xref: /openbmc/qemu/target/arm/vfp_helper.c (revision f147ed37)
1 /*
2  * ARM VFP floating-point operations
3  *
4  *  Copyright (c) 2003 Fabrice Bellard
5  *
6  * This library is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * This library is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with this library; if not, see <http://www.gnu.org/licenses/>.
18  */
19 
20 #include "qemu/osdep.h"
21 #include "cpu.h"
22 #include "exec/helper-proto.h"
23 #include "internals.h"
24 #include "cpu-features.h"
25 #ifdef CONFIG_TCG
26 #include "qemu/log.h"
27 #include "fpu/softfloat.h"
28 #endif
29 
30 /* VFP support.  We follow the convention used for VFP instructions:
31    Single precision routines have a "s" suffix, double precision a
32    "d" suffix.  */
33 
34 #ifdef CONFIG_TCG
35 
36 /* Convert host exception flags to vfp form.  */
37 static inline int vfp_exceptbits_from_host(int host_bits)
38 {
39     int target_bits = 0;
40 
41     if (host_bits & float_flag_invalid) {
42         target_bits |= 1;
43     }
44     if (host_bits & float_flag_divbyzero) {
45         target_bits |= 2;
46     }
47     if (host_bits & float_flag_overflow) {
48         target_bits |= 4;
49     }
50     if (host_bits & (float_flag_underflow | float_flag_output_denormal)) {
51         target_bits |= 8;
52     }
53     if (host_bits & float_flag_inexact) {
54         target_bits |= 0x10;
55     }
56     if (host_bits & float_flag_input_denormal) {
57         target_bits |= 0x80;
58     }
59     return target_bits;
60 }
61 
62 static uint32_t vfp_get_fpsr_from_host(CPUARMState *env)
63 {
64     uint32_t i;
65 
66     i = get_float_exception_flags(&env->vfp.fp_status);
67     i |= get_float_exception_flags(&env->vfp.standard_fp_status);
68     /* FZ16 does not generate an input denormal exception.  */
69     i |= (get_float_exception_flags(&env->vfp.fp_status_f16)
70           & ~float_flag_input_denormal);
71     i |= (get_float_exception_flags(&env->vfp.standard_fp_status_f16)
72           & ~float_flag_input_denormal);
73     return vfp_exceptbits_from_host(i);
74 }
75 
76 static void vfp_clear_float_status_exc_flags(CPUARMState *env)
77 {
78     /*
79      * Clear out all the exception-flag information in the float_status
80      * values. The caller should have arranged for env->vfp.fpsr to
81      * be the architecturally up-to-date exception flag information first.
82      */
83     set_float_exception_flags(0, &env->vfp.fp_status);
84     set_float_exception_flags(0, &env->vfp.fp_status_f16);
85     set_float_exception_flags(0, &env->vfp.standard_fp_status);
86     set_float_exception_flags(0, &env->vfp.standard_fp_status_f16);
87 }
88 
89 static void vfp_set_fpcr_to_host(CPUARMState *env, uint32_t val, uint32_t mask)
90 {
91     uint64_t changed = env->vfp.fpcr;
92 
93     changed ^= val;
94     changed &= mask;
95     if (changed & (3 << 22)) {
96         int i = (val >> 22) & 3;
97         switch (i) {
98         case FPROUNDING_TIEEVEN:
99             i = float_round_nearest_even;
100             break;
101         case FPROUNDING_POSINF:
102             i = float_round_up;
103             break;
104         case FPROUNDING_NEGINF:
105             i = float_round_down;
106             break;
107         case FPROUNDING_ZERO:
108             i = float_round_to_zero;
109             break;
110         }
111         set_float_rounding_mode(i, &env->vfp.fp_status);
112         set_float_rounding_mode(i, &env->vfp.fp_status_f16);
113     }
114     if (changed & FPCR_FZ16) {
115         bool ftz_enabled = val & FPCR_FZ16;
116         set_flush_to_zero(ftz_enabled, &env->vfp.fp_status_f16);
117         set_flush_to_zero(ftz_enabled, &env->vfp.standard_fp_status_f16);
118         set_flush_inputs_to_zero(ftz_enabled, &env->vfp.fp_status_f16);
119         set_flush_inputs_to_zero(ftz_enabled, &env->vfp.standard_fp_status_f16);
120     }
121     if (changed & FPCR_FZ) {
122         bool ftz_enabled = val & FPCR_FZ;
123         set_flush_to_zero(ftz_enabled, &env->vfp.fp_status);
124         set_flush_inputs_to_zero(ftz_enabled, &env->vfp.fp_status);
125     }
126     if (changed & FPCR_DN) {
127         bool dnan_enabled = val & FPCR_DN;
128         set_default_nan_mode(dnan_enabled, &env->vfp.fp_status);
129         set_default_nan_mode(dnan_enabled, &env->vfp.fp_status_f16);
130     }
131 }
132 
133 #else
134 
135 static uint32_t vfp_get_fpsr_from_host(CPUARMState *env)
136 {
137     return 0;
138 }
139 
140 static void vfp_clear_float_status_exc_flags(CPUARMState *env)
141 {
142 }
143 
144 static void vfp_set_fpcr_to_host(CPUARMState *env, uint32_t val, uint32_t mask)
145 {
146 }
147 
148 #endif
149 
150 uint32_t vfp_get_fpcr(CPUARMState *env)
151 {
152     uint32_t fpcr = env->vfp.fpcr
153         | (env->vfp.vec_len << 16)
154         | (env->vfp.vec_stride << 20);
155 
156     /*
157      * M-profile LTPSIZE is the same bits [18:16] as A-profile Len; whichever
158      * of the two is not applicable to this CPU will always be zero.
159      */
160     fpcr |= env->v7m.ltpsize << 16;
161 
162     return fpcr;
163 }
164 
165 uint32_t vfp_get_fpsr(CPUARMState *env)
166 {
167     uint32_t fpsr = env->vfp.fpsr;
168     uint32_t i;
169 
170     fpsr |= vfp_get_fpsr_from_host(env);
171 
172     i = env->vfp.qc[0] | env->vfp.qc[1] | env->vfp.qc[2] | env->vfp.qc[3];
173     fpsr |= i ? FPSR_QC : 0;
174     return fpsr;
175 }
176 
177 uint32_t HELPER(vfp_get_fpscr)(CPUARMState *env)
178 {
179     return (vfp_get_fpcr(env) & FPSCR_FPCR_MASK) |
180         (vfp_get_fpsr(env) & FPSCR_FPSR_MASK);
181 }
182 
183 uint32_t vfp_get_fpscr(CPUARMState *env)
184 {
185     return HELPER(vfp_get_fpscr)(env);
186 }
187 
188 void vfp_set_fpsr(CPUARMState *env, uint32_t val)
189 {
190     ARMCPU *cpu = env_archcpu(env);
191 
192     if (arm_feature(env, ARM_FEATURE_NEON) ||
193         cpu_isar_feature(aa32_mve, cpu)) {
194         /*
195          * The bit we set within vfp.qc[] is arbitrary; the array as a
196          * whole being zero/non-zero is what counts.
