target/arm: Use macros instead of open-coding fp16 conversion helpers
[qemu.git] / target / arm / vfp_helper.c
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 /* Set the current fp rounding mode in the standard fp status and return
463 * the old one. This is for NEON instructions that need to change the
464 * rounding mode but wish to use the standard FPSCR values for everything
465 * else. Always set the rounding mode back to the correct value after
466 * modifying it.
467 * The argument is a softfloat float_round_ value.
468 */
469 uint32_t HELPER(set_neon_rmode)(uint32_t rmode, CPUARMState *env)
470 {
471 float_status *fp_status = &env->vfp.standard_fp_status;
472
473 uint32_t prev_rmode = get_float_rounding_mode(fp_status);
474 set_float_rounding_mode(rmode, fp_status);
475
476 return prev_rmode;
477 }
478
479 /* Half precision conversions. */
480 float32 HELPER(vfp_fcvt_f16_to_f32)(uint32_t a, void *fpstp, uint32_t ahp_mode)
481 {
482 /* Squash FZ16 to 0 for the duration of conversion. In this case,
483 * it would affect flushing input denormals.
484 */
485 float_status *fpst = fpstp;
486 bool save = get_flush_inputs_to_zero(fpst);
487 set_flush_inputs_to_zero(false, fpst);
488 float32 r = float16_to_float32(a, !ahp_mode, fpst);
489 set_flush_inputs_to_zero(save, fpst);
490 return r;
491 }
492
493 uint32_t HELPER(vfp_fcvt_f32_to_f16)(float32 a, void *fpstp, uint32_t ahp_mode)
494 {
495 /* Squash FZ16 to 0 for the duration of conversion. In this case,
496 * it would affect flushing output denormals.
497 */
498 float_status *fpst = fpstp;
499 bool save = get_flush_to_zero(fpst);
500 set_flush_to_zero(false, fpst);
501 float16 r = float32_to_float16(a, !ahp_mode, fpst);
502 set_flush_to_zero(save, fpst);
503 return r;
504 }
505
506 float64 HELPER(vfp_fcvt_f16_to_f64)(uint32_t a, void *fpstp, uint32_t ahp_mode)
507 {
508 /* Squash FZ16 to 0 for the duration of conversion. In this case,
509 * it would affect flushing input denormals.
510 */
511 float_status *fpst = fpstp;
512 bool save = get_flush_inputs_to_zero(fpst);
513 set_flush_inputs_to_zero(false, fpst);
514 float64 r = float16_to_float64(a, !ahp_mode, fpst);
515 set_flush_inputs_to_zero(save, fpst);
516 return r;
517 }
518
519 uint32_t HELPER(vfp_fcvt_f64_to_f16)(float64 a, void *fpstp, uint32_t ahp_mode)
520 {
521 /* Squash FZ16 to 0 for the duration of conversion. In this case,
522 * it would affect flushing output denormals.
523 */
524 float_status *fpst = fpstp;
525 bool save = get_flush_to_zero(fpst);
526 set_flush_to_zero(false, fpst);
527 float16 r = float64_to_float16(a, !ahp_mode, fpst);
528 set_flush_to_zero(save, fpst);
529 return r;
530 }
531
532 float32 HELPER(recps_f32)(CPUARMState *env, float32 a, float32 b)
533 {
534 float_status *s = &env->vfp.standard_fp_status;
535 if ((float32_is_infinity(a) && float32_is_zero_or_denormal(b)) ||
536 (float32_is_infinity(b) && float32_is_zero_or_denormal(a))) {
537 if (!(float32_is_zero(a) || float32_is_zero(b))) {
538 float_raise(float_flag_input_denormal, s);
539 }
540 return float32_two;
541 }
542 return float32_sub(float32_two, float32_mul(a, b, s), s);
543 }
544
545 float32 HELPER(rsqrts_f32)(CPUARMState *env, float32 a, float32 b)
546 {
547 float_status *s = &env->vfp.standard_fp_status;
548 float32 product;
549 if ((float32_is_infinity(a) && float32_is_zero_or_denormal(b)) ||
550 (float32_is_infinity(b) && float32_is_zero_or_denormal(a))) {
551 if (!(float32_is_zero(a) || float32_is_zero(b))) {
552 float_raise(float_flag_input_denormal, s);
553 }
554 return float32_one_point_five;
555 }
556 product = float32_mul(a, b, s);
557 return float32_div(float32_sub(float32_three, product, s), float32_two, s);
558 }
559
560 /* NEON helpers. */
561
562 /* Constants 256 and 512 are used in some helpers; we avoid relying on
563 * int->float conversions at run-time. */
564 #define float64_256 make_float64(0x4070000000000000LL)
565 #define float64_512 make_float64(0x4080000000000000LL)
566 #define float16_maxnorm make_float16(0x7bff)
567 #define float32_maxnorm make_float32(0x7f7fffff)
568 #define float64_maxnorm make_float64(0x7fefffffffffffffLL)
569
570 /* Reciprocal functions
571 *
572 * The algorithm that must be used to calculate the estimate
573 * is specified by the ARM ARM, see FPRecipEstimate()/RecipEstimate
574 */
575
576 /* See RecipEstimate()
577 *
578 * input is a 9 bit fixed point number
579 * input range 256 .. 511 for a number from 0.5 <= x < 1.0.
