xref: /openbmc/qemu/target/arm/vfp_helper.c (revision 69430111)
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     /* When ARMv8.2-FP16 is not supported, FZ16 is RES0.  */
199     if (!cpu_isar_feature(any_fp16, env_archcpu(env))) {
200         val &= ~FPCR_FZ16;
201     }
202 
203     vfp_set_fpscr_to_host(env, val);
204 
205     if (!arm_feature(env, ARM_FEATURE_M)) {
206         /*
207          * Short-vector length and stride; on M-profile these bits
208          * are used for different purposes.
209          * We can't make this conditional be "if MVFR0.FPShVec != 0",
210          * because in v7A no-short-vector-support cores still had to
211          * allow Stride/Len to be written with the only effect that
212          * some insns are required to UNDEF if the guest sets them.
213          *
214          * TODO: if M-profile MVE implemented, set LTPSIZE.
215          */
216         env->vfp.vec_len = extract32(val, 16, 3);
217         env->vfp.vec_stride = extract32(val, 20, 2);
218     }
219 
220     if (arm_feature(env, ARM_FEATURE_NEON)) {
221         /*
222          * The bit we set within fpscr_q is arbitrary; the register as a
223          * whole being zero/non-zero is what counts.
224          * TODO: M-profile MVE also has a QC bit.
225          */
226         env->vfp.qc[0] = val & FPCR_QC;
227         env->vfp.qc[1] = 0;
228         env->vfp.qc[2] = 0;
229         env->vfp.qc[3] = 0;
230     }
231 
232     /*
233      * We don't implement trapped exception handling, so the
234      * trap enable bits, IDE|IXE|UFE|OFE|DZE|IOE are all RAZ/WI (not RES0!)
235      *
236      * The exception flags IOC|DZC|OFC|UFC|IXC|IDC are stored in
237      * fp_status; QC, Len and Stride are stored separately earlier.
238      * Clear out all of those and the RES0 bits: only NZCV, AHP, DN,
239      * FZ, RMode and FZ16 are kept in vfp.xregs[FPSCR].
240      */
241     env->vfp.xregs[ARM_VFP_FPSCR] = val & 0xf7c80000;
242 }
243 
244 void vfp_set_fpscr(CPUARMState *env, uint32_t val)
245 {
246     HELPER(vfp_set_fpscr)(env, val);
247 }
248 
249 #ifdef CONFIG_TCG
250 
251 #define VFP_HELPER(name, p) HELPER(glue(glue(vfp_,name),p))
252 
253 #define VFP_BINOP(name) \
254 dh_ctype_f16 VFP_HELPER(name, h)(dh_ctype_f16 a, dh_ctype_f16 b, void *fpstp) \
255 { \
256     float_status *fpst = fpstp; \
257     return float16_ ## name(a, b, fpst); \
258 } \
259 float32 VFP_HELPER(name, s)(float32 a, float32 b, void *fpstp) \
260 { \
261     float_status *fpst = fpstp; \
262     return float32_ ## name(a, b, fpst); \
263 } \
264 float64 VFP_HELPER(name, d)(float64 a, float64 b, void *fpstp) \
265 { \
266     float_status *fpst = fpstp; \
267     return float64_ ## name(a, b, fpst); \
268 }
269 VFP_BINOP(add)
270 VFP_BINOP(sub)
271 VFP_BINOP(mul)
272 VFP_BINOP(div)
273 VFP_BINOP(min)
274 VFP_BINOP(max)
275 VFP_BINOP(minnum)
276 VFP_BINOP(maxnum)
277 #undef VFP_BINOP
278 
279 dh_ctype_f16 VFP_HELPER(neg, h)(dh_ctype_f16 a)
280 {
281     return float16_chs(a);
282 }
283 
284 float32 VFP_HELPER(neg, s)(float32 a)
285 {
286     return float32_chs(a);
287 }
288 
289 float64 VFP_HELPER(neg, d)(float64 a)
290 {
291     return float64_chs(a);
292 }
293 
294 dh_ctype_f16 VFP_HELPER(abs, h)(dh_ctype_f16 a)
295 {
296     return float16_abs(a);
297 }
298 
299 float32 VFP_HELPER(abs, s)(float32 a)
300 {
301     return float32_abs(a);
302 }
303 
304 float64 VFP_HELPER(abs, d)(float64 a)
305 {
306     return float64_abs(a);
307 }
308 
309 dh_ctype_f16 VFP_HELPER(sqrt, h)(dh_ctype_f16 a, CPUARMState *env)
310 {
311     return float16_sqrt(a, &env->vfp.fp_status_f16);
312 }
313 
314 float32 VFP_HELPER(sqrt, s)(float32 a, CPUARMState *env)
315 {
316     return float32_sqrt(a, &env->vfp.fp_status);
317 }
318 
319 float64 VFP_HELPER(sqrt, d)(float64 a, CPUARMState *env)
320 {
321     return float64_sqrt(a, &env->vfp.fp_status);
322 }
323 
324 static void softfloat_to_vfp_compare(CPUARMState *env, FloatRelation cmp)
325 {
326     uint32_t flags;
327     switch (cmp) {
328     case float_relation_equal:
329         flags = 0x6;
330         break;
331     case float_relation_less:
332         flags = 0x8;
333         break;
334     case float_relation_greater:
335         flags = 0x2;
336         break;
337     case float_relation_unordered:
338         flags = 0x3;
339         break;
340     default:
341         g_assert_not_reached();
342     }
343     env->vfp.xregs[ARM_VFP_FPSCR] =
344         deposit32(env->vfp.xregs[ARM_VFP_FPSCR], 28, 4, flags);
345 }
346 
347 /* XXX: check quiet/signaling case */
348 #define DO_VFP_cmp(P, FLOATTYPE, ARGTYPE, FPST) \
349 void VFP_HELPER(cmp, P)(ARGTYPE a, ARGTYPE b, CPUARMState *env)  \
350 { \
351     softfloat_to_vfp_compare(env, \
352         FLOATTYPE ## _compare_quiet(a, b, &env->vfp.FPST)); \
353 } \
354 void VFP_HELPER(cmpe, P)(ARGTYPE a, ARGTYPE b, CPUARMState *env) \
355 { \
356     softfloat_to_vfp_compare(env, \
357         FLOATTYPE ## _compare(a, b, &env->vfp.FPST)); \
358 }
