xref: /openbmc/qemu/target/arm/vfp_helper.c (revision a6caeee8)
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 "cpu.h"
22 #include "exec/helper-proto.h"
23 #include "internals.h"
24 #ifdef CONFIG_TCG
25 #include "qemu/log.h"
26 #include "fpu/softfloat.h"
27 #endif
28 
29 /* VFP support.  We follow the convention used for VFP instructions:
30    Single precision routines have a "s" suffix, double precision a
31    "d" suffix.  */
32 
33 #ifdef CONFIG_TCG
34 
35 /* Convert host exception flags to vfp form.  */
36 static inline int vfp_exceptbits_from_host(int host_bits)
37 {
38     int target_bits = 0;
39 
40     if (host_bits & float_flag_invalid) {
41         target_bits |= 1;
42     }
43     if (host_bits & float_flag_divbyzero) {
44         target_bits |= 2;
45     }
46     if (host_bits & float_flag_overflow) {
47         target_bits |= 4;
48     }
49     if (host_bits & (float_flag_underflow | float_flag_output_denormal)) {
50         target_bits |= 8;
51     }
52     if (host_bits & float_flag_inexact) {
53         target_bits |= 0x10;
54     }
55     if (host_bits & float_flag_input_denormal) {
56         target_bits |= 0x80;
57     }
58     return target_bits;
59 }
60 
61 /* Convert vfp exception flags to target form.  */
62 static inline int vfp_exceptbits_to_host(int target_bits)
63 {
64     int host_bits = 0;
65 
66     if (target_bits & 1) {
67         host_bits |= float_flag_invalid;
68     }
69     if (target_bits & 2) {
70         host_bits |= float_flag_divbyzero;
71     }
72     if (target_bits & 4) {
73         host_bits |= float_flag_overflow;
74     }
75     if (target_bits & 8) {
76         host_bits |= float_flag_underflow;
77     }
78     if (target_bits & 0x10) {
79         host_bits |= float_flag_inexact;
80     }
81     if (target_bits & 0x80) {
82         host_bits |= float_flag_input_denormal;
83     }
84     return host_bits;
85 }
86 
87 static uint32_t vfp_get_fpscr_from_host(CPUARMState *env)
88 {
89     uint32_t i;
90 
91     i = get_float_exception_flags(&env->vfp.fp_status);
92     i |= get_float_exception_flags(&env->vfp.standard_fp_status);
93     /* FZ16 does not generate an input denormal exception.  */
94     i |= (get_float_exception_flags(&env->vfp.fp_status_f16)
95           & ~float_flag_input_denormal);
96     i |= (get_float_exception_flags(&env->vfp.standard_fp_status_f16)
97           & ~float_flag_input_denormal);
98     return vfp_exceptbits_from_host(i);
99 }
100 
101 static void vfp_set_fpscr_to_host(CPUARMState *env, uint32_t val)
102 {
103     int i;
104     uint32_t changed = env->vfp.xregs[ARM_VFP_FPSCR];
105 
106     changed ^= val;
107     if (changed & (3 << 22)) {
108         i = (val >> 22) & 3;
109         switch (i) {
110         case FPROUNDING_TIEEVEN:
111             i = float_round_nearest_even;
112             break;
113         case FPROUNDING_POSINF:
114             i = float_round_up;
115             break;
116         case FPROUNDING_NEGINF:
117             i = float_round_down;
118             break;
119         case FPROUNDING_ZERO:
120             i = float_round_to_zero;
121             break;
122         }
123         set_float_rounding_mode(i, &env->vfp.fp_status);
124         set_float_rounding_mode(i, &env->vfp.fp_status_f16);
125     }
126     if (changed & FPCR_FZ16) {
127         bool ftz_enabled = val & FPCR_FZ16;
128         set_flush_to_zero(ftz_enabled, &env->vfp.fp_status_f16);
129         set_flush_to_zero(ftz_enabled, &env->vfp.standard_fp_status_f16);
130         set_flush_inputs_to_zero(ftz_enabled, &env->vfp.fp_status_f16);
131         set_flush_inputs_to_zero(ftz_enabled, &env->vfp.standard_fp_status_f16);
132     }
133     if (changed & FPCR_FZ) {
134         bool ftz_enabled = val & FPCR_FZ;
135         set_flush_to_zero(ftz_enabled, &env->vfp.fp_status);
136         set_flush_inputs_to_zero(ftz_enabled, &env->vfp.fp_status);
137     }
138     if (changed & FPCR_DN) {
139         bool dnan_enabled = val & FPCR_DN;
140         set_default_nan_mode(dnan_enabled, &env->vfp.fp_status);
141         set_default_nan_mode(dnan_enabled, &env->vfp.fp_status_f16);
142     }
143 
144     /*
145      * The exception flags are ORed together when we read fpscr so we
146      * only need to preserve the current state in one of our
147      * float_status values.
148      */
149     i = vfp_exceptbits_to_host(val);
150     set_float_exception_flags(i, &env->vfp.fp_status);
151     set_float_exception_flags(0, &env->vfp.fp_status_f16);
152     set_float_exception_flags(0, &env->vfp.standard_fp_status);
153     set_float_exception_flags(0, &env->vfp.standard_fp_status_f16);
154 }
155 
156 #else
157 
158 static uint32_t vfp_get_fpscr_from_host(CPUARMState *env)
159 {
160     return 0;
161 }
162 
163 static void vfp_set_fpscr_to_host(CPUARMState *env, uint32_t val)
164 {
165 }
166 
167 #endif
168 
169 uint32_t HELPER(vfp_get_fpscr)(CPUARMState *env)
170 {
171     uint32_t i, fpscr;
172 
173     fpscr = env->vfp.xregs[ARM_VFP_FPSCR]
174             | (env->vfp.vec_len << 16)
175             | (env->vfp.vec_stride << 20);
176 
177     /*
178      * M-profile LTPSIZE overlaps A-profile Stride; whichever of the
179      * two is not applicable to this CPU will always be zero.
180      */
181     fpscr |= env->v7m.ltpsize << 16;
182 
183     fpscr |= vfp_get_fpscr_from_host(env);
184 
185     i = env->vfp.qc[0] | env->vfp.qc[1] | env->vfp.qc[2] | env->vfp.qc[3];
186     fpscr |= i ? FPCR_QC : 0;
187 
188     return fpscr;
189 }
190 
191 uint32_t vfp_get_fpscr(CPUARMState *env)
192 {
193     return HELPER(vfp_get_fpscr)(env);
194 }
195 
196 void HELPER(vfp_set_fpscr)(CPUARMState *env, uint32_t val)
197 {
198     ARMCPU *cpu = env_archcpu(env);
199 
200     /* When ARMv8.2-FP16 is not supported, FZ16 is RES0.  */
201     if (!cpu_isar_feature(any_fp16, cpu)) {
202         val &= ~FPCR_FZ16;
203     }
204 
205     vfp_set_fpscr_to_host(env, val);
206 
207     if (!arm_feature(env, ARM_FEATURE_M)) {
208         /*
209          * Short-vector length and stride; on M-profile these bits
210          * are used for different purposes.
