xref: /openbmc/qemu/target/arm/vfp_helper.c (revision 709395f8)
1 /*
2  * ARM VFP floating-point operations
3  *
4  *  Copyright (c) 2003 Fabrice Bellard
5  *
6  * This library is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * This library is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with this library; if not, see <http://www.gnu.org/licenses/>.
18  */
19 
20 #include "qemu/osdep.h"
21 #include "qemu/log.h"
22 #include "cpu.h"
23 #include "exec/helper-proto.h"
24 #include "fpu/softfloat.h"
25 #include "internals.h"
26 
27 
28 /* VFP support.  We follow the convention used for VFP instructions:
29    Single precision routines have a "s" suffix, double precision a
30    "d" suffix.  */
31 
32 /* Convert host exception flags to vfp form.  */
33 static inline int vfp_exceptbits_from_host(int host_bits)
34 {
35     int target_bits = 0;
36 
37     if (host_bits & float_flag_invalid)
38         target_bits |= 1;
39     if (host_bits & float_flag_divbyzero)
40         target_bits |= 2;
41     if (host_bits & float_flag_overflow)
42         target_bits |= 4;
43     if (host_bits & (float_flag_underflow | float_flag_output_denormal))
44         target_bits |= 8;
45     if (host_bits & float_flag_inexact)
46         target_bits |= 0x10;
47     if (host_bits & float_flag_input_denormal)
48         target_bits |= 0x80;
49     return target_bits;
50 }
51 
52 uint32_t HELPER(vfp_get_fpscr)(CPUARMState *env)
53 {
54     uint32_t i, fpscr;
55 
56     fpscr = env->vfp.xregs[ARM_VFP_FPSCR]
57             | (env->vfp.vec_len << 16)
58             | (env->vfp.vec_stride << 20);
59 
60     i = get_float_exception_flags(&env->vfp.fp_status);
61     i |= get_float_exception_flags(&env->vfp.standard_fp_status);
62     /* FZ16 does not generate an input denormal exception.  */
63     i |= (get_float_exception_flags(&env->vfp.fp_status_f16)
64           & ~float_flag_input_denormal);
65     fpscr |= vfp_exceptbits_from_host(i);
66 
67     i = env->vfp.qc[0] | env->vfp.qc[1] | env->vfp.qc[2] | env->vfp.qc[3];
68     fpscr |= i ? FPCR_QC : 0;
69 
70     return fpscr;
71 }
72 
73 uint32_t vfp_get_fpscr(CPUARMState *env)
74 {
75     return HELPER(vfp_get_fpscr)(env);
76 }
77 
78 /* Convert vfp exception flags to target form.  */
79 static inline int vfp_exceptbits_to_host(int target_bits)
80 {
81     int host_bits = 0;
82 
83     if (target_bits & 1)
84         host_bits |= float_flag_invalid;
85     if (target_bits & 2)
86         host_bits |= float_flag_divbyzero;
87     if (target_bits & 4)
88         host_bits |= float_flag_overflow;
89     if (target_bits & 8)
90         host_bits |= float_flag_underflow;
91     if (target_bits & 0x10)
92         host_bits |= float_flag_inexact;
93     if (target_bits & 0x80)
94         host_bits |= float_flag_input_denormal;
95     return host_bits;
96 }
97 
98 void HELPER(vfp_set_fpscr)(CPUARMState *env, uint32_t val)
99 {
100     int i;
101     uint32_t changed = env->vfp.xregs[ARM_VFP_FPSCR];
102 
103     /* When ARMv8.2-FP16 is not supported, FZ16 is RES0.  */
104     if (!cpu_isar_feature(aa64_fp16, arm_env_get_cpu(env))) {
105         val &= ~FPCR_FZ16;
106     }
107 
108     /*
109      * We don't implement trapped exception handling, so the
110      * trap enable bits, IDE|IXE|UFE|OFE|DZE|IOE are all RAZ/WI (not RES0!)
111      *
112      * If we exclude the exception flags, IOC|DZC|OFC|UFC|IXC|IDC
113      * (which are stored in fp_status), and the other RES0 bits
114      * in between, then we clear all of the low 16 bits.
115      */
116     env->vfp.xregs[ARM_VFP_FPSCR] = val & 0xf7c80000;
117     env->vfp.vec_len = (val >> 16) & 7;
118     env->vfp.vec_stride = (val >> 20) & 3;
119 
120     /*
121      * The bit we set within fpscr_q is arbitrary; the register as a
122      * whole being zero/non-zero is what counts.
123      */
124     env->vfp.qc[0] = val & FPCR_QC;
125     env->vfp.qc[1] = 0;
126     env->vfp.qc[2] = 0;
127     env->vfp.qc[3] = 0;
128 
129     changed ^= val;
130     if (changed & (3 << 22)) {
131         i = (val >> 22) & 3;
132         switch (i) {
133         case FPROUNDING_TIEEVEN:
134             i = float_round_nearest_even;
135             break;
136         case FPROUNDING_POSINF:
137             i = float_round_up;
138             break;
139         case FPROUNDING_NEGINF:
140             i = float_round_down;
141             break;
142         case FPROUNDING_ZERO:
143             i = float_round_to_zero;
144             break;
145         }
146         set_float_rounding_mode(i, &env->vfp.fp_status);
147         set_float_rounding_mode(i, &env->vfp.fp_status_f16);
148     }
149     if (changed & FPCR_FZ16) {
150         bool ftz_enabled = val & FPCR_FZ16;
151         set_flush_to_zero(ftz_enabled, &env->vfp.fp_status_f16);
152         set_flush_inputs_to_zero(ftz_enabled, &env->vfp.fp_status_f16);
153     }
154     if (changed & FPCR_FZ) {
155         bool ftz_enabled = val & FPCR_FZ;
156         set_flush_to_zero(ftz_enabled, &env->vfp.fp_status);
157         set_flush_inputs_to_zero(ftz_enabled, &env->vfp.fp_status);
158     }
159     if (changed & FPCR_DN) {
160         bool dnan_enabled = val & FPCR_DN;
161         set_default_nan_mode(dnan_enabled, &env->vfp.fp_status);
162         set_default_nan_mode(dnan_enabled, &env->vfp.fp_status_f16);
163     }
164 
165     /* The exception flags are ORed together when we read fpscr so we
166      * only need to preserve the current state in one of our
167      * float_status values.
