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 header 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 | #ifndef SOFTFLOAT_H |
83 | #define SOFTFLOAT_H |
84 | |
85 | /*---------------------------------------------------------------------------- |
86 | | Software IEC/IEEE floating-point ordering relations |
87 | *----------------------------------------------------------------------------*/ |
88 | enum { |
89 | float_relation_less = -1, |
90 | float_relation_equal = 0, |
91 | float_relation_greater = 1, |
92 | float_relation_unordered = 2 |
93 | }; |
94 | |
95 | #include "fpu/softfloat-types.h" |
96 | #include "fpu/softfloat-helpers.h" |
97 | |
98 | /*---------------------------------------------------------------------------- |
99 | | Routine to raise any or all of the software IEC/IEEE floating-point |
100 | | exception flags. |
101 | *----------------------------------------------------------------------------*/ |
102 | void float_raise(uint8_t flags, float_status *status); |
103 | |
104 | /*---------------------------------------------------------------------------- |
105 | | If `a' is denormal and we are in flush-to-zero mode then set the |
106 | | input-denormal exception and return zero. Otherwise just return the value. |
107 | *----------------------------------------------------------------------------*/ |
108 | float16 float16_squash_input_denormal(float16 a, float_status *status); |
109 | float32 float32_squash_input_denormal(float32 a, float_status *status); |
110 | float64 float64_squash_input_denormal(float64 a, float_status *status); |
111 | |
112 | /*---------------------------------------------------------------------------- |
113 | | Options to indicate which negations to perform in float*_muladd() |
114 | | Using these differs from negating an input or output before calling |
115 | | the muladd function in that this means that a NaN doesn't have its |
116 | | sign bit inverted before it is propagated. |
117 | | We also support halving the result before rounding, as a special |
118 | | case to support the ARM fused-sqrt-step instruction FRSQRTS. |
119 | *----------------------------------------------------------------------------*/ |
120 | enum { |
121 | float_muladd_negate_c = 1, |
122 | float_muladd_negate_product = 2, |
123 | float_muladd_negate_result = 4, |
124 | float_muladd_halve_result = 8, |
125 | }; |
126 | |
127 | /*---------------------------------------------------------------------------- |
128 | | Software IEC/IEEE integer-to-floating-point conversion routines. |
129 | *----------------------------------------------------------------------------*/ |
130 | |
131 | float16 int16_to_float16_scalbn(int16_t a, int, float_status *status); |
132 | float16 int32_to_float16_scalbn(int32_t a, int, float_status *status); |
133 | float16 int64_to_float16_scalbn(int64_t a, int, float_status *status); |
134 | float16 uint16_to_float16_scalbn(uint16_t a, int, float_status *status); |
135 | float16 uint32_to_float16_scalbn(uint32_t a, int, float_status *status); |
136 | float16 uint64_to_float16_scalbn(uint64_t a, int, float_status *status); |
137 | |
138 | float16 int16_to_float16(int16_t a, float_status *status); |
139 | float16 int32_to_float16(int32_t a, float_status *status); |
140 | float16 int64_to_float16(int64_t a, float_status *status); |
141 | float16 uint16_to_float16(uint16_t a, float_status *status); |
142 | float16 uint32_to_float16(uint32_t a, float_status *status); |
143 | float16 uint64_to_float16(uint64_t a, float_status *status); |
144 | |
145 | float32 int16_to_float32_scalbn(int16_t, int, float_status *status); |
146 | float32 int32_to_float32_scalbn(int32_t, int, float_status *status); |
147 | float32 int64_to_float32_scalbn(int64_t, int, float_status *status); |
148 | float32 uint16_to_float32_scalbn(uint16_t, int, float_status *status); |
149 | float32 uint32_to_float32_scalbn(uint32_t, int, float_status *status); |
150 | float32 uint64_to_float32_scalbn(uint64_t, int, float_status *status); |
151 | |
152 | float32 int16_to_float32(int16_t, float_status *status); |
153 | float32 int32_to_float32(int32_t, float_status *status); |
154 | float32 int64_to_float32(int64_t, float_status *status); |
155 | float32 uint16_to_float32(uint16_t, float_status *status); |
156 | float32 uint32_to_float32(uint32_t, float_status *status); |
157 | float32 uint64_to_float32(uint64_t, float_status *status); |
158 | |
159 | float64 int16_to_float64_scalbn(int16_t, int, float_status *status); |
160 | float64 int32_to_float64_scalbn(int32_t, int, float_status *status); |
161 | float64 int64_to_float64_scalbn(int64_t, int, float_status *status); |
162 | float64 uint16_to_float64_scalbn(uint16_t, int, float_status *status); |
163 | float64 uint32_to_float64_scalbn(uint32_t, int, float_status *status); |
164 | float64 uint64_to_float64_scalbn(uint64_t, int, float_status *status); |
165 | |
166 | float64 int16_to_float64(int16_t, float_status *status); |
167 | float64 int32_to_float64(int32_t, float_status *status); |
168 | float64 int64_to_float64(int64_t, float_status *status); |
169 | float64 uint16_to_float64(uint16_t, float_status *status); |
170 | float64 uint32_to_float64(uint32_t, float_status *status); |
