xref: /openbmc/qemu/tcg/tcg-op-gvec.c (revision 06831001)
1 /*
2  * Generic vector operation expansion
3  *
4  * Copyright (c) 2018 Linaro
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 "tcg/tcg.h"
22 #include "tcg/tcg-temp-internal.h"
23 #include "tcg/tcg-op-common.h"
24 #include "tcg/tcg-op-gvec-common.h"
25 #include "tcg/tcg-gvec-desc.h"
26 
27 #define MAX_UNROLL  4
28 
29 #ifdef CONFIG_DEBUG_TCG
30 static const TCGOpcode vecop_list_empty[1] = { 0 };
31 #else
32 #define vecop_list_empty NULL
33 #endif
34 
35 
36 /* Verify vector size and alignment rules.  OFS should be the OR of all
37    of the operand offsets so that we can check them all at once.  */
38 static void check_size_align(uint32_t oprsz, uint32_t maxsz, uint32_t ofs)
39 {
40     uint32_t max_align;
41 
42     switch (oprsz) {
43     case 8:
44     case 16:
45     case 32:
46         tcg_debug_assert(oprsz <= maxsz);
47         break;
48     default:
49         tcg_debug_assert(oprsz == maxsz);
50         break;
51     }
52     tcg_debug_assert(maxsz <= (8 << SIMD_MAXSZ_BITS));
53 
54     max_align = maxsz >= 16 ? 15 : 7;
55     tcg_debug_assert((maxsz & max_align) == 0);
56     tcg_debug_assert((ofs & max_align) == 0);
57 }
58 
59 /* Verify vector overlap rules for two operands.  */
60 static void check_overlap_2(uint32_t d, uint32_t a, uint32_t s)
61 {
62     tcg_debug_assert(d == a || d + s <= a || a + s <= d);
63 }
64 
65 /* Verify vector overlap rules for three operands.  */
66 static void check_overlap_3(uint32_t d, uint32_t a, uint32_t b, uint32_t s)
67 {
68     check_overlap_2(d, a, s);
69     check_overlap_2(d, b, s);
70     check_overlap_2(a, b, s);
71 }
72 
73 /* Verify vector overlap rules for four operands.  */
74 static void check_overlap_4(uint32_t d, uint32_t a, uint32_t b,
75                             uint32_t c, uint32_t s)
76 {
77     check_overlap_2(d, a, s);
78     check_overlap_2(d, b, s);
79     check_overlap_2(d, c, s);
80     check_overlap_2(a, b, s);
81     check_overlap_2(a, c, s);
82     check_overlap_2(b, c, s);
83 }
84 
85 /* Create a descriptor from components.  */
86 uint32_t simd_desc(uint32_t oprsz, uint32_t maxsz, int32_t data)
87 {
88     uint32_t desc = 0;
89 
90     check_size_align(oprsz, maxsz, 0);
91     tcg_debug_assert(data == sextract32(data, 0, SIMD_DATA_BITS));
92 
93     oprsz = (oprsz / 8) - 1;
94     maxsz = (maxsz / 8) - 1;
95 
96     /*
97      * We have just asserted in check_size_align that either
98      * oprsz is {8,16,32} or matches maxsz.  Encode the final
99      * case with '2', as that would otherwise map to 24.
100      */
101     if (oprsz == maxsz) {
102         oprsz = 2;
103     }
104 
105     desc = deposit32(desc, SIMD_OPRSZ_SHIFT, SIMD_OPRSZ_BITS, oprsz);
106     desc = deposit32(desc, SIMD_MAXSZ_SHIFT, SIMD_MAXSZ_BITS, maxsz);
107     desc = deposit32(desc, SIMD_DATA_SHIFT, SIMD_DATA_BITS, data);
108 
109     return desc;
110 }
111 
112 /* Generate a call to a gvec-style helper with two vector operands.  */
113 void tcg_gen_gvec_2_ool(uint32_t dofs, uint32_t aofs,
114                         uint32_t oprsz, uint32_t maxsz, int32_t data,
115                         gen_helper_gvec_2 *fn)
116 {
117     TCGv_ptr a0, a1;
118     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
119 
120     a0 = tcg_temp_ebb_new_ptr();
121     a1 = tcg_temp_ebb_new_ptr();
122 
123     tcg_gen_addi_ptr(a0, cpu_env, dofs);
124     tcg_gen_addi_ptr(a1, cpu_env, aofs);
125 
126     fn(a0, a1, desc);
127 
128     tcg_temp_free_ptr(a0);
129     tcg_temp_free_ptr(a1);
130 }
131 
132 /* Generate a call to a gvec-style helper with two vector operands
133    and one scalar operand.  */
134 void tcg_gen_gvec_2i_ool(uint32_t dofs, uint32_t aofs, TCGv_i64 c,
135                          uint32_t oprsz, uint32_t maxsz, int32_t data,
136                          gen_helper_gvec_2i *fn)
137 {
138     TCGv_ptr a0, a1;
139     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
140 
141     a0 = tcg_temp_ebb_new_ptr();
142     a1 = tcg_temp_ebb_new_ptr();
143 
144     tcg_gen_addi_ptr(a0, cpu_env, dofs);
145     tcg_gen_addi_ptr(a1, cpu_env, aofs);
146 
147     fn(a0, a1, c, desc);
148 
149     tcg_temp_free_ptr(a0);
150     tcg_temp_free_ptr(a1);
151 }
152 
153 /* Generate a call to a gvec-style helper with three vector operands.  */
154 void tcg_gen_gvec_3_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
155                         uint32_t oprsz, uint32_t maxsz, int32_t data,
156                         gen_helper_gvec_3 *fn)
157 {
158     TCGv_ptr a0, a1, a2;
159     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
160 
161     a0 = tcg_temp_ebb_new_ptr();
162     a1 = tcg_temp_ebb_new_ptr();
163     a2 = tcg_temp_ebb_new_ptr();
164 
165     tcg_gen_addi_ptr(a0, cpu_env, dofs);
166     tcg_gen_addi_ptr(a1, cpu_env, aofs);
167     tcg_gen_addi_ptr(a2, cpu_env, bofs);
168 
169     fn(a0, a1, a2, desc);
170 
171     tcg_temp_free_ptr(a0);
172     tcg_temp_free_ptr(a1);
173     tcg_temp_free_ptr(a2);
174 }
175 
176 /* Generate a call to a gvec-style helper with four vector operands.  */
177 void tcg_gen_gvec_4_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
178                         uint32_t cofs, uint32_t oprsz, uint32_t maxsz,
179                         int32_t data, gen_helper_gvec_4 *fn)
180 {
181     TCGv_ptr a0, a1, a2, a3;
182     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
183 
184     a0 = tcg_temp_ebb_new_ptr();
185     a1 = tcg_temp_ebb_new_ptr();
186     a2 = tcg_temp_ebb_new_ptr();
187     a3 = tcg_temp_ebb_new_ptr();
188 
189     tcg_gen_addi_ptr(a0, cpu_env, dofs);
190     tcg_gen_addi_ptr(a1, cpu_env, aofs);
191     tcg_gen_addi_ptr(a2, cpu_env, bofs);
192     tcg_gen_addi_ptr(a3, cpu_env, cofs);
193 
194     fn(a0, a1, a2, a3, desc);
195 
196     tcg_temp_free_ptr(a0);
197     tcg_temp_free_ptr(a1);
198     tcg_temp_free_ptr(a2);
199     tcg_temp_free_ptr(a3);
200 }
201 
202 /* Generate a call to a gvec-style helper with five vector operands.  */
203 void tcg_gen_gvec_5_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
204                         uint32_t cofs, uint32_t xofs, uint32_t oprsz,
205                         uint32_t maxsz, int32_t data, gen_helper_gvec_5 *fn)
206 {
207     TCGv_ptr a0, a1, a2, a3, a4;
208     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
209 
210     a0 = tcg_temp_ebb_new_ptr();
211     a1 = tcg_temp_ebb_new_ptr();
212     a2 = tcg_temp_ebb_new_ptr();
213     a3 = tcg_temp_ebb_new_ptr();
214     a4 = tcg_temp_ebb_new_ptr();
215 
216     tcg_gen_addi_ptr(a0, cpu_env, dofs);
217     tcg_gen_addi_ptr(a1, cpu_env, aofs);
218     tcg_gen_addi_ptr(a2, cpu_env, bofs);
219     tcg_gen_addi_ptr(a3, cpu_env, cofs);
220     tcg_gen_addi_ptr(a4, cpu_env, xofs);
221 
222     fn(a0, a1, a2, a3, a4, desc);
223 
224     tcg_temp_free_ptr(a0);
225     tcg_temp_free_ptr(a1);
226     tcg_temp_free_ptr(a2);
227     tcg_temp_free_ptr(a3);
228     tcg_temp_free_ptr(a4);
229 }
230 
231 /* Generate a call to a gvec-style helper with three vector operands
232    and an extra pointer operand.  */
233 void tcg_gen_gvec_2_ptr(uint32_t dofs, uint32_t aofs,
234                         TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
235                         int32_t data, gen_helper_gvec_2_ptr *fn)
236 {
237     TCGv_ptr a0, a1;
238     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
239 
240     a0 = tcg_temp_ebb_new_ptr();
241     a1 = tcg_temp_ebb_new_ptr();
242 
243     tcg_gen_addi_ptr(a0, cpu_env, dofs);
244     tcg_gen_addi_ptr(a1, cpu_env, aofs);
245 
246     fn(a0, a1, ptr, desc);
247 
248     tcg_temp_free_ptr(a0);
249     tcg_temp_free_ptr(a1);
250 }
251 
252 /* Generate a call to a gvec-style helper with three vector operands
253    and an extra pointer operand.  */
254 void tcg_gen_gvec_3_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
255                         TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
256                         int32_t data, gen_helper_gvec_3_ptr *fn)
257 {
258     TCGv_ptr a0, a1, a2;
259     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
260 
261     a0 = tcg_temp_ebb_new_ptr();
262     a1 = tcg_temp_ebb_new_ptr();
263     a2 = tcg_temp_ebb_new_ptr();
264 
265     tcg_gen_addi_ptr(a0, cpu_env, dofs);
266     tcg_gen_addi_ptr(a1, cpu_env, aofs);
267     tcg_gen_addi_ptr(a2, cpu_env, bofs);
268 
269     fn(a0, a1, a2, ptr, desc);
270 
271     tcg_temp_free_ptr(a0);
272     tcg_temp_free_ptr(a1);
273     tcg_temp_free_ptr(a2);
274 }
275 
276 /* Generate a call to a gvec-style helper with four vector operands
277    and an extra pointer operand.  */
278 void tcg_gen_gvec_4_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
279                         uint32_t cofs, TCGv_ptr ptr, uint32_t oprsz,
280                         uint32_t maxsz, int32_t data,
281                         gen_helper_gvec_4_ptr *fn)
282 {
283     TCGv_ptr a0, a1, a2, a3;
284     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
285 
286     a0 = tcg_temp_ebb_new_ptr();
287     a1 = tcg_temp_ebb_new_ptr();
288     a2 = tcg_temp_ebb_new_ptr();
289     a3 = tcg_temp_ebb_new_ptr();
290 
291     tcg_gen_addi_ptr(a0, cpu_env, dofs);
292     tcg_gen_addi_ptr(a1, cpu_env, aofs);
293     tcg_gen_addi_ptr(a2, cpu_env, bofs);
294     tcg_gen_addi_ptr(a3, cpu_env, cofs);
295 
296     fn(a0, a1, a2, a3, ptr, desc);
297 
298     tcg_temp_free_ptr(a0);
299     tcg_temp_free_ptr(a1);
300     tcg_temp_free_ptr(a2);
301     tcg_temp_free_ptr(a3);
302 }
303 
304 /* Generate a call to a gvec-style helper with five vector operands
305    and an extra pointer operand.  */
306 void tcg_gen_gvec_5_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
307                         uint32_t cofs, uint32_t eofs, TCGv_ptr ptr,
308                         uint32_t oprsz, uint32_t maxsz, int32_t data,
309                         gen_helper_gvec_5_ptr *fn)
310 {
311     TCGv_ptr a0, a1, a2, a3, a4;
312     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
313 
314     a0 = tcg_temp_ebb_new_ptr();
315     a1 = tcg_temp_ebb_new_ptr();
316     a2 = tcg_temp_ebb_new_ptr();
317     a3 = tcg_temp_ebb_new_ptr();
318     a4 = tcg_temp_ebb_new_ptr();
319 
320     tcg_gen_addi_ptr(a0, cpu_env, dofs);
321     tcg_gen_addi_ptr(a1, cpu_env, aofs);
322     tcg_gen_addi_ptr(a2, cpu_env, bofs);
323     tcg_gen_addi_ptr(a3, cpu_env, cofs);
324     tcg_gen_addi_ptr(a4, cpu_env, eofs);
325 
326     fn(a0, a1, a2, a3, a4, ptr, desc);
327 
328     tcg_temp_free_ptr(a0);
329     tcg_temp_free_ptr(a1);
330     tcg_temp_free_ptr(a2);
331     tcg_temp_free_ptr(a3);
332     tcg_temp_free_ptr(a4);
333 }
334 
335 /* Return true if we want to implement something of OPRSZ bytes
336    in units of LNSZ.  This limits the expansion of inline code.  */
337 static inline bool check_size_impl(uint32_t oprsz, uint32_t lnsz)
338 {
339     uint32_t q, r;
340 
341     if (oprsz < lnsz) {
342         return false;
343     }
344 
345     q = oprsz / lnsz;
346     r = oprsz % lnsz;
347     tcg_debug_assert((r & 7) == 0);
348 
349     if (lnsz < 16) {
350         /* For sizes below 16, accept no remainder. */
351         if (r != 0) {
352             return false;
353         }
354     } else {
355         /*
356          * Recall that ARM SVE allows vector sizes that are not a
357          * power of 2, but always a multiple of 16.  The intent is
358          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
359          * In addition, expand_clr needs to handle a multiple of 8.
360          * Thus we can handle the tail with one more operation per
361          * diminishing power of 2.
362          */
363         q += ctpop32(r);
364     }
365 
366     return q <= MAX_UNROLL;
367 }
368 
369 static void expand_clr(uint32_t dofs, uint32_t maxsz);
370 
371 /* Duplicate C as per VECE.  */
372 uint64_t (dup_const)(unsigned vece, uint64_t c)
373 {
374     switch (vece) {
375     case MO_8:
376         return 0x0101010101010101ull * (uint8_t)c;
377     case MO_16:
378         return 0x0001000100010001ull * (uint16_t)c;
379     case MO_32:
380         return 0x0000000100000001ull * (uint32_t)c;
381     case MO_64:
382         return c;
383     default:
384         g_assert_not_reached();
385     }
386 }
387 
388 /* Duplicate IN into OUT as per VECE.  */
389 void tcg_gen_dup_i32(unsigned vece, TCGv_i32 out, TCGv_i32 in)
390 {
391     switch (vece) {
392     case MO_8:
393         tcg_gen_ext8u_i32(out, in);
394         tcg_gen_muli_i32(out, out, 0x01010101);
395         break;
396     case MO_16:
397         tcg_gen_deposit_i32(out, in, in, 16, 16);
398         break;
399     case MO_32:
400         tcg_gen_mov_i32(out, in);
401         break;
402     default:
403         g_assert_not_reached();
404     }
405 }
406 
407 void tcg_gen_dup_i64(unsigned vece, TCGv_i64 out, TCGv_i64 in)
408 {
409     switch (vece) {
410     case MO_8:
411         tcg_gen_ext8u_i64(out, in);
412         tcg_gen_muli_i64(out, out, 0x0101010101010101ull);
413         break;
414     case MO_16:
415         tcg_gen_ext16u_i64(out, in);
416         tcg_gen_muli_i64(out, out, 0x0001000100010001ull);
417         break;
418     case MO_32:
419         tcg_gen_deposit_i64(out, in, in, 32, 32);
420         break;
421     case MO_64:
422         tcg_gen_mov_i64(out, in);
423         break;
424     default:
425         g_assert_not_reached();
426     }
427 }
428 
429 /* Select a supported vector type for implementing an operation on SIZE
430  * bytes.  If OP is 0, assume that the real operation to be performed is
431  * required by all backends.  Otherwise, make sure than OP can be performed
432  * on elements of size VECE in the selected type.  Do not select V64 if
433  * PREFER_I64 is true.  Return 0 if no vector type is selected.
434  */
435 static TCGType choose_vector_type(const TCGOpcode *list, unsigned vece,
436                                   uint32_t size, bool prefer_i64)
437 {
438     /*
439      * Recall that ARM SVE allows vector sizes that are not a
440      * power of 2, but always a multiple of 16.  The intent is
441      * that e.g. size == 80 would be expanded with 2x32 + 1x16.
442      * It is hard to imagine a case in which v256 is supported
443      * but v128 is not, but check anyway.
444      * In addition, expand_clr needs to handle a multiple of 8.
445      */
446     if (TCG_TARGET_HAS_v256 &&
447         check_size_impl(size, 32) &&
448         tcg_can_emit_vecop_list(list, TCG_TYPE_V256, vece) &&
449         (!(size & 16) ||
450          (TCG_TARGET_HAS_v128 &&
451           tcg_can_emit_vecop_list(list, TCG_TYPE_V128, vece))) &&
452         (!(size & 8) ||
453          (TCG_TARGET_HAS_v64 &&
454           tcg_can_emit_vecop_list(list, TCG_TYPE_V64, vece)))) {
455         return TCG_TYPE_V256;
456     }
457     if (TCG_TARGET_HAS_v128 &&
458         check_size_impl(size, 16) &&
459         tcg_can_emit_vecop_list(list, TCG_TYPE_V128, vece) &&
460         (!(size & 8) ||
461          (TCG_TARGET_HAS_v64 &&
462           tcg_can_emit_vecop_list(list, TCG_TYPE_V64, vece)))) {
463         return TCG_TYPE_V128;
464     }
465     if (TCG_TARGET_HAS_v64 && !prefer_i64 && check_size_impl(size, 8)
466         && tcg_can_emit_vecop_list(list, TCG_TYPE_V64, vece)) {
467         return TCG_TYPE_V64;
468     }
469     return 0;
470 }
471 
472 static void do_dup_store(TCGType type, uint32_t dofs, uint32_t oprsz,
473                          uint32_t maxsz, TCGv_vec t_vec)
474 {
475     uint32_t i = 0;
476 
477     tcg_debug_assert(oprsz >= 8);
478 
479     /*
480      * This may be expand_clr for the tail of an operation, e.g.
481      * oprsz == 8 && maxsz == 64.  The first 8 bytes of this store
482      * are misaligned wrt the maximum vector size, so do that first.