359 DO_VFP_cmp(h, float16, dh_ctype_f16, fp_status_f16)
360 DO_VFP_cmp(s, float32, float32, fp_status)
361 DO_VFP_cmp(d, float64, float64, fp_status)
362 #undef DO_VFP_cmp
363 
364 /* Integer to float and float to integer conversions */
365 
366 #define CONV_ITOF(name, ftype, fsz, sign)                           \
367 ftype HELPER(name)(uint32_t x, void *fpstp)                         \
368 {                                                                   \
369     float_status *fpst = fpstp;                                     \
370     return sign##int32_to_##float##fsz((sign##int32_t)x, fpst);     \
371 }
372 
373 #define CONV_FTOI(name, ftype, fsz, sign, round)                \
374 sign##int32_t HELPER(name)(ftype x, void *fpstp)                \
375 {                                                               \
376     float_status *fpst = fpstp;                                 \
377     if (float##fsz##_is_any_nan(x)) {                           \
378         float_raise(float_flag_invalid, fpst);                  \
379         return 0;                                               \
380     }                                                           \
381     return float##fsz##_to_##sign##int32##round(x, fpst);       \
382 }
383 
384 #define FLOAT_CONVS(name, p, ftype, fsz, sign)            \
385     CONV_ITOF(vfp_##name##to##p, ftype, fsz, sign)        \
386     CONV_FTOI(vfp_to##name##p, ftype, fsz, sign, )        \
387     CONV_FTOI(vfp_to##name##z##p, ftype, fsz, sign, _round_to_zero)
388 
389 FLOAT_CONVS(si, h, uint32_t, 16, )
390 FLOAT_CONVS(si, s, float32, 32, )
391 FLOAT_CONVS(si, d, float64, 64, )
392 FLOAT_CONVS(ui, h, uint32_t, 16, u)
393 FLOAT_CONVS(ui, s, float32, 32, u)
394 FLOAT_CONVS(ui, d, float64, 64, u)
395 
396 #undef CONV_ITOF
397 #undef CONV_FTOI
398 #undef FLOAT_CONVS
399 
400 /* floating point conversion */
401 float64 VFP_HELPER(fcvtd, s)(float32 x, CPUARMState *env)
402 {
403     return float32_to_float64(x, &env->vfp.fp_status);
404 }
405 
406 float32 VFP_HELPER(fcvts, d)(float64 x, CPUARMState *env)
407 {
408     return float64_to_float32(x, &env->vfp.fp_status);
409 }
410 
411 /*
412  * VFP3 fixed point conversion. The AArch32 versions of fix-to-float
413  * must always round-to-nearest; the AArch64 ones honour the FPSCR
414  * rounding mode. (For AArch32 Neon the standard-FPSCR is set to
415  * round-to-nearest so either helper will work.) AArch32 float-to-fix
416  * must round-to-zero.
417  */
418 #define VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)            \
419 ftype HELPER(vfp_##name##to##p)(uint##isz##_t  x, uint32_t shift,      \
420                                      void *fpstp) \
421 { return itype##_to_##float##fsz##_scalbn(x, -shift, fpstp); }
422 
423 #define VFP_CONV_FIX_FLOAT_ROUND(name, p, fsz, ftype, isz, itype)      \
424     ftype HELPER(vfp_##name##to##p##_round_to_nearest)(uint##isz##_t  x, \
425                                                      uint32_t shift,   \
426                                                      void *fpstp)      \
427     {                                                                  \
428         ftype ret;                                                     \
429         float_status *fpst = fpstp;                                    \
430         FloatRoundMode oldmode = fpst->float_rounding_mode;            \
431         fpst->float_rounding_mode = float_round_nearest_even;          \
432         ret = itype##_to_##float##fsz##_scalbn(x, -shift, fpstp);      \
433         fpst->float_rounding_mode = oldmode;                           \
434         return ret;                                                    \
435     }
436 
437 #define VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype, ROUND, suff) \
438 uint##isz##_t HELPER(vfp_to##name##p##suff)(ftype x, uint32_t shift,      \
439                                             void *fpst)                   \
440 {                                                                         \
441     if (unlikely(float##fsz##_is_any_nan(x))) {                           \
442         float_raise(float_flag_invalid, fpst);                            \
443         return 0;                                                         \
444     }                                                                     \
445     return float##fsz##_to_##itype##_scalbn(x, ROUND, shift, fpst);       \
446 }
447 
448 #define VFP_CONV_FIX(name, p, fsz, ftype, isz, itype)            \
449 VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)              \
450 VFP_CONV_FIX_FLOAT_ROUND(name, p, fsz, ftype, isz, itype)        \
451 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
452                          float_round_to_zero, _round_to_zero)    \
453 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
454                          get_float_rounding_mode(fpst), )
455 
456 #define VFP_CONV_FIX_A64(name, p, fsz, ftype, isz, itype)        \