171 | float64 uint64_to_float64(uint64_t, float_status *status); |
172 | |
173 | floatx80 int32_to_floatx80(int32_t, float_status *status); |
174 | floatx80 int64_to_floatx80(int64_t, float_status *status); |
175 | |
176 | float128 int32_to_float128(int32_t, float_status *status); |
177 | float128 int64_to_float128(int64_t, float_status *status); |
178 | float128 uint64_to_float128(uint64_t, float_status *status); |
179 | |
180 | /*---------------------------------------------------------------------------- |
181 | | Software half-precision conversion routines. |
182 | *----------------------------------------------------------------------------*/ |
183 | |
184 | float16 float32_to_float16(float32, bool ieee, float_status *status); |
185 | float32 float16_to_float32(float16, bool ieee, float_status *status); |
186 | float16 float64_to_float16(float64 a, bool ieee, float_status *status); |
187 | float64 float16_to_float64(float16 a, bool ieee, float_status *status); |
188 | |
189 | int16_t float16_to_int16_scalbn(float16, int, int, float_status *status); |
190 | int32_t float16_to_int32_scalbn(float16, int, int, float_status *status); |
191 | int64_t float16_to_int64_scalbn(float16, int, int, float_status *status); |
192 | |
193 | int16_t float16_to_int16(float16, float_status *status); |
194 | int32_t float16_to_int32(float16, float_status *status); |
195 | int64_t float16_to_int64(float16, float_status *status); |
196 | |
197 | int16_t float16_to_int16_round_to_zero(float16, float_status *status); |
198 | int32_t float16_to_int32_round_to_zero(float16, float_status *status); |
199 | int64_t float16_to_int64_round_to_zero(float16, float_status *status); |
200 | |
201 | uint16_t float16_to_uint16_scalbn(float16 a, int, int, float_status *status); |
202 | uint32_t float16_to_uint32_scalbn(float16 a, int, int, float_status *status); |
203 | uint64_t float16_to_uint64_scalbn(float16 a, int, int, float_status *status); |
204 | |
205 | uint16_t float16_to_uint16(float16 a, float_status *status); |
206 | uint32_t float16_to_uint32(float16 a, float_status *status); |
207 | uint64_t float16_to_uint64(float16 a, float_status *status); |
208 | |
209 | uint16_t float16_to_uint16_round_to_zero(float16 a, float_status *status); |
210 | uint32_t float16_to_uint32_round_to_zero(float16 a, float_status *status); |
211 | uint64_t float16_to_uint64_round_to_zero(float16 a, float_status *status); |
212 | |
213 | /*---------------------------------------------------------------------------- |
214 | | Software half-precision operations. |
215 | *----------------------------------------------------------------------------*/ |
216 | |
217 | float16 float16_round_to_int(float16, float_status *status); |
218 | float16 float16_add(float16, float16, float_status *status); |
219 | float16 float16_sub(float16, float16, float_status *status); |
220 | float16 float16_mul(float16, float16, float_status *status); |
221 | float16 float16_muladd(float16, float16, float16, int, float_status *status); |
222 | float16 float16_div(float16, float16, float_status *status); |
223 | float16 float16_scalbn(float16, int, float_status *status); |
224 | float16 float16_min(float16, float16, float_status *status); |
225 | float16 float16_max(float16, float16, float_status *status); |
226 | float16 float16_minnum(float16, float16, float_status *status); |
227 | float16 float16_maxnum(float16, float16, float_status *status); |
228 | float16 float16_minnummag(float16, float16, float_status *status); |
229 | float16 float16_maxnummag(float16, float16, float_status *status); |
230 | float16 float16_sqrt(float16, float_status *status); |
231 | int float16_compare(float16, float16, float_status *status); |
232 | int float16_compare_quiet(float16, float16, float_status *status); |
233 | |
234 | int float16_is_quiet_nan(float16, float_status *status); |
235 | int float16_is_signaling_nan(float16, float_status *status); |
236 | float16 float16_silence_nan(float16, float_status *status); |
237 | |
238 | static inline int float16_is_any_nan(float16 a) |
239 | { |
240 | return ((float16_val(a) & ~0x8000) > 0x7c00); |
241 | } |
242 | |
243 | static inline int float16_is_neg(float16 a) |
244 | { |
245 | return float16_val(a) >> 15; |
246 | } |
247 | |
248 | static inline int float16_is_infinity(float16 a) |
249 | { |
250 | return (float16_val(a) & 0x7fff) == 0x7c00; |
251 | } |
252 | |
253 | static inline int float16_is_zero(float16 a) |
254 | { |
255 | return (float16_val(a) & 0x7fff) == 0; |
256 | } |
257 | |
258 | static inline int float16_is_zero_or_denormal(float16 a) |
259 | { |
260 | return (float16_val(a) & 0x7c00) == 0; |
261 | } |
262 | |
263 | static inline float16 float16_abs(float16 a) |
264 | { |
265 | /* Note that abs does *not* handle NaN specially, nor does |
266 | * it flush denormal inputs to zero. |
267 | */ |
268 | return make_float16(float16_val(a) & 0x7fff); |
269 | } |
270 | |
271 | static inline float16 float16_chs(float16 a) |
272 | { |
273 | /* Note that chs does *not* handle NaN specially, nor does |
274 | * it flush denormal inputs to zero. |
275 | */ |
276 | return make_float16(float16_val(a) ^ 0x8000); |
277 | } |
