xref: /openbmc/qemu/target/arm/tcg/helper-a64.c (revision a3ef070e)
1 /*
2  *  AArch64 specific helpers
3  *
4  *  Copyright (c) 2013 Alexander Graf <agraf@suse.de>
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 "qemu/units.h"
22 #include "cpu.h"
23 #include "exec/gdbstub.h"
24 #include "exec/helper-proto.h"
25 #include "qemu/host-utils.h"
26 #include "qemu/log.h"
27 #include "qemu/main-loop.h"
28 #include "qemu/bitops.h"
29 #include "internals.h"
30 #include "qemu/crc32c.h"
31 #include "exec/exec-all.h"
32 #include "exec/cpu_ldst.h"
33 #include "qemu/int128.h"
34 #include "qemu/atomic128.h"
35 #include "fpu/softfloat.h"
36 #include <zlib.h> /* For crc32 */
37 
38 /* C2.4.7 Multiply and divide */
39 /* special cases for 0 and LLONG_MIN are mandated by the standard */
40 uint64_t HELPER(udiv64)(uint64_t num, uint64_t den)
41 {
42     if (den == 0) {
43         return 0;
44     }
45     return num / den;
46 }
47 
48 int64_t HELPER(sdiv64)(int64_t num, int64_t den)
49 {
50     if (den == 0) {
51         return 0;
52     }
53     if (num == LLONG_MIN && den == -1) {
54         return LLONG_MIN;
55     }
56     return num / den;
57 }
58 
59 uint64_t HELPER(rbit64)(uint64_t x)
60 {
61     return revbit64(x);
62 }
63 
64 void HELPER(msr_i_spsel)(CPUARMState *env, uint32_t imm)
65 {
66     update_spsel(env, imm);
67 }
68 
69 static void daif_check(CPUARMState *env, uint32_t op,
70                        uint32_t imm, uintptr_t ra)
71 {
72     /* DAIF update to PSTATE. This is OK from EL0 only if UMA is set.  */
73     if (arm_current_el(env) == 0 && !(arm_sctlr(env, 0) & SCTLR_UMA)) {
74         raise_exception_ra(env, EXCP_UDEF,
75                            syn_aa64_sysregtrap(0, extract32(op, 0, 3),
76                                                extract32(op, 3, 3), 4,
77                                                imm, 0x1f, 0),
78                            exception_target_el(env), ra);
79     }
80 }
81 
82 void HELPER(msr_i_daifset)(CPUARMState *env, uint32_t imm)
83 {
84     daif_check(env, 0x1e, imm, GETPC());
85     env->daif |= (imm << 6) & PSTATE_DAIF;
86     arm_rebuild_hflags(env);
87 }
88 
89 void HELPER(msr_i_daifclear)(CPUARMState *env, uint32_t imm)
90 {
91     daif_check(env, 0x1f, imm, GETPC());
92     env->daif &= ~((imm << 6) & PSTATE_DAIF);
93     arm_rebuild_hflags(env);
94 }
95 
96 /* Convert a softfloat float_relation_ (as returned by
97  * the float*_compare functions) to the correct ARM
98  * NZCV flag state.
99  */
100 static inline uint32_t float_rel_to_flags(int res)
101 {
102     uint64_t flags;
103     switch (res) {
104     case float_relation_equal:
105         flags = PSTATE_Z | PSTATE_C;
106         break;
107     case float_relation_less:
108         flags = PSTATE_N;
109         break;
110     case float_relation_greater:
111         flags = PSTATE_C;
112         break;
113     case float_relation_unordered:
114     default:
115         flags = PSTATE_C | PSTATE_V;
116         break;
117     }
118     return flags;
119 }
120 
121 uint64_t HELPER(vfp_cmph_a64)(uint32_t x, uint32_t y, void *fp_status)
122 {
123     return float_rel_to_flags(float16_compare_quiet(x, y, fp_status));
124 }
125 
126 uint64_t HELPER(vfp_cmpeh_a64)(uint32_t x, uint32_t y, void *fp_status)
127 {
128     return float_rel_to_flags(float16_compare(x, y, fp_status));
129 }
130 
131 uint64_t HELPER(vfp_cmps_a64)(float32 x, float32 y, void *fp_status)
132 {
133     return float_rel_to_flags(float32_compare_quiet(x, y, fp_status));
134 }
135 
136 uint64_t HELPER(vfp_cmpes_a64)(float32 x, float32 y, void *fp_status)
137 {
138     return float_rel_to_flags(float32_compare(x, y, fp_status));
139 }
140 
141 uint64_t HELPER(vfp_cmpd_a64)(float64 x, float64 y, void *fp_status)
142 {