197          */
198         env->vfp.qc[0] = val & FPSR_QC;
199         env->vfp.qc[1] = 0;
200         env->vfp.qc[2] = 0;
201         env->vfp.qc[3] = 0;
202     }
203 
204     /*
205      * NZCV lives only in env->vfp.fpsr. The cumulative exception flags
206      * IOC|DZC|OFC|UFC|IXC|IDC also live in env->vfp.fpsr, with possible
207      * extra pending exception information that hasn't yet been folded in
208      * living in the float_status values (for TCG).
209      * Since this FPSR write gives us the up to date values of the exception
210      * flags, we want to store into vfp.fpsr the NZCV and CEXC bits, zeroing
211      * anything else. We also need to clear out the float_status exception
212      * information so that the next vfp_get_fpsr does not fold in stale data.
213      */
214     val &= FPSR_NZCV_MASK | FPSR_CEXC_MASK;
215     env->vfp.fpsr = val;
216     vfp_clear_float_status_exc_flags(env);
217 }
218 
219 static void vfp_set_fpcr_masked(CPUARMState *env, uint32_t val, uint32_t mask)
220 {
221     /*
222      * We only set FPCR bits defined by mask, and leave the others alone.
223      * We assume the mask is sensible (e.g. doesn't try to set only
224      * part of a field)
225      */
226     ARMCPU *cpu = env_archcpu(env);
227 
228     /* When ARMv8.2-FP16 is not supported, FZ16 is RES0.  */
229     if (!cpu_isar_feature(any_fp16, cpu)) {
230         val &= ~FPCR_FZ16;
231     }
232 
233     vfp_set_fpcr_to_host(env, val, mask);
234 
235     if (mask & (FPCR_LEN_MASK | FPCR_STRIDE_MASK)) {
236         if (!arm_feature(env, ARM_FEATURE_M)) {
237             /*
238              * Short-vector length and stride; on M-profile these bits
239              * are used for different purposes.
240              * We can't make this conditional be "if MVFR0.FPShVec != 0",
241              * because in v7A no-short-vector-support cores still had to
242              * allow Stride/Len to be written with the only effect that
243              * some insns are required to UNDEF if the guest sets them.
244              */
245             env->vfp.vec_len = extract32(val, 16, 3);
246             env->vfp.vec_stride = extract32(val, 20, 2);
247         } else if (cpu_isar_feature(aa32_mve, cpu)) {
248             env->v7m.ltpsize = extract32(val, FPCR_LTPSIZE_SHIFT,
249                                          FPCR_LTPSIZE_LENGTH);
250         }
251     }
252 
253     /*
254      * We don't implement trapped exception handling, so the
255      * trap enable bits, IDE|IXE|UFE|OFE|DZE|IOE are all RAZ/WI (not RES0!)
256      *
257      * The FPCR bits we keep in vfp.fpcr are AHP, DN, FZ, RMode
258      * and FZ16. Len, Stride and LTPSIZE we just handled. Store those bits
259      * there, and zero any of the other FPCR bits and the RES0 and RAZ/WI
260      * bits.
261      */
262     val &= FPCR_AHP | FPCR_DN | FPCR_FZ | FPCR_RMODE_MASK | FPCR_FZ16;
263     env->vfp.fpcr &= ~mask;
264     env->vfp.fpcr |= val;
265 }
266 
267 void vfp_set_fpcr(CPUARMState *env, uint32_t val)
268 {
269     vfp_set_fpcr_masked(env, val, MAKE_64BIT_MASK(0, 32));
270 }
271 
272 void HELPER(vfp_set_fpscr)(CPUARMState *env, uint32_t val)
273 {
274     vfp_set_fpcr_masked(env, val, FPSCR_FPCR_MASK);
275     vfp_set_fpsr(env, val & FPSCR_FPSR_MASK);
276 }
277 
278 void vfp_set_fpscr(CPUARMState *env, uint32_t val)
279 {
280     HELPER(vfp_set_fpscr)(env, val);
281 }
282 
283 #ifdef CONFIG_TCG
284 
285 #define VFP_HELPER(name, p) HELPER(glue(glue(vfp_,name),p))
286 
287 #define VFP_BINOP(name) \
288 dh_ctype_f16 VFP_HELPER(name, h)(dh_ctype_f16 a, dh_ctype_f16 b, void *fpstp) \
289 { \
290     float_status *fpst = fpstp; \
291     return float16_ ## name(a, b, fpst); \
292 } \
293 float32 VFP_HELPER(name, s)(float32 a, float32 b, void *fpstp) \
294 { \
295     float_status *fpst = fpstp; \
296     return float32_ ## name(a, b, fpst); \
297 } \
298 float64 VFP_HELPER(name, d)(float64 a, float64 b, void *fpstp) \
299 { \
300     float_status *fpst = fpstp; \
301     return float64_ ## name(a, b, fpst); \
302 }
303 VFP_BINOP(add)
304 VFP_BINOP(sub)
305 VFP_BINOP(mul)
306 VFP_BINOP(div)
307 VFP_BINOP(min)
308 VFP_BINOP(max)
309 VFP_BINOP(minnum)
310 VFP_BINOP(maxnum)
311 #undef VFP_BINOP
312 
313 dh_ctype_f16 VFP_HELPER(sqrt, h)(dh_ctype_f16 a, CPUARMState *env)
314 {
315     return float16_sqrt(a, &env->vfp.fp_status_f16);
316 }
317 
318 float32 VFP_HELPER(sqrt, s)(float32 a, CPUARMState *env)
319 {
320     return float32_sqrt(a, &env->vfp.fp_status);
321 }
322 
323 float64 VFP_HELPER(sqrt, d)(float64 a, CPUARMState *env)
324 {
325     return float64_sqrt(a, &env->vfp.fp_status);
326 }
327 
328 static void softfloat_to_vfp_compare(CPUARMState *env, FloatRelation cmp)
329 {
330     uint32_t flags;
331     switch (cmp) {
332     case float_relation_equal:
333         flags = 0x6;
334         break;
335     case float_relation_less:
336         flags = 0x8;
337         break;
338     case float_relation_greater:
339         flags = 0x2;
340         break;
341     case float_relation_unordered:
342         flags = 0x3;
343         break;
344     default:
345         g_assert_not_reached();
346     }
347     env->vfp.fpsr = deposit64(env->vfp.fpsr, 28, 4, flags); /* NZCV */
348 }
349 
350 /* XXX: check quiet/signaling case */
351 #define DO_VFP_cmp(P, FLOATTYPE, ARGTYPE, FPST) \
