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 * This header holds definitions for code that might be dealing with 14 * softfloat types but not need access to the actual library functions. 15 */ 16 /* 17 =============================================================================== 18 This C header file is part of the SoftFloat IEC/IEEE Floating-point 19 Arithmetic Package, Release 2a. 20 21 Written by John R. Hauser. This work was made possible in part by the 22 International Computer Science Institute, located at Suite 600, 1947 Center 23 Street, Berkeley, California 94704. Funding was partially provided by the 24 National Science Foundation under grant MIP-9311980. The original version 25 of this code was written as part of a project to build a fixed-point vector 26 processor in collaboration with the University of California at Berkeley, 27 overseen by Profs. Nelson Morgan and John Wawrzynek. More information 28 is available through the Web page `http://HTTP.CS.Berkeley.EDU/~jhauser/ 29 arithmetic/SoftFloat.html'. 30 31 THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE. Although reasonable effort 32 has been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT 33 TIMES RESULT IN INCORRECT BEHAVIOR. USE OF THIS SOFTWARE IS RESTRICTED TO 34 PERSONS AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ANY 35 AND ALL LOSSES, COSTS, OR OTHER PROBLEMS ARISING FROM ITS USE. 36 37 Derivative works are acceptable, even for commercial purposes, so long as 38 (1) they include prominent notice that the work is derivative, and (2) they 39 include prominent notice akin to these four paragraphs for those parts of 40 this code that are retained. 41 42 =============================================================================== 43 */ 44 45 /* BSD licensing: 46 * Copyright (c) 2006, Fabrice Bellard 47 * All rights reserved. 48 * 49 * Redistribution and use in source and binary forms, with or without 50 * modification, are permitted provided that the following conditions are met: 51 * 52 * 1. Redistributions of source code must retain the above copyright notice, 53 * this list of conditions and the following disclaimer. 54 * 55 * 2. Redistributions in binary form must reproduce the above copyright notice, 56 * this list of conditions and the following disclaimer in the documentation 57 * and/or other materials provided with the distribution. 58 * 59 * 3. Neither the name of the copyright holder nor the names of its contributors 60 * may be used to endorse or promote products derived from this software without 61 * specific prior written permission. 62 * 63 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" 64 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 65 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 66 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE 67 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 68 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 69 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 70 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 71 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 72 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF 73 * THE POSSIBILITY OF SUCH DAMAGE. 74 */ 75 76 /* Portions of this work are licensed under the terms of the GNU GPL, 77 * version 2 or later. See the COPYING file in the top-level directory. 78 */ 79 80 #ifndef SOFTFLOAT_TYPES_H 81 #define SOFTFLOAT_TYPES_H 82 83 #include "hw/registerfields.h" 84 85 /* 86 * Software IEC/IEEE floating-point types. 87 */ 88 89 typedef uint16_t float16; 90 typedef uint32_t float32; 91 typedef uint64_t float64; 92 #define float16_val(x) (x) 93 #define float32_val(x) (x) 94 #define float64_val(x) (x) 95 #define make_float16(x) (x) 96 #define make_float32(x) (x) 97 #define make_float64(x) (x) 98 #define const_float16(x) (x) 99 #define const_float32(x) (x) 100 #define const_float64(x) (x) 101 typedef struct { 102 uint64_t low; 103 uint16_t high; 104 } floatx80; 105 #define make_floatx80(exp, mant) ((floatx80) { mant, exp }) 106 #define make_floatx80_init(exp, mant) { .low = mant, .high = exp } 107 typedef struct { 108 #if HOST_BIG_ENDIAN 109 uint64_t high, low; 110 #else 111 uint64_t low, high; 112 #endif 113 } float128; 114 #define make_float128(high_, low_) ((float128) { .high = high_, .low = low_ }) 115 #define make_float128_init(high_, low_) { .high = high_, .low = low_ } 116 117 /* 118 * Software neural-network floating-point types. 