168      */
169     i = vfp_exceptbits_to_host(val);
170     set_float_exception_flags(i, &env->vfp.fp_status);
171     set_float_exception_flags(0, &env->vfp.fp_status_f16);
172     set_float_exception_flags(0, &env->vfp.standard_fp_status);
173 }
174 
175 void vfp_set_fpscr(CPUARMState *env, uint32_t val)
176 {
177     HELPER(vfp_set_fpscr)(env, val);
178 }
179 
180 #define VFP_HELPER(name, p) HELPER(glue(glue(vfp_,name),p))
181 
182 #define VFP_BINOP(name) \
183 float32 VFP_HELPER(name, s)(float32 a, float32 b, void *fpstp) \
184 { \
185     float_status *fpst = fpstp; \
186     return float32_ ## name(a, b, fpst); \
187 } \
188 float64 VFP_HELPER(name, d)(float64 a, float64 b, void *fpstp) \
189 { \
190     float_status *fpst = fpstp; \
191     return float64_ ## name(a, b, fpst); \
192 }
193 VFP_BINOP(add)
194 VFP_BINOP(sub)
195 VFP_BINOP(mul)
196 VFP_BINOP(div)
197 VFP_BINOP(min)
198 VFP_BINOP(max)
199 VFP_BINOP(minnum)
200 VFP_BINOP(maxnum)
201 #undef VFP_BINOP
202 
203 float32 VFP_HELPER(neg, s)(float32 a)
204 {
205     return float32_chs(a);
206 }
207 
208 float64 VFP_HELPER(neg, d)(float64 a)
209 {
210     return float64_chs(a);
211 }
212 
213 float32 VFP_HELPER(abs, s)(float32 a)
214 {
215     return float32_abs(a);
216 }
217 
218 float64 VFP_HELPER(abs, d)(float64 a)
219 {
220     return float64_abs(a);
221 }
222 
223 float32 VFP_HELPER(sqrt, s)(float32 a, CPUARMState *env)
224 {
225     return float32_sqrt(a, &env->vfp.fp_status);
226 }
227 
228 float64 VFP_HELPER(sqrt, d)(float64 a, CPUARMState *env)
229 {
230     return float64_sqrt(a, &env->vfp.fp_status);
231 }
232 
233 static void softfloat_to_vfp_compare(CPUARMState *env, int cmp)
234 {
235     uint32_t flags;
236     switch (cmp) {
237     case float_relation_equal:
238         flags = 0x6;
239         break;
240     case float_relation_less:
241         flags = 0x8;
242         break;
243     case float_relation_greater:
244         flags = 0x2;
245         break;
246     case float_relation_unordered:
247         flags = 0x3;
248         break;
249     default:
250         g_assert_not_reached();
251     }
252     env->vfp.xregs[ARM_VFP_FPSCR] =
253         deposit32(env->vfp.xregs[ARM_VFP_FPSCR], 28, 4, flags);
254 }
255 
256 /* XXX: check quiet/signaling case */
257 #define DO_VFP_cmp(p, type) \
258 void VFP_HELPER(cmp, p)(type a, type b, CPUARMState *env)  \
259 { \
260     softfloat_to_vfp_compare(env, \
261         type ## _compare_quiet(a, b, &env->vfp.fp_status)); \
262 } \
263 void VFP_HELPER(cmpe, p)(type a, type b, CPUARMState *env) \
264 { \
265     softfloat_to_vfp_compare(env, \
266         type ## _compare(a, b, &env->vfp.fp_status)); \
267 }
268 DO_VFP_cmp(s, float32)
269 DO_VFP_cmp(d, float64)
270 #undef DO_VFP_cmp
271 
272 /* Integer to float and float to integer conversions */
273 
274 #define CONV_ITOF(name, ftype, fsz, sign)                           \
275 ftype HELPER(name)(uint32_t x, void *fpstp)                         \
276 {                                                                   \
277     float_status *fpst = fpstp;                                     \
278     return sign##int32_to_##float##fsz((sign##int32_t)x, fpst);     \
279 }
280 
281 #define CONV_FTOI(name, ftype, fsz, sign, round)                \
282 sign##int32_t HELPER(name)(ftype x, void *fpstp)                \
283 {                                                               \
284     float_status *fpst = fpstp;                                 \
285     if (float##fsz##_is_any_nan(x)) {                           \
286         float_raise(float_flag_invalid, fpst);                  \
287         return 0;                                               \
288     }                                                           \
289     return float##fsz##_to_##sign##int32##round(x, fpst);       \
290 }
291 
292 #define FLOAT_CONVS(name, p, ftype, fsz, sign)            \
293     CONV_ITOF(vfp_##name##to##p, ftype, fsz, sign)        \
294     CONV_FTOI(vfp_to##name##p, ftype, fsz, sign, )        \
295     CONV_FTOI(vfp_to##name##z##p, ftype, fsz, sign, _round_to_zero)
296 
297 FLOAT_CONVS(si, h, uint32_t, 16, )
298 FLOAT_CONVS(si, s, float32, 32, )
299 FLOAT_CONVS(si, d, float64, 64, )
300 FLOAT_CONVS(ui, h, uint32_t, 16, u)
301 FLOAT_CONVS(ui, s, float32, 32, u)
302 FLOAT_CONVS(ui, d, float64, 64, u)
303 
304 #undef CONV_ITOF
305 #undef CONV_FTOI
306 #undef FLOAT_CONVS
307 
308 /* floating point conversion */
309 float64 VFP_HELPER(fcvtd, s)(float32 x, CPUARMState *env)
310 {
311     return float32_to_float64(x, &env->vfp.fp_status);
312 }
313 
314 float32 VFP_HELPER(fcvts, d)(float64 x, CPUARMState *env)
315 {
316     return float64_to_float32(x, &env->vfp.fp_status);
317 }
318 
319 /* VFP3 fixed point conversion.  */
320 #define VFP_CONV_FIX_FLOAT(name, p, fsz, isz, itype) \
321 float##fsz HELPER(vfp_##name##to##p)(uint##isz##_t  x, uint32_t shift, \