457 VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)              \
458 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
459                          get_float_rounding_mode(fpst), )
460 
461 VFP_CONV_FIX(sh, d, 64, float64, 64, int16)
462 VFP_CONV_FIX(sl, d, 64, float64, 64, int32)
463 VFP_CONV_FIX_A64(sq, d, 64, float64, 64, int64)
464 VFP_CONV_FIX(uh, d, 64, float64, 64, uint16)
465 VFP_CONV_FIX(ul, d, 64, float64, 64, uint32)
466 VFP_CONV_FIX_A64(uq, d, 64, float64, 64, uint64)
467 VFP_CONV_FIX(sh, s, 32, float32, 32, int16)
468 VFP_CONV_FIX(sl, s, 32, float32, 32, int32)
469 VFP_CONV_FIX_A64(sq, s, 32, float32, 64, int64)
470 VFP_CONV_FIX(uh, s, 32, float32, 32, uint16)
471 VFP_CONV_FIX(ul, s, 32, float32, 32, uint32)
472 VFP_CONV_FIX_A64(uq, s, 32, float32, 64, uint64)
473 VFP_CONV_FIX(sh, h, 16, dh_ctype_f16, 32, int16)
474 VFP_CONV_FIX(sl, h, 16, dh_ctype_f16, 32, int32)
475 VFP_CONV_FIX_A64(sq, h, 16, dh_ctype_f16, 64, int64)
476 VFP_CONV_FIX(uh, h, 16, dh_ctype_f16, 32, uint16)
477 VFP_CONV_FIX(ul, h, 16, dh_ctype_f16, 32, uint32)
478 VFP_CONV_FIX_A64(uq, h, 16, dh_ctype_f16, 64, uint64)
479 
480 #undef VFP_CONV_FIX
481 #undef VFP_CONV_FIX_FLOAT
482 #undef VFP_CONV_FLOAT_FIX_ROUND
483 #undef VFP_CONV_FIX_A64
484 
485 /* Set the current fp rounding mode and return the old one.
486  * The argument is a softfloat float_round_ value.
487  */
488 uint32_t HELPER(set_rmode)(uint32_t rmode, void *fpstp)
489 {
490     float_status *fp_status = fpstp;
491 
492     uint32_t prev_rmode = get_float_rounding_mode(fp_status);
493     set_float_rounding_mode(rmode, fp_status);
494 
495     return prev_rmode;
496 }
497 
498 /* Half precision conversions.  */
499 float32 HELPER(vfp_fcvt_f16_to_f32)(uint32_t a, void *fpstp, uint32_t ahp_mode)
500 {
501     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
502      * it would affect flushing input denormals.
503      */
504     float_status *fpst = fpstp;
505     bool save = get_flush_inputs_to_zero(fpst);
506     set_flush_inputs_to_zero(false, fpst);
507     float32 r = float16_to_float32(a, !ahp_mode, fpst);
508     set_flush_inputs_to_zero(save, fpst);
509     return r;
510 }
511 
512 uint32_t HELPER(vfp_fcvt_f32_to_f16)(float32 a, void *fpstp, uint32_t ahp_mode)
513 {
514     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
515      * it would affect flushing output denormals.
516      */
517     float_status *fpst = fpstp;
518     bool save = get_flush_to_zero(fpst);
519     set_flush_to_zero(false, fpst);
520     float16 r = float32_to_float16(a, !ahp_mode, fpst);
521     set_flush_to_zero(save, fpst);
522     return r;
523 }
524 
525 float64 HELPER(vfp_fcvt_f16_to_f64)(uint32_t a, void *fpstp, uint32_t ahp_mode)
526 {
527     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
528      * it would affect flushing input denormals.
529      */
530     float_status *fpst = fpstp;
531     bool save = get_flush_inputs_to_zero(fpst);
532     set_flush_inputs_to_zero(false, fpst);
533     float64 r = float16_to_float64(a, !ahp_mode, fpst);
534     set_flush_inputs_to_zero(save, fpst);
535     return r;
536 }
537 
538 uint32_t HELPER(vfp_fcvt_f64_to_f16)(float64 a, void *fpstp, uint32_t ahp_mode)
539 {
540     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
541      * it would affect flushing output denormals.
542      */
543     float_status *fpst = fpstp;
544     bool save = get_flush_to_zero(fpst);
545     set_flush_to_zero(false, fpst);
546     float16 r = float64_to_float16(a, !ahp_mode, fpst);
547     set_flush_to_zero(save, fpst);
548     return r;
549 }
550 
551 /* NEON helpers.  */
552 
553 /* Constants 256 and 512 are used in some helpers; we avoid relying on
554  * int->float conversions at run-time.  */
555 #define float64_256 make_float64(0x4070000000000000LL)
556 #define float64_512 make_float64(0x4080000000000000LL)
557 #define float16_maxnorm make_float16(0x7bff)
558 #define float32_maxnorm make_float32(0x7f7fffff)
559 #define float64_maxnorm make_float64(0x7fefffffffffffffLL)
560 
561 /* Reciprocal functions
562  *
563  * The algorithm that must be used to calculate the estimate
564  * is specified by the ARM ARM, see FPRecipEstimate()/RecipEstimate
565  */
566 
567 /* See RecipEstimate()
568  *
569  * input is a 9 bit fixed point number
570  * input range 256 .. 511 for a number from 0.5 <= x < 1.0.
571  * result range 256 .. 511 for a number from 1.0 to 511/256.
572  */
573 
574 static int recip_estimate(int input)
575 {
576     int a, b, r;
577     assert(256 <= input && input < 512);
578     a = (input * 2) + 1;
579     b = (1 << 19) / a;
580     r = (b + 1) >> 1;
581     assert(256 <= r && r < 512);
582     return r;
583 }
584 
585 /*
586  * Common wrapper to call recip_estimate
587  *
588  * The parameters are exponent and 64 bit fraction (without implicit
589  * bit) where the binary point is nominally at bit 52. Returns a
590  * float64 which can then be rounded to the appropriate size by the
591  * callee.