211          * We can't make this conditional be "if MVFR0.FPShVec != 0",
212          * because in v7A no-short-vector-support cores still had to
213          * allow Stride/Len to be written with the only effect that
214          * some insns are required to UNDEF if the guest sets them.
215          */
216         env->vfp.vec_len = extract32(val, 16, 3);
217         env->vfp.vec_stride = extract32(val, 20, 2);
218     } else if (cpu_isar_feature(aa32_mve, cpu)) {
219         env->v7m.ltpsize = extract32(val, FPCR_LTPSIZE_SHIFT,
220                                      FPCR_LTPSIZE_LENGTH);
221     }
222 
223     if (arm_feature(env, ARM_FEATURE_NEON) ||
224         cpu_isar_feature(aa32_mve, cpu)) {
225         /*
226          * The bit we set within fpscr_q is arbitrary; the register as a
227          * whole being zero/non-zero is what counts.
228          * TODO: M-profile MVE also has a QC bit.
229          */
230         env->vfp.qc[0] = val & FPCR_QC;
231         env->vfp.qc[1] = 0;
232         env->vfp.qc[2] = 0;
233         env->vfp.qc[3] = 0;
234     }
235 
236     /*
237      * We don't implement trapped exception handling, so the
238      * trap enable bits, IDE|IXE|UFE|OFE|DZE|IOE are all RAZ/WI (not RES0!)
239      *
240      * The exception flags IOC|DZC|OFC|UFC|IXC|IDC are stored in
241      * fp_status; QC, Len and Stride are stored separately earlier.
242      * Clear out all of those and the RES0 bits: only NZCV, AHP, DN,
243      * FZ, RMode and FZ16 are kept in vfp.xregs[FPSCR].
244      */
245     env->vfp.xregs[ARM_VFP_FPSCR] = val & 0xf7c80000;
246 }
247 
248 void vfp_set_fpscr(CPUARMState *env, uint32_t val)
249 {
250     HELPER(vfp_set_fpscr)(env, val);
251 }
252 
253 #ifdef CONFIG_TCG
254 
255 #define VFP_HELPER(name, p) HELPER(glue(glue(vfp_,name),p))
256 
257 #define VFP_BINOP(name) \
258 dh_ctype_f16 VFP_HELPER(name, h)(dh_ctype_f16 a, dh_ctype_f16 b, void *fpstp) \
259 { \
260     float_status *fpst = fpstp; \
261     return float16_ ## name(a, b, fpst); \
262 } \
263 float32 VFP_HELPER(name, s)(float32 a, float32 b, void *fpstp) \
264 { \
265     float_status *fpst = fpstp; \
266     return float32_ ## name(a, b, fpst); \
267 } \
268 float64 VFP_HELPER(name, d)(float64 a, float64 b, void *fpstp) \
269 { \
270     float_status *fpst = fpstp; \
271     return float64_ ## name(a, b, fpst); \
272 }
273 VFP_BINOP(add)
274 VFP_BINOP(sub)
275 VFP_BINOP(mul)
276 VFP_BINOP(div)
277 VFP_BINOP(min)
278 VFP_BINOP(max)
279 VFP_BINOP(minnum)
280 VFP_BINOP(maxnum)
281 #undef VFP_BINOP
282 
283 dh_ctype_f16 VFP_HELPER(neg, h)(dh_ctype_f16 a)
284 {
285     return float16_chs(a);
286 }
287 
288 float32 VFP_HELPER(neg, s)(float32 a)
289 {
290     return float32_chs(a);
291 }
292 
293 float64 VFP_HELPER(neg, d)(float64 a)
294 {
295     return float64_chs(a);
296 }
297 
298 dh_ctype_f16 VFP_HELPER(abs, h)(dh_ctype_f16 a)
299 {
300     return float16_abs(a);
301 }
302 
303 float32 VFP_HELPER(abs, s)(float32 a)
304 {
305     return float32_abs(a);
306 }
307 
308 float64 VFP_HELPER(abs, d)(float64 a)
309 {
310     return float64_abs(a);
311 }
312 
313 dh_ctype_f16 VFP_HELPER(sqrt, h)(dh_ctype_f16 a, CPUARMState *env)
314 {
315     return float16_sqrt(a, &env->vfp.fp_status_f16);
316 }
317 
318 float32 VFP_HELPER(sqrt, s)(float32 a, CPUARMState *env)
319 {
320     return float32_sqrt(a, &env->vfp.fp_status);
321 }
322 
323 float64 VFP_HELPER(sqrt, d)(float64 a, CPUARMState *env)
324 {
325     return float64_sqrt(a, &env->vfp.fp_status);
326 }
327 
328 static void softfloat_to_vfp_compare(CPUARMState *env, FloatRelation cmp)
329 {
330     uint32_t flags;
331     switch (cmp) {
332     case float_relation_equal:
333         flags = 0x6;
334         break;
335     case float_relation_less:
336         flags = 0x8;
337         break;
338     case float_relation_greater:
339         flags = 0x2;
340         break;
341     case float_relation_unordered:
342         flags = 0x3;
343         break;
344     default:
345         g_assert_not_reached();
346     }
347     env->vfp.xregs[ARM_VFP_FPSCR] =
348         deposit32(env->vfp.xregs[ARM_VFP_FPSCR], 28, 4, flags);
349 }
350 
351 /* XXX: check quiet/signaling case */
352 #define DO_VFP_cmp(P, FLOATTYPE, ARGTYPE, FPST) \
353 void VFP_HELPER(cmp, P)(ARGTYPE a, ARGTYPE b, CPUARMState *env)  \
354 { \
355     softfloat_to_vfp_compare(env, \
356         FLOATTYPE ## _compare_quiet(a, b, &env->vfp.FPST)); \
357 } \
358 void VFP_HELPER(cmpe, P)(ARGTYPE a, ARGTYPE b, CPUARMState *env) \
359 { \
360     softfloat_to_vfp_compare(env, \
361         FLOATTYPE ## _compare(a, b, &env->vfp.FPST)); \
362 }
363 DO_VFP_cmp(h, float16, dh_ctype_f16, fp_status_f16)
364 DO_VFP_cmp(s, float32, float32, fp_status)
365 DO_VFP_cmp(d, float64, float64, fp_status)
366 #undef DO_VFP_cmp
367 
368 /* Integer to float and float to integer conversions */
369 
370 #define CONV_ITOF(name, ftype, fsz, sign)                           \
371 ftype HELPER(name)(uint32_t x, void *fpstp)                         \
372 {                                                                   \
373     float_status *fpst = fpstp;                                     \
374     return sign##int32_to_##float##fsz((sign##int32_t)x, fpst);     \
375 }
376 
377 #define CONV_FTOI(name, ftype, fsz, sign, round)                \
378 sign##int32_t HELPER(name)(ftype x, void *fpstp)                \
379 {                                                               \
380     float_status *fpst = fpstp;                                 \
381     if (float##fsz##_is_any_nan(x)) {                           \
382         float_raise(float_flag_invalid, fpst);                  \