322                                      void *fpstp) \
323 { return itype##_to_##float##fsz##_scalbn(x, -shift, fpstp); }
324 
325 #define VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, isz, itype, ROUND, suff)   \
326 uint##isz##_t HELPER(vfp_to##name##p##suff)(float##fsz x, uint32_t shift, \
327                                             void *fpst)                   \
328 {                                                                         \
329     if (unlikely(float##fsz##_is_any_nan(x))) {                           \
330         float_raise(float_flag_invalid, fpst);                            \
331         return 0;                                                         \
332     }                                                                     \
333     return float##fsz##_to_##itype##_scalbn(x, ROUND, shift, fpst);       \
334 }
335 
336 #define VFP_CONV_FIX(name, p, fsz, isz, itype)                   \
337 VFP_CONV_FIX_FLOAT(name, p, fsz, isz, itype)                     \
338 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, isz, itype,               \
339                          float_round_to_zero, _round_to_zero)    \
340 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, isz, itype,               \
341                          get_float_rounding_mode(fpst), )
342 
343 #define VFP_CONV_FIX_A64(name, p, fsz, isz, itype)               \
344 VFP_CONV_FIX_FLOAT(name, p, fsz, isz, itype)                     \
345 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, isz, itype,               \
346                          get_float_rounding_mode(fpst), )
347 
348 VFP_CONV_FIX(sh, d, 64, 64, int16)
349 VFP_CONV_FIX(sl, d, 64, 64, int32)
350 VFP_CONV_FIX_A64(sq, d, 64, 64, int64)
351 VFP_CONV_FIX(uh, d, 64, 64, uint16)
352 VFP_CONV_FIX(ul, d, 64, 64, uint32)
353 VFP_CONV_FIX_A64(uq, d, 64, 64, uint64)
354 VFP_CONV_FIX(sh, s, 32, 32, int16)
355 VFP_CONV_FIX(sl, s, 32, 32, int32)
356 VFP_CONV_FIX_A64(sq, s, 32, 64, int64)
357 VFP_CONV_FIX(uh, s, 32, 32, uint16)
358 VFP_CONV_FIX(ul, s, 32, 32, uint32)
359 VFP_CONV_FIX_A64(uq, s, 32, 64, uint64)
360 
361 #undef VFP_CONV_FIX
362 #undef VFP_CONV_FIX_FLOAT
363 #undef VFP_CONV_FLOAT_FIX_ROUND
364 #undef VFP_CONV_FIX_A64
365 
366 uint32_t HELPER(vfp_sltoh)(uint32_t x, uint32_t shift, void *fpst)
367 {
368     return int32_to_float16_scalbn(x, -shift, fpst);
369 }
370 
371 uint32_t HELPER(vfp_ultoh)(uint32_t x, uint32_t shift, void *fpst)
372 {
373     return uint32_to_float16_scalbn(x, -shift, fpst);
374 }
375 
376 uint32_t HELPER(vfp_sqtoh)(uint64_t x, uint32_t shift, void *fpst)
377 {
378     return int64_to_float16_scalbn(x, -shift, fpst);
379 }
380 
381 uint32_t HELPER(vfp_uqtoh)(uint64_t x, uint32_t shift, void *fpst)
382 {
383     return uint64_to_float16_scalbn(x, -shift, fpst);
384 }
385 
386 uint32_t HELPER(vfp_toshh)(uint32_t x, uint32_t shift, void *fpst)
387 {
388     if (unlikely(float16_is_any_nan(x))) {
389         float_raise(float_flag_invalid, fpst);
390         return 0;
391     }
392     return float16_to_int16_scalbn(x, get_float_rounding_mode(fpst),
393                                    shift, fpst);
394 }
395 
396 uint32_t HELPER(vfp_touhh)(uint32_t x, uint32_t shift, void *fpst)
397 {
398     if (unlikely(float16_is_any_nan(x))) {
399         float_raise(float_flag_invalid, fpst);
400         return 0;
401     }
402     return float16_to_uint16_scalbn(x, get_float_rounding_mode(fpst),
403                                     shift, fpst);
404 }
405 
406 uint32_t HELPER(vfp_toslh)(uint32_t x, uint32_t shift, void *fpst)
407 {
408     if (unlikely(float16_is_any_nan(x))) {
409         float_raise(float_flag_invalid, fpst);
410         return 0;
411     }
412     return float16_to_int32_scalbn(x, get_float_rounding_mode(fpst),
413                                    shift, fpst);
414 }
415 
416 uint32_t HELPER(vfp_toulh)(uint32_t x, uint32_t shift, void *fpst)
417 {
418     if (unlikely(float16_is_any_nan(x))) {
419         float_raise(float_flag_invalid, fpst);
420         return 0;
421     }
422     return float16_to_uint32_scalbn(x, get_float_rounding_mode(fpst),
423                                     shift, fpst);
424 }
425 
426 uint64_t HELPER(vfp_tosqh)(uint32_t x, uint32_t shift, void *fpst)
427 {
428     if (unlikely(float16_is_any_nan(x))) {
429         float_raise(float_flag_invalid, fpst);
430         return 0;
431     }
432     return float16_to_int64_scalbn(x, get_float_rounding_mode(fpst),
433                                    shift, fpst);
434 }
435 
436 uint64_t HELPER(vfp_touqh)(uint32_t x, uint32_t shift, void *fpst)
437 {
438     if (unlikely(float16_is_any_nan(x))) {
439         float_raise(float_flag_invalid, fpst);
440         return 0;
441     }
442     return float16_to_uint64_scalbn(x, get_float_rounding_mode(fpst),
443                                     shift, fpst);
444 }
445 
446 /* Set the current fp rounding mode and return the old one.