592  */
593 
594 static uint64_t call_recip_estimate(int *exp, int exp_off, uint64_t frac)
595 {
596     uint32_t scaled, estimate;
597     uint64_t result_frac;
598     int result_exp;
599 
600     /* Handle sub-normals */
601     if (*exp == 0) {
602         if (extract64(frac, 51, 1) == 0) {
603             *exp = -1;
604             frac <<= 2;
605         } else {
606             frac <<= 1;
607         }
608     }
609 
610     /* scaled = UInt('1':fraction<51:44>) */
611     scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
612     estimate = recip_estimate(scaled);
613 
614     result_exp = exp_off - *exp;
615     result_frac = deposit64(0, 44, 8, estimate);
616     if (result_exp == 0) {
617         result_frac = deposit64(result_frac >> 1, 51, 1, 1);
618     } else if (result_exp == -1) {
619         result_frac = deposit64(result_frac >> 2, 50, 2, 1);
620         result_exp = 0;
621     }
622 
623     *exp = result_exp;
624 
625     return result_frac;
626 }
627 
628 static bool round_to_inf(float_status *fpst, bool sign_bit)
629 {
630     switch (fpst->float_rounding_mode) {
631     case float_round_nearest_even: /* Round to Nearest */
632         return true;
633     case float_round_up: /* Round to +Inf */
634         return !sign_bit;
635     case float_round_down: /* Round to -Inf */
636         return sign_bit;
637     case float_round_to_zero: /* Round to Zero */
638         return false;
639     default:
640         g_assert_not_reached();
641     }
642 }
643 
644 uint32_t HELPER(recpe_f16)(uint32_t input, void *fpstp)
645 {
646     float_status *fpst = fpstp;
647     float16 f16 = float16_squash_input_denormal(input, fpst);
648     uint32_t f16_val = float16_val(f16);
649     uint32_t f16_sign = float16_is_neg(f16);
650     int f16_exp = extract32(f16_val, 10, 5);
651     uint32_t f16_frac = extract32(f16_val, 0, 10);
652     uint64_t f64_frac;
653 
654     if (float16_is_any_nan(f16)) {
655         float16 nan = f16;
656         if (float16_is_signaling_nan(f16, fpst)) {
657             float_raise(float_flag_invalid, fpst);
658             nan = float16_silence_nan(f16, fpst);
659         }
660         if (fpst->default_nan_mode) {
661             nan =  float16_default_nan(fpst);
662         }
663         return nan;
664     } else if (float16_is_infinity(f16)) {
665         return float16_set_sign(float16_zero, float16_is_neg(f16));
666     } else if (float16_is_zero(f16)) {
667         float_raise(float_flag_divbyzero, fpst);
668         return float16_set_sign(float16_infinity, float16_is_neg(f16));
669     } else if (float16_abs(f16) < (1 << 8)) {
670         /* Abs(value) < 2.0^-16 */
671         float_raise(float_flag_overflow | float_flag_inexact, fpst);
672         if (round_to_inf(fpst, f16_sign)) {
673             return float16_set_sign(float16_infinity, f16_sign);
674         } else {
675             return float16_set_sign(float16_maxnorm, f16_sign);
676         }
677     } else if (f16_exp >= 29 && fpst->flush_to_zero) {
678         float_raise(float_flag_underflow, fpst);
679         return float16_set_sign(float16_zero, float16_is_neg(f16));
680     }
681 
682     f64_frac = call_recip_estimate(&f16_exp, 29,
683                                    ((uint64_t) f16_frac) << (52 - 10));
684 
685     /* result = sign : result_exp<4:0> : fraction<51:42> */
686     f16_val = deposit32(0, 15, 1, f16_sign);
687     f16_val = deposit32(f16_val, 10, 5, f16_exp);
688     f16_val = deposit32(f16_val, 0, 10, extract64(f64_frac, 52 - 10, 10));
689     return make_float16(f16_val);
690 }
691 
692 float32 HELPER(recpe_f32)(float32 input, void *fpstp)
693 {
694     float_status *fpst = fpstp;
695     float32 f32 = float32_squash_input_denormal(input, fpst);
696     uint32_t f32_val = float32_val(f32);
697     bool f32_sign = float32_is_neg(f32);
698     int f32_exp = extract32(f32_val, 23, 8);
699     uint32_t f32_frac = extract32(f32_val, 0, 23);
700     uint64_t f64_frac;
701 
702     if (float32_is_any_nan(f32)) {
703         float32 nan = f32;
704         if (float32_is_signaling_nan(f32, fpst)) {
705             float_raise(float_flag_invalid, fpst);
706             nan = float32_silence_nan(f32, fpst);
707         }
708         if (fpst->default_nan_mode) {
709             nan =  float32_default_nan(fpst);
710         }
711         return nan;
712     } else if (float32_is_infinity(f32)) {
713         return float32_set_sign(float32_zero, float32_is_neg(f32));
714     } else if (float32_is_zero(f32)) {
715         float_raise(float_flag_divbyzero, fpst);
716         return float32_set_sign(float32_infinity, float32_is_neg(f32));
717     } else if (float32_abs(f32) < (1ULL << 21)) {
718         /* Abs(value) < 2.0^-128 */
719         float_raise(float_flag_overflow | float_flag_inexact, fpst);
720         if (round_to_inf(fpst, f32_sign)) {
721             return float32_set_sign(float32_infinity, f32_sign);
722         } else {
723             return float32_set_sign(float32_maxnorm, f32_sign);