383         return 0;                                               \
384     }                                                           \
385     return float##fsz##_to_##sign##int32##round(x, fpst);       \
386 }
387 
388 #define FLOAT_CONVS(name, p, ftype, fsz, sign)            \
389     CONV_ITOF(vfp_##name##to##p, ftype, fsz, sign)        \
390     CONV_FTOI(vfp_to##name##p, ftype, fsz, sign, )        \
391     CONV_FTOI(vfp_to##name##z##p, ftype, fsz, sign, _round_to_zero)
392 
393 FLOAT_CONVS(si, h, uint32_t, 16, )
394 FLOAT_CONVS(si, s, float32, 32, )
395 FLOAT_CONVS(si, d, float64, 64, )
396 FLOAT_CONVS(ui, h, uint32_t, 16, u)
397 FLOAT_CONVS(ui, s, float32, 32, u)
398 FLOAT_CONVS(ui, d, float64, 64, u)
399 
400 #undef CONV_ITOF
401 #undef CONV_FTOI
402 #undef FLOAT_CONVS
403 
404 /* floating point conversion */
405 float64 VFP_HELPER(fcvtd, s)(float32 x, CPUARMState *env)
406 {
407     return float32_to_float64(x, &env->vfp.fp_status);
408 }
409 
410 float32 VFP_HELPER(fcvts, d)(float64 x, CPUARMState *env)
411 {
412     return float64_to_float32(x, &env->vfp.fp_status);
413 }
414 
415 uint32_t HELPER(bfcvt)(float32 x, void *status)
416 {
417     return float32_to_bfloat16(x, status);
418 }
419 
420 uint32_t HELPER(bfcvt_pair)(uint64_t pair, void *status)
421 {
422     bfloat16 lo = float32_to_bfloat16(extract64(pair, 0, 32), status);
423     bfloat16 hi = float32_to_bfloat16(extract64(pair, 32, 32), status);
424     return deposit32(lo, 16, 16, hi);
425 }
426 
427 /*
428  * VFP3 fixed point conversion. The AArch32 versions of fix-to-float
429  * must always round-to-nearest; the AArch64 ones honour the FPSCR
430  * rounding mode. (For AArch32 Neon the standard-FPSCR is set to
431  * round-to-nearest so either helper will work.) AArch32 float-to-fix
432  * must round-to-zero.
433  */
434 #define VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)            \
435 ftype HELPER(vfp_##name##to##p)(uint##isz##_t  x, uint32_t shift,      \
436                                      void *fpstp) \
437 { return itype##_to_##float##fsz##_scalbn(x, -shift, fpstp); }
438 
439 #define VFP_CONV_FIX_FLOAT_ROUND(name, p, fsz, ftype, isz, itype)      \
440     ftype HELPER(vfp_##name##to##p##_round_to_nearest)(uint##isz##_t  x, \
441                                                      uint32_t shift,   \
442                                                      void *fpstp)      \
443     {                                                                  \
444         ftype ret;                                                     \
445         float_status *fpst = fpstp;                                    \
446         FloatRoundMode oldmode = fpst->float_rounding_mode;            \
447         fpst->float_rounding_mode = float_round_nearest_even;          \
448         ret = itype##_to_##float##fsz##_scalbn(x, -shift, fpstp);      \
449         fpst->float_rounding_mode = oldmode;                           \
450         return ret;                                                    \
451     }
452 
453 #define VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype, ROUND, suff) \
454 uint##isz##_t HELPER(vfp_to##name##p##suff)(ftype x, uint32_t shift,      \
455                                             void *fpst)                   \
456 {                                                                         \
457     if (unlikely(float##fsz##_is_any_nan(x))) {                           \
458         float_raise(float_flag_invalid, fpst);                            \
459         return 0;                                                         \
460     }                                                                     \
461     return float##fsz##_to_##itype##_scalbn(x, ROUND, shift, fpst);       \
462 }
463 
464 #define VFP_CONV_FIX(name, p, fsz, ftype, isz, itype)            \
465 VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)              \
466 VFP_CONV_FIX_FLOAT_ROUND(name, p, fsz, ftype, isz, itype)        \
467 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
468                          float_round_to_zero, _round_to_zero)    \
469 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
470                          get_float_rounding_mode(fpst), )
471 
472 #define VFP_CONV_FIX_A64(name, p, fsz, ftype, isz, itype)        \
473 VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)              \
474 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
475                          get_float_rounding_mode(fpst), )
476 
477 VFP_CONV_FIX(sh, d, 64, float64, 64, int16)
478 VFP_CONV_FIX(sl, d, 64, float64, 64, int32)
479 VFP_CONV_FIX_A64(sq, d, 64, float64, 64, int64)
480 VFP_CONV_FIX(uh, d, 64, float64, 64, uint16)
481 VFP_CONV_FIX(ul, d, 64, float64, 64, uint32)
482 VFP_CONV_FIX_A64(uq, d, 64, float64, 64, uint64)
483 VFP_CONV_FIX(sh, s, 32, float32, 32, int16)
484 VFP_CONV_FIX(sl, s, 32, float32, 32, int32)
485 VFP_CONV_FIX_A64(sq, s, 32, float32, 64, int64)
486 VFP_CONV_FIX(uh, s, 32, float32, 32, uint16)
487 VFP_CONV_FIX(ul, s, 32, float32, 32, uint32)
488 VFP_CONV_FIX_A64(uq, s, 32, float32, 64, uint64)
489 VFP_CONV_FIX(sh, h, 16, dh_ctype_f16, 32, int16)
490 VFP_CONV_FIX(sl, h, 16, dh_ctype_f16, 32, int32)
491 VFP_CONV_FIX_A64(sq, h, 16, dh_ctype_f16, 64, int64)
492 VFP_CONV_FIX(uh, h, 16, dh_ctype_f16, 32, uint16)
493 VFP_CONV_FIX(ul, h, 16, dh_ctype_f16, 32, uint32)
494 VFP_CONV_FIX_A64(uq, h, 16, dh_ctype_f16, 64, uint64)
495 
496 #undef VFP_CONV_FIX
497 #undef VFP_CONV_FIX_FLOAT
498 #undef VFP_CONV_FLOAT_FIX_ROUND
499 #undef VFP_CONV_FIX_A64
500 
501 /* Set the current fp rounding mode and return the old one.