143     return float_rel_to_flags(float64_compare_quiet(x, y, fp_status));
144 }
145 
146 uint64_t HELPER(vfp_cmped_a64)(float64 x, float64 y, void *fp_status)
147 {
148     return float_rel_to_flags(float64_compare(x, y, fp_status));
149 }
150 
151 float32 HELPER(vfp_mulxs)(float32 a, float32 b, void *fpstp)
152 {
153     float_status *fpst = fpstp;
154 
155     a = float32_squash_input_denormal(a, fpst);
156     b = float32_squash_input_denormal(b, fpst);
157 
158     if ((float32_is_zero(a) && float32_is_infinity(b)) ||
159         (float32_is_infinity(a) && float32_is_zero(b))) {
160         /* 2.0 with the sign bit set to sign(A) XOR sign(B) */
161         return make_float32((1U << 30) |
162                             ((float32_val(a) ^ float32_val(b)) & (1U << 31)));
163     }
164     return float32_mul(a, b, fpst);
165 }
166 
167 float64 HELPER(vfp_mulxd)(float64 a, float64 b, void *fpstp)
168 {
169     float_status *fpst = fpstp;
170 
171     a = float64_squash_input_denormal(a, fpst);
172     b = float64_squash_input_denormal(b, fpst);
173 
174     if ((float64_is_zero(a) && float64_is_infinity(b)) ||
175         (float64_is_infinity(a) && float64_is_zero(b))) {
176         /* 2.0 with the sign bit set to sign(A) XOR sign(B) */
177         return make_float64((1ULL << 62) |
178                             ((float64_val(a) ^ float64_val(b)) & (1ULL << 63)));
179     }
180     return float64_mul(a, b, fpst);
181 }
182 
183 /* 64bit/double versions of the neon float compare functions */
184 uint64_t HELPER(neon_ceq_f64)(float64 a, float64 b, void *fpstp)
185 {
186     float_status *fpst = fpstp;
187     return -float64_eq_quiet(a, b, fpst);
188 }
189 
190 uint64_t HELPER(neon_cge_f64)(float64 a, float64 b, void *fpstp)
191 {
192     float_status *fpst = fpstp;
193     return -float64_le(b, a, fpst);
194 }
195 
196 uint64_t HELPER(neon_cgt_f64)(float64 a, float64 b, void *fpstp)
197 {
198     float_status *fpst = fpstp;
199     return -float64_lt(b, a, fpst);
200 }
201 
202 /* Reciprocal step and sqrt step. Note that unlike the A32/T32
203  * versions, these do a fully fused multiply-add or
204  * multiply-add-and-halve.
205  */
206 
207 uint32_t HELPER(recpsf_f16)(uint32_t a, uint32_t b, void *fpstp)
208 {
209     float_status *fpst = fpstp;
210 
211     a = float16_squash_input_denormal(a, fpst);
212     b = float16_squash_input_denormal(b, fpst);
213 
214     a = float16_chs(a);
215     if ((float16_is_infinity(a) && float16_is_zero(b)) ||
216         (float16_is_infinity(b) && float16_is_zero(a))) {
217         return float16_two;
218     }
219     return float16_muladd(a, b, float16_two, 0, fpst);
220 }
221 
222 float32 HELPER(recpsf_f32)(float32 a, float32 b, void *fpstp)
223 {
224     float_status *fpst = fpstp;
225 
226     a = float32_squash_input_denormal(a, fpst);
227     b = float32_squash_input_denormal(b, fpst);
228 
229     a = float32_chs(a);
230     if ((float32_is_infinity(a) && float32_is_zero(b)) ||
231         (float32_is_infinity(b) && float32_is_zero(a))) {
232         return float32_two;
233     }
234     return float32_muladd(a, b, float32_two, 0, fpst);
235 }
236 
237 float64 HELPER(recpsf_f64)(float64 a, float64 b, void *fpstp)
238 {
239     float_status *fpst = fpstp;
240 
241     a = float64_squash_input_denormal(a, fpst);
242     b = float64_squash_input_denormal(b, fpst);
243 
244     a = float64_chs(a);
245     if ((float64_is_infinity(a) && float64_is_zero(b)) ||
246         (float64_is_infinity(b) && float64_is_zero(a))) {
247         return float64_two;
248     }
249     return float64_muladd(a, b, float64_two, 0, fpst);
250 }
251 
252 uint32_t HELPER(rsqrtsf_f16)(uint32_t a, uint32_t b, void *fpstp)
253 {
254     float_status *fpst = fpstp;
255 
256     a = float16_squash_input_denormal(a, fpst);
257     b = float16_squash_input_denormal(b, fpst);
258 
259     a = float16_chs(a);
260     if ((float16_is_infinity(a) && float16_is_zero(b)) ||
261         (float16_is_infinity(b) && float16_is_zero(a))) {
262         return float16_one_point_five;
263     }
264     return float16_muladd(a, b, float16_three, float_muladd_halve_result, fpst);
265 }
266 
267 float32 HELPER(rsqrtsf_f32)(float32 a, float32 b, void *fpstp)
268 {
269     float_status *fpst = fpstp;
270 
271     a = float32_squash_input_denormal(a, fpst);
272     b = float32_squash_input_denormal(b, fpst);
273 
274     a = float32_chs(a);
275     if ((float32_is_infinity(a) && float32_is_zero(b)) ||
276         (float32_is_infinity(b) && float32_is_zero(a))) {
277         return float32_one_point_five;
278     }
279     return float32_muladd(a, b, float32_three, float_muladd_halve_result, fpst);
280 }
281 
282 float64 HELPER(rsqrtsf_f64)(float64 a, float64 b, void *fpstp)
283 {
284     float_status *fpst = fpstp;
285 
286     a = float64_squash_input_denormal(a, fpst);
287     b = float64_squash_input_denormal(b, fpst);
288 
289     a = float64_chs(a);
290     if ((float64_is_infinity(a) && float64_is_zero(b)) ||
291         (float64_is_infinity(b) && float64_is_zero(a))) {
292         return float64_one_point_five;
293     }
294     return float64_muladd(a, b, float64_three, float_muladd_halve_result, fpst);
295 }
296 
297 /* Pairwise long add: add pairs of adjacent elements into
298  * double-width elements in the result (eg _s8 is an 8x8->16 op)
299  */
300 uint64_t HELPER(neon_addlp_s8)(uint64_t a)
301 {
302     uint64_t nsignmask = 0x0080008000800080ULL;
303     uint64_t wsignmask = 0x8000800080008000ULL;
304     uint64_t elementmask = 0x00ff00ff00ff00ffULL;
305     uint64_t tmp1, tmp2;
306     uint64_t res, signres;
307 
308     /* Extract odd elements, sign extend each to a 16 bit field */
309     tmp1 = a & elementmask;
310     tmp1 ^= nsignmask;
311     tmp1 |= wsignmask;
312     tmp1 = (tmp1 - nsignmask) ^ wsignmask;
313     /* Ditto for the even elements */
314     tmp2 = (a >> 8) & elementmask;
315     tmp2 ^= nsignmask;
316     tmp2 |= wsignmask;
317     tmp2 = (tmp2 - nsignmask) ^ wsignmask;
318 
319     /* calculate the result by summing bits 0..14, 16..22, etc,
320      * and then adjusting the sign bits 15, 23, etc manually.
321      * This ensures the addition can't overflow the 16 bit field.
322      */
323     signres = (tmp1 ^ tmp2) & wsignmask;
324     res = (tmp1 & ~wsignmask) + (tmp2 & ~wsignmask);
325     res ^= signres;
326 
327     return res;
328 }
329 
330 uint64_t HELPER(neon_addlp_u8)(uint64_t a)
331 {
332     uint64_t tmp;
333 
334     tmp = a & 0x00ff00ff00ff00ffULL;
335     tmp += (a >> 8) & 0x00ff00ff00ff00ffULL;
336     return tmp;
337 }
338 
339 uint64_t HELPER(neon_addlp_s16)(uint64_t a)
340 {
341     int32_t reslo, reshi;
342 
343     reslo = (int32_t)(int16_t)a + (int32_t)(int16_t)(a >> 16);
344     reshi = (int32_t)(int16_t)(a >> 32) + (int32_t)(int16_t)(a >> 48);
345 
346     return (uint32_t)reslo | (((uint64_t)reshi) << 32);
347 }
348 
349 uint64_t HELPER(neon_addlp_u16)(uint64_t a)
350 {
351     uint64_t tmp;
352 
353     tmp = a & 0x0000ffff0000ffffULL;
354     tmp += (a >> 16) & 0x0000ffff0000ffffULL;
355     return tmp;
356 }
357 
358 /* Floating-point reciprocal exponent - see FPRecpX in ARM ARM */
359 uint32_t HELPER(frecpx_f16)(uint32_t a, void *fpstp)
360 {
361     float_status *fpst = fpstp;
362     uint16_t val16, sbit;
363     int16_t exp;
364 
365     if (float16_is_any_nan(a)) {
366         float16 nan = a;
367         if (float16_is_signaling_nan(a, fpst)) {
368             float_raise(float_flag_invalid, fpst);
369             if (!fpst->default_nan_mode) {
370                 nan = float16_silence_nan(a, fpst);
