xref: /openbmc/qemu/target/arm/vfp_helper.c (revision b15e402f)
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     fpscr |= vfp_get_fpscr_from_host(env);
178 
179     i = env->vfp.qc[0] | env->vfp.qc[1] | env->vfp.qc[2] | env->vfp.qc[3];
180     fpscr |= i ? FPCR_QC : 0;
181 
182     return fpscr;
183 }
184 
185 uint32_t vfp_get_fpscr(CPUARMState *env)
186 {
187     return HELPER(vfp_get_fpscr)(env);
188 }
189 
190 void HELPER(vfp_set_fpscr)(CPUARMState *env, uint32_t val)
191 {
192     /* When ARMv8.2-FP16 is not supported, FZ16 is RES0.  */
193     if (!cpu_isar_feature(any_fp16, env_archcpu(env))) {
194         val &= ~FPCR_FZ16;
195     }
196 
197     if (arm_feature(env, ARM_FEATURE_M)) {
198         /*
199          * M profile FPSCR is RES0 for the QC, STRIDE, FZ16, LEN bits
200          * and also for the trapped-exception-handling bits IxE.
201          */
202         val &= 0xf7c0009f;
203     }
204 
205     vfp_set_fpscr_to_host(env, val);
206 
207     /*
208      * We don't implement trapped exception handling, so the
209      * trap enable bits, IDE|IXE|UFE|OFE|DZE|IOE are all RAZ/WI (not RES0!)
210      *
211      * If we exclude the exception flags, IOC|DZC|OFC|UFC|IXC|IDC
212      * (which are stored in fp_status), and the other RES0 bits
213      * in between, then we clear all of the low 16 bits.
214      */
215     env->vfp.xregs[ARM_VFP_FPSCR] = val & 0xf7c80000;
216     env->vfp.vec_len = (val >> 16) & 7;
217     env->vfp.vec_stride = (val >> 20) & 3;
218 
219     /*
220      * The bit we set within fpscr_q is arbitrary; the register as a
221      * whole being zero/non-zero is what counts.
222      */
223     env->vfp.qc[0] = val & FPCR_QC;
224     env->vfp.qc[1] = 0;
225     env->vfp.qc[2] = 0;
226     env->vfp.qc[3] = 0;
227 }
228 
229 void vfp_set_fpscr(CPUARMState *env, uint32_t val)
230 {
231     HELPER(vfp_set_fpscr)(env, val);
232 }
233 
234 #ifdef CONFIG_TCG
235 
236 #define VFP_HELPER(name, p) HELPER(glue(glue(vfp_,name),p))
237 
238 #define VFP_BINOP(name) \
239 dh_ctype_f16 VFP_HELPER(name, h)(dh_ctype_f16 a, dh_ctype_f16 b, void *fpstp) \
240 { \
241     float_status *fpst = fpstp; \
242     return float16_ ## name(a, b, fpst); \
243 } \
244 float32 VFP_HELPER(name, s)(float32 a, float32 b, void *fpstp) \
245 { \
246     float_status *fpst = fpstp; \
247     return float32_ ## name(a, b, fpst); \
248 } \
249 float64 VFP_HELPER(name, d)(float64 a, float64 b, void *fpstp) \
250 { \
251     float_status *fpst = fpstp; \
252     return float64_ ## name(a, b, fpst); \
253 }
254 VFP_BINOP(add)
255 VFP_BINOP(sub)
256 VFP_BINOP(mul)
257 VFP_BINOP(div)
258 VFP_BINOP(min)
259 VFP_BINOP(max)
260 VFP_BINOP(minnum)
261 VFP_BINOP(maxnum)
262 #undef VFP_BINOP
263 
264 dh_ctype_f16 VFP_HELPER(neg, h)(dh_ctype_f16 a)
265 {
266     return float16_chs(a);
267 }
268 
269 float32 VFP_HELPER(neg, s)(float32 a)
270 {
271     return float32_chs(a);
272 }
273 
274 float64 VFP_HELPER(neg, d)(float64 a)
275 {
276     return float64_chs(a);
277 }
278 
279 dh_ctype_f16 VFP_HELPER(abs, h)(dh_ctype_f16 a)
280 {
281     return float16_abs(a);
282 }
283 
284 float32 VFP_HELPER(abs, s)(float32 a)
285 {
286     return float32_abs(a);
287 }
288 
289 float64 VFP_HELPER(abs, d)(float64 a)
290 {
291     return float64_abs(a);
292 }
293 
294 dh_ctype_f16 VFP_HELPER(sqrt, h)(dh_ctype_f16 a, CPUARMState *env)
295 {
296     return float16_sqrt(a, &env->vfp.fp_status_f16);
297 }
298 
299 float32 VFP_HELPER(sqrt, s)(float32 a, CPUARMState *env)
300 {
301     return float32_sqrt(a, &env->vfp.fp_status);
302 }
303 
304 float64 VFP_HELPER(sqrt, d)(float64 a, CPUARMState *env)
305 {
306     return float64_sqrt(a, &env->vfp.fp_status);
307 }
308 
309 static void softfloat_to_vfp_compare(CPUARMState *env, FloatRelation cmp)
310 {
311     uint32_t flags;
312     switch (cmp) {
313     case float_relation_equal:
314         flags = 0x6;
315         break;
316     case float_relation_less:
317         flags = 0x8;
318         break;
319     case float_relation_greater:
320         flags = 0x2;
321         break;
322     case float_relation_unordered:
323         flags = 0x3;
324         break;
325     default:
326         g_assert_not_reached();
327     }
328     env->vfp.xregs[ARM_VFP_FPSCR] =
329         deposit32(env->vfp.xregs[ARM_VFP_FPSCR], 28, 4, flags);
330 }
331 
332 /* XXX: check quiet/signaling case */
333 #define DO_VFP_cmp(P, FLOATTYPE, ARGTYPE, FPST) \
334 void VFP_HELPER(cmp, P)(ARGTYPE a, ARGTYPE b, CPUARMState *env)  \
335 { \
336     softfloat_to_vfp_compare(env, \
337         FLOATTYPE ## _compare_quiet(a, b, &env->vfp.FPST)); \
338 } \
339 void VFP_HELPER(cmpe, P)(ARGTYPE a, ARGTYPE b, CPUARMState *env) \
340 { \
341     softfloat_to_vfp_compare(env, \
342         FLOATTYPE ## _compare(a, b, &env->vfp.FPST)); \
343 }
344 DO_VFP_cmp(h, float16, dh_ctype_f16, fp_status_f16)
345 DO_VFP_cmp(s, float32, float32, fp_status)
346 DO_VFP_cmp(d, float64, float64, fp_status)
347 #undef DO_VFP_cmp
348 
349 /* Integer to float and float to integer conversions */
350 
351 #define CONV_ITOF(name, ftype, fsz, sign)                           \
352 ftype HELPER(name)(uint32_t x, void *fpstp)                         \
353 {                                                                   \
354     float_status *fpst = fpstp;                                     \
355     return sign##int32_to_##float##fsz((sign##int32_t)x, fpst);     \
356 }
357 
358 #define CONV_FTOI(name, ftype, fsz, sign, round)                \
359 sign##int32_t HELPER(name)(ftype x, void *fpstp)                \
360 {                                                               \
361     float_status *fpst = fpstp;                                 \
362     if (float##fsz##_is_any_nan(x)) {                           \
363         float_raise(float_flag_invalid, fpst);                  \
364         return 0;                                               \
365     }                                                           \
366     return float##fsz##_to_##sign##int32##round(x, fpst);       \
367 }
368 
369 #define FLOAT_CONVS(name, p, ftype, fsz, sign)            \
370     CONV_ITOF(vfp_##name##to##p, ftype, fsz, sign)        \
371     CONV_FTOI(vfp_to##name##p, ftype, fsz, sign, )        \
372     CONV_FTOI(vfp_to##name##z##p, ftype, fsz, sign, _round_to_zero)
373 
374 FLOAT_CONVS(si, h, uint32_t, 16, )
375 FLOAT_CONVS(si, s, float32, 32, )
376 FLOAT_CONVS(si, d, float64, 64, )
377 FLOAT_CONVS(ui, h, uint32_t, 16, u)
378 FLOAT_CONVS(ui, s, float32, 32, u)
379 FLOAT_CONVS(ui, d, float64, 64, u)
380 
381 #undef CONV_ITOF
382 #undef CONV_FTOI
383 #undef FLOAT_CONVS
384 
385 /* floating point conversion */
386 float64 VFP_HELPER(fcvtd, s)(float32 x, CPUARMState *env)
387 {
388     return float32_to_float64(x, &env->vfp.fp_status);
389 }
390 
391 float32 VFP_HELPER(fcvts, d)(float64 x, CPUARMState *env)
392 {
393     return float64_to_float32(x, &env->vfp.fp_status);
394 }
395 
396 /* VFP3 fixed point conversion.  */
397 #define VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)            \
398 ftype HELPER(vfp_##name##to##p)(uint##isz##_t  x, uint32_t shift,      \
399                                      void *fpstp) \
400 { return itype##_to_##float##fsz##_scalbn(x, -shift, fpstp); }
401 
402 #define VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype, ROUND, suff) \
403 uint##isz##_t HELPER(vfp_to##name##p##suff)(ftype x, uint32_t shift,      \
404                                             void *fpst)                   \
405 {                                                                         \
406     if (unlikely(float##fsz##_is_any_nan(x))) {                           \
407         float_raise(float_flag_invalid, fpst);                            \
408         return 0;                                                         \