580 * result range 256 .. 511 for a number from 1.0 to 511/256.
581 */
582
583 static int recip_estimate(int input)
584 {
585 int a, b, r;
586 assert(256 <= input && input < 512);
587 a = (input * 2) + 1;
588 b = (1 << 19) / a;
589 r = (b + 1) >> 1;
590 assert(256 <= r && r < 512);
591 return r;
592 }
593
594 /*
595 * Common wrapper to call recip_estimate
596 *
597 * The parameters are exponent and 64 bit fraction (without implicit
598 * bit) where the binary point is nominally at bit 52. Returns a
599 * float64 which can then be rounded to the appropriate size by the
600 * callee.
601 */
602
603 static uint64_t call_recip_estimate(int *exp, int exp_off, uint64_t frac)
604 {
605 uint32_t scaled, estimate;
606 uint64_t result_frac;
607 int result_exp;
608
609 /* Handle sub-normals */
610 if (*exp == 0) {
611 if (extract64(frac, 51, 1) == 0) {
612 *exp = -1;
613 frac <<= 2;
614 } else {
615 frac <<= 1;
616 }
617 }
618
619 /* scaled = UInt('1':fraction<51:44>) */
620 scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
621 estimate = recip_estimate(scaled);
622
623 result_exp = exp_off - *exp;
624 result_frac = deposit64(0, 44, 8, estimate);
625 if (result_exp == 0) {
626 result_frac = deposit64(result_frac >> 1, 51, 1, 1);
627 } else if (result_exp == -1) {
628 result_frac = deposit64(result_frac >> 2, 50, 2, 1);
629 result_exp = 0;
630 }
631
632 *exp = result_exp;
633
634 return result_frac;
635 }
636
637 static bool round_to_inf(float_status *fpst, bool sign_bit)
638 {
639 switch (fpst->float_rounding_mode) {
640 case float_round_nearest_even: /* Round to Nearest */
641 return true;
642 case float_round_up: /* Round to +Inf */
643 return !sign_bit;
644 case float_round_down: /* Round to -Inf */
645 return sign_bit;
646 case float_round_to_zero: /* Round to Zero */
647 return false;
648 default:
649 g_assert_not_reached();
650 }
651 }
652
653 uint32_t HELPER(recpe_f16)(uint32_t input, void *fpstp)
654 {
655 float_status *fpst = fpstp;
656 float16 f16 = float16_squash_input_denormal(input, fpst);
657 uint32_t f16_val = float16_val(f16);
658 uint32_t f16_sign = float16_is_neg(f16);
659 int f16_exp = extract32(f16_val, 10, 5);
660 uint32_t f16_frac = extract32(f16_val, 0, 10);
661 uint64_t f64_frac;
662
663 if (float16_is_any_nan(f16)) {
664 float16 nan = f16;
665 if (float16_is_signaling_nan(f16, fpst)) {
666 float_raise(float_flag_invalid, fpst);
667 nan = float16_silence_nan(f16, fpst);
668 }
669 if (fpst->default_nan_mode) {
670 nan = float16_default_nan(fpst);
671 }
672 return nan;
673 } else if (float16_is_infinity(f16)) {
674 return float16_set_sign(float16_zero, float16_is_neg(f16));
675 } else if (float16_is_zero(f16)) {
676 float_raise(float_flag_divbyzero, fpst);
677 return float16_set_sign(float16_infinity, float16_is_neg(f16));
678 } else if (float16_abs(f16) < (1 << 8)) {
679 /* Abs(value) < 2.0^-16 */
680 float_raise(float_flag_overflow | float_flag_inexact, fpst);
681 if (round_to_inf(fpst, f16_sign)) {
682 return float16_set_sign(float16_infinity, f16_sign);
683 } else {
684 return float16_set_sign(float16_maxnorm, f16_sign);
685 }
686 } else if (f16_exp >= 29 && fpst->flush_to_zero) {