371             }
372         }
373         if (fpst->default_nan_mode) {
374             nan = float16_default_nan(fpst);
375         }
376         return nan;
377     }
378 
379     a = float16_squash_input_denormal(a, fpst);
380 
381     val16 = float16_val(a);
382     sbit = 0x8000 & val16;
383     exp = extract32(val16, 10, 5);
384 
385     if (exp == 0) {
386         return make_float16(deposit32(sbit, 10, 5, 0x1e));
387     } else {
388         return make_float16(deposit32(sbit, 10, 5, ~exp));
389     }
390 }
391 
392 float32 HELPER(frecpx_f32)(float32 a, void *fpstp)
393 {
394     float_status *fpst = fpstp;
395     uint32_t val32, sbit;
396     int32_t exp;
397 
398     if (float32_is_any_nan(a)) {
399         float32 nan = a;
400         if (float32_is_signaling_nan(a, fpst)) {
401             float_raise(float_flag_invalid, fpst);
402             if (!fpst->default_nan_mode) {
403                 nan = float32_silence_nan(a, fpst);
404             }
405         }
406         if (fpst->default_nan_mode) {
407             nan = float32_default_nan(fpst);
408         }
409         return nan;
410     }
411 
412     a = float32_squash_input_denormal(a, fpst);
413 
414     val32 = float32_val(a);
415     sbit = 0x80000000ULL & val32;
416     exp = extract32(val32, 23, 8);
417 
418     if (exp == 0) {
419         return make_float32(sbit | (0xfe << 23));
420     } else {
421         return make_float32(sbit | (~exp & 0xff) << 23);
422     }
423 }
424 
425 float64 HELPER(frecpx_f64)(float64 a, void *fpstp)
426 {
427     float_status *fpst = fpstp;
428     uint64_t val64, sbit;
429     int64_t exp;
430 
431     if (float64_is_any_nan(a)) {
432         float64 nan = a;
433         if (float64_is_signaling_nan(a, fpst)) {
434             float_raise(float_flag_invalid, fpst);
435             if (!fpst->default_nan_mode) {
436                 nan = float64_silence_nan(a, fpst);
437             }
438         }
439         if (fpst->default_nan_mode) {
440             nan = float64_default_nan(fpst);
441         }
442         return nan;
443     }
444 
445     a = float64_squash_input_denormal(a, fpst);
446 
447     val64 = float64_val(a);
448     sbit = 0x8000000000000000ULL & val64;
449     exp = extract64(float64_val(a), 52, 11);
450 
451     if (exp == 0) {
452         return make_float64(sbit | (0x7feULL << 52));
453     } else {
454         return make_float64(sbit | (~exp & 0x7ffULL) << 52);
455     }
456 }
457 
458 float32 HELPER(fcvtx_f64_to_f32)(float64 a, CPUARMState *env)
459 {
460     /* Von Neumann rounding is implemented by using round-to-zero
461      * and then setting the LSB of the result if Inexact was raised.
462      */
463     float32 r;
464     float_status *fpst = &env->vfp.fp_status;
465     float_status tstat = *fpst;
466     int exflags;
467 
468     set_float_rounding_mode(float_round_to_zero, &tstat);
469     set_float_exception_flags(0, &tstat);
470     r = float64_to_float32(a, &tstat);
471     exflags = get_float_exception_flags(&tstat);
472     if (exflags & float_flag_inexact) {
473         r = make_float32(float32_val(r) | 1);
474     }
475     exflags |= get_float_exception_flags(fpst);
476     set_float_exception_flags(exflags, fpst);
477     return r;
478 }
479 
480 /* 64-bit versions of the CRC helpers. Note that although the operation
481  * (and the prototypes of crc32c() and crc32() mean that only the bottom
482  * 32 bits of the accumulator and result are used, we pass and return
483  * uint64_t for convenience of the generated code. Unlike the 32-bit
484  * instruction set versions, val may genuinely have 64 bits of data in it.
485  * The upper bytes of val (above the number specified by 'bytes') must have
486  * been zeroed out by the caller.