278 | |
279 | static inline float16 float16_set_sign(float16 a, int sign) |
280 | { |
281 | return make_float16((float16_val(a) & 0x7fff) | (sign << 15)); |
282 | } |
283 | |
284 | #define float16_zero make_float16(0) |
285 | #define float16_half make_float16(0x3800) |
286 | #define float16_one make_float16(0x3c00) |
287 | #define float16_one_point_five make_float16(0x3e00) |
288 | #define float16_two make_float16(0x4000) |
289 | #define float16_three make_float16(0x4200) |
290 | #define float16_infinity make_float16(0x7c00) |
291 | |
292 | /*---------------------------------------------------------------------------- |
293 | | The pattern for a default generated half-precision NaN. |
294 | *----------------------------------------------------------------------------*/ |
295 | float16 float16_default_nan(float_status *status); |
296 | |
297 | /*---------------------------------------------------------------------------- |
298 | | Software IEC/IEEE single-precision conversion routines. |
299 | *----------------------------------------------------------------------------*/ |
300 | |
301 | int16_t float32_to_int16_scalbn(float32, int, int, float_status *status); |
302 | int32_t float32_to_int32_scalbn(float32, int, int, float_status *status); |
303 | int64_t float32_to_int64_scalbn(float32, int, int, float_status *status); |
304 | |
305 | int16_t float32_to_int16(float32, float_status *status); |
306 | int32_t float32_to_int32(float32, float_status *status); |
307 | int64_t float32_to_int64(float32, float_status *status); |
308 | |
309 | int16_t float32_to_int16_round_to_zero(float32, float_status *status); |
310 | int32_t float32_to_int32_round_to_zero(float32, float_status *status); |
311 | int64_t float32_to_int64_round_to_zero(float32, float_status *status); |
312 | |
313 | uint16_t float32_to_uint16_scalbn(float32, int, int, float_status *status); |
314 | uint32_t float32_to_uint32_scalbn(float32, int, int, float_status *status); |
315 | uint64_t float32_to_uint64_scalbn(float32, int, int, float_status *status); |
316 | |
317 | uint16_t float32_to_uint16(float32, float_status *status); |
318 | uint32_t float32_to_uint32(float32, float_status *status); |
319 | uint64_t float32_to_uint64(float32, float_status *status); |
320 | |
321 | uint16_t float32_to_uint16_round_to_zero(float32, float_status *status); |
322 | uint32_t float32_to_uint32_round_to_zero(float32, float_status *status); |
323 | uint64_t float32_to_uint64_round_to_zero(float32, float_status *status); |
324 | |
325 | float64 float32_to_float64(float32, float_status *status); |
326 | floatx80 float32_to_floatx80(float32, float_status *status); |
327 | float128 float32_to_float128(float32, float_status *status); |
328 | |
329 | /*---------------------------------------------------------------------------- |
330 | | Software IEC/IEEE single-precision operations. |
331 | *----------------------------------------------------------------------------*/ |
332 | float32 float32_round_to_int(float32, float_status *status); |
333 | float32 float32_add(float32, float32, float_status *status); |
334 | float32 float32_sub(float32, float32, float_status *status); |
335 | float32 float32_mul(float32, float32, float_status *status); |
336 | float32 float32_div(float32, float32, float_status *status); |
337 | float32 float32_rem(float32, float32, float_status *status); |
338 | float32 float32_muladd(float32, float32, float32, int, float_status *status); |
339 | float32 float32_sqrt(float32, float_status *status); |
340 | float32 float32_exp2(float32, float_status *status); |
341 | float32 float32_log2(float32, float_status *status); |
342 | int float32_eq(float32, float32, float_status *status); |
343 | int float32_le(float32, float32, float_status *status); |
344 | int float32_lt(float32, float32, float_status *status); |
345 | int float32_unordered(float32, float32, float_status *status); |
346 | int float32_eq_quiet(float32, float32, float_status *status); |
347 | int float32_le_quiet(float32, float32, float_status *status); |
348 | int float32_lt_quiet(float32, float32, float_status *status); |
349 | int float32_unordered_quiet(float32, float32, float_status *status); |
350 | int float32_compare(float32, float32, float_status *status); |
351 | int float32_compare_quiet(float32, float32, float_status *status); |
352 | float32 float32_min(float32, float32, float_status *status); |
353 | float32 float32_max(float32, float32, float_status *status); |
354 | float32 float32_minnum(float32, float32, float_status *status); |
355 | float32 float32_maxnum(float32, float32, float_status *status); |
356 | float32 float32_minnummag(float32, float32, float_status *status); |
357 | float32 float32_maxnummag(float32, float32, float_status *status); |
358 | int float32_is_quiet_nan(float32, float_status *status); |
359 | int float32_is_signaling_nan(float32, float_status *status); |
360 | float32 float32_silence_nan(float32, float_status *status); |
361 | float32 float32_scalbn(float32, int, float_status *status); |
362 | |
363 | static inline float32 float32_abs(float32 a) |
364 | { |
365 | /* Note that abs does *not* handle NaN specially, nor does |