409     }                                                                     \
410     return float##fsz##_to_##itype##_scalbn(x, ROUND, shift, fpst);       \
411 }
412 
413 #define VFP_CONV_FIX(name, p, fsz, ftype, isz, itype)            \
414 VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)              \
415 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
416                          float_round_to_zero, _round_to_zero)    \
417 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
418                          get_float_rounding_mode(fpst), )
419 
420 #define VFP_CONV_FIX_A64(name, p, fsz, ftype, isz, itype)        \
421 VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)              \
422 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
423                          get_float_rounding_mode(fpst), )
424 
425 VFP_CONV_FIX(sh, d, 64, float64, 64, int16)
426 VFP_CONV_FIX(sl, d, 64, float64, 64, int32)
427 VFP_CONV_FIX_A64(sq, d, 64, float64, 64, int64)
428 VFP_CONV_FIX(uh, d, 64, float64, 64, uint16)
429 VFP_CONV_FIX(ul, d, 64, float64, 64, uint32)
430 VFP_CONV_FIX_A64(uq, d, 64, float64, 64, uint64)
431 VFP_CONV_FIX(sh, s, 32, float32, 32, int16)
432 VFP_CONV_FIX(sl, s, 32, float32, 32, int32)
433 VFP_CONV_FIX_A64(sq, s, 32, float32, 64, int64)
434 VFP_CONV_FIX(uh, s, 32, float32, 32, uint16)
435 VFP_CONV_FIX(ul, s, 32, float32, 32, uint32)
436 VFP_CONV_FIX_A64(uq, s, 32, float32, 64, uint64)
437 VFP_CONV_FIX(sh, h, 16, dh_ctype_f16, 32, int16)
438 VFP_CONV_FIX(sl, h, 16, dh_ctype_f16, 32, int32)
439 VFP_CONV_FIX_A64(sq, h, 16, dh_ctype_f16, 64, int64)
440 VFP_CONV_FIX(uh, h, 16, dh_ctype_f16, 32, uint16)
441 VFP_CONV_FIX(ul, h, 16, dh_ctype_f16, 32, uint32)
442 VFP_CONV_FIX_A64(uq, h, 16, dh_ctype_f16, 64, uint64)
443 
444 #undef VFP_CONV_FIX
445 #undef VFP_CONV_FIX_FLOAT
446 #undef VFP_CONV_FLOAT_FIX_ROUND
447 #undef VFP_CONV_FIX_A64
448 
449 /* Set the current fp rounding mode and return the old one.
450  * The argument is a softfloat float_round_ value.
451  */
452 uint32_t HELPER(set_rmode)(uint32_t rmode, void *fpstp)
453 {
454     float_status *fp_status = fpstp;
455 
456     uint32_t prev_rmode = get_float_rounding_mode(fp_status);
457     set_float_rounding_mode(rmode, fp_status);
458 
459     return prev_rmode;
460 }
461 
462 /* Half precision conversions.  */
463 float32 HELPER(vfp_fcvt_f16_to_f32)(uint32_t a, void *fpstp, uint32_t ahp_mode)
464 {
465     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
466      * it would affect flushing input denormals.
467      */
468     float_status *fpst = fpstp;
469     bool save = get_flush_inputs_to_zero(fpst);
470     set_flush_inputs_to_zero(false, fpst);
471     float32 r = float16_to_float32(a, !ahp_mode, fpst);
472     set_flush_inputs_to_zero(save, fpst);
473     return r;
474 }
475 
476 uint32_t HELPER(vfp_fcvt_f32_to_f16)(float32 a, void *fpstp, uint32_t ahp_mode)
477 {
478     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
479      * it would affect flushing output denormals.
480      */
481     float_status *fpst = fpstp;
482     bool save = get_flush_to_zero(fpst);
483     set_flush_to_zero(false, fpst);
484     float16 r = float32_to_float16(a, !ahp_mode, fpst);
485     set_flush_to_zero(save, fpst);
486     return r;
487 }
488 
489 float64 HELPER(vfp_fcvt_f16_to_f64)(uint32_t a, void *fpstp, uint32_t ahp_mode)
490 {
491     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
492      * it would affect flushing input denormals.
493      */
494     float_status *fpst = fpstp;
495     bool save = get_flush_inputs_to_zero(fpst);
496     set_flush_inputs_to_zero(false, fpst);
497     float64 r = float16_to_float64(a, !ahp_mode, fpst);
498     set_flush_inputs_to_zero(save, fpst);
499     return r;
500 }
501 
502 uint32_t HELPER(vfp_fcvt_f64_to_f16)(float64 a, void *fpstp, uint32_t ahp_mode)
503 {
504     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
505      * it would affect flushing output denormals.