487  */
488 uint64_t HELPER(crc32_64)(uint64_t acc, uint64_t val, uint32_t bytes)
489 {
490     uint8_t buf[8];
491 
492     stq_le_p(buf, val);
493 
494     /* zlib crc32 converts the accumulator and output to one's complement.  */
495     return crc32(acc ^ 0xffffffff, buf, bytes) ^ 0xffffffff;
496 }
497 
498 uint64_t HELPER(crc32c_64)(uint64_t acc, uint64_t val, uint32_t bytes)
499 {
500     uint8_t buf[8];
501 
502     stq_le_p(buf, val);
503 
504     /* Linux crc32c converts the output to one's complement.  */
505     return crc32c(acc, buf, bytes) ^ 0xffffffff;
506 }
507 
508 /*
509  * AdvSIMD half-precision
510  */
511 
512 #define ADVSIMD_HELPER(name, suffix) HELPER(glue(glue(advsimd_, name), suffix))
513 
514 #define ADVSIMD_HALFOP(name) \
515 uint32_t ADVSIMD_HELPER(name, h)(uint32_t a, uint32_t b, void *fpstp) \
516 { \
517     float_status *fpst = fpstp; \
518     return float16_ ## name(a, b, fpst);    \
519 }
520 
521 ADVSIMD_HALFOP(add)
522 ADVSIMD_HALFOP(sub)
523 ADVSIMD_HALFOP(mul)
524 ADVSIMD_HALFOP(div)
525 ADVSIMD_HALFOP(min)
526 ADVSIMD_HALFOP(max)
527 ADVSIMD_HALFOP(minnum)
528 ADVSIMD_HALFOP(maxnum)
529 
530 #define ADVSIMD_TWOHALFOP(name)                                         \
531 uint32_t ADVSIMD_HELPER(name, 2h)(uint32_t two_a, uint32_t two_b, void *fpstp) \
532 { \
533     float16  a1, a2, b1, b2;                        \
534     uint32_t r1, r2;                                \
535     float_status *fpst = fpstp;                     \
536     a1 = extract32(two_a, 0, 16);                   \
537     a2 = extract32(two_a, 16, 16);                  \
538     b1 = extract32(two_b, 0, 16);                   \
539     b2 = extract32(two_b, 16, 16);                  \
540     r1 = float16_ ## name(a1, b1, fpst);            \
541     r2 = float16_ ## name(a2, b2, fpst);            \
542     return deposit32(r1, 16, 16, r2);               \
543 }
544 
545 ADVSIMD_TWOHALFOP(add)
546 ADVSIMD_TWOHALFOP(sub)
547 ADVSIMD_TWOHALFOP(mul)
548 ADVSIMD_TWOHALFOP(div)
549 ADVSIMD_TWOHALFOP(min)
550 ADVSIMD_TWOHALFOP(max)
551 ADVSIMD_TWOHALFOP(minnum)
552 ADVSIMD_TWOHALFOP(maxnum)
553 
554 /* Data processing - scalar floating-point and advanced SIMD */
555 static float16 float16_mulx(float16 a, float16 b, void *fpstp)
556 {
557     float_status *fpst = fpstp;
558 
559     a = float16_squash_input_denormal(a, fpst);
560     b = float16_squash_input_denormal(b, fpst);
561 
562     if ((float16_is_zero(a) && float16_is_infinity(b)) ||
563         (float16_is_infinity(a) && float16_is_zero(b))) {
564         /* 2.0 with the sign bit set to sign(A) XOR sign(B) */
565         return make_float16((1U << 14) |
566                             ((float16_val(a) ^ float16_val(b)) & (1U << 15)));
567     }
568     return float16_mul(a, b, fpst);
569 }
570 
571 ADVSIMD_HALFOP(mulx)
572 ADVSIMD_TWOHALFOP(mulx)
573 
574 /* fused multiply-accumulate */
575 uint32_t HELPER(advsimd_muladdh)(uint32_t a, uint32_t b, uint32_t c,
576                                  void *fpstp)
577 {
578     float_status *fpst = fpstp;
579     return float16_muladd(a, b, c, 0, fpst);
580 }
581 
582 uint32_t HELPER(advsimd_muladd2h)(uint32_t two_a, uint32_t two_b,
583                                   uint32_t two_c, void *fpstp)
584 {
585     float_status *fpst = fpstp;
586     float16  a1, a2, b1, b2, c1, c2;
587     uint32_t r1, r2;
588     a1 = extract32(two_a, 0, 16);
589     a2 = extract32(two_a, 16, 16);
590     b1 = extract32(two_b, 0, 16);
591     b2 = extract32(two_b, 16, 16);
592     c1 = extract32(two_c, 0, 16);
593     c2 = extract32(two_c, 16, 16);
594     r1 = float16_muladd(a1, b1, c1, 0, fpst);
595     r2 = float16_muladd(a2, b2, c2, 0, fpst);
596     return deposit32(r1, 16, 16, r2);
597 }
598 
599 /*
600  * Floating point comparisons produce an integer result. Softfloat
601  * routines return float_relation types which we convert to the 0/-1
602  * Neon requires.
