 1312 mul128By64To192(b0, b1, q, &t0, &t1, &t2); 1313 sub192(a0, a1, a2, t0, t1, t2, &a0, &a1, &a2); 1314 shortShift192Left(a0, a1, a2, 61, &a0, &a1, &a2); 1315 exp_diff -= 61; 1316 quot = (quot << 61) + q; 1317 } 1318 1319 exp_diff += 64; 1320 if (exp_diff > 0) { 1321 q = estimateDiv128To64(a0, a1, b0); 1322 q = q > 4 ? (q - 4) >> (64 - exp_diff) : 0; 1323 mul128By64To192(b0, b1, q << (64 - exp_diff), &t0, &t1, &t2); 1324 sub192(a0, a1, a2, t0, t1, t2, &a0, &a1, &a2); 1325 shortShift192Left(0, b0, b1, 64 - exp_diff, &t0, &t1, &t2); 1326 while (le192(t0, t1, t2, a0, a1, a2)) { 1327 ++q; 1328 sub192(a0, a1, a2, t0, t1, t2, &a0, &a1, &a2); 1329 } 1330 quot = (exp_diff < 64 ? quot << exp_diff : 0) + q; 1331 } else { 1332 t0 = b0; 1333 t1 = b1; 1334 t2 = 0; 1335 } 1336 1337 if (mod_quot) { 1338 *mod_quot = quot; 1339 } else { 1340 sub192(t0, t1, t2, a0, a1, a2, &t0, &t1, &t2); 1341 if (lt192(t0, t1, t2, a0, a1, a2) || 1342 (eq192(t0, t1, t2, a0, a1, a2) && (q & 1))) { 1343 a0 = t0; 1344 a1 = t1; 1345 a2 = t2; 1346 a->sign = !a->sign; 1347 } 1348 } 1349 1350 if (likely(a0)) { 1351 shift = clz64(a0); 1352 shortShift192Left(a0, a1, a2, shift, &a0, &a1, &a2); 1353 } else if (likely(a1)) { 1354 shift = clz64(a1); 1355 shortShift128Left(a1, a2, shift, &a0, &a1); 1356 a2 = 0; 1357 shift += 64; 1358 } else if (likely(a2)) { 1359 shift = clz64(a2); 1360 a0 = a2 << shift; 1361 a1 = a2 = 0; 1362 shift += 128; 1363 } else { 1364 a->cls = float_class_zero; 1365 return; 1366 } 1367 1368 a->exp = b->exp + exp_diff - shift; 1369 a->frac_hi = a0; 1370 a->frac_lo = a1 | (a2 != 0); 1371 } 1372 1373 #define frac_modrem(A, B, Q) FRAC_GENERIC_64_128(modrem, A)(A, B, Q) 1374 1375 static void frac64_shl(FloatParts64 *a, int c) 1376 { 1377 a->frac <<= c; 1378 } 1379 1380 static void frac128_shl(FloatParts128 *a, int c) 1381 { 1382 uint64_t a0 = a->frac_hi, a1 = a->frac_lo; 1383 1384 if (c & 64) { 1385 a0 = a1, a1 = 0; 1386 } 1387 1388 c &= 63; 1389 if (c) { 1390 a0 = shl_double(a0, a1, c); 1391 a1 = a1 << c; 1392 } 1393 1394 a->frac_hi = a0; 1395 a->frac_lo = a1; 1396 } 1397 1398 #define frac_shl(A, C) FRAC_GENERIC_64_128(shl, A)(A, C) 1399 1400 static void frac64_shr(FloatParts64 *a, int c) 1401 { 1402 a->frac >>= c; 1403 } 1404 1405 static void frac128_shr(FloatParts128 *a, int c) 1406 { 1407 uint64_t a0 = a->frac_hi, a1 = a->frac_lo; 1408 1409 if (c & 64) { 1410 a1 = a0, a0 = 0; 1411 } 1412 1413 c &= 63; 1414 if (c) { 1415 a1 = shr_double(a0, a1, c); 1416 a0 = a0 >> c; 1417 } 1418 1419 a->frac_hi = a0; 1420 a->frac_lo = a1; 1421 } 1422 1423 #define frac_shr(A, C) FRAC_GENERIC_64_128(shr, A)(A, C) 1424 1425 static void frac64_shrjam(FloatParts64 *a, int c) 1426 { 1427 uint64_t a0 = a->frac; 1428 1429 if (likely(c != 0)) { 1430 if (likely(c < 64)) { 1431 a0 = (a0 >> c) | (shr_double(a0, 0, c) != 0); 1432 } else { 1433 a0 = a0 != 0; 1434 } 1435 a->frac = a0; 1436 } 1437 } 1438 1439 static void frac128_shrjam(FloatParts128 *a, int c) 1440 { 1441 uint64_t a0 = a->frac_hi, a1 = a->frac_lo; 1442 uint64_t sticky = 0; 1443 1444 if (unlikely(c == 0)) { 1445 return; 1446 } else if (likely(c < 64)) { 1447 /* nothing */ 1448 } else if (likely(c < 128)) { 1449 sticky = a1; 1450 a1 = a0; 1451 a0 = 0; 1452 c &= 63; 1453 if (c == 0) { 1454 goto done; 1455 } 1456 } else { 1457 sticky = a0 | a1; 1458 a0 = a1 = 0; 1459 goto done; 1460 } 1461 1462 sticky |= shr_double(a1, 0, c); 1463 a1 = shr_double(a0, a1, c); 1464 a0 = a0 >> c; 1465 1466 done: 1467 a->frac_lo = a1 | (sticky != 0); 1468 a->frac_hi = a0; 1469 } 1470 1471 static void frac256_shrjam(FloatParts256 *a, int c) 1472 { 1473 uint64_t a0 = a->frac_hi, a1 = a->frac_hm; 1474 uint64_t a2 = a->frac_lm, a3 = a->frac_lo; 1475 uint64_t sticky = 0; 1476 1477 if (unlikely(c == 0)) { 1478 return; 1479 } else if (likely(c < 64)) { 1480 /* nothing */ 1481 } else if (likely(c < 256)) { 1482 if (unlikely(c & 128)) { 1483 sticky |= a2 | a3; 1484 a3 = a1, a2 = a0, a1 = 0, a0 = 0; 1485 } 1486 if (unlikely(c & 64)) { 1487 sticky |= a3; 1488 a3 = a2, a2 = a1, a1 = a0, a0 = 0; 1489 } 1490 c &= 63; 1491 if (c == 0) { 1492 goto done; 1493 } 1494 } else { 1495 sticky = a0 | a1 | a2 | a3; 1496 a0 = a1 = a2 = a3 = 0; 1497 goto done; 1498 } 1499 1500 sticky |= shr_double(a3, 0, c); 1501 a3 = shr_double(a2, a3, c); 1502 a2 = shr_double(a1, a2, c); 1503 a1 = shr_double(a0, a1, c); 1504 a0 = a0 >> c; 1505 1506 done: 1507 a->frac_lo = a3 | (sticky != 0); 1508 a->frac_lm = a2; 1509 a->frac_hm = a1; 1510 a->frac_hi = a0; 1511 } 1512 1513 #define frac_shrjam(A, C) FRAC_GENERIC_64_128_256(shrjam, A)(A, C) 1514 1515 static bool frac64_sub(FloatParts64 *r, FloatParts64 *a, FloatParts64 *b) 1516 { 1517 return usub64_overflow(a->frac, b->frac, &r->frac); 1518 } 1519 1520 static bool frac128_sub(FloatParts128 *r, FloatParts128 *a, FloatParts128 *b) 1521 { 1522 bool c = 0; 1523 r->frac_lo = usub64_borrow(a->frac_lo, b->frac_lo, &c); 1524 r->frac_hi = usub64_borrow(a->frac_hi, b->frac_hi, &c); 1525 return c; 1526 } 1527 1528 static bool frac256_sub(FloatParts256 *r, FloatParts256 *a, FloatParts256 *b) 1529 { 1530 bool c = 0; 1531 r->frac_lo = usub64_borrow(a->frac_lo, b->frac_lo, &c); 1532 r->frac_lm = usub64_borrow(a->frac_lm, b->frac_lm, &c); 1533 r->frac_hm = usub64_borrow(a->frac_hm, b->frac_hm, &c); 1534 r->frac_hi = usub64_borrow(a->frac_hi, b->frac_hi, &c); 1535 return c; 1536 } 1537 1538 #define frac_sub(R, A, B) FRAC_GENERIC_64_128_256(sub, R)(R, A, B) 1539 1540 static void frac64_truncjam(FloatParts64 *r, FloatParts128 *a) 1541 { 1542 r->frac = a->frac_hi | (a->frac_lo != 0); 1543 } 1544 1545 static void frac128_truncjam(FloatParts128 *r, FloatParts256 *a) 1546 { 1547 r->frac_hi = a->frac_hi; 1548 r->frac_lo = a->frac_hm | ((a->frac_lm | a->frac_lo) != 0); 1549 } 1550 1551 #define frac_truncjam(R, A) FRAC_GENERIC_64_128(truncjam, R)(R, A) 1552 1553 static void frac64_widen(FloatParts128 *r, FloatParts64 *a) 1554 { 1555 r->frac_hi = a->frac; 1556 r->frac_lo = 0; 1557 } 1558 1559 static void frac128_widen(FloatParts256 *r, FloatParts128 *a) 1560 { 1561 r->frac_hi = a->frac_hi; 1562 r->frac_hm = a->frac_lo; 1563 r->frac_lm = 0; 1564 r->frac_lo = 0; 1565 } 1566 1567 #define frac_widen(A, B) FRAC_GENERIC_64_128(widen, B)(A, B) 1568 1569 /* 1570 * Reciprocal sqrt table. 1 bit of exponent, 6-bits of mantessa. 1571 * From https://git.musl-libc.org/cgit/musl/tree/src/math/sqrt_data.c 1572 * and thus MIT licenced. 1573 */ 1574 static const uint16_t rsqrt_tab[128] = { 1575 0xb451, 0xb2f0, 0xb196, 0xb044, 0xaef9, 0xadb6, 0xac79, 0xab43, 1576 0xaa14, 0xa8eb, 0xa7c8, 0xa6aa, 0xa592, 0xa480, 0xa373, 0xa26b, 1577 0xa168, 0xa06a, 0x9f70, 0x9e7b, 0x9d8a, 0x9c9d, 0x9bb5, 0x9ad1, 1578 0x99f0, 0x9913, 0x983a, 0x9765, 0x9693, 0x95c4, 0x94f8, 0x9430, 1579 0x936b, 0x92a9, 0x91ea, 0x912e, 0x9075, 0x8fbe, 0x8f0a, 0x8e59, 1580 0x8daa, 0x8cfe, 0x8c54, 0x8bac, 0x8b07, 0x8a64, 0x89c4, 0x8925, 1581 0x8889, 0x87ee, 0x8756, 0x86c0, 0x862b, 0x8599, 0x8508, 0x8479, 1582 0x83ec, 0x8361, 0x82d8, 0x8250, 0x81c9, 0x8145, 0x80c2, 0x8040, 1583 0xff02, 0xfd0e, 0xfb25, 0xf947, 0xf773, 0xf5aa, 0xf3ea, 0xf234, 1584 0xf087, 0xeee3, 0xed47, 0xebb3, 0xea27, 0xe8a3, 0xe727, 0xe5b2, 1585 0xe443, 0xe2dc, 0xe17a, 0xe020, 0xdecb, 0xdd7d, 0xdc34, 0xdaf1, 1586 0xd9b3, 0xd87b, 0xd748, 0xd61a, 0xd4f1, 0xd3cd, 0xd2ad, 0xd192, 1587 0xd07b, 0xcf69, 0xce5b, 0xcd51, 0xcc4a, 0xcb48, 0xca4a, 0xc94f, 1588 0xc858, 0xc764, 0xc674, 0xc587, 0xc49d, 0xc3b7, 0xc2d4, 0xc1f4, 1589 0xc116, 0xc03c, 0xbf65, 0xbe90, 0xbdbe, 0xbcef, 0xbc23, 0xbb59, 1590 0xba91, 0xb9cc, 0xb90a, 0xb84a, 0xb78c, 0xb6d0, 0xb617, 0xb560, 1591 }; 1592 1593 #define partsN(NAME) glue(glue(glue(parts,N),_),NAME) 1594 #define FloatPartsN glue(FloatParts,N) 1595 #define FloatPartsW glue(FloatParts,W) 1596 1597 #define N 64 1598 #define W 128 1599 1600 #include "softfloat-parts-addsub.c.inc" 1601 #include "softfloat-parts.c.inc" 1602 1603 #undef N 1604 #undef W 1605 #define N 128 1606 #define W 256 1607 1608 #include "softfloat-parts-addsub.c.inc" 1609 #include "softfloat-parts.c.inc" 1610 1611 #undef N 1612 #undef W 1613 #define N 256 1614 1615 #include "softfloat-parts-addsub.c.inc" 1616 1617 #undef N 1618 #undef W 1619 #undef partsN 1620 #undef FloatPartsN 1621 #undef FloatPartsW 1622 1623 /* 1624 * Pack/unpack routines with a specific FloatFmt. 1625 */ 1626 1627 static void float16a_unpack_canonical(FloatParts64 *p, float16 f, 1628 float_status *s, const FloatFmt *params) 1629 { 1630 float16_unpack_raw(p, f); 1631 parts_canonicalize(p, s, params); 1632 } 1633 1634 static void float16_unpack_canonical(FloatParts64 *p, float16 f, 1635 float_status *s) 1636 { 1637 float16a_unpack_canonical(p, f, s, &float16_params); 1638 } 1639 1640 static void bfloat16_unpack_canonical(FloatParts64 *p, bfloat16 f, 1641 float_status *s) 1642 { 1643 bfloat16_unpack_raw(p, f); 1644 parts_canonicalize(p, s, &bfloat16_params); 1645 } 1646 1647 static float16 float16a_round_pack_canonical(FloatParts64 *p, 1648 float_status *s, 1649 const FloatFmt *params) 1650 { 1651 parts_uncanon(p, s, params); 1652 return float16_pack_raw(p); 1653 } 1654 1655 static float16 float16_round_pack_canonical(FloatParts64 *p, 1656 float_status *s) 1657 { 1658 return float16a_round_pack_canonical(p, s, &float16_params); 1659 } 1660 1661 static bfloat16 bfloat16_round_pack_canonical(FloatParts64 *p, 1662 float_status *s) 1663 { 1664 parts_uncanon(p, s, &bfloat16_params); 1665 return bfloat16_pack_raw(p); 1666 } 1667 1668 static void float32_unpack_canonical(FloatParts64 *p, float32 f, 1669 float_status *s) 1670 { 1671 float32_unpack_raw(p, f); 1672 parts_canonicalize(p, s, &float32_params); 1673 } 1674 1675 static float32 float32_round_pack_canonical(FloatParts64 *p, 1676 float_status *s) 1677 { 1678 parts_uncanon(p, s, &float32_params); 1679 return float32_pack_raw(p); 1680 } 1681 1682 static void float64_unpack_canonical(FloatParts64 *p, float64 f, 1683 float_status *s) 1684 { 1685 float64_unpack_raw(p, f); 1686 parts_canonicalize(p, s, &float64_params); 1687 } 1688 1689 static float64 float64_round_pack_canonical(FloatParts64 *p, 1690 float_status *s) 1691 { 1692 parts_uncanon(p, s, &float64_params); 1693 return float64_pack_raw(p); 1694 } 1695 1696 static float64 float64r32_round_pack_canonical(FloatParts64 *p, 1697 float_status *s) 1698 { 1699 parts_uncanon(p, s, &float32_params); 1700 1701 /* 1702 * In parts_uncanon, we placed the fraction for float32 at the lsb. 1703 * We need to adjust the fraction higher so that the least N bits are 1704 * zero, and the fraction is adjacent to the float64 implicit bit. 1705 */ 1706 switch (p->cls) { 1707 case float_class_normal: 1708 if (unlikely(p->exp == 0)) { 1709 /* 1710 * The result is denormal for float32, but can be represented 1711 * in normalized form for float64. Adjust, per canonicalize. 1712 */ 1713 int shift = frac_normalize(p); 1714 p->exp = (float32_params.frac_shift - 1715 float32_params.exp_bias - shift + 1 + 1716 float64_params.exp_bias); 1717 frac_shr(p, float64_params.frac_shift); 1718 } else { 1719 frac_shl(p, float32_params.frac_shift - float64_params.frac_shift); 1720 p->exp += float64_params.exp_bias - float32_params.exp_bias; 1721 } 1722 break; 1723 case float_class_snan: 1724 case float_class_qnan: 1725 frac_shl(p, float32_params.frac_shift - float64_params.frac_shift); 1726 p->exp = float64_params.exp_max; 1727 break; 1728 case float_class_inf: 1729 p->exp = float64_params.exp_max; 1730 break; 1731 case float_class_zero: 1732 break; 1733 default: 1734 g_assert_not_reached(); 1735 } 1736 1737 return float64_pack_raw(p); 1738 } 1739 1740 static void float128_unpack_canonical(FloatParts128 *p, float128 f, 1741 float_status *s) 1742 { 1743 float128_unpack_raw(p, f); 1744 parts_canonicalize(p, s, &float128_params); 1745 } 1746 1747 static float128 float128_round_pack_canonical(FloatParts128 *p, 1748 float_status *s) 1749 { 1750 parts_uncanon(p, s, &float128_params); 1751 return float128_pack_raw(p); 1752 } 1753 1754 /* Returns false if the encoding is invalid. */ 1755 static bool floatx80_unpack_canonical(FloatParts128 *p, floatx80 f, 1756 float_status *s) 1757 { 1758 /* Ensure rounding precision is set before beginning. */ 1759 switch (s->floatx80_rounding_precision) { 1760 case floatx80_precision_x: 1761 case floatx80_precision_d: 1762 case floatx80_precision_s: 1763 break; 1764 default: 1765 g_assert_not_reached(); 1766 } 1767 1768 if (unlikely(floatx80_invalid_encoding(f))) { 1769 float_raise(float_flag_invalid, s); 1770 return false; 1771 } 1772 1773 floatx80_unpack_raw(p, f); 1774 1775 if (likely(p->exp != floatx80_params[floatx80_precision_x].exp_max)) { 1776 parts_canonicalize(p, s, &floatx80_params[floatx80_precision_x]); 1777 } else { 1778 /* The explicit integer bit is ignored, after invalid checks. */ 1779 p->frac_hi &= MAKE_64BIT_MASK(0, 63); 1780 p->cls = (p->frac_hi == 0 ? float_class_inf 1781 : parts_is_snan_frac(p->frac_hi, s) 1782 ? float_class_snan : float_class_qnan); 1783 } 1784 return true; 1785 } 1786 1787 static floatx80 floatx80_round_pack_canonical(FloatParts128 *p, 1788 float_status *s) 1789 { 1790 const FloatFmt *fmt = &floatx80_params[s->floatx80_rounding_precision]; 1791 uint64_t frac; 1792 int exp; 1793 1794 switch (p->cls) { 1795 case float_class_normal: 1796 if (s->floatx80_rounding_precision == floatx80_precision_x) { 1797 parts_uncanon_normal(p, s, fmt); 1798 frac = p->frac_hi; 1799 exp = p->exp; 1800 } else { 1801 FloatParts64 p64; 1802 1803 p64.sign = p->sign; 1804 p64.exp = p->exp; 1805 frac_truncjam(&p64, p); 1806 parts_uncanon_normal(&p64, s, fmt); 1807 frac = p64.frac; 1808 exp = p64.exp; 1809 } 1810 if (exp != fmt->exp_max) { 1811 break; 1812 } 1813 /* rounded to inf -- fall through to set frac correctly */ 1814 1815 case float_class_inf: 1816 /* x86 and m68k differ in the setting of the integer bit. */ 1817 frac = floatx80_infinity_low; 1818 exp = fmt->exp_max; 1819 break; 1820 1821 case float_class_zero: 1822 frac = 0; 1823 exp = 0; 1824 break; 1825 1826 case float_class_snan: 1827 case float_class_qnan: 1828 /* NaNs have the integer bit set. */ 1829 frac = p->frac_hi | (1ull << 63); 1830 exp = fmt->exp_max; 1831 break; 1832 1833 default: 1834 g_assert_not_reached(); 1835 } 1836 1837 return packFloatx80(p->sign, exp, frac); 1838 } 1839 1840 /* 1841 * Addition and subtraction 1842 */ 1843 1844 static float16 QEMU_FLATTEN 1845 float16_addsub(float16 a, float16 b, float_status *status, bool subtract) 1846 { 1847 FloatParts64 pa, pb, *pr; 1848 1849 float16_unpack_canonical(&pa, a, status); 1850 float16_unpack_canonical(&pb, b, status); 1851 pr = parts_addsub(&pa, &pb, status, subtract); 1852 1853 return float16_round_pack_canonical(pr, status); 1854 } 1855 1856 float16 float16_add(float16 a, float16 b, float_status *status) 1857 { 1858 return float16_addsub(a, b, status, false); 1859 } 1860 1861 float16 float16_sub(float16 a, float16 b, float_status *status) 1862 { 1863 return float16_addsub(a, b, status, true); 1864 } 1865 1866 static float32 QEMU_SOFTFLOAT_ATTR 1867 soft_f32_addsub(float32 a, float32 b, float_status *status, bool subtract) 1868 { 1869 FloatParts64 pa, pb, *pr; 1870 1871 float32_unpack_canonical(&pa, a, status); 1872 float32_unpack_canonical(&pb, b, status); 1873 pr = parts_addsub(&pa, &pb, status, subtract); 1874 1875 return float32_round_pack_canonical(pr, status); 1876 } 1877 1878 static float32 soft_f32_add(float32 a, float32 b, float_status *status) 1879 { 1880 return soft_f32_addsub(a, b, status, false); 1881 } 1882 1883 static float32 soft_f32_sub(float32 a, float32 b, float_status *status) 1884 { 1885 return soft_f32_addsub(a, b, status, true); 1886 } 1887 1888 static float64 QEMU_SOFTFLOAT_ATTR 1889 soft_f64_addsub(float64 a, float64 b, float_status *status, bool subtract) 1890 { 1891 FloatParts64 pa, pb, *pr; 1892 1893 float64_unpack_canonical(&pa, a, status); 1894 float64_unpack_canonical(&pb, b, status); 1895 pr = parts_addsub(&pa, &pb, status, subtract); 1896 1897 return float64_round_pack_canonical(pr, status); 1898 } 1899 1900 static float64 soft_f64_add(float64 a, float64 b, float_status *status) 1901 { 1902 return soft_f64_addsub(a, b, status, false); 1903 } 1904 1905 static float64 soft_f64_sub(float64 a, float64 b, float_status *status) 1906 { 1907 return soft_f64_addsub(a, b, status, true); 1908 } 1909 1910 static float hard_f32_add(float a, float b) 1911 { 1912 return a + b; 1913 } 1914 1915 static float hard_f32_sub(float a, float b) 1916 { 1917 return a - b; 1918 } 1919 1920 static double hard_f64_add(double a, double b) 1921 { 1922 return a + b; 1923 } 1924 1925 static double hard_f64_sub(double a, double b) 1926 { 1927 return a - b; 1928 } 1929 1930 static bool f32_addsubmul_post(union_float32 a, union_float32 b) 1931 { 1932 if (QEMU_HARDFLOAT_2F32_USE_FP) { 1933 return !(fpclassify(a.h) == FP_ZERO && fpclassify(b.h) == FP_ZERO); 1934 } 1935 return !(float32_is_zero(a.s) && float32_is_zero(b.s)); 1936 } 1937 1938 static bool f64_addsubmul_post(union_float64 a, union_float64 b) 1939 { 1940 if (QEMU_HARDFLOAT_2F64_USE_FP) { 1941 return !(fpclassify(a.h) == FP_ZERO && fpclassify(b.h) == FP_ZERO); 1942 } else { 1943 return !(float64_is_zero(a.s) && float64_is_zero(b.s)); 1944 } 1945 } 1946 1947 static float32 float32_addsub(float32 a, float32 b, float_status *s, 1948 hard_f32_op2_fn hard, soft_f32_op2_fn soft) 1949 { 1950 return float32_gen2(a, b, s, hard, soft, 1951 f32_is_zon2, f32_addsubmul_post); 1952 } 1953 1954 static float64 float64_addsub(float64 a, float64 b, float_status *s, 1955 hard_f64_op2_fn hard, soft_f64_op2_fn soft) 1956 { 1957 return float64_gen2(a, b, s, hard, soft, 1958 f64_is_zon2, f64_addsubmul_post); 1959 } 1960 1961 float32 QEMU_FLATTEN 1962 float32_add(float32 a, float32 b, float_status *s) 1963 { 1964 return float32_addsub(a, b, s, hard_f32_add, soft_f32_add); 1965 } 1966 1967 float32 QEMU_FLATTEN 1968 float32_sub(float32 a, float32 b, float_status *s) 1969 { 1970 return float32_addsub(a, b, s, hard_f32_sub, soft_f32_sub); 1971 } 1972 1973 float64 QEMU_FLATTEN 1974 float64_add(float64 a, float64 b, float_status *s) 1975 { 1976 return float64_addsub(a, b, s, hard_f64_add, soft_f64_add); 1977 } 1978 1979 float64 QEMU_FLATTEN 1980 float64_sub(float64 a, float64 b, float_status *s) 1981 { 1982 return float64_addsub(a, b, s, hard_f64_sub, soft_f64_sub); 1983 } 1984 1985 static float64 float64r32_addsub(float64 a, float64 b, float_status *status, 1986 bool subtract) 1987 { 1988 FloatParts64 pa, pb, *pr; 1989 1990 float64_unpack_canonical(&pa, a, status); 1991 float64_unpack_canonical(&pb, b, status); 1992 pr = parts_addsub(&pa, &pb, status, subtract); 1993 1994 return float64r32_round_pack_canonical(pr, status); 1995 } 1996 1997 float64 float64r32_add(float64 a, float64 b, float_status *status) 1998 { 1999 return float64r32_addsub(a, b, status, false); 2000 } 2001 2002 float64 float64r32_sub(float64 a, float64 b, float_status *status) 2003 { 2004 return float64r32_addsub(a, b, status, true); 2005 } 2006 2007 static bfloat16 QEMU_FLATTEN 2008 bfloat16_addsub(bfloat16 a, bfloat16 b, float_status *status, bool subtract) 2009 { 2010 FloatParts64 pa, pb, *pr; 2011 2012 bfloat16_unpack_canonical(&pa, a, status); 2013 bfloat16_unpack_canonical(&pb, b, status); 2014 pr = parts_addsub(&pa, &pb, status, subtract); 2015 2016 return bfloat16_round_pack_canonical(pr, status); 2017 } 2018 2019 bfloat16 bfloat16_add(bfloat16 a, bfloat16 b, float_status *status) 2020 { 2021 return bfloat16_addsub(a, b, status, false); 2022 } 2023 2024 bfloat16 bfloat16_sub(bfloat16 a, bfloat16 b, float_status *status) 2025 { 2026 return bfloat16_addsub(a, b, status, true); 2027 } 2028 2029 static float128 QEMU_FLATTEN 2030 float128_addsub(float128 a, float128 b, float_status *status, bool subtract) 2031 { 2032 FloatParts128 pa, pb, *pr; 2033 2034 float128_unpack_canonical(&pa, a, status); 2035 float128_unpack_canonical(&pb, b, status); 2036 pr = parts_addsub(&pa, &pb, status, subtract); 2037 2038 return float128_round_pack_canonical(pr, status); 2039 } 2040 2041 float128 float128_add(float128 a, float128 b, float_status *status) 2042 { 2043 return float128_addsub(a, b, status, false); 2044 } 2045 2046 float128 float128_sub(float128 a, float128 b, float_status *status) 2047 { 2048 return float128_addsub(a, b, status, true); 2049 } 2050 2051 static floatx80 QEMU_FLATTEN 2052 floatx80_addsub(floatx80 a, floatx80 b, float_status *status, bool subtract) 2053 { 2054 FloatParts128 pa, pb, *pr; 2055 2056 if (!floatx80_unpack_canonical(&pa, a, status) || 2057 !floatx80_unpack_canonical(&pb, b, status)) { 2058 return floatx80_default_nan(status); 2059 } 2060 2061 pr = parts_addsub(&pa, &pb, status, subtract); 2062 return floatx80_round_pack_canonical(pr, status); 2063 } 2064 2065 floatx80 floatx80_add(floatx80 a, floatx80 b, float_status *status) 2066 { 2067 return floatx80_addsub(a, b, status, false); 2068 } 2069 2070 floatx80 floatx80_sub(floatx80 a, floatx80 b, float_status *status) 2071 { 2072 return floatx80_addsub(a, b, status, true); 2073 } 2074 2075 /* 2076 * Multiplication 2077 */ 2078 2079 float16 QEMU_FLATTEN float16_mul(float16 a, float16 b, float_status *status) 2080 { 2081 FloatParts64 pa, pb, *pr; 2082 2083 float16_unpack_canonical(&pa, a, status); 2084 float16_unpack_canonical(&pb, b, status); 2085 pr = parts_mul(&pa, &pb, status); 2086 2087 return float16_round_pack_canonical(pr, status); 2088 } 2089 2090 static float32 QEMU_SOFTFLOAT_ATTR 2091 soft_f32_mul(float32 a, float32 b, float_status *status) 2092 { 2093 FloatParts64 pa, pb, *pr; 2094 2095 float32_unpack_canonical(&pa, a, status); 2096 float32_unpack_canonical(&pb, b, status); 2097 pr = parts_mul(&pa, &pb, status); 2098 2099 return float32_round_pack_canonical(pr, status); 2100 } 2101 2102 static float64 QEMU_SOFTFLOAT_ATTR 2103 soft_f64_mul(float64 a, float64 b, float_status *status) 2104 { 2105 FloatParts64 pa, pb, *pr; 2106 2107 float64_unpack_canonical(&pa, a, status); 2108 float64_unpack_canonical(&pb, b, status); 2109 pr = parts_mul(&pa, &pb, status); 2110 2111 return float64_round_pack_canonical(pr, status); 2112 } 2113 2114 static float hard_f32_mul(float a, float b) 2115 { 2116 return a * b; 2117 } 2118 2119 static double hard_f64_mul(double a, double b) 2120 { 2121 return a * b; 2122 } 2123 2124 float32 QEMU_FLATTEN 2125 float32_mul(float32 a, float32 b, float_status *s) 2126 { 2127 return float32_gen2(a, b, s, hard_f32_mul, soft_f32_mul, 2128 f32_is_zon2, f32_addsubmul_post); 2129 } 2130 2131 float64 QEMU_FLATTEN 2132 float64_mul(float64 a, float64 b, float_status *s) 2133 { 2134 return float64_gen2(a, b, s, hard_f64_mul, soft_f64_mul, 2135 f64_is_zon2, f64_addsubmul_post); 2136 } 2137 2138 float64 float64r32_mul(float64 a, float64 b, float_status *status) 2139 { 2140 FloatParts64 pa, pb, *pr; 2141 2142 float64_unpack_canonical(&pa, a, status); 2143 float64_unpack_canonical(&pb, b, status); 2144 pr = parts_mul(&pa, &pb, status); 2145 2146 return float64r32_round_pack_canonical(pr, status); 2147 } 2148 2149 bfloat16 QEMU_FLATTEN 2150 bfloat16_mul(bfloat16 a, bfloat16 b, float_status *status) 2151 { 2152 FloatParts64 pa, pb, *pr; 2153 2154 bfloat16_unpack_canonical(&pa, a, status); 2155 bfloat16_unpack_canonical(&pb, b, status); 2156 pr = parts_mul(&pa, &pb, status); 2157 2158 return bfloat16_round_pack_canonical(pr, status); 2159 } 2160 2161 float128 QEMU_FLATTEN 2162 float128_mul(float128 a, float128 b, float_status *status) 2163 { 2164 FloatParts128 pa, pb, *pr; 2165 2166 float128_unpack_canonical(&pa, a, status); 2167 float128_unpack_canonical(&pb, b, status); 2168 pr = parts_mul(&pa, &pb, status); 2169 2170 return float128_round_pack_canonical(pr, status); 2171 } 2172 2173 floatx80 QEMU_FLATTEN 2174 floatx80_mul(floatx80 a, floatx80 b, float_status *status) 2175 { 2176 FloatParts128 pa, pb, *pr; 2177 2178 if (!floatx80_unpack_canonical(&pa, a, status) || 2179 !floatx80_unpack_canonical(&pb, b, status)) { 2180 return floatx80_default_nan(status); 2181 } 2182 2183 pr = parts_mul(&pa, &pb, status); 2184 return floatx80_round_pack_canonical(pr, status); 2185 } 2186 2187 /* 2188 * Fused multiply-add 2189 */ 2190 2191 float16 QEMU_FLATTEN float16_muladd(float16 a, float16 b, float16 c, 2192 int flags, float_status *status) 2193 { 2194 FloatParts64 pa, pb, pc, *pr; 2195 2196 float16_unpack_canonical(&pa, a, status); 2197 float16_unpack_canonical(&pb, b, status); 2198 float16_unpack_canonical(&pc, c, status); 2199 pr = parts_muladd(&pa, &pb, &pc, flags, status); 2200 2201 return float16_round_pack_canonical(pr, status); 2202 } 2203 2204 static float32 QEMU_SOFTFLOAT_ATTR 2205 soft_f32_muladd(float32 a, float32 b, float32 c, int flags, 2206 float_status *status) 2207 { 2208 FloatParts64 pa, pb, pc, *pr; 2209 2210 float32_unpack_canonical(&pa, a, status); 2211 float32_unpack_canonical(&pb, b, status); 2212 float32_unpack_canonical(&pc, c, status); 2213 pr = parts_muladd(&pa, &pb, &pc, flags, status); 2214 2215 return float32_round_pack_canonical(pr, status); 2216 } 2217 2218 static float64 QEMU_SOFTFLOAT_ATTR 2219 soft_f64_muladd(float64 a, float64 b, float64 c, int flags, 2220 float_status *status) 2221 { 2222 FloatParts64 pa, pb, pc, *pr; 2223 2224 float64_unpack_canonical(&pa, a, status); 2225 float64_unpack_canonical(&pb, b, status); 2226 float64_unpack_canonical(&pc, c, status); 2227 pr = parts_muladd(&pa, &pb, &pc, flags, status); 2228 2229 return float64_round_pack_canonical(pr, status); 2230 } 2231 2232 static bool force_soft_fma; 2233 2234 float32 QEMU_FLATTEN 2235 float32_muladd(float32 xa, float32 xb, float32 xc, int flags, float_status *s) 2236 { 2237 union_float32 ua, ub, uc, ur; 2238 2239 ua.s = xa; 2240 ub.s = xb; 2241 uc.s = xc; 2242 2243 if (unlikely(!can_use_fpu(s))) { 2244 goto soft; 2245 } 2246 if (unlikely(flags & float_muladd_halve_result)) { 2247 goto soft; 2248 } 2249 2250 float32_input_flush3(&ua.s, &ub.s, &uc.s, s); 2251 if (unlikely(!f32_is_zon3(ua, ub, uc))) { 2252 goto soft; 2253 } 2254 2255 if (unlikely(force_soft_fma)) { 2256 goto soft; 2257 } 2258 2259 /* 2260 * When (a || b) == 0, there's no need to check for under/over flow, 2261 * since we know the addend is (normal || 0) and the product is 0. 2262 */ 2263 if (float32_is_zero(ua.s) || float32_is_zero(ub.s)) { 2264 union_float32 up; 2265 bool prod_sign; 2266 2267 prod_sign = float32_is_neg(ua.s) ^ float32_is_neg(ub.s); 2268 prod_sign ^= !!(flags & float_muladd_negate_product); 2269 up.s = float32_set_sign(float32_zero, prod_sign); 2270 2271 if (flags & float_muladd_negate_c) { 2272 uc.h = -uc.h; 2273 } 2274 ur.h = up.h + uc.h; 2275 } else { 2276 union_float32 ua_orig = ua; 2277 union_float32 uc_orig = uc; 2278 2279 if (flags & float_muladd_negate_product) { 2280 ua.h = -ua.h; 2281 } 2282 if (flags & float_muladd_negate_c) { 2283 uc.h = -uc.h; 2284 } 2285 2286 ur.h = fmaf(ua.h, ub.h, uc.h); 2287 2288 if (unlikely(f32_is_inf(ur))) { 2289 float_raise(float_flag_overflow, s); 2290 } else if (unlikely(fabsf(ur.h) <= FLT_MIN)) { 2291 ua = ua_orig; 2292 uc = uc_orig; 2293 goto soft; 2294 } 2295 } 2296 if (flags & float_muladd_negate_result) { 2297 return float32_chs(ur.s); 2298 } 2299 return ur.s; 2300 2301 soft: 2302 return soft_f32_muladd(ua.s, ub.s, uc.s, flags, s); 2303 } 2304 2305 float64 QEMU_FLATTEN 2306 float64_muladd(float64 xa, float64 xb, float64 xc, int flags, float_status *s) 2307 { 2308 union_float64 ua, ub, uc, ur; 2309 2310 ua.s = xa; 2311 ub.s = xb; 2312 uc.s = xc; 2313 2314 if (unlikely(!can_use_fpu(s))) { 2315 goto soft; 2316 } 2317 if (unlikely(flags & float_muladd_halve_result)) { 2318 goto soft; 2319 } 2320 2321 float64_input_flush3(&ua.s, &ub.s, &uc.s, s); 2322 if (unlikely(!f64_is_zon3(ua, ub, uc))) { 2323 goto soft; 2324 } 2325 2326 if (unlikely(force_soft_fma)) { 2327 goto soft; 2328 } 2329 2330 /* 2331 * When (a || b) == 0, there's no need to check for under/over flow, 2332 * since we know the addend is (normal || 0) and the product is 0. 