687 float_raise(float_flag_underflow, fpst);
688 return float16_set_sign(float16_zero, float16_is_neg(f16));
689 }
690
691 f64_frac = call_recip_estimate(&f16_exp, 29,
692 ((uint64_t) f16_frac) << (52 - 10));
693
694 /* result = sign : result_exp<4:0> : fraction<51:42> */
695 f16_val = deposit32(0, 15, 1, f16_sign);
696 f16_val = deposit32(f16_val, 10, 5, f16_exp);
697 f16_val = deposit32(f16_val, 0, 10, extract64(f64_frac, 52 - 10, 10));
698 return make_float16(f16_val);
699 }
700
701 float32 HELPER(recpe_f32)(float32 input, void *fpstp)
702 {
703 float_status *fpst = fpstp;
704 float32 f32 = float32_squash_input_denormal(input, fpst);
705 uint32_t f32_val = float32_val(f32);
706 bool f32_sign = float32_is_neg(f32);
707 int f32_exp = extract32(f32_val, 23, 8);
708 uint32_t f32_frac = extract32(f32_val, 0, 23);
709 uint64_t f64_frac;
710
711 if (float32_is_any_nan(f32)) {
712 float32 nan = f32;
713 if (float32_is_signaling_nan(f32, fpst)) {
714 float_raise(float_flag_invalid, fpst);
715 nan = float32_silence_nan(f32, fpst);
716 }
717 if (fpst->default_nan_mode) {
718 nan = float32_default_nan(fpst);
719 }
720 return nan;
721 } else if (float32_is_infinity(f32)) {
722 return float32_set_sign(float32_zero, float32_is_neg(f32));
723 } else if (float32_is_zero(f32)) {
724 float_raise(float_flag_divbyzero, fpst);
725 return float32_set_sign(float32_infinity, float32_is_neg(f32));
726 } else if (float32_abs(f32) < (1ULL << 21)) {
727 /* Abs(value) < 2.0^-128 */
728 float_raise(float_flag_overflow | float_flag_inexact, fpst);
729 if (round_to_inf(fpst, f32_sign)) {
730 return float32_set_sign(float32_infinity, f32_sign);
731 } else {
732 return float32_set_sign(float32_maxnorm, f32_sign);
733 }
734 } else if (f32_exp >= 253 && fpst->flush_to_zero) {
735 float_raise(float_flag_underflow, fpst);
736 return float32_set_sign(float32_zero, float32_is_neg(f32));
737 }
738
739 f64_frac = call_recip_estimate(&f32_exp, 253,
740 ((uint64_t) f32_frac) << (52 - 23));
741
742 /* result = sign : result_exp<7:0> : fraction<51:29> */
743 f32_val = deposit32(0, 31, 1, f32_sign);
744 f32_val = deposit32(f32_val, 23, 8, f32_exp);
745 f32_val = deposit32(f32_val, 0, 23, extract64(f64_frac, 52 - 23, 23));
746 return make_float32(f32_val);
747 }
748
749 float64 HELPER(recpe_f64)(float64 input, void *fpstp)
750 {
751 float_status *fpst = fpstp;
752 float64 f64 = float64_squash_input_denormal(input, fpst);
753 uint64_t f64_val = float64_val(f64);
754 bool f64_sign = float64_is_neg(f64);
755 int f64_exp = extract64(f64_val, 52, 11);
756 uint64_t f64_frac = extract64(f64_val, 0, 52);
757
758 /* Deal with any special cases */
759 if (float64_is_any_nan(f64)) {
760 float64 nan = f64;
761 if (float64_is_signaling_nan(f64, fpst)) {
762 float_raise(float_flag_invalid, fpst);
763 nan = float64_silence_nan(f64, fpst);
764 }
765 if (fpst->default_nan_mode) {
766 nan = float64_default_nan(fpst);
767 }
768 return nan;
769 } else if (float64_is_infinity(f64)) {
770 return float64_set_sign(float64_zero, float64_is_neg(f64));
771 } else if (float64_is_zero(f64)) {
772 float_raise(float_flag_divbyzero, fpst);
773 return float64_set_sign(float64_infinity, float64_is_neg(f64));