352 void VFP_HELPER(cmp, P)(ARGTYPE a, ARGTYPE b, CPUARMState *env)  \
353 { \
354     softfloat_to_vfp_compare(env, \
355         FLOATTYPE ## _compare_quiet(a, b, &env->vfp.FPST)); \
356 } \
357 void VFP_HELPER(cmpe, P)(ARGTYPE a, ARGTYPE b, CPUARMState *env) \
358 { \
359     softfloat_to_vfp_compare(env, \
360         FLOATTYPE ## _compare(a, b, &env->vfp.FPST)); \
361 }
362 DO_VFP_cmp(h, float16, dh_ctype_f16, fp_status_f16)
363 DO_VFP_cmp(s, float32, float32, fp_status)
364 DO_VFP_cmp(d, float64, float64, fp_status)
365 #undef DO_VFP_cmp
366 
367 /* Integer to float and float to integer conversions */
368 
369 #define CONV_ITOF(name, ftype, fsz, sign)                           \
370 ftype HELPER(name)(uint32_t x, void *fpstp)                         \
371 {                                                                   \
372     float_status *fpst = fpstp;                                     \
373     return sign##int32_to_##float##fsz((sign##int32_t)x, fpst);     \
374 }
375 
376 #define CONV_FTOI(name, ftype, fsz, sign, round)                \
377 sign##int32_t HELPER(name)(ftype x, void *fpstp)                \
378 {                                                               \
379     float_status *fpst = fpstp;                                 \
380     if (float##fsz##_is_any_nan(x)) {                           \
381         float_raise(float_flag_invalid, fpst);                  \
382         return 0;                                               \
383     }                                                           \
384     return float##fsz##_to_##sign##int32##round(x, fpst);       \
385 }
386 
387 #define FLOAT_CONVS(name, p, ftype, fsz, sign)            \
388     CONV_ITOF(vfp_##name##to##p, ftype, fsz, sign)        \
389     CONV_FTOI(vfp_to##name##p, ftype, fsz, sign, )        \
390     CONV_FTOI(vfp_to##name##z##p, ftype, fsz, sign, _round_to_zero)
391 
392 FLOAT_CONVS(si, h, uint32_t, 16, )
393 FLOAT_CONVS(si, s, float32, 32, )
394 FLOAT_CONVS(si, d, float64, 64, )
395 FLOAT_CONVS(ui, h, uint32_t, 16, u)
396 FLOAT_CONVS(ui, s, float32, 32, u)
397 FLOAT_CONVS(ui, d, float64, 64, u)
398 
399 #undef CONV_ITOF
400 #undef CONV_FTOI
401 #undef FLOAT_CONVS
402 
403 /* floating point conversion */
404 float64 VFP_HELPER(fcvtd, s)(float32 x, CPUARMState *env)
405 {
406     return float32_to_float64(x, &env->vfp.fp_status);
407 }
408 
409 float32 VFP_HELPER(fcvts, d)(float64 x, CPUARMState *env)
410 {
411     return float64_to_float32(x, &env->vfp.fp_status);
412 }
413 
414 uint32_t HELPER(bfcvt)(float32 x, void *status)
415 {
416     return float32_to_bfloat16(x, status);
417 }
418 
419 uint32_t HELPER(bfcvt_pair)(uint64_t pair, void *status)
420 {
421     bfloat16 lo = float32_to_bfloat16(extract64(pair, 0, 32), status);
422     bfloat16 hi = float32_to_bfloat16(extract64(pair, 32, 32), status);
423     return deposit32(lo, 16, 16, hi);
424 }
425 
426 /*
427  * VFP3 fixed point conversion. The AArch32 versions of fix-to-float
428  * must always round-to-nearest; the AArch64 ones honour the FPSCR
429  * rounding mode. (For AArch32 Neon the standard-FPSCR is set to
430  * round-to-nearest so either helper will work.) AArch32 float-to-fix
431  * must round-to-zero.
432  */
433 #define VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)            \
434 ftype HELPER(vfp_##name##to##p)(uint##isz##_t  x, uint32_t shift,      \
435                                      void *fpstp) \
436 { return itype##_to_##float##fsz##_scalbn(x, -shift, fpstp); }
437 
438 #define VFP_CONV_FIX_FLOAT_ROUND(name, p, fsz, ftype, isz, itype)      \
439     ftype HELPER(vfp_##name##to##p##_round_to_nearest)(uint##isz##_t  x, \
440                                                      uint32_t shift,   \
441                                                      void *fpstp)      \
442     {                                                                  \
443         ftype ret;                                                     \
444         float_status *fpst = fpstp;                                    \
445         FloatRoundMode oldmode = fpst->float_rounding_mode;            \
446         fpst->float_rounding_mode = float_round_nearest_even;          \
447         ret = itype##_to_##float##fsz##_scalbn(x, -shift, fpstp);      \
448         fpst->float_rounding_mode = oldmode;                           \
449         return ret;                                                    \
450     }
451 
452 #define VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype, ROUND, suff) \
453 uint##isz##_t HELPER(vfp_to##name##p##suff)(ftype x, uint32_t shift,      \
454                                             void *fpst)                   \
455 {                                                                         \
456     if (unlikely(float##fsz##_is_any_nan(x))) {                           \
457         float_raise(float_flag_invalid, fpst);                            \
458         return 0;                                                         \
459     }                                                                     \
460     return float##fsz##_to_##itype##_scalbn(x, ROUND, shift, fpst);       \
461 }
462 
463 #define VFP_CONV_FIX(name, p, fsz, ftype, isz, itype)            \
464 VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)              \
465 VFP_CONV_FIX_FLOAT_ROUND(name, p, fsz, ftype, isz, itype)        \