447  * The argument is a softfloat float_round_ value.
448  */
449 uint32_t HELPER(set_rmode)(uint32_t rmode, void *fpstp)
450 {
451     float_status *fp_status = fpstp;
452 
453     uint32_t prev_rmode = get_float_rounding_mode(fp_status);
454     set_float_rounding_mode(rmode, fp_status);
455 
456     return prev_rmode;
457 }
458 
459 /* Set the current fp rounding mode in the standard fp status and return
460  * the old one. This is for NEON instructions that need to change the
461  * rounding mode but wish to use the standard FPSCR values for everything
462  * else. Always set the rounding mode back to the correct value after
463  * modifying it.
464  * The argument is a softfloat float_round_ value.
465  */
466 uint32_t HELPER(set_neon_rmode)(uint32_t rmode, CPUARMState *env)
467 {
468     float_status *fp_status = &env->vfp.standard_fp_status;
469 
470     uint32_t prev_rmode = get_float_rounding_mode(fp_status);
471     set_float_rounding_mode(rmode, fp_status);
472 
473     return prev_rmode;
474 }
475 
476 /* Half precision conversions.  */
477 float32 HELPER(vfp_fcvt_f16_to_f32)(uint32_t a, void *fpstp, uint32_t ahp_mode)
478 {
479     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
480      * it would affect flushing input denormals.
481      */
482     float_status *fpst = fpstp;
483     flag save = get_flush_inputs_to_zero(fpst);
484     set_flush_inputs_to_zero(false, fpst);
485     float32 r = float16_to_float32(a, !ahp_mode, fpst);
486     set_flush_inputs_to_zero(save, fpst);
487     return r;
488 }
489 
490 uint32_t HELPER(vfp_fcvt_f32_to_f16)(float32 a, void *fpstp, uint32_t ahp_mode)
491 {
492     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
493      * it would affect flushing output denormals.
494      */
495     float_status *fpst = fpstp;
496     flag save = get_flush_to_zero(fpst);
497     set_flush_to_zero(false, fpst);
498     float16 r = float32_to_float16(a, !ahp_mode, fpst);
499     set_flush_to_zero(save, fpst);
500     return r;
501 }
502 
503 float64 HELPER(vfp_fcvt_f16_to_f64)(uint32_t a, void *fpstp, uint32_t ahp_mode)
504 {
505     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
506      * it would affect flushing input denormals.
507      */
508     float_status *fpst = fpstp;
509     flag save = get_flush_inputs_to_zero(fpst);
510     set_flush_inputs_to_zero(false, fpst);
511     float64 r = float16_to_float64(a, !ahp_mode, fpst);
512     set_flush_inputs_to_zero(save, fpst);
513     return r;
514 }
515 
516 uint32_t HELPER(vfp_fcvt_f64_to_f16)(float64 a, void *fpstp, uint32_t ahp_mode)
517 {
518     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
519      * it would affect flushing output denormals.
520      */
521     float_status *fpst = fpstp;
522     flag save = get_flush_to_zero(fpst);
523     set_flush_to_zero(false, fpst);
524     float16 r = float64_to_float16(a, !ahp_mode, fpst);
525     set_flush_to_zero(save, fpst);
526     return r;
527 }
528 
529 #define float32_two make_float32(0x40000000)
530 #define float32_three make_float32(0x40400000)
531 #define float32_one_point_five make_float32(0x3fc00000)
532 
533 float32 HELPER(recps_f32)(float32 a, float32 b, CPUARMState *env)
534 {
535     float_status *s = &env->vfp.standard_fp_status;
536     if ((float32_is_infinity(a) && float32_is_zero_or_denormal(b)) ||
537         (float32_is_infinity(b) && float32_is_zero_or_denormal(a))) {
538         if (!(float32_is_zero(a) || float32_is_zero(b))) {
539             float_raise(float_flag_input_denormal, s);
540         }
541         return float32_two;
542     }
543     return float32_sub(float32_two, float32_mul(a, b, s), s);
544 }
545 
546 float32 HELPER(rsqrts_f32)(float32 a, float32 b, CPUARMState *env)
547 {
548     float_status *s = &env->vfp.standard_fp_status;
549     float32 product;
550     if ((float32_is_infinity(a) && float32_is_zero_or_denormal(b)) ||
551         (float32_is_infinity(b) && float32_is_zero_or_denormal(a))) {
552         if (!(float32_is_zero(a) || float32_is_zero(b))) {
553             float_raise(float_flag_input_denormal, s);
554         }
555         return float32_one_point_five;
556     }
557     product = float32_mul(a, b, s);
558     return float32_div(float32_sub(float32_three, product, s), float32_two, s);
559 }
560 
561 /* NEON helpers.  */
562 
563 /* Constants 256 and 512 are used in some helpers; we avoid relying on
564  * int->float conversions at run-time.  */
565 #define float64_256 make_float64(0x4070000000000000LL)
566 #define float64_512 make_float64(0x4080000000000000LL)
567 #define float16_maxnorm make_float16(0x7bff)
568 #define float32_maxnorm make_float32(0x7f7fffff)
569 #define float64_maxnorm make_float64(0x7fefffffffffffffLL)
570 
571 /* Reciprocal functions
572  *
573  * The algorithm that must be used to calculate the estimate
574  * is specified by the ARM ARM, see FPRecipEstimate()/RecipEstimate
575  */
576 
577 /* See RecipEstimate()
578  *
579  * input is a 9 bit fixed point number
580  * input range 256 .. 511 for a number from 0.5 <= x < 1.0.