506      */
507     float_status *fpst = fpstp;
508     bool save = get_flush_to_zero(fpst);
509     set_flush_to_zero(false, fpst);
510     float16 r = float64_to_float16(a, !ahp_mode, fpst);
511     set_flush_to_zero(save, fpst);
512     return r;
513 }
514 
515 /* NEON helpers.  */
516 
517 /* Constants 256 and 512 are used in some helpers; we avoid relying on
518  * int->float conversions at run-time.  */
519 #define float64_256 make_float64(0x4070000000000000LL)
520 #define float64_512 make_float64(0x4080000000000000LL)
521 #define float16_maxnorm make_float16(0x7bff)
522 #define float32_maxnorm make_float32(0x7f7fffff)
523 #define float64_maxnorm make_float64(0x7fefffffffffffffLL)
524 
525 /* Reciprocal functions
526  *
527  * The algorithm that must be used to calculate the estimate
528  * is specified by the ARM ARM, see FPRecipEstimate()/RecipEstimate
529  */
530 
531 /* See RecipEstimate()
532  *
533  * input is a 9 bit fixed point number
534  * input range 256 .. 511 for a number from 0.5 <= x < 1.0.
535  * result range 256 .. 511 for a number from 1.0 to 511/256.
536  */
537 
538 static int recip_estimate(int input)
539 {
540     int a, b, r;
541     assert(256 <= input && input < 512);
542     a = (input * 2) + 1;
543     b = (1 << 19) / a;
544     r = (b + 1) >> 1;
545     assert(256 <= r && r < 512);
546     return r;
547 }
548 
549 /*
550  * Common wrapper to call recip_estimate
551  *
552  * The parameters are exponent and 64 bit fraction (without implicit
553  * bit) where the binary point is nominally at bit 52. Returns a
554  * float64 which can then be rounded to the appropriate size by the
555  * callee.
556  */
557 
558 static uint64_t call_recip_estimate(int *exp, int exp_off, uint64_t frac)
559 {
560     uint32_t scaled, estimate;
561     uint64_t result_frac;
562     int result_exp;
563 
564     /* Handle sub-normals */
565     if (*exp == 0) {
566         if (extract64(frac, 51, 1) == 0) {
567             *exp = -1;
568             frac <<= 2;
569         } else {
570             frac <<= 1;
571         }
572     }
573 
574     /* scaled = UInt('1':fraction<51:44>) */
575     scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
576     estimate = recip_estimate(scaled);
577 
578     result_exp = exp_off - *exp;
579     result_frac = deposit64(0, 44, 8, estimate);
580     if (result_exp == 0) {
581         result_frac = deposit64(result_frac >> 1, 51, 1, 1);
582     } else if (result_exp == -1) {
583         result_frac = deposit64(result_frac >> 2, 50, 2, 1);
584         result_exp = 0;
585     }
586 
587     *exp = result_exp;
588 
589     return result_frac;
590 }
591 
592 static bool round_to_inf(float_status *fpst, bool sign_bit)
593 {
594     switch (fpst->float_rounding_mode) {
595     case float_round_nearest_even: /* Round to Nearest */
596         return true;
597     case float_round_up: /* Round to +Inf */
598         return !sign_bit;
599     case float_round_down: /* Round to -Inf */
600         return sign_bit;
601     case float_round_to_zero: /* Round to Zero */
602         return false;
603     default:
604         g_assert_not_reached();
605     }
606 }
607 
608 uint32_t HELPER(recpe_f16)(uint32_t input, void *fpstp)
609 {
610     float_status *fpst = fpstp;
611     float16 f16 = float16_squash_input_denormal(input, fpst);
612     uint32_t f16_val = float16_val(f16);
613     uint32_t f16_sign = float16_is_neg(f16);
614     int f16_exp = extract32(f16_val, 10, 5);
615     uint32_t f16_frac = extract32(f16_val, 0, 10);
616     uint64_t f64_frac;
617 
618     if (float16_is_any_nan(f16)) {
619         float16 nan = f16;
620         if (float16_is_signaling_nan(f16, fpst)) {
621             float_raise(float_flag_invalid, fpst);
622             nan = float16_silence_nan(f16, fpst);
623         }
624         if (fpst->default_nan_mode) {
625             nan =  float16_default_nan(fpst);
626         }
627         return nan;
628     } else if (float16_is_infinity(f16)) {
629         return float16_set_sign(float16_zero, float16_is_neg(f16));
630     } else if (float16_is_zero(f16)) {
631         float_raise(float_flag_divbyzero, fpst);
632         return float16_set_sign(float16_infinity, float16_is_neg(f16));
633     } else if (float16_abs(f16) < (1 << 8)) {
634         /* Abs(value) < 2.0^-16 */
635         float_raise(float_flag_overflow | float_flag_inexact, fpst);
636         if (round_to_inf(fpst, f16_sign)) {
637             return float16_set_sign(float16_infinity, f16_sign);
638         } else {
639             return float16_set_sign(float16_maxnorm, f16_sign);
640         }
641     } else if (f16_exp >= 29 && fpst->flush_to_zero) {
642         float_raise(float_flag_underflow, fpst);
