xref: /openbmc/qemu/target/arm/tcg/helper-a64.c (revision f14eced5)
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 "gdbstub/helpers.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     /*
784      * FEAT_RME forbids return from EL3 with an invalid security state.
785      * We don't need an explicit check for FEAT_RME here because we enforce
786      * in scr_write() that you can't set the NSE bit without it.
787      */
788     if (cur_el == 3 && (env->cp15.scr_el3 & (SCR_NS | SCR_NSE)) == SCR_NSE) {
789         goto illegal_return;
790     }
791 
792     new_el = el_from_spsr(spsr);
793     if (new_el == -1) {
794         goto illegal_return;
795     }
796     if (new_el > cur_el || (new_el == 2 && !arm_is_el2_enabled(env))) {
797         /* Disallow return to an EL which is unimplemented or higher
798          * than the current one.
799          */
800         goto illegal_return;
801     }
802 
803     if (new_el != 0 && arm_el_is_aa64(env, new_el) != return_to_aa64) {
804         /* Return to an EL which is configured for a different register width */
805         goto illegal_return;
806     }
807 
808     if (new_el == 1 && (arm_hcr_el2_eff(env) & HCR_TGE)) {
809         goto illegal_return;
810     }
811 
812     qemu_mutex_lock_iothread();
813     arm_call_pre_el_change_hook(env_archcpu(env));
814     qemu_mutex_unlock_iothread();
815 
816     if (!return_to_aa64) {
817         env->aarch64 = false;
818         /* We do a raw CPSR write because aarch64_sync_64_to_32()
819          * will sort the register banks out for us, and we've already
820          * caught all the bad-mode cases in el_from_spsr().
821          */
822         cpsr_write_from_spsr_elx(env, spsr);
823         if (!arm_singlestep_active(env)) {
824             env->pstate &= ~PSTATE_SS;
825         }
826         aarch64_sync_64_to_32(env);
827 
828         if (spsr & CPSR_T) {
829             env->regs[15] = new_pc & ~0x1;
830         } else {
831             env->regs[15] = new_pc & ~0x3;
832         }
833         helper_rebuild_hflags_a32(env, new_el);
834         qemu_log_mask(CPU_LOG_INT, "Exception return from AArch64 EL%d to "
835                       "AArch32 EL%d PC 0x%" PRIx32 "\n",
836                       cur_el, new_el, env->regs[15]);
837     } else {
838         int tbii;
839 
840         env->aarch64 = true;
841         spsr &= aarch64_pstate_valid_mask(&env_archcpu(env)->isar);
842         pstate_write(env, spsr);
843         if (!arm_singlestep_active(env)) {
844             env->pstate &= ~PSTATE_SS;
845         }
846         aarch64_restore_sp(env, new_el);
847         helper_rebuild_hflags_a64(env, new_el);
848 
849         /*
850          * Apply TBI to the exception return address.  We had to delay this
851          * until after we selected the new EL, so that we could select the
852          * correct TBI+TBID bits.  This is made easier by waiting until after
853          * the hflags rebuild, since we can pull the composite TBII field
854          * from there.
855          */
856         tbii = EX_TBFLAG_A64(env->hflags, TBII);
857         if ((tbii >> extract64(new_pc, 55, 1)) & 1) {
858             /* TBI is enabled. */
859             int core_mmu_idx = cpu_mmu_index(env, false);
860             if (regime_has_2_ranges(core_to_aa64_mmu_idx(core_mmu_idx))) {
861                 new_pc = sextract64(new_pc, 0, 56);
862             } else {
863                 new_pc = extract64(new_pc, 0, 56);
864             }
865         }
866         env->pc = new_pc;
867 
868         qemu_log_mask(CPU_LOG_INT, "Exception return from AArch64 EL%d to "
869                       "AArch64 EL%d PC 0x%" PRIx64 "\n",
870                       cur_el, new_el, env->pc);
871     }
872 
873     /*
874      * Note that cur_el can never be 0.  If new_el is 0, then
875      * el0_a64 is return_to_aa64, else el0_a64 is ignored.
876      */
877     aarch64_sve_change_el(env, cur_el, new_el, return_to_aa64);
878 
879     qemu_mutex_lock_iothread();
880     arm_call_el_change_hook(env_archcpu(env));
881     qemu_mutex_unlock_iothread();
882 
883     return;
884 
885 illegal_return:
886     /* Illegal return events of various kinds have architecturally
887      * mandated behaviour:
888      * restore NZCV and DAIF from SPSR_ELx
889      * set PSTATE.IL
890      * restore PC from ELR_ELx
891      * no change to exception level, execution state or stack pointer
892      */
893     env->pstate |= PSTATE_IL;
894     env->pc = new_pc;
895     spsr &= PSTATE_NZCV | PSTATE_DAIF;
896     spsr |= pstate_read(env) & ~(PSTATE_NZCV | PSTATE_DAIF);
897     pstate_write(env, spsr);
898     if (!arm_singlestep_active(env)) {
899         env->pstate &= ~PSTATE_SS;
900     }
901     helper_rebuild_hflags_a64(env, cur_el);
902     qemu_log_mask(LOG_GUEST_ERROR, "Illegal exception return at EL%d: "
903                   "resuming execution at 0x%" PRIx64 "\n", cur_el, env->pc);
904 }
905 
906 /*
907  * Square Root and Reciprocal square root
908  */
909 
910 uint32_t HELPER(sqrt_f16)(uint32_t a, void *fpstp)
911 {
912     float_status *s = fpstp;
913 
914     return float16_sqrt(a, s);
915 }
916 
917 void HELPER(dc_zva)(CPUARMState *env, uint64_t vaddr_in)
918 {
919     /*
920      * Implement DC ZVA, which zeroes a fixed-length block of memory.