2333 */ 2334 if (float64_is_zero(ua.s) || float64_is_zero(ub.s)) { 2335 union_float64 up; 2336 bool prod_sign; 2337 2338 prod_sign = float64_is_neg(ua.s) ^ float64_is_neg(ub.s); 2339 prod_sign ^= !!(flags & float_muladd_negate_product); 2340 up.s = float64_set_sign(float64_zero, prod_sign); 2341 2342 if (flags & float_muladd_negate_c) { 2343 uc.h = -uc.h; 2344 } 2345 ur.h = up.h + uc.h; 2346 } else { 2347 union_float64 ua_orig = ua; 2348 union_float64 uc_orig = uc; 2349 2350 if (flags & float_muladd_negate_product) { 2351 ua.h = -ua.h; 2352 } 2353 if (flags & float_muladd_negate_c) { 2354 uc.h = -uc.h; 2355 } 2356 2357 ur.h = fma(ua.h, ub.h, uc.h); 2358 2359 if (unlikely(f64_is_inf(ur))) { 2360 float_raise(float_flag_overflow, s); 2361 } else if (unlikely(fabs(ur.h) <= FLT_MIN)) { 2362 ua = ua_orig; 2363 uc = uc_orig; 2364 goto soft; 2365 } 2366 } 2367 if (flags & float_muladd_negate_result) { 2368 return float64_chs(ur.s); 2369 } 2370 return ur.s; 2371 2372 soft: 2373 return soft_f64_muladd(ua.s, ub.s, uc.s, flags, s); 2374 } 2375 2376 float64 float64r32_muladd(float64 a, float64 b, float64 c, 2377 int flags, float_status *status) 2378 { 2379 FloatParts64 pa, pb, pc, *pr; 2380 2381 float64_unpack_canonical(&pa, a, status); 2382 float64_unpack_canonical(&pb, b, status); 2383 float64_unpack_canonical(&pc, c, status); 2384 pr = parts_muladd(&pa, &pb, &pc, flags, status); 2385 2386 return float64r32_round_pack_canonical(pr, status); 2387 } 2388 2389 bfloat16 QEMU_FLATTEN bfloat16_muladd(bfloat16 a, bfloat16 b, bfloat16 c, 2390 int flags, float_status *status) 2391 { 2392 FloatParts64 pa, pb, pc, *pr; 2393 2394 bfloat16_unpack_canonical(&pa, a, status); 2395 bfloat16_unpack_canonical(&pb, b, status); 2396 bfloat16_unpack_canonical(&pc, c, status); 2397 pr = parts_muladd(&pa, &pb, &pc, flags, status); 2398 2399 return bfloat16_round_pack_canonical(pr, status); 2400 } 2401 2402 float128 QEMU_FLATTEN float128_muladd(float128 a, float128 b, float128 c, 2403 int flags, float_status *status) 2404 { 2405 FloatParts128 pa, pb, pc, *pr; 2406 2407 float128_unpack_canonical(&pa, a, status); 2408 float128_unpack_canonical(&pb, b, status); 2409 float128_unpack_canonical(&pc, c, status); 2410 pr = parts_muladd(&pa, &pb, &pc, flags, status); 2411 2412 return float128_round_pack_canonical(pr, status); 2413 } 2414 2415 /* 2416 * Division 2417 */ 2418 2419 float16 float16_div(float16 a, float16 b, float_status *status) 2420 { 2421 FloatParts64 pa, pb, *pr; 2422 2423 float16_unpack_canonical(&pa, a, status); 2424 float16_unpack_canonical(&pb, b, status); 2425 pr = parts_div(&pa, &pb, status); 2426 2427 return float16_round_pack_canonical(pr, status); 2428 } 2429 2430 static float32 QEMU_SOFTFLOAT_ATTR 2431 soft_f32_div(float32 a, float32 b, float_status *status) 2432 { 2433 FloatParts64 pa, pb, *pr; 2434 2435 float32_unpack_canonical(&pa, a, status); 2436 float32_unpack_canonical(&pb, b, status); 2437 pr = parts_div(&pa, &pb, status); 2438 2439 return float32_round_pack_canonical(pr, status); 2440 } 2441 2442 static float64 QEMU_SOFTFLOAT_ATTR 2443 soft_f64_div(float64 a, float64 b, float_status *status) 2444 { 2445 FloatParts64 pa, pb, *pr; 2446 2447 float64_unpack_canonical(&pa, a, status); 2448 float64_unpack_canonical(&pb, b, status); 2449 pr = parts_div(&pa, &pb, status); 2450 2451 return float64_round_pack_canonical(pr, status); 2452 } 2453 2454 static float hard_f32_div(float a, float b) 2455 { 2456 return a / b; 2457 } 2458 2459 static double hard_f64_div(double a, double b) 2460 { 2461 return a / b; 2462 } 2463 2464 static bool f32_div_pre(union_float32 a, union_float32 b) 2465 { 2466 if (QEMU_HARDFLOAT_2F32_USE_FP) { 2467 return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) && 2468 fpclassify(b.h) == FP_NORMAL; 2469 } 2470 return float32_is_zero_or_normal(a.s) && float32_is_normal(b.s); 2471 } 2472 2473 static bool f64_div_pre(union_float64 a, union_float64 b) 2474 { 2475 if (QEMU_HARDFLOAT_2F64_USE_FP) { 2476 return (fpclassify(a.h) == FP_NORMAL || fpclassify(a.h) == FP_ZERO) && 2477 fpclassify(b.h) == FP_NORMAL; 2478 } 2479 return float64_is_zero_or_normal(a.s) && float64_is_normal(b.s); 2480 } 2481 2482 static bool f32_div_post(union_float32 a, union_float32 b) 2483 { 2484 if (QEMU_HARDFLOAT_2F32_USE_FP) { 2485 return fpclassify(a.h) != FP_ZERO; 2486 } 2487 return !float32_is_zero(a.s); 2488 } 2489 2490 static bool f64_div_post(union_float64 a, union_float64 b) 2491 { 2492 if (QEMU_HARDFLOAT_2F64_USE_FP) { 2493 return fpclassify(a.h) != FP_ZERO; 2494 } 2495 return !float64_is_zero(a.s); 2496 } 2497 2498 float32 QEMU_FLATTEN 2499 float32_div(float32 a, float32 b, float_status *s) 2500 { 2501 return float32_gen2(a, b, s, hard_f32_div, soft_f32_div, 2502 f32_div_pre, f32_div_post); 2503 } 2504 2505 float64 QEMU_FLATTEN 2506 float64_div(float64 a, float64 b, float_status *s) 2507 { 2508 return float64_gen2(a, b, s, hard_f64_div, soft_f64_div, 2509 f64_div_pre, f64_div_post); 2510 } 2511 2512 float64 float64r32_div(float64 a, float64 b, float_status *status) 2513 { 2514 FloatParts64 pa, pb, *pr; 2515 2516 float64_unpack_canonical(&pa, a, status); 2517 float64_unpack_canonical(&pb, b, status); 2518 pr = parts_div(&pa, &pb, status); 2519 2520 return float64r32_round_pack_canonical(pr, status); 2521 } 2522 2523 bfloat16 QEMU_FLATTEN 2524 bfloat16_div(bfloat16 a, bfloat16 b, float_status *status) 2525 { 2526 FloatParts64 pa, pb, *pr; 2527 2528 bfloat16_unpack_canonical(&pa, a, status); 2529 bfloat16_unpack_canonical(&pb, b, status); 2530 pr = parts_div(&pa, &pb, status); 2531 2532 return bfloat16_round_pack_canonical(pr, status); 2533 } 2534 2535 float128 QEMU_FLATTEN 2536 float128_div(float128 a, float128 b, float_status *status) 2537 { 2538 FloatParts128 pa, pb, *pr; 2539 2540 float128_unpack_canonical(&pa, a, status); 2541 float128_unpack_canonical(&pb, b, status); 2542 pr = parts_div(&pa, &pb, status); 2543 2544 return float128_round_pack_canonical(pr, status); 2545 } 2546 2547 floatx80 floatx80_div(floatx80 a, floatx80 b, float_status *status) 2548 { 2549 FloatParts128 pa, pb, *pr; 2550 2551 if (!floatx80_unpack_canonical(&pa, a, status) || 2552 !floatx80_unpack_canonical(&pb, b, status)) { 2553 return floatx80_default_nan(status); 2554 } 2555 2556 pr = parts_div(&pa, &pb, status); 2557 return floatx80_round_pack_canonical(pr, status); 2558 } 2559 2560 /* 2561 * Remainder 2562 */ 2563 2564 float32 float32_rem(float32 a, float32 b, float_status *status) 2565 { 2566 FloatParts64 pa, pb, *pr; 2567 2568 float32_unpack_canonical(&pa, a, status); 2569 float32_unpack_canonical(&pb, b, status); 2570 pr = parts_modrem(&pa, &pb, NULL, status); 2571 2572 return float32_round_pack_canonical(pr, status); 2573 } 2574 2575 float64 float64_rem(float64 a, float64 b, float_status *status) 2576 { 2577 FloatParts64 pa, pb, *pr; 2578 2579 float64_unpack_canonical(&pa, a, status); 2580 float64_unpack_canonical(&pb, b, status); 2581 pr = parts_modrem(&pa, &pb, NULL, status); 2582 2583 return float64_round_pack_canonical(pr, status); 2584 } 2585 2586 float128 float128_rem(float128 a, float128 b, float_status *status) 2587 { 2588 FloatParts128 pa, pb, *pr; 2589 2590 float128_unpack_canonical(&pa, a, status); 2591 float128_unpack_canonical(&pb, b, status); 2592 pr = parts_modrem(&pa, &pb, NULL, status); 2593 2594 return float128_round_pack_canonical(pr, status); 2595 } 2596 2597 /* 2598 * Returns the remainder of the extended double-precision floating-point value 2599 * `a' with respect to the corresponding value `b'. 2600 * If 'mod' is false, the operation is performed according to the IEC/IEEE 2601 * Standard for Binary Floating-Point Arithmetic. If 'mod' is true, return 2602 * the remainder based on truncating the quotient toward zero instead and 2603 * *quotient is set to the low 64 bits of the absolute value of the integer 2604 * quotient. 2605 */ 2606 floatx80 floatx80_modrem(floatx80 a, floatx80 b, bool mod, 2607 uint64_t *quotient, float_status *status) 2608 { 2609 FloatParts128 pa, pb, *pr; 2610 2611 *quotient = 0; 2612 if (!floatx80_unpack_canonical(&pa, a, status) || 2613 !floatx80_unpack_canonical(&pb, b, status)) { 2614 return floatx80_default_nan(status); 2615 } 2616 pr = parts_modrem(&pa, &pb, mod ? quotient : NULL, status); 2617 2618 return floatx80_round_pack_canonical(pr, status); 2619 } 2620 2621 floatx80 floatx80_rem(floatx80 a, floatx80 b, float_status *status) 2622 { 2623 uint64_t quotient; 2624 return floatx80_modrem(a, b, false, "ient, status); 2625 } 2626 2627 floatx80 floatx80_mod(floatx80 a, floatx80 b, float_status *status) 2628 { 2629 uint64_t quotient; 2630 return floatx80_modrem(a, b, true, "ient, status); 2631 } 2632 2633 /* 2634 * Float to Float conversions 2635 * 2636 * Returns the result of converting one float format to another. The 2637 * conversion is performed according to the IEC/IEEE Standard for 2638 * Binary Floating-Point Arithmetic. 2639 * 2640 * Usually this only needs to take care of raising invalid exceptions 2641 * and handling the conversion on NaNs. 2642 */ 2643 2644 static void parts_float_to_ahp(FloatParts64 *a, float_status *s) 2645 { 2646 switch (a->cls) { 2647 case float_class_snan: 2648 float_raise(float_flag_invalid_snan, s); 2649 /* fall through */ 2650 case float_class_qnan: 2651 /* 2652 * There is no NaN in the destination format. Raise Invalid 2653 * and return a zero with the sign of the input NaN. 2654 */ 2655 float_raise(float_flag_invalid, s); 2656 a->cls = float_class_zero; 2657 break; 2658 2659 case float_class_inf: 2660 /* 2661 * There is no Inf in the destination format. Raise Invalid 2662 * and return the maximum normal with the correct sign. 2663 */ 2664 float_raise(float_flag_invalid, s); 2665 a->cls = float_class_normal; 2666 a->exp = float16_params_ahp.exp_max; 2667 a->frac = MAKE_64BIT_MASK(float16_params_ahp.frac_shift, 2668 float16_params_ahp.frac_size + 1); 2669 break; 2670 2671 case float_class_normal: 2672 case float_class_zero: 2673 break; 2674 2675 default: 2676 g_assert_not_reached(); 2677 } 2678 } 2679 2680 static void parts64_float_to_float(FloatParts64 *a, float_status *s) 2681 { 2682 if (is_nan(a->cls)) { 2683 parts_return_nan(a, s); 2684 } 2685 } 2686 2687 static void parts128_float_to_float(FloatParts128 *a, float_status *s) 2688 { 2689 if (is_nan(a->cls)) { 2690 parts_return_nan(a, s); 2691 } 2692 } 2693 2694 #define parts_float_to_float(P, S) \ 2695 PARTS_GENERIC_64_128(float_to_float, P)(P, S) 2696 2697 static void parts_float_to_float_narrow(FloatParts64 *a, FloatParts128 *b, 2698 float_status *s) 2699 { 2700 a->cls = b->cls; 2701 a->sign = b->sign; 2702 a->exp = b->exp; 2703 2704 if (a->cls == float_class_normal) { 2705 frac_truncjam(a, b); 2706 } else if (is_nan(a->cls)) { 2707 /* Discard the low bits of the NaN. */ 2708 a->frac = b->frac_hi; 2709 parts_return_nan(a, s); 2710 } 2711 } 2712 2713 static void parts_float_to_float_widen(FloatParts128 *a, FloatParts64 *b, 2714 float_status *s) 2715 { 2716 a->cls = b->cls; 2717 a->sign = b->sign; 2718 a->exp = b->exp; 2719 frac_widen(a, b); 2720 2721 if (is_nan(a->cls)) { 2722 parts_return_nan(a, s); 2723 } 2724 } 2725 2726 float32 float16_to_float32(float16 a, bool ieee, float_status *s) 2727 { 2728 const FloatFmt *fmt16 = ieee ? &float16_params : &float16_params_ahp; 2729 FloatParts64 p; 2730 2731 float16a_unpack_canonical(&p, a, s, fmt16); 2732 parts_float_to_float(&p, s); 2733 return float32_round_pack_canonical(&p, s); 2734 } 2735 2736 float64 float16_to_float64(float16 a, bool ieee, float_status *s) 2737 { 2738 const FloatFmt *fmt16 = ieee ? &float16_params : &float16_params_ahp; 2739 FloatParts64 p; 2740 2741 float16a_unpack_canonical(&p, a, s, fmt16); 2742 parts_float_to_float(&p, s); 2743 return float64_round_pack_canonical(&p, s); 2744 } 2745 2746 float16 float32_to_float16(float32 a, bool ieee, float_status *s) 2747 { 2748 FloatParts64 p; 2749 const FloatFmt *fmt; 2750 2751 float32_unpack_canonical(&p, a, s); 2752 if (ieee) { 2753 parts_float_to_float(&p, s); 2754 fmt = &float16_params; 2755 } else { 2756 parts_float_to_ahp(&p, s); 2757 fmt = &float16_params_ahp; 2758 } 2759 return float16a_round_pack_canonical(&p, s, fmt); 2760 } 2761 2762 static float64 QEMU_SOFTFLOAT_ATTR 2763 soft_float32_to_float64(float32 a, float_status *s) 2764 { 2765 FloatParts64 p; 2766 2767 float32_unpack_canonical(&p, a, s); 2768 parts_float_to_float(&p, s); 2769 return float64_round_pack_canonical(&p, s); 2770 } 2771 2772 float64 float32_to_float64(float32 a, float_status *s) 2773 { 2774 if (likely(float32_is_normal(a))) { 2775 /* Widening conversion can never produce inexact results. */ 2776 union_float32 uf; 2777 union_float64 ud; 2778 uf.s = a; 2779 ud.h = uf.h; 2780 return ud.s; 2781 } else if (float32_is_zero(a)) { 2782 return float64_set_sign(float64_zero, float32_is_neg(a)); 2783 } else { 2784 return soft_float32_to_float64(a, s); 2785 } 2786 } 2787 2788 float16 float64_to_float16(float64 a, bool ieee, float_status *s) 2789 { 2790 FloatParts64 p; 2791 const FloatFmt *fmt; 2792 2793 float64_unpack_canonical(&p, a, s); 2794 if (ieee) { 2795 parts_float_to_float(&p, s); 2796 fmt = &float16_params; 2797 } else { 2798 parts_float_to_ahp(&p, s); 2799 fmt = &float16_params_ahp; 2800 } 2801 return float16a_round_pack_canonical(&p, s, fmt); 2802 } 2803 2804 float32 float64_to_float32(float64 a, float_status *s) 2805 { 2806 FloatParts64 