724         }
725     } else if (f32_exp >= 253 && fpst->flush_to_zero) {
726         float_raise(float_flag_underflow, fpst);
727         return float32_set_sign(float32_zero, float32_is_neg(f32));
728     }
729 
730     f64_frac = call_recip_estimate(&f32_exp, 253,
731                                    ((uint64_t) f32_frac) << (52 - 23));
732 
733     /* result = sign : result_exp<7:0> : fraction<51:29> */
734     f32_val = deposit32(0, 31, 1, f32_sign);
735     f32_val = deposit32(f32_val, 23, 8, f32_exp);
736     f32_val = deposit32(f32_val, 0, 23, extract64(f64_frac, 52 - 23, 23));
737     return make_float32(f32_val);
738 }
739 
740 float64 HELPER(recpe_f64)(float64 input, void *fpstp)
741 {
742     float_status *fpst = fpstp;
743     float64 f64 = float64_squash_input_denormal(input, fpst);
744     uint64_t f64_val = float64_val(f64);
745     bool f64_sign = float64_is_neg(f64);
746     int f64_exp = extract64(f64_val, 52, 11);
747     uint64_t f64_frac = extract64(f64_val, 0, 52);
748 
749     /* Deal with any special cases */
750     if (float64_is_any_nan(f64)) {
751         float64 nan = f64;
752         if (float64_is_signaling_nan(f64, fpst)) {
753             float_raise(float_flag_invalid, fpst);
754             nan = float64_silence_nan(f64, fpst);
755         }
756         if (fpst->default_nan_mode) {
757             nan =  float64_default_nan(fpst);
758         }
759         return nan;
760     } else if (float64_is_infinity(f64)) {
761         return float64_set_sign(float64_zero, float64_is_neg(f64));
762     } else if (float64_is_zero(f64)) {
763         float_raise(float_flag_divbyzero, fpst);
764         return float64_set_sign(float64_infinity, float64_is_neg(f64));
765     } else if ((f64_val & ~(1ULL << 63)) < (1ULL << 50)) {
766         /* Abs(value) < 2.0^-1024 */
767         float_raise(float_flag_overflow | float_flag_inexact, fpst);
768         if (round_to_inf(fpst, f64_sign)) {
769             return float64_set_sign(float64_infinity, f64_sign);
770         } else {
771             return float64_set_sign(float64_maxnorm, f64_sign);
772         }
773     } else if (f64_exp >= 2045 && fpst->flush_to_zero) {
774         float_raise(float_flag_underflow, fpst);
775         return float64_set_sign(float64_zero, float64_is_neg(f64));
776     }
777 
778     f64_frac = call_recip_estimate(&f64_exp, 2045, f64_frac);
779 
780     /* result = sign : result_exp<10:0> : fraction<51:0>; */
781     f64_val = deposit64(0, 63, 1, f64_sign);
782     f64_val = deposit64(f64_val, 52, 11, f64_exp);
783     f64_val = deposit64(f64_val, 0, 52, f64_frac);
784     return make_float64(f64_val);
785 }
786 
787 /* The algorithm that must be used to calculate the estimate
788  * is specified by the ARM ARM.
789  */
790 
791 static int do_recip_sqrt_estimate(int a)
792 {
793     int b, estimate;
794 
795     assert(128 <= a && a < 512);
796     if (a < 256) {
797         a = a * 2 + 1;
798     } else {
799         a = (a >> 1) << 1;
800         a = (a + 1) * 2;
801     }
802     b = 512;
803     while (a * (b + 1) * (b + 1) < (1 << 28)) {
804         b += 1;
805     }
806     estimate = (b + 1) / 2;
807     assert(256 <= estimate && estimate < 512);
808 
809     return estimate;
810 }
811 
812 
813 static uint64_t recip_sqrt_estimate(int *exp , int exp_off, uint64_t frac)
814 {
815     int estimate;
816     uint32_t scaled;
817 
818     if (*exp == 0) {
819         while (extract64(frac, 51, 1) == 0) {
820             frac = frac << 1;
821             *exp -= 1;
822         }
823         frac = extract64(frac, 0, 51) << 1;
824     }
825 
826     if (*exp & 1) {
827         /* scaled = UInt('01':fraction<51:45>) */
828         scaled = deposit32(1 << 7, 0, 7, extract64(frac, 45, 7));
829     } else {
830         /* scaled = UInt('1':fraction<51:44>) */
831         scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
832     }
833     estimate = do_recip_sqrt_estimate(scaled);
834 
835     *exp = (exp_off - *exp) / 2;
836     return extract64(estimate, 0, 8) << 44;
837 }
838 
839 uint32_t HELPER(rsqrte_f16)(uint32_t input, void *fpstp)
840 {
841     float_status *s = fpstp;
842     float16 f16 = float16_squash_input_denormal(input, s);
843     uint16_t val = float16_val(f16);
844     bool f16_sign = float16_is_neg(f16);
845     int f16_exp = extract32(val, 10, 5);
846     uint16_t f16_frac = extract32(val, 0, 10);
847     uint64_t f64_frac;
848 
849     if (float16_is_any_nan(f16)) {
850         float16 nan = f16;
851         if (float16_is_signaling_nan(f16, s)) {
852             float_raise(float_flag_invalid, s);
853             nan = float16_silence_nan(f16, s);
854         }
855         if (s->default_nan_mode) {
856             nan =  float16_default_nan(s);
857         }
858         return nan;
859     } else if (float16_is_zero(f16)) {
860         float_raise(float_flag_divbyzero, s);
861         return float16_set_sign(float16_infinity, f16_sign);
862     } else if (f16_sign) {