921      * Note that we do not implement the (architecturally mandated)
922      * alignment fault for attempts to use this on Device memory
923      * (which matches the usual QEMU behaviour of not implementing either
924      * alignment faults or any memory attribute handling).
925      */
926     int blocklen = 4 << env_archcpu(env)->dcz_blocksize;
927     uint64_t vaddr = vaddr_in & ~(blocklen - 1);
928     int mmu_idx = cpu_mmu_index(env, false);
929     void *mem;
930 
931     /*
932      * Trapless lookup.  In addition to actual invalid page, may
933      * return NULL for I/O, watchpoints, clean pages, etc.
934      */
935     mem = tlb_vaddr_to_host(env, vaddr, MMU_DATA_STORE, mmu_idx);
936 
937 #ifndef CONFIG_USER_ONLY
938     if (unlikely(!mem)) {
939         uintptr_t ra = GETPC();
940 
941         /*
942          * Trap if accessing an invalid page.  DC_ZVA requires that we supply
943          * the original pointer for an invalid page.  But watchpoints require
944          * that we probe the actual space.  So do both.
945          */
946         (void) probe_write(env, vaddr_in, 1, mmu_idx, ra);
947         mem = probe_write(env, vaddr, blocklen, mmu_idx, ra);
948 
949         if (unlikely(!mem)) {
950             /*
951              * The only remaining reason for mem == NULL is I/O.
952              * Just do a series of byte writes as the architecture demands.
953              */
954             for (int i = 0; i < blocklen; i++) {
955                 cpu_stb_mmuidx_ra(env, vaddr + i, 0, mmu_idx, ra);
956             }
957             return;
958         }
959     }
960 #endif
961 
962     memset(mem, 0, blocklen);
963 }
964 
965 void HELPER(unaligned_access)(CPUARMState *env, uint64_t addr,
966                               uint32_t access_type, uint32_t mmu_idx)
967 {
968     arm_cpu_do_unaligned_access(env_cpu(env), addr, access_type,
969                                 mmu_idx, GETPC());
970 }
971 
972 /* Memory operations (memset, memmove, memcpy) */
973 
974 /*
975  * Return true if the CPY* and SET* insns can execute; compare
976  * pseudocode CheckMOPSEnabled(), though we refactor it a little.
977  */
978 static bool mops_enabled(CPUARMState *env)
979 {
980     int el = arm_current_el(env);
981 
982     if (el < 2 &&
983         (arm_hcr_el2_eff(env) & (HCR_E2H | HCR_TGE)) != (HCR_E2H | HCR_TGE) &&
984         !(arm_hcrx_el2_eff(env) & HCRX_MSCEN)) {
985         return false;
986     }
987 
988     if (el == 0) {
989         if (!el_is_in_host(env, 0)) {
990             return env->cp15.sctlr_el[1] & SCTLR_MSCEN;
991         } else {
992             return env->cp15.sctlr_el[2] & SCTLR_MSCEN;
993         }
994     }
995     return true;
996 }
997 
998 static void check_mops_enabled(CPUARMState *env, uintptr_t ra)
999 {
1000     if (!mops_enabled(env)) {
1001         raise_exception_ra(env, EXCP_UDEF, syn_uncategorized(),
1002                            exception_target_el(env), ra);
1003     }
1004 }
1005 
1006 /*
1007  * Return the target exception level for an exception due
1008  * to mismatched arguments in a FEAT_MOPS copy or set.
1009  * Compare pseudocode MismatchedCpySetTargetEL()
1010  */
1011 static int mops_mismatch_exception_target_el(CPUARMState *env)
1012 {
1013     int el = arm_current_el(env);
1014 
1015     if (el > 1) {
1016         return el;
1017     }
1018     if (el == 0 && (arm_hcr_el2_eff(env) & HCR_TGE)) {
1019         return 2;
1020     }
1021     if (el == 1 && (arm_hcrx_el2_eff(env) & HCRX_MCE2)) {
1022         return 2;
1023     }
1024     return 1;
1025 }
1026 
1027 /*
1028  * Check whether an M or E instruction was executed with a CF value
1029  * indicating the wrong option for this implementation.
1030  * Assumes we are always Option A.
1031  */
1032 static void check_mops_wrong_option(CPUARMState *env, uint32_t syndrome,
1033                                     uintptr_t ra)
1034 {
1035     if (env->CF != 0) {
1036         syndrome |= 1 << 17; /* Set the wrong-option bit */
1037         raise_exception_ra(env, EXCP_UDEF, syndrome,
1038                            mops_mismatch_exception_target_el(env), ra);
1039     }
1040 }
1041 
1042 /*
1043  * Return the maximum number of bytes we can transfer starting at addr
1044  * without crossing a page boundary.
1045  */
1046 static uint64_t page_limit(uint64_t addr)
1047 {
1048     return TARGET_PAGE_ALIGN(addr + 1) - addr;
1049 }
1050 
1051 /*
1052  * Return the number of bytes we can copy starting from addr and working
1053  * backwards without crossing a page boundary.
1054  */
1055 static uint64_t page_limit_rev(uint64_t addr)
1056 {
1057     return (addr & ~TARGET_PAGE_MASK) + 1;
1058 }
1059 
1060 /*
1061  * Perform part of a memory set on an area of guest memory starting at
1062  * toaddr (a dirty address) and extending for setsize bytes.
1063  *
1064  * Returns the number of bytes actually set, which might be less than
1065  * setsize; the caller should loop until the whole set has been done.