366 | * it flush denormal inputs to zero. |
367 | */ |
368 | return make_float32(float32_val(a) & 0x7fffffff); |
369 | } |
370 | |
371 | static inline float32 float32_chs(float32 a) |
372 | { |
373 | /* Note that chs does *not* handle NaN specially, nor does |
374 | * it flush denormal inputs to zero. |
375 | */ |
376 | return make_float32(float32_val(a) ^ 0x80000000); |
377 | } |
378 | |
379 | static inline int float32_is_infinity(float32 a) |
380 | { |
381 | return (float32_val(a) & 0x7fffffff) == 0x7f800000; |
382 | } |
383 | |
384 | static inline int float32_is_neg(float32 a) |
385 | { |
386 | return float32_val(a) >> 31; |
387 | } |
388 | |
389 | static inline int float32_is_zero(float32 a) |
390 | { |
391 | return (float32_val(a) & 0x7fffffff) == 0; |
392 | } |
393 | |
394 | static inline int float32_is_any_nan(float32 a) |
395 | { |
396 | return ((float32_val(a) & ~(1 << 31)) > 0x7f800000UL); |
397 | } |
398 | |
399 | static inline int float32_is_zero_or_denormal(float32 a) |
400 | { |
401 | return (float32_val(a) & 0x7f800000) == 0; |
402 | } |
403 | |
404 | static inline bool float32_is_normal(float32 a) |
405 | { |
406 | return (((float32_val(a) >> 23) + 1) & 0xff) >= 2; |
407 | } |
408 | |
409 | static inline bool float32_is_denormal(float32 a) |
410 | { |
411 | return float32_is_zero_or_denormal(a) && !float32_is_zero(a); |
412 | } |
413 | |
414 | static inline bool float32_is_zero_or_normal(float32 a) |
415 | { |
416 | return float32_is_normal(a) || float32_is_zero(a); |
417 | } |
418 | |
419 | static inline float32 float32_set_sign(float32 a, int sign) |
420 | { |
421 | return make_float32((float32_val(a) & 0x7fffffff) | (sign << 31)); |
422 | } |
423 | |
424 | #define float32_zero make_float32(0) |
425 | #define float32_half make_float32(0x3f000000) |
426 | #define float32_one make_float32(0x3f800000) |
427 | #define float32_one_point_five make_float32(0x3fc00000) |
428 | #define float32_two make_float32(0x40000000) |
429 | #define float32_three make_float32(0x40400000) |
430 | #define float32_infinity make_float32(0x7f800000) |
431 | |
432 | /*---------------------------------------------------------------------------- |
433 | | Packs the sign `zSign', exponent `zExp', and significand `zSig' into a |
434 | | single-precision floating-point value, returning the result. After being |
435 | | shifted into the proper positions, the three fields are simply added |
436 | | together to form the result. This means that any integer portion of `zSig' |
437 | | will be added into the exponent. Since a properly normalized significand |
438 | | will have an integer portion equal to 1, the `zExp' input should be 1 less |
439 | | than the desired result exponent whenever `zSig' is a complete, normalized |
440 | | significand. |
441 | *----------------------------------------------------------------------------*/ |
442 | |
443 | static inline float32 packFloat32(flag zSign, int zExp, uint32_t zSig) |
444 | { |
445 | return make_float32( |
446 | (((uint32_t)zSign) << 31) + (((uint32_t)zExp) << 23) + zSig); |
447 | } |
448 | |
449 | /*---------------------------------------------------------------------------- |
450 | | The pattern for a default generated single-precision NaN. |
451 | *----------------------------------------------------------------------------*/ |
452 | float32 float32_default_nan(float_status *status); |
453 | |
454 | /*---------------------------------------------------------------------------- |
455 | | Software IEC/IEEE double-precision conversion routines. |
456 | *----------------------------------------------------------------------------*/ |
457 | |
458 | int16_t float64_to_int16_scalbn(float64, int, int, float_status *status); |
459 | int32_t float64_to_int32_scalbn(float64, int, int, float_status *status); |
460 | int64_t float64_to_int64_scalbn(float64, int, int, float_status *status); |
461 | |
462 | int16_t float64_to_int16(float64, float_status *status); |
463 | int32_t float64_to_int32(float64, float_status *status); |
464 | int64_t float64_to_int64(float64, float_status *status); |
465 | |
466 | int16_t float64_to_int16_round_to_zero(float64, float_status *status); |
467 | int32_t float64_to_int32_round_to_zero(float64, float_status *status); |
468 | int64_t float64_to_int64_round_to_zero(float64, float_status *status); |
469 | |
470 | uint16_t float64_to_uint16_scalbn(float64, int, int, float_status *status); |
471 | uint32_t float64_to_uint32_scalbn(float64, int, int, float_status *status); |
472 | uint64_t float64_to_uint64_scalbn(float64, int, int, float_status *status); |
473 | |
474 | uint16_t float64_to_uint16(float64, float_status *status); |
475 | uint32_t float64_to_uint32(float64, float_status *status); |
476 | uint64_t float64_to_uint64(float64, float_status *status); |
477 | |
478 | uint16_t float64_to_uint16_round_to_zero(float64, float_status *status); |
479 | uint32_t float64_to_uint32_round_to_zero(float64, float_status *status); |
480 | uint64_t float64_to_uint64_round_to_zero(float64, float_status *status); |
481 | |
482 | float32 float64_to_float32(float64, float_status *status); |
483 | floatx80 float64_to_floatx80(float64, float_status *status); |
