xref: /openbmc/qemu/target/arm/vfp_helper.c (revision 646b5378)
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 #include "cpu-features.h"
25 #ifdef CONFIG_TCG
26 #include "qemu/log.h"
27 #include "fpu/softfloat.h"
28 #endif
29 
30 /* VFP support.  We follow the convention used for VFP instructions:
31    Single precision routines have a "s" suffix, double precision a
32    "d" suffix.  */
33 
34 #ifdef CONFIG_TCG
35 
36 /* Convert host exception flags to vfp form.  */
37 static inline int vfp_exceptbits_from_host(int host_bits)
38 {
39     int target_bits = 0;
40 
41     if (host_bits & float_flag_invalid) {
42         target_bits |= 1;
43     }
44     if (host_bits & float_flag_divbyzero) {
45         target_bits |= 2;
46     }
47     if (host_bits & float_flag_overflow) {
48         target_bits |= 4;
49     }
50     if (host_bits & (float_flag_underflow | float_flag_output_denormal)) {
51         target_bits |= 8;
52     }
53     if (host_bits & float_flag_inexact) {
54         target_bits |= 0x10;
55     }
56     if (host_bits & float_flag_input_denormal) {
57         target_bits |= 0x80;
58     }
59     return target_bits;
60 }
61 
62 static uint32_t vfp_get_fpsr_from_host(CPUARMState *env)
63 {
64     uint32_t i;
65 
66     i = get_float_exception_flags(&env->vfp.fp_status);
67     i |= get_float_exception_flags(&env->vfp.standard_fp_status);
68     /* FZ16 does not generate an input denormal exception.  */
69     i |= (get_float_exception_flags(&env->vfp.fp_status_f16)
70           & ~float_flag_input_denormal);
71     i |= (get_float_exception_flags(&env->vfp.standard_fp_status_f16)
72           & ~float_flag_input_denormal);
73     return vfp_exceptbits_from_host(i);
74 }
75 
76 static void vfp_clear_float_status_exc_flags(CPUARMState *env)
77 {
78     /*
79      * Clear out all the exception-flag information in the float_status
80      * values. The caller should have arranged for env->vfp.fpsr to
81      * be the architecturally up-to-date exception flag information first.
82      */
83     set_float_exception_flags(0, &env->vfp.fp_status);
84     set_float_exception_flags(0, &env->vfp.fp_status_f16);
85     set_float_exception_flags(0, &env->vfp.standard_fp_status);
86     set_float_exception_flags(0, &env->vfp.standard_fp_status_f16);
87 }
88 
89 static void vfp_set_fpcr_to_host(CPUARMState *env, uint32_t val, uint32_t mask)
90 {
91     uint64_t changed = env->vfp.fpcr;
92 
93     changed ^= val;
94     changed &= mask;
95     if (changed & (3 << 22)) {
96         int i = (val >> 22) & 3;
97         switch (i) {
98         case FPROUNDING_TIEEVEN:
99             i = float_round_nearest_even;
100             break;
101         case FPROUNDING_POSINF:
102             i = float_round_up;
103             break;
104         case FPROUNDING_NEGINF:
105             i = float_round_down;
106             break;
107         case FPROUNDING_ZERO:
108             i = float_round_to_zero;
109             break;
110         }
111         set_float_rounding_mode(i, &env->vfp.fp_status);
112         set_float_rounding_mode(i, &env->vfp.fp_status_f16);
113     }
114     if (changed & FPCR_FZ16) {
115         bool ftz_enabled = val & FPCR_FZ16;
116         set_flush_to_zero(ftz_enabled, &env->vfp.fp_status_f16);
117         set_flush_to_zero(ftz_enabled, &env->vfp.standard_fp_status_f16);
118         set_flush_inputs_to_zero(ftz_enabled, &env->vfp.fp_status_f16);
119         set_flush_inputs_to_zero(ftz_enabled, &env->vfp.standard_fp_status_f16);
120     }
121     if (changed & FPCR_FZ) {
122         bool ftz_enabled = val & FPCR_FZ;
123         set_flush_to_zero(ftz_enabled, &env->vfp.fp_status);
124         set_flush_inputs_to_zero(ftz_enabled, &env->vfp.fp_status);
125     }
126     if (changed & FPCR_DN) {
127         bool dnan_enabled = val & FPCR_DN;
128         set_default_nan_mode(dnan_enabled, &env->vfp.fp_status);
129         set_default_nan_mode(dnan_enabled, &env->vfp.fp_status_f16);
130     }
131 }
132 
133 #else
134 
135 static uint32_t vfp_get_fpsr_from_host(CPUARMState *env)
136 {
137     return 0;
138 }
139 
140 static void vfp_clear_float_status_exc_flags(CPUARMState *env)
141 {
142 }
143 
144 static void vfp_set_fpcr_to_host(CPUARMState *env, uint32_t val, uint32_t mask)
145 {
146 }
147 
148 #endif
149 
150 uint32_t vfp_get_fpcr(CPUARMState *env)
151 {
152     uint32_t fpcr = env->vfp.fpcr
153         | (env->vfp.vec_len << 16)
154         | (env->vfp.vec_stride << 20);
155 
156     /*
157      * M-profile LTPSIZE is the same bits [18:16] as A-profile Len; whichever
158      * of the two is not applicable to this CPU will always be zero.
159      */
160     fpcr |= env->v7m.ltpsize << 16;
161 
162     return fpcr;
163 }
164 
165 uint32_t vfp_get_fpsr(CPUARMState *env)
166 {
167     uint32_t fpsr = env->vfp.fpsr;
168     uint32_t i;
169 
170     fpsr |= vfp_get_fpsr_from_host(env);
171 
172     i = env->vfp.qc[0] | env->vfp.qc[1] | env->vfp.qc[2] | env->vfp.qc[3];
173     fpsr |= i ? FPSR_QC : 0;
174     return fpsr;
175 }
176 
177 uint32_t HELPER(vfp_get_fpscr)(CPUARMState *env)
178 {
179     return (vfp_get_fpcr(env) & FPSCR_FPCR_MASK) |
180         (vfp_get_fpsr(env) & FPSCR_FPSR_MASK);
181 }
182 
183 uint32_t vfp_get_fpscr(CPUARMState *env)
184 {
185     return HELPER(vfp_get_fpscr)(env);
186 }
187 
188 void vfp_set_fpsr(CPUARMState *env, uint32_t val)
189 {
190     ARMCPU *cpu = env_archcpu(env);
191 
192     if (arm_feature(env, ARM_FEATURE_NEON) ||
193         cpu_isar_feature(aa32_mve, cpu)) {
194         /*
195          * The bit we set within vfp.qc[] is arbitrary; the array as a
196          * whole being zero/non-zero is what counts.