p; 2807 2808 float64_unpack_canonical(&p, a, s); 2809 parts_float_to_float(&p, s); 2810 return float32_round_pack_canonical(&p, s); 2811 } 2812 2813 float32 bfloat16_to_float32(bfloat16 a, float_status *s) 2814 { 2815 FloatParts64 p; 2816 2817 bfloat16_unpack_canonical(&p, a, s); 2818 parts_float_to_float(&p, s); 2819 return float32_round_pack_canonical(&p, s); 2820 } 2821 2822 float64 bfloat16_to_float64(bfloat16 a, float_status *s) 2823 { 2824 FloatParts64 p; 2825 2826 bfloat16_unpack_canonical(&p, a, s); 2827 parts_float_to_float(&p, s); 2828 return float64_round_pack_canonical(&p, s); 2829 } 2830 2831 bfloat16 float32_to_bfloat16(float32 a, float_status *s) 2832 { 2833 FloatParts64 p; 2834 2835 float32_unpack_canonical(&p, a, s); 2836 parts_float_to_float(&p, s); 2837 return bfloat16_round_pack_canonical(&p, s); 2838 } 2839 2840 bfloat16 float64_to_bfloat16(float64 a, float_status *s) 2841 { 2842 FloatParts64 p; 2843 2844 float64_unpack_canonical(&p, a, s); 2845 parts_float_to_float(&p, s); 2846 return bfloat16_round_pack_canonical(&p, s); 2847 } 2848 2849 float32 float128_to_float32(float128 a, float_status *s) 2850 { 2851 FloatParts64 p64; 2852 FloatParts128 p128; 2853 2854 float128_unpack_canonical(&p128, a, s); 2855 parts_float_to_float_narrow(&p64, &p128, s); 2856 return float32_round_pack_canonical(&p64, s); 2857 } 2858 2859 float64 float128_to_float64(float128 a, float_status *s) 2860 { 2861 FloatParts64 p64; 2862 FloatParts128 p128; 2863 2864 float128_unpack_canonical(&p128, a, s); 2865 parts_float_to_float_narrow(&p64, &p128, s); 2866 return float64_round_pack_canonical(&p64, s); 2867 } 2868 2869 float128 float32_to_float128(float32 a, float_status *s) 2870 { 2871 FloatParts64 p64; 2872 FloatParts128 p128; 2873 2874 float32_unpack_canonical(&p64, a, s); 2875 parts_float_to_float_widen(&p128, &p64, s); 2876 return float128_round_pack_canonical(&p128, s); 2877 } 2878 2879 float128 float64_to_float128(float64 a, float_status *s) 2880 { 2881 FloatParts64 p64; 2882 FloatParts128 p128; 2883 2884 float64_unpack_canonical(&p64, a, s); 2885 parts_float_to_float_widen(&p128, &p64, s); 2886 return float128_round_pack_canonical(&p128, s); 2887 } 2888 2889 float32 floatx80_to_float32(floatx80 a, float_status *s) 2890 { 2891 FloatParts64 p64; 2892 FloatParts128 p128; 2893 2894 if (floatx80_unpack_canonical(&p128, a, s)) { 2895 parts_float_to_float_narrow(&p64, &p128, s); 2896 } else { 2897 parts_default_nan(&p64, s); 2898 } 2899 return float32_round_pack_canonical(&p64, s); 2900 } 2901 2902 float64 floatx80_to_float64(floatx80 a, float_status *s) 2903 { 2904 FloatParts64 p64; 2905 FloatParts128 p128; 2906 2907 if (floatx80_unpack_canonical(&p128, a, s)) { 2908 parts_float_to_float_narrow(&p64, &p128, s); 2909 } else { 2910 parts_default_nan(&p64, s); 2911 } 2912 return float64_round_pack_canonical(&p64, s); 2913 } 2914 2915 float128 floatx80_to_float128(floatx80 a, float_status *s) 2916 { 2917 FloatParts128 p; 2918 2919 if (floatx80_unpack_canonical(&p, a, s)) { 2920 parts_float_to_float(&p, s); 2921 } else { 2922 parts_default_nan(&p, s); 2923 } 2924 return float128_round_pack_canonical(&p, s); 2925 } 2926 2927 floatx80 float32_to_floatx80(float32 a, float_status *s) 2928 { 2929 FloatParts64 p64; 2930 FloatParts128 p128; 2931 2932 float32_unpack_canonical(&p64, a, s); 2933 parts_float_to_float_widen(&p128, &p64, s); 2934 return floatx80_round_pack_canonical(&p128, s); 2935 } 2936 2937 floatx80 float64_to_floatx80(float64 a, float_status *s) 2938 { 2939 FloatParts64 p64; 2940 FloatParts128 p128; 2941 2942 float64_unpack_canonical(&p64, a, s); 2943 parts_float_to_float_widen(&p128, &p64, s); 2944 return floatx80_round_pack_canonical(&p128, s); 2945 } 2946 2947 floatx80 float128_to_floatx80(float128 a, float_status *s) 2948 { 2949 FloatParts128 p; 2950 2951 float128_unpack_canonical(&p, a, s); 2952 parts_float_to_float(&p, s); 2953 return floatx80_round_pack_canonical(&p, s); 2954 } 2955 2956 /* 2957 * Round to integral value 2958 */ 2959 2960 float16 float16_round_to_int(float16 a, float_status *s) 2961 { 2962 FloatParts64 p; 2963 2964 float16_unpack_canonical(&p, a, s); 2965 parts_round_to_int(&p, s->float_rounding_mode, 0, s, &float16_params); 2966 return float16_round_pack_canonical(&p, s); 2967 } 2968 2969 float32 float32_round_to_int(float32 a, float_status *s) 2970 { 2971 FloatParts64 p; 2972 2973 float32_unpack_canonical(&p, a, s); 2974 parts_round_to_int(&p, s->float_rounding_mode, 0, s, &float32_params); 2975 return float32_round_pack_canonical(&p, s); 2976 } 2977 2978 float64 float64_round_to_int(float64 a, float_status *s) 2979 { 2980 FloatParts64 p; 2981 2982 float64_unpack_canonical(&p, a, s); 2983 parts_round_to_int(&p, s->float_rounding_mode, 0, s, &float64_params); 2984 return float64_round_pack_canonical(&p, s); 2985 } 2986 2987 bfloat16 bfloat16_round_to_int(bfloat16 a, float_status *s) 2988 { 2989 FloatParts64 p; 2990 2991 bfloat16_unpack_canonical(&p, a, s); 2992 parts_round_to_int(&p, s->float_rounding_mode, 0, s, &bfloat16_params); 2993 return bfloat16_round_pack_canonical(&p, s); 2994 } 2995 2996 float128 float128_round_to_int(float128 a, float_status *s) 2997 { 2998 FloatParts128 p; 2999 3000 float128_unpack_canonical(&p, a, s); 3001 parts_round_to_int(&p, s->float_rounding_mode, 0, s, &float128_params); 3002 return float128_round_pack_canonical(&p, s); 3003 } 3004 3005 floatx80 floatx80_round_to_int(floatx80 a, float_status *status) 3006 { 3007 FloatParts128 p; 3008 3009 if (!floatx80_unpack_canonical(&p, a, status)) { 3010 return floatx80_default_nan(status); 3011 } 3012 3013 parts_round_to_int(&p, status->float_rounding_mode, 0, status, 3014 &floatx80_params[status->floatx80_rounding_precision]); 3015 return floatx80_round_pack_canonical(&p, status); 3016 } 3017 3018 /* 3019 * Floating-point to signed integer conversions 3020 */ 3021 3022 int8_t float16_to_int8_scalbn(float16 a, FloatRoundMode rmode, int scale, 3023 float_status *s) 3024 { 3025 FloatParts64 p; 3026 3027 float16_unpack_canonical(&p, a, s); 3028 return parts_float_to_sint(&p, rmode, scale, INT8_MIN, INT8_MAX, s); 3029 } 3030 3031 int16_t float16_to_int16_scalbn(float16 a, FloatRoundMode rmode, int scale, 3032 float_status *s) 3033 { 3034 FloatParts64 p; 3035 3036 float16_unpack_canonical(&p, a, s); 3037 return parts_float_to_sint(&p, rmode, scale, INT16_MIN, INT16_MAX, s); 3038 } 3039 3040 int32_t float16_to_int32_scalbn(float16 a, FloatRoundMode rmode, int scale, 3041 float_status *s) 3042 { 3043 FloatParts64 p; 3044 3045 float16_unpack_canonical(&p, a, s); 3046 return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s); 3047 } 3048 3049 int64_t float16_to_int64_scalbn(float16 a, FloatRoundMode rmode, int scale, 3050 float_status *s) 3051 { 3052 FloatParts64 p; 3053 3054 float16_unpack_canonical(&p, a, s); 3055 return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s); 3056 } 3057 3058 int16_t float32_to_int16_scalbn(float32 a, FloatRoundMode rmode, int scale, 3059 float_status *s) 3060 { 3061 FloatParts64 p; 3062 3063 float32_unpack_canonical(&p, a, s); 3064 return parts_float_to_sint(&p, rmode, scale, INT16_MIN, INT16_MAX, s); 3065 } 3066 3067 int32_t float32_to_int32_scalbn(float32 a, FloatRoundMode rmode, int scale, 3068 float_status *s) 3069 { 3070 FloatParts64 p; 3071 3072 float32_unpack_canonical(&p, a, s); 3073 return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s); 3074 } 3075 3076 int64_t float32_to_int64_scalbn(float32 a, FloatRoundMode rmode, int scale, 3077 float_status *s) 3078 { 3079 FloatParts64 p; 3080 3081 float32_unpack_canonical(&p, a, s); 3082 return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s); 3083 } 3084 3085 int16_t float64_to_int16_scalbn(float64 a, FloatRoundMode rmode, int scale, 3086 float_status *s) 3087 { 3088 FloatParts64 p; 3089 3090 float64_unpack_canonical(&p, a, s); 3091 return parts_float_to_sint(&p, rmode, scale, INT16_MIN, INT16_MAX, s); 3092 } 3093 3094 int32_t float64_to_int32_scalbn(float64 a, FloatRoundMode rmode, int scale, 3095 float_status *s) 3096 { 3097 FloatParts64 p; 3098 3099 float64_unpack_canonical(&p, a, s); 3100 return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s); 3101 } 3102 3103 int64_t float64_to_int64_scalbn(float64 a, FloatRoundMode rmode, int scale, 3104 float_status *s) 3105 { 3106 FloatParts64 p; 3107 3108 float64_unpack_canonical(&p, a, s); 3109 return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s); 3110 } 3111 3112 int16_t bfloat16_to_int16_scalbn(bfloat16 a, FloatRoundMode rmode, int scale, 3113 float_status *s) 3114 { 3115 FloatParts64 p; 3116 3117 bfloat16_unpack_canonical(&p, a, s); 3118 return parts_float_to_sint(&p, rmode, scale, INT16_MIN, INT16_MAX, s); 3119 } 3120 3121 int32_t bfloat16_to_int32_scalbn(bfloat16 a, FloatRoundMode rmode, int scale, 3122 float_status *s) 3123 { 3124 FloatParts64 p; 3125 3126 bfloat16_unpack_canonical(&p, a, s); 3127 return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s); 3128 } 3129 3130 int64_t bfloat16_to_int64_scalbn(bfloat16 a, FloatRoundMode rmode, int scale, 3131 float_status *s) 3132 { 3133 FloatParts64 p; 3134 3135 bfloat16_unpack_canonical(&p, a, s); 3136 return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s); 3137 } 3138 3139 static int32_t float128_to_int32_scalbn(float128 a, FloatRoundMode rmode, 3140 int scale, float_status *s) 3141 { 3142 FloatParts128 p; 3143 3144 float128_unpack_canonical(&p, a, s); 3145 return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s); 3146 } 3147 3148 static int64_t float128_to_int64_scalbn(float128 a, FloatRoundMode rmode, 3149 int scale, float_status *s) 3150 { 3151 FloatParts128 p; 3152 3153 float128_unpack_canonical(&p, a, s); 3154 return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s); 3155 } 3156 3157 static Int128 float128_to_int128_scalbn(float128 a, FloatRoundMode rmode, 3158 int scale, float_status *s) 3159 { 3160 int flags = 0; 3161 Int128 r; 3162 FloatParts128 p; 3163 3164 float128_unpack_canonical(&p, a, s); 3165 3166 switch (p.cls) { 3167 case float_class_snan: 3168 flags |= float_flag_invalid_snan; 3169 /* fall through */ 3170 case float_class_qnan: 3171 flags |= float_flag_invalid; 3172 r = UINT128_MAX; 3173 break; 3174 3175 case float_class_inf: 3176 flags = float_flag_invalid | float_flag_invalid_cvti; 3177 r = p.sign ? INT128_MIN : INT128_MAX; 3178 break; 3179 3180 case float_class_zero: 3181 return int128_zero(); 3182 3183 case float_class_normal: 3184 if (parts_round_to_int_normal(&p, rmode, scale, 128 - 2)) { 3185 flags = float_flag_inexact; 3186 } 3187 3188 if (p.exp < 127) { 3189 int shift = 127 - p.exp; 3190 r = int128_urshift(int128_make128(p.frac_lo, p.frac_hi), shift); 3191 if (p.sign) { 3192 r = int128_neg(r); 3193 } 3194 } else if (p.exp == 127 && p.sign && p.frac_lo == 0 && 3195 p.frac_hi == DECOMPOSED_IMPLICIT_BIT) { 3196 r = INT128_MIN; 3197 } else { 3198 flags = float_flag_invalid | float_flag_invalid_cvti; 3199 r = p.sign ? INT128_MIN : INT128_MAX; 3200 } 3201 break; 3202 3203 default: 3204 g_assert_not_reached(); 3205 } 3206 3207 float_raise(flags, s); 3208 return r; 3209 } 3210 3211 static int32_t floatx80_to_int32_scalbn(floatx80 a, FloatRoundMode rmode, 3212 int scale, float_status *s) 3213 { 3214 FloatParts128 p; 3215 3216 if (!floatx80_unpack_canonical(&p, a, s)) { 3217 parts_default_nan(&p, s); 3218 } 3219 return parts_float_to_sint(&p, rmode, scale, INT32_MIN, INT32_MAX, s); 3220 } 3221 3222 static int64_t floatx80_to_int64_scalbn(floatx80 a, FloatRoundMode rmode, 3223 int scale, float_status *s) 3224 { 3225 FloatParts128 p; 3226 3227 if (!floatx80_unpack_canonical(&p, a, s)) { 3228 parts_default_nan(&p, s); 3229 } 3230 return parts_float_to_sint(&p, rmode, scale, INT64_MIN, INT64_MAX, s); 3231 } 3232 3233 int8_t float16_to_int8(float16 a, float_status *s) 3234 { 3235 return float16_to_int8_scalbn(a, s->float_rounding_mode, 0, s); 3236 } 3237 3238 int16_t float16_to_int16(float16 a, float_status *s) 3239 { 3240 return float16_to_int16_scalbn(a, s->float_rounding_mode, 0, s); 3241 } 3242 3243 int32_t float16_to_int32(float16 a, float_status *s) 3244 { 3245 return float16_to_int32_scalbn(a, s->float_rounding_mode, 0, s); 3246 } 3247 3248 int64_t float16_to_int64(float16 a, float_status *s) 3249 { 3250 return float16_to_int64_scalbn(a, s->float_rounding_mode, 0, s); 3251 } 3252 3253 int16_t float32_to_int16(float32 a, float_status *s) 3254 { 3255 return float32_to_int16_scalbn(a, s->float_rounding_mode, 0, s); 3256 } 3257 3258 int32_t float32_to_int32(float32 a, float_status *s) 3259 { 3260 return float32_to_int32_scalbn(a, s->float_rounding_mode, 0, s); 3261 } 3262 3263 int64_t float32_to_int64(float32 a, float_status *s) 3264 { 3265 return float32_to_int64_scalbn(a, s->float_rounding_mode, 0, s); 3266 } 3267 3268 int16_t float64_to_int16(float64 a, float_status *s) 3269 { 3270 return float64_to_int16_scalbn(a, s->float_rounding_mode, 0, s); 3271 } 3272 3273 int32_t float64_to_int32(float64 a, float_status *s) 3274 { 3275 return float64_to_int32_scalbn(a, s->float_rounding_mode, 0, s); 3276 } 3277 3278 int64_t float64_to_int64(float64 a, float_status *s) 3279 { 3280 return float64_to_int64_scalbn(a, s->float_rounding_mode, 0, s); 3281 } 3282 3283 int32_t float128_to_int32(float128 a, float_status *s) 3284 { 3285 return float128_to_int32_scalbn(a, s->float_rounding_mode, 0, s); 3286 } 3287 3288 int64_t float128_to_int64(float128 a, float_status *s) 3289 { 3290 return float128_to_int64_scalbn(a, s->float_rounding_mode, 0, s); 3291 } 3292 3293 Int128 float128_to_int128(float128 a, float_status *s) 3294 { 3295 return float128_to_int128_scalbn(a, s->float_rounding_mode, 0, s); 3296 } 3297 3298 int32_t floatx80_to_int32(floatx80 a, float_status *s) 3299 { 3300 return