483      */
484     if (dofs & 8) {
485         tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V64);
486         i += 8;
487     }
488 
489     switch (type) {
490     case TCG_TYPE_V256:
491         /*
492          * Recall that ARM SVE allows vector sizes that are not a
493          * power of 2, but always a multiple of 16.  The intent is
494          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
495          */
496         for (; i + 32 <= oprsz; i += 32) {
497             tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V256);
498         }
499         /* fallthru */
500     case TCG_TYPE_V128:
501         for (; i + 16 <= oprsz; i += 16) {
502             tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V128);
503         }
504         break;
505     case TCG_TYPE_V64:
506         for (; i < oprsz; i += 8) {
507             tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V64);
508         }
509         break;
510     default:
511         g_assert_not_reached();
512     }
513 
514     if (oprsz < maxsz) {
515         expand_clr(dofs + oprsz, maxsz - oprsz);
516     }
517 }
518 
519 /* Set OPRSZ bytes at DOFS to replications of IN_32, IN_64 or IN_C.
520  * Only one of IN_32 or IN_64 may be set;
521  * IN_C is used if IN_32 and IN_64 are unset.
522  */
523 static void do_dup(unsigned vece, uint32_t dofs, uint32_t oprsz,
524                    uint32_t maxsz, TCGv_i32 in_32, TCGv_i64 in_64,
525                    uint64_t in_c)
526 {
527     TCGType type;
528     TCGv_i64 t_64;
529     TCGv_i32 t_32, t_desc;
530     TCGv_ptr t_ptr;
531     uint32_t i;
532 
533     assert(vece <= (in_32 ? MO_32 : MO_64));
534     assert(in_32 == NULL || in_64 == NULL);
535 
536     /* If we're storing 0, expand oprsz to maxsz.  */
537     if (in_32 == NULL && in_64 == NULL) {
538         in_c = dup_const(vece, in_c);
539         if (in_c == 0) {
540             oprsz = maxsz;
541             vece = MO_8;
542         } else if (in_c == dup_const(MO_8, in_c)) {
543             vece = MO_8;
544         }
545     }
546 
547     /* Implement inline with a vector type, if possible.
548      * Prefer integer when 64-bit host and no variable dup.
549      */
550     type = choose_vector_type(NULL, vece, oprsz,
551                               (TCG_TARGET_REG_BITS == 64 && in_32 == NULL
552                                && (in_64 == NULL || vece == MO_64)));
553     if (type != 0) {
554         TCGv_vec t_vec = tcg_temp_new_vec(type);
555 
556         if (in_32) {
557             tcg_gen_dup_i32_vec(vece, t_vec, in_32);
558         } else if (in_64) {
559             tcg_gen_dup_i64_vec(vece, t_vec, in_64);
560         } else {
561             tcg_gen_dupi_vec(vece, t_vec, in_c);
562         }
563         do_dup_store(type, dofs, oprsz, maxsz, t_vec);
564         tcg_temp_free_vec(t_vec);
565         return;
566     }
567 
568     /* Otherwise, inline with an integer type, unless "large".  */
569     if (check_size_impl(oprsz, TCG_TARGET_REG_BITS / 8)) {
570         t_64 = NULL;
571         t_32 = NULL;
572 
573         if (in_32) {
574             /* We are given a 32-bit variable input.  For a 64-bit host,
575                use a 64-bit operation unless the 32-bit operation would
576                be simple enough.  */
577             if (TCG_TARGET_REG_BITS == 64
578                 && (vece != MO_32 || !check_size_impl(oprsz, 4))) {
579                 t_64 = tcg_temp_ebb_new_i64();
580                 tcg_gen_extu_i32_i64(t_64, in_32);
581                 tcg_gen_dup_i64(vece, t_64, t_64);
582             } else {
583                 t_32 = tcg_temp_ebb_new_i32();
584                 tcg_gen_dup_i32(vece, t_32, in_32);
585             }
586         } else if (in_64) {
587             /* We are given a 64-bit variable input.  */
588             t_64 = tcg_temp_ebb_new_i64();
589             tcg_gen_dup_i64(vece, t_64, in_64);
590         } else {
591             /* We are given a constant input.  */
592             /* For 64-bit hosts, use 64-bit constants for "simple" constants
593                or when we'd need too many 32-bit stores, or when a 64-bit
594                constant is really required.  */
595             if (vece == MO_64
596                 || (TCG_TARGET_REG_BITS == 64
597                     && (in_c == 0 || in_c == -1
598                         || !check_size_impl(oprsz, 4)))) {
599                 t_64 = tcg_constant_i64(in_c);
600             } else {
601                 t_32 = tcg_constant_i32(in_c);
602             }
603         }
604 
605         /* Implement inline if we picked an implementation size above.  */
606         if (t_32) {
607             for (i = 0; i < oprsz; i += 4) {
608                 tcg_gen_st_i32(t_32, cpu_env, dofs + i);
609             }
610             tcg_temp_free_i32(t_32);
611             goto done;
612         }
613         if (t_64) {
614             for (i = 0; i < oprsz; i += 8) {
615                 tcg_gen_st_i64(t_64, cpu_env, dofs + i);
616             }
617             tcg_temp_free_i64(t_64);
618             goto done;
619         }
620     }
621 
622     /* Otherwise implement out of line.  */
623     t_ptr = tcg_temp_ebb_new_ptr();
624     tcg_gen_addi_ptr(t_ptr, cpu_env, dofs);
625 
626     /*
627      * This may be expand_clr for the tail of an operation, e.g.
628      * oprsz == 8 && maxsz == 64.  The size of the clear is misaligned
629      * wrt simd_desc and will assert.  Simply pass all replicated byte
630      * stores through to memset.
631      */
632     if (oprsz == maxsz && vece == MO_8) {
633         TCGv_ptr t_size = tcg_constant_ptr(oprsz);
634         TCGv_i32 t_val;
635 
636         if (in_32) {
637             t_val = in_32;
638         } else if (in_64) {
639             t_val = tcg_temp_ebb_new_i32();
640             tcg_gen_extrl_i64_i32(t_val, in_64);
641         } else {
642             t_val = tcg_constant_i32(in_c);
643         }
644         gen_helper_memset(t_ptr, t_ptr, t_val, t_size);
645 
646         if (in_64) {
647             tcg_temp_free_i32(t_val);
648         }
649         tcg_temp_free_ptr(t_ptr);
650         return;
651     }
652 
653     t_desc = tcg_constant_i32(simd_desc(oprsz, maxsz, 0));
654 
655     if (vece == MO_64) {
656         if (in_64) {
657             gen_helper_gvec_dup64(t_ptr, t_desc, in_64);
658         } else {
659             t_64 = tcg_constant_i64(in_c);
660             gen_helper_gvec_dup64(t_ptr, t_desc, t_64);
661         }
662     } else {
663         typedef void dup_fn(TCGv_ptr, TCGv_i32, TCGv_i32);
664         static dup_fn * const fns[3] = {
665             gen_helper_gvec_dup8,
666             gen_helper_gvec_dup16,
667             gen_helper_gvec_dup32
668         };
669 
670         if (in_32) {
671             fns[vece](t_ptr, t_desc, in_32);
672         } else if (in_64) {
673             t_32 = tcg_temp_ebb_new_i32();
674             tcg_gen_extrl_i64_i32(t_32, in_64);
675             fns[vece](t_ptr, t_desc, t_32);
676             tcg_temp_free_i32(t_32);
677         } else {
678             if (vece == MO_8) {
679                 in_c &= 0xff;
680             } else if (vece == MO_16) {
681                 in_c &= 0xffff;
682             }
683             t_32 = tcg_constant_i32(in_c);
684             fns[vece](t_ptr, t_desc, t_32);
685         }
686     }
687 
688     tcg_temp_free_ptr(t_ptr);
689     return;
690 
691  done:
692     if (oprsz < maxsz) {
693         expand_clr(dofs + oprsz, maxsz - oprsz);
694     }
695 }
696 
697 /* Likewise, but with zero.  */
698 static void expand_clr(uint32_t dofs, uint32_t maxsz)
699 {
700     do_dup(MO_8, dofs, maxsz, maxsz, NULL, NULL, 0);
701 }
702 
703 /* Expand OPSZ bytes worth of two-operand operations using i32 elements.  */
704 static void expand_2_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
705                          bool load_dest, void (*fni)(TCGv_i32, TCGv_i32))
706 {
707     TCGv_i32 t0 = tcg_temp_new_i32();
708     TCGv_i32 t1 = tcg_temp_new_i32();
709     uint32_t i;
710 
711     for (i = 0; i < oprsz; i += 4) {
712         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
713         if (load_dest) {
714             tcg_gen_ld_i32(t1, cpu_env, dofs + i);
715         }
716         fni(t1, t0);
717         tcg_gen_st_i32(t1, cpu_env, dofs + i);
718     }
719     tcg_temp_free_i32(t0);
720     tcg_temp_free_i32(t1);
721 }
722 
723 static void expand_2i_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
724                           int32_t c, bool load_dest,
725                           void (*fni)(TCGv_i32, TCGv_i32, int32_t))
726 {
727     TCGv_i32 t0 = tcg_temp_new_i32();
728     TCGv_i32 t1 = tcg_temp_new_i32();
729     uint32_t i;
730 
731     for (i = 0; i < oprsz; i += 4) {
732         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
733         if (load_dest) {
734             tcg_gen_ld_i32(t1, cpu_env, dofs + i);
735         }
736         fni(t1, t0, c);
737         tcg_gen_st_i32(t1, cpu_env, dofs + i);
738     }
739     tcg_temp_free_i32(t0);
740     tcg_temp_free_i32(t1);
741 }
742 
743 static void expand_2s_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
744                           TCGv_i32 c, bool scalar_first,
745                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
746 {
747     TCGv_i32 t0 = tcg_temp_new_i32();
748     TCGv_i32 t1 = tcg_temp_new_i32();
749     uint32_t i;
750 
751     for (i = 0; i < oprsz; i += 4) {
752         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
753         if (scalar_first) {
754             fni(t1, c, t0);
755         } else {
756             fni(t1, t0, c);
757         }
758         tcg_gen_st_i32(t1, cpu_env, dofs + i);
759     }
760     tcg_temp_free_i32(t0);
761     tcg_temp_free_i32(t1);
762 }
763 
764 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
765 static void expand_3_i32(uint32_t dofs, uint32_t aofs,
766                          uint32_t bofs, uint32_t oprsz, bool load_dest,
767                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
768 {
769     TCGv_i32 t0 = tcg_temp_new_i32();
770     TCGv_i32 t1 = tcg_temp_new_i32();
771     TCGv_i32 t2 = tcg_temp_new_i32();
772     uint32_t i;
773 
774     for (i = 0; i < oprsz; i += 4) {
775         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
776         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
777         if (load_dest) {
778             tcg_gen_ld_i32(t2, cpu_env, dofs + i);
779         }
780         fni(t2, t0, t1);
781         tcg_gen_st_i32(t2, cpu_env, dofs + i);
782     }
783     tcg_temp_free_i32(t2);
784     tcg_temp_free_i32(t1);
785     tcg_temp_free_i32(t0);
786 }
787 
788 static void expand_3i_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
789                           uint32_t oprsz, int32_t c, bool load_dest,
790                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, int32_t))
791 {
792     TCGv_i32 t0 = tcg_temp_new_i32();
793     TCGv_i32 t1 = tcg_temp_new_i32();
794     TCGv_i32 t2 = tcg_temp_new_i32();
795     uint32_t i;
796 
797     for (i = 0; i < oprsz; i += 4) {
798         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
799         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
800         if (load_dest) {
801             tcg_gen_ld_i32(t2, cpu_env, dofs + i);
802         }
803         fni(t2, t0, t1, c);
804         tcg_gen_st_i32(t2, cpu_env, dofs + i);
805     }
806     tcg_temp_free_i32(t0);
807     tcg_temp_free_i32(t1);
808     tcg_temp_free_i32(t2);
809 }
810 
811 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
812 static void expand_4_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
813                          uint32_t cofs, uint32_t oprsz, bool write_aofs,
814                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32))
815 {
816     TCGv_i32 t0 = tcg_temp_new_i32();
817     TCGv_i32 t1 = tcg_temp_new_i32();
818     TCGv_i32 t2 = tcg_temp_new_i32();
819     TCGv_i32 t3 = tcg_temp_new_i32();
820     uint32_t i;
821 
822     for (i = 0; i < oprsz; i += 4) {
823         tcg_gen_ld_i32(t1, cpu_env, aofs + i);
824         tcg_gen_ld_i32(t2, cpu_env, bofs + i);
825         tcg_gen_ld_i32(t3, cpu_env, cofs + i);
826         fni(t0, t1, t2, t3);
827         tcg_gen_st_i32(t0, cpu_env, dofs + i);
828         if (write_aofs) {
829             tcg_gen_st_i32(t1, cpu_env, aofs + i);
830         }
831     }
832     tcg_temp_free_i32(t3);
833     tcg_temp_free_i32(t2);
834     tcg_temp_free_i32(t1);
835     tcg_temp_free_i32(t0);
836 }
837 
838 static void expand_4i_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
839                           uint32_t cofs, uint32_t oprsz, int32_t c,
840                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32,
841                                       int32_t))
842 {
843     TCGv_i32 t0 = tcg_temp_new_i32();
844     TCGv_i32 t1 = tcg_temp_new_i32();
845     TCGv_i32 t2 = tcg_temp_new_i32();
846     TCGv_i32 t3 = tcg_temp_new_i32();
847     uint32_t i;
848 
849     for (i = 0; i < oprsz; i += 4) {
850         tcg_gen_ld_i32(t1, cpu_env, aofs + i);
851         tcg_gen_ld_i32(t2, cpu_env, bofs + i);
852         tcg_gen_ld_i32(t3, cpu_env, cofs + i);
853         fni(t0, t1, t2, t3, c);
854         tcg_gen_st_i32(t0, cpu_env, dofs + i);
855     }
856     tcg_temp_free_i32(t3);
857     tcg_temp_free_i32(t2);
858     tcg_temp_free_i32(t1);
859     tcg_temp_free_i32(t0);
860 }
861 
862 /* Expand OPSZ bytes worth of two-operand operations using i64 elements.  */
863 static void expand_2_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
864                          bool load_dest, void (*fni)(TCGv_i64, TCGv_i64))
865 {
866     TCGv_i64 t0 = tcg_temp_new_i64();
867     TCGv_i64 t1 = tcg_temp_new_i64();
868     uint32_t i;
869 
870     for (i = 0; i < oprsz; i += 8) {
871         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
872         if (load_dest) {
873             tcg_gen_ld_i64(t1, cpu_env, dofs + i);
874         }
875         fni(t1, t0);
876         tcg_gen_st_i64(t1, cpu_env, dofs + i);
877     }
878     tcg_temp_free_i64(t0);
879     tcg_temp_free_i64(t1);
880 }
881 
882 static void expand_2i_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
883                           int64_t c, bool load_dest,
884                           void (*fni)(TCGv_i64, TCGv_i64, int64_t))
885 {
886     TCGv_i64 t0 = tcg_temp_new_i64();
887     TCGv_i64 t1 = tcg_temp_new_i64();
888     uint32_t i;
889 
890     for (i = 0; i < oprsz; i += 8) {
891         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
892         if (load_dest) {
893             tcg_gen_ld_i64(t1, cpu_env, dofs + i);
894         }
895         fni(t1, t0, c);
896         tcg_gen_st_i64(t1, cpu_env, dofs + i);
897     }
898     tcg_temp_free_i64(t0);
899     tcg_temp_free_i64(t1);
900 }
901 
902 static void expand_2s_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
903                           TCGv_i64 c, bool scalar_first,
904                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
905 {
906     TCGv_i64 t0 = tcg_temp_new_i64();
907     TCGv_i64 t1 = tcg_temp_new_i64();
908     uint32_t i;
909 
910     for (i = 0; i < oprsz; i += 8) {
911         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
912         if (scalar_first) {
913             fni(t1, c, t0);
914         } else {
915             fni(t1, t0, c);
916         }
917         tcg_gen_st_i64(t1, cpu_env, dofs + i);
918     }
919     tcg_temp_free_i64(t0);
920     tcg_temp_free_i64(t1);
921 }
922 
923 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
924 static void expand_3_i64(uint32_t dofs, uint32_t aofs,
925                          uint32_t bofs, uint32_t oprsz, bool load_dest,
926                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
927 {
928     TCGv_i64 t0 = tcg_temp_new_i64();
929     TCGv_i64 t1 = tcg_temp_new_i64();
930     TCGv_i64 t2 = tcg_temp_new_i64();
931     uint32_t i;
932 
933     for (i = 0; i < oprsz; i += 8) {
934         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
935         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
936         if (load_dest) {
937             tcg_gen_ld_i64(t2, cpu_env, dofs + i);
938         }
939         fni(t2, t0, t1);
940         tcg_gen_st_i64(t2, cpu_env, dofs + i);
941     }
942     tcg_temp_free_i64(t2);
943     tcg_temp_free_i64(t1);
944     tcg_temp_free_i64(t0);
945 }
946 
947 static void expand_3i_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
948                           uint32_t oprsz, int64_t c, bool load_dest,
949                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, int64_t))
950 {
951     TCGv_i64 t0 = tcg_temp_new_i64();
952     TCGv_i64 t1 = tcg_temp_new_i64();
953     TCGv_i64 t2 = tcg_temp_new_i64();
954     uint32_t i;
955 
956     for (i = 0; i < oprsz; i += 8) {
957         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
958         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
959         if (load_dest) {
960             tcg_gen_ld_i64(t2, cpu_env, dofs + i);
961         }
962         fni(t2, t0, t1, c);
963         tcg_gen_st_i64(t2, cpu_env, dofs + i);
964     }
965     tcg_temp_free_i64(t0);
966     tcg_temp_free_i64(t1);
967     tcg_temp_free_i64(t2);
968 }
969 
970 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
971 static void expand_4_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
972                          uint32_t cofs, uint32_t oprsz, bool write_aofs,
973                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64))
974 {
975     TCGv_i64 t0 = tcg_temp_new_i64();
976     TCGv_i64 t1 = tcg_temp_new_i64();
977     TCGv_i64 t2 = tcg_temp_new_i64();
978     TCGv_i64 t3 = tcg_temp_new_i64();
979     uint32_t i;
980 
981     for (i = 0; i < oprsz; i += 8) {
982         tcg_gen_ld_i64(t1, cpu_env, aofs + i);
983         tcg_gen_ld_i64(t2, cpu_env, bofs + i);
984         tcg_gen_ld_i64(t3, cpu_env, cofs + i);
985         fni(t0, t1, t2, t3);
986         tcg_gen_st_i64(t0, cpu_env, dofs + i);
987         if (write_aofs) {
988             tcg_gen_st_i64(t1, cpu_env, aofs + i);
989         }
990     }
991     tcg_temp_free_i64(t3);
992     tcg_temp_free_i64(t2);
993     tcg_temp_free_i64(t1);
994     tcg_temp_free_i64(t0);
995 }
996 
997 static void expand_4i_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
998                           uint32_t cofs, uint32_t oprsz, int64_t c,
999                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64,
1000                                       int64_t))
1001 {
1002     TCGv_i64 t0 = tcg_temp_new_i64();
1003     TCGv_i64 t1 = tcg_temp_new_i64();
1004     TCGv_i64 t2 = tcg_temp_new_i64();
1005     TCGv_i64 t3 = tcg_temp_new_i64();
1006     uint32_t i;
1007 
1008     for (i = 0; i < oprsz; i += 8) {
1009         tcg_gen_ld_i64(t1, cpu_env, aofs + i);
1010         tcg_gen_ld_i64(t2, cpu_env, bofs + i);
1011         tcg_gen_ld_i64(t3, cpu_env, cofs + i);
1012         fni(t0, t1, t2, t3, c);
1013         tcg_gen_st_i64(t0, cpu_env, dofs + i);
1014     }
1015     tcg_temp_free_i64(t3);
1016     tcg_temp_free_i64(t2);
1017     tcg_temp_free_i64(t1);
1018     tcg_temp_free_i64(t0);
1019 }
1020 
1021 /* Expand OPSZ bytes worth of two-operand operations using host vectors.  */
1022 static void expand_2_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1023                          uint32_t oprsz, uint32_t tysz, TCGType type,
1024                          bool load_dest,
1025                          void (*fni)(unsigned, TCGv_vec, TCGv_vec))
1026 {
1027     TCGv_vec t0 = tcg_temp_new_vec(type);
1028     TCGv_vec t1 = tcg_temp_new_vec(type);
1029     uint32_t i;
1030 
1031     for (i = 0; i < oprsz; i += tysz) {
1032         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1033         if (load_dest) {
1034             tcg_gen_ld_vec(t1, cpu_env, dofs + i);
1035         }
1036         fni(vece, t1, t0);
1037         tcg_gen_st_vec(t1, cpu_env, dofs + i);
1038     }
1039     tcg_temp_free_vec(t0);
1040     tcg_temp_free_vec(t1);
1041 }
1042 
1043 /* Expand OPSZ bytes worth of two-vector operands and an immediate operand
1044    using host vectors.  */
1045 static void expand_2i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1046                           uint32_t oprsz, uint32_t tysz, TCGType type,
1047                           int64_t c, bool load_dest,
1048                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, int64_t))
1049 {
1050     TCGv_vec t0 = tcg_temp_new_vec(type);
1051     TCGv_vec t1 = tcg_temp_new_vec(type);
1052     uint32_t i;
1053 
1054     for (i = 0; i < oprsz; i += tysz) {
1055         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1056         if (load_dest) {
1057             tcg_gen_ld_vec(t1, cpu_env, dofs + i);
1058         }
1059         fni(vece, t1, t0, c);
1060         tcg_gen_st_vec(t1, cpu_env, dofs + i);
1061     }
1062     tcg_temp_free_vec(t0);
1063     tcg_temp_free_vec(t1);
1064 }
1065 
1066 static void expand_2s_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1067                           uint32_t oprsz, uint32_t tysz, TCGType type,
1068                           TCGv_vec c, bool scalar_first,
1069                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
1070 {
1071     TCGv_vec t0 = tcg_temp_new_vec(type);
1072     TCGv_vec t1 = tcg_temp_new_vec(type);
1073     uint32_t i;
1074 
1075     for (i = 0; i < oprsz; i += tysz) {
1076         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1077         if (scalar_first) {
1078             fni(vece, t1, c, t0);
1079         } else {
1080             fni(vece, t1, t0, c);
1081         }
1082         tcg_gen_st_vec(t1, cpu_env, dofs + i);
1083     }
1084     tcg_temp_free_vec(t0);
1085     tcg_temp_free_vec(t1);
1086 }
1087 
1088 /* Expand OPSZ bytes worth of three-operand operations using host vectors.  */
1089 static void expand_3_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1090                          uint32_t bofs, uint32_t oprsz,
1091                          uint32_t tysz, TCGType type, bool load_dest,
1092                          void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
1093 {
1094     TCGv_vec t0 = tcg_temp_new_vec(type);
1095     TCGv_vec t1 = tcg_temp_new_vec(type);
1096     TCGv_vec t2 = tcg_temp_new_vec(type);
1097     uint32_t i;
1098 
1099     for (i = 0; i < oprsz; i += tysz) {
1100         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1101         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
1102         if (load_dest) {
1103             tcg_gen_ld_vec(t2, cpu_env, dofs + i);
1104         }
1105         fni(vece, t2, t0, t1);
1106         tcg_gen_st_vec(t2, cpu_env, dofs + i);
1107     }
1108     tcg_temp_free_vec(t2);
1109     tcg_temp_free_vec(t1);
1110     tcg_temp_free_vec(t0);
1111 }
1112 
1113 /*
1114  * Expand OPSZ bytes worth of three-vector operands and an immediate operand
1115  * using host vectors.