197          */
198         env->vfp.qc[0] = val & FPSR_QC;
199         env->vfp.qc[1] = 0;
200         env->vfp.qc[2] = 0;
201         env->vfp.qc[3] = 0;
202     }
203 
204     /*
205      * NZCV lives only in env->vfp.fpsr. The cumulative exception flags
206      * IOC|DZC|OFC|UFC|IXC|IDC also live in env->vfp.fpsr, with possible
207      * extra pending exception information that hasn't yet been folded in
208      * living in the float_status values (for TCG).
209      * Since this FPSR write gives us the up to date values of the exception
210      * flags, we want to store into vfp.fpsr the NZCV and CEXC bits, zeroing
211      * anything else. We also need to clear out the float_status exception
212      * information so that the next vfp_get_fpsr does not fold in stale data.
213      */
214     val &= FPSR_NZCV_MASK | FPSR_CEXC_MASK;
215     env->vfp.fpsr = val;
216     vfp_clear_float_status_exc_flags(env);
217 }
218 
219 static void vfp_set_fpcr_masked(CPUARMState *env, uint32_t val, uint32_t mask)
220 {
221     /*
222      * We only set FPCR bits defined by mask, and leave the others alone.
223      * We assume the mask is sensible (e.g. doesn't try to set only
224      * part of a field)
225      */
226     ARMCPU *cpu = env_archcpu(env);
227 
228     /* When ARMv8.2-FP16 is not supported, FZ16 is RES0.  */
229     if (!cpu_isar_feature(any_fp16, cpu)) {
230         val &= ~FPCR_FZ16;
231     }
232 
233     if (!cpu_isar_feature(aa64_ebf16, cpu)) {
234         val &= ~FPCR_EBF;
235     }
236 
237     vfp_set_fpcr_to_host(env, val, mask);
238 
239     if (mask & (FPCR_LEN_MASK | FPCR_STRIDE_MASK)) {
240         if (!arm_feature(env, ARM_FEATURE_M)) {
241             /*
242              * Short-vector length and stride; on M-profile these bits
243              * are used for different purposes.
244              * We can't make this conditional be "if MVFR0.FPShVec != 0",
245              * because in v7A no-short-vector-support cores still had to
246              * allow Stride/Len to be written with the only effect that
247              * some insns are required to UNDEF if the guest sets them.
248              */
249             env->vfp.vec_len = extract32(val, 16, 3);
250             env->vfp.vec_stride = extract32(val, 20, 2);
251         } else if (cpu_isar_feature(aa32_mve, cpu)) {
252             env->v7m.ltpsize = extract32(val, FPCR_LTPSIZE_SHIFT,
253                                          FPCR_LTPSIZE_LENGTH);
254         }
255     }
256 
257     /*
258      * We don't implement trapped exception handling, so the
259      * trap enable bits, IDE|IXE|UFE|OFE|DZE|IOE are all RAZ/WI (not RES0!)
260      *
261      * The FPCR bits we keep in vfp.fpcr are AHP, DN, FZ, RMode, EBF
262      * and FZ16. Len, Stride and LTPSIZE we just handled. Store those bits
263      * there, and zero any of the other FPCR bits and the RES0 and RAZ/WI
264      * bits.
265      */
266     val &= FPCR_AHP | FPCR_DN | FPCR_FZ | FPCR_RMODE_MASK | FPCR_FZ16 | FPCR_EBF;
267     env->vfp.fpcr &= ~mask;
268     env->vfp.fpcr |= val;
269 }
270 
271 void vfp_set_fpcr(CPUARMState *env, uint32_t val)
272 {
273     vfp_set_fpcr_masked(env, val, MAKE_64BIT_MASK(0, 32));
274 }
275 
276 void HELPER(vfp_set_fpscr)(CPUARMState *env, uint32_t val)
277 {
278     vfp_set_fpcr_masked(env, val, FPSCR_FPCR_MASK);
279     vfp_set_fpsr(env, val & FPSCR_FPSR_MASK);
280 }
281 
282 void vfp_set_fpscr(CPUARMState *env, uint32_t val)
283 {
284     HELPER(vfp_set_fpscr)(env, val);
285 }
286 
287 #ifdef CONFIG_TCG
288 
289 #define VFP_HELPER(name, p) HELPER(glue(glue(vfp_,name),p))
290 
291 #define VFP_BINOP(name) \
292 dh_ctype_f16 VFP_HELPER(name, h)(dh_ctype_f16 a, dh_ctype_f16 b, void *fpstp) \
293 { \
294     float_status *fpst = fpstp; \
295     return float16_ ## name(a, b, fpst); \
296 } \
297 float32 VFP_HELPER(name, s)(float32 a, float32 b, void *fpstp) \
298 { \
299     float_status *fpst = fpstp; \
300     return float32_ ## name(a, b, fpst); \
301 } \
302 float64 VFP_HELPER(name, d)(float64 a, float64 b, void *fpstp) \
303 { \
304     float_status *fpst = fpstp; \
305     return float64_ ## name(a, b, fpst); \
306 }
307 VFP_BINOP(add)
308 VFP_BINOP(sub)
309 VFP_BINOP(mul)
310 VFP_BINOP(div)
311 VFP_BINOP(min)
312 VFP_BINOP(max)
313 VFP_BINOP(minnum)
314 VFP_BINOP(maxnum)
315 #undef VFP_BINOP
316 
317 dh_ctype_f16 VFP_HELPER(sqrt, h)(dh_ctype_f16 a, CPUARMState *env)
318 {
319     return float16_sqrt(a, &env->vfp.fp_status_f16);
320 }
321 
322 float32 VFP_HELPER(sqrt, s)(float32 a, CPUARMState *env)
323 {
324     return float32_sqrt(a, &env->vfp.fp_status);
325 }
326 
327 float64 VFP_HELPER(sqrt, d)(float64 a, CPUARMState *env)
328 {
329     return float64_sqrt(a, &env->vfp.fp_status);
330 }
331 
332 static void softfloat_to_vfp_compare(CPUARMState *env, FloatRelation cmp)
333 {
334     uint32_t flags;
335     switch (cmp) {
336     case float_relation_equal:
337         flags = 0x6;
338         break;
339     case float_relation_less:
340         flags = 0x8;
341         break;
342     case float_relation_greater:
343         flags = 0x2;
344         break;
345     case float_relation_unordered:
346         flags = 0x3;
347         break;
348     default:
349         g_assert_not_reached();
350     }
351     env->vfp.fpsr = deposit64(env->vfp.fpsr, 28, 4, flags); /* NZCV */
352 }
353 
354 /* XXX: check quiet/signaling case */