484 | float128 float64_to_float128(float64, float_status *status); |
485 | |
486 | /*---------------------------------------------------------------------------- |
487 | | Software IEC/IEEE double-precision operations. |
488 | *----------------------------------------------------------------------------*/ |
489 | float64 float64_round_to_int(float64, float_status *status); |
490 | float64 float64_add(float64, float64, float_status *status); |
491 | float64 float64_sub(float64, float64, float_status *status); |
492 | float64 float64_mul(float64, float64, float_status *status); |
493 | float64 float64_div(float64, float64, float_status *status); |
494 | float64 float64_rem(float64, float64, float_status *status); |
495 | float64 float64_muladd(float64, float64, float64, int, float_status *status); |
496 | float64 float64_sqrt(float64, float_status *status); |
497 | float64 float64_log2(float64, float_status *status); |
498 | int float64_eq(float64, float64, float_status *status); |
499 | int float64_le(float64, float64, float_status *status); |
500 | int float64_lt(float64, float64, float_status *status); |
501 | int float64_unordered(float64, float64, float_status *status); |
502 | int float64_eq_quiet(float64, float64, float_status *status); |
503 | int float64_le_quiet(float64, float64, float_status *status); |
504 | int float64_lt_quiet(float64, float64, float_status *status); |
505 | int float64_unordered_quiet(float64, float64, float_status *status); |
506 | int float64_compare(float64, float64, float_status *status); |
507 | int float64_compare_quiet(float64, float64, float_status *status); |
508 | float64 float64_min(float64, float64, float_status *status); |
509 | float64 float64_max(float64, float64, float_status *status); |
510 | float64 float64_minnum(float64, float64, float_status *status); |
511 | float64 float64_maxnum(float64, float64, float_status *status); |
512 | float64 float64_minnummag(float64, float64, float_status *status); |
513 | float64 float64_maxnummag(float64, float64, float_status *status); |
514 | int float64_is_quiet_nan(float64 a, float_status *status); |
515 | int float64_is_signaling_nan(float64, float_status *status); |
516 | float64 float64_silence_nan(float64, float_status *status); |
517 | float64 float64_scalbn(float64, int, float_status *status); |
518 | |
519 | static inline float64 float64_abs(float64 a) |
520 | { |
521 | /* Note that abs does *not* handle NaN specially, nor does |
522 | * it flush denormal inputs to zero. |
523 | */ |
524 | return make_float64(float64_val(a) & 0x7fffffffffffffffLL); |
525 | } |
526 | |
527 | static inline float64 float64_chs(float64 a) |
528 | { |
529 | /* Note that chs does *not* handle NaN specially, nor does |
530 | * it flush denormal inputs to zero. |
531 | */ |
532 | return make_float64(float64_val(a) ^ 0x8000000000000000LL); |
533 | } |
534 | |
535 | static inline int float64_is_infinity(float64 a) |
536 | { |
537 | return (float64_val(a) & 0x7fffffffffffffffLL ) == 0x7ff0000000000000LL; |
538 | } |
539 | |
540 | static inline int float64_is_neg(float64 a) |
541 | { |
542 | return float64_val(a) >> 63; |
543 | } |
544 | |
545 | static inline int float64_is_zero(float64 a) |
546 | { |
547 | return (float64_val(a) & 0x7fffffffffffffffLL) == 0; |
548 | } |
549 | |
550 | static inline int float64_is_any_nan(float64 a) |
551 | { |
552 | return ((float64_val(a) & ~(1ULL << 63)) > 0x7ff0000000000000ULL); |
553 | } |
554 | |
555 | static inline int float64_is_zero_or_denormal(float64 a) |
556 | { |
557 | return (float64_val(a) & 0x7ff0000000000000LL) == 0; |
558 | } |
559 | |
560 | static inline bool float64_is_normal(float64 a) |
561 | { |
562 | return (((float64_val(a) >> 52) + 1) & 0x7ff) >= 2; |
563 | } |
564 | |
565 | static inline bool float64_is_denormal(float64 a) |
566 | { |
567 | return float64_is_zero_or_denormal(a) && !float64_is_zero(a); |
568 | } |
569 | |
570 | static inline bool float64_is_zero_or_normal(float64 a) |
571 | { |
572 | return float64_is_normal(a) || float64_is_zero(a); |
573 | } |
574 | |
575 | static inline float64 float64_set_sign(float64 a, int sign) |
576 | { |
577 | return make_float64((float64_val(a) & 0x7fffffffffffffffULL) |
578 | | ((int64_t)sign << 63)); |
579 | } |
580 | |
581 | #define float64_zero make_float64(0) |
582 | #define float64_half make_float64(0x3fe0000000000000LL) |
583 | #define float64_one make_float64(0x3ff0000000000000LL) |
584 | #define float64_one_point_five make_float64(0x3FF8000000000000ULL) |
585 | #define float64_two make_float64(0x4000000000000000ULL) |
586 | #define float64_three make_float64(0x4008000000000000ULL) |
587 | #define float64_ln2 make_float64(0x3fe62e42fefa39efLL) |
588 | #define float64_infinity make_float64(0x7ff0000000000000LL) |
589 | |
590 | /*---------------------------------------------------------------------------- |
591 | | The pattern for a default generated double-precision NaN. |
592 | *----------------------------------------------------------------------------*/ |
593 | float64 float64_default_nan(float_status *status); |
594 | |
595 | /*---------------------------------------------------------------------------- |
596 | | Software IEC/IEEE extended double-precision conversion routines. |