466 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
467                          float_round_to_zero, _round_to_zero)    \
468 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
469                          get_float_rounding_mode(fpst), )
470 
471 #define VFP_CONV_FIX_A64(name, p, fsz, ftype, isz, itype)        \
472 VFP_CONV_FIX_FLOAT(name, p, fsz, ftype, isz, itype)              \
473 VFP_CONV_FLOAT_FIX_ROUND(name, p, fsz, ftype, isz, itype,        \
474                          get_float_rounding_mode(fpst), )
475 
476 VFP_CONV_FIX(sh, d, 64, float64, 64, int16)
477 VFP_CONV_FIX(sl, d, 64, float64, 64, int32)
478 VFP_CONV_FIX_A64(sq, d, 64, float64, 64, int64)
479 VFP_CONV_FIX(uh, d, 64, float64, 64, uint16)
480 VFP_CONV_FIX(ul, d, 64, float64, 64, uint32)
481 VFP_CONV_FIX_A64(uq, d, 64, float64, 64, uint64)
482 VFP_CONV_FIX(sh, s, 32, float32, 32, int16)
483 VFP_CONV_FIX(sl, s, 32, float32, 32, int32)
484 VFP_CONV_FIX_A64(sq, s, 32, float32, 64, int64)
485 VFP_CONV_FIX(uh, s, 32, float32, 32, uint16)
486 VFP_CONV_FIX(ul, s, 32, float32, 32, uint32)
487 VFP_CONV_FIX_A64(uq, s, 32, float32, 64, uint64)
488 VFP_CONV_FIX(sh, h, 16, dh_ctype_f16, 32, int16)
489 VFP_CONV_FIX(sl, h, 16, dh_ctype_f16, 32, int32)
490 VFP_CONV_FIX_A64(sq, h, 16, dh_ctype_f16, 64, int64)
491 VFP_CONV_FIX(uh, h, 16, dh_ctype_f16, 32, uint16)
492 VFP_CONV_FIX(ul, h, 16, dh_ctype_f16, 32, uint32)
493 VFP_CONV_FIX_A64(uq, h, 16, dh_ctype_f16, 64, uint64)
494 
495 #undef VFP_CONV_FIX
496 #undef VFP_CONV_FIX_FLOAT
497 #undef VFP_CONV_FLOAT_FIX_ROUND
498 #undef VFP_CONV_FIX_A64
499 
500 /* Set the current fp rounding mode and return the old one.
501  * The argument is a softfloat float_round_ value.
502  */
503 uint32_t HELPER(set_rmode)(uint32_t rmode, void *fpstp)
504 {
505     float_status *fp_status = fpstp;
506 
507     uint32_t prev_rmode = get_float_rounding_mode(fp_status);
508     set_float_rounding_mode(rmode, fp_status);
509 
510     return prev_rmode;
511 }
512 
513 /* Half precision conversions.  */
514 float32 HELPER(vfp_fcvt_f16_to_f32)(uint32_t a, void *fpstp, uint32_t ahp_mode)
515 {
516     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
517      * it would affect flushing input denormals.
518      */
519     float_status *fpst = fpstp;
520     bool save = get_flush_inputs_to_zero(fpst);
521     set_flush_inputs_to_zero(false, fpst);
522     float32 r = float16_to_float32(a, !ahp_mode, fpst);
523     set_flush_inputs_to_zero(save, fpst);
524     return r;
525 }
526 
527 uint32_t HELPER(vfp_fcvt_f32_to_f16)(float32 a, void *fpstp, uint32_t ahp_mode)
528 {
529     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
530      * it would affect flushing output denormals.
531      */
532     float_status *fpst = fpstp;
533     bool save = get_flush_to_zero(fpst);
534     set_flush_to_zero(false, fpst);
535     float16 r = float32_to_float16(a, !ahp_mode, fpst);
536     set_flush_to_zero(save, fpst);
537     return r;
538 }
539 
540 float64 HELPER(vfp_fcvt_f16_to_f64)(uint32_t a, void *fpstp, uint32_t ahp_mode)
541 {
542     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
543      * it would affect flushing input denormals.
544      */
545     float_status *fpst = fpstp;
546     bool save = get_flush_inputs_to_zero(fpst);
547     set_flush_inputs_to_zero(false, fpst);
548     float64 r = float16_to_float64(a, !ahp_mode, fpst);
549     set_flush_inputs_to_zero(save, fpst);
550     return r;
551 }
552 
553 uint32_t HELPER(vfp_fcvt_f64_to_f16)(float64 a, void *fpstp, uint32_t ahp_mode)
554 {
555     /* Squash FZ16 to 0 for the duration of conversion.  In this case,
556      * it would affect flushing output denormals.
557      */
558     float_status *fpst = fpstp;
559     bool save = get_flush_to_zero(fpst);
560     set_flush_to_zero(false, fpst);
561     float16 r = float64_to_float16(a, !ahp_mode, fpst);
562     set_flush_to_zero(save, fpst);
563     return r;
564 }
565 
566 /* NEON helpers.  */
567 
568 /* Constants 256 and 512 are used in some helpers; we avoid relying on
569  * int->float conversions at run-time.  */
570 #define float64_256 make_float64(0x4070000000000000LL)
571 #define float64_512 make_float64(0x4080000000000000LL)
572 #define float16_maxnorm make_float16(0x7bff)
573 #define float32_maxnorm make_float32(0x7f7fffff)
574 #define float64_maxnorm make_float64(0x7fefffffffffffffLL)
575 
576 /* Reciprocal functions
577  *
578  * The algorithm that must be used to calculate the estimate
579  * is specified by the ARM ARM, see FPRecipEstimate()/RecipEstimate
580  */
581 
582 /* See RecipEstimate()
583  *
584  * input is a 9 bit fixed point number
585  * input range 256 .. 511 for a number from 0.5 <= x < 1.0.
586  * result range 256 .. 511 for a number from 1.0 to 511/256.
587  */
588 
589 static int recip_estimate(int input)
590 {
591     int a, b, r;
592     assert(256 <= input && input < 512);
593     a = (input * 2) + 1;
594     b = (1 << 19) / a;
595     r = (b + 1) >> 1;
596     assert(256 <= r && r < 512);
597     return r;
598 }
599 
600 /*
601  * Common wrapper to call recip_estimate
602  *
603  * The parameters are exponent and 64 bit fraction (without implicit
604  * bit) where the binary point is nominally at bit 52. Returns a
605  * float64 which can then be rounded to the appropriate size by the
606  * callee.