502  * The argument is a softfloat float_round_ value.
503  */
504 uint32_t HELPER(set_rmode)(uint32_t rmode, void *fpstp)
505 {
506     float_status *fp_status = fpstp;
507 
508     uint32_t prev_rmode = get_float_rounding_mode(fp_status);
509     set_float_rounding_mode(rmode, fp_status);
510 
511     return prev_rmode;
512 }
513 
514 /* Half precision conversions.  */
515 float32 HELPER(vfp_fcvt_f16_to_f32)(uint32_t a, void *fpstp, uint32_t ahp_mode)
516 {
517     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
518      * it would affect flushing input denormals.
519      */
520     float_status *fpst = fpstp;
521     bool save = get_flush_inputs_to_zero(fpst);
522     set_flush_inputs_to_zero(false, fpst);
523     float32 r = float16_to_float32(a, !ahp_mode, fpst);
524     set_flush_inputs_to_zero(save, fpst);
525     return r;
526 }
527 
528 uint32_t HELPER(vfp_fcvt_f32_to_f16)(float32 a, void *fpstp, uint32_t ahp_mode)
529 {
530     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
531      * it would affect flushing output denormals.
532      */
533     float_status *fpst = fpstp;
534     bool save = get_flush_to_zero(fpst);
535     set_flush_to_zero(false, fpst);
536     float16 r = float32_to_float16(a, !ahp_mode, fpst);
537     set_flush_to_zero(save, fpst);
538     return r;
539 }
540 
541 float64 HELPER(vfp_fcvt_f16_to_f64)(uint32_t a, void *fpstp, uint32_t ahp_mode)
542 {
543     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
544      * it would affect flushing input denormals.
545      */
546     float_status *fpst = fpstp;
547     bool save = get_flush_inputs_to_zero(fpst);
548     set_flush_inputs_to_zero(false, fpst);
549     float64 r = float16_to_float64(a, !ahp_mode, fpst);
550     set_flush_inputs_to_zero(save, fpst);
551     return r;
552 }
553 
554 uint32_t HELPER(vfp_fcvt_f64_to_f16)(float64 a, void *fpstp, uint32_t ahp_mode)
555 {
556     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
557      * it would affect flushing output denormals.
558      */
559     float_status *fpst = fpstp;
560     bool save = get_flush_to_zero(fpst);
561     set_flush_to_zero(false, fpst);
562     float16 r = float64_to_float16(a, !ahp_mode, fpst);
563     set_flush_to_zero(save, fpst);
564     return r;
565 }
566 
567 /* NEON helpers.  */
568 
569 /* Constants 256 and 512 are used in some helpers; we avoid relying on
570  * int->float conversions at run-time.  */
571 #define float64_256 make_float64(0x4070000000000000LL)
572 #define float64_512 make_float64(0x4080000000000000LL)
573 #define float16_maxnorm make_float16(0x7bff)
574 #define float32_maxnorm make_float32(0x7f7fffff)
575 #define float64_maxnorm make_float64(0x7fefffffffffffffLL)
576 
577 /* Reciprocal functions
578  *
579  * The algorithm that must be used to calculate the estimate
580  * is specified by the ARM ARM, see FPRecipEstimate()/RecipEstimate
581  */
582 
583 /* See RecipEstimate()
584  *
585  * input is a 9 bit fixed point number
586  * input range 256 .. 511 for a number from 0.5 <= x < 1.0.
587  * result range 256 .. 511 for a number from 1.0 to 511/256.
588  */
589 
590 static int recip_estimate(int input)
591 {
592     int a, b, r;
593     assert(256 <= input && input < 512);
594     a = (input * 2) + 1;
595     b = (1 << 19) / a;
596     r = (b + 1) >> 1;
597     assert(256 <= r && r < 512);
598     return r;
599 }
600 
601 /*
602  * Common wrapper to call recip_estimate
603  *
604  * The parameters are exponent and 64 bit fraction (without implicit
605  * bit) where the binary point is nominally at bit 52. Returns a
606  * float64 which can then be rounded to the appropriate size by the
607  * callee.