597 | *----------------------------------------------------------------------------*/ |
598 | int32_t floatx80_to_int32(floatx80, float_status *status); |
599 | int32_t floatx80_to_int32_round_to_zero(floatx80, float_status *status); |
600 | int64_t floatx80_to_int64(floatx80, float_status *status); |
601 | int64_t floatx80_to_int64_round_to_zero(floatx80, float_status *status); |
602 | float32 floatx80_to_float32(floatx80, float_status *status); |
603 | float64 floatx80_to_float64(floatx80, float_status *status); |
604 | float128 floatx80_to_float128(floatx80, float_status *status); |
605 | |
606 | /*---------------------------------------------------------------------------- |
607 | | The pattern for an extended double-precision inf. |
608 | *----------------------------------------------------------------------------*/ |
609 | extern const floatx80 floatx80_infinity; |
610 | |
611 | /*---------------------------------------------------------------------------- |
612 | | Software IEC/IEEE extended double-precision operations. |
613 | *----------------------------------------------------------------------------*/ |
614 | floatx80 floatx80_round(floatx80 a, float_status *status); |
615 | floatx80 floatx80_round_to_int(floatx80, float_status *status); |
616 | floatx80 floatx80_add(floatx80, floatx80, float_status *status); |
617 | floatx80 floatx80_sub(floatx80, floatx80, float_status *status); |
618 | floatx80 floatx80_mul(floatx80, floatx80, float_status *status); |
619 | floatx80 floatx80_div(floatx80, floatx80, float_status *status); |
620 | floatx80 floatx80_rem(floatx80, floatx80, float_status *status); |
621 | floatx80 floatx80_sqrt(floatx80, float_status *status); |
622 | int floatx80_eq(floatx80, floatx80, float_status *status); |
623 | int floatx80_le(floatx80, floatx80, float_status *status); |
624 | int floatx80_lt(floatx80, floatx80, float_status *status); |
625 | int floatx80_unordered(floatx80, floatx80, float_status *status); |
626 | int floatx80_eq_quiet(floatx80, floatx80, float_status *status); |
627 | int floatx80_le_quiet(floatx80, floatx80, float_status *status); |
628 | int floatx80_lt_quiet(floatx80, floatx80, float_status *status); |
629 | int floatx80_unordered_quiet(floatx80, floatx80, float_status *status); |
630 | int floatx80_compare(floatx80, floatx80, float_status *status); |
631 | int floatx80_compare_quiet(floatx80, floatx80, float_status *status); |
632 | int floatx80_is_quiet_nan(floatx80, float_status *status); |
633 | int floatx80_is_signaling_nan(floatx80, float_status *status); |
634 | floatx80 floatx80_silence_nan(floatx80, float_status *status); |
635 | floatx80 floatx80_scalbn(floatx80, int, float_status *status); |
636 | |
637 | static inline floatx80 floatx80_abs(floatx80 a) |
638 | { |
639 | a.high &= 0x7fff; |
640 | return a; |
641 | } |
642 | |
643 | static inline floatx80 floatx80_chs(floatx80 a) |
644 | { |
645 | a.high ^= 0x8000; |
646 | return a; |
647 | } |
648 | |
649 | static inline int floatx80_is_infinity(floatx80 a) |
650 | { |
651 | #if defined(TARGET_M68K) |
652 | return (a.high & 0x7fff) == floatx80_infinity.high && !(a.low << 1); |
653 | #else |
654 | return (a.high & 0x7fff) == floatx80_infinity.high && |
655 | a.low == floatx80_infinity.low; |
656 | #endif |
657 | } |
658 | |
659 | static inline int floatx80_is_neg(floatx80 a) |
660 | { |
661 | return a.high >> 15; |
662 | } |
663 | |
664 | static inline int floatx80_is_zero(floatx80 a) |
665 | { |
666 | return (a.high & 0x7fff) == 0 && a.low == 0; |
667 | } |
668 | |
669 | static inline int floatx80_is_zero_or_denormal(floatx80 a) |
670 | { |
671 | return (a.high & 0x7fff) == 0; |
672 | } |
673 | |
674 | static inline int floatx80_is_any_nan(floatx80 a) |
675 | { |
676 | return ((a.high & 0x7fff) == 0x7fff) && (a.low<<1); |
677 | } |
678 | |
679 | /*---------------------------------------------------------------------------- |
680 | | Return whether the given value is an invalid floatx80 encoding. |
681 | | Invalid floatx80 encodings arise when the integer bit is not set, but |
682 | | the exponent is not zero. The only times the integer bit is permitted to |
683 | | be zero is in subnormal numbers and the value zero. |
684 | | This includes what the Intel software developer's manual calls pseudo-NaNs, |
685 | | pseudo-infinities and un-normal numbers. It does not include |
686 | | pseudo-denormals, which must still be correctly handled as inputs even |
687 | | if they are never generated as outputs. |
688 | *----------------------------------------------------------------------------*/ |
689 | static inline bool floatx80_invalid_encoding(floatx80 a) |
690 | { |
691 | return (a.low & (1ULL << 63)) == 0 && (a.high & 0x7FFF) != 0; |
692 | } |
693 | |
694 | #define floatx80_zero make_floatx80(0x0000, 0x0000000000000000LL) |
695 | #define floatx80_one make_floatx80(0x3fff, 0x8000000000000000LL) |
696 | #define floatx80_ln2 make_floatx80(0x3ffe, 0xb17217f7d1cf79acLL) |
697 | #define floatx80_pi make_floatx80(0x4000, 0xc90fdaa22168c235LL) |
698 | #define floatx80_half make_floatx80(0x3ffe, 0x8000000000000000LL) |
699 | |
700 | /*---------------------------------------------------------------------------- |