607  */
608 
609 static uint64_t call_recip_estimate(int *exp, int exp_off, uint64_t frac)
610 {
611     uint32_t scaled, estimate;
612     uint64_t result_frac;
613     int result_exp;
614 
615     /* Handle sub-normals */
616     if (*exp == 0) {
617         if (extract64(frac, 51, 1) == 0) {
618             *exp = -1;
619             frac <<= 2;
620         } else {
621             frac <<= 1;
622         }
623     }
624 
625     /* scaled = UInt('1':fraction<51:44>) */
626     scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
627     estimate = recip_estimate(scaled);
628 
629     result_exp = exp_off - *exp;
630     result_frac = deposit64(0, 44, 8, estimate);
631     if (result_exp == 0) {
632         result_frac = deposit64(result_frac >> 1, 51, 1, 1);
633     } else if (result_exp == -1) {
634         result_frac = deposit64(result_frac >> 2, 50, 2, 1);
635         result_exp = 0;
636     }
637 
638     *exp = result_exp;
639 
640     return result_frac;
641 }
642 
643 static bool round_to_inf(float_status *fpst, bool sign_bit)
644 {
645     switch (fpst->float_rounding_mode) {
646     case float_round_nearest_even: /* Round to Nearest */
647         return true;
648     case float_round_up: /* Round to +Inf */
649         return !sign_bit;
650     case float_round_down: /* Round to -Inf */
651         return sign_bit;
652     case float_round_to_zero: /* Round to Zero */
653         return false;
654     default:
655         g_assert_not_reached();
656     }
657 }
658 
659 uint32_t HELPER(recpe_f16)(uint32_t input, void *fpstp)
660 {
661     float_status *fpst = fpstp;
662     float16 f16 = float16_squash_input_denormal(input, fpst);
663     uint32_t f16_val = float16_val(f16);
664     uint32_t f16_sign = float16_is_neg(f16);
665     int f16_exp = extract32(f16_val, 10, 5);
666     uint32_t f16_frac = extract32(f16_val, 0, 10);
667     uint64_t f64_frac;
668 
669     if (float16_is_any_nan(f16)) {
670         float16 nan = f16;
671         if (float16_is_signaling_nan(f16, fpst)) {
672             float_raise(float_flag_invalid, fpst);
673             if (!fpst->default_nan_mode) {
674                 nan = float16_silence_nan(f16, fpst);
675             }
676         }
677         if (fpst->default_nan_mode) {
678             nan =  float16_default_nan(fpst);
679         }
680         return nan;
681     } else if (float16_is_infinity(f16)) {
682         return float16_set_sign(float16_zero, float16_is_neg(f16));
683     } else if (float16_is_zero(f16)) {
684         float_raise(float_flag_divbyzero, fpst);
685         return float16_set_sign(float16_infinity, float16_is_neg(f16));
686     } else if (float16_abs(f16) < (1 << 8)) {
687         /* Abs(value) < 2.0^-16 */
688         float_raise(float_flag_overflow | float_flag_inexact, fpst);
689         if (round_to_inf(fpst, f16_sign)) {
690             return float16_set_sign(float16_infinity, f16_sign);
691         } else {
692             return float16_set_sign(float16_maxnorm, f16_sign);
693         }
694     } else if (f16_exp >= 29 && fpst->flush_to_zero) {
695         float_raise(float_flag_underflow, fpst);
696         return float16_set_sign(float16_zero, float16_is_neg(f16));
697     }
698 
699     f64_frac = call_recip_estimate(&f16_exp, 29,
700                                    ((uint64_t) f16_frac) << (52 - 10));
701 
702     /* result = sign : result_exp<4:0> : fraction<51:42> */
703     f16_val = deposit32(0, 15, 1, f16_sign);
704     f16_val = deposit32(f16_val, 10, 5, f16_exp);
705     f16_val = deposit32(f16_val, 0, 10, extract64(f64_frac, 52 - 10, 10));
706     return make_float16(f16_val);
707 }
708 
709 float32 HELPER(recpe_f32)(float32 input, void *fpstp)
710 {
711     float_status *fpst = fpstp;
712     float32 f32 = float32_squash_input_denormal(input, fpst);
713     uint32_t f32_val = float32_val(f32);
714     bool f32_sign = float32_is_neg(f32);
715     int f32_exp = extract32(f32_val, 23, 8);
716     uint32_t f32_frac = extract32(f32_val, 0, 23);
717     uint64_t f64_frac;
718 
719     if (float32_is_any_nan(f32)) {
720         float32 nan = f32;
721         if (float32_is_signaling_nan(f32, fpst)) {
722             float_raise(float_flag_invalid, fpst);
723             if (!fpst->default_nan_mode) {
724                 nan = float32_silence_nan(f32, fpst);
725             }
726         }
727         if (fpst->default_nan_mode) {
728             nan =  float32_default_nan(fpst);
729         }
730         return nan;
731     } else if (float32_is_infinity(f32)) {
732         return float32_set_sign(float32_zero, float32_is_neg(f32));
733     } else if (float32_is_zero(f32)) {
734         float_raise(float_flag_divbyzero, fpst);
735         return float32_set_sign(float32_infinity, float32_is_neg(f32));
736     } else if (float32_abs(f32) < (1ULL << 21)) {
737         /* Abs(value) < 2.0^-128 */
738         float_raise(float_flag_overflow | float_flag_inexact, fpst);
739         if (round_to_inf(fpst, f32_sign)) {
740             return float32_set_sign(float32_infinity, f32_sign);
741         } else {
742             return float32_set_sign(float32_maxnorm, f32_sign);
743         }
744     } else if (f32_exp >= 253 && fpst->flush_to_zero) {
745         float_raise(float_flag_underflow, fpst);
746         return float32_set_sign(float32_zero, float32_is_neg(f32));
747     }
748 
749     f64_frac = call_recip_estimate(&f32_exp, 253,
750                                    ((uint64_t) f32_frac) << (52 - 23));
751 
752     /* result = sign : result_exp<7:0> : fraction<51:29> */
753     f32_val = deposit32(0, 31, 1, f32_sign);
754     f32_val = deposit32(f32_val, 23, 8, f32_exp);
755     f32_val = deposit32(f32_val, 0, 23, extract64(f64_frac, 52 - 23, 23));
756     return make_float32(f32_val);
757 }
758 
759 float64 HELPER(recpe_f64)(float64 input, void *fpstp)
760 {
761     float_status *fpst = fpstp;
762     float64 f64 = float64_squash_input_denormal(input, fpst);
763     uint64_t f64_val = float64_val(f64);
764     bool f64_sign = float64_is_neg(f64);