643         return float16_set_sign(float16_zero, float16_is_neg(f16));
644     }
645 
646     f64_frac = call_recip_estimate(&f16_exp, 29,
647                                    ((uint64_t) f16_frac) << (52 - 10));
648 
649     /* result = sign : result_exp<4:0> : fraction<51:42> */
650     f16_val = deposit32(0, 15, 1, f16_sign);
651     f16_val = deposit32(f16_val, 10, 5, f16_exp);
652     f16_val = deposit32(f16_val, 0, 10, extract64(f64_frac, 52 - 10, 10));
653     return make_float16(f16_val);
654 }
655 
656 float32 HELPER(recpe_f32)(float32 input, void *fpstp)
657 {
658     float_status *fpst = fpstp;
659     float32 f32 = float32_squash_input_denormal(input, fpst);
660     uint32_t f32_val = float32_val(f32);
661     bool f32_sign = float32_is_neg(f32);
662     int f32_exp = extract32(f32_val, 23, 8);
663     uint32_t f32_frac = extract32(f32_val, 0, 23);
664     uint64_t f64_frac;
665 
666     if (float32_is_any_nan(f32)) {
667         float32 nan = f32;
668         if (float32_is_signaling_nan(f32, fpst)) {
669             float_raise(float_flag_invalid, fpst);
670             nan = float32_silence_nan(f32, fpst);
671         }
672         if (fpst->default_nan_mode) {
673             nan =  float32_default_nan(fpst);
674         }
675         return nan;
676     } else if (float32_is_infinity(f32)) {
677         return float32_set_sign(float32_zero, float32_is_neg(f32));
678     } else if (float32_is_zero(f32)) {
679         float_raise(float_flag_divbyzero, fpst);
680         return float32_set_sign(float32_infinity, float32_is_neg(f32));
681     } else if (float32_abs(f32) < (1ULL << 21)) {
682         /* Abs(value) < 2.0^-128 */
683         float_raise(float_flag_overflow | float_flag_inexact, fpst);
684         if (round_to_inf(fpst, f32_sign)) {
685             return float32_set_sign(float32_infinity, f32_sign);
686         } else {
687             return float32_set_sign(float32_maxnorm, f32_sign);
688         }
689     } else if (f32_exp >= 253 && fpst->flush_to_zero) {
690         float_raise(float_flag_underflow, fpst);
691         return float32_set_sign(float32_zero, float32_is_neg(f32));
692     }
693 
694     f64_frac = call_recip_estimate(&f32_exp, 253,
695                                    ((uint64_t) f32_frac) << (52 - 23));
696 
697     /* result = sign : result_exp<7:0> : fraction<51:29> */
698     f32_val = deposit32(0, 31, 1, f32_sign);
699     f32_val = deposit32(f32_val, 23, 8, f32_exp);
700     f32_val = deposit32(f32_val, 0, 23, extract64(f64_frac, 52 - 23, 23));
701     return make_float32(f32_val);
702 }
703 
704 float64 HELPER(recpe_f64)(float64 input, void *fpstp)
705 {
706     float_status *fpst = fpstp;
707     float64 f64 = float64_squash_input_denormal(input, fpst);
708     uint64_t f64_val = float64_val(f64);
709     bool f64_sign = float64_is_neg(f64);
710     int f64_exp = extract64(f64_val, 52, 11);
711     uint64_t f64_frac = extract64(f64_val, 0, 52);
712 
713     /* Deal with any special cases */
714     if (float64_is_any_nan(f64)) {
715         float64 nan = f64;
716         if (float64_is_signaling_nan(f64, fpst)) {
717             float_raise(float_flag_invalid, fpst);
718             nan = float64_silence_nan(f64, fpst);
719         }
720         if (fpst->default_nan_mode) {
721             nan =  float64_default_nan(fpst);
722         }
723         return nan;
724     } else if (float64_is_infinity(f64)) {
725         return float64_set_sign(float64_zero, float64_is_neg(f64));
726     } else if (float64_is_zero(f64)) {
727         float_raise(float_flag_divbyzero, fpst);
728         return float64_set_sign(float64_infinity, float64_is_neg(f64));
729     } else if ((f64_val & ~(1ULL << 63)) < (1ULL << 50)) {
730         /* Abs(value) < 2.0^-1024 */
731         float_raise(float_flag_overflow | float_flag_inexact, fpst);
732         if (round_to_inf(fpst, f64_sign)) {
733             return float64_set_sign(float64_infinity, f64_sign);
734         } else {
735             return float64_set_sign(float64_maxnorm, f64_sign);
736         }
737     } else if (f64_exp >= 2045 && fpst->flush_to_zero) {
738         float_raise(float_flag_underflow, fpst);
739         return float64_set_sign(float64_zero, float64_is_neg(f64));
740     }
741 
742     f64_frac = call_recip_estimate(&f64_exp, 2045, f64_frac);
743 
744     /* result = sign : result_exp<10:0> : fraction<51:0>; */
745     f64_val = deposit64(0, 63, 1, f64_sign);
746     f64_val = deposit64(f64_val, 52, 11, f64_exp);
747     f64_val = deposit64(f64_val, 0, 52, f64_frac);
748     return make_float64(f64_val);
749 }
750 
751 /* The algorithm that must be used to calculate the estimate
752  * is specified by the ARM ARM.