701 | | Returns the fraction bits of the extended double-precision floating-point |
702 | | value `a'. |
703 | *----------------------------------------------------------------------------*/ |
704 | |
705 | static inline uint64_t (floatx80 a) |
706 | { |
707 | return a.low; |
708 | } |
709 | |
710 | /*---------------------------------------------------------------------------- |
711 | | Returns the exponent bits of the extended double-precision floating-point |
712 | | value `a'. |
713 | *----------------------------------------------------------------------------*/ |
714 | |
715 | static inline int32_t (floatx80 a) |
716 | { |
717 | return a.high & 0x7FFF; |
718 | } |
719 | |
720 | /*---------------------------------------------------------------------------- |
721 | | Returns the sign bit of the extended double-precision floating-point value |
722 | | `a'. |
723 | *----------------------------------------------------------------------------*/ |
724 | |
725 | static inline flag (floatx80 a) |
726 | { |
727 | return a.high >> 15; |
728 | } |
729 | |
730 | /*---------------------------------------------------------------------------- |
731 | | Packs the sign `zSign', exponent `zExp', and significand `zSig' into an |
732 | | extended double-precision floating-point value, returning the result. |
733 | *----------------------------------------------------------------------------*/ |
734 | |
735 | static inline floatx80 packFloatx80(flag zSign, int32_t zExp, uint64_t zSig) |
736 | { |
737 | floatx80 z; |
738 | |
739 | z.low = zSig; |
740 | z.high = (((uint16_t)zSign) << 15) + zExp; |
741 | return z; |
742 | } |
743 | |
744 | /*---------------------------------------------------------------------------- |
745 | | Normalizes the subnormal extended double-precision floating-point value |
746 | | represented by the denormalized significand `aSig'. The normalized exponent |
747 | | and significand are stored at the locations pointed to by `zExpPtr' and |
748 | | `zSigPtr', respectively. |
749 | *----------------------------------------------------------------------------*/ |
750 | |
751 | void normalizeFloatx80Subnormal(uint64_t aSig, int32_t *zExpPtr, |
752 | uint64_t *zSigPtr); |
753 | |
754 | /*---------------------------------------------------------------------------- |
755 | | Takes two extended double-precision floating-point values `a' and `b', one |
756 | | of which is a NaN, and returns the appropriate NaN result. If either `a' or |
757 | | `b' is a signaling NaN, the invalid exception is raised. |
758 | *----------------------------------------------------------------------------*/ |
759 | |
760 | floatx80 propagateFloatx80NaN(floatx80 a, floatx80 b, float_status *status); |
761 | |
762 | /*---------------------------------------------------------------------------- |
763 | | Takes an abstract floating-point value having sign `zSign', exponent `zExp', |
764 | | and extended significand formed by the concatenation of `zSig0' and `zSig1', |
765 | | and returns the proper extended double-precision floating-point value |
766 | | corresponding to the abstract input. Ordinarily, the abstract value is |
767 | | rounded and packed into the extended double-precision format, with the |
768 | | inexact exception raised if the abstract input cannot be represented |
769 | | exactly. However, if the abstract value is too large, the overflow and |
770 | | inexact exceptions are raised and an infinity or maximal finite value is |
771 | | returned. If the abstract value is too small, the input value is rounded to |
772 | | a subnormal number, and the underflow and inexact exceptions are raised if |
773 | | the abstract input cannot be represented exactly as a subnormal extended |
774 | | double-precision floating-point number. |
775 | | If `roundingPrecision' is 32 or 64, the result is rounded to the same |
776 | | number of bits as single or double precision, respectively. Otherwise, the |
777 | | result is rounded to the full precision of the extended double-precision |
778 | | format. |
779 | | The input significand must be normalized or smaller. If the input |
780 | | significand is not normalized, `zExp' must be 0; in that case, the result |
781 | | returned is a subnormal number, and it must not require rounding. The |
782 | | handling of underflow and overflow follows the IEC/IEEE Standard for Binary |
783 | | Floating-Point Arithmetic. |
784 | *----------------------------------------------------------------------------*/ |
785 | |
786 | floatx80 roundAndPackFloatx80(int8_t roundingPrecision, flag zSign, |
787 | int32_t zExp, uint64_t zSig0, uint64_t zSig1, |
788 | float_status *status); |
789 | |
790 | /*---------------------------------------------------------------------------- |
791 | | Takes an abstract floating-point value having sign `zSign', exponent |
792 | | `zExp', and significand formed by the concatenation of `zSig0' and `zSig1', |
793 | | and returns the proper extended double-precision floating-point value |
794 | | corresponding to the abstract input. This routine is just like |
795 | | `roundAndPackFloatx80' except that the input significand does not have to be |