863         float_raise(float_flag_invalid, s);
864         return float16_default_nan(s);
865     } else if (float16_is_infinity(f16)) {
866         return float16_zero;
867     }
868 
869     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
870      * preserving the parity of the exponent.  */
871 
872     f64_frac = ((uint64_t) f16_frac) << (52 - 10);
873 
874     f64_frac = recip_sqrt_estimate(&f16_exp, 44, f64_frac);
875 
876     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(2) */
877     val = deposit32(0, 15, 1, f16_sign);
878     val = deposit32(val, 10, 5, f16_exp);
879     val = deposit32(val, 2, 8, extract64(f64_frac, 52 - 8, 8));
880     return make_float16(val);
881 }
882 
883 float32 HELPER(rsqrte_f32)(float32 input, void *fpstp)
884 {
885     float_status *s = fpstp;
886     float32 f32 = float32_squash_input_denormal(input, s);
887     uint32_t val = float32_val(f32);
888     uint32_t f32_sign = float32_is_neg(f32);
889     int f32_exp = extract32(val, 23, 8);
890     uint32_t f32_frac = extract32(val, 0, 23);
891     uint64_t f64_frac;
892 
893     if (float32_is_any_nan(f32)) {
894         float32 nan = f32;
895         if (float32_is_signaling_nan(f32, s)) {
896             float_raise(float_flag_invalid, s);
897             nan = float32_silence_nan(f32, s);
898         }
899         if (s->default_nan_mode) {
900             nan =  float32_default_nan(s);
901         }
902         return nan;
903     } else if (float32_is_zero(f32)) {
904         float_raise(float_flag_divbyzero, s);
905         return float32_set_sign(float32_infinity, float32_is_neg(f32));
906     } else if (float32_is_neg(f32)) {
907         float_raise(float_flag_invalid, s);
908         return float32_default_nan(s);
909     } else if (float32_is_infinity(f32)) {
910         return float32_zero;
911     }
912 
913     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
914      * preserving the parity of the exponent.  */
915 
916     f64_frac = ((uint64_t) f32_frac) << 29;
917 
918     f64_frac = recip_sqrt_estimate(&f32_exp, 380, f64_frac);
919 
920     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(15) */
921     val = deposit32(0, 31, 1, f32_sign);
922     val = deposit32(val, 23, 8, f32_exp);
923     val = deposit32(val, 15, 8, extract64(f64_frac, 52 - 8, 8));
924     return make_float32(val);
925 }
926 
927 float64 HELPER(rsqrte_f64)(float64 input, void *fpstp)
928 {
929     float_status *s = fpstp;
930     float64 f64 = float64_squash_input_denormal(input, s);
931     uint64_t val = float64_val(f64);
932     bool f64_sign = float64_is_neg(f64);
933     int f64_exp = extract64(val, 52, 11);
934     uint64_t f64_frac = extract64(val, 0, 52);
935 
936     if (float64_is_any_nan(f64)) {
937         float64 nan = f64;
938         if (float64_is_signaling_nan(f64, s)) {
939             float_raise(float_flag_invalid, s);
940             nan = float64_silence_nan(f64, s);
941         }
942         if (s->default_nan_mode) {
943             nan =  float64_default_nan(s);
944         }
945         return nan;
946     } else if (float64_is_zero(f64)) {
947         float_raise(float_flag_divbyzero, s);
948         return float64_set_sign(float64_infinity, float64_is_neg(f64));
949     } else if (float64_is_neg(f64)) {
950         float_raise(float_flag_invalid, s);
951         return float64_default_nan(s);
952     } else if (float64_is_infinity(f64)) {
953         return float64_zero;
954     }
955 
956     f64_frac = recip_sqrt_estimate(&f64_exp, 3068, f64_frac);
957 
958     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(44) */
959     val = deposit64(0, 61, 1, f64_sign);
960     val = deposit64(val, 52, 11, f64_exp);
961     val = deposit64(val, 44, 8, extract64(f64_frac, 52 - 8, 8));
962     return make_float64(val);
963 }
964 
965 uint32_t HELPER(recpe_u32)(uint32_t a)
966 {
967     int input, estimate;
968 
969     if ((a & 0x80000000) == 0) {
970         return 0xffffffff;
971     }
972 
973     input = extract32(a, 23, 9);
974     estimate = recip_estimate(input);
975 
976     return deposit32(0, (32 - 9), 9, estimate);
977 }
978 
979 uint32_t HELPER(rsqrte_u32)(uint32_t a)
980 {
981     int estimate;
982 
983     if ((a & 0xc0000000) == 0) {
984         return 0xffffffff;
985     }
986 
987     estimate = do_recip_sqrt_estimate(extract32(a, 23, 9));
988 
989     return deposit32(0, 23, 9, estimate);
990 }
991 
992 /* VFPv4 fused multiply-accumulate */
993 dh_ctype_f16 VFP_HELPER(muladd, h)(dh_ctype_f16 a, dh_ctype_f16 b,
994                                    dh_ctype_f16 c, void *fpstp)
995 {
996     float_status *fpst = fpstp;
997     return float16_muladd(a, b, c, 0, fpst);
998 }
999 
1000 float32 VFP_HELPER(muladd, s)(float32 a, float32 b, float32 c, void *fpstp)
1001 {
1002     float_status *fpst = fpstp;
1003     return float32_muladd(a, b, c, 0, fpst);
1004 }
1005 
1006 float64 VFP_HELPER(muladd, d)(float64 a, float64 b, float64 c, void *fpstp)
1007 {
1008     float_status *fpst = fpstp;