355 #define DO_VFP_cmp(P, FLOATTYPE, ARGTYPE, FPST) \
356 void VFP_HELPER(cmp, P)(ARGTYPE a, ARGTYPE b, CPUARMState *env)  \
357 { \
358     softfloat_to_vfp_compare(env, \
359         FLOATTYPE ## _compare_quiet(a, b, &env->vfp.FPST)); \
360 } \
361 void VFP_HELPER(cmpe, P)(ARGTYPE a, ARGTYPE b, CPUARMState *env) \
362 { \
363     softfloat_to_vfp_compare(env, \
364         FLOATTYPE ## _compare(a, b, &env->vfp.FPST)); \
365 }
366 DO_VFP_cmp(h, float16, dh_ctype_f16, fp_status_f16)
367 DO_VFP_cmp(s, float32, float32, fp_status)
368 DO_VFP_cmp(d, float64, float64, fp_status)
369 #undef DO_VFP_cmp
370 
371 /* Integer to float and float to integer conversions */
372 
373 #define CONV_ITOF(name, ftype, fsz, sign)                           \
374 ftype HELPER(name)(uint32_t x, void *fpstp)                         \
375 {                                                                   \
376     float_status *fpst = fpstp;                                     \
377     return sign##int32_to_##float##fsz((sign##int32_t)x, fpst);     \
378 }
379 
380 #define CONV_FTOI(name, ftype, fsz, sign, round)                \
381 sign##int32_t HELPER(name)(ftype x, void *fpstp)                \
382 {                                                               \
383     float_status *fpst = fpstp;                                 \
384     if (float##fsz##_is_any_nan(x)) {                           \
385         float_raise(float_flag_invalid, fpst);                  \
386         return 0;                                               \
387     }                                                           \
388     return float##fsz##_to_##sign##int32##round(x, fpst);       \
389 }
390 
391 #define FLOAT_CONVS(name, p, ftype, fsz, sign)            \
392     CONV_ITOF(vfp_##name##to##p, ftype, fsz, sign)        \
393     CONV_FTOI(vfp_to##name##p, ftype, fsz, sign, )        \
394     CONV_FTOI(vfp_to##name##z##p, ftype, fsz, sign, _round_to_zero)
395 
396 FLOAT_CONVS(si, h, uint32_t, 16, )
397 FLOAT_CONVS(si, s, float32, 32, )
398 FLOAT_CONVS(si, d, float64, 64, )
399 FLOAT_CONVS(ui, h, uint32_t, 16, u)
400 FLOAT_CONVS(ui, s, float32, 32, u)
401 FLOAT_CONVS(ui, d, float64, 64, u)
402 
403 #undef CONV_ITOF
404 #undef CONV_FTOI
405 #undef FLOAT_CONVS
406 
407 /* floating point conversion */
408 float64 VFP_HELPER(fcvtd, s)(float32 x, CPUARMState *env)
409 {
410     return float32_to_float64(x, &env->vfp.fp_status);
411 }
412 
413 float32 VFP_HELPER(fcvts, d)(float64 x, CPUARMState *env)
414 {
415     return float64_to_float32(x, &env->vfp.fp_status);
416 }
417 
418 uint32_t HELPER(bfcvt)(float32 x, void *status)
419 {
420     return float32_to_bfloat16(x, status);
421 }
422 
423 uint32_t HELPER(bfcvt_pair)(uint64_t pair, void *status)
424 {
425     bfloat16 lo = float32_to_bfloat16(extract64(pair, 0, 32), status);
426     bfloat16 hi = float32_to_bfloat16(extract64(pair, 32, 32), status);
427     return deposit32(lo, 16, 16, hi);
428 }
429 
430 /*
431  * VFP3 fixed point conversion. The AArch32 versions of fix-to-float
432  * must always round-to-nearest; the AArch64 ones honour the FPSCR
433  * rounding mode. (For AArch32 Neon the standard-FPSCR is set to
434  * round-to-nearest so either helper will work.) AArch32 float-to-fix
435  * must round-to-zero.
436  */
437 #define VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)            \
438 ftype HELPER(vfp_##name##to##p)(uint##isz##_t  x, uint32_t shift,      \
439                                      void *fpstp) \
440 { return itype##_to_##float##fsz##_scalbn(x, -shift, fpstp); }
441 
442 #define VFP_CONV_FIX_FLOAT_ROUND(name, p, fsz, ftype, isz, itype)      \
443     ftype HELPER(vfp_##name##to##p##_round_to_nearest)(uint##isz##_t  x, \
444                                                      uint32_t shift,   \
445                                                      void *fpstp)      \
446     {                                                                  \
447         ftype ret;                                                     \
448         float_status *fpst = fpstp;                                    \
449         FloatRoundMode oldmode = fpst->float_rounding_mode;            \
450         fpst->float_rounding_mode = float_round_nearest_even;          \
451         ret = itype##_to_##float##fsz##_scalbn(x, -shift, fpstp);      \
452         fpst->float_rounding_mode = oldmode;                           \
453         return ret;                                                    \
454     }
455 
456 #define VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype, ROUND, suff) \
457 uint##isz##_t HELPER(vfp_to##name##p##suff)(ftype x, uint32_t shift,      \
458                                             void *fpst)                   \
459 {                                                                         \
460     if (unlikely(float##fsz##_is_any_nan(x))) {                           \
461         float_raise(float_flag_invalid, fpst);                            \
462         return 0;                                                         \
463     }                                                                     \
464     return float##fsz##_to_##itype##_scalbn(x, ROUND, shift, fpst);       \
465 }
466 
467 #define VFP_CONV_FIX(name, p, fsz, ftype, isz, itype)            \
468 VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)              \