753  */
754 
755 static int do_recip_sqrt_estimate(int a)
756 {
757     int b, estimate;
758 
759     assert(128 <= a && a < 512);
760     if (a < 256) {
761         a = a * 2 + 1;
762     } else {
763         a = (a >> 1) << 1;
764         a = (a + 1) * 2;
765     }
766     b = 512;
767     while (a * (b + 1) * (b + 1) < (1 << 28)) {
768         b += 1;
769     }
770     estimate = (b + 1) / 2;
771     assert(256 <= estimate && estimate < 512);
772 
773     return estimate;
774 }
775 
776 
777 static uint64_t recip_sqrt_estimate(int *exp , int exp_off, uint64_t frac)
778 {
779     int estimate;
780     uint32_t scaled;
781 
782     if (*exp == 0) {
783         while (extract64(frac, 51, 1) == 0) {
784             frac = frac << 1;
785             *exp -= 1;
786         }
787         frac = extract64(frac, 0, 51) << 1;
788     }
789 
790     if (*exp & 1) {
791         /* scaled = UInt('01':fraction<51:45>) */
792         scaled = deposit32(1 << 7, 0, 7, extract64(frac, 45, 7));
793     } else {
794         /* scaled = UInt('1':fraction<51:44>) */
795         scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
796     }
797     estimate = do_recip_sqrt_estimate(scaled);
798 
799     *exp = (exp_off - *exp) / 2;
800     return extract64(estimate, 0, 8) << 44;
801 }
802 
803 uint32_t HELPER(rsqrte_f16)(uint32_t input, void *fpstp)
804 {
805     float_status *s = fpstp;
806     float16 f16 = float16_squash_input_denormal(input, s);
807     uint16_t val = float16_val(f16);
808     bool f16_sign = float16_is_neg(f16);
809     int f16_exp = extract32(val, 10, 5);
810     uint16_t f16_frac = extract32(val, 0, 10);
811     uint64_t f64_frac;
812 
813     if (float16_is_any_nan(f16)) {
814         float16 nan = f16;
815         if (float16_is_signaling_nan(f16, s)) {
816             float_raise(float_flag_invalid, s);
817             nan = float16_silence_nan(f16, s);
818         }
819         if (s->default_nan_mode) {
820             nan =  float16_default_nan(s);
821         }
822         return nan;
823     } else if (float16_is_zero(f16)) {
824         float_raise(float_flag_divbyzero, s);
825         return float16_set_sign(float16_infinity, f16_sign);
826     } else if (f16_sign) {
827         float_raise(float_flag_invalid, s);
828         return float16_default_nan(s);
829     } else if (float16_is_infinity(f16)) {
830         return float16_zero;
831     }
832 
833     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
834      * preserving the parity of the exponent.  */
835 
836     f64_frac = ((uint64_t) f16_frac) << (52 - 10);
837 
838     f64_frac = recip_sqrt_estimate(&f16_exp, 44, f64_frac);
839 
840     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(2) */
841     val = deposit32(0, 15, 1, f16_sign);
842     val = deposit32(val, 10, 5, f16_exp);
843     val = deposit32(val, 2, 8, extract64(f64_frac, 52 - 8, 8));
844     return make_float16(val);
845 }
846 
847 float32 HELPER(rsqrte_f32)(float32 input, void *fpstp)
848 {
849     float_status *s = fpstp;
850     float32 f32 = float32_squash_input_denormal(input, s);
851     uint32_t val = float32_val(f32);
852     uint32_t f32_sign = float32_is_neg(f32);
853     int f32_exp = extract32(val, 23, 8);
854     uint32_t f32_frac = extract32(val, 0, 23);
855     uint64_t f64_frac;
856 
857     if (float32_is_any_nan(f32)) {
858         float32 nan = f32;
859         if (float32_is_signaling_nan(f32, s)) {
860             float_raise(float_flag_invalid, s);
861             nan = float32_silence_nan(f32, s);
862         }
863         if (s->default_nan_mode) {
864             nan =  float32_default_nan(s);
865         }
866         return nan;
867     } else if (float32_is_zero(f32)) {
868         float_raise(float_flag_divbyzero, s);
869         return float32_set_sign(float32_infinity, float32_is_neg(f32));
870     } else if (float32_is_neg(f32)) {
871         float_raise(float_flag_invalid, s);
872         return float32_default_nan(s);
873     } else if (float32_is_infinity(f32)) {
874         return float32_zero;
875     }
876 
877     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
878      * preserving the parity of the exponent.  */
879 
880     f64_frac = ((uint64_t) f32_frac) << 29;
881 
882     f64_frac = recip_sqrt_estimate(&f32_exp, 380, f64_frac);
883 
884     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(15) */
885     val = deposit32(0, 31, 1, f32_sign);
886     val = deposit32(val, 23, 8, f32_exp);