774 } else if ((f64_val & ~(1ULL << 63)) < (1ULL << 50)) {
775 /* Abs(value) < 2.0^-1024 */
776 float_raise(float_flag_overflow | float_flag_inexact, fpst);
777 if (round_to_inf(fpst, f64_sign)) {
778 return float64_set_sign(float64_infinity, f64_sign);
779 } else {
780 return float64_set_sign(float64_maxnorm, f64_sign);
781 }
782 } else if (f64_exp >= 2045 && fpst->flush_to_zero) {
783 float_raise(float_flag_underflow, fpst);
784 return float64_set_sign(float64_zero, float64_is_neg(f64));
785 }
786
787 f64_frac = call_recip_estimate(&f64_exp, 2045, f64_frac);
788
789 /* result = sign : result_exp<10:0> : fraction<51:0>; */
790 f64_val = deposit64(0, 63, 1, f64_sign);
791 f64_val = deposit64(f64_val, 52, 11, f64_exp);
792 f64_val = deposit64(f64_val, 0, 52, f64_frac);
793 return make_float64(f64_val);
794 }
795
796 /* The algorithm that must be used to calculate the estimate
797 * is specified by the ARM ARM.
798 */
799
800 static int do_recip_sqrt_estimate(int a)
801 {
802 int b, estimate;
803
804 assert(128 <= a && a < 512);
805 if (a < 256) {
806 a = a * 2 + 1;
807 } else {
808 a = (a >> 1) << 1;
809 a = (a + 1) * 2;
810 }
811 b = 512;
812 while (a * (b + 1) * (b + 1) < (1 << 28)) {
813 b += 1;
814 }
815 estimate = (b + 1) / 2;
816 assert(256 <= estimate && estimate < 512);
817
818 return estimate;
819 }
820
821
822 static uint64_t recip_sqrt_estimate(int *exp , int exp_off, uint64_t frac)
823 {
824 int estimate;
825 uint32_t scaled;
826
827 if (*exp == 0) {
828 while (extract64(frac, 51, 1) == 0) {
829 frac = frac << 1;
830 *exp -= 1;
831 }
832 frac = extract64(frac, 0, 51) << 1;
833 }
834
835 if (*exp & 1) {
836 /* scaled = UInt('01':fraction<51:45>) */
837 scaled = deposit32(1 << 7, 0, 7, extract64(frac, 45, 7));
838 } else {
839 /* scaled = UInt('1':fraction<51:44>) */
840 scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
841 }
842 estimate = do_recip_sqrt_estimate(scaled);
843
844 *exp = (exp_off - *exp) / 2;
845 return extract64(estimate, 0, 8) << 44;
846 }
847
848 uint32_t HELPER(rsqrte_f16)(uint32_t input, void *fpstp)
849 {
850 float_status *s = fpstp;
851 float16 f16 = float16_squash_input_denormal(input, s);
852 uint16_t val = float16_val(f16);
853 bool f16_sign = float16_is_neg(f16);
854 int f16_exp = extract32(val, 10, 5);
855 uint16_t f16_frac = extract32(val, 0, 10);
856 uint64_t f64_frac;
857
858 if (float16_is_any_nan(f16)) {
859 float16 nan = f16;
860 if (float16_is_signaling_nan(f16, s)) {
861 float_raise(float_flag_invalid, s);
862 nan = float16_silence_nan(f16, s);
863 }
864 if (s->default_nan_mode) {
865 nan = float16_default_nan(s);
866 }
867 return nan;
868 } else if (float16_is_zero(f16)) {
869 float_raise(float_flag_divbyzero, s);
870 return float16_set_sign(float16_infinity, f16_sign);
871 } else if (f16_sign) {
872 float_raise(float_flag_invalid, s);
873 return float16_default_nan(s);
874 } else if (float16_is_infinity(f16)) {
875 return float16_zero;
876 }
877
878 /* Scale and normalize to a double-precision value between 0.25 and 1.0,
879 * preserving the parity of the exponent. */
880
881 f64_frac = ((uint64_t) f16_frac) << (52 - 10);
882