469 VFP_CONV_FIX_FLOAT_ROUND(name, p, fsz, ftype, isz, itype)        \
470 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
471                          float_round_to_zero, _round_to_zero)    \
472 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
473                          get_float_rounding_mode(fpst), )
474 
475 #define VFP_CONV_FIX_A64(name, p, fsz, ftype, isz, itype)        \
476 VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)              \
477 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
478                          get_float_rounding_mode(fpst), )
479 
480 VFP_CONV_FIX(sh, d, 64, float64, 64, int16)
481 VFP_CONV_FIX(sl, d, 64, float64, 64, int32)
482 VFP_CONV_FIX_A64(sq, d, 64, float64, 64, int64)
483 VFP_CONV_FIX(uh, d, 64, float64, 64, uint16)
484 VFP_CONV_FIX(ul, d, 64, float64, 64, uint32)
485 VFP_CONV_FIX_A64(uq, d, 64, float64, 64, uint64)
486 VFP_CONV_FIX(sh, s, 32, float32, 32, int16)
487 VFP_CONV_FIX(sl, s, 32, float32, 32, int32)
488 VFP_CONV_FIX_A64(sq, s, 32, float32, 64, int64)
489 VFP_CONV_FIX(uh, s, 32, float32, 32, uint16)
490 VFP_CONV_FIX(ul, s, 32, float32, 32, uint32)
491 VFP_CONV_FIX_A64(uq, s, 32, float32, 64, uint64)
492 VFP_CONV_FIX(sh, h, 16, dh_ctype_f16, 32, int16)
493 VFP_CONV_FIX(sl, h, 16, dh_ctype_f16, 32, int32)
494 VFP_CONV_FIX_A64(sq, h, 16, dh_ctype_f16, 64, int64)
495 VFP_CONV_FIX(uh, h, 16, dh_ctype_f16, 32, uint16)
496 VFP_CONV_FIX(ul, h, 16, dh_ctype_f16, 32, uint32)
497 VFP_CONV_FIX_A64(uq, h, 16, dh_ctype_f16, 64, uint64)
498 
499 #undef VFP_CONV_FIX
500 #undef VFP_CONV_FIX_FLOAT
501 #undef VFP_CONV_FLOAT_FIX_ROUND
502 #undef VFP_CONV_FIX_A64
503 
504 /* Set the current fp rounding mode and return the old one.
505  * The argument is a softfloat float_round_ value.
506  */
507 uint32_t HELPER(set_rmode)(uint32_t rmode, void *fpstp)
508 {
509     float_status *fp_status = fpstp;
510 
511     uint32_t prev_rmode = get_float_rounding_mode(fp_status);
512     set_float_rounding_mode(rmode, fp_status);
513 
514     return prev_rmode;
515 }
516 
517 /* Half precision conversions.  */
518 float32 HELPER(vfp_fcvt_f16_to_f32)(uint32_t a, void *fpstp, uint32_t ahp_mode)
519 {
520     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
521      * it would affect flushing input denormals.
522      */
523     float_status *fpst = fpstp;
524     bool save = get_flush_inputs_to_zero(fpst);
525     set_flush_inputs_to_zero(false, fpst);
526     float32 r = float16_to_float32(a, !ahp_mode, fpst);
527     set_flush_inputs_to_zero(save, fpst);
528     return r;
529 }
530 
531 uint32_t HELPER(vfp_fcvt_f32_to_f16)(float32 a, void *fpstp, uint32_t ahp_mode)
532 {
533     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
534      * it would affect flushing output denormals.
535      */
536     float_status *fpst = fpstp;
537     bool save = get_flush_to_zero(fpst);
538     set_flush_to_zero(false, fpst);
539     float16 r = float32_to_float16(a, !ahp_mode, fpst);
540     set_flush_to_zero(save, fpst);
541     return r;
542 }
543 
544 float64 HELPER(vfp_fcvt_f16_to_f64)(uint32_t a, void *fpstp, uint32_t ahp_mode)
545 {
546     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
547      * it would affect flushing input denormals.
548      */
549     float_status *fpst = fpstp;
550     bool save = get_flush_inputs_to_zero(fpst);
551     set_flush_inputs_to_zero(false, fpst);
552     float64 r = float16_to_float64(a, !ahp_mode, fpst);
553     set_flush_inputs_to_zero(save, fpst);
554     return r;
555 }
556 
557 uint32_t HELPER(vfp_fcvt_f64_to_f16)(float64 a, void *fpstp, uint32_t ahp_mode)
558 {
559     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
560      * it would affect flushing output denormals.
561      */
562     float_status *fpst = fpstp;
563     bool save = get_flush_to_zero(fpst);
564     set_flush_to_zero(false, fpst);
565     float16 r = float64_to_float16(a, !ahp_mode, fpst);
566     set_flush_to_zero(save, fpst);
567     return r;
568 }
569 
570 /* NEON helpers.  */
571 
572 /* Constants 256 and 512 are used in some helpers; we avoid relying on
573  * int->float conversions at run-time.  */
574 #define float64_256 make_float64(0x4070000000000000LL)
575 #define float64_512 make_float64(0x4080000000000000LL)
576 #define float16_maxnorm make_float16(0x7bff)
577 #define float32_maxnorm make_float32(0x7f7fffff)
578 #define float64_maxnorm make_float64(0x7fefffffffffffffLL)
579 
580 /* Reciprocal functions
581  *
582  * The algorithm that must be used to calculate the estimate
583  * is specified by the ARM ARM, see FPRecipEstimate()/RecipEstimate
584  */
585 
586 /* See RecipEstimate()
587  *
588  * input is a 9 bit fixed point number
589  * input range 256 .. 511 for a number from 0.5 <= x < 1.0.
590  * result range 256 .. 511 for a number from 1.0 to 511/256.
591  */
592 
593 static int recip_estimate(int input)
594 {
595     int a, b, r;
596     assert(256 <= input && input < 512);
597     a = (input * 2) + 1;
598     b = (1 << 19) / a;
599     r = (b + 1) >> 1;
600     assert(256 <= r && r < 512);
601     return r;
602 }
603 
604 /*
605  * Common wrapper to call recip_estimate
606  *
607  * The parameters are exponent and 64 bit fraction (without implicit
608  * bit) where the binary point is nominally at bit 52. Returns a
609  * float64 which can then be rounded to the appropriate size by the
610  * callee.