floatx80_to_int32_scalbn(a, s->float_rounding_mode, 0, s); 3301 } 3302 3303 int64_t floatx80_to_int64(floatx80 a, float_status *s) 3304 { 3305 return floatx80_to_int64_scalbn(a, s->float_rounding_mode, 0, s); 3306 } 3307 3308 int16_t float16_to_int16_round_to_zero(float16 a, float_status *s) 3309 { 3310 return float16_to_int16_scalbn(a, float_round_to_zero, 0, s); 3311 } 3312 3313 int32_t float16_to_int32_round_to_zero(float16 a, float_status *s) 3314 { 3315 return float16_to_int32_scalbn(a, float_round_to_zero, 0, s); 3316 } 3317 3318 int64_t float16_to_int64_round_to_zero(float16 a, float_status *s) 3319 { 3320 return float16_to_int64_scalbn(a, float_round_to_zero, 0, s); 3321 } 3322 3323 int16_t float32_to_int16_round_to_zero(float32 a, float_status *s) 3324 { 3325 return float32_to_int16_scalbn(a, float_round_to_zero, 0, s); 3326 } 3327 3328 int32_t float32_to_int32_round_to_zero(float32 a, float_status *s) 3329 { 3330 return float32_to_int32_scalbn(a, float_round_to_zero, 0, s); 3331 } 3332 3333 int64_t float32_to_int64_round_to_zero(float32 a, float_status *s) 3334 { 3335 return float32_to_int64_scalbn(a, float_round_to_zero, 0, s); 3336 } 3337 3338 int16_t float64_to_int16_round_to_zero(float64 a, float_status *s) 3339 { 3340 return float64_to_int16_scalbn(a, float_round_to_zero, 0, s); 3341 } 3342 3343 int32_t float64_to_int32_round_to_zero(float64 a, float_status *s) 3344 { 3345 return float64_to_int32_scalbn(a, float_round_to_zero, 0, s); 3346 } 3347 3348 int64_t float64_to_int64_round_to_zero(float64 a, float_status *s) 3349 { 3350 return float64_to_int64_scalbn(a, float_round_to_zero, 0, s); 3351 } 3352 3353 int32_t float128_to_int32_round_to_zero(float128 a, float_status *s) 3354 { 3355 return float128_to_int32_scalbn(a, float_round_to_zero, 0, s); 3356 } 3357 3358 int64_t float128_to_int64_round_to_zero(float128 a, float_status *s) 3359 { 3360 return float128_to_int64_scalbn(a, float_round_to_zero, 0, s); 3361 } 3362 3363 Int128 float128_to_int128_round_to_zero(float128 a, float_status *s) 3364 { 3365 return float128_to_int128_scalbn(a, float_round_to_zero, 0, s); 3366 } 3367 3368 int32_t floatx80_to_int32_round_to_zero(floatx80 a, float_status *s) 3369 { 3370 return floatx80_to_int32_scalbn(a, float_round_to_zero, 0, s); 3371 } 3372 3373 int64_t floatx80_to_int64_round_to_zero(floatx80 a, float_status *s) 3374 { 3375 return floatx80_to_int64_scalbn(a, float_round_to_zero, 0, s); 3376 } 3377 3378 int16_t bfloat16_to_int16(bfloat16 a, float_status *s) 3379 { 3380 return bfloat16_to_int16_scalbn(a, s->float_rounding_mode, 0, s); 3381 } 3382 3383 int32_t bfloat16_to_int32(bfloat16 a, float_status *s) 3384 { 3385 return bfloat16_to_int32_scalbn(a, s->float_rounding_mode, 0, s); 3386 } 3387 3388 int64_t bfloat16_to_int64(bfloat16 a, float_status *s) 3389 { 3390 return bfloat16_to_int64_scalbn(a, s->float_rounding_mode, 0, s); 3391 } 3392 3393 int16_t bfloat16_to_int16_round_to_zero(bfloat16 a, float_status *s) 3394 { 3395 return bfloat16_to_int16_scalbn(a, float_round_to_zero, 0, s); 3396 } 3397 3398 int32_t bfloat16_to_int32_round_to_zero(bfloat16 a, float_status *s) 3399 { 3400 return bfloat16_to_int32_scalbn(a, float_round_to_zero, 0, s); 3401 } 3402 3403 int64_t bfloat16_to_int64_round_to_zero(bfloat16 a, float_status *s) 3404 { 3405 return bfloat16_to_int64_scalbn(a, float_round_to_zero, 0, s); 3406 } 3407 3408 /* 3409 * Floating-point to unsigned integer conversions 3410 */ 3411 3412 uint8_t float16_to_uint8_scalbn(float16 a, FloatRoundMode rmode, int scale, 3413 float_status *s) 3414 { 3415 FloatParts64 p; 3416 3417 float16_unpack_canonical(&p, a, s); 3418 return parts_float_to_uint(&p, rmode, scale, UINT8_MAX, s); 3419 } 3420 3421 uint16_t float16_to_uint16_scalbn(float16 a, FloatRoundMode rmode, int scale, 3422 float_status *s) 3423 { 3424 FloatParts64 p; 3425 3426 float16_unpack_canonical(&p, a, s); 3427 return parts_float_to_uint(&p, rmode, scale, UINT16_MAX, s); 3428 } 3429 3430 uint32_t float16_to_uint32_scalbn(float16 a, FloatRoundMode rmode, int scale, 3431 float_status *s) 3432 { 3433 FloatParts64 p; 3434 3435 float16_unpack_canonical(&p, a, s); 3436 return parts_float_to_uint(&p, rmode, scale, UINT32_MAX, s); 3437 } 3438 3439 uint64_t float16_to_uint64_scalbn(float16 a, FloatRoundMode rmode, int scale, 3440 float_status *s) 3441 { 3442 FloatParts64 p; 3443 3444 float16_unpack_canonical(&p, a, s); 3445 return parts_float_to_uint(&p, rmode, scale, UINT64_MAX, s); 3446 } 3447 3448 uint16_t float32_to_uint16_scalbn(float32 a, FloatRoundMode rmode, int scale, 3449 float_status *s) 3450 { 3451 FloatParts64 p; 3452 3453 float32_unpack_canonical(&p, a, s); 3454 return parts_float_to_uint(&p, rmode, scale, UINT16_MAX, s); 3455 } 3456 3457 uint32_t float32_to_uint32_scalbn(float32 a, FloatRoundMode rmode, int scale, 3458 float_status *s) 3459 { 3460 FloatParts64 p; 3461 3462 float32_unpack_canonical(&p, a, s); 3463 return parts_float_to_uint(&p, rmode, scale, UINT32_MAX, s); 3464 } 3465 3466 uint64_t float32_to_uint64_scalbn(float32 a, FloatRoundMode rmode, int scale, 3467 float_status *s) 3468 { 3469 FloatParts64 p; 3470 3471 float32_unpack_canonical(&p, a, s); 3472 return parts_float_to_uint(&p, rmode, scale, UINT64_MAX, s); 3473 } 3474 3475 uint16_t float64_to_uint16_scalbn(float64 a, FloatRoundMode rmode, int scale, 3476 float_status *s) 3477 { 3478 FloatParts64 p; 3479 3480 float64_unpack_canonical(&p, a, s); 3481 return parts_float_to_uint(&p, rmode, scale, UINT16_MAX, s); 3482 } 3483 3484 uint32_t float64_to_uint32_scalbn(float64 a, FloatRoundMode rmode, int scale, 3485 float_status *s) 3486 { 3487 FloatParts64 p; 3488 3489 float64_unpack_canonical(&p, a, s); 3490 return parts_float_to_uint(&p, rmode, scale, UINT32_MAX, s); 3491 } 3492 3493 uint64_t float64_to_uint64_scalbn(float64 a, FloatRoundMode rmode, int scale, 3494 float_status *s) 3495 { 3496 FloatParts64 p; 3497 3498 float64_unpack_canonical(&p, a, s); 3499 return parts_float_to_uint(&p, rmode, scale, UINT64_MAX, s); 3500 } 3501 3502 uint16_t bfloat16_to_uint16_scalbn(bfloat16 a, FloatRoundMode rmode, 3503 int scale, float_status *s) 3504 { 3505 FloatParts64 p; 3506 3507 bfloat16_unpack_canonical(&p, a, s); 3508 return parts_float_to_uint(&p, rmode, scale, UINT16_MAX, s); 3509 } 3510 3511 uint32_t bfloat16_to_uint32_scalbn(bfloat16 a, FloatRoundMode rmode, 3512 int scale, float_status *s) 3513 { 3514 FloatParts64 p; 3515 3516 bfloat16_unpack_canonical(&p, a, s); 3517 return parts_float_to_uint(&p, rmode, scale, UINT32_MAX, s); 3518 } 3519 3520 uint64_t bfloat16_to_uint64_scalbn(bfloat16 a, FloatRoundMode rmode, 3521 int scale, float_status *s) 3522 { 3523 FloatParts64 p; 3524 3525 bfloat16_unpack_canonical(&p, a, s); 3526 return parts_float_to_uint(&p, rmode, scale, UINT64_MAX, s); 3527 } 3528 3529 static uint32_t float128_to_uint32_scalbn(float128 a, FloatRoundMode rmode, 3530 int scale, float_status *s) 3531 { 3532 FloatParts128 p; 3533 3534 float128_unpack_canonical(&p, a, s); 3535 return parts_float_to_uint(&p, rmode, scale, UINT32_MAX, s); 3536 } 3537 3538 static uint64_t float128_to_uint64_scalbn(float128 a, FloatRoundMode rmode, 3539 int scale, float_status *s) 3540 { 3541 FloatParts128 p; 3542 3543 float128_unpack_canonical(&p, a, s); 3544 return parts_float_to_uint(&p, rmode, scale, UINT64_MAX, s); 3545 } 3546 3547 static Int128 float128_to_uint128_scalbn(float128 a, FloatRoundMode rmode, 3548 int scale, float_status *s) 3549 { 3550 int flags = 0; 3551 Int128 r; 3552 FloatParts128 p; 3553 3554 float128_unpack_canonical(&p, a, s); 3555 3556 switch (p.cls) { 3557 case float_class_snan: 3558 flags |= float_flag_invalid_snan; 3559 /* fall through */ 3560 case float_class_qnan: 3561 flags |= float_flag_invalid; 3562 r = UINT128_MAX; 3563 break; 3564 3565 case float_class_inf: 3566 flags = float_flag_invalid | float_flag_invalid_cvti; 3567 r = p.sign ? int128_zero() : UINT128_MAX; 3568 break; 3569 3570 case float_class_zero: 3571 return int128_zero(); 3572 3573 case float_class_normal: 3574 if (parts_round_to_int_normal(&p, rmode, scale, 128 - 2)) { 3575 flags = float_flag_inexact; 3576 if (p.cls == float_class_zero) { 3577 r = int128_zero(); 3578 break; 3579 } 3580 } 3581 3582 if (p.sign) { 3583 flags = float_flag_invalid | float_flag_invalid_cvti; 3584 r = int128_zero(); 3585 } else if (p.exp <= 127) { 3586 int shift = 127 - p.exp; 3587 r = int128_urshift(int128_make128(p.frac_lo, p.frac_hi), shift); 3588 } else { 3589 flags = float_flag_invalid | float_flag_invalid_cvti; 3590 r = UINT128_MAX; 3591 } 3592 break; 3593 3594 default: 3595 g_assert_not_reached(); 3596 } 3597 3598 float_raise(flags, s); 3599 return r; 3600 } 3601 3602 uint8_t float16_to_uint8(float16 a, float_status *s) 3603 { 3604 return float16_to_uint8_scalbn(a, s->float_rounding_mode, 0, s); 3605 } 3606 3607 uint16_t float16_to_uint16(float16 a, float_status *s) 3608 { 3609 return float16_to_uint16_scalbn(a, s->float_rounding_mode, 0, s); 3610 } 3611 3612 uint32_t float16_to_uint32(float16 a, float_status *s) 3613 { 3614 return float16_to_uint32_scalbn(a, s->float_rounding_mode, 0, s); 3615 } 3616 3617 uint64_t float16_to_uint64(float16 a, float_status *s) 3618 { 3619 return float16_to_uint64_scalbn(a, s->float_rounding_mode, 0, s); 3620 } 3621 3622 uint16_t float32_to_uint16(float32 a, float_status *s) 3623 { 3624 return float32_to_uint16_scalbn(a, s->float_rounding_mode, 0, s); 3625 } 3626 3627 uint32_t float32_to_uint32(float32 a, float_status *s) 3628 { 3629 return float32_to_uint32_scalbn(a, s->float_rounding_mode, 0, s); 3630 } 3631 3632 uint64_t float32_to_uint64(float32 a, float_status *s) 3633 { 3634 return float32_to_uint64_scalbn(a, s->float_rounding_mode, 0, s); 3635 } 3636 3637 uint16_t float64_to_uint16(float64 a, float_status *s) 3638 { 3639 return float64_to_uint16_scalbn(a, s->float_rounding_mode, 0, s); 3640 } 3641 3642 uint32_t float64_to_uint32(float64 a, float_status *s) 3643 { 3644 return float64_to_uint32_scalbn(a, s->float_rounding_mode, 0, s); 3645 } 3646 3647 uint64_t float64_to_uint64(float64 a, float_status *s) 3648 { 3649 return float64_to_uint64_scalbn(a, s->float_rounding_mode, 0, s); 3650 } 3651 3652 uint32_t float128_to_uint32(float128 a, float_status *s) 3653 { 3654 return float128_to_uint32_scalbn(a, s->float_rounding_mode, 0, s); 3655 } 3656 3657 uint64_t float128_to_uint64(float128 a, float_status *s) 3658 { 3659 return float128_to_uint64_scalbn(a, s->float_rounding_mode, 0, s); 3660 } 3661 3662 Int128 float128_to_uint128(float128 a, float_status *s) 3663 { 3664 return float128_to_uint128_scalbn(a, s->float_rounding_mode, 0, s); 3665 } 3666 3667 uint16_t float16_to_uint16_round_to_zero(float16 a, float_status *s) 3668 { 3669 return float16_to_uint16_scalbn(a, float_round_to_zero, 0, s); 3670 } 3671 3672 uint32_t float16_to_uint32_round_to_zero(float16 a, float_status *s) 3673 { 3674 return float16_to_uint32_scalbn(a, float_round_to_zero, 0, s); 3675 } 3676 3677 uint64_t float16_to_uint64_round_to_zero(float16 a, float_status *s) 3678 { 3679 return float16_to_uint64_scalbn(a, float_round_to_zero, 0, s); 3680 } 3681 3682 uint16_t float32_to_uint16_round_to_zero(float32 a, float_status *s) 3683 { 3684 return float32_to_uint16_scalbn(a, float_round_to_zero, 0, s); 3685 } 3686 3687 uint32_t float32_to_uint32_round_to_zero(float32 a, float_status *s) 3688 { 3689 return float32_to_uint32_scalbn(a, float_round_to_zero, 0, s); 3690 } 3691 3692 uint64_t float32_to_uint64_round_to_zero(float32 a, float_status *s) 3693 { 3694 return float32_to_uint64_scalbn(a, float_round_to_zero, 0, s); 3695 } 3696 3697 uint16_t float64_to_uint16_round_to_zero(float64 a, float_status *s) 3698 { 3699 return float64_to_uint16_scalbn(a, float_round_to_zero, 0, s); 3700 } 3701 3702 uint32_t float64_to_uint32_round_to_zero(float64 a, float_status *s) 3703 { 3704 return float64_to_uint32_scalbn(a, float_round_to_zero, 0, s); 3705 } 3706 3707 uint64_t float64_to_uint64_round_to_zero(float64 a, float_status *s) 3708 { 3709 return float64_to_uint64_scalbn(a, float_round_to_zero, 0, s); 3710 } 3711 3712 uint32_t float128_to_uint32_round_to_zero(float128 a, float_status *s) 3713 { 3714 return float128_to_uint32_scalbn(a, float_round_to_zero, 0, s); 3715 } 3716 3717 uint64_t float128_to_uint64_round_to_zero(float128 a, float_status *s) 3718 { 3719 return float128_to_uint64_scalbn(a, float_round_to_zero, 0, s); 3720 } 3721 3722 Int128 float128_to_uint128_round_to_zero(float128 a, float_status *s) 3723 { 3724 return float128_to_uint128_scalbn(a, float_round_to_zero, 0, s); 3725 } 3726 3727 uint16_t bfloat16_to_uint16(bfloat16 a, float_status *s) 3728 { 3729 return bfloat16_to_uint16_scalbn(a, s->float_rounding_mode, 0, s); 3730 } 3731 3732 uint32_t bfloat16_to_uint32(bfloat16 a, float_status *s) 3733 { 3734 return bfloat16_to_uint32_scalbn(a, s->float_rounding_mode, 0, s); 3735 } 3736 3737 uint64_t bfloat16_to_uint64(bfloat16 a, float_status *s) 3738 { 3739 return bfloat16_to_uint64_scalbn(a, s->float_rounding_mode, 0, s); 3740 } 3741 3742 uint16_t bfloat16_to_uint16_round_to_zero(bfloat16 a, float_status *s) 3743 { 3744 return bfloat16_to_uint16_scalbn(a, float_round_to_zero, 0, s); 3745 } 3746 3747 uint32_t bfloat16_to_uint32_round_to_zero(bfloat16 a, float_status *s) 3748 { 3749 return bfloat16_to_uint32_scalbn(a, float_round_to_zero, 0, s); 3750 } 3751 3752 uint64_t bfloat16_to_uint64_round_to_zero(bfloat16 a, float_status *s) 3753 { 3754 return bfloat16_to_uint64_scalbn(a, float_round_to_zero, 0, s); 3755 } 3756 3757 /* 3758 * Signed integer to floating-point conversions 3759 */ 3760 3761 float16 int64_to_float16_scalbn(int64_t a, int scale, float_status *status) 3762 { 3763 FloatParts64 p; 3764 3765 parts_sint_to_float(&p, a, scale, status); 3766 return