796 | | normalized. |
797 | *----------------------------------------------------------------------------*/ |
798 | |
799 | floatx80 normalizeRoundAndPackFloatx80(int8_t roundingPrecision, |
800 | flag zSign, int32_t zExp, |
801 | uint64_t zSig0, uint64_t zSig1, |
802 | float_status *status); |
803 | |
804 | /*---------------------------------------------------------------------------- |
805 | | The pattern for a default generated extended double-precision NaN. |
806 | *----------------------------------------------------------------------------*/ |
807 | floatx80 floatx80_default_nan(float_status *status); |
808 | |
809 | /*---------------------------------------------------------------------------- |
810 | | Software IEC/IEEE quadruple-precision conversion routines. |
811 | *----------------------------------------------------------------------------*/ |
812 | int32_t float128_to_int32(float128, float_status *status); |
813 | int32_t float128_to_int32_round_to_zero(float128, float_status *status); |
814 | int64_t float128_to_int64(float128, float_status *status); |
815 | int64_t float128_to_int64_round_to_zero(float128, float_status *status); |
816 | uint64_t float128_to_uint64(float128, float_status *status); |
817 | uint64_t float128_to_uint64_round_to_zero(float128, float_status *status); |
818 | uint32_t float128_to_uint32(float128, float_status *status); |
819 | uint32_t float128_to_uint32_round_to_zero(float128, float_status *status); |
820 | float32 float128_to_float32(float128, float_status *status); |
821 | float64 float128_to_float64(float128, float_status *status); |
822 | floatx80 float128_to_floatx80(float128, float_status *status); |
823 | |
824 | /*---------------------------------------------------------------------------- |
825 | | Software IEC/IEEE quadruple-precision operations. |
826 | *----------------------------------------------------------------------------*/ |
827 | float128 float128_round_to_int(float128, float_status *status); |
828 | float128 float128_add(float128, float128, float_status *status); |
829 | float128 float128_sub(float128, float128, float_status *status); |
830 | float128 float128_mul(float128, float128, float_status *status); |
831 | float128 float128_div(float128, float128, float_status *status); |
832 | float128 float128_rem(float128, float128, float_status *status); |
833 | float128 float128_sqrt(float128, float_status *status); |
834 | int float128_eq(float128, float128, float_status *status); |
835 | int float128_le(float128, float128, float_status *status); |
836 | int float128_lt(float128, float128, float_status *status); |
837 | int float128_unordered(float128, float128, float_status *status); |
838 | int float128_eq_quiet(float128, float128, float_status *status); |
839 | int float128_le_quiet(float128, float128, float_status *status); |
840 | int float128_lt_quiet(float128, float128, float_status *status); |
841 | int float128_unordered_quiet(float128, float128, float_status *status); |
842 | int float128_compare(float128, float128, float_status *status); |
843 | int float128_compare_quiet(float128, float128, float_status *status); |
844 | int float128_is_quiet_nan(float128, float_status *status); |
845 | int float128_is_signaling_nan(float128, float_status *status); |
846 | float128 float128_silence_nan(float128, float_status *status); |
847 | float128 float128_scalbn(float128, int, float_status *status); |
848 | |
849 | static inline float128 float128_abs(float128 a) |
850 | { |
851 | a.high &= 0x7fffffffffffffffLL; |
852 | return a; |
853 | } |
854 | |
855 | static inline float128 float128_chs(float128 a) |
856 | { |
857 | a.high ^= 0x8000000000000000LL; |
858 | return a; |
859 | } |
860 | |
861 | static inline int float128_is_infinity(float128 a) |
862 | { |
863 | return (a.high & 0x7fffffffffffffffLL) == 0x7fff000000000000LL && a.low == 0; |
864 | } |
865 | |
866 | static inline int float128_is_neg(float128 a) |
867 | { |
868 | return a.high >> 63; |
869 | } |
870 | |
871 | static inline int float128_is_zero(float128 a) |
872 | { |
873 | return (a.high & 0x7fffffffffffffffLL) == 0 && a.low == 0; |
874 | } |
875 | |
876 | static inline int float128_is_zero_or_denormal(float128 a) |
877 | { |
878 | return (a.high & 0x7fff000000000000LL) == 0; |
879 | } |
880 | |
881 | static inline bool float128_is_normal(float128 a) |
882 | { |
883 | return (((a.high >> 48) + 1) & 0x7fff) >= 2; |
884 | } |
885 | |
886 | static inline bool float128_is_denormal(float128 a) |
887 | { |
888 | return float128_is_zero_or_denormal(a) && !float128_is_zero(a); |
889 | } |
890 | |
891 | static inline int float128_is_any_nan(float128 a) |
892 | { |
893 | return ((a.high >> 48) & 0x7fff) == 0x7fff && |
894 | ((a.low != 0) || ((a.high & 0xffffffffffffLL) != 0)); |
895 | } |
896 | |
897 | #define float128_zero make_float128(0, 0) |
898 | |
899 | /*---------------------------------------------------------------------------- |
900 | | The pattern for a default generated quadruple-precision NaN. |
901 | *----------------------------------------------------------------------------*/ |
902 | float128 float128_default_nan(float_status *status); |
903 | |
904 | #endif /* SOFTFLOAT_H */ |
905 | |