611  */
612 
613 static uint64_t call_recip_estimate(int *exp, int exp_off, uint64_t frac)
614 {
615     uint32_t scaled, estimate;
616     uint64_t result_frac;
617     int result_exp;
618 
619     /* Handle sub-normals */
620     if (*exp == 0) {
621         if (extract64(frac, 51, 1) == 0) {
622             *exp = -1;
623             frac <<= 2;
624         } else {
625             frac <<= 1;
626         }
627     }
628 
629     /* scaled = UInt('1':fraction<51:44>) */
630     scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
631     estimate = recip_estimate(scaled);
632 
633     result_exp = exp_off - *exp;
634     result_frac = deposit64(0, 44, 8, estimate);
635     if (result_exp == 0) {
636         result_frac = deposit64(result_frac >> 1, 51, 1, 1);
637     } else if (result_exp == -1) {
638         result_frac = deposit64(result_frac >> 2, 50, 2, 1);
639         result_exp = 0;
640     }
641 
642     *exp = result_exp;
643 
644     return result_frac;
645 }
646 
647 static bool round_to_inf(float_status *fpst, bool sign_bit)
648 {
649     switch (fpst->float_rounding_mode) {
650     case float_round_nearest_even: /* Round to Nearest */
651         return true;
652     case float_round_up: /* Round to +Inf */
653         return !sign_bit;
654     case float_round_down: /* Round to -Inf */
655         return sign_bit;
656     case float_round_to_zero: /* Round to Zero */
657         return false;
658     default:
659         g_assert_not_reached();
660     }
661 }
662 
663 uint32_t HELPER(recpe_f16)(uint32_t input, void *fpstp)
664 {
665     float_status *fpst = fpstp;
666     float16 f16 = float16_squash_input_denormal(input, fpst);
667     uint32_t f16_val = float16_val(f16);
668     uint32_t f16_sign = float16_is_neg(f16);
669     int f16_exp = extract32(f16_val, 10, 5);
670     uint32_t f16_frac = extract32(f16_val, 0, 10);
671     uint64_t f64_frac;
672 
673     if (float16_is_any_nan(f16)) {
674         float16 nan = f16;
675         if (float16_is_signaling_nan(f16, fpst)) {
676             float_raise(float_flag_invalid, fpst);
677             if (!fpst->default_nan_mode) {
678                 nan = float16_silence_nan(f16, fpst);
679             }
680         }
681         if (fpst->default_nan_mode) {
682             nan =  float16_default_nan(fpst);
683         }
684         return nan;
685     } else if (float16_is_infinity(f16)) {
686         return float16_set_sign(float16_zero, float16_is_neg(f16));
687     } else if (float16_is_zero(f16)) {
688         float_raise(float_flag_divbyzero, fpst);
689         return float16_set_sign(float16_infinity, float16_is_neg(f16));
690     } else if (float16_abs(f16) < (1 << 8)) {
691         /* Abs(value) < 2.0^-16 */
692         float_raise(float_flag_overflow | float_flag_inexact, fpst);
693         if (round_to_inf(fpst, f16_sign)) {
694             return float16_set_sign(float16_infinity, f16_sign);
695         } else {
696             return float16_set_sign(float16_maxnorm, f16_sign);
697         }
698     } else if (f16_exp >= 29 && fpst->flush_to_zero) {
699         float_raise(float_flag_underflow, fpst);
700         return float16_set_sign(float16_zero, float16_is_neg(f16));
701     }
702 
703     f64_frac = call_recip_estimate(&f16_exp, 29,
704                                    ((uint64_t) f16_frac) << (52 - 10));
705 
706     /* result = sign : result_exp<4:0> : fraction<51:42> */
707     f16_val = deposit32(0, 15, 1, f16_sign);
708     f16_val = deposit32(f16_val, 10, 5, f16_exp);
709     f16_val = deposit32(f16_val, 0, 10, extract64(f64_frac, 52 - 10, 10));
710     return make_float16(f16_val);
711 }
712 
713 float32 HELPER(recpe_f32)(float32 input, void *fpstp)
714 {
715     float_status *fpst = fpstp;
716     float32 f32 = float32_squash_input_denormal(input, fpst);
717     uint32_t f32_val = float32_val(f32);
718     bool f32_sign = float32_is_neg(f32);
719     int f32_exp = extract32(f32_val, 23, 8);
720     uint32_t f32_frac = extract32(f32_val, 0, 23);
721     uint64_t f64_frac;
722 
723     if (float32_is_any_nan(f32)) {
724         float32 nan = f32;
725         if (float32_is_signaling_nan(f32, fpst)) {
726             float_raise(float_flag_invalid, fpst);
727             if (!fpst->default_nan_mode) {
728                 nan = float32_silence_nan(f32, fpst);
729             }
730         }
731         if (fpst->default_nan_mode) {
732             nan =  float32_default_nan(fpst);
733         }
734         return nan;
735     } else if (float32_is_infinity(f32)) {
736         return float32_set_sign(float32_zero, float32_is_neg(f32));
737     } else if (float32_is_zero(f32)) {
738         float_raise(float_flag_divbyzero, fpst);
739         return float32_set_sign(float32_infinity, float32_is_neg(f32));
740     } else if (float32_abs(f32) < (1ULL << 21)) {
741         /* Abs(value) < 2.0^-128 */
742         float_raise(float_flag_overflow | float_flag_inexact, fpst);
743         if (round_to_inf(fpst, f32_sign)) {
744             return float32_set_sign(float32_infinity, f32_sign);
745         } else {
746             return float32_set_sign(float32_maxnorm, f32_sign);
747         }
748     } else if (f32_exp >= 253 && fpst->flush_to_zero) {
749         float_raise(float_flag_underflow, fpst);
750         return float32_set_sign(float32_zero, float32_is_neg(f32));
751     }
752 
753     f64_frac = call_recip_estimate(&f32_exp, 253,
754                                    ((uint64_t) f32_frac) << (52 - 23));
755 
756     /* result = sign : result_exp<7:0> : fraction<51:29> */
757     f32_val = deposit32(0, 31, 1, f32_sign);
758     f32_val = deposit32(f32_val, 23, 8, f32_exp);
759     f32_val = deposit32(f32_val, 0, 23, extract64(f64_frac, 52 - 23, 23));
760     return make_float32(f32_val);
761 }
762 
763 float64 HELPER(recpe_f64)(float64 input, void *fpstp)
764 {
765     float_status *fpst = fpstp;
766     float64 f64 = float64_squash_input_denormal(input, fpst);
767     uint64_t f64_val = float64_val(f64);
768     bool f64_sign = float64_is_neg(f64);
769     int f64_exp = extract64(f64_val, 52, 11);
770     uint64_t f64_frac = extract64(f64_val, 0, 52);
771 
772     /* Deal with any special cases */
773     if (float64_is_any_nan(f64)) {
774         float64 nan = f64;
775         if (float64_is_signaling_nan(f64, fpst)) {
776             float_raise(float_flag_invalid, fpst);
777             if (!fpst->default_nan_mode) {
778                 nan = float64_silence_nan(f64, fpst);
779             }
780         }
781         if (fpst->default_nan_mode) {
782             nan =  float64_default_nan(fpst);
783         }
784         return nan;
785     } else if (float64_is_infinity(f64)) {
786         return float64_set_sign(float64_zero, float64_is_neg(f64));
787     } else if (float64_is_zero(f64)) {
788         float_raise(float_flag_divbyzero, fpst);
789         return float64_set_sign(float64_infinity, float64_is_neg(f64));
790     } else if ((f64_val & ~(1ULL << 63)) < (1ULL << 50)) {
791         /* Abs(value) < 2.0^-1024 */
792         float_raise(float_flag_overflow | float_flag_inexact, fpst);
793         if (round_to_inf(fpst, f64_sign)) {
794             return float64_set_sign(float64_infinity, f64_sign);
795         } else {
796             return float64_set_sign(float64_maxnorm, f64_sign);
797         }
798     } else if (f64_exp >= 2045 && fpst->flush_to_zero) {
799         float_raise(float_flag_underflow, fpst);
800         return float64_set_sign(float64_zero, float64_is_neg(f64));
801     }
802 
803     f64_frac = call_recip_estimate(&f64_exp, 2045, f64_frac);
804 
805     /* result = sign : result_exp<10:0> : fraction<51:0>; */
806     f64_val = deposit64(0, 63, 1, f64_sign);
807     f64_val = deposit64(f64_val, 52, 11, f64_exp);
808     f64_val = deposit64(f64_val, 0, 52, f64_frac);
809     return make_float64(f64_val);
810 }
811 
812 /* The algorithm that must be used to calculate the estimate
813  * is specified by the ARM ARM.