887     val = deposit32(val, 15, 8, extract64(f64_frac, 52 - 8, 8));
888     return make_float32(val);
889 }
890 
891 float64 HELPER(rsqrte_f64)(float64 input, void *fpstp)
892 {
893     float_status *s = fpstp;
894     float64 f64 = float64_squash_input_denormal(input, s);
895     uint64_t val = float64_val(f64);
896     bool f64_sign = float64_is_neg(f64);
897     int f64_exp = extract64(val, 52, 11);
898     uint64_t f64_frac = extract64(val, 0, 52);
899 
900     if (float64_is_any_nan(f64)) {
901         float64 nan = f64;
902         if (float64_is_signaling_nan(f64, s)) {
903             float_raise(float_flag_invalid, s);
904             nan = float64_silence_nan(f64, s);
905         }
906         if (s->default_nan_mode) {
907             nan =  float64_default_nan(s);
908         }
909         return nan;
910     } else if (float64_is_zero(f64)) {
911         float_raise(float_flag_divbyzero, s);
912         return float64_set_sign(float64_infinity, float64_is_neg(f64));
913     } else if (float64_is_neg(f64)) {
914         float_raise(float_flag_invalid, s);
915         return float64_default_nan(s);
916     } else if (float64_is_infinity(f64)) {
917         return float64_zero;
918     }
919 
920     f64_frac = recip_sqrt_estimate(&f64_exp, 3068, f64_frac);
921 
922     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(44) */
923     val = deposit64(0, 61, 1, f64_sign);
924     val = deposit64(val, 52, 11, f64_exp);
925     val = deposit64(val, 44, 8, extract64(f64_frac, 52 - 8, 8));
926     return make_float64(val);
927 }
928 
929 uint32_t HELPER(recpe_u32)(uint32_t a)
930 {
931     int input, estimate;
932 
933     if ((a & 0x80000000) == 0) {
934         return 0xffffffff;
935     }
936 
937     input = extract32(a, 23, 9);
938     estimate = recip_estimate(input);
939 
940     return deposit32(0, (32 - 9), 9, estimate);
941 }
942 
943 uint32_t HELPER(rsqrte_u32)(uint32_t a)
944 {
945     int estimate;
946 
947     if ((a & 0xc0000000) == 0) {
948         return 0xffffffff;
949     }
950 
951     estimate = do_recip_sqrt_estimate(extract32(a, 23, 9));
952 
953     return deposit32(0, 23, 9, estimate);
954 }
955 
956 /* VFPv4 fused multiply-accumulate */
957 dh_ctype_f16 VFP_HELPER(muladd, h)(dh_ctype_f16 a, dh_ctype_f16 b,
958                                    dh_ctype_f16 c, void *fpstp)
959 {
960     float_status *fpst = fpstp;
961     return float16_muladd(a, b, c, 0, fpst);
962 }
963 
964 float32 VFP_HELPER(muladd, s)(float32 a, float32 b, float32 c, void *fpstp)
965 {
966     float_status *fpst = fpstp;
967     return float32_muladd(a, b, c, 0, fpst);
968 }
969 
970 float64 VFP_HELPER(muladd, d)(float64 a, float64 b, float64 c, void *fpstp)
971 {
972     float_status *fpst = fpstp;
973     return float64_muladd(a, b, c, 0, fpst);
974 }
975 
976 /* ARMv8 round to integral */
977 dh_ctype_f16 HELPER(rinth_exact)(dh_ctype_f16 x, void *fp_status)
978 {
979     return float16_round_to_int(x, fp_status);
980 }
981 
982 float32 HELPER(rints_exact)(float32 x, void *fp_status)
983 {
984     return float32_round_to_int(x, fp_status);
985 }
986 
987 float64 HELPER(rintd_exact)(float64 x, void *fp_status)
988 {
989     return float64_round_to_int(x, fp_status);
990 }
991 
992 dh_ctype_f16 HELPER(rinth)(dh_ctype_f16 x, void *fp_status)
993 {
994     int old_flags = get_float_exception_flags(fp_status), new_flags;
995     float16 ret;
996 
997     ret = float16_round_to_int(x, fp_status);
998 
999     /* Suppress any inexact exceptions the conversion produced */
1000     if (!(old_flags & float_flag_inexact)) {
1001         new_flags = get_float_exception_flags(fp_status);
1002         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1003     }
1004 
1005     return ret;
1006 }
1007 
1008 float32 HELPER(rints)(float32 x, void *fp_status)
1009 {
1010     int old_flags = get_float_exception_flags(fp_status), new_flags;
1011     float32 ret;
1012 
1013     ret = float32_round_to_int(x, fp_status);
1014 
1015     /* Suppress any inexact exceptions the conversion produced */
1016     if (!(old_flags & float_flag_inexact)) {
1017         new_flags = get_float_exception_flags(fp_status);
1018         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1019     }
1020 
1021     return ret;
1022 }
1023 
1024 float64 HELPER(rintd)(float64 x, void *fp_status)
1025 {
1026     int old_flags = get_float_exception_flags(fp_status), new_flags;
1027     float64 ret;
1028 