603  */
604 
605 #define ADVSIMD_CMPRES(test) (test) ? 0xffff : 0
606 
607 uint32_t HELPER(advsimd_ceq_f16)(uint32_t a, uint32_t b, void *fpstp)
608 {
609     float_status *fpst = fpstp;
610     int compare = float16_compare_quiet(a, b, fpst);
611     return ADVSIMD_CMPRES(compare == float_relation_equal);
612 }
613 
614 uint32_t HELPER(advsimd_cge_f16)(uint32_t a, uint32_t b, void *fpstp)
615 {
616     float_status *fpst = fpstp;
617     int compare = float16_compare(a, b, fpst);
618     return ADVSIMD_CMPRES(compare == float_relation_greater ||
619                           compare == float_relation_equal);
620 }
621 
622 uint32_t HELPER(advsimd_cgt_f16)(uint32_t a, uint32_t b, void *fpstp)
623 {
624     float_status *fpst = fpstp;
625     int compare = float16_compare(a, b, fpst);
626     return ADVSIMD_CMPRES(compare == float_relation_greater);
627 }
628 
629 uint32_t HELPER(advsimd_acge_f16)(uint32_t a, uint32_t b, void *fpstp)
630 {
631     float_status *fpst = fpstp;
632     float16 f0 = float16_abs(a);
633     float16 f1 = float16_abs(b);
634     int compare = float16_compare(f0, f1, fpst);
635     return ADVSIMD_CMPRES(compare == float_relation_greater ||
636                           compare == float_relation_equal);
637 }
638 
639 uint32_t HELPER(advsimd_acgt_f16)(uint32_t a, uint32_t b, void *fpstp)
640 {
641     float_status *fpst = fpstp;
642     float16 f0 = float16_abs(a);
643     float16 f1 = float16_abs(b);
644     int compare = float16_compare(f0, f1, fpst);
645     return ADVSIMD_CMPRES(compare == float_relation_greater);
646 }
647 
648 /* round to integral */
649 uint32_t HELPER(advsimd_rinth_exact)(uint32_t x, void *fp_status)
650 {
651     return float16_round_to_int(x, fp_status);
652 }
653 
654 uint32_t HELPER(advsimd_rinth)(uint32_t x, void *fp_status)
655 {
656     int old_flags = get_float_exception_flags(fp_status), new_flags;
657     float16 ret;
658 
659     ret = float16_round_to_int(x, fp_status);
660 
661     /* Suppress any inexact exceptions the conversion produced */
662     if (!(old_flags & float_flag_inexact)) {
663         new_flags = get_float_exception_flags(fp_status);
664         set_float_exception_flags(new_flags & ~float_flag_inexact, fp_status);
665     }
666 
667     return ret;
668 }
669 
670 /*
671  * Half-precision floating point conversion functions
672  *
673  * There are a multitude of conversion functions with various
674  * different rounding modes. This is dealt with by the calling code
675  * setting the mode appropriately before calling the helper.
676  */
677 
678 uint32_t HELPER(advsimd_f16tosinth)(uint32_t a, void *fpstp)
679 {
680     float_status *fpst = fpstp;
681 
682     /* Invalid if we are passed a NaN */
683     if (float16_is_any_nan(a)) {
684         float_raise(float_flag_invalid, fpst);
685         return 0;
686     }
687     return float16_to_int16(a, fpst);
688 }
689 
690 uint32_t HELPER(advsimd_f16touinth)(uint32_t a, void *fpstp)
691 {
692     float_status *fpst = fpstp;
693 
694     /* Invalid if we are passed a NaN */
695     if (float16_is_any_nan(a)) {
696         float_raise(float_flag_invalid, fpst);
697         return 0;
698     }
699     return float16_to_uint16(a, fpst);
700 }
701 
702 static int el_from_spsr(uint32_t spsr)
703 {
704     /* Return the exception level that this SPSR is requesting a return to,
705      * or -1 if it is invalid (an illegal return)
706      */
707     if (spsr & PSTATE_nRW) {
708         switch (spsr & CPSR_M) {
709         case ARM_CPU_MODE_USR:
710             return 0;
711         case ARM_CPU_MODE_HYP:
712             return 2;
713         case ARM_CPU_MODE_FIQ:
714         case ARM_CPU_MODE_IRQ:
715         case ARM_CPU_MODE_SVC:
716         case ARM_CPU_MODE_ABT:
717         case ARM_CPU_MODE_UND:
718         case ARM_CPU_MODE_SYS:
719             return 1;
720         case ARM_CPU_MODE_MON:
721             /* Returning to Mon from AArch64 is never possible,
722              * so this is an illegal return.