608  */
609 
610 static uint64_t call_recip_estimate(int *exp, int exp_off, uint64_t frac)
611 {
612     uint32_t scaled, estimate;
613     uint64_t result_frac;
614     int result_exp;
615 
616     /* Handle sub-normals */
617     if (*exp == 0) {
618         if (extract64(frac, 51, 1) == 0) {
619             *exp = -1;
620             frac <<= 2;
621         } else {
622             frac <<= 1;
623         }
624     }
625 
626     /* scaled = UInt('1':fraction<51:44>) */
627     scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
628     estimate = recip_estimate(scaled);
629 
630     result_exp = exp_off - *exp;
631     result_frac = deposit64(0, 44, 8, estimate);
632     if (result_exp == 0) {
633         result_frac = deposit64(result_frac >> 1, 51, 1, 1);
634     } else if (result_exp == -1) {
635         result_frac = deposit64(result_frac >> 2, 50, 2, 1);
636         result_exp = 0;
637     }
638 
639     *exp = result_exp;
640 
641     return result_frac;
642 }
643 
644 static bool round_to_inf(float_status *fpst, bool sign_bit)
645 {
646     switch (fpst->float_rounding_mode) {
647     case float_round_nearest_even: /* Round to Nearest */
648         return true;
649     case float_round_up: /* Round to +Inf */
650         return !sign_bit;
651     case float_round_down: /* Round to -Inf */
652         return sign_bit;
653     case float_round_to_zero: /* Round to Zero */
654         return false;
655     default:
656         g_assert_not_reached();
657     }
658 }
659 
660 uint32_t HELPER(recpe_f16)(uint32_t input, void *fpstp)
661 {
662     float_status *fpst = fpstp;
663     float16 f16 = float16_squash_input_denormal(input, fpst);
664     uint32_t f16_val = float16_val(f16);
665     uint32_t f16_sign = float16_is_neg(f16);
666     int f16_exp = extract32(f16_val, 10, 5);
667     uint32_t f16_frac = extract32(f16_val, 0, 10);
668     uint64_t f64_frac;
669 
670     if (float16_is_any_nan(f16)) {
671         float16 nan = f16;
672         if (float16_is_signaling_nan(f16, fpst)) {
673             float_raise(float_flag_invalid, fpst);
674             if (!fpst->default_nan_mode) {
675                 nan = float16_silence_nan(f16, fpst);
676             }
677         }
678         if (fpst->default_nan_mode) {
679             nan =  float16_default_nan(fpst);
680         }
681         return nan;
682     } else if (float16_is_infinity(f16)) {
683         return float16_set_sign(float16_zero, float16_is_neg(f16));
684     } else if (float16_is_zero(f16)) {
685         float_raise(float_flag_divbyzero, fpst);
686         return float16_set_sign(float16_infinity, float16_is_neg(f16));
687     } else if (float16_abs(f16) < (1 << 8)) {
688         /* Abs(value) < 2.0^-16 */
689         float_raise(float_flag_overflow | float_flag_inexact, fpst);
690         if (round_to_inf(fpst, f16_sign)) {
691             return float16_set_sign(float16_infinity, f16_sign);
692         } else {
693             return float16_set_sign(float16_maxnorm, f16_sign);
694         }
695     } else if (f16_exp >= 29 && fpst->flush_to_zero) {
696         float_raise(float_flag_underflow, fpst);
697         return float16_set_sign(float16_zero, float16_is_neg(f16));
698     }
699 
700     f64_frac = call_recip_estimate(&f16_exp, 29,
701                                    ((uint64_t) f16_frac) << (52 - 10));
702 
703     /* result = sign : result_exp<4:0> : fraction<51:42> */
704     f16_val = deposit32(0, 15, 1, f16_sign);
705     f16_val = deposit32(f16_val, 10, 5, f16_exp);
706     f16_val = deposit32(f16_val, 0, 10, extract64(f64_frac, 52 - 10, 10));
707     return make_float16(f16_val);
708 }
709 
710 float32 HELPER(recpe_f32)(float32 input, void *fpstp)
711 {
712     float_status *fpst = fpstp;
713     float32 f32 = float32_squash_input_denormal(input, fpst);
714     uint32_t f32_val = float32_val(f32);
715     bool f32_sign = float32_is_neg(f32);
716     int f32_exp = extract32(f32_val, 23, 8);
717     uint32_t f32_frac = extract32(f32_val, 0, 23);
718     uint64_t f64_frac;
719 
720     if (float32_is_any_nan(f32)) {
721         float32 nan = f32;
722         if (float32_is_signaling_nan(f32, fpst)) {
723             float_raise(float_flag_invalid, fpst);
724             if (!fpst->default_nan_mode) {
725                 nan = float32_silence_nan(f32, fpst);
726             }
727         }
728         if (fpst->default_nan_mode) {
729             nan =  float32_default_nan(fpst);
730         }
731         return nan;
732     } else if (float32_is_infinity(f32)) {
733         return float32_set_sign(float32_zero, float32_is_neg(f32));
734     } else if (float32_is_zero(f32)) {
735         float_raise(float_flag_divbyzero, fpst);
736         return float32_set_sign(float32_infinity, float32_is_neg(f32));
737     } else if (float32_abs(f32) < (1ULL << 21)) {
738         /* Abs(value) < 2.0^-128 */
739         float_raise(float_flag_overflow | float_flag_inexact, fpst);
740         if (round_to_inf(fpst, f32_sign)) {
741             return float32_set_sign(float32_infinity, f32_sign);
742         } else {
743             return float32_set_sign(float32_maxnorm, f32_sign);
744         }
745     } else if (f32_exp >= 253 && fpst->flush_to_zero) {
746         float_raise(float_flag_underflow, fpst);
747         return float32_set_sign(float32_zero, float32_is_neg(f32));
748     }
749 
750     f64_frac = call_recip_estimate(&f32_exp, 253,
751                                    ((uint64_t) f32_frac) << (52 - 23));
752 
753     /* result = sign : result_exp<7:0> : fraction<51:29> */
754     f32_val = deposit32(0, 31, 1, f32_sign);
755     f32_val = deposit32(f32_val, 23, 8, f32_exp);
756     f32_val = deposit32(f32_val, 0, 23, extract64(f64_frac, 52 - 23, 23));
757     return make_float32(f32_val);
758 }
759 
760 float64 HELPER(recpe_f64)(float64 input, void *fpstp)
761 {
762     float_status *fpst = fpstp;
763     float64 f64 = float64_squash_input_denormal(input, fpst);
764     uint64_t f64_val = float64_val(f64);
765     bool f64_sign = float64_is_neg(f64);
766     int f64_exp = extract64(f64_val, 52, 11);
767     uint64_t f64_frac = extract64(f64_val, 0, 52);
768 
769     /* Deal with any special cases */
770     if (float64_is_any_nan(f64)) {
771         float64 nan = f64;
772         if (float64_is_signaling_nan(f64, fpst)) {
773             float_raise(float_flag_invalid, fpst);
774             if (!fpst->default_nan_mode) {
775                 nan = float64_silence_nan(f64, fpst);
776             }
777         }
778         if (fpst->default_nan_mode) {
779             nan =  float64_default_nan(fpst);
780         }
781         return nan;
782     } else if (float64_is_infinity(f64)) {
783         return float64_set_sign(float64_zero, float64_is_neg(f64));
784     } else if (float64_is_zero(f64)) {
785         float_raise(float_flag_divbyzero, fpst);
786         return float64_set_sign(float64_infinity, float64_is_neg(f64));
787     } else if ((f64_val & ~(1ULL << 63)) < (1ULL << 50)) {
788         /* Abs(value) < 2.0^-1024 */
789         float_raise(float_flag_overflow | float_flag_inexact, fpst);
790         if (round_to_inf(fpst, f64_sign)) {
791             return float64_set_sign(float64_infinity, f64_sign);
792         } else {
793             return float64_set_sign(float64_maxnorm, f64_sign);
794         }
795     } else if (f64_exp >= 2045 && fpst->flush_to_zero) {
796         float_raise(float_flag_underflow, fpst);
797         return float64_set_sign(float64_zero, float64_is_neg(f64));
798     }
799 
800     f64_frac = call_recip_estimate(&f64_exp, 2045, f64_frac);
801 
802     /* result = sign : result_exp<10:0> : fraction<51:0>; */
803     f64_val = deposit64(0, 63, 1, f64_sign);
804     f64_val = deposit64(f64_val, 52, 11, f64_exp);
805     f64_val = deposit64(f64_val, 0, 52, f64_frac);
806     return make_float64(f64_val);
807 }
808 
809 /* The algorithm that must be used to calculate the estimate
810  * is specified by the ARM ARM.