1009     return float64_muladd(a, b, c, 0, fpst);
1010 }
1011 
1012 /* ARMv8 round to integral */
1013 dh_ctype_f16 HELPER(rinth_exact)(dh_ctype_f16 x, void *fp_status)
1014 {
1015     return float16_round_to_int(x, fp_status);
1016 }
1017 
1018 float32 HELPER(rints_exact)(float32 x, void *fp_status)
1019 {
1020     return float32_round_to_int(x, fp_status);
1021 }
1022 
1023 float64 HELPER(rintd_exact)(float64 x, void *fp_status)
1024 {
1025     return float64_round_to_int(x, fp_status);
1026 }
1027 
1028 dh_ctype_f16 HELPER(rinth)(dh_ctype_f16 x, void *fp_status)
1029 {
1030     int old_flags = get_float_exception_flags(fp_status), new_flags;
1031     float16 ret;
1032 
1033     ret = float16_round_to_int(x, fp_status);
1034 
1035     /* Suppress any inexact exceptions the conversion produced */
1036     if (!(old_flags & float_flag_inexact)) {
1037         new_flags = get_float_exception_flags(fp_status);
1038         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1039     }
1040 
1041     return ret;
1042 }
1043 
1044 float32 HELPER(rints)(float32 x, void *fp_status)
1045 {
1046     int old_flags = get_float_exception_flags(fp_status), new_flags;
1047     float32 ret;
1048 
1049     ret = float32_round_to_int(x, fp_status);
1050 
1051     /* Suppress any inexact exceptions the conversion produced */
1052     if (!(old_flags & float_flag_inexact)) {
1053         new_flags = get_float_exception_flags(fp_status);
1054         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1055     }
1056 
1057     return ret;
1058 }
1059 
1060 float64 HELPER(rintd)(float64 x, void *fp_status)
1061 {
1062     int old_flags = get_float_exception_flags(fp_status), new_flags;
1063     float64 ret;
1064 
1065     ret = float64_round_to_int(x, fp_status);
1066 
1067     new_flags = get_float_exception_flags(fp_status);
1068 
1069     /* Suppress any inexact exceptions the conversion produced */
1070     if (!(old_flags & float_flag_inexact)) {
1071         new_flags = get_float_exception_flags(fp_status);
1072         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1073     }
1074 
1075     return ret;
1076 }
1077 
1078 /* Convert ARM rounding mode to softfloat */
1079 int arm_rmode_to_sf(int rmode)
1080 {
1081     switch (rmode) {
1082     case FPROUNDING_TIEAWAY:
1083         rmode = float_round_ties_away;
1084         break;
1085     case FPROUNDING_ODD:
1086         /* FIXME: add support for TIEAWAY and ODD */
1087         qemu_log_mask(LOG_UNIMP, "arm: unimplemented rounding mode: %d\n",
1088                       rmode);
1089         /* fall through for now */
1090     case FPROUNDING_TIEEVEN:
1091     default:
1092         rmode = float_round_nearest_even;
1093         break;
1094     case FPROUNDING_POSINF:
1095         rmode = float_round_up;
1096         break;
1097     case FPROUNDING_NEGINF:
1098         rmode = float_round_down;
1099         break;
1100     case FPROUNDING_ZERO:
1101         rmode = float_round_to_zero;
1102         break;
1103     }
1104     return rmode;
1105 }
1106 
1107 /*
1108  * Implement float64 to int32_t conversion without saturation;
1109  * the result is supplied modulo 2^32.
1110  */
1111 uint64_t HELPER(fjcvtzs)(float64 value, void *vstatus)
1112 {
1113     float_status *status = vstatus;
1114     uint32_t exp, sign;
1115     uint64_t frac;
1116     uint32_t inexact = 1; /* !Z */
1117 
1118     sign = extract64(value, 63, 1);
1119     exp = extract64(value, 52, 11);
1120     frac = extract64(value, 0, 52);
1121 
1122     if (exp == 0) {
1123         /* While not inexact for IEEE FP, -0.0 is inexact for JavaScript.  */
1124         inexact = sign;
1125         if (frac != 0) {
1126             if (status->flush_inputs_to_zero) {
1127                 float_raise(float_flag_input_denormal, status);
1128             } else {
1129                 float_raise(float_flag_inexact, status);
1130                 inexact = 1;
1131             }
1132         }
1133         frac = 0;
1134     } else if (exp == 0x7ff) {
1135         /* This operation raises Invalid for both NaN and overflow (Inf).  */
1136         float_raise(float_flag_invalid, status);
1137         frac = 0;
1138     } else {
1139         int true_exp = exp - 1023;
1140         int shift = true_exp - 52;
1141 
1142         /* Restore implicit bit.  */
1143         frac |= 1ull << 52;
1144 
1145         /* Shift the fraction into place.  */
1146         if (shift >= 0) {
1147             /* The number is so large we must shift the fraction left.  */
1148             if (shift >= 64) {
1149                 /* The fraction is shifted out entirely.  */
1150                 frac = 0;
1151             } else {
1152                 frac <<= shift;
1153             }
1154         } else if (shift > -64) {
1155             /* Normal case -- shift right and notice if bits shift out.  */
1156             inexact = (frac << (64 + shift)) != 0;
1157             frac >>= -shift;
1158         } else {
1159             /* The fraction is shifted out entirely.  */