1066  * The caller should ensure that the guest registers are correct
1067  * for the possibility that the first byte of the set encounters
1068  * an exception or watchpoint. We guarantee not to take any faults
1069  * for bytes other than the first.
1070  */
1071 static uint64_t set_step(CPUARMState *env, uint64_t toaddr,
1072                          uint64_t setsize, uint32_t data, int memidx,
1073                          uint32_t *mtedesc, uintptr_t ra)
1074 {
1075     void *mem;
1076 
1077     setsize = MIN(setsize, page_limit(toaddr));
1078     if (*mtedesc) {
1079         uint64_t mtesize = mte_mops_probe(env, toaddr, setsize, *mtedesc);
1080         if (mtesize == 0) {
1081             /* Trap, or not. All CPU state is up to date */
1082             mte_check_fail(env, *mtedesc, toaddr, ra);
1083             /* Continue, with no further MTE checks required */
1084             *mtedesc = 0;
1085         } else {
1086             /* Advance to the end, or to the tag mismatch */
1087             setsize = MIN(setsize, mtesize);
1088         }
1089     }
1090 
1091     toaddr = useronly_clean_ptr(toaddr);
1092     /*
1093      * Trapless lookup: returns NULL for invalid page, I/O,
1094      * watchpoints, clean pages, etc.
1095      */
1096     mem = tlb_vaddr_to_host(env, toaddr, MMU_DATA_STORE, memidx);
1097 
1098 #ifndef CONFIG_USER_ONLY
1099     if (unlikely(!mem)) {
1100         /*
1101          * Slow-path: just do one byte write. This will handle the
1102          * watchpoint, invalid page, etc handling correctly.
1103          * For clean code pages, the next iteration will see
1104          * the page dirty and will use the fast path.
1105          */
1106         cpu_stb_mmuidx_ra(env, toaddr, data, memidx, ra);
1107         return 1;
1108     }
1109 #endif
1110     /* Easy case: just memset the host memory */
1111     memset(mem, data, setsize);
1112     return setsize;
1113 }
1114 
1115 /*
1116  * Similar, but setting tags. The architecture requires us to do this
1117  * in 16-byte chunks. SETP accesses are not tag checked; they set
1118  * the tags.
1119  */
1120 static uint64_t set_step_tags(CPUARMState *env, uint64_t toaddr,
1121                               uint64_t setsize, uint32_t data, int memidx,
1122                               uint32_t *mtedesc, uintptr_t ra)
1123 {
1124     void *mem;
1125     uint64_t cleanaddr;
1126 
1127     setsize = MIN(setsize, page_limit(toaddr));
1128 
1129     cleanaddr = useronly_clean_ptr(toaddr);
1130     /*
1131      * Trapless lookup: returns NULL for invalid page, I/O,
1132      * watchpoints, clean pages, etc.
1133      */
1134     mem = tlb_vaddr_to_host(env, cleanaddr, MMU_DATA_STORE, memidx);
1135 
1136 #ifndef CONFIG_USER_ONLY
1137     if (unlikely(!mem)) {
1138         /*
1139          * Slow-path: just do one write. This will handle the
1140          * watchpoint, invalid page, etc handling correctly.
1141          * The architecture requires that we do 16 bytes at a time,
1142          * and we know both ptr and size are 16 byte aligned.
1143          * For clean code pages, the next iteration will see
1144          * the page dirty and will use the fast path.
1145          */
1146         uint64_t repldata = data * 0x0101010101010101ULL;
1147         MemOpIdx oi16 = make_memop_idx(MO_TE | MO_128, memidx);
1148         cpu_st16_mmu(env, toaddr, int128_make128(repldata, repldata), oi16, ra);
1149         mte_mops_set_tags(env, toaddr, 16, *mtedesc);
1150         return 16;
1151     }
1152 #endif
1153     /* Easy case: just memset the host memory */
1154     memset(mem, data, setsize);
1155     mte_mops_set_tags(env, toaddr, setsize, *mtedesc);
1156     return setsize;
1157 }
1158 
1159 typedef uint64_t StepFn(CPUARMState *env, uint64_t toaddr,
1160                         uint64_t setsize, uint32_t data,
1161                         int memidx, uint32_t *mtedesc, uintptr_t ra);
1162 
1163 /* Extract register numbers from a MOPS exception syndrome value */
1164 static int mops_destreg(uint32_t syndrome)
1165 {
1166     return extract32(syndrome, 10, 5);
1167 }
1168 
1169 static int mops_srcreg(uint32_t syndrome)
1170 {
1171     return extract32(syndrome, 5, 5);
1172 }
1173 
1174 static int mops_sizereg(uint32_t syndrome)
1175 {
1176     return extract32(syndrome, 0, 5);
1177 }
1178 
1179 /*
1180  * Return true if TCMA and TBI bits mean we need to do MTE checks.
1181  * We only need to do this once per MOPS insn, not for every page.
1182  */
1183 static bool mte_checks_needed(uint64_t ptr, uint32_t desc)
1184 {
1185     int bit55 = extract64(ptr, 55, 1);
1186 
1187     /*
1188      * Note that tbi_check() returns true for "access checked" but
1189      * tcma_check() returns true for "access unchecked".