811  */
812 
813 static int do_recip_sqrt_estimate(int a)
814 {
815     int b, estimate;
816 
817     assert(128 <= a && a < 512);
818     if (a < 256) {
819         a = a * 2 + 1;
820     } else {
821         a = (a >> 1) << 1;
822         a = (a + 1) * 2;
823     }
824     b = 512;
825     while (a * (b + 1) * (b + 1) < (1 << 28)) {
826         b += 1;
827     }
828     estimate = (b + 1) / 2;
829     assert(256 <= estimate && estimate < 512);
830 
831     return estimate;
832 }
833 
834 
835 static uint64_t recip_sqrt_estimate(int *exp , int exp_off, uint64_t frac)
836 {
837     int estimate;
838     uint32_t scaled;
839 
840     if (*exp == 0) {
841         while (extract64(frac, 51, 1) == 0) {
842             frac = frac << 1;
843             *exp -= 1;
844         }
845         frac = extract64(frac, 0, 51) << 1;
846     }
847 
848     if (*exp & 1) {
849         /* scaled = UInt('01':fraction<51:45>) */
850         scaled = deposit32(1 << 7, 0, 7, extract64(frac, 45, 7));
851     } else {
852         /* scaled = UInt('1':fraction<51:44>) */
853         scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
854     }
855     estimate = do_recip_sqrt_estimate(scaled);
856 
857     *exp = (exp_off - *exp) / 2;
858     return extract64(estimate, 0, 8) << 44;
859 }
860 
861 uint32_t HELPER(rsqrte_f16)(uint32_t input, void *fpstp)
862 {
863     float_status *s = fpstp;
864     float16 f16 = float16_squash_input_denormal(input, s);
865     uint16_t val = float16_val(f16);
866     bool f16_sign = float16_is_neg(f16);
867     int f16_exp = extract32(val, 10, 5);
868     uint16_t f16_frac = extract32(val, 0, 10);
869     uint64_t f64_frac;
870 
871     if (float16_is_any_nan(f16)) {
872         float16 nan = f16;
873         if (float16_is_signaling_nan(f16, s)) {
874             float_raise(float_flag_invalid, s);
875             if (!s->default_nan_mode) {
876                 nan = float16_silence_nan(f16, fpstp);
877             }
878         }
879         if (s->default_nan_mode) {
880             nan =  float16_default_nan(s);
881         }
882         return nan;
883     } else if (float16_is_zero(f16)) {
884         float_raise(float_flag_divbyzero, s);
885         return float16_set_sign(float16_infinity, f16_sign);
886     } else if (f16_sign) {
887         float_raise(float_flag_invalid, s);
888         return float16_default_nan(s);
889     } else if (float16_is_infinity(f16)) {
890         return float16_zero;
891     }
892 
893     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
894      * preserving the parity of the exponent.  */
895 
896     f64_frac = ((uint64_t) f16_frac) << (52 - 10);
897 
898     f64_frac = recip_sqrt_estimate(&f16_exp, 44, f64_frac);
899 
900     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(2) */
901     val = deposit32(0, 15, 1, f16_sign);
902     val = deposit32(val, 10, 5, f16_exp);
903     val = deposit32(val, 2, 8, extract64(f64_frac, 52 - 8, 8));
904     return make_float16(val);
905 }
906 
907 float32 HELPER(rsqrte_f32)(float32 input, void *fpstp)
908 {
909     float_status *s = fpstp;
910     float32 f32 = float32_squash_input_denormal(input, s);
911     uint32_t val = float32_val(f32);
912     uint32_t f32_sign = float32_is_neg(f32);
913     int f32_exp = extract32(val, 23, 8);
914     uint32_t f32_frac = extract32(val, 0, 23);
915     uint64_t f64_frac;
916 
917     if (float32_is_any_nan(f32)) {
918         float32 nan = f32;
919         if (float32_is_signaling_nan(f32, s)) {
920             float_raise(float_flag_invalid, s);
921             if (!s->default_nan_mode) {
922                 nan = float32_silence_nan(f32, fpstp);
923             }
924         }
925         if (s->default_nan_mode) {
926             nan =  float32_default_nan(s);
927         }
928         return nan;
929     } else if (float32_is_zero(f32)) {
930         float_raise(float_flag_divbyzero, s);
931         return float32_set_sign(float32_infinity, float32_is_neg(f32));
932     } else if (float32_is_neg(f32)) {
933         float_raise(float_flag_invalid, s);
934         return float32_default_nan(s);
935     } else if (float32_is_infinity(f32)) {
936         return float32_zero;
937     }
938 
939     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
940      * preserving the parity of the exponent.  */
941 
942     f64_frac = ((uint64_t) f32_frac) << 29;
943 
944     f64_frac = recip_sqrt_estimate(&f32_exp, 380, f64_frac);
945 
946     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(15) */
947     val = deposit32(0, 31, 1, f32_sign);
948     val = deposit32(val, 23, 8, f32_exp);
949     val = deposit32(val, 15, 8, extract64(f64_frac, 52 - 8, 8));
950     return make_float32(val);
951 }
952 
953 float64 HELPER(rsqrte_f64)(float64 input, void *fpstp)
954 {
955     float_status *s = fpstp;
956     float64 f64 = float64_squash_input_denormal(input, s);
957     uint64_t val = float64_val(f64);
958     bool f64_sign = float64_is_neg(f64);
959     int f64_exp = extract64(val, 52, 11);
960     uint64_t f64_frac = extract64(val, 0, 52);
961 
962     if (float64_is_any_nan(f64)) {
963         float64 nan = f64;
964         if (float64_is_signaling_nan(f64, s)) {
965             float_raise(float_flag_invalid, s);
966             if (!s->default_nan_mode) {