119 */ 120 typedef uint16_t bfloat16; 121 122 /* 123 * Software IEC/IEEE floating-point underflow tininess-detection mode. 124 */ 125 126 #define float_tininess_after_rounding false 127 #define float_tininess_before_rounding true 128 129 /* 130 *Software IEC/IEEE floating-point rounding mode. 131 */ 132 133 typedef enum __attribute__((__packed__)) { 134 float_round_nearest_even = 0, 135 float_round_down = 1, 136 float_round_up = 2, 137 float_round_to_zero = 3, 138 float_round_ties_away = 4, 139 /* Not an IEEE rounding mode: round to closest odd, overflow to max */ 140 float_round_to_odd = 5, 141 /* Not an IEEE rounding mode: round to closest odd, overflow to inf */ 142 float_round_to_odd_inf = 6, 143 /* Not an IEEE rounding mode: round to nearest even, overflow to max */ 144 float_round_nearest_even_max = 7, 145 } FloatRoundMode; 146 147 /* 148 * Software IEC/IEEE floating-point exception flags. 149 */ 150 151 enum { 152 float_flag_invalid = 0x0001, 153 float_flag_divbyzero = 0x0002, 154 float_flag_overflow = 0x0004, 155 float_flag_underflow = 0x0008, 156 float_flag_inexact = 0x0010, 157 /* We flushed an input denormal to 0 (because of flush_inputs_to_zero) */ 158 float_flag_input_denormal_flushed = 0x0020, 159 /* We flushed an output denormal to 0 (because of flush_to_zero) */ 160 float_flag_output_denormal_flushed = 0x0040, 161 float_flag_invalid_isi = 0x0080, /* inf - inf */ 162 float_flag_invalid_imz = 0x0100, /* inf * 0 */ 163 float_flag_invalid_idi = 0x0200, /* inf / inf */ 164 float_flag_invalid_zdz = 0x0400, /* 0 / 0 */ 165 float_flag_invalid_sqrt = 0x0800, /* sqrt(-x) */ 166 float_flag_invalid_cvti = 0x1000, /* non-nan to integer */ 167 float_flag_invalid_snan = 0x2000, /* any operand was snan */ 168 }; 169 170 /* 171 * Rounding precision for floatx80. 172 */ 173 typedef enum __attribute__((__packed__)) { 174 floatx80_precision_x, 175 floatx80_precision_d, 176 floatx80_precision_s, 177 } FloatX80RoundPrec; 178 179 /* 180 * 2-input NaN propagation rule. Individual architectures have 181 * different rules for which input NaN is propagated to the output 182 * when there is more than one NaN on the input. 183 * 184 * If default_nan_mode is enabled then it is valid not to set a 185 * NaN propagation rule, because the softfloat code guarantees 186 * not to try to pick a NaN to propagate in default NaN mode. 187 * When not in default-NaN mode, it is an error for the target 188 * not to set the rule in float_status, and we will assert if 189 * we need to handle an input NaN and no rule was selected. 190 */ 191 typedef enum __attribute__((__packed__)) { 192 /* No propagation rule specified */ 193 float_2nan_prop_none = 0, 194 /* Prefer SNaN over QNaN, then operand A over B */ 195 float_2nan_prop_s_ab, 196 /* Prefer SNaN over QNaN, then operand B over A */ 197 float_2nan_prop_s_ba, 198 /* Prefer A over B regardless of SNaN vs QNaN */ 199 float_2nan_prop_ab, 200 /* Prefer B over A regardless of SNaN vs QNaN */ 201 float_2nan_prop_ba, 202 /* 203 * This implements x87 NaN propagation rules: 204 * SNaN + QNaN => return the QNaN 205 * two SNaNs => return the one with the larger significand, silenced 206 * two QNaNs => return the one with the larger significand 207 * SNaN and a non-NaN => return the SNaN, silenced 208 * QNaN and a non-NaN => return the QNaN 209 * 210 * If we get down to comparing significands and they are the same, 211 * return the NaN with the positive sign bit (if any). 212 */ 213 float_2nan_prop_x87, 214 } Float2NaNPropRule; 215 216 /* 217 * 3-input NaN propagation rule, for fused multiply-add. Individual 218 * architectures have different rules for which input NaN is 219 * propagated to the output when there is more than one NaN on the 220 * input. 221 * 222 * If default_nan_mode is enabled then it is valid not to set a NaN 223 * propagation rule, because the softfloat code guarantees not to try 224 * to pick a NaN to propagate in default NaN mode. When not in 225 * default-NaN mode, it is an error for the target not to set the rule 226 * in float_status if it uses a muladd, and we will assert if we need 227 * to handle an input NaN and no rule was selected. 