1116  */
1117 static void expand_3i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1118                           uint32_t bofs, uint32_t oprsz, uint32_t tysz,
1119                           TCGType type, int64_t c, bool load_dest,
1120                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec,
1121                                       int64_t))
1122 {
1123     TCGv_vec t0 = tcg_temp_new_vec(type);
1124     TCGv_vec t1 = tcg_temp_new_vec(type);
1125     TCGv_vec t2 = tcg_temp_new_vec(type);
1126     uint32_t i;
1127 
1128     for (i = 0; i < oprsz; i += tysz) {
1129         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1130         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
1131         if (load_dest) {
1132             tcg_gen_ld_vec(t2, cpu_env, dofs + i);
1133         }
1134         fni(vece, t2, t0, t1, c);
1135         tcg_gen_st_vec(t2, cpu_env, dofs + i);
1136     }
1137     tcg_temp_free_vec(t0);
1138     tcg_temp_free_vec(t1);
1139     tcg_temp_free_vec(t2);
1140 }
1141 
1142 /* Expand OPSZ bytes worth of four-operand operations using host vectors.  */
1143 static void expand_4_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1144                          uint32_t bofs, uint32_t cofs, uint32_t oprsz,
1145                          uint32_t tysz, TCGType type, bool write_aofs,
1146                          void (*fni)(unsigned, TCGv_vec, TCGv_vec,
1147                                      TCGv_vec, TCGv_vec))
1148 {
1149     TCGv_vec t0 = tcg_temp_new_vec(type);
1150     TCGv_vec t1 = tcg_temp_new_vec(type);
1151     TCGv_vec t2 = tcg_temp_new_vec(type);
1152     TCGv_vec t3 = tcg_temp_new_vec(type);
1153     uint32_t i;
1154 
1155     for (i = 0; i < oprsz; i += tysz) {
1156         tcg_gen_ld_vec(t1, cpu_env, aofs + i);
1157         tcg_gen_ld_vec(t2, cpu_env, bofs + i);
1158         tcg_gen_ld_vec(t3, cpu_env, cofs + i);
1159         fni(vece, t0, t1, t2, t3);
1160         tcg_gen_st_vec(t0, cpu_env, dofs + i);
1161         if (write_aofs) {
1162             tcg_gen_st_vec(t1, cpu_env, aofs + i);
1163         }
1164     }
1165     tcg_temp_free_vec(t3);
1166     tcg_temp_free_vec(t2);
1167     tcg_temp_free_vec(t1);
1168     tcg_temp_free_vec(t0);
1169 }
1170 
1171 /*
1172  * Expand OPSZ bytes worth of four-vector operands and an immediate operand
1173  * using host vectors.
1174  */
1175 static void expand_4i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1176                           uint32_t bofs, uint32_t cofs, uint32_t oprsz,
1177                           uint32_t tysz, TCGType type, int64_t c,
1178                           void (*fni)(unsigned, TCGv_vec, TCGv_vec,
1179                                      TCGv_vec, TCGv_vec, int64_t))
1180 {
1181     TCGv_vec t0 = tcg_temp_new_vec(type);
1182     TCGv_vec t1 = tcg_temp_new_vec(type);
1183     TCGv_vec t2 = tcg_temp_new_vec(type);
1184     TCGv_vec t3 = tcg_temp_new_vec(type);
1185     uint32_t i;
1186 
1187     for (i = 0; i < oprsz; i += tysz) {
1188         tcg_gen_ld_vec(t1, cpu_env, aofs + i);
1189         tcg_gen_ld_vec(t2, cpu_env, bofs + i);
1190         tcg_gen_ld_vec(t3, cpu_env, cofs + i);
1191         fni(vece, t0, t1, t2, t3, c);
1192         tcg_gen_st_vec(t0, cpu_env, dofs + i);
1193     }
1194     tcg_temp_free_vec(t3);
1195     tcg_temp_free_vec(t2);
1196     tcg_temp_free_vec(t1);
1197     tcg_temp_free_vec(t0);
1198 }
1199 
1200 /* Expand a vector two-operand operation.  */
1201 void tcg_gen_gvec_2(uint32_t dofs, uint32_t aofs,
1202                     uint32_t oprsz, uint32_t maxsz, const GVecGen2 *g)
1203 {
1204     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1205     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1206     TCGType type;
1207     uint32_t some;
1208 
1209     check_size_align(oprsz, maxsz, dofs | aofs);
1210     check_overlap_2(dofs, aofs, maxsz);
1211 
1212     type = 0;
1213     if (g->fniv) {
1214         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1215     }
1216     switch (type) {
1217     case TCG_TYPE_V256:
1218         /* Recall that ARM SVE allows vector sizes that are not a
1219          * power of 2, but always a multiple of 16.  The intent is
1220          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1221          */
1222         some = QEMU_ALIGN_DOWN(oprsz, 32);
1223         expand_2_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1224                      g->load_dest, g->fniv);
1225         if (some == oprsz) {
1226             break;
1227         }
1228         dofs += some;
1229         aofs += some;
1230         oprsz -= some;
1231         maxsz -= some;
1232         /* fallthru */
1233     case TCG_TYPE_V128:
1234         expand_2_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1235                      g->load_dest, g->fniv);
1236         break;
1237     case TCG_TYPE_V64:
1238         expand_2_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1239                      g->load_dest, g->fniv);
1240         break;
1241 
1242     case 0:
1243         if (g->fni8 && check_size_impl(oprsz, 8)) {
1244             expand_2_i64(dofs, aofs, oprsz, g->load_dest, g->fni8);
1245         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1246             expand_2_i32(dofs, aofs, oprsz, g->load_dest, g->fni4);
1247         } else {
1248             assert(g->fno != NULL);
1249             tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, g->data, g->fno);
1250             oprsz = maxsz;
1251         }
1252         break;
1253 
1254     default:
1255         g_assert_not_reached();
1256     }
1257     tcg_swap_vecop_list(hold_list);
1258 
1259     if (oprsz < maxsz) {
1260         expand_clr(dofs + oprsz, maxsz - oprsz);
1261     }
1262 }
1263 
1264 /* Expand a vector operation with two vectors and an immediate.  */
1265 void tcg_gen_gvec_2i(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
1266                      uint32_t maxsz, int64_t c, const GVecGen2i *g)
1267 {
1268     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1269     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1270     TCGType type;
1271     uint32_t some;
1272 
1273     check_size_align(oprsz, maxsz, dofs | aofs);
1274     check_overlap_2(dofs, aofs, maxsz);
1275 
1276     type = 0;
1277     if (g->fniv) {
1278         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1279     }
1280     switch (type) {
1281     case TCG_TYPE_V256:
1282         /* Recall that ARM SVE allows vector sizes that are not a
1283          * power of 2, but always a multiple of 16.  The intent is
1284          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1285          */
1286         some = QEMU_ALIGN_DOWN(oprsz, 32);
1287         expand_2i_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1288                       c, g->load_dest, g->fniv);
1289         if (some == oprsz) {
1290             break;
1291         }
1292         dofs += some;
1293         aofs += some;
1294         oprsz -= some;
1295         maxsz -= some;
1296         /* fallthru */
1297     case TCG_TYPE_V128:
1298         expand_2i_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1299                       c, g->load_dest, g->fniv);
1300         break;
1301     case TCG_TYPE_V64:
1302         expand_2i_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1303                       c, g->load_dest, g->fniv);
1304         break;
1305 
1306     case 0:
1307         if (g->fni8 && check_size_impl(oprsz, 8)) {
1308             expand_2i_i64(dofs, aofs, oprsz, c, g->load_dest, g->fni8);
1309         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1310             expand_2i_i32(dofs, aofs, oprsz, c, g->load_dest, g->fni4);
1311         } else {
1312             if (g->fno) {
1313                 tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, c, g->fno);
1314             } else {
1315                 TCGv_i64 tcg_c = tcg_constant_i64(c);
1316                 tcg_gen_gvec_2i_ool(dofs, aofs, tcg_c, oprsz,
1317                                     maxsz, c, g->fnoi);
1318             }
1319             oprsz = maxsz;
1320         }
1321         break;
1322 
1323     default:
1324         g_assert_not_reached();
1325     }
1326     tcg_swap_vecop_list(hold_list);
1327 
1328     if (oprsz < maxsz) {
1329         expand_clr(dofs + oprsz, maxsz - oprsz);
1330     }
1331 }
1332 
1333 /* Expand a vector operation with two vectors and a scalar.  */
1334 void tcg_gen_gvec_2s(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
1335                      uint32_t maxsz, TCGv_i64 c, const GVecGen2s *g)
1336 {
1337     TCGType type;
1338 
1339     check_size_align(oprsz, maxsz, dofs | aofs);
1340     check_overlap_2(dofs, aofs, maxsz);
1341 
1342     type = 0;
1343     if (g->fniv) {
1344         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1345     }
1346     if (type != 0) {
1347         const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1348         const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1349         TCGv_vec t_vec = tcg_temp_new_vec(type);
1350         uint32_t some;
1351 
1352         tcg_gen_dup_i64_vec(g->vece, t_vec, c);
1353 
1354         switch (type) {
1355         case TCG_TYPE_V256:
1356             /* Recall that ARM SVE allows vector sizes that are not a
1357              * power of 2, but always a multiple of 16.  The intent is
1358              * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1359              */
1360             some = QEMU_ALIGN_DOWN(oprsz, 32);
1361             expand_2s_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1362                           t_vec, g->scalar_first, g->fniv);
1363             if (some == oprsz) {
1364                 break;
1365             }
1366             dofs += some;
1367             aofs += some;
1368             oprsz -= some;
1369             maxsz -= some;
1370             /* fallthru */
1371 
1372         case TCG_TYPE_V128:
1373             expand_2s_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1374                           t_vec, g->scalar_first, g->fniv);
1375             break;
1376 
1377         case TCG_TYPE_V64:
1378             expand_2s_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1379                           t_vec, g->scalar_first, g->fniv);
1380             break;
1381 
1382         default:
1383             g_assert_not_reached();
1384         }
1385         tcg_temp_free_vec(t_vec);
1386         tcg_swap_vecop_list(hold_list);
1387     } else if (g->fni8 && check_size_impl(oprsz, 8)) {
1388         TCGv_i64 t64 = tcg_temp_new_i64();
1389 
1390         tcg_gen_dup_i64(g->vece, t64, c);
1391         expand_2s_i64(dofs, aofs, oprsz, t64, g->scalar_first, g->fni8);
1392         tcg_temp_free_i64(t64);
1393     } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1394         TCGv_i32 t32 = tcg_temp_new_i32();
1395 
1396         tcg_gen_extrl_i64_i32(t32, c);
1397         tcg_gen_dup_i32(g->vece, t32, t32);
1398         expand_2s_i32(dofs, aofs, oprsz, t32, g->scalar_first, g->fni4);
1399         tcg_temp_free_i32(t32);
1400     } else {
1401         tcg_gen_gvec_2i_ool(dofs, aofs, c, oprsz, maxsz, 0, g->fno);
1402         return;
1403     }
1404 
1405     if (oprsz < maxsz) {
1406         expand_clr(dofs + oprsz, maxsz - oprsz);
1407     }
1408 }
1409 
1410 /* Expand a vector three-operand operation.  */
1411 void tcg_gen_gvec_3(uint32_t dofs, uint32_t aofs, uint32_t bofs,
1412                     uint32_t oprsz, uint32_t maxsz, const GVecGen3 *g)
1413 {
1414     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1415     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1416     TCGType type;
1417     uint32_t some;
1418 
1419     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
1420     check_overlap_3(dofs, aofs, bofs, maxsz);
1421 
1422     type = 0;
1423     if (g->fniv) {
1424         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1425     }
1426     switch (type) {
1427     case TCG_TYPE_V256:
1428         /* Recall that ARM SVE allows vector sizes that are not a
1429          * power of 2, but always a multiple of 16.  The intent is
1430          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1431          */
1432         some = QEMU_ALIGN_DOWN(oprsz, 32);
1433         expand_3_vec(g->vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256,
1434                      g->load_dest, g->fniv);
1435         if (some == oprsz) {
1436             break;
1437         }
1438         dofs += some;
1439         aofs += some;
1440         bofs += some;
1441         oprsz -= some;
1442         maxsz -= some;
1443         /* fallthru */
1444     case TCG_TYPE_V128:
1445         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128,
1446                      g->load_dest, g->fniv);
1447         break;
1448     case TCG_TYPE_V64:
1449         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64,
1450                      g->load_dest, g->fniv);
1451         break;
1452 
1453     case 0:
1454         if (g->fni8 && check_size_impl(oprsz, 8)) {
1455             expand_3_i64(dofs, aofs, bofs, oprsz, g->load_dest, g->fni8);
1456         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1457             expand_3_i32(dofs, aofs, bofs, oprsz, g->load_dest, g->fni4);
1458         } else {
1459             assert(g->fno != NULL);
1460             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz,
1461                                maxsz, g->data, g->fno);
1462             oprsz = maxsz;
1463         }
1464         break;
1465 
1466     default:
1467         g_assert_not_reached();
1468     }
1469     tcg_swap_vecop_list(hold_list);
1470 
1471     if (oprsz < maxsz) {
1472         expand_clr(dofs + oprsz, maxsz - oprsz);
1473     }
1474 }
1475 
1476 /* Expand a vector operation with three vectors and an immediate.  */
1477 void tcg_gen_gvec_3i(uint32_t dofs, uint32_t aofs, uint32_t bofs,
1478                      uint32_t oprsz, uint32_t maxsz, int64_t c,
1479                      const GVecGen3i *g)
1480 {
1481     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1482     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1483     TCGType type;
1484     uint32_t some;
1485 
1486     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
1487     check_overlap_3(dofs, aofs, bofs, maxsz);
1488 
1489     type = 0;
1490     if (g->fniv) {
1491         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1492     }
1493     switch (type) {
1494     case TCG_TYPE_V256:
1495         /*
1496          * Recall that ARM SVE allows vector sizes that are not a
1497          * power of 2, but always a multiple of 16.  The intent is
1498          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1499          */
1500         some = QEMU_ALIGN_DOWN(oprsz, 32);
1501         expand_3i_vec(g->vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256,
1502                       c, g->load_dest, g->fniv);
1503         if (some == oprsz) {
1504             break;
1505         }
1506         dofs += some;
1507         aofs += some;
1508         bofs += some;
1509         oprsz -= some;
1510         maxsz -= some;
1511         /* fallthru */
1512     case TCG_TYPE_V128:
1513         expand_3i_vec(g->vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128,
1514                       c, g->load_dest, g->fniv);
1515         break;
1516     case TCG_TYPE_V64:
1517         expand_3i_vec(g->vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64,
1518                       c, g->load_dest, g->fniv);
1519         break;
1520 
1521     case 0:
1522         if (g->fni8 && check_size_impl(oprsz, 8)) {
1523             expand_3i_i64(dofs, aofs, bofs, oprsz, c, g->load_dest, g->fni8);
1524         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1525             expand_3i_i32(dofs, aofs, bofs, oprsz, c, g->load_dest, g->fni4);
1526         } else {
1527             assert(g->fno != NULL);
1528             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, c, g->fno);
1529             oprsz = maxsz;
1530         }
1531         break;
1532 
1533     default:
1534         g_assert_not_reached();
1535     }
1536     tcg_swap_vecop_list(hold_list);
1537 
1538     if (oprsz < maxsz) {
1539         expand_clr(dofs + oprsz, maxsz - oprsz);
1540     }
1541 }
1542 
1543 /* Expand a vector four-operand operation.  */
1544 void tcg_gen_gvec_4(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
1545                     uint32_t oprsz, uint32_t maxsz, const GVecGen4 *g)
1546 {
1547     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1548     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1549     TCGType type;
1550     uint32_t some;
1551 
1552     check_size_align(oprsz, maxsz, dofs | aofs | bofs | cofs);
1553     check_overlap_4(dofs, aofs, bofs, cofs, maxsz);
1554 
1555     type = 0;
1556     if (g->fniv) {
1557         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1558     }
1559     switch (type) {
1560     case TCG_TYPE_V256:
1561         /* Recall that ARM SVE allows vector sizes that are not a
1562          * power of 2, but always a multiple of 16.  The intent is
1563          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1564          */
1565         some = QEMU_ALIGN_DOWN(oprsz, 32);
1566         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, some,
1567                      32, TCG_TYPE_V256, g->write_aofs, g->fniv);
1568         if (some == oprsz) {
1569             break;
1570         }
1571         dofs += some;