1190      */
1191     if (!tbi_check(desc, bit55)) {
1192         return false;
1193     }
1194     return !tcma_check(desc, bit55, allocation_tag_from_addr(ptr));
1195 }
1196 
1197 /* Take an exception if the SETG addr/size are not granule aligned */
1198 static void check_setg_alignment(CPUARMState *env, uint64_t ptr, uint64_t size,
1199                                  uint32_t memidx, uintptr_t ra)
1200 {
1201     if ((size != 0 && !QEMU_IS_ALIGNED(ptr, TAG_GRANULE)) ||
1202         !QEMU_IS_ALIGNED(size, TAG_GRANULE)) {
1203         arm_cpu_do_unaligned_access(env_cpu(env), ptr, MMU_DATA_STORE,
1204                                     memidx, ra);
1205 
1206     }
1207 }
1208 
1209 /*
1210  * For the Memory Set operation, our implementation chooses
1211  * always to use "option A", where we update Xd to the final
1212  * address in the SETP insn, and set Xn to be -(bytes remaining).
1213  * On SETM and SETE insns we only need update Xn.
1214  *
1215  * @env: CPU
1216  * @syndrome: syndrome value for mismatch exceptions
1217  * (also contains the register numbers we need to use)
1218  * @mtedesc: MTE descriptor word
1219  * @stepfn: function which does a single part of the set operation
1220  * @is_setg: true if this is the tag-setting SETG variant
1221  */
1222 static void do_setp(CPUARMState *env, uint32_t syndrome, uint32_t mtedesc,
1223                     StepFn *stepfn, bool is_setg, uintptr_t ra)
1224 {
1225     /* Prologue: we choose to do up to the next page boundary */
1226     int rd = mops_destreg(syndrome);
1227     int rs = mops_srcreg(syndrome);
1228     int rn = mops_sizereg(syndrome);
1229     uint8_t data = env->xregs[rs];
1230     uint32_t memidx = FIELD_EX32(mtedesc, MTEDESC, MIDX);
1231     uint64_t toaddr = env->xregs[rd];
1232     uint64_t setsize = env->xregs[rn];
1233     uint64_t stagesetsize, step;
1234 
1235     check_mops_enabled(env, ra);
1236 
1237     if (setsize > INT64_MAX) {
1238         setsize = INT64_MAX;
1239         if (is_setg) {
1240             setsize &= ~0xf;
1241         }
1242     }
1243 
1244     if (unlikely(is_setg)) {
1245         check_setg_alignment(env, toaddr, setsize, memidx, ra);
1246     } else if (!mte_checks_needed(toaddr, mtedesc)) {
1247         mtedesc = 0;
1248     }
1249 
1250     stagesetsize = MIN(setsize, page_limit(toaddr));
1251     while (stagesetsize) {
1252         env->xregs[rd] = toaddr;
1253         env->xregs[rn] = setsize;
1254         step = stepfn(env, toaddr, stagesetsize, data, memidx, &mtedesc, ra);
1255         toaddr += step;
1256         setsize -= step;
1257         stagesetsize -= step;
1258     }
1259     /* Insn completed, so update registers to the Option A format */
1260     env->xregs[rd] = toaddr + setsize;
1261     env->xregs[rn] = -setsize;
1262 
1263     /* Set NZCV = 0000 to indicate we are an Option A implementation */
1264     env->NF = 0;
1265     env->ZF = 1; /* our env->ZF encoding is inverted */
1266     env->CF = 0;
1267     env->VF = 0;
1268     return;
1269 }
1270 
1271 void HELPER(setp)(CPUARMState *env, uint32_t syndrome, uint32_t mtedesc)
1272 {
1273     do_setp(env, syndrome, mtedesc, set_step, false, GETPC());
1274 }
1275 
1276 void HELPER(setgp)(CPUARMState *env, uint32_t syndrome, uint32_t mtedesc)
1277 {
1278     do_setp(env, syndrome, mtedesc, set_step_tags, true, GETPC());
1279 }
1280 
1281 static void do_setm(CPUARMState *env, uint32_t syndrome, uint32_t mtedesc,
1282                     StepFn *stepfn, bool is_setg, uintptr_t ra)
1283 {
1284     /* Main: we choose to do all the full-page chunks */
1285     CPUState *cs = env_cpu(env);
1286     int rd = mops_destreg(syndrome);
1287     int rs = mops_srcreg(syndrome);
1288     int rn = mops_sizereg(syndrome);
1289     uint8_t data = env->xregs[rs];
1290     uint64_t toaddr = env->xregs[rd] + env->xregs[rn];
1291     uint64_t setsize = -env->xregs[rn];
1292     uint32_t memidx = FIELD_EX32(mtedesc, MTEDESC, MIDX);
1293     uint64_t step, stagesetsize;
1294 
1295     check_mops_enabled(env, ra);
1296 
1297     /*
1298      * We're allowed to NOP out "no data to copy" before the consistency
1299      * checks; we choose to do so.
1300      */
1301     if (env->xregs[rn] == 0) {
1302         return;
1303     }
1304 
1305     check_mops_wrong_option(env, syndrome, ra);
1306 
1307     /*
1308      * Our implementation will work fine even if we have an unaligned
1309      * destination address, and because we update Xn every time around
1310      * the loop below and the return value from stepfn() may be less
1311      * than requested, we might find toaddr is unaligned. So we don't
1312      * have an IMPDEF check for alignment here.