883 f64_frac = recip_sqrt_estimate(&f16_exp, 44, f64_frac);
884
885 /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(2) */
886 val = deposit32(0, 15, 1, f16_sign);
887 val = deposit32(val, 10, 5, f16_exp);
888 val = deposit32(val, 2, 8, extract64(f64_frac, 52 - 8, 8));
889 return make_float16(val);
890 }
891
892 float32 HELPER(rsqrte_f32)(float32 input, void *fpstp)
893 {
894 float_status *s = fpstp;
895 float32 f32 = float32_squash_input_denormal(input, s);
896 uint32_t val = float32_val(f32);
897 uint32_t f32_sign = float32_is_neg(f32);
898 int f32_exp = extract32(val, 23, 8);
899 uint32_t f32_frac = extract32(val, 0, 23);
900 uint64_t f64_frac;
901
902 if (float32_is_any_nan(f32)) {
903 float32 nan = f32;
904 if (float32_is_signaling_nan(f32, s)) {
905 float_raise(float_flag_invalid, s);
906 nan = float32_silence_nan(f32, s);
907 }
908 if (s->default_nan_mode) {
909 nan = float32_default_nan(s);
910 }
911 return nan;
912 } else if (float32_is_zero(f32)) {
913 float_raise(float_flag_divbyzero, s);
914 return float32_set_sign(float32_infinity, float32_is_neg(f32));
915 } else if (float32_is_neg(f32)) {
916 float_raise(float_flag_invalid, s);
917 return float32_default_nan(s);
918 } else if (float32_is_infinity(f32)) {
919 return float32_zero;
920 }
921
922 /* Scale and normalize to a double-precision value between 0.25 and 1.0,
923 * preserving the parity of the exponent. */
924
925 f64_frac = ((uint64_t) f32_frac) << 29;
926
927 f64_frac = recip_sqrt_estimate(&f32_exp, 380, f64_frac);
928
929 /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(15) */
930 val = deposit32(0, 31, 1, f32_sign);
931 val = deposit32(val, 23, 8, f32_exp);
932 val = deposit32(val, 15, 8, extract64(f64_frac, 52 - 8, 8));
933 return make_float32(val);
934 }
935
936 float64 HELPER(rsqrte_f64)(float64 input, void *fpstp)
937 {
938 float_status *s = fpstp;
939 float64 f64 = float64_squash_input_denormal(input, s);
940 uint64_t val = float64_val(f64);
941 bool f64_sign = float64_is_neg(f64);
942 int f64_exp = extract64(val, 52, 11);
943 uint64_t f64_frac = extract64(val, 0, 52);
944
945 if (float64_is_any_nan(f64)) {
946 float64 nan = f64;
947 if (float64_is_signaling_nan(f64, s)) {
948 float_raise(float_flag_invalid, s);
949 nan = float64_silence_nan(f64, s);
950 }
951 if (s->default_nan_mode) {
952 nan = float64_default_nan(s);
953 }
954 return nan;
955 } else if (float64_is_zero(f64)) {
956 float_raise(float_flag_divbyzero, s);
957 return float64_set_sign(float64_infinity, float64_is_neg(f64));
958 } else if (float64_is_neg(f64)) {
959 float_raise(float_flag_invalid, s);
960 return float64_default_nan(s);
961 } else if (float64_is_infinity(f64)) {
962 return float64_zero;
963 }
964
965 f64_frac = recip_sqrt_estimate(&f64_exp, 3068, f64_frac);
966
967 /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(44) */
968 val = deposit64(0, 61, 1, f64_sign);
969 val = deposit64(val, 52, 11, f64_exp);
970 val = deposit64(val, 44, 8, extract64(f64_frac, 52 - 8, 8));
971 return make_float64(val);
972 }
973
974 uint32_t HELPER(recpe_u32)(uint32_t a)
975 {
976 int input, estimate;
977
978 if ((a & 0x80000000) == 0) {
979 return 0xffffffff;