765     int f64_exp = extract64(f64_val, 52, 11);
766     uint64_t f64_frac = extract64(f64_val, 0, 52);
767 
768     /* Deal with any special cases */
769     if (float64_is_any_nan(f64)) {
770         float64 nan = f64;
771         if (float64_is_signaling_nan(f64, fpst)) {
772             float_raise(float_flag_invalid, fpst);
773             if (!fpst->default_nan_mode) {
774                 nan = float64_silence_nan(f64, fpst);
775             }
776         }
777         if (fpst->default_nan_mode) {
778             nan =  float64_default_nan(fpst);
779         }
780         return nan;
781     } else if (float64_is_infinity(f64)) {
782         return float64_set_sign(float64_zero, float64_is_neg(f64));
783     } else if (float64_is_zero(f64)) {
784         float_raise(float_flag_divbyzero, fpst);
785         return float64_set_sign(float64_infinity, float64_is_neg(f64));
786     } else if ((f64_val & ~(1ULL << 63)) < (1ULL << 50)) {
787         /* Abs(value) < 2.0^-1024 */
788         float_raise(float_flag_overflow | float_flag_inexact, fpst);
789         if (round_to_inf(fpst, f64_sign)) {
790             return float64_set_sign(float64_infinity, f64_sign);
791         } else {
792             return float64_set_sign(float64_maxnorm, f64_sign);
793         }
794     } else if (f64_exp >= 2045 && fpst->flush_to_zero) {
795         float_raise(float_flag_underflow, fpst);
796         return float64_set_sign(float64_zero, float64_is_neg(f64));
797     }
798 
799     f64_frac = call_recip_estimate(&f64_exp, 2045, f64_frac);
800 
801     /* result = sign : result_exp<10:0> : fraction<51:0>; */
802     f64_val = deposit64(0, 63, 1, f64_sign);
803     f64_val = deposit64(f64_val, 52, 11, f64_exp);
804     f64_val = deposit64(f64_val, 0, 52, f64_frac);
805     return make_float64(f64_val);
806 }
807 
808 /* The algorithm that must be used to calculate the estimate
809  * is specified by the ARM ARM.
810  */
811 
812 static int do_recip_sqrt_estimate(int a)
813 {
814     int b, estimate;
815 
816     assert(128 <= a && a < 512);
817     if (a < 256) {
818         a = a * 2 + 1;
819     } else {
820         a = (a >> 1) << 1;
821         a = (a + 1) * 2;
822     }
823     b = 512;
824     while (a * (b + 1) * (b + 1) < (1 << 28)) {
825         b += 1;
826     }
827     estimate = (b + 1) / 2;
828     assert(256 <= estimate && estimate < 512);
829 
830     return estimate;
831 }
832 
833 
834 static uint64_t recip_sqrt_estimate(int *exp , int exp_off, uint64_t frac)
835 {
836     int estimate;
837     uint32_t scaled;
838 
839     if (*exp == 0) {
840         while (extract64(frac, 51, 1) == 0) {
841             frac = frac << 1;
842             *exp -= 1;
843         }
844         frac = extract64(frac, 0, 51) << 1;
845     }
846 
847     if (*exp & 1) {
848         /* scaled = UInt('01':fraction<51:45>) */
849         scaled = deposit32(1 << 7, 0, 7, extract64(frac, 45, 7));
850     } else {
851         /* scaled = UInt('1':fraction<51:44>) */
852         scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
853     }
854     estimate = do_recip_sqrt_estimate(scaled);
855 
856     *exp = (exp_off - *exp) / 2;
857     return extract64(estimate, 0, 8) << 44;
858 }
859 
860 uint32_t HELPER(rsqrte_f16)(uint32_t input, void *fpstp)
861 {
862     float_status *s = fpstp;
863     float16 f16 = float16_squash_input_denormal(input, s);
864     uint16_t val = float16_val(f16);
865     bool f16_sign = float16_is_neg(f16);
866     int f16_exp = extract32(val, 10, 5);
867     uint16_t f16_frac = extract32(val, 0, 10);
868     uint64_t f64_frac;
869 
870     if (float16_is_any_nan(f16)) {
871         float16 nan = f16;
872         if (float16_is_signaling_nan(f16, s)) {
873             float_raise(float_flag_invalid, s);
874             if (!s->default_nan_mode) {
875                 nan = float16_silence_nan(f16, fpstp);
876             }
877         }
878         if (s->default_nan_mode) {
879             nan =  float16_default_nan(s);
880         }
881         return nan;
882     } else if (float16_is_zero(f16)) {
883         float_raise(float_flag_divbyzero, s);
884         return float16_set_sign(float16_infinity, f16_sign);
885     } else if (f16_sign) {
886         float_raise(float_flag_invalid, s);
887         return float16_default_nan(s);
888     } else if (float16_is_infinity(f16)) {
889         return float16_zero;
890     }
891 
892     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
893      * preserving the parity of the exponent.  */
894 
895     f64_frac = ((uint64_t) f16_frac) << (52 - 10);
896 
897     f64_frac = recip_sqrt_estimate(&f16_exp, 44, f64_frac);
898 
899     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(2) */
900     val = deposit32(0, 15, 1, f16_sign);
901     val = deposit32(val, 10, 5, f16_exp);
902     val = deposit32(val, 2, 8, extract64(f64_frac, 52 - 8, 8));
903     return make_float16(val);
904 }
905 
906 float32 HELPER(rsqrte_f32)(float32 input, void *fpstp)
907 {
908     float_status *s = fpstp;
909     float32 f32 = float32_squash_input_denormal(input, s);
910     uint32_t val = float32_val(f32);
911     uint32_t f32_sign = float32_is_neg(f32);
912     int f32_exp = extract32(val, 23, 8);
913     uint32_t f32_frac = extract32(val, 0, 23);
914     uint64_t f64_frac;
915 
916     if (float32_is_any_nan(f32)) {
917         float32 nan = f32;
918         if (float32_is_signaling_nan(f32, s)) {
919             float_raise(float_flag_invalid, s);
920             if (!s->default_nan_mode) {
921                 nan = float32_silence_nan(f32, fpstp);
922             }
923         }
924         if (s->default_nan_mode) {
925             nan =  float32_default_nan(s);
926         }
927         return nan;
928     } else if (float32_is_zero(f32)) {
929         float_raise(float_flag_divbyzero, s);
930         return float32_set_sign(float32_infinity, float32_is_neg(f32));
931     } else if (float32_is_neg(f32)) {
932         float_raise(float_flag_invalid, s);
933         return float32_default_nan(s);