581  * result range 256 .. 511 for a number from 1.0 to 511/256.
582  */
583 
584 static int recip_estimate(int input)
585 {
586     int a, b, r;
587     assert(256 <= input && input < 512);
588     a = (input * 2) + 1;
589     b = (1 << 19) / a;
590     r = (b + 1) >> 1;
591     assert(256 <= r && r < 512);
592     return r;
593 }
594 
595 /*
596  * Common wrapper to call recip_estimate
597  *
598  * The parameters are exponent and 64 bit fraction (without implicit
599  * bit) where the binary point is nominally at bit 52. Returns a
600  * float64 which can then be rounded to the appropriate size by the
601  * callee.
602  */
603 
604 static uint64_t call_recip_estimate(int *exp, int exp_off, uint64_t frac)
605 {
606     uint32_t scaled, estimate;
607     uint64_t result_frac;
608     int result_exp;
609 
610     /* Handle sub-normals */
611     if (*exp == 0) {
612         if (extract64(frac, 51, 1) == 0) {
613             *exp = -1;
614             frac <<= 2;
615         } else {
616             frac <<= 1;
617         }
618     }
619 
620     /* scaled = UInt('1':fraction<51:44>) */
621     scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
622     estimate = recip_estimate(scaled);
623 
624     result_exp = exp_off - *exp;
625     result_frac = deposit64(0, 44, 8, estimate);
626     if (result_exp == 0) {
627         result_frac = deposit64(result_frac >> 1, 51, 1, 1);
628     } else if (result_exp == -1) {
629         result_frac = deposit64(result_frac >> 2, 50, 2, 1);
630         result_exp = 0;
631     }
632 
633     *exp = result_exp;
634 
635     return result_frac;
636 }
637 
638 static bool round_to_inf(float_status *fpst, bool sign_bit)
639 {
640     switch (fpst->float_rounding_mode) {
641     case float_round_nearest_even: /* Round to Nearest */
642         return true;
643     case float_round_up: /* Round to +Inf */
644         return !sign_bit;
645     case float_round_down: /* Round to -Inf */
646         return sign_bit;
647     case float_round_to_zero: /* Round to Zero */
648         return false;
649     }
650 
651     g_assert_not_reached();
652 }
653 
654 uint32_t HELPER(recpe_f16)(uint32_t input, void *fpstp)
655 {
656     float_status *fpst = fpstp;
657     float16 f16 = float16_squash_input_denormal(input, fpst);
658     uint32_t f16_val = float16_val(f16);
659     uint32_t f16_sign = float16_is_neg(f16);
660     int f16_exp = extract32(f16_val, 10, 5);
661     uint32_t f16_frac = extract32(f16_val, 0, 10);
662     uint64_t f64_frac;
663 
664     if (float16_is_any_nan(f16)) {
665         float16 nan = f16;
666         if (float16_is_signaling_nan(f16, fpst)) {
667             float_raise(float_flag_invalid, fpst);
668             nan = float16_silence_nan(f16, fpst);
669         }
670         if (fpst->default_nan_mode) {
671             nan =  float16_default_nan(fpst);
672         }
673         return nan;
674     } else if (float16_is_infinity(f16)) {
675         return float16_set_sign(float16_zero, float16_is_neg(f16));
676     } else if (float16_is_zero(f16)) {
677         float_raise(float_flag_divbyzero, fpst);
678         return float16_set_sign(float16_infinity, float16_is_neg(f16));
679     } else if (float16_abs(f16) < (1 << 8)) {
680         /* Abs(value) < 2.0^-16 */
681         float_raise(float_flag_overflow | float_flag_inexact, fpst);
682         if (round_to_inf(fpst, f16_sign)) {
683             return float16_set_sign(float16_infinity, f16_sign);
684         } else {
685             return float16_set_sign(float16_maxnorm, f16_sign);
686         }
687     } else if (f16_exp >= 29 && fpst->flush_to_zero) {
688         float_raise(float_flag_underflow, fpst);
689         return float16_set_sign(float16_zero, float16_is_neg(f16));
690     }
691 
692     f64_frac = call_recip_estimate(&f16_exp, 29,
693                                    ((uint64_t) f16_frac) << (52 - 10));
694 
695     /* result = sign : result_exp<4:0> : fraction<51:42> */
696     f16_val = deposit32(0, 15, 1, f16_sign);
697     f16_val = deposit32(f16_val, 10, 5, f16_exp);
698     f16_val = deposit32(f16_val, 0, 10, extract64(f64_frac, 52 - 10, 10));
699     return make_float16(f16_val);
700 }
701 
702 float32 HELPER(recpe_f32)(float32 input, void *fpstp)
703 {
704     float_status *fpst = fpstp;
705     float32 f32 = float32_squash_input_denormal(input, fpst);
706     uint32_t f32_val = float32_val(f32);
707     bool f32_sign = float32_is_neg(f32);
708     int f32_exp = extract32(f32_val, 23, 8);
709     uint32_t f32_frac = extract32(f32_val, 0, 23);
710     uint64_t f64_frac;
711 
712     if (float32_is_any_nan(f32)) {
713         float32 nan = f32;
714         if (float32_is_signaling_nan(f32, fpst)) {
715             float_raise(float_flag_invalid, fpst);
716             nan = float32_silence_nan(f32, fpst);
717         }
718         if (fpst->default_nan_mode) {
719             nan =  float32_default_nan(fpst);
720         }
721         return nan;
722     } else if (float32_is_infinity(f32)) {
723         return float32_set_sign(float32_zero, float32_is_neg(f32));
724     } else if (float32_is_zero(f32)) {
725         float_raise(float_flag_divbyzero, fpst);
726         return float32_set_sign(float32_infinity, float32_is_neg(f32));