float16_round_pack_canonical(&p, status); 3767 } 3768 3769 float16 int32_to_float16_scalbn(int32_t a, int scale, float_status *status) 3770 { 3771 return int64_to_float16_scalbn(a, scale, status); 3772 } 3773 3774 float16 int16_to_float16_scalbn(int16_t a, int scale, float_status *status) 3775 { 3776 return int64_to_float16_scalbn(a, scale, status); 3777 } 3778 3779 float16 int64_to_float16(int64_t a, float_status *status) 3780 { 3781 return int64_to_float16_scalbn(a, 0, status); 3782 } 3783 3784 float16 int32_to_float16(int32_t a, float_status *status) 3785 { 3786 return int64_to_float16_scalbn(a, 0, status); 3787 } 3788 3789 float16 int16_to_float16(int16_t a, float_status *status) 3790 { 3791 return int64_to_float16_scalbn(a, 0, status); 3792 } 3793 3794 float16 int8_to_float16(int8_t a, float_status *status) 3795 { 3796 return int64_to_float16_scalbn(a, 0, status); 3797 } 3798 3799 float32 int64_to_float32_scalbn(int64_t a, int scale, float_status *status) 3800 { 3801 FloatParts64 p; 3802 3803 /* Without scaling, there are no overflow concerns. */ 3804 if (likely(scale == 0) && can_use_fpu(status)) { 3805 union_float32 ur; 3806 ur.h = a; 3807 return ur.s; 3808 } 3809 3810 parts64_sint_to_float(&p, a, scale, status); 3811 return float32_round_pack_canonical(&p, status); 3812 } 3813 3814 float32 int32_to_float32_scalbn(int32_t a, int scale, float_status *status) 3815 { 3816 return int64_to_float32_scalbn(a, scale, status); 3817 } 3818 3819 float32 int16_to_float32_scalbn(int16_t a, int scale, float_status *status) 3820 { 3821 return int64_to_float32_scalbn(a, scale, status); 3822 } 3823 3824 float32 int64_to_float32(int64_t a, float_status *status) 3825 { 3826 return int64_to_float32_scalbn(a, 0, status); 3827 } 3828 3829 float32 int32_to_float32(int32_t a, float_status *status) 3830 { 3831 return int64_to_float32_scalbn(a, 0, status); 3832 } 3833 3834 float32 int16_to_float32(int16_t a, float_status *status) 3835 { 3836 return int64_to_float32_scalbn(a, 0, status); 3837 } 3838 3839 float64 int64_to_float64_scalbn(int64_t a, int scale, float_status *status) 3840 { 3841 FloatParts64 p; 3842 3843 /* Without scaling, there are no overflow concerns. */ 3844 if (likely(scale == 0) && can_use_fpu(status)) { 3845 union_float64 ur; 3846 ur.h = a; 3847 return ur.s; 3848 } 3849 3850 parts_sint_to_float(&p, a, scale, status); 3851 return float64_round_pack_canonical(&p, status); 3852 } 3853 3854 float64 int32_to_float64_scalbn(int32_t a, int scale, float_status *status) 3855 { 3856 return int64_to_float64_scalbn(a, scale, status); 3857 } 3858 3859 float64 int16_to_float64_scalbn(int16_t a, int scale, float_status *status) 3860 { 3861 return int64_to_float64_scalbn(a, scale, status); 3862 } 3863 3864 float64 int64_to_float64(int64_t a, float_status *status) 3865 { 3866 return int64_to_float64_scalbn(a, 0, status); 3867 } 3868 3869 float64 int32_to_float64(int32_t a, float_status *status) 3870 { 3871 return int64_to_float64_scalbn(a, 0, status); 3872 } 3873 3874 float64 int16_to_float64(int16_t a, float_status *status) 3875 { 3876 return int64_to_float64_scalbn(a, 0, status); 3877 } 3878 3879 bfloat16 int64_to_bfloat16_scalbn(int64_t a, int scale, float_status *status) 3880 { 3881 FloatParts64 p; 3882 3883 parts_sint_to_float(&p, a, scale, status); 3884 return bfloat16_round_pack_canonical(&p, status); 3885 } 3886 3887 bfloat16 int32_to_bfloat16_scalbn(int32_t a, int scale, float_status *status) 3888 { 3889 return int64_to_bfloat16_scalbn(a, scale, status); 3890 } 3891 3892 bfloat16 int16_to_bfloat16_scalbn(int16_t a, int scale, float_status *status) 3893 { 3894 return int64_to_bfloat16_scalbn(a, scale, status); 3895 } 3896 3897 bfloat16 int64_to_bfloat16(int64_t a, float_status *status) 3898 { 3899 return int64_to_bfloat16_scalbn(a, 0, status); 3900 } 3901 3902 bfloat16 int32_to_bfloat16(int32_t a, float_status *status) 3903 { 3904 return int64_to_bfloat16_scalbn(a, 0, status); 3905 } 3906 3907 bfloat16 int16_to_bfloat16(int16_t a, float_status *status) 3908 { 3909 return int64_to_bfloat16_scalbn(a, 0, status); 3910 } 3911 3912 float128 int128_to_float128(Int128 a, float_status *status) 3913 { 3914 FloatParts128 p = { }; 3915 int shift; 3916 3917 if (int128_nz(a)) { 3918 p.cls = float_class_normal; 3919 if (!int128_nonneg(a)) { 3920 p.sign = true; 3921 a = int128_neg(a); 3922 } 3923 3924 shift = clz64(int128_gethi(a)); 3925 if (shift == 64) { 3926 shift += clz64(int128_getlo(a)); 3927 } 3928 3929 p.exp = 127 - shift; 3930 a = int128_lshift(a, shift); 3931 3932 p.frac_hi = int128_gethi(a); 3933 p.frac_lo = int128_getlo(a); 3934 } else { 3935 p.cls = float_class_zero; 3936 } 3937 3938 return float128_round_pack_canonical(&p, status); 3939 } 3940 3941 float128 int64_to_float128(int64_t a, float_status *status) 3942 { 3943 FloatParts128 p; 3944 3945 parts_sint_to_float(&p, a, 0, status); 3946 return float128_round_pack_canonical(&p, status); 3947 } 3948 3949 float128 int32_to_float128(int32_t a, float_status *status) 3950 { 3951 return int64_to_float128(a, status); 3952 } 3953 3954 floatx80 int64_to_floatx80(int64_t a, float_status *status) 3955 { 3956 FloatParts128 p; 3957 3958 parts_sint_to_float(&p, a, 0, status); 3959 return floatx80_round_pack_canonical(&p, status); 3960 } 3961 3962 floatx80 int32_to_floatx80(int32_t a, float_status *status) 3963 { 3964 return int64_to_floatx80(a, status); 3965 } 3966 3967 /* 3968 * Unsigned Integer to floating-point conversions 3969 */ 3970 3971 float16 uint64_to_float16_scalbn(uint64_t a, int scale, float_status *status) 3972 { 3973 FloatParts64 p; 3974 3975 parts_uint_to_float(&p, a, scale, status); 3976 return float16_round_pack_canonical(&p, status); 3977 } 3978 3979 float16 uint32_to_float16_scalbn(uint32_t a, int scale, float_status *status) 3980 { 3981 return uint64_to_float16_scalbn(a, scale, status); 3982 } 3983 3984 float16 uint16_to_float16_scalbn(uint16_t a, int scale, float_status *status) 3985 { 3986 return uint64_to_float16_scalbn(a, scale, status); 3987 } 3988 3989 float16 uint64_to_float16(uint64_t a, float_status *status) 3990 { 3991 return uint64_to_float16_scalbn(a, 0, status); 3992 } 3993 3994 float16 uint32_to_float16(uint32_t a, float_status *status) 3995 { 3996 return uint64_to_float16_scalbn(a, 0, status); 3997 } 3998 3999 float16 uint16_to_float16(uint16_t a, float_status *status) 4000 { 4001 return uint64_to_float16_scalbn(a, 0, status); 4002 } 4003 4004 float16 uint8_to_float16(uint8_t a, float_status *status) 4005 { 4006 return uint64_to_float16_scalbn(a, 0, status); 4007 } 4008 4009 float32 uint64_to_float32_scalbn(uint64_t a, int scale, float_status *status) 4010 { 4011 FloatParts64 p; 4012 4013 /* Without scaling, there are no overflow concerns. */ 4014 if (likely(scale == 0) && can_use_fpu(status)) { 4015 union_float32 ur; 4016 ur.h = a; 4017 return ur.s; 4018 } 4019 4020 parts_uint_to_float(&p, a, scale, status); 4021 return float32_round_pack_canonical(&p, status); 4022 } 4023 4024 float32 uint32_to_float32_scalbn(uint32_t a, int scale, float_status *status) 4025 { 4026 return uint64_to_float32_scalbn(a, scale, status); 4027 } 4028 4029 float32 uint16_to_float32_scalbn(uint16_t a, int scale, float_status *status) 4030 { 4031 return uint64_to_float32_scalbn(a, scale, status); 4032 } 4033 4034 float32 uint64_to_float32(uint64_t a, float_status *status) 4035 { 4036 return uint64_to_float32_scalbn(a, 0, status); 4037 } 4038 4039 float32 uint32_to_float32(uint32_t a, float_status *status) 4040 { 4041 return uint64_to_float32_scalbn(a, 0, status); 4042 } 4043 4044 float32 uint16_to_float32(uint16_t a, float_status *status) 4045 { 4046 return uint64_to_float32_scalbn(a, 0, status); 4047 } 4048 4049 float64 uint64_to_float64_scalbn(uint64_t a, int scale, float_status *status) 4050 { 4051 FloatParts64 p; 4052 4053 /* Without scaling, there are no overflow concerns. */ 4054 if (likely(scale == 0) && can_use_fpu(status)) { 4055 union_float64 ur; 4056 ur.h = a; 4057 return ur.s; 4058 } 4059 4060 parts_uint_to_float(&p, a, scale, status); 4061 return float64_round_pack_canonical(&p, status); 4062 } 4063 4064 float64 uint32_to_float64_scalbn(uint32_t a, int scale, float_status *status) 4065 { 4066 return uint64_to_float64_scalbn(a, scale, status); 4067 } 4068 4069 float64 uint16_to_float64_scalbn(uint16_t a, int scale, float_status *status) 4070 { 4071 return uint64_to_float64_scalbn(a, scale, status); 4072 } 4073 4074 float64 uint64_to_float64(uint64_t a, float_status *status) 4075 { 4076 return uint64_to_float64_scalbn(a, 0, status); 4077 } 4078 4079 float64 uint32_to_float64(uint32_t a, float_status *status) 4080 { 4081 return uint64_to_float64_scalbn(a, 0, status); 4082 } 4083 4084 float64 uint16_to_float64(uint16_t a, float_status *status) 4085 { 4086 return uint64_to_float64_scalbn(a, 0, status); 4087 } 4088 4089 bfloat16 uint64_to_bfloat16_scalbn(uint64_t a, int scale, float_status *status) 4090 { 4091 FloatParts64 p; 4092 4093 parts_uint_to_float(&p, a, scale, status); 4094 return bfloat16_round_pack_canonical(&p, status); 4095 } 4096 4097 bfloat16 uint32_to_bfloat16_scalbn(uint32_t a, int scale, float_status *status) 4098 { 4099 return uint64_to_bfloat16_scalbn(a, scale, status); 4100 } 4101 4102 bfloat16 uint16_to_bfloat16_scalbn(uint16_t a, int scale, float_status *status) 4103 { 4104 return uint64_to_bfloat16_scalbn(a, scale, status); 4105 } 4106 4107 bfloat16 uint64_to_bfloat16(uint64_t a, float_status *status) 4108 { 4109 return uint64_to_bfloat16_scalbn(a, 0, status); 4110 } 4111 4112 bfloat16 uint32_to_bfloat16(uint32_t a, float_status *status) 4113 { 4114 return uint64_to_bfloat16_scalbn(a, 0, status); 4115 } 4116 4117 bfloat16 uint16_to_bfloat16(uint16_t a, float_status *status) 4118 { 4119 return uint64_to_bfloat16_scalbn(a, 0, status); 4120 } 4121 4122 float128 uint64_to_float128(uint64_t a, float_status *status) 4123 { 4124 FloatParts128 p; 4125 4126 parts_uint_to_float(&p, a, 0, status); 4127 return float128_round_pack_canonical(&p, status); 4128 } 4129 4130 float128 uint128_to_float128(Int128 a, float_status *status) 4131 { 4132 FloatParts128 p = { }; 4133 int shift; 4134 4135 if (int128_nz(a)) { 4136 p.cls = float_class_normal; 4137 4138 shift = clz64(int128_gethi(a)); 4139 if (shift == 64) { 4140 shift += clz64(int128_getlo(a)); 4141 } 4142 4143 p.exp = 127 - shift; 4144 a = int128_lshift(a, shift); 4145 4146 p.frac_hi = int128_gethi(a); 4147 p.frac_lo = int128_getlo(a); 4148 } else { 4149 p.cls = float_class_zero; 4150 } 4151 4152 return float128_round_pack_canonical(&p, status); 4153 } 4154 4155 /* 4156 * Minimum and maximum 4157 */ 4158 4159 static float16 float16_minmax(float16 a, float16 b, float_status *s, int flags) 4160 { 4161 FloatParts64 pa, pb, *pr; 4162 4163 float16_unpack_canonical(&pa, a, s); 4164 float16_unpack_canonical(&pb, b, s); 4165 pr = parts_minmax(&pa, &pb, s, flags); 4166 4167 return float16_round_pack_canonical(pr, s); 4168 } 4169 4170 static bfloat16 bfloat16_minmax(bfloat16 a, bfloat16 b, 4171 float_status *s, int flags) 4172 { 4173 FloatParts64 pa, pb, *pr; 4174 4175 bfloat16_unpack_canonical(&pa, a, s); 4176 bfloat16_unpack_canonical(&pb, b, s); 4177 pr = parts_minmax(&pa, &pb, s, flags); 4178 4179 return bfloat16_round_pack_canonical(pr, s); 4180 } 4181 4182 static float32 float32_minmax(float32 a, float32 b, float_status *s, int flags) 4183 { 4184 FloatParts64 pa, pb, *pr; 4185 4186 float32_unpack_canonical(&pa, a, s); 4187 float32_unpack_canonical(&pb, b, s); 4188 pr = parts_minmax(&pa, &pb, s, flags); 4189 4190 return float32_round_pack_canonical(pr, s); 4191 } 4192 4193 static float64 float64_minmax(float64 a, float64 b, float_status *s, int flags) 4194 { 4195 FloatParts64 pa, pb, *pr; 4196 4197 float64_unpack_canonical(&pa, a, s); 4198 float64_unpack_canonical(&pb, b, s); 4199 pr = parts_minmax(&pa, &pb, s, flags); 4200 4201 return float64_round_pack_canonical(pr, s); 4202 } 4203 4204 static float128 float128_minmax(float128 a, float128 b, 4205 float_status *s, int flags) 4206 { 4207 FloatParts128 pa, pb, *pr; 4208 4209 float128_unpack_canonical(&pa, a, s); 4210 float128_unpack_canonical(&pb, b, s); 4211 pr = parts_minmax(&pa, &pb, s, flags); 4212 4213 return float128_round_pack_canonical(pr, s); 4214 } 4215 4216 #define MINMAX_1(type, name, flags) \ 4217 type type##_##name(type a, type b, float_status *s) \ 4218 { return type##_minmax(a, b, s, flags); } 4219 4220 #define MINMAX_2(type) \ 4221 MINMAX_1(type, max, 0) \ 4222 MINMAX_1(type, maxnum, minmax_isnum) \ 4223 MINMAX_1(type, maxnummag, minmax_isnum | minmax_ismag) \ 4224 MINMAX_1(type, maximum_number, minmax_isnumber) \ 4225 MINMAX_1(type, min, minmax_ismin) \ 4226 MINMAX_1(type, minnum, minmax_ismin | minmax_isnum) \ 4227 MINMAX_1(type, minnummag, minmax_ismin | minmax_isnum | minmax_ismag) \ 4228 MINMAX_1(type, minimum_number, minmax_ismin | minmax_isnumber) \ 4229 4230 MINMAX_2(float16) 4231 MINMAX_2(bfloat16) 4232 MINMAX_2(float32) 4233 MINMAX_2(float64) 4234 MINMAX_2(float128) 4235 4236 #undef MINMAX_1 4237 #undef MINMAX_2 4238 4239 /* 4240 * Floating point compare 4241 */ 4242 4243 static FloatRelation QEMU_FLATTEN 4244 float16_do_compare(float16 a, float16 b, float_status *s, bool is_quiet) 4245 { 4246 FloatParts64 pa, pb; 4247 4248 float16_unpack_canonical(&pa, a, s); 4249 float16_unpack_canonical(&pb, b, s); 4250 return parts_compare(&pa, &pb, s, is_quiet); 4251 } 4252 4253 FloatRelation float16_compare(float16 a, float16 b, float_status *s) 4254 { 4255 return float16_do_compare(a, b, s, false); 4256 } 4257 4258 FloatRelation float16_compare_quiet(float16 a, float16 b, float_status *s) 4259 { 4260 return float16_do_compare(a, b, s, true); 4261 } 4262 4263 static FloatRelation QEMU_SOFTFLOAT_ATTR 4264 float32_do_compare(float32 a, float32 b, float_status *s, bool is_quiet) 4265 { 4266 FloatParts64 pa, pb; 4267 4268 float32_unpack_canonical(&pa, a, s); 4269 float32_unpack_canonical(&pb, b, s); 4270 return parts_compare(&pa, &pb, s, is_quiet); 4271 } 4272 4273 static FloatRelation QEMU_FLATTEN 4274 float32_hs_compare(float32 xa, float32 xb, float_status *s, bool is_quiet) 4275 { 4276 union_float32 ua, ub; 4277 4278 ua.s = xa; 4279 ub.s = xb; 4280 4281 if (QEMU_NO_HARDFLOAT) { 4282 goto soft; 4283 } 4284 4285 float32_input_flush2(&ua.s, &ub.s, s); 4286 if (isgreaterequal(ua.h, ub.h)) { 4287 if (isgreater(ua.h, ub.h)) { 4288 return float_relation_greater; 4289 } 4290 return float_relation_equal; 4291 } 4292 if (likely(isless(ua.h, ub.h))) { 4293 return float_relation_less; 4294 } 4295 /* 4296 * The only condition remaining is unordered. 