1313      */
1314 
1315     if (unlikely(is_setg)) {
1316         check_setg_alignment(env, toaddr, setsize, memidx, ra);
1317     } else if (!mte_checks_needed(toaddr, mtedesc)) {
1318         mtedesc = 0;
1319     }
1320 
1321     /* Do the actual memset: we leave the last partial page to SETE */
1322     stagesetsize = setsize & TARGET_PAGE_MASK;
1323     while (stagesetsize > 0) {
1324         step = stepfn(env, toaddr, setsize, data, memidx, &mtedesc, ra);
1325         toaddr += step;
1326         setsize -= step;
1327         stagesetsize -= step;
1328         env->xregs[rn] = -setsize;
1329         if (stagesetsize > 0 && unlikely(cpu_loop_exit_requested(cs))) {
1330             cpu_loop_exit_restore(cs, ra);
1331         }
1332     }
1333 }
1334 
1335 void HELPER(setm)(CPUARMState *env, uint32_t syndrome, uint32_t mtedesc)
1336 {
1337     do_setm(env, syndrome, mtedesc, set_step, false, GETPC());
1338 }
1339 
1340 void HELPER(setgm)(CPUARMState *env, uint32_t syndrome, uint32_t mtedesc)
1341 {
1342     do_setm(env, syndrome, mtedesc, set_step_tags, true, GETPC());
1343 }
1344 
1345 static void do_sete(CPUARMState *env, uint32_t syndrome, uint32_t mtedesc,
1346                     StepFn *stepfn, bool is_setg, uintptr_t ra)
1347 {
1348     /* Epilogue: do the last partial page */
1349     int rd = mops_destreg(syndrome);
1350     int rs = mops_srcreg(syndrome);
1351     int rn = mops_sizereg(syndrome);
1352     uint8_t data = env->xregs[rs];
1353     uint64_t toaddr = env->xregs[rd] + env->xregs[rn];
1354     uint64_t setsize = -env->xregs[rn];
1355     uint32_t memidx = FIELD_EX32(mtedesc, MTEDESC, MIDX);
1356     uint64_t step;
1357 
1358     check_mops_enabled(env, ra);
1359 
1360     /*
1361      * We're allowed to NOP out "no data to copy" before the consistency
1362      * checks; we choose to do so.
1363      */
1364     if (setsize == 0) {
1365         return;
1366     }
1367 
1368     check_mops_wrong_option(env, syndrome, ra);
1369 
1370     /*
1371      * Our implementation has no address alignment requirements, but
1372      * we do want to enforce the "less than a page" size requirement,
1373      * so we don't need to have the "check for interrupts" here.
1374      */
1375     if (setsize >= TARGET_PAGE_SIZE) {
1376         raise_exception_ra(env, EXCP_UDEF, syndrome,
1377                            mops_mismatch_exception_target_el(env), ra);
1378     }
1379 
1380     if (unlikely(is_setg)) {
1381         check_setg_alignment(env, toaddr, setsize, memidx, ra);
1382     } else if (!mte_checks_needed(toaddr, mtedesc)) {
1383         mtedesc = 0;
1384     }
1385 
1386     /* Do the actual memset */
1387     while (setsize > 0) {
1388         step = stepfn(env, toaddr, setsize, data, memidx, &mtedesc, ra);
1389         toaddr += step;
1390         setsize -= step;
1391         env->xregs[rn] = -setsize;
1392     }
1393 }
1394 
1395 void HELPER(sete)(CPUARMState *env, uint32_t syndrome, uint32_t mtedesc)
1396 {
1397     do_sete(env, syndrome, mtedesc, set_step, false, GETPC());
1398 }
1399 
1400 void HELPER(setge)(CPUARMState *env, uint32_t syndrome, uint32_t mtedesc)
1401 {
1402     do_sete(env, syndrome, mtedesc, set_step_tags, true, GETPC());
1403 }
1404 
1405 /*
1406  * Perform part of a memory copy from the guest memory at fromaddr
1407  * and extending for copysize bytes, to the guest memory at
1408  * toaddr. Both addreses are dirty.
1409  *
1410  * Returns the number of bytes actually set, which might be less than
1411  * copysize; the caller should loop until the whole copy has been done.
1412  * The caller should ensure that the guest registers are correct
1413  * for the possibility that the first byte of the copy encounters
1414  * an exception or watchpoint. We guarantee not to take any faults
1415  * for bytes other than the first.
1416  */
1417 static uint64_t copy_step(CPUARMState *env, uint64_t toaddr, uint64_t fromaddr,
1418                           uint64_t copysize, int wmemidx, int rmemidx,
1419                           uint32_t *wdesc, uint32_t *rdesc, uintptr_t ra)
1420 {
1421     void *rmem;
1422     void *wmem;
1423 
1424     /* Don't cross a page boundary on either source or destination */
1425     copysize = MIN(copysize, page_limit(toaddr));
1426     copysize = MIN(copysize, page_limit(fromaddr));
1427     /*
1428      * Handle MTE tag checks: either handle the tag mismatch for byte 0,
1429      * or else copy up to but not including the byte with the mismatch.
1430      */
1431     if (*rdesc) {
1432         uint64_t mtesize = mte_mops_probe(env, fromaddr, copysize, *rdesc);
1433         if (mtesize == 0) {
1434             mte_check_fail(env, *rdesc, fromaddr, ra);
1435             *rdesc = 0;
1436         } else {
1437             copysize = MIN(copysize, mtesize);
1438         }
1439     }
1440     if (*wdesc) {
1441         uint64_t mtesize = mte_mops_probe(env, toaddr, copysize, *wdesc);
1442         if (mtesize == 0) {
1443             mte_check_fail(env, *wdesc, toaddr, ra);
1444             *wdesc = 0;
1445         } else {
1446             copysize = MIN(copysize, mtesize);
1447         }
1448     }
1449 
1450     toaddr = useronly_clean_ptr(toaddr);
1451     fromaddr = useronly_clean_ptr(fromaddr);
1452     /* Trapless lookup of whether we can get a host memory pointer */
1453     wmem = tlb_vaddr_to_host(env, toaddr, MMU_DATA_STORE, wmemidx);
1454     rmem = tlb_vaddr_to_host(env, fromaddr, MMU_DATA_LOAD, rmemidx);
1455 
1456 #ifndef CONFIG_USER_ONLY
1457     /*
1458      * If we don't have host memory for both source and dest then just
1459      * do a single byte copy. This will handle watchpoints, invalid pages,
1460      * etc correctly. For clean code pages, the next iteration will see
1461      * the page dirty and will use the fast path.