980 }
981
982 input = extract32(a, 23, 9);
983 estimate = recip_estimate(input);
984
985 return deposit32(0, (32 - 9), 9, estimate);
986 }
987
988 uint32_t HELPER(rsqrte_u32)(uint32_t a)
989 {
990 int estimate;
991
992 if ((a & 0xc0000000) == 0) {
993 return 0xffffffff;
994 }
995
996 estimate = do_recip_sqrt_estimate(extract32(a, 23, 9));
997
998 return deposit32(0, 23, 9, estimate);
999 }
1000
1001 /* VFPv4 fused multiply-accumulate */
1002 dh_ctype_f16 VFP_HELPER(muladd, h)(dh_ctype_f16 a, dh_ctype_f16 b,
1003 dh_ctype_f16 c, void *fpstp)
1004 {
1005 float_status *fpst = fpstp;
1006 return float16_muladd(a, b, c, 0, fpst);
1007 }
1008
1009 float32 VFP_HELPER(muladd, s)(float32 a, float32 b, float32 c, void *fpstp)
1010 {
1011 float_status *fpst = fpstp;
1012 return float32_muladd(a, b, c, 0, fpst);
1013 }
1014
1015 float64 VFP_HELPER(muladd, d)(float64 a, float64 b, float64 c, void *fpstp)
1016 {
1017 float_status *fpst = fpstp;
1018 return float64_muladd(a, b, c, 0, fpst);
1019 }
1020
1021 /* ARMv8 round to integral */
1022 float32 HELPER(rints_exact)(float32 x, void *fp_status)
1023 {
1024 return float32_round_to_int(x, fp_status);
1025 }
1026
1027 float64 HELPER(rintd_exact)(float64 x, void *fp_status)
1028 {
1029 return float64_round_to_int(x, fp_status);
1030 }
1031
1032 float32 HELPER(rints)(float32 x, void *fp_status)
1033 {
1034 int old_flags = get_float_exception_flags(fp_status), new_flags;
1035 float32 ret;
1036
1037 ret = float32_round_to_int(x, fp_status);
1038
1039 /* Suppress any inexact exceptions the conversion produced */
1040 if (!(old_flags & float_flag_inexact)) {
1041 new_flags = get_float_exception_flags(fp_status);
1042 set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1043 }
1044
1045 return ret;
1046 }
1047
1048 float64 HELPER(rintd)(float64 x, void *fp_status)
1049 {
1050 int old_flags = get_float_exception_flags(fp_status), new_flags;
1051 float64 ret;
1052
1053 ret = float64_round_to_int(x, fp_status);
1054
1055 new_flags = get_float_exception_flags(fp_status);
1056
1057 /* Suppress any inexact exceptions the conversion produced */
1058 if (!(old_flags & float_flag_inexact)) {
1059 new_flags = get_float_exception_flags(fp_status);
1060 set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1061 }
1062
1063 return ret;
1064 }
1065
1066 /* Convert ARM rounding mode to softfloat */
1067 int arm_rmode_to_sf(int rmode)
1068 {
1069 switch (rmode) {
1070 case FPROUNDING_TIEAWAY:
1071 rmode = float_round_ties_away;
1072 break;
1073 case FPROUNDING_ODD:
1074 /* FIXME: add support for TIEAWAY and ODD */
1075 qemu_log_mask(LOG_UNIMP, "arm: unimplemented rounding mode: %d\n",
1076 rmode);
1077 /* fall through for now */
1078 case FPROUNDING_TIEEVEN:
1079 default:
1080 rmode = float_round_nearest_even;
1081 break;
1082 case FPROUNDING_POSINF:
1083 rmode = float_round_up;
1084 break;
1085 case FPROUNDING_NEGINF:
1086 rmode = float_round_down;
1087 break;
1088 case FPROUNDING_ZERO:
1089 rmode = float_round_to_zero;
1090 break;
1091 }
1092 return rmode;
1093 }
1094
1095 /*
1096 * Implement float64 to int32_t conversion without saturation;
1097 * the result is supplied modulo 2^32.