1572         aofs += some;
1573         bofs += some;
1574         cofs += some;
1575         oprsz -= some;
1576         maxsz -= some;
1577         /* fallthru */
1578     case TCG_TYPE_V128:
1579         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1580                      16, TCG_TYPE_V128, g->write_aofs, g->fniv);
1581         break;
1582     case TCG_TYPE_V64:
1583         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1584                      8, TCG_TYPE_V64, g->write_aofs, g->fniv);
1585         break;
1586 
1587     case 0:
1588         if (g->fni8 && check_size_impl(oprsz, 8)) {
1589             expand_4_i64(dofs, aofs, bofs, cofs, oprsz,
1590                          g->write_aofs, g->fni8);
1591         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1592             expand_4_i32(dofs, aofs, bofs, cofs, oprsz,
1593                          g->write_aofs, g->fni4);
1594         } else {
1595             assert(g->fno != NULL);
1596             tcg_gen_gvec_4_ool(dofs, aofs, bofs, cofs,
1597                                oprsz, maxsz, g->data, g->fno);
1598             oprsz = maxsz;
1599         }
1600         break;
1601 
1602     default:
1603         g_assert_not_reached();
1604     }
1605     tcg_swap_vecop_list(hold_list);
1606 
1607     if (oprsz < maxsz) {
1608         expand_clr(dofs + oprsz, maxsz - oprsz);
1609     }
1610 }
1611 
1612 /* Expand a vector four-operand operation.  */
1613 void tcg_gen_gvec_4i(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
1614                      uint32_t oprsz, uint32_t maxsz, int64_t c,
1615                      const GVecGen4i *g)
1616 {
1617     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1618     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1619     TCGType type;
1620     uint32_t some;
1621 
1622     check_size_align(oprsz, maxsz, dofs | aofs | bofs | cofs);
1623     check_overlap_4(dofs, aofs, bofs, cofs, maxsz);
1624 
1625     type = 0;
1626     if (g->fniv) {
1627         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1628     }
1629     switch (type) {
1630     case TCG_TYPE_V256:
1631         /*
1632          * Recall that ARM SVE allows vector sizes that are not a
1633          * power of 2, but always a multiple of 16.  The intent is
1634          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1635          */
1636         some = QEMU_ALIGN_DOWN(oprsz, 32);
1637         expand_4i_vec(g->vece, dofs, aofs, bofs, cofs, some,
1638                       32, TCG_TYPE_V256, c, g->fniv);
1639         if (some == oprsz) {
1640             break;
1641         }
1642         dofs += some;
1643         aofs += some;
1644         bofs += some;
1645         cofs += some;
1646         oprsz -= some;
1647         maxsz -= some;
1648         /* fallthru */
1649     case TCG_TYPE_V128:
1650         expand_4i_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1651                        16, TCG_TYPE_V128, c, g->fniv);
1652         break;
1653     case TCG_TYPE_V64:
1654         expand_4i_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1655                       8, TCG_TYPE_V64, c, g->fniv);
1656         break;
1657 
1658     case 0:
1659         if (g->fni8 && check_size_impl(oprsz, 8)) {
1660             expand_4i_i64(dofs, aofs, bofs, cofs, oprsz, c, g->fni8);
1661         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1662             expand_4i_i32(dofs, aofs, bofs, cofs, oprsz, c, g->fni4);
1663         } else {
1664             assert(g->fno != NULL);
1665             tcg_gen_gvec_4_ool(dofs, aofs, bofs, cofs,
1666                                oprsz, maxsz, c, g->fno);
1667             oprsz = maxsz;
1668         }
1669         break;
1670 
1671     default:
1672         g_assert_not_reached();
1673     }
1674     tcg_swap_vecop_list(hold_list);
1675 
1676     if (oprsz < maxsz) {
1677         expand_clr(dofs + oprsz, maxsz - oprsz);
1678     }
1679 }
1680 
1681 /*
1682  * Expand specific vector operations.
1683  */
1684 
1685 static void vec_mov2(unsigned vece, TCGv_vec a, TCGv_vec b)
1686 {
1687     tcg_gen_mov_vec(a, b);
1688 }
1689 
1690 void tcg_gen_gvec_mov(unsigned vece, uint32_t dofs, uint32_t aofs,
1691                       uint32_t oprsz, uint32_t maxsz)
1692 {
1693     static const GVecGen2 g = {
1694         .fni8 = tcg_gen_mov_i64,
1695         .fniv = vec_mov2,
1696         .fno = gen_helper_gvec_mov,
1697         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1698     };
1699     if (dofs != aofs) {
1700         tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1701     } else {
1702         check_size_align(oprsz, maxsz, dofs);
1703         if (oprsz < maxsz) {
1704             expand_clr(dofs + oprsz, maxsz - oprsz);
1705         }
1706     }
1707 }
1708 
1709 void tcg_gen_gvec_dup_i32(unsigned vece, uint32_t dofs, uint32_t oprsz,
1710                           uint32_t maxsz, TCGv_i32 in)
1711 {
1712     check_size_align(oprsz, maxsz, dofs);
1713     tcg_debug_assert(vece <= MO_32);
1714     do_dup(vece, dofs, oprsz, maxsz, in, NULL, 0);
1715 }
1716 
1717 void tcg_gen_gvec_dup_i64(unsigned vece, uint32_t dofs, uint32_t oprsz,
1718                           uint32_t maxsz, TCGv_i64 in)
1719 {
1720     check_size_align(oprsz, maxsz, dofs);
1721     tcg_debug_assert(vece <= MO_64);
1722     do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
1723 }
1724 
1725 void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
1726                           uint32_t oprsz, uint32_t maxsz)
1727 {
1728     check_size_align(oprsz, maxsz, dofs);
1729     if (vece <= MO_64) {
1730         TCGType type = choose_vector_type(NULL, vece, oprsz, 0);
1731         if (type != 0) {
1732             TCGv_vec t_vec = tcg_temp_new_vec(type);
1733             tcg_gen_dup_mem_vec(vece, t_vec, cpu_env, aofs);
1734             do_dup_store(type, dofs, oprsz, maxsz, t_vec);
1735             tcg_temp_free_vec(t_vec);
1736         } else if (vece <= MO_32) {
1737             TCGv_i32 in = tcg_temp_ebb_new_i32();
1738             switch (vece) {
1739             case MO_8:
1740                 tcg_gen_ld8u_i32(in, cpu_env, aofs);
1741                 break;
1742             case MO_16:
1743                 tcg_gen_ld16u_i32(in, cpu_env, aofs);
1744                 break;
1745             default:
1746                 tcg_gen_ld_i32(in, cpu_env, aofs);
1747                 break;
1748             }
1749             do_dup(vece, dofs, oprsz, maxsz, in, NULL, 0);
1750             tcg_temp_free_i32(in);
1751         } else {
1752             TCGv_i64 in = tcg_temp_ebb_new_i64();
1753             tcg_gen_ld_i64(in, cpu_env, aofs);
1754             do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
1755             tcg_temp_free_i64(in);
1756         }
1757     } else if (vece == 4) {
1758         /* 128-bit duplicate.  */
1759         int i;
1760 
1761         tcg_debug_assert(oprsz >= 16);
1762         if (TCG_TARGET_HAS_v128) {
1763             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V128);
1764 
1765             tcg_gen_ld_vec(in, cpu_env, aofs);
1766             for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
1767                 tcg_gen_st_vec(in, cpu_env, dofs + i);
1768             }
1769             tcg_temp_free_vec(in);
1770         } else {
1771             TCGv_i64 in0 = tcg_temp_ebb_new_i64();
1772             TCGv_i64 in1 = tcg_temp_ebb_new_i64();
1773 
1774             tcg_gen_ld_i64(in0, cpu_env, aofs);
1775             tcg_gen_ld_i64(in1, cpu_env, aofs + 8);
1776             for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
1777                 tcg_gen_st_i64(in0, cpu_env, dofs + i);
1778                 tcg_gen_st_i64(in1, cpu_env, dofs + i + 8);
1779             }
1780             tcg_temp_free_i64(in0);
1781             tcg_temp_free_i64(in1);
1782         }
1783         if (oprsz < maxsz) {
1784             expand_clr(dofs + oprsz, maxsz - oprsz);
1785         }
1786     } else if (vece == 5) {
1787         /* 256-bit duplicate.  */
1788         int i;
1789 
1790         tcg_debug_assert(oprsz >= 32);
1791         tcg_debug_assert(oprsz % 32 == 0);
1792         if (TCG_TARGET_HAS_v256) {
1793             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V256);
1794 
1795             tcg_gen_ld_vec(in, cpu_env, aofs);
1796             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1797                 tcg_gen_st_vec(in, cpu_env, dofs + i);
1798             }
1799             tcg_temp_free_vec(in);
1800         } else if (TCG_TARGET_HAS_v128) {
1801             TCGv_vec in0 = tcg_temp_new_vec(TCG_TYPE_V128);
1802             TCGv_vec in1 = tcg_temp_new_vec(TCG_TYPE_V128);
1803 
1804             tcg_gen_ld_vec(in0, cpu_env, aofs);
1805             tcg_gen_ld_vec(in1, cpu_env, aofs + 16);
1806             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1807                 tcg_gen_st_vec(in0, cpu_env, dofs + i);
1808                 tcg_gen_st_vec(in1, cpu_env, dofs + i + 16);
1809             }
1810             tcg_temp_free_vec(in0);
1811             tcg_temp_free_vec(in1);
1812         } else {
1813             TCGv_i64 in[4];
1814             int j;
1815 
1816             for (j = 0; j < 4; ++j) {
1817                 in[j] = tcg_temp_ebb_new_i64();
1818                 tcg_gen_ld_i64(in[j], cpu_env, aofs + j * 8);
1819             }
1820             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1821                 for (j = 0; j < 4; ++j) {
1822                     tcg_gen_st_i64(in[j], cpu_env, dofs + i + j * 8);
1823                 }
1824             }
1825             for (j = 0; j < 4; ++j) {
1826                 tcg_temp_free_i64(in[j]);
1827             }
1828         }
1829         if (oprsz < maxsz) {
1830             expand_clr(dofs + oprsz, maxsz - oprsz);
1831         }
1832     } else {
1833         g_assert_not_reached();
1834     }
1835 }
1836 
1837 void tcg_gen_gvec_dup_imm(unsigned vece, uint32_t dofs, uint32_t oprsz,
1838                           uint32_t maxsz, uint64_t x)
1839 {
1840     check_size_align(oprsz, maxsz, dofs);
1841     do_dup(vece, dofs, oprsz, maxsz, NULL, NULL, x);
1842 }
1843 
1844 void tcg_gen_gvec_not(unsigned vece, uint32_t dofs, uint32_t aofs,
1845                       uint32_t oprsz, uint32_t maxsz)
1846 {
1847     static const GVecGen2 g = {
1848         .fni8 = tcg_gen_not_i64,
1849         .fniv = tcg_gen_not_vec,
1850         .fno = gen_helper_gvec_not,
1851         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1852     };
1853     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1854 }
1855 
1856 /* Perform a vector addition using normal addition and a mask.  The mask
1857    should be the sign bit of each lane.  This 6-operation form is more
1858    efficient than separate additions when there are 4 or more lanes in
1859    the 64-bit operation.  */
1860 static void gen_addv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
1861 {
1862     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
1863     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
1864     TCGv_i64 t3 = tcg_temp_ebb_new_i64();
1865 
1866     tcg_gen_andc_i64(t1, a, m);
1867     tcg_gen_andc_i64(t2, b, m);
1868     tcg_gen_xor_i64(t3, a, b);
1869     tcg_gen_add_i64(d, t1, t2);
1870     tcg_gen_and_i64(t3, t3, m);
1871     tcg_gen_xor_i64(d, d, t3);
1872 
1873     tcg_temp_free_i64(t1);
1874     tcg_temp_free_i64(t2);
1875     tcg_temp_free_i64(t3);
1876 }
1877 
1878 void tcg_gen_vec_add8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1879 {
1880     TCGv_i64 m = tcg_constant_i64(dup_const(MO_8, 0x80));
1881     gen_addv_mask(d, a, b, m);
1882 }
1883 
1884 void tcg_gen_vec_add8_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
1885 {
1886     TCGv_i32 m = tcg_constant_i32((int32_t)dup_const(MO_8, 0x80));
1887     TCGv_i32 t1 = tcg_temp_ebb_new_i32();
1888     TCGv_i32 t2 = tcg_temp_ebb_new_i32();
1889     TCGv_i32 t3 = tcg_temp_ebb_new_i32();
1890 
1891     tcg_gen_andc_i32(t1, a, m);
1892     tcg_gen_andc_i32(t2, b, m);
1893     tcg_gen_xor_i32(t3, a, b);
1894     tcg_gen_add_i32(d, t1, t2);
1895     tcg_gen_and_i32(t3, t3, m);
1896     tcg_gen_xor_i32(d, d, t3);
1897 
1898     tcg_temp_free_i32(t1);
1899     tcg_temp_free_i32(t2);
1900     tcg_temp_free_i32(t3);
1901 }
1902 
1903 void tcg_gen_vec_add16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1904 {
1905     TCGv_i64 m = tcg_constant_i64(dup_const(MO_16, 0x8000));
1906     gen_addv_mask(d, a, b, m);
1907 }
1908 
1909 void tcg_gen_vec_add16_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
1910 {
1911     TCGv_i32 t1 = tcg_temp_ebb_new_i32();
1912     TCGv_i32 t2 = tcg_temp_ebb_new_i32();
1913 
1914     tcg_gen_andi_i32(t1, a, ~0xffff);
1915     tcg_gen_add_i32(t2, a, b);
1916     tcg_gen_add_i32(t1, t1, b);
1917     tcg_gen_deposit_i32(d, t1, t2, 0, 16);
1918 
1919     tcg_temp_free_i32(t1);
1920     tcg_temp_free_i32(t2);
1921 }
1922 
1923 void tcg_gen_vec_add32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1924 {
1925     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
1926     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
1927 
1928     tcg_gen_andi_i64(t1, a, ~0xffffffffull);
1929     tcg_gen_add_i64(t2, a, b);
1930     tcg_gen_add_i64(t1, t1, b);
1931     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
1932 
1933     tcg_temp_free_i64(t1);
1934     tcg_temp_free_i64(t2);
1935 }
1936 
1937 static const TCGOpcode vecop_list_add[] = { INDEX_op_add_vec, 0 };
1938 
1939 void tcg_gen_gvec_add(unsigned vece, uint32_t dofs, uint32_t aofs,
1940                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1941 {
1942     static const GVecGen3 g[4] = {
1943         { .fni8 = tcg_gen_vec_add8_i64,
1944           .fniv = tcg_gen_add_vec,
1945           .fno = gen_helper_gvec_add8,
1946           .opt_opc = vecop_list_add,
1947           .vece = MO_8 },
1948         { .fni8 = tcg_gen_vec_add16_i64,
1949           .fniv = tcg_gen_add_vec,
1950           .fno = gen_helper_gvec_add16,
1951           .opt_opc = vecop_list_add,
1952           .vece = MO_16 },
1953         { .fni4 = tcg_gen_add_i32,
1954           .fniv = tcg_gen_add_vec,
1955           .fno = gen_helper_gvec_add32,
1956           .opt_opc = vecop_list_add,
1957           .vece = MO_32 },
1958         { .fni8 = tcg_gen_add_i64,
1959           .fniv = tcg_gen_add_vec,
1960           .fno = gen_helper_gvec_add64,
1961           .opt_opc = vecop_list_add,
1962           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1963           .vece = MO_64 },
1964     };
1965 
1966     tcg_debug_assert(vece <= MO_64);
1967     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1968 }
1969 
1970 void tcg_gen_gvec_adds(unsigned vece, uint32_t dofs, uint32_t aofs,
1971                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1972 {
1973     static const GVecGen2s g[4] = {
1974         { .fni8 = tcg_gen_vec_add8_i64,
1975           .fniv = tcg_gen_add_vec,
1976           .fno = gen_helper_gvec_adds8,
1977           .opt_opc = vecop_list_add,
1978           .vece = MO_8 },
1979         { .fni8 = tcg_gen_vec_add16_i64,
1980           .fniv = tcg_gen_add_vec,
1981           .fno = gen_helper_gvec_adds16,
1982           .opt_opc = vecop_list_add,
1983           .vece = MO_16 },
1984         { .fni4 = tcg_gen_add_i32,
1985           .fniv = tcg_gen_add_vec,
1986           .fno = gen_helper_gvec_adds32,
1987           .opt_opc = vecop_list_add,
1988           .vece = MO_32 },
1989         { .fni8 = tcg_gen_add_i64,
1990           .fniv = tcg_gen_add_vec,
1991           .fno = gen_helper_gvec_adds64,
1992           .opt_opc = vecop_list_add,
1993           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1994           .vece = MO_64 },
1995     };
1996 
1997     tcg_debug_assert(vece <= MO_64);
1998     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
1999 }
2000 
2001 void tcg_gen_gvec_addi(unsigned vece, uint32_t dofs, uint32_t aofs,
2002                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2003 {
2004     TCGv_i64 tmp = tcg_constant_i64(c);
2005     tcg_gen_gvec_adds(vece, dofs, aofs, tmp, oprsz, maxsz);
2006 }
2007 
2008 static const TCGOpcode vecop_list_sub[] = { INDEX_op_sub_vec, 0 };
2009 
2010 void tcg_gen_gvec_subs(unsigned vece, uint32_t dofs, uint32_t aofs,
2011                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2012 {
2013     static const GVecGen2s g[4] = {
2014         { .fni8 = tcg_gen_vec_sub8_i64,
2015           .fniv = tcg_gen_sub_vec,
2016           .fno = gen_helper_gvec_subs8,
2017           .opt_opc = vecop_list_sub,
2018           .vece = MO_8 },
2019         { .fni8 = tcg_gen_vec_sub16_i64,
2020           .fniv = tcg_gen_sub_vec,
2021           .fno = gen_helper_gvec_subs16,
2022           .opt_opc = vecop_list_sub,
2023           .vece = MO_16 },
2024         { .fni4 = tcg_gen_sub_i32,
2025           .fniv = tcg_gen_sub_vec,
2026           .fno = gen_helper_gvec_subs32,
2027           .opt_opc = vecop_list_sub,
2028           .vece = MO_32 },
2029         { .fni8 = tcg_gen_sub_i64,
2030           .fniv = tcg_gen_sub_vec,
2031           .fno = gen_helper_gvec_subs64,
2032           .opt_opc = vecop_list_sub,
2033           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2034           .vece = MO_64 },
2035     };
2036 
2037     tcg_debug_assert(vece <= MO_64);
2038     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
2039 }
2040 
2041 /* Perform a vector subtraction using normal subtraction and a mask.