1462      */
1463     if (unlikely(!rmem || !wmem)) {
1464         uint8_t byte;
1465         if (rmem) {
1466             byte = *(uint8_t *)rmem;
1467         } else {
1468             byte = cpu_ldub_mmuidx_ra(env, fromaddr, rmemidx, ra);
1469         }
1470         if (wmem) {
1471             *(uint8_t *)wmem = byte;
1472         } else {
1473             cpu_stb_mmuidx_ra(env, toaddr, byte, wmemidx, ra);
1474         }
1475         return 1;
1476     }
1477 #endif
1478     /* Easy case: just memmove the host memory */
1479     memmove(wmem, rmem, copysize);
1480     return copysize;
1481 }
1482 
1483 /*
1484  * Do part of a backwards memory copy. Here toaddr and fromaddr point
1485  * to the *last* byte to be copied.
1486  */
1487 static uint64_t copy_step_rev(CPUARMState *env, uint64_t toaddr,
1488                               uint64_t fromaddr,
1489                               uint64_t copysize, int wmemidx, int rmemidx,
1490                               uint32_t *wdesc, uint32_t *rdesc, uintptr_t ra)
1491 {
1492     void *rmem;
1493     void *wmem;
1494 
1495     /* Don't cross a page boundary on either source or destination */
1496     copysize = MIN(copysize, page_limit_rev(toaddr));
1497     copysize = MIN(copysize, page_limit_rev(fromaddr));
1498 
1499     /*
1500      * Handle MTE tag checks: either handle the tag mismatch for byte 0,
1501      * or else copy up to but not including the byte with the mismatch.
1502      */
1503     if (*rdesc) {
1504         uint64_t mtesize = mte_mops_probe_rev(env, fromaddr, copysize, *rdesc);
1505         if (mtesize == 0) {
1506             mte_check_fail(env, *rdesc, fromaddr, ra);
1507             *rdesc = 0;
1508         } else {
1509             copysize = MIN(copysize, mtesize);
1510         }
1511     }
1512     if (*wdesc) {
1513         uint64_t mtesize = mte_mops_probe_rev(env, toaddr, copysize, *wdesc);
1514         if (mtesize == 0) {
1515             mte_check_fail(env, *wdesc, toaddr, ra);
1516             *wdesc = 0;
1517         } else {
1518             copysize = MIN(copysize, mtesize);
1519         }
1520     }
1521 
1522     toaddr = useronly_clean_ptr(toaddr);
1523     fromaddr = useronly_clean_ptr(fromaddr);
1524     /* Trapless lookup of whether we can get a host memory pointer */
1525     wmem = tlb_vaddr_to_host(env, toaddr, MMU_DATA_STORE, wmemidx);
1526     rmem = tlb_vaddr_to_host(env, fromaddr, MMU_DATA_LOAD, rmemidx);
1527 
1528 #ifndef CONFIG_USER_ONLY
1529     /*
1530      * If we don't have host memory for both source and dest then just
1531      * do a single byte copy. This will handle watchpoints, invalid pages,
1532      * etc correctly. For clean code pages, the next iteration will see
1533      * the page dirty and will use the fast path.
1534      */
1535     if (unlikely(!rmem || !wmem)) {
1536         uint8_t byte;
1537         if (rmem) {
1538             byte = *(uint8_t *)rmem;
1539         } else {
1540             byte = cpu_ldub_mmuidx_ra(env, fromaddr, rmemidx, ra);
1541         }
1542         if (wmem) {
1543             *(uint8_t *)wmem = byte;
1544         } else {
1545             cpu_stb_mmuidx_ra(env, toaddr, byte, wmemidx, ra);
1546         }
1547         return 1;
1548     }
1549 #endif
1550     /*
1551      * Easy case: just memmove the host memory. Note that wmem and
1552      * rmem here point to the *last* byte to copy.
1553      */
1554     memmove(wmem - (copysize - 1), rmem - (copysize - 1), copysize);
1555     return copysize;
1556 }
1557 
1558 /*
1559  * for the Memory Copy operation, our implementation chooses always
1560  * to use "option A", where we update Xd and Xs to the final addresses
1561  * in the CPYP insn, and then in CPYM and CPYE only need to update Xn.
1562  *
1563  * @env: CPU
1564  * @syndrome: syndrome value for mismatch exceptions
1565  * (also contains the register numbers we need to use)
1566  * @wdesc: MTE descriptor for the writes (destination)
1567  * @rdesc: MTE descriptor for the reads (source)
1568  * @move: true if this is CPY (memmove), false for CPYF (memcpy forwards)
1569  */
1570 static void do_cpyp(CPUARMState *env, uint32_t syndrome, uint32_t wdesc,
1571                     uint32_t rdesc, uint32_t move, uintptr_t ra)
1572 {
1573     int rd = mops_destreg(syndrome);
1574     int rs = mops_srcreg(syndrome);
1575     int rn = mops_sizereg(syndrome);
1576     uint32_t rmemidx = FIELD_EX32(rdesc, MTEDESC, MIDX);
1577     uint32_t wmemidx = FIELD_EX32(wdesc, MTEDESC, MIDX);
1578     bool forwards = true;
1579     uint64_t toaddr = env->xregs[rd];
1580     uint64_t fromaddr = env->xregs[rs];
1581     uint64_t copysize = env->xregs[rn];
1582     uint64_t stagecopysize, step;
1583 
1584     check_mops_enabled(env, ra);
1585 
1586 
1587     if (move) {
1588         /*
1589          * Copy backwards if necessary. The direction for a non-overlapping
1590          * copy is IMPDEF; we choose forwards.