814  */
815 
816 static int do_recip_sqrt_estimate(int a)
817 {
818     int b, estimate;
819 
820     assert(128 <= a && a < 512);
821     if (a < 256) {
822         a = a * 2 + 1;
823     } else {
824         a = (a >> 1) << 1;
825         a = (a + 1) * 2;
826     }
827     b = 512;
828     while (a * (b + 1) * (b + 1) < (1 << 28)) {
829         b += 1;
830     }
831     estimate = (b + 1) / 2;
832     assert(256 <= estimate && estimate < 512);
833 
834     return estimate;
835 }
836 
837 
838 static uint64_t recip_sqrt_estimate(int *exp , int exp_off, uint64_t frac)
839 {
840     int estimate;
841     uint32_t scaled;
842 
843     if (*exp == 0) {
844         while (extract64(frac, 51, 1) == 0) {
845             frac = frac << 1;
846             *exp -= 1;
847         }
848         frac = extract64(frac, 0, 51) << 1;
849     }
850 
851     if (*exp & 1) {
852         /* scaled = UInt('01':fraction<51:45>) */
853         scaled = deposit32(1 << 7, 0, 7, extract64(frac, 45, 7));
854     } else {
855         /* scaled = UInt('1':fraction<51:44>) */
856         scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
857     }
858     estimate = do_recip_sqrt_estimate(scaled);
859 
860     *exp = (exp_off - *exp) / 2;
861     return extract64(estimate, 0, 8) << 44;
862 }
863 
864 uint32_t HELPER(rsqrte_f16)(uint32_t input, void *fpstp)
865 {
866     float_status *s = fpstp;
867     float16 f16 = float16_squash_input_denormal(input, s);
868     uint16_t val = float16_val(f16);
869     bool f16_sign = float16_is_neg(f16);
870     int f16_exp = extract32(val, 10, 5);
871     uint16_t f16_frac = extract32(val, 0, 10);
872     uint64_t f64_frac;
873 
874     if (float16_is_any_nan(f16)) {
875         float16 nan = f16;
876         if (float16_is_signaling_nan(f16, s)) {
877             float_raise(float_flag_invalid, s);
878             if (!s->default_nan_mode) {
879                 nan = float16_silence_nan(f16, fpstp);
880             }
881         }
882         if (s->default_nan_mode) {
883             nan =  float16_default_nan(s);
884         }
885         return nan;
886     } else if (float16_is_zero(f16)) {
887         float_raise(float_flag_divbyzero, s);
888         return float16_set_sign(float16_infinity, f16_sign);
889     } else if (f16_sign) {
890         float_raise(float_flag_invalid, s);
891         return float16_default_nan(s);
892     } else if (float16_is_infinity(f16)) {
893         return float16_zero;
894     }
895 
896     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
897      * preserving the parity of the exponent.  */
898 
899     f64_frac = ((uint64_t) f16_frac) << (52 - 10);
900 
901     f64_frac = recip_sqrt_estimate(&f16_exp, 44, f64_frac);
902 
903     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(2) */
904     val = deposit32(0, 15, 1, f16_sign);
905     val = deposit32(val, 10, 5, f16_exp);
906     val = deposit32(val, 2, 8, extract64(f64_frac, 52 - 8, 8));
907     return make_float16(val);
908 }
909 
910 float32 HELPER(rsqrte_f32)(float32 input, void *fpstp)
911 {
912     float_status *s = fpstp;
913     float32 f32 = float32_squash_input_denormal(input, s);
914     uint32_t val = float32_val(f32);
915     uint32_t f32_sign = float32_is_neg(f32);
916     int f32_exp = extract32(val, 23, 8);
917     uint32_t f32_frac = extract32(val, 0, 23);
918     uint64_t f64_frac;
919 
920     if (float32_is_any_nan(f32)) {
921         float32 nan = f32;
922         if (float32_is_signaling_nan(f32, s)) {
923             float_raise(float_flag_invalid, s);
924             if (!s->default_nan_mode) {
925                 nan = float32_silence_nan(f32, fpstp);
926             }
927         }
928         if (s->default_nan_mode) {
929             nan =  float32_default_nan(s);
930         }
931         return nan;
932     } else if (float32_is_zero(f32)) {
933         float_raise(float_flag_divbyzero, s);
934         return float32_set_sign(float32_infinity, float32_is_neg(f32));
935     } else if (float32_is_neg(f32)) {
936         float_raise(float_flag_invalid, s);
937         return float32_default_nan(s);
938     } else if (float32_is_infinity(f32)) {
939         return float32_zero;
940     }
941 
942     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
943      * preserving the parity of the exponent.  */
944 
945     f64_frac = ((uint64_t) f32_frac) << 29;
946 
947     f64_frac = recip_sqrt_estimate(&f32_exp, 380, f64_frac);
948 
949     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(15) */
950     val = deposit32(0, 31, 1, f32_sign);
951     val = deposit32(val, 23, 8, f32_exp);
952     val = deposit32(val, 15, 8, extract64(f64_frac, 52 - 8, 8));
953     return make_float32(val);
954 }
955 
956 float64 HELPER(rsqrte_f64)(float64 input, void *fpstp)
957 {
958     float_status *s = fpstp;
959     float64 f64 = float64_squash_input_denormal(input, s);
960     uint64_t val = float64_val(f64);
961     bool f64_sign = float64_is_neg(f64);
962     int f64_exp = extract64(val, 52, 11);
963     uint64_t f64_frac = extract64(val, 0, 52);
964 