967                 nan = float64_silence_nan(f64, fpstp);
968             }
969         }
970         if (s->default_nan_mode) {
971             nan =  float64_default_nan(s);
972         }
973         return nan;
974     } else if (float64_is_zero(f64)) {
975         float_raise(float_flag_divbyzero, s);
976         return float64_set_sign(float64_infinity, float64_is_neg(f64));
977     } else if (float64_is_neg(f64)) {
978         float_raise(float_flag_invalid, s);
979         return float64_default_nan(s);
980     } else if (float64_is_infinity(f64)) {
981         return float64_zero;
982     }
983 
984     f64_frac = recip_sqrt_estimate(&f64_exp, 3068, f64_frac);
985 
986     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(44) */
987     val = deposit64(0, 61, 1, f64_sign);
988     val = deposit64(val, 52, 11, f64_exp);
989     val = deposit64(val, 44, 8, extract64(f64_frac, 52 - 8, 8));
990     return make_float64(val);
991 }
992 
993 uint32_t HELPER(recpe_u32)(uint32_t a)
994 {
995     int input, estimate;
996 
997     if ((a & 0x80000000) == 0) {
998         return 0xffffffff;
999     }
1000 
1001     input = extract32(a, 23, 9);
1002     estimate = recip_estimate(input);
1003 
1004     return deposit32(0, (32 - 9), 9, estimate);
1005 }
1006 
1007 uint32_t HELPER(rsqrte_u32)(uint32_t a)
1008 {
1009     int estimate;
1010 
1011     if ((a & 0xc0000000) == 0) {
1012         return 0xffffffff;
1013     }
1014 
1015     estimate = do_recip_sqrt_estimate(extract32(a, 23, 9));
1016 
1017     return deposit32(0, 23, 9, estimate);
1018 }
1019 
1020 /* VFPv4 fused multiply-accumulate */
1021 dh_ctype_f16 VFP_HELPER(muladd, h)(dh_ctype_f16 a, dh_ctype_f16 b,
1022                                    dh_ctype_f16 c, void *fpstp)
1023 {
1024     float_status *fpst = fpstp;
1025     return float16_muladd(a, b, c, 0, fpst);
1026 }
1027 
1028 float32 VFP_HELPER(muladd, s)(float32 a, float32 b, float32 c, void *fpstp)
1029 {
1030     float_status *fpst = fpstp;
1031     return float32_muladd(a, b, c, 0, fpst);
1032 }
1033 
1034 float64 VFP_HELPER(muladd, d)(float64 a, float64 b, float64 c, void *fpstp)
1035 {
1036     float_status *fpst = fpstp;
1037     return float64_muladd(a, b, c, 0, fpst);
1038 }
1039 
1040 /* ARMv8 round to integral */
1041 dh_ctype_f16 HELPER(rinth_exact)(dh_ctype_f16 x, void *fp_status)
1042 {
1043     return float16_round_to_int(x, fp_status);
1044 }
1045 
1046 float32 HELPER(rints_exact)(float32 x, void *fp_status)
1047 {
1048     return float32_round_to_int(x, fp_status);
1049 }
1050 
1051 float64 HELPER(rintd_exact)(float64 x, void *fp_status)
1052 {
1053     return float64_round_to_int(x, fp_status);
1054 }
1055 
1056 dh_ctype_f16 HELPER(rinth)(dh_ctype_f16 x, void *fp_status)
1057 {
1058     int old_flags = get_float_exception_flags(fp_status), new_flags;
1059     float16 ret;
1060 
1061     ret = float16_round_to_int(x, fp_status);
1062 
1063     /* Suppress any inexact exceptions the conversion produced */
1064     if (!(old_flags & float_flag_inexact)) {
1065         new_flags = get_float_exception_flags(fp_status);
1066         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1067     }
1068 
1069     return ret;
1070 }
1071 
1072 float32 HELPER(rints)(float32 x, void *fp_status)
1073 {
1074     int old_flags = get_float_exception_flags(fp_status), new_flags;
1075     float32 ret;
1076 
1077     ret = float32_round_to_int(x, fp_status);
1078 
1079     /* Suppress any inexact exceptions the conversion produced */
1080     if (!(old_flags & float_flag_inexact)) {
1081         new_flags = get_float_exception_flags(fp_status);
1082         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1083     }
1084 
1085     return ret;
1086 }
1087 
1088 float64 HELPER(rintd)(float64 x, void *fp_status)
1089 {
1090     int old_flags = get_float_exception_flags(fp_status), new_flags;
1091     float64 ret;
1092 
1093     ret = float64_round_to_int(x, fp_status);
1094 
1095     new_flags = get_float_exception_flags(fp_status);
1096 
1097     /* Suppress any inexact exceptions the conversion produced */
1098     if (!(old_flags & float_flag_inexact)) {
1099         new_flags = get_float_exception_flags(fp_status);
1100         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1101     }
1102 
1103     return ret;
1104 }
1105 
1106 /* Convert ARM rounding mode to softfloat */
1107 int arm_rmode_to_sf(int rmode)
1108 {
1109     switch (rmode) {
1110     case FPROUNDING_TIEAWAY:
1111         rmode = float_round_ties_away;
1112         break;
1113     case FPROUNDING_ODD:
1114         /* FIXME: add support for TIEAWAY and ODD */
1115         qemu_log_mask(LOG_UNIMP, "arm: unimplemented rounding mode: %d\n",
1116                       rmode);
1117         /* fall through for now */
1118     case FPROUNDING_TIEEVEN:
1119     default:
1120         rmode = float_round_nearest_even;
1121         break;
1122     case FPROUNDING_POSINF:
1123         rmode = float_round_up;
1124         break;
1125     case FPROUNDING_NEGINF:
1126         rmode = float_round_down;
1127         break;
1128     case FPROUNDING_ZERO:
1129         rmode = float_round_to_zero;
1130         break;
1131     }
1132     return rmode;
1133 }
1134 
1135 /*
1136  * Implement float64 to int32_t conversion without saturation;
1137  * the result is supplied modulo 2^32.