1098 */
1099 uint64_t HELPER(fjcvtzs)(float64 value, void *vstatus)
1100 {
1101 float_status *status = vstatus;
1102 uint32_t exp, sign;
1103 uint64_t frac;
1104 uint32_t inexact = 1; /* !Z */
1105
1106 sign = extract64(value, 63, 1);
1107 exp = extract64(value, 52, 11);
1108 frac = extract64(value, 0, 52);
1109
1110 if (exp == 0) {
1111 /* While not inexact for IEEE FP, -0.0 is inexact for JavaScript. */
1112 inexact = sign;
1113 if (frac != 0) {
1114 if (status->flush_inputs_to_zero) {
1115 float_raise(float_flag_input_denormal, status);
1116 } else {
1117 float_raise(float_flag_inexact, status);
1118 inexact = 1;
1119 }
1120 }
1121 frac = 0;
1122 } else if (exp == 0x7ff) {
1123 /* This operation raises Invalid for both NaN and overflow (Inf). */
1124 float_raise(float_flag_invalid, status);
1125 frac = 0;
1126 } else {
1127 int true_exp = exp - 1023;
1128 int shift = true_exp - 52;
1129
1130 /* Restore implicit bit. */
1131 frac |= 1ull << 52;
1132
1133 /* Shift the fraction into place. */
1134 if (shift >= 0) {
1135 /* The number is so large we must shift the fraction left. */
1136 if (shift >= 64) {
1137 /* The fraction is shifted out entirely. */
1138 frac = 0;
1139 } else {
1140 frac <<= shift;
1141 }
1142 } else if (shift > -64) {
1143 /* Normal case -- shift right and notice if bits shift out. */
1144 inexact = (frac << (64 + shift)) != 0;
1145 frac >>= -shift;
1146 } else {
1147 /* The fraction is shifted out entirely. */
1148 frac = 0;
1149 }
1150
1151 /* Notice overflow or inexact exceptions. */
1152 if (true_exp > 31 || frac > (sign ? 0x80000000ull : 0x7fffffff)) {
1153 /* Overflow, for which this operation raises invalid. */
1154 float_raise(float_flag_invalid, status);
1155 inexact = 1;
1156 } else if (inexact) {
1157 float_raise(float_flag_inexact, status);
1158 }
1159
1160 /* Honor the sign. */
1161 if (sign) {
1162 frac = -frac;
1163 }
1164 }
1165
1166 /* Pack the result and the env->ZF representation of Z together. */
1167 return deposit64(frac, 32, 32, inexact);
1168 }
1169
1170 uint32_t HELPER(vjcvt)(float64 value, CPUARMState *env)
1171 {
1172 uint64_t pair = HELPER(fjcvtzs)(value, &env->vfp.fp_status);
1173 uint32_t result = pair;
1174 uint32_t z = (pair >> 32) == 0;
1175
1176 /* Store Z, clear NCV, in FPSCR.NZCV. */
1177 env->vfp.xregs[ARM_VFP_FPSCR]
1178 = (env->vfp.xregs[ARM_VFP_FPSCR] & ~CPSR_NZCV) | (z * CPSR_Z);
1179
1180 return result;
1181 }
1182
1183 /* Round a float32 to an integer that fits in int32_t or int64_t. */
1184 static float32 frint_s(float32 f, float_status *fpst, int intsize)
1185 {
1186 int old_flags = get_float_exception_flags(fpst);
1187 uint32_t exp = extract32(f, 23, 8);
1188
1189 if (unlikely(exp == 0xff)) {
1190 /* NaN or Inf. */
1191 goto overflow;
1192 }
1193
1194 /* Round and re-extract the exponent. */
1195 f = float32_round_to_int(f, fpst);
1196 exp = extract32(f, 23, 8);
1197
1198 /* Validate the range of the result. */
1199 if (exp < 126 + intsize) {
1200 /* abs(F) <= INT{N}_MAX */
1201 return f;
1202 }
1203 if (exp == 126 + intsize) {
1204 uint32_t sign = extract32(f, 31, 1);
1205 uint32_t frac = extract32(f, 0, 23);
1206 if (sign && frac == 0) {
1207 /* F == INT{N}_MIN */
1208 return f;
1209 }
1210 }
1211
1212 overflow:
1213 /*
1214 * Raise Invalid and return INT{N}_MIN as a float. Revert any
1215 * inexact exception float32_round_to_int may have raised.