4297 * Fall through to set flags. 4298 */ 4299 soft: 4300 return float32_do_compare(ua.s, ub.s, s, is_quiet); 4301 } 4302 4303 FloatRelation float32_compare(float32 a, float32 b, float_status *s) 4304 { 4305 return float32_hs_compare(a, b, s, false); 4306 } 4307 4308 FloatRelation float32_compare_quiet(float32 a, float32 b, float_status *s) 4309 { 4310 return float32_hs_compare(a, b, s, true); 4311 } 4312 4313 static FloatRelation QEMU_SOFTFLOAT_ATTR 4314 float64_do_compare(float64 a, float64 b, float_status *s, bool is_quiet) 4315 { 4316 FloatParts64 pa, pb; 4317 4318 float64_unpack_canonical(&pa, a, s); 4319 float64_unpack_canonical(&pb, b, s); 4320 return parts_compare(&pa, &pb, s, is_quiet); 4321 } 4322 4323 static FloatRelation QEMU_FLATTEN 4324 float64_hs_compare(float64 xa, float64 xb, float_status *s, bool is_quiet) 4325 { 4326 union_float64 ua, ub; 4327 4328 ua.s = xa; 4329 ub.s = xb; 4330 4331 if (QEMU_NO_HARDFLOAT) { 4332 goto soft; 4333 } 4334 4335 float64_input_flush2(&ua.s, &ub.s, s); 4336 if (isgreaterequal(ua.h, ub.h)) { 4337 if (isgreater(ua.h, ub.h)) { 4338 return float_relation_greater; 4339 } 4340 return float_relation_equal; 4341 } 4342 if (likely(isless(ua.h, ub.h))) { 4343 return float_relation_less; 4344 } 4345 /* 4346 * The only condition remaining is unordered. 4347 * Fall through to set flags. 4348 */ 4349 soft: 4350 return float64_do_compare(ua.s, ub.s, s, is_quiet); 4351 } 4352 4353 FloatRelation float64_compare(float64 a, float64 b, float_status *s) 4354 { 4355 return float64_hs_compare(a, b, s, false); 4356 } 4357 4358 FloatRelation float64_compare_quiet(float64 a, float64 b, float_status *s) 4359 { 4360 return float64_hs_compare(a, b, s, true); 4361 } 4362 4363 static FloatRelation QEMU_FLATTEN 4364 bfloat16_do_compare(bfloat16 a, bfloat16 b, float_status *s, bool is_quiet) 4365 { 4366 FloatParts64 pa, pb; 4367 4368 bfloat16_unpack_canonical(&pa, a, s); 4369 bfloat16_unpack_canonical(&pb, b, s); 4370 return parts_compare(&pa, &pb, s, is_quiet); 4371 } 4372 4373 FloatRelation bfloat16_compare(bfloat16 a, bfloat16 b, float_status *s) 4374 { 4375 return bfloat16_do_compare(a, b, s, false); 4376 } 4377 4378 FloatRelation bfloat16_compare_quiet(bfloat16 a, bfloat16 b, float_status *s) 4379 { 4380 return bfloat16_do_compare(a, b, s, true); 4381 } 4382 4383 static FloatRelation QEMU_FLATTEN 4384 float128_do_compare(float128 a, float128 b, float_status *s, bool is_quiet) 4385 { 4386 FloatParts128 pa, pb; 4387 4388 float128_unpack_canonical(&pa, a, s); 4389 float128_unpack_canonical(&pb, b, s); 4390 return parts_compare(&pa, &pb, s, is_quiet); 4391 } 4392 4393 FloatRelation float128_compare(float128 a, float128 b, float_status *s) 4394 { 4395 return float128_do_compare(a, b, s, false); 4396 } 4397 4398 FloatRelation float128_compare_quiet(float128 a, float128 b, float_status *s) 4399 { 4400 return float128_do_compare(a, b, s, true); 4401 } 4402 4403 static FloatRelation QEMU_FLATTEN 4404 floatx80_do_compare(floatx80 a, floatx80 b, float_status *s, bool is_quiet) 4405 { 4406 FloatParts128 pa, pb; 4407 4408 if (!floatx80_unpack_canonical(&pa, a, s) || 4409 !floatx80_unpack_canonical(&pb, b, s)) { 4410 return float_relation_unordered; 4411 } 4412 return parts_compare(&pa, &pb, s, is_quiet); 4413 } 4414 4415 FloatRelation floatx80_compare(floatx80 a, floatx80 b, float_status *s) 4416 { 4417 return floatx80_do_compare(a, b, s, false); 4418 } 4419 4420 FloatRelation floatx80_compare_quiet(floatx80 a, floatx80 b, float_status *s) 4421 { 4422 return floatx80_do_compare(a, b, s, true); 4423 } 4424 4425 /* 4426 * Scale by 2**N 4427 */ 4428 4429 float16 float16_scalbn(float16 a, int n, float_status *status) 4430 { 4431 FloatParts64 p; 4432 4433 float16_unpack_canonical(&p, a, status); 4434 parts_scalbn(&p, n, status); 4435 return float16_round_pack_canonical(&p, status); 4436 } 4437 4438 float32 float32_scalbn(float32 a, int n, float_status *status) 4439 { 4440 FloatParts64 p; 4441 4442 float32_unpack_canonical(&p, a, status); 4443 parts_scalbn(&p, n, status); 4444 return float32_round_pack_canonical(&p, status); 4445 } 4446 4447 float64 float64_scalbn(float64 a, int n, float_status *status) 4448 { 4449 FloatParts64 p; 4450 4451 float64_unpack_canonical(&p, a, status); 4452 parts_scalbn(&p, n, status); 4453 return float64_round_pack_canonical(&p, status); 4454 } 4455 4456 bfloat16 bfloat16_scalbn(bfloat16 a, int n, float_status *status) 4457 { 4458 FloatParts64 p; 4459 4460 bfloat16_unpack_canonical(&p, a, status); 4461 parts_scalbn(&p, n, status); 4462 return bfloat16_round_pack_canonical(&p, status); 4463 } 4464 4465 float128 float128_scalbn(float128 a, int n, float_status *status) 4466 { 4467 FloatParts128 p; 4468 4469 float128_unpack_canonical(&p, a, status); 4470 parts_scalbn(&p, n, status); 4471 return float128_round_pack_canonical(&p, status); 4472 } 4473 4474 floatx80 floatx80_scalbn(floatx80 a, int n, float_status *status) 4475 { 4476 FloatParts128 p; 4477 4478 if (!floatx80_unpack_canonical(&p, a, status)) { 4479 return floatx80_default_nan(status); 4480 } 4481 parts_scalbn(&p, n, status); 4482 return floatx80_round_pack_canonical(&p, status); 4483 } 4484 4485 /* 4486 * Square Root 4487 */ 4488 4489 float16 QEMU_FLATTEN float16_sqrt(float16 a, float_status *status) 4490 { 4491 FloatParts64 p; 4492 4493 float16_unpack_canonical(&p, a, status); 4494 parts_sqrt(&p, status, &float16_params); 4495 return float16_round_pack_canonical(&p, status); 4496 } 4497 4498 static float32 QEMU_SOFTFLOAT_ATTR 4499 soft_f32_sqrt(float32 a, float_status *status) 4500 { 4501 FloatParts64 p; 4502 4503 float32_unpack_canonical(&p, a, status); 4504 parts_sqrt(&p, status, &float32_params); 4505 return float32_round_pack_canonical(&p, status); 4506 } 4507 4508 static float64 QEMU_SOFTFLOAT_ATTR 4509 soft_f64_sqrt(float64 a, float_status *status) 4510 { 4511 FloatParts64 p; 4512 4513 float64_unpack_canonical(&p, a, status); 4514 parts_sqrt(&p, status, &float64_params); 4515 return float64_round_pack_canonical(&p, status); 4516 } 4517 4518 float32 QEMU_FLATTEN float32_sqrt(float32 xa, float_status *s) 4519 { 4520 union_float32 ua, ur; 4521 4522 ua.s = xa; 4523 if (unlikely(!can_use_fpu(s))) { 4524 goto soft; 4525 } 4526 4527 float32_input_flush1(&ua.s, s); 4528 if (QEMU_HARDFLOAT_1F32_USE_FP) { 4529 if (unlikely(!(fpclassify(ua.h) == FP_NORMAL || 4530 fpclassify(ua.h) == FP_ZERO) || 4531 signbit(ua.h))) { 4532 goto soft; 4533 } 4534 } else if (unlikely(!float32_is_zero_or_normal(ua.s) || 4535 float32_is_neg(ua.s))) { 4536 goto soft; 4537 } 4538 ur.h = sqrtf(ua.h); 4539 return ur.s; 4540 4541 soft: 4542 return soft_f32_sqrt(ua.s, s); 4543 } 4544 4545 float64 QEMU_FLATTEN float64_sqrt(float64 xa, float_status *s) 4546 { 4547 union_float64 ua, ur; 4548 4549 ua.s = xa; 4550 if (unlikely(!can_use_fpu(s))) { 4551 goto soft; 4552 } 4553 4554 float64_input_flush1(&ua.s, s); 4555 if (QEMU_HARDFLOAT_1F64_USE_FP) { 4556 if (unlikely(!(fpclassify(ua.h) == FP_NORMAL || 4557 fpclassify(ua.h) == FP_ZERO) || 4558 signbit(ua.h))) { 4559 goto soft; 4560 } 4561 } else if (unlikely(!float64_is_zero_or_normal(ua.s) || 4562 float64_is_neg(ua.s))) { 4563 goto soft; 4564 } 4565 ur.h = sqrt(ua.h); 4566 return ur.s; 4567 4568 soft: 4569 return soft_f64_sqrt(ua.s, s); 4570 } 4571 4572 float64 float64r32_sqrt(float64 a, float_status *status) 4573 { 4574 FloatParts64 p; 4575 4576 float64_unpack_canonical(&p, a, status); 4577 parts_sqrt(&p, status, &float64_params); 4578 return float64r32_round_pack_canonical(&p, status); 4579 } 4580 4581 bfloat16 QEMU_FLATTEN bfloat16_sqrt(bfloat16 a, float_status *status) 4582 { 4583 FloatParts64 p; 4584 4585 bfloat16_unpack_canonical(&p, a, status); 4586 parts_sqrt(&p, status, &bfloat16_params); 4587 return bfloat16_round_pack_canonical(&p, status); 4588 } 4589 4590 float128 QEMU_FLATTEN float128_sqrt(float128 a, float_status *status) 4591 { 4592 FloatParts128 p; 4593 4594 float128_unpack_canonical(&p, a, status); 4595 parts_sqrt(&p, status, &float128_params); 4596 return float128_round_pack_canonical(&p, status); 4597 } 4598 4599 floatx80 floatx80_sqrt(floatx80 a, float_status *s) 4600 { 4601 FloatParts128 p; 4602 4603 if (!floatx80_unpack_canonical(&p, a, s)) { 4604 return floatx80_default_nan(s); 4605 } 4606 parts_sqrt(&p, s, &floatx80_params[s->floatx80_rounding_precision]); 4607 return floatx80_round_pack_canonical(&p, s); 4608 } 4609 4610 /* 4611 * log2 4612 */ 4613 float32 float32_log2(float32 a, float_status *status) 4614 { 4615 FloatParts64 p; 4616 4617 float32_unpack_canonical(&p, a, status); 4618 parts_log2(&p, status, &float32_params); 4619 return float32_round_pack_canonical(&p, status); 4620 } 4621 4622 float64 float64_log2(float64 a, float_status *status) 4623 { 4624 FloatParts64 p; 4625 4626 float64_unpack_canonical(&p, a, status); 4627 parts_log2(&p, status, &float64_params); 4628 return float64_round_pack_canonical(&p, status); 4629 } 4630 4631 /*---------------------------------------------------------------------------- 4632 | The pattern for a default generated NaN. 4633 *----------------------------------------------------------------------------*/ 4634 4635 float16 float16_default_nan(float_status *status) 4636 { 4637 FloatParts64 p; 4638 4639 parts_default_nan(&p, status); 4640 p.frac >>= float16_params.frac_shift; 4641 return float16_pack_raw(&p); 4642 } 4643 4644 float32 float32_default_nan(float_status *status) 4645 { 4646 FloatParts64 p; 4647 4648 parts_default_nan(&p, status); 4649 p.frac >>= float32_params.frac_shift; 4650 return float32_pack_raw(&p); 4651 } 4652 4653 float64 float64_default_nan(float_status *status) 4654 { 4655 FloatParts64 p; 4656 4657 parts_default_nan(&p, status); 4658 p.frac >>= float64_params.frac_shift; 4659 return float64_pack_raw(&p); 4660 } 4661 4662 float128 float128_default_nan(float_status *status) 4663 { 4664 FloatParts128 p; 4665 4666 parts_default_nan(&p, status); 4667 frac_shr(&p, float128_params.frac_shift); 4668 return float128_pack_raw(&p); 4669 } 4670 4671 bfloat16 bfloat16_default_nan(float_status *status) 4672 { 4673 FloatParts64 p; 4674 4675 parts_default_nan(&p, status); 4676 p.frac >>= bfloat16_params.frac_shift; 4677 return bfloat16_pack_raw(&p); 4678 } 4679 4680 /*---------------------------------------------------------------------------- 4681 | Returns a quiet NaN from a signalling NaN for the floating point value `a'. 4682 *----------------------------------------------------------------------------*/ 4683 4684 float16 float16_silence_nan(float16 a, float_status *status) 4685 { 4686 FloatParts64 p; 4687 4688 float16_unpack_raw(&p, a); 4689 p.frac <<= float16_params.frac_shift; 4690 parts_silence_nan(&p, status); 4691 p.frac >>= float16_params.frac_shift; 4692 return float16_pack_raw(&p); 4693 } 4694 4695 float32 float32_silence_nan(float32 a, float_status *status) 4696 { 4697 FloatParts64 p; 4698 4699 float32_unpack_raw(&p, a); 4700 p.frac <<= float32_params.frac_shift; 4701 parts_silence_nan(&p, status); 4702 p.frac >>= float32_params.frac_shift; 4703 return float32_pack_raw(&p); 4704 } 4705 4706 float64 float64_silence_nan(float64 a, float_status *status) 4707 { 4708 FloatParts64 p; 4709 4710 float64_unpack_raw(&p, a); 4711 p.frac <<= float64_params.frac_shift; 4712 parts_silence_nan(&p, status); 4713 p.frac >>= float64_params.frac_shift; 4714 return float64_pack_raw(&p); 4715 } 4716 4717 bfloat16 bfloat16_silence_nan(bfloat16 a, float_status *status) 4718 { 4719 FloatParts64 p; 4720 4721 bfloat16_unpack_raw(&p, a); 4722 p.frac <<= bfloat16_params.frac_shift; 4723 parts_silence_nan(&p, status); 4724 p.frac >>= bfloat16_params.frac_shift; 4725 return bfloat16_pack_raw(&p); 4726 } 4727 4728 float128 float128_silence_nan(float128 a, float_status *status) 4729 { 4730 FloatParts128 p; 4731 4732 float128_unpack_raw(&p, a); 4733 frac_shl(&p, float128_params.frac_shift); 4734 parts_silence_nan(&p, status); 4735 frac_shr(&p, float128_params.frac_shift); 4736 return float128_pack_raw(&p); 4737 } 4738 4739 /*---------------------------------------------------------------------------- 4740 | If `a' is denormal and we are in flush-to-zero mode then set the 4741 | input-denormal exception and return zero. Otherwise just return the value. 