1138  */
1139 uint64_t HELPER(fjcvtzs)(float64 value, void *vstatus)
1140 {
1141     float_status *status = vstatus;
1142     uint32_t exp, sign;
1143     uint64_t frac;
1144     uint32_t inexact = 1; /* !Z */
1145 
1146     sign = extract64(value, 63, 1);
1147     exp = extract64(value, 52, 11);
1148     frac = extract64(value, 0, 52);
1149 
1150     if (exp == 0) {
1151         /* While not inexact for IEEE FP, -0.0 is inexact for JavaScript.  */
1152         inexact = sign;
1153         if (frac != 0) {
1154             if (status->flush_inputs_to_zero) {
1155                 float_raise(float_flag_input_denormal, status);
1156             } else {
1157                 float_raise(float_flag_inexact, status);
1158                 inexact = 1;
1159             }
1160         }
1161         frac = 0;
1162     } else if (exp == 0x7ff) {
1163         /* This operation raises Invalid for both NaN and overflow (Inf).  */
1164         float_raise(float_flag_invalid, status);
1165         frac = 0;
1166     } else {
1167         int true_exp = exp - 1023;
1168         int shift = true_exp - 52;
1169 
1170         /* Restore implicit bit.  */
1171         frac |= 1ull << 52;
1172 
1173         /* Shift the fraction into place.  */
1174         if (shift >= 0) {
1175             /* The number is so large we must shift the fraction left.  */
1176             if (shift >= 64) {
1177                 /* The fraction is shifted out entirely.  */
1178                 frac = 0;
1179             } else {
1180                 frac <<= shift;
1181             }
1182         } else if (shift > -64) {
1183             /* Normal case -- shift right and notice if bits shift out.  */
1184             inexact = (frac << (64 + shift)) != 0;
1185             frac >>= -shift;
1186         } else {
1187             /* The fraction is shifted out entirely.  */
1188             frac = 0;
1189         }
1190 
1191         /* Notice overflow or inexact exceptions.  */
1192         if (true_exp > 31 || frac > (sign ? 0x80000000ull : 0x7fffffff)) {
1193             /* Overflow, for which this operation raises invalid.  */
1194             float_raise(float_flag_invalid, status);
1195             inexact = 1;
1196         } else if (inexact) {
1197             float_raise(float_flag_inexact, status);
1198         }
1199 
1200         /* Honor the sign.  */
1201         if (sign) {
1202             frac = -frac;
1203         }
1204     }
1205 
1206     /* Pack the result and the env->ZF representation of Z together.  */
1207     return deposit64(frac, 32, 32, inexact);
1208 }
1209 
1210 uint32_t HELPER(vjcvt)(float64 value, CPUARMState *env)
1211 {
1212     uint64_t pair = HELPER(fjcvtzs)(value, &env->vfp.fp_status);
1213     uint32_t result = pair;
1214     uint32_t z = (pair >> 32) == 0;
1215 
1216     /* Store Z, clear NCV, in FPSCR.NZCV.  */
1217     env->vfp.xregs[ARM_VFP_FPSCR]
1218         = (env->vfp.xregs[ARM_VFP_FPSCR] & ~CPSR_NZCV) | (z * CPSR_Z);
1219 
1220     return result;
1221 }
1222 
1223 /* Round a float32 to an integer that fits in int32_t or int64_t.  */
1224 static float32 frint_s(float32 f, float_status *fpst, int intsize)
1225 {
1226     int old_flags = get_float_exception_flags(fpst);
1227     uint32_t exp = extract32(f, 23, 8);
1228 
1229     if (unlikely(exp == 0xff)) {
1230         /* NaN or Inf.  */
1231         goto overflow;
1232     }
1233 
1234     /* Round and re-extract the exponent.  */
1235     f = float32_round_to_int(f, fpst);
1236     exp = extract32(f, 23, 8);
1237 
1238     /* Validate the range of the result.  */
1239     if (exp < 126 + intsize) {
1240         /* abs(F) <= INT{N}_MAX */
1241         return f;
1242     }
1243     if (exp == 126 + intsize) {
1244         uint32_t sign = extract32(f, 31, 1);
1245         uint32_t frac = extract32(f, 0, 23);
1246         if (sign && frac == 0) {
1247             /* F == INT{N}_MIN */
1248             return f;
1249         }
1250     }
1251 
1252  overflow:
1253     /*
1254      * Raise Invalid and return INT{N}_MIN as a float.  Revert any
1255      * inexact exception float32_round_to_int may have raised.
1256      */
1257     set_float_exception_flags(old_flags | float_flag_invalid, fpst);
1258     return (0x100u + 126u + intsize) << 23;
1259 }
1260 
1261 float32 HELPER(frint32_s)(float32 f, void *fpst)
1262 {
1263     return frint_s(f, fpst, 32);
1264 }
1265 
1266 float32 HELPER(frint64_s)(float32 f, void *fpst)
1267 {
1268     return frint_s(f, fpst, 64);
1269 }
1270 
1271 /* Round a float64 to an integer that fits in int32_t or int64_t.  */
1272 static float64 frint_d(float64 f, float_status *fpst, int intsize)
1273 {
1274     int old_flags = get_float_exception_flags(fpst);
1275     uint32_t exp = extract64(f, 52, 11);
1276 
1277     if (unlikely(exp == 0x7ff)) {
1278         /* NaN or Inf.  */
1279         goto overflow;
1280     }
1281 
1282     /* Round and re-extract the exponent.  */
1283     f = float64_round_to_int(f, fpst);
1284     exp = extract64(f, 52, 11);
1285 
1286     /* Validate the range of the result.  */
1287     if (exp < 1022 + intsize) {
1288         /* abs(F) <= INT{N}_MAX */
1289         return f;
1290     }
1291     if (exp == 1022 + intsize) {
1292         uint64_t sign = extract64(f, 63, 1);
1293         uint64_t frac = extract64(f, 0, 52);
1294         if (sign && frac == 0) {
1295             /* F == INT{N}_MIN */
1296             return f;
1297         }
1298     }
1299 
1300  overflow:
1301     /*
1302      * Raise Invalid and return INT{N}_MIN as a float.  Revert any
1303      * inexact exception float64_round_to_int may have raised.
1304      */
1305     set_float_exception_flags(old_flags | float_flag_invalid, fpst);
1306     return (uint64_t)(0x800 + 1022 + intsize) << 52;
1307 }
1308 
1309 float64 HELPER(frint32_d)(float64 f, void *fpst)
1310 {
1311     return frint_d(f, fpst, 32);
1312 }
1313 
1314 float64 HELPER(frint64_d)(float64 f, void *fpst)
1315 {
1316     return frint_d(f, fpst, 64);
1317 }
1318 
1319 void HELPER(check_hcr_el2_trap)(CPUARMState *env, uint32_t rt, uint32_t reg)
1320 {
1321     uint32_t syndrome;
1322 
1323     switch (reg) {
1324     case ARM_VFP_MVFR0:
1325     case ARM_VFP_MVFR1:
1326     case ARM_VFP_MVFR2:
1327         if (!(arm_hcr_el2_eff(env) & HCR_TID3)) {
1328             return;
1329         }
1330         break;
1331     case ARM_VFP_FPSID:
1332         if (!(arm_hcr_el2_eff(env) & HCR_TID0)) {
1333             return;
1334         }
1335         break;
1336     default:
1337         g_assert_not_reached();
1338     }
1339 
1340     syndrome = ((EC_FPIDTRAP << ARM_EL_EC_SHIFT)
1341                 | ARM_EL_IL
1342                 | (1 << 24) | (0xe << 20) | (7 << 14)
1343                 | (reg << 10) | (rt << 5) | 1);
1344 
1345     raise_exception(env, EXCP_HYP_TRAP, syndrome, 2);
1346 }
1347 
1348 #endif
1349