228 * 229 * The naming scheme for Float3NaNPropRule values is: 230 * float_3nan_prop_s_abc: 231 * = "Prefer SNaN over QNaN, then operand A over B over C" 232 * float_3nan_prop_abc: 233 * = "Prefer A over B over C regardless of SNaN vs QNAN" 234 * 235 * For QEMU, the multiply-add operation is A * B + C. 236 */ 237 238 /* 239 * We set the Float3NaNPropRule enum values up so we can select the 240 * right value in pickNaNMulAdd in a data driven way. 241 */ 242 FIELD(3NAN, 1ST, 0, 2) /* which operand is most preferred ? */ 243 FIELD(3NAN, 2ND, 2, 2) /* which operand is next most preferred ? */ 244 FIELD(3NAN, 3RD, 4, 2) /* which operand is least preferred ? */ 245 FIELD(3NAN, SNAN, 6, 1) /* do we prefer SNaN over QNaN ? */ 246 247 #define PROPRULE(X, Y, Z) \ 248 ((X << R_3NAN_1ST_SHIFT) | (Y << R_3NAN_2ND_SHIFT) | (Z << R_3NAN_3RD_SHIFT)) 249 250 typedef enum __attribute__((__packed__)) { 251 float_3nan_prop_none = 0, /* No propagation rule specified */ 252 float_3nan_prop_abc = PROPRULE(0, 1, 2), 253 float_3nan_prop_acb = PROPRULE(0, 2, 1), 254 float_3nan_prop_bac = PROPRULE(1, 0, 2), 255 float_3nan_prop_bca = PROPRULE(1, 2, 0), 256 float_3nan_prop_cab = PROPRULE(2, 0, 1), 257 float_3nan_prop_cba = PROPRULE(2, 1, 0), 258 float_3nan_prop_s_abc = float_3nan_prop_abc | R_3NAN_SNAN_MASK, 259 float_3nan_prop_s_acb = float_3nan_prop_acb | R_3NAN_SNAN_MASK, 260 float_3nan_prop_s_bac = float_3nan_prop_bac | R_3NAN_SNAN_MASK, 261 float_3nan_prop_s_bca = float_3nan_prop_bca | R_3NAN_SNAN_MASK, 262 float_3nan_prop_s_cab = float_3nan_prop_cab | R_3NAN_SNAN_MASK, 263 float_3nan_prop_s_cba = float_3nan_prop_cba | R_3NAN_SNAN_MASK, 264 } Float3NaNPropRule; 265 266 #undef PROPRULE 267 268 /* 269 * Rule for result of fused multiply-add 0 * Inf + NaN. 270 * This must be a NaN, but implementations differ on whether this 271 * is the input NaN or the default NaN. 272 * 273 * You don't need to set this if default_nan_mode is enabled. 274 * When not in default-NaN mode, it is an error for the target 275 * not to set the rule in float_status if it uses muladd, and we 276 * will assert if we need to handle an input NaN and no rule was 277 * selected. 278 */ 279 typedef enum __attribute__((__packed__)) { 280 /* No propagation rule specified */ 281 float_infzeronan_none = 0, 282 /* Result is never the default NaN (so always the input NaN) */ 283 float_infzeronan_dnan_never, 284 /* Result is always the default NaN */ 285 float_infzeronan_dnan_always, 286 /* Result is the default NaN if the input NaN is quiet */ 287 float_infzeronan_dnan_if_qnan, 288 } FloatInfZeroNaNRule; 289 290 /* 291 * Floating Point Status. Individual architectures may maintain 292 * several versions of float_status for different functions. The 293 * correct status for the operation is then passed by reference to 294 * most of the softfloat functions. 295 */ 296 297 typedef struct float_status { 298 uint16_t float_exception_flags; 299 FloatRoundMode float_rounding_mode; 300 FloatX80RoundPrec floatx80_rounding_precision; 301 Float2NaNPropRule float_2nan_prop_rule; 302 Float3NaNPropRule float_3nan_prop_rule; 303 FloatInfZeroNaNRule float_infzeronan_rule; 304 bool tininess_before_rounding; 305 /* should denormalised results go to zero and set output_denormal_flushed? */ 306 bool flush_to_zero; 307 /* should denormalised inputs go to zero and set input_denormal_flushed? */ 308 bool flush_inputs_to_zero; 309 bool default_nan_mode; 310 /* 311 * The pattern to use for the default NaN. Here the high bit specifies 312 * the default NaN's sign bit, and bits 6..0 specify the high bits of the 313 * fractional part. The low bits of the fractional part are copies of bit 0. 314 * The exponent of the default NaN is (as for any NaN) always all 1s. 315 * Note that a value of 0 here is not a valid NaN. The target must set 316 * this to the correct non-zero value, or we will assert when trying to 317 * create a default NaN. 318 */ 319 uint8_t default_nan_pattern; 320 /* 321 * The flags below are not used on all specializations and may 322 * constant fold away (see snan_bit_is_one()/no_signalling_nans() in 323 * softfloat-specialize.inc.c) 324 */ 325 bool snan_bit_is_one; 326 bool no_signaling_nans; 327 /* should overflowed results subtract re_bias to its exponent? */ 328 bool rebias_overflow; 329 /* should underflowed results add re_bias to its exponent? */ 330 bool rebias_underflow; 331 } float_status; 332 333 #endif /* SOFTFLOAT_TYPES_H */ 334