723              */
724         default:
725             return -1;
726         }
727     } else {
728         if (extract32(spsr, 1, 1)) {
729             /* Return with reserved M[1] bit set */
730             return -1;
731         }
732         if (extract32(spsr, 0, 4) == 1) {
733             /* return to EL0 with M[0] bit set */
734             return -1;
735         }
736         return extract32(spsr, 2, 2);
737     }
738 }
739 
740 static void cpsr_write_from_spsr_elx(CPUARMState *env,
741                                      uint32_t val)
742 {
743     uint32_t mask;
744 
745     /* Save SPSR_ELx.SS into PSTATE. */
746     env->pstate = (env->pstate & ~PSTATE_SS) | (val & PSTATE_SS);
747     val &= ~PSTATE_SS;
748 
749     /* Move DIT to the correct location for CPSR */
750     if (val & PSTATE_DIT) {
751         val &= ~PSTATE_DIT;
752         val |= CPSR_DIT;
753     }
754 
755     mask = aarch32_cpsr_valid_mask(env->features, \
756         &env_archcpu(env)->isar);
757     cpsr_write(env, val, mask, CPSRWriteRaw);
758 }
759 
760 void HELPER(exception_return)(CPUARMState *env, uint64_t new_pc)
761 {
762     int cur_el = arm_current_el(env);
763     unsigned int spsr_idx = aarch64_banked_spsr_index(cur_el);
764     uint32_t spsr = env->banked_spsr[spsr_idx];
765     int new_el;
766     bool return_to_aa64 = (spsr & PSTATE_nRW) == 0;
767 
768     aarch64_save_sp(env, cur_el);
769 
770     arm_clear_exclusive(env);
771 
772     /* We must squash the PSTATE.SS bit to zero unless both of the
773      * following hold:
774      *  1. debug exceptions are currently disabled
775      *  2. singlestep will be active in the EL we return to
776      * We check 1 here and 2 after we've done the pstate/cpsr write() to
777      * transition to the EL we're going to.
778      */
779     if (arm_generate_debug_exceptions(env)) {
780         spsr &= ~PSTATE_SS;
781     }
782 
783     new_el = el_from_spsr(spsr);
784     if (new_el == -1) {
785         goto illegal_return;
786     }
787     if (new_el > cur_el || (new_el == 2 && !arm_is_el2_enabled(env))) {
788         /* Disallow return to an EL which is unimplemented or higher
789          * than the current one.
790          */
791         goto illegal_return;
792     }
793 
794     if (new_el != 0 && arm_el_is_aa64(env, new_el) != return_to_aa64) {
795         /* Return to an EL which is configured for a different register width */
796         goto illegal_return;
797     }
798 
799     if (new_el == 1 && (arm_hcr_el2_eff(env) & HCR_TGE)) {
800         goto illegal_return;
801     }
802 
803     qemu_mutex_lock_iothread();
804     arm_call_pre_el_change_hook(env_archcpu(env));
805     qemu_mutex_unlock_iothread();
806 
807     if (!return_to_aa64) {
808         env->aarch64 = false;
809         /* We do a raw CPSR write because aarch64_sync_64_to_32()
810          * will sort the register banks out for us, and we've already
811          * caught all the bad-mode cases in el_from_spsr().
812          */
813         cpsr_write_from_spsr_elx(env, spsr);
814         if (!arm_singlestep_active(env)) {
815             env->pstate &= ~PSTATE_SS;
816         }
817         aarch64_sync_64_to_32(env);
818 
819         if (spsr & CPSR_T) {
820             env->regs[15] = new_pc & ~0x1;
821         } else {
822             env->regs[15] = new_pc & ~0x3;
823         }
824         helper_rebuild_hflags_a32(env, new_el);
825         qemu_log_mask(CPU_LOG_INT, "Exception return from AArch64 EL%d to "
826                       "AArch32 EL%d PC 0x%" PRIx32 "\n",
827                       cur_el, new_el, env->regs[15]);
828     } else {
829         int tbii;
830 
831         env->aarch64 = true;
832         spsr &= aarch64_pstate_valid_mask(&env_archcpu(env)->isar);
833         pstate_write(env, spsr);
834         if (!arm_singlestep_active(env)) {
835             env->pstate &= ~PSTATE_SS;
836         }
837         aarch64_restore_sp(env, new_el);
838         helper_rebuild_hflags_a64(env, new_el);
839 
840         /*
841          * Apply TBI to the exception return address.  We had to delay this
842          * until after we selected the new EL, so that we could select the
843          * correct TBI+TBID bits.  This is made easier by waiting until after
844          * the hflags rebuild, since we can pull the composite TBII field
845          * from there.