727     } else if (float32_abs(f32) < (1ULL << 21)) {
728         /* Abs(value) < 2.0^-128 */
729         float_raise(float_flag_overflow | float_flag_inexact, fpst);
730         if (round_to_inf(fpst, f32_sign)) {
731             return float32_set_sign(float32_infinity, f32_sign);
732         } else {
733             return float32_set_sign(float32_maxnorm, f32_sign);
734         }
735     } else if (f32_exp >= 253 && fpst->flush_to_zero) {
736         float_raise(float_flag_underflow, fpst);
737         return float32_set_sign(float32_zero, float32_is_neg(f32));
738     }
739 
740     f64_frac = call_recip_estimate(&f32_exp, 253,
741                                    ((uint64_t) f32_frac) << (52 - 23));
742 
743     /* result = sign : result_exp<7:0> : fraction<51:29> */
744     f32_val = deposit32(0, 31, 1, f32_sign);
745     f32_val = deposit32(f32_val, 23, 8, f32_exp);
746     f32_val = deposit32(f32_val, 0, 23, extract64(f64_frac, 52 - 23, 23));
747     return make_float32(f32_val);
748 }
749 
750 float64 HELPER(recpe_f64)(float64 input, void *fpstp)
751 {
752     float_status *fpst = fpstp;
753     float64 f64 = float64_squash_input_denormal(input, fpst);
754     uint64_t f64_val = float64_val(f64);
755     bool f64_sign = float64_is_neg(f64);
756     int f64_exp = extract64(f64_val, 52, 11);
757     uint64_t f64_frac = extract64(f64_val, 0, 52);
758 
759     /* Deal with any special cases */
760     if (float64_is_any_nan(f64)) {
761         float64 nan = f64;
762         if (float64_is_signaling_nan(f64, fpst)) {
763             float_raise(float_flag_invalid, fpst);
764             nan = float64_silence_nan(f64, fpst);
765         }
766         if (fpst->default_nan_mode) {
767             nan =  float64_default_nan(fpst);
768         }
769         return nan;
770     } else if (float64_is_infinity(f64)) {
771         return float64_set_sign(float64_zero, float64_is_neg(f64));
772     } else if (float64_is_zero(f64)) {
773         float_raise(float_flag_divbyzero, fpst);
774         return float64_set_sign(float64_infinity, float64_is_neg(f64));
775     } else if ((f64_val & ~(1ULL << 63)) < (1ULL << 50)) {
776         /* Abs(value) < 2.0^-1024 */
777         float_raise(float_flag_overflow | float_flag_inexact, fpst);
778         if (round_to_inf(fpst, f64_sign)) {
779             return float64_set_sign(float64_infinity, f64_sign);
780         } else {
781             return float64_set_sign(float64_maxnorm, f64_sign);
782         }
783     } else if (f64_exp >= 2045 && fpst->flush_to_zero) {
784         float_raise(float_flag_underflow, fpst);
785         return float64_set_sign(float64_zero, float64_is_neg(f64));
786     }
787 
788     f64_frac = call_recip_estimate(&f64_exp, 2045, f64_frac);
789 
790     /* result = sign : result_exp<10:0> : fraction<51:0>; */
791     f64_val = deposit64(0, 63, 1, f64_sign);
792     f64_val = deposit64(f64_val, 52, 11, f64_exp);
793     f64_val = deposit64(f64_val, 0, 52, f64_frac);
794     return make_float64(f64_val);
795 }
796 
797 /* The algorithm that must be used to calculate the estimate
798  * is specified by the ARM ARM.
799  */
800 
801 static int do_recip_sqrt_estimate(int a)
802 {
803     int b, estimate;
804 
805     assert(128 <= a && a < 512);
806     if (a < 256) {
807         a = a * 2 + 1;
808     } else {
809         a = (a >> 1) << 1;
810         a = (a + 1) * 2;
811     }
812     b = 512;
813     while (a * (b + 1) * (b + 1) < (1 << 28)) {
814         b += 1;
815     }
816     estimate = (b + 1) / 2;
817     assert(256 <= estimate && estimate < 512);
818 
819     return estimate;
820 }
821 
822 
823 static uint64_t recip_sqrt_estimate(int *exp , int exp_off, uint64_t frac)
824 {
825     int estimate;
826     uint32_t scaled;
827 
828     if (*exp == 0) {
829         while (extract64(frac, 51, 1) == 0) {
830             frac = frac << 1;
831             *exp -= 1;
832         }
833         frac = extract64(frac, 0, 51) << 1;
834     }
835 
836     if (*exp & 1) {
837         /* scaled = UInt('01':fraction<51:45>) */
838         scaled = deposit32(1 << 7, 0, 7, extract64(frac, 45, 7));
839     } else {
840         /* scaled = UInt('1':fraction<51:44>) */
841         scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
842     }
843     estimate = do_recip_sqrt_estimate(scaled);
844 
845     *exp = (exp_off - *exp) / 2;
846     return extract64(estimate, 0, 8) << 44;
847 }
848 
849 uint32_t HELPER(rsqrte_f16)(uint32_t input, void *fpstp)
850 {
851     float_status *s = fpstp;
852     float16 f16 = float16_squash_input_denormal(input, s);
853     uint16_t val = float16_val(f16);
854     bool f16_sign = float16_is_neg(f16);
855     int f16_exp = extract32(val, 10, 5);
856     uint16_t f16_frac = extract32(val, 0, 10);
857     uint64_t f64_frac;
858 
859     if (float16_is_any_nan(f16)) {
860         float16 nan = f16;
861         if (float16_is_signaling_nan(f16, s)) {
862             float_raise(float_flag_invalid, s);
863             nan = float16_silence_nan(f16, s);
864         }
865         if (s->default_nan_mode) {
866             nan =  float16_default_nan(s);
867         }
868         return nan;
869     } else if (float16_is_zero(f16)) {