1160             frac = 0;
1161         }
1162 
1163         /* Notice overflow or inexact exceptions.  */
1164         if (true_exp > 31 || frac > (sign ? 0x80000000ull : 0x7fffffff)) {
1165             /* Overflow, for which this operation raises invalid.  */
1166             float_raise(float_flag_invalid, status);
1167             inexact = 1;
1168         } else if (inexact) {
1169             float_raise(float_flag_inexact, status);
1170         }
1171 
1172         /* Honor the sign.  */
1173         if (sign) {
1174             frac = -frac;
1175         }
1176     }
1177 
1178     /* Pack the result and the env->ZF representation of Z together.  */
1179     return deposit64(frac, 32, 32, inexact);
1180 }
1181 
1182 uint32_t HELPER(vjcvt)(float64 value, CPUARMState *env)
1183 {
1184     uint64_t pair = HELPER(fjcvtzs)(value, &env->vfp.fp_status);
1185     uint32_t result = pair;
1186     uint32_t z = (pair >> 32) == 0;
1187 
1188     /* Store Z, clear NCV, in FPSCR.NZCV.  */
1189     env->vfp.xregs[ARM_VFP_FPSCR]
1190         = (env->vfp.xregs[ARM_VFP_FPSCR] & ~CPSR_NZCV) | (z * CPSR_Z);
1191 
1192     return result;
1193 }
1194 
1195 /* Round a float32 to an integer that fits in int32_t or int64_t.  */
1196 static float32 frint_s(float32 f, float_status *fpst, int intsize)
1197 {
1198     int old_flags = get_float_exception_flags(fpst);
1199     uint32_t exp = extract32(f, 23, 8);
1200 
1201     if (unlikely(exp == 0xff)) {
1202         /* NaN or Inf.  */
1203         goto overflow;
1204     }
1205 
1206     /* Round and re-extract the exponent.  */
1207     f = float32_round_to_int(f, fpst);
1208     exp = extract32(f, 23, 8);
1209 
1210     /* Validate the range of the result.  */
1211     if (exp < 126 + intsize) {
1212         /* abs(F) <= INT{N}_MAX */
1213         return f;
1214     }
1215     if (exp == 126 + intsize) {
1216         uint32_t sign = extract32(f, 31, 1);
1217         uint32_t frac = extract32(f, 0, 23);
1218         if (sign && frac == 0) {
1219             /* F == INT{N}_MIN */
1220             return f;
1221         }
1222     }
1223 
1224  overflow:
1225     /*
1226      * Raise Invalid and return INT{N}_MIN as a float.  Revert any
1227      * inexact exception float32_round_to_int may have raised.
1228      */
1229     set_float_exception_flags(old_flags | float_flag_invalid, fpst);
1230     return (0x100u + 126u + intsize) << 23;
1231 }
1232 
1233 float32 HELPER(frint32_s)(float32 f, void *fpst)
1234 {
1235     return frint_s(f, fpst, 32);
1236 }
1237 
1238 float32 HELPER(frint64_s)(float32 f, void *fpst)
1239 {
1240     return frint_s(f, fpst, 64);
1241 }
1242 
1243 /* Round a float64 to an integer that fits in int32_t or int64_t.  */
1244 static float64 frint_d(float64 f, float_status *fpst, int intsize)
1245 {
1246     int old_flags = get_float_exception_flags(fpst);
1247     uint32_t exp = extract64(f, 52, 11);
1248 
1249     if (unlikely(exp == 0x7ff)) {
1250         /* NaN or Inf.  */
1251         goto overflow;
1252     }
1253 
1254     /* Round and re-extract the exponent.  */
1255     f = float64_round_to_int(f, fpst);
1256     exp = extract64(f, 52, 11);
1257 
1258     /* Validate the range of the result.  */
1259     if (exp < 1022 + intsize) {
1260         /* abs(F) <= INT{N}_MAX */
1261         return f;
1262     }
1263     if (exp == 1022 + intsize) {
1264         uint64_t sign = extract64(f, 63, 1);
1265         uint64_t frac = extract64(f, 0, 52);
1266         if (sign && frac == 0) {
1267             /* F == INT{N}_MIN */
1268             return f;
1269         }
1270     }
1271 
1272  overflow:
1273     /*
1274      * Raise Invalid and return INT{N}_MIN as a float.  Revert any
1275      * inexact exception float64_round_to_int may have raised.
1276      */
1277     set_float_exception_flags(old_flags | float_flag_invalid, fpst);
1278     return (uint64_t)(0x800 + 1022 + intsize) << 52;
1279 }
1280 
1281 float64 HELPER(frint32_d)(float64 f, void *fpst)
1282 {
1283     return frint_d(f, fpst, 32);
1284 }
1285 
1286 float64 HELPER(frint64_d)(float64 f, void *fpst)
1287 {
1288     return frint_d(f, fpst, 64);
1289 }
1290 
1291 void HELPER(check_hcr_el2_trap)(CPUARMState *env, uint32_t rt, uint32_t reg)
1292 {
1293     uint32_t syndrome;
1294 
1295     switch (reg) {
1296     case ARM_VFP_MVFR0:
1297     case ARM_VFP_MVFR1:
1298     case ARM_VFP_MVFR2:
1299         if (!(arm_hcr_el2_eff(env) & HCR_TID3)) {
1300             return;
1301         }
1302         break;
1303     case ARM_VFP_FPSID:
1304         if (!(arm_hcr_el2_eff(env) & HCR_TID0)) {
1305             return;
1306         }
1307         break;
1308     default:
1309         g_assert_not_reached();
1310     }
1311 
1312     syndrome = ((EC_FPIDTRAP << ARM_EL_EC_SHIFT)
1313                 | ARM_EL_IL
1314                 | (1 << 24) | (0xe << 20) | (7 << 14)
1315                 | (reg << 10) | (rt << 5) | 1);
1316 
1317     raise_exception(env, EXCP_HYP_TRAP, syndrome, 2);
1318 }
1319 
1320 #endif
1321