4742 *----------------------------------------------------------------------------*/ 4743 4744 static bool parts_squash_denormal(FloatParts64 p, float_status *status) 4745 { 4746 if (p.exp == 0 && p.frac != 0) { 4747 float_raise(float_flag_input_denormal, status); 4748 return true; 4749 } 4750 4751 return false; 4752 } 4753 4754 float16 float16_squash_input_denormal(float16 a, float_status *status) 4755 { 4756 if (status->flush_inputs_to_zero) { 4757 FloatParts64 p; 4758 4759 float16_unpack_raw(&p, a); 4760 if (parts_squash_denormal(p, status)) { 4761 return float16_set_sign(float16_zero, p.sign); 4762 } 4763 } 4764 return a; 4765 } 4766 4767 float32 float32_squash_input_denormal(float32 a, float_status *status) 4768 { 4769 if (status->flush_inputs_to_zero) { 4770 FloatParts64 p; 4771 4772 float32_unpack_raw(&p, a); 4773 if (parts_squash_denormal(p, status)) { 4774 return float32_set_sign(float32_zero, p.sign); 4775 } 4776 } 4777 return a; 4778 } 4779 4780 float64 float64_squash_input_denormal(float64 a, float_status *status) 4781 { 4782 if (status->flush_inputs_to_zero) { 4783 FloatParts64 p; 4784 4785 float64_unpack_raw(&p, a); 4786 if (parts_squash_denormal(p, status)) { 4787 return float64_set_sign(float64_zero, p.sign); 4788 } 4789 } 4790 return a; 4791 } 4792 4793 bfloat16 bfloat16_squash_input_denormal(bfloat16 a, float_status *status) 4794 { 4795 if (status->flush_inputs_to_zero) { 4796 FloatParts64 p; 4797 4798 bfloat16_unpack_raw(&p, a); 4799 if (parts_squash_denormal(p, status)) { 4800 return bfloat16_set_sign(bfloat16_zero, p.sign); 4801 } 4802 } 4803 return a; 4804 } 4805 4806 /*---------------------------------------------------------------------------- 4807 | Normalizes the subnormal extended double-precision floating-point value 4808 | represented by the denormalized significand `aSig'. The normalized exponent 4809 | and significand are stored at the locations pointed to by `zExpPtr' and 4810 | `zSigPtr', respectively. 4811 *----------------------------------------------------------------------------*/ 4812 4813 void normalizeFloatx80Subnormal(uint64_t aSig, int32_t *zExpPtr, 4814 uint64_t *zSigPtr) 4815 { 4816 int8_t shiftCount; 4817 4818 shiftCount = clz64(aSig); 4819 *zSigPtr = aSig<<shiftCount; 4820 *zExpPtr = 1 - shiftCount; 4821 } 4822 4823 /*---------------------------------------------------------------------------- 4824 | Takes an abstract floating-point value having sign `zSign', exponent `zExp', 4825 | and extended significand formed by the concatenation of `zSig0' and `zSig1', 4826 | and returns the proper extended double-precision floating-point value 4827 | corresponding to the abstract input. Ordinarily, the abstract value is 4828 | rounded and packed into the extended double-precision format, with the 4829 | inexact exception raised if the abstract input cannot be represented 4830 | exactly. However, if the abstract value is too large, the overflow and 4831 | inexact exceptions are raised and an infinity or maximal finite value is 4832 | returned. If the abstract value is too small, the input value is rounded to 4833 | a subnormal number, and the underflow and inexact exceptions are raised if 4834 | the abstract input cannot be represented exactly as a subnormal extended 4835 | double-precision floating-point number. 4836 | If `roundingPrecision' is floatx80_precision_s or floatx80_precision_d, 4837 | the result is rounded to the same number of bits as single or double 4838 | precision, respectively. Otherwise, the result is rounded to the full 4839 | precision of the extended double-precision format. 4840 | The input significand must be normalized or smaller. If the input 4841 | significand is not normalized, `zExp' must be 0; in that case, the result 4842 | returned is a subnormal number, and it must not require rounding. The 4843 | handling of underflow and overflow follows the IEC/IEEE Standard for Binary 4844 | Floating-Point Arithmetic. 4845 *----------------------------------------------------------------------------*/ 4846 4847 floatx80 roundAndPackFloatx80(FloatX80RoundPrec roundingPrecision, bool zSign, 4848 int32_t zExp, uint64_t zSig0, uint64_t zSig1, 4849 float_status *status) 4850 { 4851 FloatRoundMode roundingMode; 4852 bool roundNearestEven, increment, isTiny; 4853 int64_t roundIncrement, roundMask, roundBits; 4854 4855 roundingMode = status->float_rounding_mode; 4856 roundNearestEven = ( roundingMode == float_round_nearest_even ); 4857 switch (roundingPrecision) { 4858 case floatx80_precision_x: 4859 goto precision80; 4860 case floatx80_precision_d: 4861 roundIncrement = UINT64_C(0x0000000000000400); 4862 roundMask = UINT64_C(0x00000000000007FF); 4863 break; 4864 case floatx80_precision_s: 4865 roundIncrement = UINT64_C(0x0000008000000000); 4866 roundMask = UINT64_C(0x000000FFFFFFFFFF); 4867 break; 4868 default: 4869 g_assert_not_reached(); 4870 } 4871 zSig0 |= ( zSig1 != 0 ); 4872 switch (roundingMode) { 4873 case float_round_nearest_even: 4874 case float_round_ties_away: 4875 break; 4876 case float_round_to_zero: 4877 roundIncrement = 0; 4878 break; 4879 case float_round_up: 4880 roundIncrement = zSign ? 0 : roundMask; 4881 break; 4882 case float_round_down: 4883 roundIncrement = zSign ? roundMask : 0; 4884 break; 4885 default: 4886 abort(); 4887 } 4888 roundBits = zSig0 & roundMask; 4889 if ( 0x7FFD <= (uint32_t) ( zExp - 1 ) ) { 4890 if ( ( 0x7FFE < zExp ) 4891 || ( ( zExp == 0x7FFE ) && ( zSig0 + roundIncrement < zSig0 ) ) 4892 ) { 4893 goto overflow; 4894 } 4895 if ( zExp <= 0 ) { 4896 if (status->flush_to_zero) { 4897 float_raise(float_flag_output_denormal, status); 4898 return packFloatx80(zSign, 0, 0); 4899 } 4900 isTiny = status->tininess_before_rounding 4901 || (zExp < 0 ) 4902 || (zSig0 <= zSig0 + roundIncrement); 4903 shift64RightJamming( zSig0, 1 - zExp, &zSig0 ); 4904 zExp = 0; 4905 roundBits = zSig0 & roundMask; 4906 if (isTiny && roundBits) { 4907 float_raise(float_flag_underflow, status); 4908 } 4909 if (roundBits) { 4910 float_raise(float_flag_inexact, status); 4911 } 4912 zSig0 += roundIncrement; 4913 if ( (int64_t) zSig0 < 0 ) zExp = 1; 4914 roundIncrement = roundMask + 1; 4915 if ( roundNearestEven && ( roundBits<<1 == roundIncrement ) ) { 4916 roundMask |= roundIncrement; 4917 } 4918 zSig0 &= ~ roundMask; 4919 return packFloatx80( zSign, zExp, zSig0 ); 4920 } 4921 } 4922 if (roundBits) { 4923 float_raise(float_flag_inexact, status); 4924 } 4925 zSig0 += roundIncrement; 4926 if ( zSig0 < roundIncrement ) { 4927 ++zExp; 4928 zSig0 = UINT64_C(0x8000000000000000); 4929 } 4930 roundIncrement = roundMask + 1; 4931 if ( roundNearestEven && ( roundBits<<1 == roundIncrement ) ) { 4932 roundMask |= roundIncrement; 4933 } 4934 zSig0 &= ~ roundMask; 4935 if ( zSig0 == 0 ) zExp = 0; 4936 return packFloatx80( zSign, zExp, zSig0 ); 4937 precision80: 4938 switch (roundingMode) { 4939 case float_round_nearest_even: 4940 case float_round_ties_away: 4941 increment = ((int64_t)zSig1 < 0); 4942 break; 4943 case float_round_to_zero: 4944 increment = 0; 4945 break; 4946 case float_round_up: 4947 increment = !zSign && zSig1; 4948 break; 4949 case float_round_down: 4950 increment = zSign && zSig1; 4951 break; 4952 default: 4953 abort(); 4954 } 4955 if ( 0x7FFD <= (uint32_t) ( zExp - 1 ) ) { 4956 if ( ( 0x7FFE < zExp ) 4957 || ( ( zExp == 0x7FFE ) 4958 && ( zSig0 == UINT64_C(0xFFFFFFFFFFFFFFFF) ) 4959 && increment 4960 ) 4961 ) { 4962 roundMask = 0; 4963 overflow: 4964 float_raise(float_flag_overflow | float_flag_inexact, status); 4965 if ( ( roundingMode == float_round_to_zero ) 4966 || ( zSign && ( roundingMode == float_round_up ) ) 4967 || ( ! zSign && ( roundingMode == float_round_down ) ) 4968 ) { 4969 return packFloatx80( zSign, 0x7FFE, ~ roundMask ); 4970 } 4971 return packFloatx80(zSign, 4972 floatx80_infinity_high, 4973 floatx80_infinity_low); 4974 } 4975 if ( zExp <= 0 ) { 4976 isTiny = status->tininess_before_rounding 4977 || (zExp < 0) 4978 || !increment 4979 || (zSig0 < UINT64_C(0xFFFFFFFFFFFFFFFF)); 4980 shift64ExtraRightJamming( zSig0, zSig1, 1 - zExp, &zSig0, &zSig1 ); 4981 zExp = 0; 4982 if (isTiny && zSig1) { 4983 float_raise(float_flag_underflow, status); 4984 } 4985 if (zSig1) { 4986 float_raise(float_flag_inexact, status); 4987 } 4988 switch (roundingMode) { 4989 case float_round_nearest_even: 4990 case float_round_ties_away: 4991 increment = ((int64_t)zSig1 < 0); 4992 break; 4993 case float_round_to_zero: 4994 increment = 0; 4995 break; 4996 case float_round_up: 4997 increment = !zSign && zSig1; 4998 break; 4999 case float_round_down: 5000 increment = zSign && zSig1; 5001 break; 5002 default: 5003 abort(); 5004 } 5005 if ( increment ) { 5006 ++zSig0; 5007 if (!(zSig1 << 1) && roundNearestEven) { 5008 zSig0 &= ~1; 5009 } 5010 if ( (int64_t) zSig0 < 0 ) zExp = 1; 5011 } 5012 return packFloatx80( zSign, zExp, zSig0 ); 5013 } 5014 } 5015 if (zSig1) { 5016 float_raise(float_flag_inexact, status); 5017 } 5018 if ( increment ) { 5019 ++zSig0; 5020 if ( zSig0 == 0 ) { 5021 ++zExp; 5022 zSig0 = UINT64_C(0x8000000000000000); 5023 } 5024 else { 5025 if (!(zSig1 << 1) && roundNearestEven) { 5026 zSig0 &= ~1; 5027 } 5028 } 5029 } 5030 else { 5031 if ( zSig0 == 0 ) zExp = 0; 5032 } 5033 return packFloatx80( zSign, zExp, zSig0 ); 5034 5035 } 5036 5037 /*---------------------------------------------------------------------------- 5038 | Takes an abstract floating-point value having sign `zSign', exponent 5039 | `zExp', and significand formed by the concatenation of `zSig0' and `zSig1', 5040 | and returns the proper extended double-precision floating-point value 5041 | corresponding to the abstract input. This routine is just like 5042 | `roundAndPackFloatx80' except that the input significand does not have to be 5043 | normalized. 5044 *----------------------------------------------------------------------------*/ 5045 5046 floatx80 normalizeRoundAndPackFloatx80(FloatX80RoundPrec roundingPrecision, 5047 bool zSign, int32_t zExp, 5048 uint64_t zSig0, uint64_t zSig1, 5049 float_status *status) 5050 { 5051 int8_t shiftCount; 5052 5053 if ( zSig0 == 0 ) { 5054 zSig0 = zSig1; 5055 zSig1 = 0; 5056 zExp -= 64; 5057 } 5058 shiftCount = clz64(zSig0); 5059 shortShift128Left( zSig0, zSig1, shiftCount, &zSig0, &zSig1 ); 5060 zExp -= shiftCount; 5061 return roundAndPackFloatx80(roundingPrecision, zSign, zExp, 5062 zSig0, zSig1, status); 5063 5064 } 5065 5066 /*---------------------------------------------------------------------------- 5067 | Returns the binary exponential of the single-precision floating-point value 5068 | `a'. The operation is performed according to the IEC/IEEE Standard for 5069 | Binary Floating-Point Arithmetic. 5070 | 5071 | Uses the following identities: 5072 | 5073 | 1. ------------------------------------------------------------------------- 5074 | x x*ln(2) 5075 | 2 = e 5076 | 5077 | 2. ------------------------------------------------------------------------- 5078 | 2 3 4 5 n 5079 | x x x x x x x 5080 | e = 1 + --- + --- + --- + --- + --- + ... + --- + ... 5081 | 1! 2! 3! 4! 5! n! 5082 *----------------------------------------------------------------------------*/ 5083 5084 static const float64 float32_exp2_coefficients[15] = 5085 { 5086 const_float64( 0x3ff0000000000000ll ), /* 1 */ 5087 const_float64( 0x3fe0000000000000ll ), /* 2 */ 5088 const_float64( 0x3fc5555555555555ll ), /* 3 */ 5089 const_float64( 0x3fa5555555555555ll ), /* 4 */ 5090 const_float64( 0x3f81111111111111ll ), /* 5 */ 5091 const_float64( 0x3f56c16c16c16c17ll ), /* 6 */ 5092 const_float64( 0x3f2a01a01a01a01all ), /* 7 */ 5093 const_float64( 0x3efa01a01a01a01all ), /* 8 */ 5094 const_float64( 0x3ec71de3a556c734ll ), /* 9 */ 5095 const_float64( 0x3e927e4fb7789f5cll ), /* 10 */ 5096 const_float64( 0x3e5ae64567f544e4ll ), /* 11 */ 5097 const_float64( 0x3e21eed8eff8d898ll ), /* 12 */ 5098 const_float64( 0x3de6124613a86d09ll ), /* 13 */ 5099 const_float64( 0x3da93974a8c07c9dll ), /* 14 */ 5100 const_float64( 0x3d6ae7f3e733b81fll ), /* 15 */ 5101 }; 5102 5103 float32 float32_exp2(float32 a, float_status *status) 5104 { 5105 FloatParts64 xp, xnp, tp, rp; 5106 int i; 5107 5108 float32_unpack_canonical(&xp, a, status); 5109 if (unlikely(xp.cls != float_class_normal)) { 5110 switch (xp.cls) { 5111 case float_class_snan: 5112 case float_class_qnan: 5113 parts_return_nan(&xp, status); 5114 return float32_round_pack_canonical(&xp, status); 5115 case float_class_inf: 5116 return xp.sign ? float32_zero : a; 5117 case float_class_zero: 5118 return float32_one; 5119 default: 5120 break; 5121 } 5122 g_assert_not_reached(); 5123 } 5124 5125 float_raise(float_flag_inexact, status); 5126 5127 float64_unpack_canonical(&tp, float64_ln2, status); 5128 xp = *parts_mul(&xp, &tp, status); 5129 xnp = xp; 5130 5131 float64_unpack_canonical(&rp, float64_one, status); 5132 for (i = 0 ; i < 15 ; i++) { 5133 float64_unpack_canonical(&tp, float32_exp2_coefficients[i], status); 5134 rp = *parts_muladd(&tp, &xp, &rp, 0, status); 5135 xnp = *parts_mul(&xnp, &xp, status); 5136 } 5137 5138 return float32_round_pack_canonical(&rp, status); 5139 } 5140 5141 /*---------------------------------------------------------------------------- 5142 | Rounds the extended double-precision floating-point value `a' 5143 | to the precision provided by floatx80_rounding_precision and returns the 5144 | result as an extended double-precision floating-point value. 5145 | The operation is performed according to the IEC/IEEE Standard for Binary 5146 | Floating-Point Arithmetic. 5147 *----------------------------------------------------------------------------*/ 5148 5149 floatx80 floatx80_round(floatx80 a, float_status *status) 5150 { 5151 FloatParts128 p; 5152 5153 if (!floatx80_unpack_canonical(&p, a, status)) { 5154 return floatx80_default_nan(status); 5155 } 5156 return floatx80_round_pack_canonical(&p, status); 5157 } 5158 5159 static void __attribute__((constructor)) softfloat_init(void) 5160 { 5161 union_float64 ua, ub, uc, ur; 5162 5163 if (QEMU_NO_HARDFLOAT) { 5164 return; 5165 } 5166 /* 5167 * Test that the host's FMA is not obviously broken. For example, 5168 * glibc < 2.23 can perform an incorrect FMA on certain hosts; see 5169 * https://sourceware.org/bugzilla/show_bug.cgi?id=13304 5170 */ 5171 ua.s = 0x0020000000000001ULL; 5172 ub.s = 0x3ca0000000000000ULL; 5173 uc.s = 0x0020000000000000ULL; 5174 ur.h = fma(ua.h, ub.h, uc.h); 5175 if (ur.s != 0x0020000000000001ULL) { 5176 force_soft_fma = true; 5177 } 5178 } 5179