870         float_raise(float_flag_divbyzero, s);
871         return float16_set_sign(float16_infinity, f16_sign);
872     } else if (f16_sign) {
873         float_raise(float_flag_invalid, s);
874         return float16_default_nan(s);
875     } else if (float16_is_infinity(f16)) {
876         return float16_zero;
877     }
878 
879     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
880      * preserving the parity of the exponent.  */
881 
882     f64_frac = ((uint64_t) f16_frac) << (52 - 10);
883 
884     f64_frac = recip_sqrt_estimate(&f16_exp, 44, f64_frac);
885 
886     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(2) */
887     val = deposit32(0, 15, 1, f16_sign);
888     val = deposit32(val, 10, 5, f16_exp);
889     val = deposit32(val, 2, 8, extract64(f64_frac, 52 - 8, 8));
890     return make_float16(val);
891 }
892 
893 float32 HELPER(rsqrte_f32)(float32 input, void *fpstp)
894 {
895     float_status *s = fpstp;
896     float32 f32 = float32_squash_input_denormal(input, s);
897     uint32_t val = float32_val(f32);
898     uint32_t f32_sign = float32_is_neg(f32);
899     int f32_exp = extract32(val, 23, 8);
900     uint32_t f32_frac = extract32(val, 0, 23);
901     uint64_t f64_frac;
902 
903     if (float32_is_any_nan(f32)) {
904         float32 nan = f32;
905         if (float32_is_signaling_nan(f32, s)) {
906             float_raise(float_flag_invalid, s);
907             nan = float32_silence_nan(f32, s);
908         }
909         if (s->default_nan_mode) {
910             nan =  float32_default_nan(s);
911         }
912         return nan;
913     } else if (float32_is_zero(f32)) {
914         float_raise(float_flag_divbyzero, s);
915         return float32_set_sign(float32_infinity, float32_is_neg(f32));
916     } else if (float32_is_neg(f32)) {
917         float_raise(float_flag_invalid, s);
918         return float32_default_nan(s);
919     } else if (float32_is_infinity(f32)) {
920         return float32_zero;
921     }
922 
923     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
924      * preserving the parity of the exponent.  */
925 
926     f64_frac = ((uint64_t) f32_frac) << 29;
927 
928     f64_frac = recip_sqrt_estimate(&f32_exp, 380, f64_frac);
929 
930     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(15) */
931     val = deposit32(0, 31, 1, f32_sign);
932     val = deposit32(val, 23, 8, f32_exp);
933     val = deposit32(val, 15, 8, extract64(f64_frac, 52 - 8, 8));
934     return make_float32(val);
935 }
936 
937 float64 HELPER(rsqrte_f64)(float64 input, void *fpstp)
938 {
939     float_status *s = fpstp;
940     float64 f64 = float64_squash_input_denormal(input, s);
941     uint64_t val = float64_val(f64);
942     bool f64_sign = float64_is_neg(f64);
943     int f64_exp = extract64(val, 52, 11);
944     uint64_t f64_frac = extract64(val, 0, 52);
945 
946     if (float64_is_any_nan(f64)) {
947         float64 nan = f64;
948         if (float64_is_signaling_nan(f64, s)) {
949             float_raise(float_flag_invalid, s);
950             nan = float64_silence_nan(f64, s);
951         }
952         if (s->default_nan_mode) {
953             nan =  float64_default_nan(s);
954         }
955         return nan;
956     } else if (float64_is_zero(f64)) {
957         float_raise(float_flag_divbyzero, s);
958         return float64_set_sign(float64_infinity, float64_is_neg(f64));
959     } else if (float64_is_neg(f64)) {
960         float_raise(float_flag_invalid, s);
961         return float64_default_nan(s);
962     } else if (float64_is_infinity(f64)) {
963         return float64_zero;
964     }
965 
966     f64_frac = recip_sqrt_estimate(&f64_exp, 3068, f64_frac);
967 
968     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(44) */
969     val = deposit64(0, 61, 1, f64_sign);
970     val = deposit64(val, 52, 11, f64_exp);
971     val = deposit64(val, 44, 8, extract64(f64_frac, 52 - 8, 8));
972     return make_float64(val);
973 }
974 
975 uint32_t HELPER(recpe_u32)(uint32_t a, void *fpstp)
976 {
977     /* float_status *s = fpstp; */
978     int input, estimate;
979 
980     if ((a & 0x80000000) == 0) {
981         return 0xffffffff;
982     }
983 
984     input = extract32(a, 23, 9);
985     estimate = recip_estimate(input);
986 
987     return deposit32(0, (32 - 9), 9, estimate);
988 }
989 
990 uint32_t HELPER(rsqrte_u32)(uint32_t a, void *fpstp)
991 {
992     int estimate;
993 
994     if ((a & 0xc0000000) == 0) {
995         return 0xffffffff;
996     }
997 
998     estimate = do_recip_sqrt_estimate(extract32(a, 23, 9));
999 
1000     return deposit32(0, 23, 9, estimate);
1001 }
1002 
1003 /* VFPv4 fused multiply-accumulate */
1004 float32 VFP_HELPER(muladd, s)(float32 a, float32 b, float32 c, void *fpstp)
1005 {
1006     float_status *fpst = fpstp;
1007     return float32_muladd(a, b, c, 0, fpst);
1008 }
1009 
1010 float64 VFP_HELPER(muladd, d)(float64 a, float64 b, float64 c, void *fpstp)
1011 {
1012     float_status *fpst = fpstp;
1013     return float64_muladd(a, b, c, 0, fpst);
1014 }
1015 
1016 /* ARMv8 round to integral */
1017 float32 HELPER(rints_exact)(float32 x, void *fp_status)
1018 {
1019     return float32_round_to_int(x, fp_status);
1020 }
1021 
1022 float64 HELPER(rintd_exact)(float64 x, void *fp_status)