1591          */
1592         if (copysize > 0x007FFFFFFFFFFFFFULL) {
1593             copysize = 0x007FFFFFFFFFFFFFULL;
1594         }
1595         uint64_t fs = extract64(fromaddr, 0, 56);
1596         uint64_t ts = extract64(toaddr, 0, 56);
1597         uint64_t fe = extract64(fromaddr + copysize, 0, 56);
1598 
1599         if (fs < ts && fe > ts) {
1600             forwards = false;
1601         }
1602     } else {
1603         if (copysize > INT64_MAX) {
1604             copysize = INT64_MAX;
1605         }
1606     }
1607 
1608     if (!mte_checks_needed(fromaddr, rdesc)) {
1609         rdesc = 0;
1610     }
1611     if (!mte_checks_needed(toaddr, wdesc)) {
1612         wdesc = 0;
1613     }
1614 
1615     if (forwards) {
1616         stagecopysize = MIN(copysize, page_limit(toaddr));
1617         stagecopysize = MIN(stagecopysize, page_limit(fromaddr));
1618         while (stagecopysize) {
1619             env->xregs[rd] = toaddr;
1620             env->xregs[rs] = fromaddr;
1621             env->xregs[rn] = copysize;
1622             step = copy_step(env, toaddr, fromaddr, stagecopysize,
1623                              wmemidx, rmemidx, &wdesc, &rdesc, ra);
1624             toaddr += step;
1625             fromaddr += step;
1626             copysize -= step;
1627             stagecopysize -= step;
1628         }
1629         /* Insn completed, so update registers to the Option A format */
1630         env->xregs[rd] = toaddr + copysize;
1631         env->xregs[rs] = fromaddr + copysize;
1632         env->xregs[rn] = -copysize;
1633     } else {
1634         /*
1635          * In a reverse copy the to and from addrs in Xs and Xd are the start
1636          * of the range, but it's more convenient for us to work with pointers
1637          * to the last byte being copied.
1638          */
1639         toaddr += copysize - 1;
1640         fromaddr += copysize - 1;
1641         stagecopysize = MIN(copysize, page_limit_rev(toaddr));
1642         stagecopysize = MIN(stagecopysize, page_limit_rev(fromaddr));
1643         while (stagecopysize) {
1644             env->xregs[rn] = copysize;
1645             step = copy_step_rev(env, toaddr, fromaddr, stagecopysize,
1646                                  wmemidx, rmemidx, &wdesc, &rdesc, ra);
1647             copysize -= step;
1648             stagecopysize -= step;
1649             toaddr -= step;
1650             fromaddr -= step;
1651         }
1652         /*
1653          * Insn completed, so update registers to the Option A format.
1654          * For a reverse copy this is no different to the CPYP input format.
1655          */
1656         env->xregs[rn] = copysize;
1657     }
1658 
1659     /* Set NZCV = 0000 to indicate we are an Option A implementation */
1660     env->NF = 0;
1661     env->ZF = 1; /* our env->ZF encoding is inverted */
1662     env->CF = 0;
1663     env->VF = 0;
1664     return;
1665 }
1666 
1667 void HELPER(cpyp)(CPUARMState *env, uint32_t syndrome, uint32_t wdesc,
1668                   uint32_t rdesc)
1669 {
1670     do_cpyp(env, syndrome, wdesc, rdesc, true, GETPC());
1671 }
1672 
1673 void HELPER(cpyfp)(CPUARMState *env, uint32_t syndrome, uint32_t wdesc,
1674                    uint32_t rdesc)
1675 {
1676     do_cpyp(env, syndrome, wdesc, rdesc, false, GETPC());
1677 }
1678 
1679 static void do_cpym(CPUARMState *env, uint32_t syndrome, uint32_t wdesc,
1680                     uint32_t rdesc, uint32_t move, uintptr_t ra)
1681 {
1682     /* Main: we choose to copy until less than a page remaining */
1683     CPUState *cs = env_cpu(env);
1684     int rd = mops_destreg(syndrome);
1685     int rs = mops_srcreg(syndrome);
1686     int rn = mops_sizereg(syndrome);
1687     uint32_t rmemidx = FIELD_EX32(rdesc, MTEDESC, MIDX);
1688     uint32_t wmemidx = FIELD_EX32(wdesc, MTEDESC, MIDX);
1689     bool forwards = true;
1690     uint64_t toaddr, fromaddr, copysize, step;
1691 
1692     check_mops_enabled(env, ra);
1693 
1694     /* We choose to NOP out "no data to copy" before consistency checks */
1695     if (env->xregs[rn] == 0) {
1696         return;
1697     }
1698 
1699     check_mops_wrong_option(env, syndrome, ra);
1700 
1701     if (move) {
1702         forwards = (int64_t)env->xregs[rn] < 0;
1703     }
1704 
1705     if (forwards) {
1706         toaddr = env->xregs[rd] + env->xregs[rn];
1707         fromaddr = env->xregs[rs] + env->xregs[rn];
1708         copysize = -env->xregs[rn];
1709     } else {
1710         copysize = env->xregs[rn];
1711         /* This toaddr and fromaddr point to the *last* byte to copy */
1712         toaddr = env->xregs[rd] + copysize - 1;
1713         fromaddr = env->xregs[rs] + copysize - 1;
1714     }
1715 
1716     if (!mte_checks_needed(fromaddr, rdesc)) {