965     if (float64_is_any_nan(f64)) {
966         float64 nan = f64;
967         if (float64_is_signaling_nan(f64, s)) {
968             float_raise(float_flag_invalid, s);
969             if (!s->default_nan_mode) {
970                 nan = float64_silence_nan(f64, fpstp);
971             }
972         }
973         if (s->default_nan_mode) {
974             nan =  float64_default_nan(s);
975         }
976         return nan;
977     } else if (float64_is_zero(f64)) {
978         float_raise(float_flag_divbyzero, s);
979         return float64_set_sign(float64_infinity, float64_is_neg(f64));
980     } else if (float64_is_neg(f64)) {
981         float_raise(float_flag_invalid, s);
982         return float64_default_nan(s);
983     } else if (float64_is_infinity(f64)) {
984         return float64_zero;
985     }
986 
987     f64_frac = recip_sqrt_estimate(&f64_exp, 3068, f64_frac);
988 
989     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(44) */
990     val = deposit64(0, 61, 1, f64_sign);
991     val = deposit64(val, 52, 11, f64_exp);
992     val = deposit64(val, 44, 8, extract64(f64_frac, 52 - 8, 8));
993     return make_float64(val);
994 }
995 
996 uint32_t HELPER(recpe_u32)(uint32_t a)
997 {
998     int input, estimate;
999 
1000     if ((a & 0x80000000) == 0) {
1001         return 0xffffffff;
1002     }
1003 
1004     input = extract32(a, 23, 9);
1005     estimate = recip_estimate(input);
1006 
1007     return deposit32(0, (32 - 9), 9, estimate);
1008 }
1009 
1010 uint32_t HELPER(rsqrte_u32)(uint32_t a)
1011 {
1012     int estimate;
1013 
1014     if ((a & 0xc0000000) == 0) {
1015         return 0xffffffff;
1016     }
1017 
1018     estimate = do_recip_sqrt_estimate(extract32(a, 23, 9));
1019 
1020     return deposit32(0, 23, 9, estimate);
1021 }
1022 
1023 /* VFPv4 fused multiply-accumulate */
1024 dh_ctype_f16 VFP_HELPER(muladd, h)(dh_ctype_f16 a, dh_ctype_f16 b,
1025                                    dh_ctype_f16 c, void *fpstp)
1026 {
1027     float_status *fpst = fpstp;
1028     return float16_muladd(a, b, c, 0, fpst);
1029 }
1030 
1031 float32 VFP_HELPER(muladd, s)(float32 a, float32 b, float32 c, void *fpstp)
1032 {
1033     float_status *fpst = fpstp;
1034     return float32_muladd(a, b, c, 0, fpst);
1035 }
1036 
1037 float64 VFP_HELPER(muladd, d)(float64 a, float64 b, float64 c, void *fpstp)
1038 {
1039     float_status *fpst = fpstp;
1040     return float64_muladd(a, b, c, 0, fpst);
1041 }
1042 
1043 /* ARMv8 round to integral */
1044 dh_ctype_f16 HELPER(rinth_exact)(dh_ctype_f16 x, void *fp_status)
1045 {
1046     return float16_round_to_int(x, fp_status);
1047 }
1048 
1049 float32 HELPER(rints_exact)(float32 x, void *fp_status)
1050 {
1051     return float32_round_to_int(x, fp_status);
1052 }
1053 
1054 float64 HELPER(rintd_exact)(float64 x, void *fp_status)
1055 {
1056     return float64_round_to_int(x, fp_status);
1057 }
1058 
1059 dh_ctype_f16 HELPER(rinth)(dh_ctype_f16 x, void *fp_status)
1060 {
1061     int old_flags = get_float_exception_flags(fp_status), new_flags;
1062     float16 ret;
1063 
1064     ret = float16_round_to_int(x, fp_status);
1065 
1066     /* Suppress any inexact exceptions the conversion produced */
1067     if (!(old_flags & float_flag_inexact)) {
1068         new_flags = get_float_exception_flags(fp_status);
1069         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1070     }
1071 
1072     return ret;
1073 }
1074 
1075 float32 HELPER(rints)(float32 x, void *fp_status)
1076 {
1077     int old_flags = get_float_exception_flags(fp_status), new_flags;
1078     float32 ret;
1079 
1080     ret = float32_round_to_int(x, fp_status);
1081 
1082     /* Suppress any inexact exceptions the conversion produced */
1083     if (!(old_flags & float_flag_inexact)) {
1084         new_flags = get_float_exception_flags(fp_status);
1085         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1086     }
1087 
1088     return ret;
1089 }
1090 
1091 float64 HELPER(rintd)(float64 x, void *fp_status)
1092 {
1093     int old_flags = get_float_exception_flags(fp_status), new_flags;
1094     float64 ret;
1095 
1096     ret = float64_round_to_int(x, fp_status);
1097 
1098     new_flags = get_float_exception_flags(fp_status);
1099 
1100     /* Suppress any inexact exceptions the conversion produced */
1101     if (!(old_flags & float_flag_inexact)) {
1102         new_flags = get_float_exception_flags(fp_status);
1103         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1104     }
1105 
1106     return ret;
1107 }
1108 
1109 /* Convert ARM rounding mode to softfloat */
1110 const FloatRoundMode arm_rmode_to_sf_map[] = {
1111     [FPROUNDING_TIEEVEN] = float_round_nearest_even,
1112     [FPROUNDING_POSINF] = float_round_up,
1113     [FPROUNDING_NEGINF] = float_round_down,
1114     [FPROUNDING_ZERO] = float_round_to_zero,
1115     [FPROUNDING_TIEAWAY] = float_round_ties_away,
1116     [FPROUNDING_ODD] = float_round_to_odd,
1117 };
1118 
1119 /*
1120  * Implement float64 to int32_t conversion without saturation;
1121  * the result is supplied modulo 2^32.