2042    Compare gen_addv_mask above.  */
2043 static void gen_subv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
2044 {
2045     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
2046     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
2047     TCGv_i64 t3 = tcg_temp_ebb_new_i64();
2048 
2049     tcg_gen_or_i64(t1, a, m);
2050     tcg_gen_andc_i64(t2, b, m);
2051     tcg_gen_eqv_i64(t3, a, b);
2052     tcg_gen_sub_i64(d, t1, t2);
2053     tcg_gen_and_i64(t3, t3, m);
2054     tcg_gen_xor_i64(d, d, t3);
2055 
2056     tcg_temp_free_i64(t1);
2057     tcg_temp_free_i64(t2);
2058     tcg_temp_free_i64(t3);
2059 }
2060 
2061 void tcg_gen_vec_sub8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2062 {
2063     TCGv_i64 m = tcg_constant_i64(dup_const(MO_8, 0x80));
2064     gen_subv_mask(d, a, b, m);
2065 }
2066 
2067 void tcg_gen_vec_sub8_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2068 {
2069     TCGv_i32 m = tcg_constant_i32((int32_t)dup_const(MO_8, 0x80));
2070     TCGv_i32 t1 = tcg_temp_ebb_new_i32();
2071     TCGv_i32 t2 = tcg_temp_ebb_new_i32();
2072     TCGv_i32 t3 = tcg_temp_ebb_new_i32();
2073 
2074     tcg_gen_or_i32(t1, a, m);
2075     tcg_gen_andc_i32(t2, b, m);
2076     tcg_gen_eqv_i32(t3, a, b);
2077     tcg_gen_sub_i32(d, t1, t2);
2078     tcg_gen_and_i32(t3, t3, m);
2079     tcg_gen_xor_i32(d, d, t3);
2080 
2081     tcg_temp_free_i32(t1);
2082     tcg_temp_free_i32(t2);
2083     tcg_temp_free_i32(t3);
2084 }
2085 
2086 void tcg_gen_vec_sub16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2087 {
2088     TCGv_i64 m = tcg_constant_i64(dup_const(MO_16, 0x8000));
2089     gen_subv_mask(d, a, b, m);
2090 }
2091 
2092 void tcg_gen_vec_sub16_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2093 {
2094     TCGv_i32 t1 = tcg_temp_ebb_new_i32();
2095     TCGv_i32 t2 = tcg_temp_ebb_new_i32();
2096 
2097     tcg_gen_andi_i32(t1, b, ~0xffff);
2098     tcg_gen_sub_i32(t2, a, b);
2099     tcg_gen_sub_i32(t1, a, t1);
2100     tcg_gen_deposit_i32(d, t1, t2, 0, 16);
2101 
2102     tcg_temp_free_i32(t1);
2103     tcg_temp_free_i32(t2);
2104 }
2105 
2106 void tcg_gen_vec_sub32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2107 {
2108     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
2109     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
2110 
2111     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
2112     tcg_gen_sub_i64(t2, a, b);
2113     tcg_gen_sub_i64(t1, a, t1);
2114     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
2115 
2116     tcg_temp_free_i64(t1);
2117     tcg_temp_free_i64(t2);
2118 }
2119 
2120 void tcg_gen_gvec_sub(unsigned vece, uint32_t dofs, uint32_t aofs,
2121                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2122 {
2123     static const GVecGen3 g[4] = {
2124         { .fni8 = tcg_gen_vec_sub8_i64,
2125           .fniv = tcg_gen_sub_vec,
2126           .fno = gen_helper_gvec_sub8,
2127           .opt_opc = vecop_list_sub,
2128           .vece = MO_8 },
2129         { .fni8 = tcg_gen_vec_sub16_i64,
2130           .fniv = tcg_gen_sub_vec,
2131           .fno = gen_helper_gvec_sub16,
2132           .opt_opc = vecop_list_sub,
2133           .vece = MO_16 },
2134         { .fni4 = tcg_gen_sub_i32,
2135           .fniv = tcg_gen_sub_vec,
2136           .fno = gen_helper_gvec_sub32,
2137           .opt_opc = vecop_list_sub,
2138           .vece = MO_32 },
2139         { .fni8 = tcg_gen_sub_i64,
2140           .fniv = tcg_gen_sub_vec,
2141           .fno = gen_helper_gvec_sub64,
2142           .opt_opc = vecop_list_sub,
2143           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2144           .vece = MO_64 },
2145     };
2146 
2147     tcg_debug_assert(vece <= MO_64);
2148     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2149 }
2150 
2151 static const TCGOpcode vecop_list_mul[] = { INDEX_op_mul_vec, 0 };
2152 
2153 void tcg_gen_gvec_mul(unsigned vece, uint32_t dofs, uint32_t aofs,
2154                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2155 {
2156     static const GVecGen3 g[4] = {
2157         { .fniv = tcg_gen_mul_vec,
2158           .fno = gen_helper_gvec_mul8,
2159           .opt_opc = vecop_list_mul,
2160           .vece = MO_8 },
2161         { .fniv = tcg_gen_mul_vec,
2162           .fno = gen_helper_gvec_mul16,
2163           .opt_opc = vecop_list_mul,
2164           .vece = MO_16 },
2165         { .fni4 = tcg_gen_mul_i32,
2166           .fniv = tcg_gen_mul_vec,
2167           .fno = gen_helper_gvec_mul32,
2168           .opt_opc = vecop_list_mul,
2169           .vece = MO_32 },
2170         { .fni8 = tcg_gen_mul_i64,
2171           .fniv = tcg_gen_mul_vec,
2172           .fno = gen_helper_gvec_mul64,
2173           .opt_opc = vecop_list_mul,
2174           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2175           .vece = MO_64 },
2176     };
2177 
2178     tcg_debug_assert(vece <= MO_64);
2179     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2180 }
2181 
2182 void tcg_gen_gvec_muls(unsigned vece, uint32_t dofs, uint32_t aofs,
2183                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2184 {
2185     static const GVecGen2s g[4] = {
2186         { .fniv = tcg_gen_mul_vec,
2187           .fno = gen_helper_gvec_muls8,
2188           .opt_opc = vecop_list_mul,
2189           .vece = MO_8 },
2190         { .fniv = tcg_gen_mul_vec,
2191           .fno = gen_helper_gvec_muls16,
2192           .opt_opc = vecop_list_mul,
2193           .vece = MO_16 },
2194         { .fni4 = tcg_gen_mul_i32,
2195           .fniv = tcg_gen_mul_vec,
2196           .fno = gen_helper_gvec_muls32,
2197           .opt_opc = vecop_list_mul,
2198           .vece = MO_32 },
2199         { .fni8 = tcg_gen_mul_i64,
2200           .fniv = tcg_gen_mul_vec,
2201           .fno = gen_helper_gvec_muls64,
2202           .opt_opc = vecop_list_mul,
2203           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2204           .vece = MO_64 },
2205     };
2206 
2207     tcg_debug_assert(vece <= MO_64);
2208     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
2209 }
2210 
2211 void tcg_gen_gvec_muli(unsigned vece, uint32_t dofs, uint32_t aofs,
2212                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2213 {
2214     TCGv_i64 tmp = tcg_constant_i64(c);
2215     tcg_gen_gvec_muls(vece, dofs, aofs, tmp, oprsz, maxsz);
2216 }
2217 
2218 void tcg_gen_gvec_ssadd(unsigned vece, uint32_t dofs, uint32_t aofs,
2219                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2220 {
2221     static const TCGOpcode vecop_list[] = { INDEX_op_ssadd_vec, 0 };
2222     static const GVecGen3 g[4] = {
2223         { .fniv = tcg_gen_ssadd_vec,
2224           .fno = gen_helper_gvec_ssadd8,
2225           .opt_opc = vecop_list,
2226           .vece = MO_8 },
2227         { .fniv = tcg_gen_ssadd_vec,
2228           .fno = gen_helper_gvec_ssadd16,
2229           .opt_opc = vecop_list,
2230           .vece = MO_16 },
2231         { .fniv = tcg_gen_ssadd_vec,
2232           .fno = gen_helper_gvec_ssadd32,
2233           .opt_opc = vecop_list,
2234           .vece = MO_32 },
2235         { .fniv = tcg_gen_ssadd_vec,
2236           .fno = gen_helper_gvec_ssadd64,
2237           .opt_opc = vecop_list,
2238           .vece = MO_64 },
2239     };
2240     tcg_debug_assert(vece <= MO_64);
2241     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2242 }
2243 
2244 void tcg_gen_gvec_sssub(unsigned vece, uint32_t dofs, uint32_t aofs,
2245                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2246 {
2247     static const TCGOpcode vecop_list[] = { INDEX_op_sssub_vec, 0 };
2248     static const GVecGen3 g[4] = {
2249         { .fniv = tcg_gen_sssub_vec,
2250           .fno = gen_helper_gvec_sssub8,
2251           .opt_opc = vecop_list,
2252           .vece = MO_8 },
2253         { .fniv = tcg_gen_sssub_vec,
2254           .fno = gen_helper_gvec_sssub16,
2255           .opt_opc = vecop_list,
2256           .vece = MO_16 },
2257         { .fniv = tcg_gen_sssub_vec,
2258           .fno = gen_helper_gvec_sssub32,
2259           .opt_opc = vecop_list,
2260           .vece = MO_32 },
2261         { .fniv = tcg_gen_sssub_vec,
2262           .fno = gen_helper_gvec_sssub64,
2263           .opt_opc = vecop_list,
2264           .vece = MO_64 },
2265     };
2266     tcg_debug_assert(vece <= MO_64);
2267     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2268 }
2269 
2270 static void tcg_gen_usadd_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2271 {
2272     TCGv_i32 max = tcg_constant_i32(-1);
2273     tcg_gen_add_i32(d, a, b);
2274     tcg_gen_movcond_i32(TCG_COND_LTU, d, d, a, max, d);
2275 }
2276 
2277 static void tcg_gen_usadd_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2278 {
2279     TCGv_i64 max = tcg_constant_i64(-1);
2280     tcg_gen_add_i64(d, a, b);
2281     tcg_gen_movcond_i64(TCG_COND_LTU, d, d, a, max, d);
2282 }
2283 
2284 void tcg_gen_gvec_usadd(unsigned vece, uint32_t dofs, uint32_t aofs,
2285                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2286 {
2287     static const TCGOpcode vecop_list[] = { INDEX_op_usadd_vec, 0 };
2288     static const GVecGen3 g[4] = {
2289         { .fniv = tcg_gen_usadd_vec,
2290           .fno = gen_helper_gvec_usadd8,
2291           .opt_opc = vecop_list,
2292           .vece = MO_8 },
2293         { .fniv = tcg_gen_usadd_vec,
2294           .fno = gen_helper_gvec_usadd16,
2295           .opt_opc = vecop_list,
2296           .vece = MO_16 },
2297         { .fni4 = tcg_gen_usadd_i32,
2298           .fniv = tcg_gen_usadd_vec,
2299           .fno = gen_helper_gvec_usadd32,
2300           .opt_opc = vecop_list,
2301           .vece = MO_32 },
2302         { .fni8 = tcg_gen_usadd_i64,
2303           .fniv = tcg_gen_usadd_vec,
2304           .fno = gen_helper_gvec_usadd64,
2305           .opt_opc = vecop_list,
2306           .vece = MO_64 }
2307     };
2308     tcg_debug_assert(vece <= MO_64);
2309     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2310 }
2311 
2312 static void tcg_gen_ussub_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2313 {
2314     TCGv_i32 min = tcg_constant_i32(0);
2315     tcg_gen_sub_i32(d, a, b);
2316     tcg_gen_movcond_i32(TCG_COND_LTU, d, a, b, min, d);
2317 }
2318 
2319 static void tcg_gen_ussub_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2320 {
2321     TCGv_i64 min = tcg_constant_i64(0);
2322     tcg_gen_sub_i64(d, a, b);
2323     tcg_gen_movcond_i64(TCG_COND_LTU, d, a, b, min, d);
2324 }
2325 
2326 void tcg_gen_gvec_ussub(unsigned vece, uint32_t dofs, uint32_t aofs,
2327                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2328 {
2329     static const TCGOpcode vecop_list[] = { INDEX_op_ussub_vec, 0 };
2330     static const GVecGen3 g[4] = {
2331         { .fniv = tcg_gen_ussub_vec,
2332           .fno = gen_helper_gvec_ussub8,
2333           .opt_opc = vecop_list,
2334           .vece = MO_8 },
2335         { .fniv = tcg_gen_ussub_vec,
2336           .fno = gen_helper_gvec_ussub16,
2337           .opt_opc = vecop_list,
2338           .vece = MO_16 },
2339         { .fni4 = tcg_gen_ussub_i32,
2340           .fniv = tcg_gen_ussub_vec,
2341           .fno = gen_helper_gvec_ussub32,
2342           .opt_opc = vecop_list,
2343           .vece = MO_32 },
2344         { .fni8 = tcg_gen_ussub_i64,
2345           .fniv = tcg_gen_ussub_vec,
2346           .fno = gen_helper_gvec_ussub64,
2347           .opt_opc = vecop_list,
2348           .vece = MO_64 }
2349     };
2350     tcg_debug_assert(vece <= MO_64);
2351     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2352 }
2353 
2354 void tcg_gen_gvec_smin(unsigned vece, uint32_t dofs, uint32_t aofs,
2355                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2356 {
2357     static const TCGOpcode vecop_list[] = { INDEX_op_smin_vec, 0 };
2358     static const GVecGen3 g[4] = {
2359         { .fniv = tcg_gen_smin_vec,
2360           .fno = gen_helper_gvec_smin8,
2361           .opt_opc = vecop_list,
2362           .vece = MO_8 },
2363         { .fniv = tcg_gen_smin_vec,
2364           .fno = gen_helper_gvec_smin16,
2365           .opt_opc = vecop_list,
2366           .vece = MO_16 },
2367         { .fni4 = tcg_gen_smin_i32,
2368           .fniv = tcg_gen_smin_vec,
2369           .fno = gen_helper_gvec_smin32,
2370           .opt_opc = vecop_list,
2371           .vece = MO_32 },
2372         { .fni8 = tcg_gen_smin_i64,
2373           .fniv = tcg_gen_smin_vec,
2374           .fno = gen_helper_gvec_smin64,
2375           .opt_opc = vecop_list,
2376           .vece = MO_64 }
2377     };
2378     tcg_debug_assert(vece <= MO_64);
2379     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2380 }
2381 
2382 void tcg_gen_gvec_umin(unsigned vece, uint32_t dofs, uint32_t aofs,
2383                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2384 {
2385     static const TCGOpcode vecop_list[] = { INDEX_op_umin_vec, 0 };
2386     static const GVecGen3 g[4] = {
2387         { .fniv = tcg_gen_umin_vec,
2388           .fno = gen_helper_gvec_umin8,
2389           .opt_opc = vecop_list,
2390           .vece = MO_8 },
2391         { .fniv = tcg_gen_umin_vec,
2392           .fno = gen_helper_gvec_umin16,
2393           .opt_opc = vecop_list,
2394           .vece = MO_16 },
2395         { .fni4 = tcg_gen_umin_i32,
2396           .fniv = tcg_gen_umin_vec,
2397           .fno = gen_helper_gvec_umin32,
2398           .opt_opc = vecop_list,
2399           .vece = MO_32 },
2400         { .fni8 = tcg_gen_umin_i64,
2401           .fniv = tcg_gen_umin_vec,
2402           .fno = gen_helper_gvec_umin64,
2403           .opt_opc = vecop_list,
2404           .vece = MO_64 }
2405     };
2406     tcg_debug_assert(vece <= MO_64);
2407     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2408 }
2409 
2410 void tcg_gen_gvec_smax(unsigned vece, uint32_t dofs, uint32_t aofs,
2411                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2412 {
2413     static const TCGOpcode vecop_list[] = { INDEX_op_smax_vec, 0 };
2414     static const GVecGen3 g[4] = {
2415         { .fniv = tcg_gen_smax_vec,
2416           .fno = gen_helper_gvec_smax8,
2417           .opt_opc = vecop_list,
2418           .vece = MO_8 },
2419         { .fniv = tcg_gen_smax_vec,
2420           .fno = gen_helper_gvec_smax16,
2421           .opt_opc = vecop_list,
2422           .vece = MO_16 },
2423         { .fni4 = tcg_gen_smax_i32,
2424           .fniv = tcg_gen_smax_vec,
2425           .fno = gen_helper_gvec_smax32,
2426           .opt_opc = vecop_list,
2427           .vece = MO_32 },
2428         { .fni8 = tcg_gen_smax_i64,
2429           .fniv = tcg_gen_smax_vec,
2430           .fno = gen_helper_gvec_smax64,
2431           .opt_opc = vecop_list,
2432           .vece = MO_64 }
2433     };
2434     tcg_debug_assert(vece <= MO_64);
2435     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2436 }
2437 
2438 void tcg_gen_gvec_umax(unsigned vece, uint32_t dofs, uint32_t aofs,
2439                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2440 {
2441     static const TCGOpcode vecop_list[] = { INDEX_op_umax_vec, 0 };
2442     static const GVecGen3 g[4] = {
2443         { .fniv = tcg_gen_umax_vec,
2444           .fno = gen_helper_gvec_umax8,
2445           .opt_opc = vecop_list,
2446           .vece = MO_8 },
2447         { .fniv = tcg_gen_umax_vec,
2448           .fno = gen_helper_gvec_umax16,
2449           .opt_opc = vecop_list,
2450           .vece = MO_16 },
2451         { .fni4 = tcg_gen_umax_i32,
2452           .fniv = tcg_gen_umax_vec,
2453           .fno = gen_helper_gvec_umax32,
2454           .opt_opc = vecop_list,
2455           .vece = MO_32 },
2456         { .fni8 = tcg_gen_umax_i64,
2457           .fniv = tcg_gen_umax_vec,
2458           .fno = gen_helper_gvec_umax64,
2459           .opt_opc = vecop_list,
2460           .vece = MO_64 }
2461     };
2462     tcg_debug_assert(vece <= MO_64);
2463     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2464 }
2465 
2466 /* Perform a vector negation using normal negation and a mask.