1717         rdesc = 0;
1718     }
1719     if (!mte_checks_needed(toaddr, wdesc)) {
1720         wdesc = 0;
1721     }
1722 
1723     /* Our implementation has no particular parameter requirements for CPYM */
1724 
1725     /* Do the actual memmove */
1726     if (forwards) {
1727         while (copysize >= TARGET_PAGE_SIZE) {
1728             step = copy_step(env, toaddr, fromaddr, copysize,
1729                              wmemidx, rmemidx, &wdesc, &rdesc, ra);
1730             toaddr += step;
1731             fromaddr += step;
1732             copysize -= step;
1733             env->xregs[rn] = -copysize;
1734             if (copysize >= TARGET_PAGE_SIZE &&
1735                 unlikely(cpu_loop_exit_requested(cs))) {
1736                 cpu_loop_exit_restore(cs, ra);
1737             }
1738         }
1739     } else {
1740         while (copysize >= TARGET_PAGE_SIZE) {
1741             step = copy_step_rev(env, toaddr, fromaddr, copysize,
1742                                  wmemidx, rmemidx, &wdesc, &rdesc, ra);
1743             toaddr -= step;
1744             fromaddr -= step;
1745             copysize -= step;
1746             env->xregs[rn] = copysize;
1747             if (copysize >= TARGET_PAGE_SIZE &&
1748                 unlikely(cpu_loop_exit_requested(cs))) {
1749                 cpu_loop_exit_restore(cs, ra);
1750             }
1751         }
1752     }
1753 }
1754 
1755 void HELPER(cpym)(CPUARMState *env, uint32_t syndrome, uint32_t wdesc,
1756                   uint32_t rdesc)
1757 {
1758     do_cpym(env, syndrome, wdesc, rdesc, true, GETPC());
1759 }
1760 
1761 void HELPER(cpyfm)(CPUARMState *env, uint32_t syndrome, uint32_t wdesc,
1762                    uint32_t rdesc)
1763 {
1764     do_cpym(env, syndrome, wdesc, rdesc, false, GETPC());
1765 }
1766 
1767 static void do_cpye(CPUARMState *env, uint32_t syndrome, uint32_t wdesc,
1768                     uint32_t rdesc, uint32_t move, uintptr_t ra)
1769 {
1770     /* Epilogue: do the last partial page */
1771     int rd = mops_destreg(syndrome);
1772     int rs = mops_srcreg(syndrome);
1773     int rn = mops_sizereg(syndrome);
1774     uint32_t rmemidx = FIELD_EX32(rdesc, MTEDESC, MIDX);
1775     uint32_t wmemidx = FIELD_EX32(wdesc, MTEDESC, MIDX);
1776     bool forwards = true;
1777     uint64_t toaddr, fromaddr, copysize, step;
1778 
1779     check_mops_enabled(env, ra);
1780 
1781     /* We choose to NOP out "no data to copy" before consistency checks */
1782     if (env->xregs[rn] == 0) {
1783         return;
1784     }
1785 
1786     check_mops_wrong_option(env, syndrome, ra);
1787 
1788     if (move) {
1789         forwards = (int64_t)env->xregs[rn] < 0;
1790     }
1791 
1792     if (forwards) {
1793         toaddr = env->xregs[rd] + env->xregs[rn];
1794         fromaddr = env->xregs[rs] + env->xregs[rn];
1795         copysize = -env->xregs[rn];
1796     } else {
1797         copysize = env->xregs[rn];
1798         /* This toaddr and fromaddr point to the *last* byte to copy */
1799         toaddr = env->xregs[rd] + copysize - 1;
1800         fromaddr = env->xregs[rs] + copysize - 1;
1801     }
1802 
1803     if (!mte_checks_needed(fromaddr, rdesc)) {
1804         rdesc = 0;
1805     }
1806     if (!mte_checks_needed(toaddr, wdesc)) {
1807         wdesc = 0;
1808     }
1809 
1810     /* Check the size; we don't want to have do a check-for-interrupts */
1811     if (copysize >= TARGET_PAGE_SIZE) {
1812         raise_exception_ra(env, EXCP_UDEF, syndrome,
1813                            mops_mismatch_exception_target_el(env), ra);
1814     }
1815 
1816     /* Do the actual memmove */
1817     if (forwards) {
1818         while (copysize > 0) {
1819             step = copy_step(env, toaddr, fromaddr, copysize,
1820                              wmemidx, rmemidx, &wdesc, &rdesc, ra);
1821             toaddr += step;
1822             fromaddr += step;
1823             copysize -= step;
1824             env->xregs[rn] = -copysize;
1825         }
1826     } else {
1827         while (copysize > 0) {
1828             step = copy_step_rev(env, toaddr, fromaddr, copysize,
1829                                  wmemidx, rmemidx, &wdesc, &rdesc, ra);
1830             toaddr -= step;
1831             fromaddr -= step;
1832             copysize -= step;
1833             env->xregs[rn] = copysize;
1834         }
1835     }
1836 }
1837 
1838 void HELPER(cpye)(CPUARMState *env, uint32_t syndrome, uint32_t wdesc,
1839                   uint32_t rdesc)
1840 {
1841     do_cpye(env, syndrome, wdesc, rdesc, true, GETPC());
1842 }
1843 
1844 void HELPER(cpyfe)(CPUARMState *env, uint32_t syndrome, uint32_t wdesc,
1845                    uint32_t rdesc)
1846 {
1847     do_cpye(env, syndrome, wdesc, rdesc, false, GETPC());
1848 }
1849