846          */
847         tbii = EX_TBFLAG_A64(env->hflags, TBII);
848         if ((tbii >> extract64(new_pc, 55, 1)) & 1) {
849             /* TBI is enabled. */
850             int core_mmu_idx = cpu_mmu_index(env, false);
851             if (regime_has_2_ranges(core_to_aa64_mmu_idx(core_mmu_idx))) {
852                 new_pc = sextract64(new_pc, 0, 56);
853             } else {
854                 new_pc = extract64(new_pc, 0, 56);
855             }
856         }
857         env->pc = new_pc;
858 
859         qemu_log_mask(CPU_LOG_INT, "Exception return from AArch64 EL%d to "
860                       "AArch64 EL%d PC 0x%" PRIx64 "\n",
861                       cur_el, new_el, env->pc);
862     }
863 
864     /*
865      * Note that cur_el can never be 0.  If new_el is 0, then
866      * el0_a64 is return_to_aa64, else el0_a64 is ignored.
867      */
868     aarch64_sve_change_el(env, cur_el, new_el, return_to_aa64);
869 
870     qemu_mutex_lock_iothread();
871     arm_call_el_change_hook(env_archcpu(env));
872     qemu_mutex_unlock_iothread();
873 
874     return;
875 
876 illegal_return:
877     /* Illegal return events of various kinds have architecturally
878      * mandated behaviour:
879      * restore NZCV and DAIF from SPSR_ELx
880      * set PSTATE.IL
881      * restore PC from ELR_ELx
882      * no change to exception level, execution state or stack pointer
883      */
884     env->pstate |= PSTATE_IL;
885     env->pc = new_pc;
886     spsr &= PSTATE_NZCV | PSTATE_DAIF;
887     spsr |= pstate_read(env) & ~(PSTATE_NZCV | PSTATE_DAIF);
888     pstate_write(env, spsr);
889     if (!arm_singlestep_active(env)) {
890         env->pstate &= ~PSTATE_SS;
891     }
892     helper_rebuild_hflags_a64(env, cur_el);
893     qemu_log_mask(LOG_GUEST_ERROR, "Illegal exception return at EL%d: "
894                   "resuming execution at 0x%" PRIx64 "\n", cur_el, env->pc);
895 }
896 
897 /*
898  * Square Root and Reciprocal square root
899  */
900 
901 uint32_t HELPER(sqrt_f16)(uint32_t a, void *fpstp)
902 {
903     float_status *s = fpstp;
904 
905     return float16_sqrt(a, s);
906 }
907 
908 void HELPER(dc_zva)(CPUARMState *env, uint64_t vaddr_in)
909 {
910     /*
911      * Implement DC ZVA, which zeroes a fixed-length block of memory.
912      * Note that we do not implement the (architecturally mandated)
913      * alignment fault for attempts to use this on Device memory
914      * (which matches the usual QEMU behaviour of not implementing either
915      * alignment faults or any memory attribute handling).
916      */
917     int blocklen = 4 << env_archcpu(env)->dcz_blocksize;
918     uint64_t vaddr = vaddr_in & ~(blocklen - 1);
919     int mmu_idx = cpu_mmu_index(env, false);
920     void *mem;
921 
922     /*
923      * Trapless lookup.  In addition to actual invalid page, may
924      * return NULL for I/O, watchpoints, clean pages, etc.
925      */
926     mem = tlb_vaddr_to_host(env, vaddr, MMU_DATA_STORE, mmu_idx);
927 
928 #ifndef CONFIG_USER_ONLY
929     if (unlikely(!mem)) {
930         uintptr_t ra = GETPC();
931 
932         /*
933          * Trap if accessing an invalid page.  DC_ZVA requires that we supply
934          * the original pointer for an invalid page.  But watchpoints require
935          * that we probe the actual space.  So do both.
936          */
937         (void) probe_write(env, vaddr_in, 1, mmu_idx, ra);
938         mem = probe_write(env, vaddr, blocklen, mmu_idx, ra);
939 
940         if (unlikely(!mem)) {
941             /*
942              * The only remaining reason for mem == NULL is I/O.
943              * Just do a series of byte writes as the architecture demands.
944              */
945             for (int i = 0; i < blocklen; i++) {
946                 cpu_stb_mmuidx_ra(env, vaddr + i, 0, mmu_idx, ra);
947             }
948             return;
949         }
950     }
951 #endif
952 
953     memset(mem, 0, blocklen);
954 }
955