1216 */
1217 set_float_exception_flags(old_flags | float_flag_invalid, fpst);
1218 return (0x100u + 126u + intsize) << 23;
1219 }
1220
1221 float32 HELPER(frint32_s)(float32 f, void *fpst)
1222 {
1223 return frint_s(f, fpst, 32);
1224 }
1225
1226 float32 HELPER(frint64_s)(float32 f, void *fpst)
1227 {
1228 return frint_s(f, fpst, 64);
1229 }
1230
1231 /* Round a float64 to an integer that fits in int32_t or int64_t. */
1232 static float64 frint_d(float64 f, float_status *fpst, int intsize)
1233 {
1234 int old_flags = get_float_exception_flags(fpst);
1235 uint32_t exp = extract64(f, 52, 11);
1236
1237 if (unlikely(exp == 0x7ff)) {
1238 /* NaN or Inf. */
1239 goto overflow;
1240 }
1241
1242 /* Round and re-extract the exponent. */
1243 f = float64_round_to_int(f, fpst);
1244 exp = extract64(f, 52, 11);
1245
1246 /* Validate the range of the result. */
1247 if (exp < 1022 + intsize) {
1248 /* abs(F) <= INT{N}_MAX */
1249 return f;
1250 }
1251 if (exp == 1022 + intsize) {
1252 uint64_t sign = extract64(f, 63, 1);
1253 uint64_t frac = extract64(f, 0, 52);
1254 if (sign && frac == 0) {
1255 /* F == INT{N}_MIN */
1256 return f;
1257 }
1258 }
1259
1260 overflow:
1261 /*
1262 * Raise Invalid and return INT{N}_MIN as a float. Revert any
1263 * inexact exception float64_round_to_int may have raised.
1264 */
1265 set_float_exception_flags(old_flags | float_flag_invalid, fpst);
1266 return (uint64_t)(0x800 + 1022 + intsize) << 52;
1267 }
1268
1269 float64 HELPER(frint32_d)(float64 f, void *fpst)
1270 {
1271 return frint_d(f, fpst, 32);
1272 }
1273
1274 float64 HELPER(frint64_d)(float64 f, void *fpst)
1275 {
1276 return frint_d(f, fpst, 64);
1277 }
1278
1279 void HELPER(check_hcr_el2_trap)(CPUARMState *env, uint32_t rt, uint32_t reg)
1280 {
1281 uint32_t syndrome;
1282
1283 switch (reg) {
1284 case ARM_VFP_MVFR0:
1285 case ARM_VFP_MVFR1:
1286 case ARM_VFP_MVFR2:
1287 if (!(arm_hcr_el2_eff(env) & HCR_TID3)) {
1288 return;
1289 }
1290 break;
1291 case ARM_VFP_FPSID:
1292 if (!(arm_hcr_el2_eff(env) & HCR_TID0)) {
1293 return;
1294 }
1295 break;
1296 default:
1297 g_assert_not_reached();
1298 }
1299
1300 syndrome = ((EC_FPIDTRAP << ARM_EL_EC_SHIFT)
1301 | ARM_EL_IL
1302 | (1 << 24) | (0xe << 20) | (7 << 14)
1303 | (reg << 10) | (rt << 5) | 1);
1304
1305 raise_exception(env, EXCP_HYP_TRAP, syndrome, 2);
1306 }
1307
1308 #endif