934     } else if (float32_is_infinity(f32)) {
935         return float32_zero;
936     }
937 
938     /* Scale and normalize to a double-precision value between 0.25 and 1.0,
939      * preserving the parity of the exponent.  */
940 
941     f64_frac = ((uint64_t) f32_frac) << 29;
942 
943     f64_frac = recip_sqrt_estimate(&f32_exp, 380, f64_frac);
944 
945     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(15) */
946     val = deposit32(0, 31, 1, f32_sign);
947     val = deposit32(val, 23, 8, f32_exp);
948     val = deposit32(val, 15, 8, extract64(f64_frac, 52 - 8, 8));
949     return make_float32(val);
950 }
951 
952 float64 HELPER(rsqrte_f64)(float64 input, void *fpstp)
953 {
954     float_status *s = fpstp;
955     float64 f64 = float64_squash_input_denormal(input, s);
956     uint64_t val = float64_val(f64);
957     bool f64_sign = float64_is_neg(f64);
958     int f64_exp = extract64(val, 52, 11);
959     uint64_t f64_frac = extract64(val, 0, 52);
960 
961     if (float64_is_any_nan(f64)) {
962         float64 nan = f64;
963         if (float64_is_signaling_nan(f64, s)) {
964             float_raise(float_flag_invalid, s);
965             if (!s->default_nan_mode) {
966                 nan = float64_silence_nan(f64, fpstp);
967             }
968         }
969         if (s->default_nan_mode) {
970             nan =  float64_default_nan(s);
971         }
972         return nan;
973     } else if (float64_is_zero(f64)) {
974         float_raise(float_flag_divbyzero, s);
975         return float64_set_sign(float64_infinity, float64_is_neg(f64));
976     } else if (float64_is_neg(f64)) {
977         float_raise(float_flag_invalid, s);
978         return float64_default_nan(s);
979     } else if (float64_is_infinity(f64)) {
980         return float64_zero;
981     }
982 
983     f64_frac = recip_sqrt_estimate(&f64_exp, 3068, f64_frac);
984 
985     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(44) */
986     val = deposit64(0, 61, 1, f64_sign);
987     val = deposit64(val, 52, 11, f64_exp);
988     val = deposit64(val, 44, 8, extract64(f64_frac, 52 - 8, 8));
989     return make_float64(val);
990 }
991 
992 uint32_t HELPER(recpe_u32)(uint32_t a)
993 {
994     int input, estimate;
995 
996     if ((a & 0x80000000) == 0) {
997         return 0xffffffff;
998     }
999 
1000     input = extract32(a, 23, 9);
1001     estimate = recip_estimate(input);
1002 
1003     return deposit32(0, (32 - 9), 9, estimate);
1004 }
1005 
1006 uint32_t HELPER(rsqrte_u32)(uint32_t a)
1007 {
1008     int estimate;
1009 
1010     if ((a & 0xc0000000) == 0) {
1011         return 0xffffffff;
1012     }
1013 
1014     estimate = do_recip_sqrt_estimate(extract32(a, 23, 9));
1015 
1016     return deposit32(0, 23, 9, estimate);
1017 }
1018 
1019 /* VFPv4 fused multiply-accumulate */
1020 dh_ctype_f16 VFP_HELPER(muladd, h)(dh_ctype_f16 a, dh_ctype_f16 b,
1021                                    dh_ctype_f16 c, void *fpstp)
1022 {
1023     float_status *fpst = fpstp;
1024     return float16_muladd(a, b, c, 0, fpst);
1025 }
1026 
1027 float32 VFP_HELPER(muladd, s)(float32 a, float32 b, float32 c, void *fpstp)
1028 {
1029     float_status *fpst = fpstp;
1030     return float32_muladd(a, b, c, 0, fpst);
1031 }
1032 
1033 float64 VFP_HELPER(muladd, d)(float64 a, float64 b, float64 c, void *fpstp)
1034 {
1035     float_status *fpst = fpstp;
1036     return float64_muladd(a, b, c, 0, fpst);
1037 }
1038 
1039 /* ARMv8 round to integral */
1040 dh_ctype_f16 HELPER(rinth_exact)(dh_ctype_f16 x, void *fp_status)
1041 {
1042     return float16_round_to_int(x, fp_status);
1043 }
1044 
1045 float32 HELPER(rints_exact)(float32 x, void *fp_status)
1046 {
1047     return float32_round_to_int(x, fp_status);
1048 }
1049 
1050 float64 HELPER(rintd_exact)(float64 x, void *fp_status)
1051 {
1052     return float64_round_to_int(x, fp_status);
1053 }
1054 
1055 dh_ctype_f16 HELPER(rinth)(dh_ctype_f16 x, void *fp_status)
1056 {
1057     int old_flags = get_float_exception_flags(fp_status), new_flags;
1058     float16 ret;
1059 
1060     ret = float16_round_to_int(x, fp_status);
1061 
1062     /* Suppress any inexact exceptions the conversion produced */
1063     if (!(old_flags & float_flag_inexact)) {
1064         new_flags = get_float_exception_flags(fp_status);
1065         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1066     }
1067 
1068     return ret;
1069 }
1070 
1071 float32 HELPER(rints)(float32 x, void *fp_status)
1072 {
1073     int old_flags = get_float_exception_flags(fp_status), new_flags;
1074     float32 ret;
1075 
1076     ret = float32_round_to_int(x, fp_status);
1077 
1078     /* Suppress any inexact exceptions the conversion produced */
1079     if (!(old_flags & float_flag_inexact)) {
1080         new_flags = get_float_exception_flags(fp_status);
1081         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1082     }
1083 
1084     return ret;
1085 }
1086 
1087 float64 HELPER(rintd)(float64 x, void *fp_status)
1088 {
1089     int old_flags = get_float_exception_flags(fp_status), new_flags;
1090     float64 ret;
1091 
1092     ret = float64_round_to_int(x, fp_status);
1093 
1094     new_flags = get_float_exception_flags(fp_status);
1095 
1096     /* Suppress any inexact exceptions the conversion produced */
1097     if (!(old_flags & float_flag_inexact)) {
1098         new_flags = get_float_exception_flags(fp_status);
1099         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
1100     }
1101 
1102     return ret;
1103 }
1104 
1105 /* Convert ARM rounding mode to softfloat */
1106 const FloatRoundMode arm_rmode_to_sf_map[] = {
1107     [FPROUNDING_TIEEVEN] = float_round_nearest_even,
1108     [FPROUNDING_POSINF] = float_round_up,
1109     [FPROUNDING_NEGINF] = float_round_down,
1110     [FPROUNDING_ZERO] = float_round_to_zero,
1111     [FPROUNDING_TIEAWAY] = float_round_ties_away,
1112     [FPROUNDING_ODD] = float_round_to_odd,
1113 };
1114 
1115 /*
1116  * Implement float64 to int32_t conversion without saturation;
1117  * the result is supplied modulo 2^32.