2467    Compare gen_subv_mask above.  */
2468 static void gen_negv_mask(TCGv_i64 d, TCGv_i64 b, TCGv_i64 m)
2469 {
2470     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
2471     TCGv_i64 t3 = tcg_temp_ebb_new_i64();
2472 
2473     tcg_gen_andc_i64(t3, m, b);
2474     tcg_gen_andc_i64(t2, b, m);
2475     tcg_gen_sub_i64(d, m, t2);
2476     tcg_gen_xor_i64(d, d, t3);
2477 
2478     tcg_temp_free_i64(t2);
2479     tcg_temp_free_i64(t3);
2480 }
2481 
2482 void tcg_gen_vec_neg8_i64(TCGv_i64 d, TCGv_i64 b)
2483 {
2484     TCGv_i64 m = tcg_constant_i64(dup_const(MO_8, 0x80));
2485     gen_negv_mask(d, b, m);
2486 }
2487 
2488 void tcg_gen_vec_neg16_i64(TCGv_i64 d, TCGv_i64 b)
2489 {
2490     TCGv_i64 m = tcg_constant_i64(dup_const(MO_16, 0x8000));
2491     gen_negv_mask(d, b, m);
2492 }
2493 
2494 void tcg_gen_vec_neg32_i64(TCGv_i64 d, TCGv_i64 b)
2495 {
2496     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
2497     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
2498 
2499     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
2500     tcg_gen_neg_i64(t2, b);
2501     tcg_gen_neg_i64(t1, t1);
2502     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
2503 
2504     tcg_temp_free_i64(t1);
2505     tcg_temp_free_i64(t2);
2506 }
2507 
2508 void tcg_gen_gvec_neg(unsigned vece, uint32_t dofs, uint32_t aofs,
2509                       uint32_t oprsz, uint32_t maxsz)
2510 {
2511     static const TCGOpcode vecop_list[] = { INDEX_op_neg_vec, 0 };
2512     static const GVecGen2 g[4] = {
2513         { .fni8 = tcg_gen_vec_neg8_i64,
2514           .fniv = tcg_gen_neg_vec,
2515           .fno = gen_helper_gvec_neg8,
2516           .opt_opc = vecop_list,
2517           .vece = MO_8 },
2518         { .fni8 = tcg_gen_vec_neg16_i64,
2519           .fniv = tcg_gen_neg_vec,
2520           .fno = gen_helper_gvec_neg16,
2521           .opt_opc = vecop_list,
2522           .vece = MO_16 },
2523         { .fni4 = tcg_gen_neg_i32,
2524           .fniv = tcg_gen_neg_vec,
2525           .fno = gen_helper_gvec_neg32,
2526           .opt_opc = vecop_list,
2527           .vece = MO_32 },
2528         { .fni8 = tcg_gen_neg_i64,
2529           .fniv = tcg_gen_neg_vec,
2530           .fno = gen_helper_gvec_neg64,
2531           .opt_opc = vecop_list,
2532           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2533           .vece = MO_64 },
2534     };
2535 
2536     tcg_debug_assert(vece <= MO_64);
2537     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g[vece]);
2538 }
2539 
2540 static void gen_absv_mask(TCGv_i64 d, TCGv_i64 b, unsigned vece)
2541 {
2542     TCGv_i64 t = tcg_temp_ebb_new_i64();
2543     int nbit = 8 << vece;
2544 
2545     /* Create -1 for each negative element.  */
2546     tcg_gen_shri_i64(t, b, nbit - 1);
2547     tcg_gen_andi_i64(t, t, dup_const(vece, 1));
2548     tcg_gen_muli_i64(t, t, (1 << nbit) - 1);
2549 
2550     /*
2551      * Invert (via xor -1) and add one.
2552      * Because of the ordering the msb is cleared,
2553      * so we never have carry into the next element.
2554      */
2555     tcg_gen_xor_i64(d, b, t);
2556     tcg_gen_andi_i64(t, t, dup_const(vece, 1));
2557     tcg_gen_add_i64(d, d, t);
2558 
2559     tcg_temp_free_i64(t);
2560 }
2561 
2562 static void tcg_gen_vec_abs8_i64(TCGv_i64 d, TCGv_i64 b)
2563 {
2564     gen_absv_mask(d, b, MO_8);
2565 }
2566 
2567 static void tcg_gen_vec_abs16_i64(TCGv_i64 d, TCGv_i64 b)
2568 {
2569     gen_absv_mask(d, b, MO_16);
2570 }
2571 
2572 void tcg_gen_gvec_abs(unsigned vece, uint32_t dofs, uint32_t aofs,
2573                       uint32_t oprsz, uint32_t maxsz)
2574 {
2575     static const TCGOpcode vecop_list[] = { INDEX_op_abs_vec, 0 };
2576     static const GVecGen2 g[4] = {
2577         { .fni8 = tcg_gen_vec_abs8_i64,
2578           .fniv = tcg_gen_abs_vec,
2579           .fno = gen_helper_gvec_abs8,
2580           .opt_opc = vecop_list,
2581           .vece = MO_8 },
2582         { .fni8 = tcg_gen_vec_abs16_i64,
2583           .fniv = tcg_gen_abs_vec,
2584           .fno = gen_helper_gvec_abs16,
2585           .opt_opc = vecop_list,
2586           .vece = MO_16 },
2587         { .fni4 = tcg_gen_abs_i32,
2588           .fniv = tcg_gen_abs_vec,
2589           .fno = gen_helper_gvec_abs32,
2590           .opt_opc = vecop_list,
2591           .vece = MO_32 },
2592         { .fni8 = tcg_gen_abs_i64,
2593           .fniv = tcg_gen_abs_vec,
2594           .fno = gen_helper_gvec_abs64,
2595           .opt_opc = vecop_list,
2596           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2597           .vece = MO_64 },
2598     };
2599 
2600     tcg_debug_assert(vece <= MO_64);
2601     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g[vece]);
2602 }
2603 
2604 void tcg_gen_gvec_and(unsigned vece, uint32_t dofs, uint32_t aofs,
2605                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2606 {
2607     static const GVecGen3 g = {
2608         .fni8 = tcg_gen_and_i64,
2609         .fniv = tcg_gen_and_vec,
2610         .fno = gen_helper_gvec_and,
2611         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2612     };
2613 
2614     if (aofs == bofs) {
2615         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2616     } else {
2617         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2618     }
2619 }
2620 
2621 void tcg_gen_gvec_or(unsigned vece, uint32_t dofs, uint32_t aofs,
2622                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2623 {
2624     static const GVecGen3 g = {
2625         .fni8 = tcg_gen_or_i64,
2626         .fniv = tcg_gen_or_vec,
2627         .fno = gen_helper_gvec_or,
2628         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2629     };
2630 
2631     if (aofs == bofs) {
2632         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2633     } else {
2634         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2635     }
2636 }
2637 
2638 void tcg_gen_gvec_xor(unsigned vece, uint32_t dofs, uint32_t aofs,
2639                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2640 {
2641     static const GVecGen3 g = {
2642         .fni8 = tcg_gen_xor_i64,
2643         .fniv = tcg_gen_xor_vec,
2644         .fno = gen_helper_gvec_xor,
2645         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2646     };
2647 
2648     if (aofs == bofs) {
2649         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, 0);
2650     } else {
2651         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2652     }
2653 }
2654 
2655 void tcg_gen_gvec_andc(unsigned vece, uint32_t dofs, uint32_t aofs,
2656                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2657 {
2658     static const GVecGen3 g = {
2659         .fni8 = tcg_gen_andc_i64,
2660         .fniv = tcg_gen_andc_vec,
2661         .fno = gen_helper_gvec_andc,
2662         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2663     };
2664 
2665     if (aofs == bofs) {
2666         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, 0);
2667     } else {
2668         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2669     }
2670 }
2671 
2672 void tcg_gen_gvec_orc(unsigned vece, uint32_t dofs, uint32_t aofs,
2673                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2674 {
2675     static const GVecGen3 g = {
2676         .fni8 = tcg_gen_orc_i64,
2677         .fniv = tcg_gen_orc_vec,
2678         .fno = gen_helper_gvec_orc,
2679         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2680     };
2681 
2682     if (aofs == bofs) {
2683         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, -1);
2684     } else {
2685         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2686     }
2687 }
2688 
2689 void tcg_gen_gvec_nand(unsigned vece, uint32_t dofs, uint32_t aofs,
2690                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2691 {
2692     static const GVecGen3 g = {
2693         .fni8 = tcg_gen_nand_i64,
2694         .fniv = tcg_gen_nand_vec,
2695         .fno = gen_helper_gvec_nand,
2696         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2697     };
2698 
2699     if (aofs == bofs) {
2700         tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
2701     } else {
2702         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2703     }
2704 }
2705 
2706 void tcg_gen_gvec_nor(unsigned vece, uint32_t dofs, uint32_t aofs,
2707                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2708 {
2709     static const GVecGen3 g = {
2710         .fni8 = tcg_gen_nor_i64,
2711         .fniv = tcg_gen_nor_vec,
2712         .fno = gen_helper_gvec_nor,
2713         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2714     };
2715 
2716     if (aofs == bofs) {
2717         tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
2718     } else {
2719         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2720     }
2721 }
2722 
2723 void tcg_gen_gvec_eqv(unsigned vece, uint32_t dofs, uint32_t aofs,
2724                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2725 {
2726     static const GVecGen3 g = {
2727         .fni8 = tcg_gen_eqv_i64,
2728         .fniv = tcg_gen_eqv_vec,
2729         .fno = gen_helper_gvec_eqv,
2730         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2731     };
2732 
2733     if (aofs == bofs) {
2734         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, -1);
2735     } else {
2736         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2737     }
2738 }
2739 
2740 static const GVecGen2s gop_ands = {
2741     .fni8 = tcg_gen_and_i64,
2742     .fniv = tcg_gen_and_vec,
2743     .fno = gen_helper_gvec_ands,
2744     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2745     .vece = MO_64
2746 };
2747 
2748 void tcg_gen_gvec_ands(unsigned vece, uint32_t dofs, uint32_t aofs,
2749                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2750 {
2751     TCGv_i64 tmp = tcg_temp_ebb_new_i64();
2752     tcg_gen_dup_i64(vece, tmp, c);
2753     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
2754     tcg_temp_free_i64(tmp);
2755 }
2756 
2757 void tcg_gen_gvec_andi(unsigned vece, uint32_t dofs, uint32_t aofs,
2758                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2759 {
2760     TCGv_i64 tmp = tcg_constant_i64(dup_const(vece, c));
2761     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
2762 }
2763 
2764 void tcg_gen_gvec_andcs(unsigned vece, uint32_t dofs, uint32_t aofs,
2765                         TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2766 {
2767     static GVecGen2s g = {
2768         .fni8 = tcg_gen_andc_i64,
2769         .fniv = tcg_gen_andc_vec,
2770         .fno = gen_helper_gvec_andcs,
2771         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2772         .vece = MO_64
2773     };
2774 
2775     TCGv_i64 tmp = tcg_temp_ebb_new_i64();
2776     tcg_gen_dup_i64(vece, tmp, c);
2777     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g);
2778     tcg_temp_free_i64(tmp);
2779 }
2780 
2781 static const GVecGen2s gop_xors = {
2782     .fni8 = tcg_gen_xor_i64,
2783     .fniv = tcg_gen_xor_vec,
2784     .fno = gen_helper_gvec_xors,
2785     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2786     .vece = MO_64
2787 };
2788 
2789 void tcg_gen_gvec_xors(unsigned vece, uint32_t dofs, uint32_t aofs,
2790                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2791 {
2792     TCGv_i64 tmp = tcg_temp_ebb_new_i64();
2793     tcg_gen_dup_i64(vece, tmp, c);
2794     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
2795     tcg_temp_free_i64(tmp);
2796 }
2797 
2798 void tcg_gen_gvec_xori(unsigned vece, uint32_t dofs, uint32_t aofs,
2799                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2800 {
2801     TCGv_i64 tmp = tcg_constant_i64(dup_const(vece, c));
2802     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
2803 }
2804 
2805 static const GVecGen2s gop_ors = {
2806     .fni8 = tcg_gen_or_i64,
2807     .fniv = tcg_gen_or_vec,
2808     .fno = gen_helper_gvec_ors,
2809     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2810     .vece = MO_64
2811 };
2812 
2813 void tcg_gen_gvec_ors(unsigned vece, uint32_t dofs, uint32_t aofs,
2814                       TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2815 {
2816     TCGv_i64 tmp = tcg_temp_ebb_new_i64();
2817     tcg_gen_dup_i64(vece, tmp, c);
2818     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
2819     tcg_temp_free_i64(tmp);
2820 }
2821 
2822 void tcg_gen_gvec_ori(unsigned vece, uint32_t dofs, uint32_t aofs,
2823                       int64_t c, uint32_t oprsz, uint32_t maxsz)
2824 {
2825     TCGv_i64 tmp = tcg_constant_i64(dup_const(vece, c));
2826     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
2827 }
2828 
2829 void tcg_gen_vec_shl8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2830 {
2831     uint64_t mask = dup_const(MO_8, 0xff << c);
2832     tcg_gen_shli_i64(d, a, c);
2833     tcg_gen_andi_i64(d, d, mask);
2834 }
2835 
2836 void tcg_gen_vec_shl16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2837 {
2838     uint64_t mask = dup_const(MO_16, 0xffff << c);
2839     tcg_gen_shli_i64(d, a, c);
2840     tcg_gen_andi_i64(d, d, mask);
2841 }
2842 
2843 void tcg_gen_vec_shl8i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2844 {
2845     uint32_t mask = dup_const(MO_8, 0xff << c);
2846     tcg_gen_shli_i32(d, a, c);
2847     tcg_gen_andi_i32(d, d, mask);
2848 }
2849 
2850 void tcg_gen_vec_shl16i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2851 {
2852     uint32_t mask = dup_const(MO_16, 0xffff << c);
2853     tcg_gen_shli_i32(d, a, c);
2854     tcg_gen_andi_i32(d, d, mask);
2855 }
2856 
2857 void tcg_gen_gvec_shli(unsigned vece, uint32_t dofs, uint32_t aofs,
2858                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2859 {
2860     static const TCGOpcode vecop_list[] = { INDEX_op_shli_vec, 0 };
2861     static const GVecGen2i g[4] = {
2862         { .fni8 = tcg_gen_vec_shl8i_i64,
2863           .fniv = tcg_gen_shli_vec,
2864           .fno = gen_helper_gvec_shl8i,
2865           .opt_opc = vecop_list,
2866           .vece = MO_8 },
2867         { .fni8 = tcg_gen_vec_shl16i_i64,
2868           .fniv = tcg_gen_shli_vec,
2869           .fno = gen_helper_gvec_shl16i,
2870           .opt_opc = vecop_list,
2871           .vece = MO_16 },
2872         { .fni4 = tcg_gen_shli_i32,
2873           .fniv = tcg_gen_shli_vec,
2874           .fno = gen_helper_gvec_shl32i,
2875           .opt_opc = vecop_list,
2876           .vece = MO_32 },
2877         { .fni8 = tcg_gen_shli_i64,
2878           .fniv = tcg_gen_shli_vec,
2879           .fno = gen_helper_gvec_shl64i,
2880           .opt_opc = vecop_list,
2881           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2882           .vece = MO_64 },
2883     };
2884 
2885     tcg_debug_assert(vece <= MO_64);
2886     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2887     if (shift == 0) {
2888         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2889     } else {
2890         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2891     }
2892 }
2893 
2894 void tcg_gen_vec_shr8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2895 {
2896     uint64_t mask = dup_const(MO_8, 0xff >> c);
2897     tcg_gen_shri_i64(d, a, c);
2898     tcg_gen_andi_i64(d, d, mask);
2899 }
2900 
2901 void tcg_gen_vec_shr16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2902 {
2903     uint64_t mask = dup_const(MO_16, 0xffff >> c);
2904     tcg_gen_shri_i64(d, a, c);
2905     tcg_gen_andi_i64(d, d, mask);
2906 }
2907 
2908 void tcg_gen_vec_shr8i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2909 {
2910     uint32_t mask = dup_const(MO_8, 0xff >> c);
2911     tcg_gen_shri_i32(d, a, c);
2912     tcg_gen_andi_i32(d, d, mask);
2913 }
2914 
2915 void tcg_gen_vec_shr16i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2916 {
2917     uint32_t mask = dup_const(MO_16, 0xffff >> c);
2918     tcg_gen_shri_i32(d, a, c);
2919     tcg_gen_andi_i32(d, d, mask);
2920 }
2921 
2922 void tcg_gen_gvec_shri(unsigned vece, uint32_t dofs, uint32_t aofs,
2923                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2924 {
2925     static const TCGOpcode vecop_list[] = { INDEX_op_shri_vec, 0 };
2926     static const GVecGen2i g[4] = {
2927         { .fni8 = tcg_gen_vec_shr8i_i64,
2928           .fniv = tcg_gen_shri_vec,
2929           .fno = gen_helper_gvec_shr8i,
2930           .opt_opc = vecop_list,
2931           .vece = MO_8 },
2932         { .fni8 = tcg_gen_vec_shr16i_i64,
2933           .fniv = tcg_gen_shri_vec,
2934           .fno = gen_helper_gvec_shr16i,
2935           .opt_opc = vecop_list,
2936           .vece = MO_16 },
2937         { .fni4 = tcg_gen_shri_i32,
2938           .fniv = tcg_gen_shri_vec,
2939           .fno = gen_helper_gvec_shr32i,
2940           .opt_opc = vecop_list,
2941           .vece = MO_32 },
2942         { .fni8 = tcg_gen_shri_i64,
2943           .fniv = tcg_gen_shri_vec,
2944           .fno = gen_helper_gvec_shr64i,
2945           .opt_opc = vecop_list,
2946           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2947           .vece = MO_64 },
2948     };
2949 
2950     tcg_debug_assert(vece <= MO_64);
2951     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2952     if (shift == 0) {
2953         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2954     } else {
2955         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2956     }
2957 }
2958 
2959 void tcg_gen_vec_sar8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2960 {
2961     uint64_t s_mask = dup_const(MO_8, 0x80 >> c);
2962     uint64_t c_mask = dup_const(MO_8, 0xff >> c);
2963     TCGv_i64 s = tcg_temp_ebb_new_i64();
2964 
2965     tcg_gen_shri_i64(d, a, c);