1029     ret = float64_round_to_int(x, fp_status);
1030 
1031     new_flags = get_float_exception_flags(fp_status);
1032 
1033     /* Suppress any inexact exceptions the conversion produced */
1034     if (!(old_flags & float_flag_inexact)) {
1035         new_flags = get_float_exception_flags(fp_status);
1036         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1037     }
1038 
1039     return ret;
1040 }
1041 
1042 /* Convert ARM rounding mode to softfloat */
1043 int arm_rmode_to_sf(int rmode)
1044 {
1045     switch (rmode) {
1046     case FPROUNDING_TIEAWAY:
1047         rmode = float_round_ties_away;
1048         break;
1049     case FPROUNDING_ODD:
1050         /* FIXME: add support for TIEAWAY and ODD */
1051         qemu_log_mask(LOG_UNIMP, "arm: unimplemented rounding mode: %d\n",
1052                       rmode);
1053         /* fall through for now */
1054     case FPROUNDING_TIEEVEN:
1055     default:
1056         rmode = float_round_nearest_even;
1057         break;
1058     case FPROUNDING_POSINF:
1059         rmode = float_round_up;
1060         break;
1061     case FPROUNDING_NEGINF:
1062         rmode = float_round_down;
1063         break;
1064     case FPROUNDING_ZERO:
1065         rmode = float_round_to_zero;
1066         break;
1067     }
1068     return rmode;
1069 }
1070 
1071 /*
1072  * Implement float64 to int32_t conversion without saturation;
1073  * the result is supplied modulo 2^32.
1074  */
1075 uint64_t HELPER(fjcvtzs)(float64 value, void *vstatus)
1076 {
1077     float_status *status = vstatus;
1078     uint32_t exp, sign;
1079     uint64_t frac;
1080     uint32_t inexact = 1; /* !Z */
1081 
1082     sign = extract64(value, 63, 1);
1083     exp = extract64(value, 52, 11);
1084     frac = extract64(value, 0, 52);
1085 
1086     if (exp == 0) {
1087         /* While not inexact for IEEE FP, -0.0 is inexact for JavaScript.  */
1088         inexact = sign;
1089         if (frac != 0) {
1090             if (status->flush_inputs_to_zero) {
1091                 float_raise(float_flag_input_denormal, status);
1092             } else {
1093                 float_raise(float_flag_inexact, status);
1094                 inexact = 1;
1095             }
1096         }
1097         frac = 0;
1098     } else if (exp == 0x7ff) {
1099         /* This operation raises Invalid for both NaN and overflow (Inf).  */
1100         float_raise(float_flag_invalid, status);
1101         frac = 0;
1102     } else {
1103         int true_exp = exp - 1023;
1104         int shift = true_exp - 52;
1105 
1106         /* Restore implicit bit.  */
1107         frac |= 1ull << 52;
1108 
1109         /* Shift the fraction into place.  */
1110         if (shift >= 0) {
1111             /* The number is so large we must shift the fraction left.  */
1112             if (shift >= 64) {
1113                 /* The fraction is shifted out entirely.  */
1114                 frac = 0;
1115             } else {
1116                 frac <<= shift;
1117             }
1118         } else if (shift > -64) {
1119             /* Normal case -- shift right and notice if bits shift out.  */
1120             inexact = (frac << (64 + shift)) != 0;
1121             frac >>= -shift;
1122         } else {
1123             /* The fraction is shifted out entirely.  */
1124             frac = 0;
1125         }
1126 
1127         /* Notice overflow or inexact exceptions.  */
1128         if (true_exp > 31 || frac > (sign ? 0x80000000ull : 0x7fffffff)) {
1129             /* Overflow, for which this operation raises invalid.  */
1130             float_raise(float_flag_invalid, status);
1131             inexact = 1;
1132         } else if (inexact) {
1133             float_raise(float_flag_inexact, status);
1134         }
1135 
1136         /* Honor the sign.  */
1137         if (sign) {
1138             frac = -frac;
1139         }
1140     }
1141 
1142     /* Pack the result and the env->ZF representation of Z together.  */
1143     return deposit64(frac, 32, 32, inexact);
1144 }
1145 
1146 uint32_t HELPER(vjcvt)(float64 value, CPUARMState *env)
1147 {
1148     uint64_t pair = HELPER(fjcvtzs)(value, &env->vfp.fp_status);
1149     uint32_t result = pair;
1150     uint32_t z = (pair >> 32) == 0;
1151 
1152     /* Store Z, clear NCV, in FPSCR.NZCV.  */
1153     env->vfp.xregs[ARM_VFP_FPSCR]
1154         = (env->vfp.xregs[ARM_VFP_FPSCR] & ~CPSR_NZCV) | (z * CPSR_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