1122  */
1123 uint64_t HELPER(fjcvtzs)(float64 value, void *vstatus)
1124 {
1125     float_status *status = vstatus;
1126     uint32_t frac, e_old, e_new;
1127     bool inexact;
1128 
1129     e_old = get_float_exception_flags(status);
1130     set_float_exception_flags(0, status);
1131     frac = float64_to_int32_modulo(value, float_round_to_zero, status);
1132     e_new = get_float_exception_flags(status);
1133     set_float_exception_flags(e_old | e_new, status);
1134 
1135     /* Normal inexact, denormal with flush-to-zero, or overflow or NaN */
1136     inexact = e_new & (float_flag_inexact |
1137                        float_flag_input_denormal |
1138                        float_flag_invalid);
1139 
1140     /* While not inexact for IEEE FP, -0.0 is inexact for JavaScript. */
1141     inexact |= value == float64_chs(float64_zero);
1142 
1143     /* Pack the result and the env->ZF representation of Z together.  */
1144     return deposit64(frac, 32, 32, inexact);
1145 }
1146 
1147 uint32_t HELPER(vjcvt)(float64 value, CPUARMState *env)
1148 {
1149     uint64_t pair = HELPER(fjcvtzs)(value, &env->vfp.fp_status);
1150     uint32_t result = pair;
1151     uint32_t z = (pair >> 32) == 0;
1152 
1153     /* Store Z, clear NCV, in FPSCR.NZCV.  */
1154     env->vfp.fpsr = (env->vfp.fpsr & ~FPSR_NZCV_MASK) | (z * FPSR_Z);
1155 
1156     return result;
1157 }
1158 
1159 /* Round a float32 to an integer that fits in int32_t or int64_t.  */
1160 static float32 frint_s(float32 f, float_status *fpst, int intsize)
1161 {
1162     int old_flags = get_float_exception_flags(fpst);
1163     uint32_t exp = extract32(f, 23, 8);
1164 
1165     if (unlikely(exp == 0xff)) {
1166         /* NaN or Inf.  */
1167         goto overflow;
1168     }
1169 
1170     /* Round and re-extract the exponent.  */
1171     f = float32_round_to_int(f, fpst);
1172     exp = extract32(f, 23, 8);
1173 
1174     /* Validate the range of the result.  */
1175     if (exp < 126 + intsize) {
1176         /* abs(F) <= INT{N}_MAX */
1177         return f;
1178     }
1179     if (exp == 126 + intsize) {
1180         uint32_t sign = extract32(f, 31, 1);
1181         uint32_t frac = extract32(f, 0, 23);
1182         if (sign && frac == 0) {
1183             /* F == INT{N}_MIN */
1184             return f;
1185         }
1186     }
1187 
1188  overflow:
1189     /*
1190      * Raise Invalid and return INT{N}_MIN as a float.  Revert any
1191      * inexact exception float32_round_to_int may have raised.
1192      */
1193     set_float_exception_flags(old_flags | float_flag_invalid, fpst);
1194     return (0x100u + 126u + intsize) << 23;
1195 }
1196 
1197 float32 HELPER(frint32_s)(float32 f, void *fpst)
1198 {
1199     return frint_s(f, fpst, 32);
1200 }
1201 
1202 float32 HELPER(frint64_s)(float32 f, void *fpst)
1203 {
1204     return frint_s(f, fpst, 64);
1205 }
1206 
1207 /* Round a float64 to an integer that fits in int32_t or int64_t.  */
1208 static float64 frint_d(float64 f, float_status *fpst, int intsize)
1209 {
1210     int old_flags = get_float_exception_flags(fpst);
1211     uint32_t exp = extract64(f, 52, 11);
1212 
1213     if (unlikely(exp == 0x7ff)) {
1214         /* NaN or Inf.  */
1215         goto overflow;
1216     }
1217 
1218     /* Round and re-extract the exponent.  */
1219     f = float64_round_to_int(f, fpst);
1220     exp = extract64(f, 52, 11);
1221 
1222     /* Validate the range of the result.  */
1223     if (exp < 1022 + intsize) {
1224         /* abs(F) <= INT{N}_MAX */
1225         return f;
1226     }
1227     if (exp == 1022 + intsize) {
1228         uint64_t sign = extract64(f, 63, 1);
1229         uint64_t frac = extract64(f, 0, 52);
1230         if (sign && frac == 0) {
1231             /* F == INT{N}_MIN */
1232             return f;
1233         }
1234     }
1235 
1236  overflow:
1237     /*
1238      * Raise Invalid and return INT{N}_MIN as a float.  Revert any
1239      * inexact exception float64_round_to_int may have raised.
1240      */
1241     set_float_exception_flags(old_flags | float_flag_invalid, fpst);
1242     return (uint64_t)(0x800 + 1022 + intsize) << 52;
1243 }
1244 
1245 float64 HELPER(frint32_d)(float64 f, void *fpst)
1246 {
1247     return frint_d(f, fpst, 32);
1248 }
1249 
1250 float64 HELPER(frint64_d)(float64 f, void *fpst)
1251 {
1252     return frint_d(f, fpst, 64);
1253 }
1254 
1255 void HELPER(check_hcr_el2_trap)(CPUARMState *env, uint32_t rt, uint32_t reg)
1256 {
1257     uint32_t syndrome;
1258 
1259     switch (reg) {
1260     case ARM_VFP_MVFR0:
1261     case ARM_VFP_MVFR1:
1262     case ARM_VFP_MVFR2:
1263         if (!(arm_hcr_el2_eff(env) & HCR_TID3)) {
1264             return;
1265         }
1266         break;
1267     case ARM_VFP_FPSID:
1268         if (!(arm_hcr_el2_eff(env) & HCR_TID0)) {
1269             return;
1270         }
1271         break;
1272     default:
1273         g_assert_not_reached();
1274     }
1275 
1276     syndrome = ((EC_FPIDTRAP << ARM_EL_EC_SHIFT)
1277                 | ARM_EL_IL
1278                 | (1 << 24) | (0xe << 20) | (7 << 14)
1279                 | (reg << 10) | (rt << 5) | 1);
1280 
1281     raise_exception(env, EXCP_HYP_TRAP, syndrome, 2);
1282 }
1283 
1284 #endif
1285