2966     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2967     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2968     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2969     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2970     tcg_temp_free_i64(s);
2971 }
2972 
2973 void tcg_gen_vec_sar16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2974 {
2975     uint64_t s_mask = dup_const(MO_16, 0x8000 >> c);
2976     uint64_t c_mask = dup_const(MO_16, 0xffff >> c);
2977     TCGv_i64 s = tcg_temp_ebb_new_i64();
2978 
2979     tcg_gen_shri_i64(d, a, c);
2980     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2981     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2982     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2983     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2984     tcg_temp_free_i64(s);
2985 }
2986 
2987 void tcg_gen_vec_sar8i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2988 {
2989     uint32_t s_mask = dup_const(MO_8, 0x80 >> c);
2990     uint32_t c_mask = dup_const(MO_8, 0xff >> c);
2991     TCGv_i32 s = tcg_temp_ebb_new_i32();
2992 
2993     tcg_gen_shri_i32(d, a, c);
2994     tcg_gen_andi_i32(s, d, s_mask);  /* isolate (shifted) sign bit */
2995     tcg_gen_muli_i32(s, s, (2 << c) - 2); /* replicate isolated signs */
2996     tcg_gen_andi_i32(d, d, c_mask);  /* clear out bits above sign  */
2997     tcg_gen_or_i32(d, d, s);         /* include sign extension */
2998     tcg_temp_free_i32(s);
2999 }
3000 
3001 void tcg_gen_vec_sar16i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
3002 {
3003     uint32_t s_mask = dup_const(MO_16, 0x8000 >> c);
3004     uint32_t c_mask = dup_const(MO_16, 0xffff >> c);
3005     TCGv_i32 s = tcg_temp_ebb_new_i32();
3006 
3007     tcg_gen_shri_i32(d, a, c);
3008     tcg_gen_andi_i32(s, d, s_mask);  /* isolate (shifted) sign bit */
3009     tcg_gen_andi_i32(d, d, c_mask);  /* clear out bits above sign  */
3010     tcg_gen_muli_i32(s, s, (2 << c) - 2); /* replicate isolated signs */
3011     tcg_gen_or_i32(d, d, s);         /* include sign extension */
3012     tcg_temp_free_i32(s);
3013 }
3014 
3015 void tcg_gen_gvec_sari(unsigned vece, uint32_t dofs, uint32_t aofs,
3016                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
3017 {
3018     static const TCGOpcode vecop_list[] = { INDEX_op_sari_vec, 0 };
3019     static const GVecGen2i g[4] = {
3020         { .fni8 = tcg_gen_vec_sar8i_i64,
3021           .fniv = tcg_gen_sari_vec,
3022           .fno = gen_helper_gvec_sar8i,
3023           .opt_opc = vecop_list,
3024           .vece = MO_8 },
3025         { .fni8 = tcg_gen_vec_sar16i_i64,
3026           .fniv = tcg_gen_sari_vec,
3027           .fno = gen_helper_gvec_sar16i,
3028           .opt_opc = vecop_list,
3029           .vece = MO_16 },
3030         { .fni4 = tcg_gen_sari_i32,
3031           .fniv = tcg_gen_sari_vec,
3032           .fno = gen_helper_gvec_sar32i,
3033           .opt_opc = vecop_list,
3034           .vece = MO_32 },
3035         { .fni8 = tcg_gen_sari_i64,
3036           .fniv = tcg_gen_sari_vec,
3037           .fno = gen_helper_gvec_sar64i,
3038           .opt_opc = vecop_list,
3039           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3040           .vece = MO_64 },
3041     };
3042 
3043     tcg_debug_assert(vece <= MO_64);
3044     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
3045     if (shift == 0) {
3046         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
3047     } else {
3048         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
3049     }
3050 }
3051 
3052 void tcg_gen_vec_rotl8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
3053 {
3054     uint64_t mask = dup_const(MO_8, 0xff << c);
3055 
3056     tcg_gen_shli_i64(d, a, c);
3057     tcg_gen_shri_i64(a, a, 8 - c);
3058     tcg_gen_andi_i64(d, d, mask);
3059     tcg_gen_andi_i64(a, a, ~mask);
3060     tcg_gen_or_i64(d, d, a);
3061 }
3062 
3063 void tcg_gen_vec_rotl16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
3064 {
3065     uint64_t mask = dup_const(MO_16, 0xffff << c);
3066 
3067     tcg_gen_shli_i64(d, a, c);
3068     tcg_gen_shri_i64(a, a, 16 - c);
3069     tcg_gen_andi_i64(d, d, mask);
3070     tcg_gen_andi_i64(a, a, ~mask);
3071     tcg_gen_or_i64(d, d, a);
3072 }
3073 
3074 void tcg_gen_gvec_rotli(unsigned vece, uint32_t dofs, uint32_t aofs,
3075                         int64_t shift, uint32_t oprsz, uint32_t maxsz)
3076 {
3077     static const TCGOpcode vecop_list[] = { INDEX_op_rotli_vec, 0 };
3078     static const GVecGen2i g[4] = {
3079         { .fni8 = tcg_gen_vec_rotl8i_i64,
3080           .fniv = tcg_gen_rotli_vec,
3081           .fno = gen_helper_gvec_rotl8i,
3082           .opt_opc = vecop_list,
3083           .vece = MO_8 },
3084         { .fni8 = tcg_gen_vec_rotl16i_i64,
3085           .fniv = tcg_gen_rotli_vec,
3086           .fno = gen_helper_gvec_rotl16i,
3087           .opt_opc = vecop_list,
3088           .vece = MO_16 },
3089         { .fni4 = tcg_gen_rotli_i32,
3090           .fniv = tcg_gen_rotli_vec,
3091           .fno = gen_helper_gvec_rotl32i,
3092           .opt_opc = vecop_list,
3093           .vece = MO_32 },
3094         { .fni8 = tcg_gen_rotli_i64,
3095           .fniv = tcg_gen_rotli_vec,
3096           .fno = gen_helper_gvec_rotl64i,
3097           .opt_opc = vecop_list,
3098           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3099           .vece = MO_64 },
3100     };
3101 
3102     tcg_debug_assert(vece <= MO_64);
3103     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
3104     if (shift == 0) {
3105         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
3106     } else {
3107         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
3108     }
3109 }
3110 
3111 void tcg_gen_gvec_rotri(unsigned vece, uint32_t dofs, uint32_t aofs,
3112                         int64_t shift, uint32_t oprsz, uint32_t maxsz)
3113 {
3114     tcg_debug_assert(vece <= MO_64);
3115     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
3116     tcg_gen_gvec_rotli(vece, dofs, aofs, -shift & ((8 << vece) - 1),
3117                        oprsz, maxsz);
3118 }
3119 
3120 /*
3121  * Specialized generation vector shifts by a non-constant scalar.
3122  */
3123 
3124 typedef struct {
3125     void (*fni4)(TCGv_i32, TCGv_i32, TCGv_i32);
3126     void (*fni8)(TCGv_i64, TCGv_i64, TCGv_i64);
3127     void (*fniv_s)(unsigned, TCGv_vec, TCGv_vec, TCGv_i32);
3128     void (*fniv_v)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec);
3129     gen_helper_gvec_2 *fno[4];
3130     TCGOpcode s_list[2];
3131     TCGOpcode v_list[2];
3132 } GVecGen2sh;
3133 
3134 static void expand_2sh_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
3135                            uint32_t oprsz, uint32_t tysz, TCGType type,
3136                            TCGv_i32 shift,
3137                            void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_i32))
3138 {
3139     TCGv_vec t0 = tcg_temp_new_vec(type);
3140     uint32_t i;
3141 
3142     for (i = 0; i < oprsz; i += tysz) {
3143         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
3144         fni(vece, t0, t0, shift);
3145         tcg_gen_st_vec(t0, cpu_env, dofs + i);
3146     }
3147     tcg_temp_free_vec(t0);
3148 }
3149 
3150 static void
3151 do_gvec_shifts(unsigned vece, uint32_t dofs, uint32_t aofs, TCGv_i32 shift,
3152                uint32_t oprsz, uint32_t maxsz, const GVecGen2sh *g)
3153 {
3154     TCGType type;
3155     uint32_t some;
3156 
3157     check_size_align(oprsz, maxsz, dofs | aofs);
3158     check_overlap_2(dofs, aofs, maxsz);
3159 
3160     /* If the backend has a scalar expansion, great.  */
3161     type = choose_vector_type(g->s_list, vece, oprsz, vece == MO_64);
3162     if (type) {
3163         const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL);
3164         switch (type) {
3165         case TCG_TYPE_V256:
3166             some = QEMU_ALIGN_DOWN(oprsz, 32);
3167             expand_2sh_vec(vece, dofs, aofs, some, 32,
3168                            TCG_TYPE_V256, shift, g->fniv_s);
3169             if (some == oprsz) {
3170                 break;
3171             }
3172             dofs += some;
3173             aofs += some;
3174             oprsz -= some;
3175             maxsz -= some;
3176             /* fallthru */
3177         case TCG_TYPE_V128:
3178             expand_2sh_vec(vece, dofs, aofs, oprsz, 16,
3179                            TCG_TYPE_V128, shift, g->fniv_s);
3180             break;
3181         case TCG_TYPE_V64:
3182             expand_2sh_vec(vece, dofs, aofs, oprsz, 8,
3183                            TCG_TYPE_V64, shift, g->fniv_s);
3184             break;
3185         default:
3186             g_assert_not_reached();
3187         }
3188         tcg_swap_vecop_list(hold_list);
3189         goto clear_tail;
3190     }
3191 
3192     /* If the backend supports variable vector shifts, also cool.  */
3193     type = choose_vector_type(g->v_list, vece, oprsz, vece == MO_64);
3194     if (type) {
3195         const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL);
3196         TCGv_vec v_shift = tcg_temp_new_vec(type);
3197 
3198         if (vece == MO_64) {
3199             TCGv_i64 sh64 = tcg_temp_ebb_new_i64();
3200             tcg_gen_extu_i32_i64(sh64, shift);
3201             tcg_gen_dup_i64_vec(MO_64, v_shift, sh64);
3202             tcg_temp_free_i64(sh64);
3203         } else {
3204             tcg_gen_dup_i32_vec(vece, v_shift, shift);
3205         }
3206 
3207         switch (type) {
3208         case TCG_TYPE_V256:
3209             some = QEMU_ALIGN_DOWN(oprsz, 32);
3210             expand_2s_vec(vece, dofs, aofs, some, 32, TCG_TYPE_V256,
3211                           v_shift, false, g->fniv_v);
3212             if (some == oprsz) {
3213                 break;
3214             }
3215             dofs += some;
3216             aofs += some;
3217             oprsz -= some;
3218             maxsz -= some;
3219             /* fallthru */
3220         case TCG_TYPE_V128:
3221             expand_2s_vec(vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
3222                           v_shift, false, g->fniv_v);
3223             break;
3224         case TCG_TYPE_V64:
3225             expand_2s_vec(vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
3226                           v_shift, false, g->fniv_v);
3227             break;
3228         default:
3229             g_assert_not_reached();
3230         }
3231         tcg_temp_free_vec(v_shift);
3232         tcg_swap_vecop_list(hold_list);
3233         goto clear_tail;
3234     }
3235 
3236     /* Otherwise fall back to integral... */
3237     if (vece == MO_32 && check_size_impl(oprsz, 4)) {
3238         expand_2s_i32(dofs, aofs, oprsz, shift, false, g->fni4);
3239     } else if (vece == MO_64 && check_size_impl(oprsz, 8)) {
3240         TCGv_i64 sh64 = tcg_temp_ebb_new_i64();
3241         tcg_gen_extu_i32_i64(sh64, shift);
3242         expand_2s_i64(dofs, aofs, oprsz, sh64, false, g->fni8);
3243         tcg_temp_free_i64(sh64);
3244     } else {
3245         TCGv_ptr a0 = tcg_temp_ebb_new_ptr();
3246         TCGv_ptr a1 = tcg_temp_ebb_new_ptr();
3247         TCGv_i32 desc = tcg_temp_ebb_new_i32();
3248 
3249         tcg_gen_shli_i32(desc, shift, SIMD_DATA_SHIFT);
3250         tcg_gen_ori_i32(desc, desc, simd_desc(oprsz, maxsz, 0));
3251         tcg_gen_addi_ptr(a0, cpu_env, dofs);
3252         tcg_gen_addi_ptr(a1, cpu_env, aofs);
3253 
3254         g->fno[vece](a0, a1, desc);
3255 
3256         tcg_temp_free_ptr(a0);
3257         tcg_temp_free_ptr(a1);
3258         tcg_temp_free_i32(desc);
3259         return;
3260     }
3261 
3262  clear_tail:
3263     if (oprsz < maxsz) {
3264         expand_clr(dofs + oprsz, maxsz - oprsz);
3265     }
3266 }
3267 
3268 void tcg_gen_gvec_shls(unsigned vece, uint32_t dofs, uint32_t aofs,
3269                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3270 {
3271     static const GVecGen2sh g = {
3272         .fni4 = tcg_gen_shl_i32,
3273         .fni8 = tcg_gen_shl_i64,
3274         .fniv_s = tcg_gen_shls_vec,
3275         .fniv_v = tcg_gen_shlv_vec,
3276         .fno = {
3277             gen_helper_gvec_shl8i,
3278             gen_helper_gvec_shl16i,
3279             gen_helper_gvec_shl32i,
3280             gen_helper_gvec_shl64i,
3281         },
3282         .s_list = { INDEX_op_shls_vec, 0 },
3283         .v_list = { INDEX_op_shlv_vec, 0 },
3284     };
3285 
3286     tcg_debug_assert(vece <= MO_64);
3287     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3288 }
3289 
3290 void tcg_gen_gvec_shrs(unsigned vece, uint32_t dofs, uint32_t aofs,
3291                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3292 {
3293     static const GVecGen2sh g = {
3294         .fni4 = tcg_gen_shr_i32,
3295         .fni8 = tcg_gen_shr_i64,
3296         .fniv_s = tcg_gen_shrs_vec,
3297         .fniv_v = tcg_gen_shrv_vec,
3298         .fno = {
3299             gen_helper_gvec_shr8i,
3300             gen_helper_gvec_shr16i,
3301             gen_helper_gvec_shr32i,
3302             gen_helper_gvec_shr64i,
3303         },
3304         .s_list = { INDEX_op_shrs_vec, 0 },
3305         .v_list = { INDEX_op_shrv_vec, 0 },
3306     };
3307 
3308     tcg_debug_assert(vece <= MO_64);
3309     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3310 }
3311 
3312 void tcg_gen_gvec_sars(unsigned vece, uint32_t dofs, uint32_t aofs,
3313                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3314 {
3315     static const GVecGen2sh g = {
3316         .fni4 = tcg_gen_sar_i32,
3317         .fni8 = tcg_gen_sar_i64,
3318         .fniv_s = tcg_gen_sars_vec,
3319         .fniv_v = tcg_gen_sarv_vec,
3320         .fno = {
3321             gen_helper_gvec_sar8i,
3322             gen_helper_gvec_sar16i,
3323             gen_helper_gvec_sar32i,
3324             gen_helper_gvec_sar64i,
3325         },
3326         .s_list = { INDEX_op_sars_vec, 0 },
3327         .v_list = { INDEX_op_sarv_vec, 0 },
3328     };
3329 
3330     tcg_debug_assert(vece <= MO_64);
3331     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3332 }
3333 
3334 void tcg_gen_gvec_rotls(unsigned vece, uint32_t dofs, uint32_t aofs,
3335                         TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3336 {
3337     static const GVecGen2sh g = {
3338         .fni4 = tcg_gen_rotl_i32,
3339         .fni8 = tcg_gen_rotl_i64,
3340         .fniv_s = tcg_gen_rotls_vec,
3341         .fniv_v = tcg_gen_rotlv_vec,
3342         .fno = {
3343             gen_helper_gvec_rotl8i,
3344             gen_helper_gvec_rotl16i,
3345             gen_helper_gvec_rotl32i,
3346             gen_helper_gvec_rotl64i,
3347         },
3348         .s_list = { INDEX_op_rotls_vec, 0 },
3349         .v_list = { INDEX_op_rotlv_vec, 0 },
3350     };
3351 
3352     tcg_debug_assert(vece <= MO_64);
3353     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3354 }
3355 
3356 void tcg_gen_gvec_rotrs(unsigned vece, uint32_t dofs, uint32_t aofs,
3357                         TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3358 {
3359     TCGv_i32 tmp = tcg_temp_ebb_new_i32();
3360 
3361     tcg_gen_neg_i32(tmp, shift);
3362     tcg_gen_andi_i32(tmp, tmp, (8 << vece) - 1);
3363     tcg_gen_gvec_rotls(vece, dofs, aofs, tmp, oprsz, maxsz);
3364     tcg_temp_free_i32(tmp);
3365 }
3366 
3367 /*
3368  * Expand D = A << (B % element bits)
3369  *
3370  * Unlike scalar shifts, where it is easy for the target front end
3371  * to include the modulo as part of the expansion.  If the target
3372  * naturally includes the modulo as part of the operation, great!
3373  * If the target has some other behaviour from out-of-range shifts,
3374  * then it could not use this function anyway, and would need to
3375  * do it's own expansion with custom functions.
3376  */
3377 static void tcg_gen_shlv_mod_vec(unsigned vece, TCGv_vec d,
3378                                  TCGv_vec a, TCGv_vec b)
3379 {
3380     TCGv_vec t = tcg_temp_new_vec_matching(d);
3381     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3382 
3383     tcg_gen_and_vec(vece, t, b, m);
3384     tcg_gen_shlv_vec(vece, d, a, t);
3385     tcg_temp_free_vec(t);
3386 }
3387 
3388 static void tcg_gen_shl_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3389 {
3390     TCGv_i32 t = tcg_temp_ebb_new_i32();
3391 
3392     tcg_gen_andi_i32(t, b, 31);
3393     tcg_gen_shl_i32(d, a, t);
3394     tcg_temp_free_i32(t);
3395 }
3396 
3397 static void tcg_gen_shl_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3398 {
3399     TCGv_i64 t = tcg_temp_ebb_new_i64();
3400 
3401     tcg_gen_andi_i64(t, b, 63);
3402     tcg_gen_shl_i64(d, a, t);
3403     tcg_temp_free_i64(t);
3404 }
3405 
3406 void tcg_gen_gvec_shlv(unsigned vece, uint32_t dofs, uint32_t aofs,
3407                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3408 {
3409     static const TCGOpcode vecop_list[] = { INDEX_op_shlv_vec, 0 };
3410     static const GVecGen3 g[4] = {
3411         { .fniv = tcg_gen_shlv_mod_vec,
3412           .fno = gen_helper_gvec_shl8v,
3413           .opt_opc = vecop_list,
3414           .vece = MO_8 },
3415         { .fniv = tcg_gen_shlv_mod_vec,
3416           .fno = gen_helper_gvec_shl16v,
3417           .opt_opc = vecop_list,
3418           .vece = MO_16 },
3419         { .fni4 = tcg_gen_shl_mod_i32,
3420           .fniv = tcg_gen_shlv_mod_vec,
3421           .fno = gen_helper_gvec_shl32v,
3422           .opt_opc = vecop_list,
3423           .vece = MO_32 },
3424         { .fni8 = tcg_gen_shl_mod_i64,
3425           .fniv = tcg_gen_shlv_mod_vec,
3426           .fno = gen_helper_gvec_shl64v,
3427           .opt_opc = vecop_list,
3428           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3429           .vece = MO_64 },
3430     };
3431 
3432     tcg_debug_assert(vece <= MO_64);
3433     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3434 }
3435 
3436 /*
3437  * Similarly for logical right shifts.