1118  */
1119 uint64_t HELPER(fjcvtzs)(float64 value, void *vstatus)
1120 {
1121     float_status *status = vstatus;
1122     uint32_t frac, e_old, e_new;
1123     bool inexact;
1124 
1125     e_old = get_float_exception_flags(status);
1126     set_float_exception_flags(0, status);
1127     frac = float64_to_int32_modulo(value, float_round_to_zero, status);
1128     e_new = get_float_exception_flags(status);
1129     set_float_exception_flags(e_old | e_new, status);
1130 
1131     /* Normal inexact, denormal with flush-to-zero, or overflow or NaN */
1132     inexact = e_new & (float_flag_inexact |
1133                        float_flag_input_denormal |
1134                        float_flag_invalid);
1135 
1136     /* While not inexact for IEEE FP, -0.0 is inexact for JavaScript. */
1137     inexact |= value == float64_chs(float64_zero);
1138 
1139     /* Pack the result and the env->ZF representation of Z together.  */
1140     return deposit64(frac, 32, 32, inexact);
1141 }
1142 
1143 uint32_t HELPER(vjcvt)(float64 value, CPUARMState *env)
1144 {
1145     uint64_t pair = HELPER(fjcvtzs)(value, &env->vfp.fp_status);
1146     uint32_t result = pair;
1147     uint32_t z = (pair >> 32) == 0;
1148 
1149     /* Store Z, clear NCV, in FPSCR.NZCV.  */
1150     env->vfp.fpsr = (env->vfp.fpsr & ~FPSR_NZCV_MASK) | (z * FPSR_Z);
1151 
1152     return result;
1153 }
1154 
1155 /* Round a float32 to an integer that fits in int32_t or int64_t.  */
1156 static float32 frint_s(float32 f, float_status *fpst, int intsize)
1157 {
1158     int old_flags = get_float_exception_flags(fpst);
1159     uint32_t exp = extract32(f, 23, 8);
1160 
1161     if (unlikely(exp == 0xff)) {
1162         /* NaN or Inf.  */
1163         goto overflow;
1164     }
1165 
1166     /* Round and re-extract the exponent.  */
1167     f = float32_round_to_int(f, fpst);
1168     exp = extract32(f, 23, 8);
1169 
1170     /* Validate the range of the result.  */
1171     if (exp < 126 + intsize) {
1172         /* abs(F) <= INT{N}_MAX */
1173         return f;
1174     }
1175     if (exp == 126 + intsize) {
1176         uint32_t sign = extract32(f, 31, 1);
1177         uint32_t frac = extract32(f, 0, 23);
1178         if (sign && frac == 0) {
1179             /* F == INT{N}_MIN */
1180             return f;
1181         }
1182     }
1183 
1184  overflow:
1185     /*
1186      * Raise Invalid and return INT{N}_MIN as a float.  Revert any
1187      * inexact exception float32_round_to_int may have raised.
1188      */
1189     set_float_exception_flags(old_flags | float_flag_invalid, fpst);
1190     return (0x100u + 126u + intsize) << 23;
1191 }
1192 
1193 float32 HELPER(frint32_s)(float32 f, void *fpst)
1194 {
1195     return frint_s(f, fpst, 32);
1196 }
1197 
1198 float32 HELPER(frint64_s)(float32 f, void *fpst)
1199 {
1200     return frint_s(f, fpst, 64);
1201 }
1202 
1203 /* Round a float64 to an integer that fits in int32_t or int64_t.  */
1204 static float64 frint_d(float64 f, float_status *fpst, int intsize)
1205 {
1206     int old_flags = get_float_exception_flags(fpst);
1207     uint32_t exp = extract64(f, 52, 11);
1208 
1209     if (unlikely(exp == 0x7ff)) {
1210         /* NaN or Inf.  */
1211         goto overflow;
1212     }
1213 
1214     /* Round and re-extract the exponent.  */
1215     f = float64_round_to_int(f, fpst);
1216     exp = extract64(f, 52, 11);
1217 
1218     /* Validate the range of the result.  */
1219     if (exp < 1022 + intsize) {
1220         /* abs(F) <= INT{N}_MAX */
1221         return f;
1222     }
1223     if (exp == 1022 + intsize) {
1224         uint64_t sign = extract64(f, 63, 1);
1225         uint64_t frac = extract64(f, 0, 52);
1226         if (sign && frac == 0) {
1227             /* F == INT{N}_MIN */
1228             return f;
1229         }
1230     }
1231 
1232  overflow:
1233     /*
1234      * Raise Invalid and return INT{N}_MIN as a float.  Revert any
1235      * inexact exception float64_round_to_int may have raised.
1236      */
1237     set_float_exception_flags(old_flags | float_flag_invalid, fpst);
1238     return (uint64_t)(0x800 + 1022 + intsize) << 52;
1239 }
1240 
1241 float64 HELPER(frint32_d)(float64 f, void *fpst)
1242 {
1243     return frint_d(f, fpst, 32);
1244 }
1245 
1246 float64 HELPER(frint64_d)(float64 f, void *fpst)
1247 {
1248     return frint_d(f, fpst, 64);
1249 }
1250 
1251 void HELPER(check_hcr_el2_trap)(CPUARMState *env, uint32_t rt, uint32_t reg)
1252 {
1253     uint32_t syndrome;
1254 
1255     switch (reg) {
1256     case ARM_VFP_MVFR0:
1257     case ARM_VFP_MVFR1:
1258     case ARM_VFP_MVFR2:
1259         if (!(arm_hcr_el2_eff(env) & HCR_TID3)) {
1260             return;
1261         }
1262         break;
1263     case ARM_VFP_FPSID:
1264         if (!(arm_hcr_el2_eff(env) & HCR_TID0)) {
1265             return;
1266         }
1267         break;
1268     default:
1269         g_assert_not_reached();
1270     }
1271 
1272     syndrome = ((EC_FPIDTRAP << ARM_EL_EC_SHIFT)
1273                 | ARM_EL_IL
1274                 | (1 << 24) | (0xe << 20) | (7 << 14)
1275                 | (reg << 10) | (rt << 5) | 1);
1276 
1277     raise_exception(env, EXCP_HYP_TRAP, syndrome, 2);
1278 }
1279 
1280 #endif
1281