1023 {
1024     return float64_round_to_int(x, fp_status);
1025 }
1026 
1027 float32 HELPER(rints)(float32 x, void *fp_status)
1028 {
1029     int old_flags = get_float_exception_flags(fp_status), new_flags;
1030     float32 ret;
1031 
1032     ret = float32_round_to_int(x, fp_status);
1033 
1034     /* Suppress any inexact exceptions the conversion produced */
1035     if (!(old_flags & float_flag_inexact)) {
1036         new_flags = get_float_exception_flags(fp_status);
1037         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1038     }
1039 
1040     return ret;
1041 }
1042 
1043 float64 HELPER(rintd)(float64 x, void *fp_status)
1044 {
1045     int old_flags = get_float_exception_flags(fp_status), new_flags;
1046     float64 ret;
1047 
1048     ret = float64_round_to_int(x, fp_status);
1049 
1050     new_flags = get_float_exception_flags(fp_status);
1051 
1052     /* Suppress any inexact exceptions the conversion produced */
1053     if (!(old_flags & float_flag_inexact)) {
1054         new_flags = get_float_exception_flags(fp_status);
1055         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1056     }
1057 
1058     return ret;
1059 }
1060 
1061 /* Convert ARM rounding mode to softfloat */
1062 int arm_rmode_to_sf(int rmode)
1063 {
1064     switch (rmode) {
1065     case FPROUNDING_TIEAWAY:
1066         rmode = float_round_ties_away;
1067         break;
1068     case FPROUNDING_ODD:
1069         /* FIXME: add support for TIEAWAY and ODD */
1070         qemu_log_mask(LOG_UNIMP, "arm: unimplemented rounding mode: %d\n",
1071                       rmode);
1072         /* fall through for now */
1073     case FPROUNDING_TIEEVEN:
1074     default:
1075         rmode = float_round_nearest_even;
1076         break;
1077     case FPROUNDING_POSINF:
1078         rmode = float_round_up;
1079         break;
1080     case FPROUNDING_NEGINF:
1081         rmode = float_round_down;
1082         break;
1083     case FPROUNDING_ZERO:
1084         rmode = float_round_to_zero;
1085         break;
1086     }
1087     return rmode;
1088 }
1089 
1090 /*
1091  * Implement float64 to int32_t conversion without saturation;
1092  * the result is supplied modulo 2^32.
1093  */
1094 uint64_t HELPER(fjcvtzs)(float64 value, void *vstatus)
1095 {
1096     float_status *status = vstatus;
1097     uint32_t exp, sign;
1098     uint64_t frac;
1099     uint32_t inexact = 1; /* !Z */
1100 
1101     sign = extract64(value, 63, 1);
1102     exp = extract64(value, 52, 11);
1103     frac = extract64(value, 0, 52);
1104 
1105     if (exp == 0) {
1106         /* While not inexact for IEEE FP, -0.0 is inexact for JavaScript.  */
1107         inexact = sign;
1108         if (frac != 0) {
1109             if (status->flush_inputs_to_zero) {
1110                 float_raise(float_flag_input_denormal, status);
1111             } else {
1112                 float_raise(float_flag_inexact, status);
1113                 inexact = 1;
1114             }
1115         }
1116         frac = 0;
1117     } else if (exp == 0x7ff) {
1118         /* This operation raises Invalid for both NaN and overflow (Inf).  */
1119         float_raise(float_flag_invalid, status);
1120         frac = 0;
1121     } else {
1122         int true_exp = exp - 1023;
1123         int shift = true_exp - 52;
1124 
1125         /* Restore implicit bit.  */
1126         frac |= 1ull << 52;
1127 
1128         /* Shift the fraction into place.  */
1129         if (shift >= 0) {
1130             /* The number is so large we must shift the fraction left.  */
1131             if (shift >= 64) {
1132                 /* The fraction is shifted out entirely.  */
1133                 frac = 0;
1134             } else {
1135                 frac <<= shift;
1136             }
1137         } else if (shift > -64) {
1138             /* Normal case -- shift right and notice if bits shift out.  */
1139             inexact = (frac << (64 + shift)) != 0;
1140             frac >>= -shift;
1141         } else {
1142             /* The fraction is shifted out entirely.  */
1143             frac = 0;
1144         }
1145 
1146         /* Notice overflow or inexact exceptions.  */
1147         if (true_exp > 31 || frac > (sign ? 0x80000000ull : 0x7fffffff)) {
1148             /* Overflow, for which this operation raises invalid.  */
1149             float_raise(float_flag_invalid, status);
1150             inexact = 1;
1151         } else if (inexact) {
1152             float_raise(float_flag_inexact, status);
1153         }
1154 
1155         /* Honor the sign.  */
1156         if (sign) {
1157             frac = -frac;
1158         }
1159     }
1160 
1161     /* Pack the result and the env->ZF representation of Z together.  */
1162     return deposit64(frac, 32, 32, inexact);
1163 }
1164 
1165 uint32_t HELPER(vjcvt)(float64 value, CPUARMState *env)
1166 {
1167     uint64_t pair = HELPER(fjcvtzs)(value, &env->vfp.fp_status);
1168     uint32_t result = pair;
1169     uint32_t z = (pair >> 32) == 0;
1170 
1171     /* Store Z, clear NCV, in FPSCR.NZCV.  */
1172     env->vfp.xregs[ARM_VFP_FPSCR]
1173         = (env->vfp.xregs[ARM_VFP_FPSCR] & ~CPSR_NZCV) | (z * CPSR_Z);
1174 
1175     return result;
1176 }
1177