3438  */
3439 
3440 static void tcg_gen_shrv_mod_vec(unsigned vece, TCGv_vec d,
3441                                  TCGv_vec a, TCGv_vec b)
3442 {
3443     TCGv_vec t = tcg_temp_new_vec_matching(d);
3444     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3445 
3446     tcg_gen_and_vec(vece, t, b, m);
3447     tcg_gen_shrv_vec(vece, d, a, t);
3448     tcg_temp_free_vec(t);
3449 }
3450 
3451 static void tcg_gen_shr_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3452 {
3453     TCGv_i32 t = tcg_temp_ebb_new_i32();
3454 
3455     tcg_gen_andi_i32(t, b, 31);
3456     tcg_gen_shr_i32(d, a, t);
3457     tcg_temp_free_i32(t);
3458 }
3459 
3460 static void tcg_gen_shr_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3461 {
3462     TCGv_i64 t = tcg_temp_ebb_new_i64();
3463 
3464     tcg_gen_andi_i64(t, b, 63);
3465     tcg_gen_shr_i64(d, a, t);
3466     tcg_temp_free_i64(t);
3467 }
3468 
3469 void tcg_gen_gvec_shrv(unsigned vece, uint32_t dofs, uint32_t aofs,
3470                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3471 {
3472     static const TCGOpcode vecop_list[] = { INDEX_op_shrv_vec, 0 };
3473     static const GVecGen3 g[4] = {
3474         { .fniv = tcg_gen_shrv_mod_vec,
3475           .fno = gen_helper_gvec_shr8v,
3476           .opt_opc = vecop_list,
3477           .vece = MO_8 },
3478         { .fniv = tcg_gen_shrv_mod_vec,
3479           .fno = gen_helper_gvec_shr16v,
3480           .opt_opc = vecop_list,
3481           .vece = MO_16 },
3482         { .fni4 = tcg_gen_shr_mod_i32,
3483           .fniv = tcg_gen_shrv_mod_vec,
3484           .fno = gen_helper_gvec_shr32v,
3485           .opt_opc = vecop_list,
3486           .vece = MO_32 },
3487         { .fni8 = tcg_gen_shr_mod_i64,
3488           .fniv = tcg_gen_shrv_mod_vec,
3489           .fno = gen_helper_gvec_shr64v,
3490           .opt_opc = vecop_list,
3491           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3492           .vece = MO_64 },
3493     };
3494 
3495     tcg_debug_assert(vece <= MO_64);
3496     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3497 }
3498 
3499 /*
3500  * Similarly for arithmetic right shifts.
3501  */
3502 
3503 static void tcg_gen_sarv_mod_vec(unsigned vece, TCGv_vec d,
3504                                  TCGv_vec a, TCGv_vec b)
3505 {
3506     TCGv_vec t = tcg_temp_new_vec_matching(d);
3507     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3508 
3509     tcg_gen_and_vec(vece, t, b, m);
3510     tcg_gen_sarv_vec(vece, d, a, t);
3511     tcg_temp_free_vec(t);
3512 }
3513 
3514 static void tcg_gen_sar_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3515 {
3516     TCGv_i32 t = tcg_temp_ebb_new_i32();
3517 
3518     tcg_gen_andi_i32(t, b, 31);
3519     tcg_gen_sar_i32(d, a, t);
3520     tcg_temp_free_i32(t);
3521 }
3522 
3523 static void tcg_gen_sar_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3524 {
3525     TCGv_i64 t = tcg_temp_ebb_new_i64();
3526 
3527     tcg_gen_andi_i64(t, b, 63);
3528     tcg_gen_sar_i64(d, a, t);
3529     tcg_temp_free_i64(t);
3530 }
3531 
3532 void tcg_gen_gvec_sarv(unsigned vece, uint32_t dofs, uint32_t aofs,
3533                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3534 {
3535     static const TCGOpcode vecop_list[] = { INDEX_op_sarv_vec, 0 };
3536     static const GVecGen3 g[4] = {
3537         { .fniv = tcg_gen_sarv_mod_vec,
3538           .fno = gen_helper_gvec_sar8v,
3539           .opt_opc = vecop_list,
3540           .vece = MO_8 },
3541         { .fniv = tcg_gen_sarv_mod_vec,
3542           .fno = gen_helper_gvec_sar16v,
3543           .opt_opc = vecop_list,
3544           .vece = MO_16 },
3545         { .fni4 = tcg_gen_sar_mod_i32,
3546           .fniv = tcg_gen_sarv_mod_vec,
3547           .fno = gen_helper_gvec_sar32v,
3548           .opt_opc = vecop_list,
3549           .vece = MO_32 },
3550         { .fni8 = tcg_gen_sar_mod_i64,
3551           .fniv = tcg_gen_sarv_mod_vec,
3552           .fno = gen_helper_gvec_sar64v,
3553           .opt_opc = vecop_list,
3554           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3555           .vece = MO_64 },
3556     };
3557 
3558     tcg_debug_assert(vece <= MO_64);
3559     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3560 }
3561 
3562 /*
3563  * Similarly for rotates.
3564  */
3565 
3566 static void tcg_gen_rotlv_mod_vec(unsigned vece, TCGv_vec d,
3567                                   TCGv_vec a, TCGv_vec b)
3568 {
3569     TCGv_vec t = tcg_temp_new_vec_matching(d);
3570     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3571 
3572     tcg_gen_and_vec(vece, t, b, m);
3573     tcg_gen_rotlv_vec(vece, d, a, t);
3574     tcg_temp_free_vec(t);
3575 }
3576 
3577 static void tcg_gen_rotl_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3578 {
3579     TCGv_i32 t = tcg_temp_ebb_new_i32();
3580 
3581     tcg_gen_andi_i32(t, b, 31);
3582     tcg_gen_rotl_i32(d, a, t);
3583     tcg_temp_free_i32(t);
3584 }
3585 
3586 static void tcg_gen_rotl_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3587 {
3588     TCGv_i64 t = tcg_temp_ebb_new_i64();
3589 
3590     tcg_gen_andi_i64(t, b, 63);
3591     tcg_gen_rotl_i64(d, a, t);
3592     tcg_temp_free_i64(t);
3593 }
3594 
3595 void tcg_gen_gvec_rotlv(unsigned vece, uint32_t dofs, uint32_t aofs,
3596                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3597 {
3598     static const TCGOpcode vecop_list[] = { INDEX_op_rotlv_vec, 0 };
3599     static const GVecGen3 g[4] = {
3600         { .fniv = tcg_gen_rotlv_mod_vec,
3601           .fno = gen_helper_gvec_rotl8v,
3602           .opt_opc = vecop_list,
3603           .vece = MO_8 },
3604         { .fniv = tcg_gen_rotlv_mod_vec,
3605           .fno = gen_helper_gvec_rotl16v,
3606           .opt_opc = vecop_list,
3607           .vece = MO_16 },
3608         { .fni4 = tcg_gen_rotl_mod_i32,
3609           .fniv = tcg_gen_rotlv_mod_vec,
3610           .fno = gen_helper_gvec_rotl32v,
3611           .opt_opc = vecop_list,
3612           .vece = MO_32 },
3613         { .fni8 = tcg_gen_rotl_mod_i64,
3614           .fniv = tcg_gen_rotlv_mod_vec,
3615           .fno = gen_helper_gvec_rotl64v,
3616           .opt_opc = vecop_list,
3617           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3618           .vece = MO_64 },
3619     };
3620 
3621     tcg_debug_assert(vece <= MO_64);
3622     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3623 }
3624 
3625 static void tcg_gen_rotrv_mod_vec(unsigned vece, TCGv_vec d,
3626                                   TCGv_vec a, TCGv_vec b)
3627 {
3628     TCGv_vec t = tcg_temp_new_vec_matching(d);
3629     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3630 
3631     tcg_gen_and_vec(vece, t, b, m);
3632     tcg_gen_rotrv_vec(vece, d, a, t);
3633     tcg_temp_free_vec(t);
3634 }
3635 
3636 static void tcg_gen_rotr_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3637 {
3638     TCGv_i32 t = tcg_temp_ebb_new_i32();
3639 
3640     tcg_gen_andi_i32(t, b, 31);
3641     tcg_gen_rotr_i32(d, a, t);
3642     tcg_temp_free_i32(t);
3643 }
3644 
3645 static void tcg_gen_rotr_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3646 {
3647     TCGv_i64 t = tcg_temp_ebb_new_i64();
3648 
3649     tcg_gen_andi_i64(t, b, 63);
3650     tcg_gen_rotr_i64(d, a, t);
3651     tcg_temp_free_i64(t);
3652 }
3653 
3654 void tcg_gen_gvec_rotrv(unsigned vece, uint32_t dofs, uint32_t aofs,
3655                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3656 {
3657     static const TCGOpcode vecop_list[] = { INDEX_op_rotrv_vec, 0 };
3658     static const GVecGen3 g[4] = {
3659         { .fniv = tcg_gen_rotrv_mod_vec,
3660           .fno = gen_helper_gvec_rotr8v,
3661           .opt_opc = vecop_list,
3662           .vece = MO_8 },
3663         { .fniv = tcg_gen_rotrv_mod_vec,
3664           .fno = gen_helper_gvec_rotr16v,
3665           .opt_opc = vecop_list,
3666           .vece = MO_16 },
3667         { .fni4 = tcg_gen_rotr_mod_i32,
3668           .fniv = tcg_gen_rotrv_mod_vec,
3669           .fno = gen_helper_gvec_rotr32v,
3670           .opt_opc = vecop_list,
3671           .vece = MO_32 },
3672         { .fni8 = tcg_gen_rotr_mod_i64,
3673           .fniv = tcg_gen_rotrv_mod_vec,
3674           .fno = gen_helper_gvec_rotr64v,
3675           .opt_opc = vecop_list,
3676           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3677           .vece = MO_64 },
3678     };
3679 
3680     tcg_debug_assert(vece <= MO_64);
3681     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3682 }
3683 
3684 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
3685 static void expand_cmp_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
3686                            uint32_t oprsz, TCGCond cond)
3687 {
3688     TCGv_i32 t0 = tcg_temp_ebb_new_i32();
3689     TCGv_i32 t1 = tcg_temp_ebb_new_i32();
3690     uint32_t i;
3691 
3692     for (i = 0; i < oprsz; i += 4) {
3693         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
3694         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
3695         tcg_gen_setcond_i32(cond, t0, t0, t1);
3696         tcg_gen_neg_i32(t0, t0);
3697         tcg_gen_st_i32(t0, cpu_env, dofs + i);
3698     }
3699     tcg_temp_free_i32(t1);
3700     tcg_temp_free_i32(t0);
3701 }
3702 
3703 static void expand_cmp_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
3704                            uint32_t oprsz, TCGCond cond)
3705 {
3706     TCGv_i64 t0 = tcg_temp_ebb_new_i64();
3707     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
3708     uint32_t i;
3709 
3710     for (i = 0; i < oprsz; i += 8) {
3711         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
3712         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
3713         tcg_gen_setcond_i64(cond, t0, t0, t1);
3714         tcg_gen_neg_i64(t0, t0);
3715         tcg_gen_st_i64(t0, cpu_env, dofs + i);
3716     }
3717     tcg_temp_free_i64(t1);
3718     tcg_temp_free_i64(t0);
3719 }
3720 
3721 static void expand_cmp_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
3722                            uint32_t bofs, uint32_t oprsz, uint32_t tysz,
3723                            TCGType type, TCGCond cond)
3724 {
3725     TCGv_vec t0 = tcg_temp_new_vec(type);
3726     TCGv_vec t1 = tcg_temp_new_vec(type);
3727     uint32_t i;
3728 
3729     for (i = 0; i < oprsz; i += tysz) {
3730         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
3731         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
3732         tcg_gen_cmp_vec(cond, vece, t0, t0, t1);
3733         tcg_gen_st_vec(t0, cpu_env, dofs + i);
3734     }
3735     tcg_temp_free_vec(t1);
3736     tcg_temp_free_vec(t0);
3737 }
3738 
3739 void tcg_gen_gvec_cmp(TCGCond cond, unsigned vece, uint32_t dofs,
3740                       uint32_t aofs, uint32_t bofs,
3741                       uint32_t oprsz, uint32_t maxsz)
3742 {
3743     static const TCGOpcode cmp_list[] = { INDEX_op_cmp_vec, 0 };
3744     static gen_helper_gvec_3 * const eq_fn[4] = {
3745         gen_helper_gvec_eq8, gen_helper_gvec_eq16,
3746         gen_helper_gvec_eq32, gen_helper_gvec_eq64
3747     };
3748     static gen_helper_gvec_3 * const ne_fn[4] = {
3749         gen_helper_gvec_ne8, gen_helper_gvec_ne16,
3750         gen_helper_gvec_ne32, gen_helper_gvec_ne64
3751     };
3752     static gen_helper_gvec_3 * const lt_fn[4] = {
3753         gen_helper_gvec_lt8, gen_helper_gvec_lt16,
3754         gen_helper_gvec_lt32, gen_helper_gvec_lt64
3755     };
3756     static gen_helper_gvec_3 * const le_fn[4] = {
3757         gen_helper_gvec_le8, gen_helper_gvec_le16,
3758         gen_helper_gvec_le32, gen_helper_gvec_le64
3759     };
3760     static gen_helper_gvec_3 * const ltu_fn[4] = {
3761         gen_helper_gvec_ltu8, gen_helper_gvec_ltu16,
3762         gen_helper_gvec_ltu32, gen_helper_gvec_ltu64
3763     };
3764     static gen_helper_gvec_3 * const leu_fn[4] = {
3765         gen_helper_gvec_leu8, gen_helper_gvec_leu16,
3766         gen_helper_gvec_leu32, gen_helper_gvec_leu64
3767     };
3768     static gen_helper_gvec_3 * const * const fns[16] = {
3769         [TCG_COND_EQ] = eq_fn,
3770         [TCG_COND_NE] = ne_fn,
3771         [TCG_COND_LT] = lt_fn,
3772         [TCG_COND_LE] = le_fn,
3773         [TCG_COND_LTU] = ltu_fn,
3774         [TCG_COND_LEU] = leu_fn,
3775     };
3776 
3777     const TCGOpcode *hold_list;
3778     TCGType type;
3779     uint32_t some;
3780 
3781     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
3782     check_overlap_3(dofs, aofs, bofs, maxsz);
3783 
3784     if (cond == TCG_COND_NEVER || cond == TCG_COND_ALWAYS) {
3785         do_dup(MO_8, dofs, oprsz, maxsz,
3786                NULL, NULL, -(cond == TCG_COND_ALWAYS));
3787         return;
3788     }
3789 
3790     /*
3791      * Implement inline with a vector type, if possible.
3792      * Prefer integer when 64-bit host and 64-bit comparison.
3793      */
3794     hold_list = tcg_swap_vecop_list(cmp_list);
3795     type = choose_vector_type(cmp_list, vece, oprsz,
3796                               TCG_TARGET_REG_BITS == 64 && vece == MO_64);
3797     switch (type) {
3798     case TCG_TYPE_V256:
3799         /* Recall that ARM SVE allows vector sizes that are not a
3800          * power of 2, but always a multiple of 16.  The intent is
3801          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
3802          */
3803         some = QEMU_ALIGN_DOWN(oprsz, 32);
3804         expand_cmp_vec(vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256, cond);
3805         if (some == oprsz) {
3806             break;
3807         }
3808         dofs += some;
3809         aofs += some;
3810         bofs += some;
3811         oprsz -= some;
3812         maxsz -= some;
3813         /* fallthru */
3814     case TCG_TYPE_V128:
3815         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128, cond);
3816         break;
3817     case TCG_TYPE_V64:
3818         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64, cond);
3819         break;
3820 
3821     case 0:
3822         if (vece == MO_64 && check_size_impl(oprsz, 8)) {
3823             expand_cmp_i64(dofs, aofs, bofs, oprsz, cond);
3824         } else if (vece == MO_32 && check_size_impl(oprsz, 4)) {
3825             expand_cmp_i32(dofs, aofs, bofs, oprsz, cond);
3826         } else {
3827             gen_helper_gvec_3 * const *fn = fns[cond];
3828 
3829             if (fn == NULL) {
3830                 uint32_t tmp;
3831                 tmp = aofs, aofs = bofs, bofs = tmp;
3832                 cond = tcg_swap_cond(cond);
3833                 fn = fns[cond];
3834                 assert(fn != NULL);
3835             }
3836             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, 0, fn[vece]);
3837             oprsz = maxsz;
3838         }
3839         break;
3840 
3841     default:
3842         g_assert_not_reached();
3843     }
3844     tcg_swap_vecop_list(hold_list);
3845 
3846     if (oprsz < maxsz) {
3847         expand_clr(dofs + oprsz, maxsz - oprsz);
3848     }
3849 }
3850 
3851 static void tcg_gen_bitsel_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 c)
3852 {
3853     TCGv_i64 t = tcg_temp_ebb_new_i64();
3854 
3855     tcg_gen_and_i64(t, b, a);
3856     tcg_gen_andc_i64(d, c, a);
3857     tcg_gen_or_i64(d, d, t);
3858     tcg_temp_free_i64(t);
3859 }
3860 
3861 void tcg_gen_gvec_bitsel(unsigned vece, uint32_t dofs, uint32_t aofs,
3862                          uint32_t bofs, uint32_t cofs,
3863                          uint32_t oprsz, uint32_t maxsz)
3864 {
3865     static const GVecGen4 g = {
3866         .fni8 = tcg_gen_bitsel_i64,
3867         .fniv = tcg_gen_bitsel_vec,
3868         .fno = gen_helper_gvec_bitsel,
3869     };
3870 
3871     tcg_gen_gvec_4(dofs, aofs, bofs, cofs, oprsz, maxsz, &g);
3872 }
3873